feat(plugins): workspace_plugins tracking table — version-subscription foundation

Closes core#113 partial. Adds the DB foundation for the version-subscription model. Drift detection + queue + admin apply endpoint are follow-up scope (separate PR; filed as a new issue). WHY THIS PR ONLY GETS US PART-WAY Plugin install state today is filesystem-only — '/configs/plugins/<name>/' inside the container. There's no DB record of 'plugin X installed at workspace W from source S, tracking ref T'. That makes drift detection impossible: nothing to compare upstream tags against. This PR adds the table + the install-endpoint hook that writes to it. With baseline tags now on every plugin (post internal#92), the table starts collecting tracked-ref values immediately on the next install. The actual drift-check job + queue + apply endpoint layer on top. WHAT THIS ADDS workspace_plugins table: workspace_id FK → workspaces(id) ON DELETE CASCADE plugin_name canonical name from plugin.yaml source_raw full source URL the install used tracked_ref 'none' | 'tag:vX.Y.Z' | 'tag:latest' | 'sha:<full>' installed_at, updated_at installRequest gains optional 'track' field (defaults to 'none'). Install handler upserts the workspace_plugins row after delivery succeeds. DB write failure is logged but doesn't fail the install (the plugin IS in the container; surfacing 500 misleads the caller). validateTrackedRef enforces the closed set of accepted shapes: 'none' | 'tag:<non-empty>' | 'sha:<non-empty>' Bare values like 'latest' / 'main' / version-strings without prefix are rejected — the drift detector keys on prefix to know what kind of resolution to do. WHAT THIS DOES NOT ADD (filed separately) - Drift detector job (cron / on-demand) that scans 'WHERE tracked_ref != none' rows and queues updates on upstream drift - plugin_update_queue table (separate migration once detector lands) - GET /admin/plugin-updates-pending and POST .../apply endpoints - Tier-aware apply (core#115 — composes here) PHASE 4 SELF-REVIEW (FIVE-AXIS) Correctness: No finding — install endpoint behavior unchanged for callers that don't pass 'track'. DB write is best-effort + logged on failure. validateTrackedRef rejects ambiguous bare strings. Readability: No finding — separate file plugins_tracking.go isolates the new concern; install handler delta is a single 4-line block. Architecture: No finding — additive table; existing schema untouched. Migration 20260508160000_* uses the timestamp-prefixed convention. Security: No finding — INSERT params via placeholders (no string interpolation). validateTrackedRef rejects unexpected shapes before the column constraint would. Performance: No finding — one extra ExecContext per install. Install is already seconds-scale (network fetch + tar + docker exec); rounds to noise. TESTS (1 new, all green) TestValidateTrackedRef — pin closed set + structural validators REFS core#113 — this issue (foundation only; drift+queue+apply = follow-up) internal#92, internal#93 — plugin/template baseline tags (now exists for tracking) core#114 — atomic install (this PR composes — no atomicity regression) core#115 — canary tier filter (will key off the same DB foundation) Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Merge pull request 'feat(plugins): hot-reload classifier — skip restart on SKILL-content-only updates' (#121 ) from feat/plugin-hot-reload-classifier into main
2026-05-08 08:52:35 -07:00 · 2026-05-08 15:26:32 +00:00 · 2026-05-08 08:26:05 -07:00 · 2026-05-08 15:26:00 +00:00 · 2026-05-08 15:24:55 +00:00 · 2026-05-08 15:23:31 +00:00
393 changed files with 49639 additions and 6482 deletions
@@ -37,7 +37,7 @@ CANONICAL_FILE = Path(".github/workflows/secret-scan.yml")
 CONSUMERS: list[tuple[str, str]] = [
    (
        "molecule-ai-workspace-runtime/molecule_runtime/scripts/pre-commit-checks.sh",
-        "https://raw.githubusercontent.com/Molecule-AI/molecule-ai-workspace-runtime/main/molecule_runtime/scripts/pre-commit-checks.sh",
+        "https://git.moleculesai.app/molecule-ai/molecule-ai-workspace-runtime/raw/branch/main/molecule_runtime/scripts/pre-commit-checks.sh",
    ),
 ]

@@ -1,408 +0,0 @@
-name: Auto-promote :latest after main image build
-
-# Retags `ghcr.io/molecule-ai/{platform,platform-tenant}:staging-<sha>`
-# → `:latest` after either the image build or E2E completes on a `main`
-# push, gated on E2E Staging SaaS not being red for that SHA.
-#
-# Why two triggers:
-#
-#   `publish-workspace-server-image` and `e2e-staging-saas` are both
-#   paths-filtered, but with DIFFERENT path sets:
-#
-#     publish-workspace-server-image:
-#       workspace-server/**, canvas/**, manifest.json
-#
-#     e2e-staging-saas (full lifecycle):
-#       workspace-server/internal/handlers/{registry,workspace_provision,
-#       a2a_proxy}.go, workspace-server/internal/middleware/**,
-#       workspace-server/internal/provisioner/**, tests/e2e/test_staging_full_saas.sh
-#
-#   The E2E set is a strict SUBSET of the publish set. So:
-#     - canvas/** changes → publish fires, E2E does not
-#     - workspace-server/cmd/** changes → publish fires, E2E does not
-#     - workspace-server/internal/sweep/** → publish fires, E2E does not
-#
-#   The previous version triggered ONLY on E2E completion, which meant
-#   non-E2E-path changes (canvas, cmd, sweep, etc.) rebuilt the image
-#   but never advanced `:latest`. Result: as of 2026-04-28 this workflow
-#   had run zero times since merge despite eight main pushes — `:latest`
-#   was ~7 hours / 9 PRs behind main with no human realising. See
-#   `molecule-core` Slack discussion 2026-04-28.
-#
-#   Adding `publish-workspace-server-image` as a second trigger closes
-#   the gap: any image rebuild on main eligibly advances `:latest`.
-#
-# Why E2E remains a kill-switch (not the trigger):
-#
-#   When E2E DID run for this SHA and ended red, we abort — `:latest`
-#   stays on the prior known-good digest. When E2E didn't run (paths
-#   filtered out), we proceed: pre-merge gates already validated this
-#   SHA on staging via auto-promote-staging requiring CI + E2E Canvas +
-#   E2E API + CodeQL all green. Image content for non-E2E-paths
-#   (canvas, cmd, sweep) is exercised by those staging gates.
-#
-# Why `main` only:
-#
-#   `:latest` is what prod tenants pull. We only want SHAs that have
-#   reached main (via auto-promote-staging) to advance `:latest`.
-#   Triggering on staging would let a staging-only revert advance
-#   `:latest` to a SHA that never reaches main, breaking the "production
-#   runs what's on main" invariant.
-#
-# Idempotency:
-#
-#   When a SHA touches paths that match BOTH publish and E2E, both
-#   workflows fire and complete. Both trigger this workflow on
-#   completion → two runs race. Both retag `:staging-<sha>` →
-#   `:latest`. crane tag is idempotent (re-tagging the same digest is a
-#   no-op), so the second run is harmless. concurrency group serializes
-#   them anyway.
-
-on:
-  workflow_run:
-    workflows:
-      - 'E2E Staging SaaS (full lifecycle)'
-      - 'publish-workspace-server-image'
-    types: [completed]
-    branches: [main]
-  workflow_dispatch:
-    inputs:
-      sha:
-        description: 'Short sha to promote (override; defaults to upstream workflow_run head_sha)'
-        required: false
-        type: string
-
-permissions:
-  contents: read
-  packages: write
-
-concurrency:
-  # Serialize promotes per-SHA so the publish+E2E both-fired race lands
-  # cleanly. Different SHAs can promote in parallel.
-  group: auto-promote-latest-${{ github.event.workflow_run.head_sha || github.event.inputs.sha || github.sha }}
-  cancel-in-progress: false
-
-env:
-  IMAGE_NAME: ghcr.io/molecule-ai/platform
-  TENANT_IMAGE_NAME: ghcr.io/molecule-ai/platform-tenant
-
-jobs:
-  promote:
-    # Proceed if upstream succeeded OR manual dispatch. Upstream-failure
-    # paths are filtered here; the E2E-was-red kill-switch lives in the
-    # gate-check step below (covers the case where upstream is publish
-    # success but E2E for the same SHA failed).
-    if: |
-      github.event_name == 'workflow_dispatch' ||
-      (github.event_name == 'workflow_run' && github.event.workflow_run.conclusion == 'success')
-    runs-on: ubuntu-latest
-    steps:
-      - name: Compute short sha
-        id: sha
-        run: |
-          set -euo pipefail
-          if [ -n "${{ github.event.inputs.sha }}" ]; then
-            FULL="${{ github.event.inputs.sha }}"
-          else
-            FULL="${{ github.event.workflow_run.head_sha }}"
-          fi
-          echo "short=${FULL:0:7}" >> "$GITHUB_OUTPUT"
-          echo "full=${FULL}" >> "$GITHUB_OUTPUT"
-
-      - name: Gate — E2E Staging SaaS state for this SHA
-        # When upstream IS E2E success, we know it's green (filtered by
-        # the job-level `if` already). When upstream is publish, look up
-        # E2E state for the same SHA. Four buckets:
-        #
-        #   - completed/success: E2E confirmed safe → proceed
-        #   - completed/failure|cancelled|timed_out: E2E found a
-        #     regression → ABORT (exit 1), `:latest` stays put
-        #   - in_progress|queued|requested: E2E is RACING with publish
-        #     for a runtime-touching SHA. publish typically completes
-        #     ~5-10min before E2E (~10-15min). If we promote on the
-        #     publish signal here, a later E2E failure can't roll back
-        #     `:latest` — it'd already be wrongly advanced. So we DEFER:
-        #     skip subsequent steps (proceed=false) and let E2E's own
-        #     completion event re-fire this workflow, which then takes
-        #     the upstream-is-E2E path. exit 0 so the run shows as
-        #     success rather than a noisy fake-failure.
-        #   - none/none: E2E was paths-filtered out for this SHA (the
-        #     change touched canvas/cmd/sweep/etc. — paths covered by
-        #     publish but not by E2E). pre-merge gates on staging
-        #     already validated this SHA → proceed.
-        #
-        # Manual dispatch skips this check — operator override.
-        id: gate
-        env:
-          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          REPO: ${{ github.repository }}
-          SHA: ${{ steps.sha.outputs.full }}
-          UPSTREAM_NAME: ${{ github.event.workflow_run.name }}
-          EVENT_NAME: ${{ github.event_name }}
-        run: |
-          set -euo pipefail
-
-          if [ "$EVENT_NAME" = "workflow_dispatch" ]; then
-            echo "proceed=true" >> "$GITHUB_OUTPUT"
-            echo "::notice::Manual dispatch — skipping E2E gate (operator override)"
-            exit 0
-          fi
-
-          if [ "$UPSTREAM_NAME" = "E2E Staging SaaS (full lifecycle)" ]; then
-            echo "proceed=true" >> "$GITHUB_OUTPUT"
-            echo "::notice::Upstream is E2E itself (success per job-level if) — gate trivially satisfied"
-            exit 0
-          fi
-
-          # Upstream is publish-workspace-server-image. Check E2E state.
-          # The jq filter must defend against TWO empty cases that gh
-          # CLI emits indistinguishably:
-          #   1. gh exits non-zero (network blip, auth issue) → handled
-          #      by the `|| echo "none/none"` fallback below.
-          #   2. gh exits zero but returns `[]` (no E2E run on this
-          #      main SHA — the common case for canvas-only / cmd-only
-          #      / sweep-only changes whose paths don't trigger E2E).
-          #      Without `(.[0] // {})`, jq sees `null` and emits
-          #      "null/none" — which the case statement below has no
-          #      branch for, so it falls into *) → exit 1.
-          # Surfaced 2026-04-30 the first time the App-token chain
-          # (#2389) actually fired auto-promote-on-e2e from a publish
-          # upstream — every prior run was E2E-upstream which
-          # short-circuits before this gate.
-          RESULT=$(gh run list \
-            --repo "$REPO" \
-            --workflow e2e-staging-saas.yml \
-            --branch main \
-            --commit "$SHA" \
-            --limit 1 \
-            --json status,conclusion \
-            --jq '(.[0] // {}) | "\(.status // "none")/\(.conclusion // "none")"' \
-            2>/dev/null || echo "none/none")
-
-          echo "E2E Staging SaaS for ${SHA:0:7}: $RESULT"
-
-          case "$RESULT" in
-            completed/success)
-              echo "proceed=true" >> "$GITHUB_OUTPUT"
-              echo "::notice::E2E green for this SHA — proceeding with promote"
-              ;;
-            completed/failure|completed/cancelled|completed/timed_out)
-              echo "proceed=false" >> "$GITHUB_OUTPUT"
-              {
-                echo "## ❌ Auto-promote aborted — E2E Staging SaaS failed"
-                echo
-                echo "E2E Staging SaaS for \`${SHA:0:7}\`: \`$RESULT\`"
-                echo "\`:latest\` stays on the prior known-good digest."
-                echo
-                echo "If the failure was a flake, manually dispatch this workflow with the same sha to override."
-              } >> "$GITHUB_STEP_SUMMARY"
-              exit 1
-              ;;
-            in_progress/*|queued/*|requested/*|waiting/*|pending/*)
-              echo "proceed=false" >> "$GITHUB_OUTPUT"
-              {
-                echo "## ⏳ Auto-promote deferred — E2E Staging SaaS still running"
-                echo
-                echo "Publish completed before E2E for \`${SHA:0:7}\` (state: \`$RESULT\`)."
-                echo "Skipping retag here — E2E's own completion event will re-fire this workflow."
-                echo "If E2E ends green, that run promotes \`:latest\`. If red, it aborts."
-              } >> "$GITHUB_STEP_SUMMARY"
-              ;;
-            none/none)
-              echo "proceed=true" >> "$GITHUB_OUTPUT"
-              echo "::notice::E2E paths-filtered out for this SHA — pre-merge staging gates carry"
-              ;;
-            *)
-              echo "proceed=false" >> "$GITHUB_OUTPUT"
-              {
-                echo "## ❓ Auto-promote aborted — unexpected E2E state"
-                echo
-                echo "E2E Staging SaaS for \`${SHA:0:7}\`: \`$RESULT\` (unhandled)"
-                echo "Manual investigation needed; re-dispatch with the same sha once resolved."
-              } >> "$GITHUB_STEP_SUMMARY"
-              exit 1
-              ;;
-          esac
-
-      - if: steps.gate.outputs.proceed == 'true'
-        uses: imjasonh/setup-crane@6da1ae018866400525525ce74ff892880c099987 # v0.5
-
-      - name: GHCR login
-        if: steps.gate.outputs.proceed == 'true'
-        run: |
-          echo "${{ secrets.GITHUB_TOKEN }}" | \
-            crane auth login ghcr.io -u "${{ github.actor }}" --password-stdin
-
-      - name: Verify :staging-<sha> exists for both images
-        # Better to fail fast with a clear message than to half-tag
-        # (platform retagged but platform-tenant missing → tenants pull
-        # a stale image).
-        if: steps.gate.outputs.proceed == 'true'
-        run: |
-          set -euo pipefail
-          for img in "${IMAGE_NAME}" "${TENANT_IMAGE_NAME}"; do
-            tag="${img}:staging-${{ steps.sha.outputs.short }}"
-            if ! crane manifest "$tag" >/dev/null 2>&1; then
-              echo "::error::Missing tag: $tag"
-              echo "::error::publish-workspace-server-image must complete on this SHA before auto-promote can retag :latest."
-              exit 1
-            fi
-            echo "  ok: $tag exists"
-          done
-
-      - name: Ancestry check — refuse to promote :latest backwards
-        # #2244: workflow_run completions arrive in arbitrary order. If
-        # SHA-A and SHA-B both reach main within ~10 min and SHA-B's E2E
-        # completes before SHA-A's, this workflow can fire for SHA-A
-        # AFTER it already promoted SHA-B → :latest goes backwards. The
-        # orphan-reconciler "next run corrects it" doesn't apply: there's
-        # no auto-corrective re-promote, :latest stays wrong until the
-        # next main push lands.
-        #
-        # Detection: read current :latest's `org.opencontainers.image.revision`
-        # label (set by publish-workspace-server-image.yml at build time)
-        # and ask the GitHub compare API whether the candidate SHA is
-        # ahead-of / identical-to / behind / diverged-from current.
-        # Hard-fail on `behind` and `diverged` per the approved design —
-        # silent-bypass is the class we're moving away from. Workflow
-        # goes red, oncall sees it, operator decides how to recover
-        # (manual dispatch with the right SHA, force-promote, etc.).
-        #
-        # Manual dispatch skips this check — operator override semantics
-        # match the gate-check step above.
-        #
-        # Backward-compat: when current :latest carries no revision
-        # label (legacy image pre-publish-with-label), skip-with-warning.
-        # All :latest images on main are post-label as of 2026-04-29, so
-        # this branch will be dead within 90 days; remove then.
-        if: steps.gate.outputs.proceed == 'true' && github.event_name != 'workflow_dispatch'
-        id: ancestry
-        env:
-          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          REPO: ${{ github.repository }}
-          TARGET_SHA: ${{ steps.sha.outputs.full }}
-        run: |
-          set -euo pipefail
-
-          # Read the current :latest config and pull the revision label.
-          # `crane config` returns the OCI image config blob (not the manifest);
-          # labels live under `.config.Labels`. `// empty` makes jq return ""
-          # rather than the literal "null" so the test below works.
-          CURRENT_REVISION=$(crane config "${IMAGE_NAME}:latest" 2>/dev/null \
-            | jq -r '.config.Labels["org.opencontainers.image.revision"] // empty' \
-            || true)
-
-          if [ -z "$CURRENT_REVISION" ]; then
-            echo "decision=skip-no-label" >> "$GITHUB_OUTPUT"
-            {
-              echo "## ⚠ Ancestry check skipped — current :latest has no revision label"
-              echo
-              echo "Likely a legacy image built before \`org.opencontainers.image.revision\` was set."
-              echo "Falling through to retag. After all \`:latest\` images are post-label (TODO 90 days), this branch is dead and should be removed."
-            } >> "$GITHUB_STEP_SUMMARY"
-            echo "::warning::Current :latest carries no revision label — skipping ancestry check (legacy image)"
-            exit 0
-          fi
-
-          if [ "$CURRENT_REVISION" = "$TARGET_SHA" ]; then
-            echo "decision=identical" >> "$GITHUB_OUTPUT"
-            echo "::notice:::latest already at ${TARGET_SHA:0:7} — retag will be a no-op"
-            exit 0
-          fi
-
-          # Ask GitHub which side of the merge graph TARGET_SHA sits on
-          # relative to CURRENT_REVISION. Returns one of: ahead | identical
-          # | behind | diverged. Network or auth errors collapse to "error"
-          # via the explicit fallback so the case below always matches.
-          STATUS=$(gh api \
-            "repos/${REPO}/compare/${CURRENT_REVISION}...${TARGET_SHA}" \
-            --jq '.status' 2>/dev/null || echo "error")
-
-          echo "ancestry compare ${CURRENT_REVISION:0:7} → ${TARGET_SHA:0:7}: $STATUS"
-
-          case "$STATUS" in
-            ahead)
-              echo "decision=ahead" >> "$GITHUB_OUTPUT"
-              echo "::notice::Target ${TARGET_SHA:0:7} is ahead of current :latest (${CURRENT_REVISION:0:7}) — proceeding with retag"
-              ;;
-            identical)
-              echo "decision=identical" >> "$GITHUB_OUTPUT"
-              echo "::notice::Target identical to :latest — retag will be a no-op"
-              ;;
-            behind)
-              echo "decision=behind" >> "$GITHUB_OUTPUT"
-              {
-                echo "## ❌ Auto-promote refused — target is BEHIND current :latest"
-                echo
-                echo "| Field | Value |"
-                echo "|---|---|"
-                echo "| Target SHA | \`$TARGET_SHA\` |"
-                echo "| Current :latest revision | \`$CURRENT_REVISION\` |"
-                echo "| GitHub compare status | \`behind\` |"
-                echo
-                echo "This guard catches the workflow_run-completion-order race (#2244):"
-                echo "two rapid main pushes whose E2Es complete out-of-order can otherwise"
-                echo "promote \`:latest\` backwards. \`:latest\` stays on \`${CURRENT_REVISION:0:7}\`."
-                echo
-                echo "**Recovery:** if this is a legitimate revert that should land on \`:latest\`,"
-                echo "manually dispatch this workflow with the target sha as input — the manual-dispatch"
-                echo "path skips the ancestry check (operator override)."
-              } >> "$GITHUB_STEP_SUMMARY"
-              exit 1
-              ;;
-            diverged)
-              echo "decision=diverged" >> "$GITHUB_OUTPUT"
-              {
-                echo "## ❓ Auto-promote refused — history diverged"
-                echo
-                echo "| Field | Value |"
-                echo "|---|---|"
-                echo "| Target SHA | \`$TARGET_SHA\` |"
-                echo "| Current :latest revision | \`$CURRENT_REVISION\` |"
-                echo "| GitHub compare status | \`diverged\` |"
-                echo
-                echo "Likely cause: force-push rewrote main's history, leaving the previous"
-                echo "\`:latest\` revision orphaned. Needs human review before \`:latest\` advances."
-              } >> "$GITHUB_STEP_SUMMARY"
-              exit 1
-              ;;
-            error|*)
-              echo "decision=error" >> "$GITHUB_OUTPUT"
-              {
-                echo "## ❌ Auto-promote aborted — ancestry-check API error"
-                echo
-                echo "\`gh api repos/${REPO}/compare/${CURRENT_REVISION}...${TARGET_SHA}\` returned unexpected status: \`$STATUS\`"
-                echo
-                echo "Manual dispatch with the target sha bypasses this check."
-              } >> "$GITHUB_STEP_SUMMARY"
-              exit 1
-              ;;
-          esac
-
-      - name: Retag platform :staging-<sha> → :latest
-        if: steps.gate.outputs.proceed == 'true'
-        run: |
-          crane tag "${IMAGE_NAME}:staging-${{ steps.sha.outputs.short }}" latest
-
-      - name: Retag tenant :staging-<sha> → :latest
-        if: steps.gate.outputs.proceed == 'true'
-        run: |
-          crane tag "${TENANT_IMAGE_NAME}:staging-${{ steps.sha.outputs.short }}" latest
-
-      - name: Summary
-        if: steps.gate.outputs.proceed == 'true'
-        run: |
-          {
-            echo "## :latest promoted to ${{ steps.sha.outputs.short }}"
-            echo
-            if [ "${{ github.event_name }}" = "workflow_dispatch" ]; then
-              echo "- Trigger: manual dispatch"
-            else
-              echo "- Upstream: \`${{ github.event.workflow_run.name }}\` ([run](${{ github.event.workflow_run.html_url }}))"
-            fi
-            echo "- platform:staging-${{ steps.sha.outputs.short }} → :latest"
-            echo "- platform-tenant:staging-${{ steps.sha.outputs.short }} → :latest"
-            echo
-            echo "Tenant fleet auto-pulls within 5 min via IMAGE_AUTO_REFRESH=true."
-            echo "Force immediate fanout: dispatch redeploy-tenants-on-main.yml."
-          } >> "$GITHUB_STEP_SUMMARY"
@@ -1,434 +0,0 @@
-name: Auto-promote staging → main
-
-# Fires after any of the staging-branch quality gates complete. When ALL
-# required gates are green on the same staging SHA, opens (or re-uses)
-# a PR `staging → main` and enables auto-merge so the merge queue lands
-# it. Closes the gap that historically let features sit on staging for
-# weeks waiting for a bulk promotion PR (see molecule-core#1496 for the
-# 1172-commit example).
-#
-# 2026-04-28 rewrite (PR #142): the previous version did a direct
-# `git merge --ff-only origin staging && git push origin main`. That
-# breaks against main's branch-protection ruleset, which requires
-# status checks "set by the expected GitHub apps" — direct pushes
-# can't satisfy that condition (only PR merges through the queue can).
-# The workflow was failing every tick with:
-#   remote: error: GH006: Protected branch update failed for refs/heads/main.
-#   remote: - Required status checks ... were not set by the expected GitHub apps.
-# Fix: mirror the PR-based pattern from auto-sync-main-to-staging.yml
-# (the reverse-direction sync, fixed in #2234 for the same reason).
-# Both directions now use the same merge-queue path that humans use,
-# no special-case bypass.
-#
-# Safety model:
-# - Runs ONLY on workflow_run events for the staging branch.
-# - Requires EVERY named gate workflow to have the same head_sha and
-#   all be `conclusion == success`. If any of them is red, skipped,
-#   cancelled, or pending, we abort (stay on the current main).
-# - The PR base=main head=staging path lets GitHub itself enforce
-#   branch protection. If main has diverged from staging or required
-#   checks aren't satisfied, the merge queue declines the PR — no
-#   need for a manual ff-only ancestry check here.
-# - Loop safety: the auto-sync-main-to-staging workflow fires when
-#   main lands the auto-promote PR, but its merge into staging is by
-#   GITHUB_TOKEN which doesn't trigger downstream workflow_run events
-#   (GitHub Actions safety). So this workflow doesn't re-fire from
-#   its own promote landing.
-#
-# Toggle via repo variable AUTO_PROMOTE_ENABLED (true/unset). When
-# unset, the workflow logs what it would have done but doesn't open
-# the PR — useful for dry-running the gate logic without surfacing
-# a noisy PR while staging CI is still flaky.
-#
-# **One-time repo setting (load-bearing):** this workflow opens the
-# staging→main PR via `gh pr create` using the default GITHUB_TOKEN.
-# Since GitHub's 2022 default change, that token cannot create or
-# approve PRs unless the repo opts in. The toggle is at:
-#
-#   Settings → Actions → General → Workflow permissions
-#   → ✅ Allow GitHub Actions to create and approve pull requests
-#
-# Without it, every workflow_run fails with:
-#
-#   pull request create failed: GraphQL: GitHub Actions is not
-#   permitted to create or approve pull requests (createPullRequest)
-#
-# Observed 2026-04-29 01:43 UTC blocking promotion of fcd87b9 (PRs
-# #2248 + #2249); manually bridged via PR #2252. Re-check this
-# setting if auto-promote starts failing with createPullRequest
-# errors after a repo or org admin change.
-
-on:
-  workflow_run:
-    workflows:
-      - CI
-      - E2E Staging Canvas (Playwright)
-      - E2E API Smoke Test
-      - CodeQL
-    types: [completed]
-  workflow_dispatch:
-    inputs:
-      force:
-        description: "Force promote even when AUTO_PROMOTE_ENABLED is unset (manual override)"
-        required: false
-        default: "false"
-
-permissions:
-  contents: write
-  pull-requests: write
-  # actions: write is needed by the post-merge dispatch tail step
-  # (#2358 / #2357) — `gh workflow run publish-workspace-server-image.yml`
-  # POSTs to /actions/workflows/.../dispatches which requires this scope.
-  # Without it the call 403s and the publish/canary/redeploy chain still
-  # doesn't run on staging→main promotions, undoing #2358.
-  actions: write
-
-# Serialize auto-promote runs. Multiple staging gate completions can land
-# in quick succession (CI + E2E + CodeQL all finish within seconds of
-# each other on a green PR) — without this, two parallel runs both:
-#   1. Open / re-use the same promote PR.
-#   2. Both call `gh pr merge --auto` (idempotent — fine).
-#   3. Both poll for the same mergedAt and both `gh workflow run` publish
-#      → 2× redundant publish builds racing for the same `:staging-latest`
-#      retag, and 2× canary-verify chains.
-# cancel-in-progress: false because we don't want a brand-new run to kill
-# a polling-tail that's about to dispatch — the polling tail's 30 min cap
-# is the right backstop, not workflow-level cancel.
-concurrency:
-  group: auto-promote-staging
-  cancel-in-progress: false
-
-jobs:
-  check-all-gates-green:
-    # Only consider staging pushes. PRs into staging don't promote.
-    if: >
-      (github.event_name == 'workflow_run' &&
-       github.event.workflow_run.head_branch == 'staging' &&
-       github.event.workflow_run.event == 'push')
-      || github.event_name == 'workflow_dispatch'
-    runs-on: ubuntu-latest
-    outputs:
-      all_green: ${{ steps.gates.outputs.all_green }}
-      head_sha: ${{ steps.gates.outputs.head_sha }}
-    steps:
-      # Skip empty-tree promotes (the perpetual auto-promote↔auto-sync cycle
-      # observed 2026-05-03). Sequence: auto-promote merges via the staging
-      # merge-queue's MERGE strategy, creating a merge commit on main that
-      # staging doesn't have. auto-sync then merges main back into staging
-      # via another merge commit (the queue's MERGE strategy applies on
-      # the staging side too, even when the workflow's local FF would
-      # have sufficed). Now staging has a new merge-commit SHA whose
-      # tree == main's tree — but auto-promote sees "staging ahead of
-      # main by 1" and opens YET another empty promote PR. Each round
-      # costs ~30-40 min wallclock, ~2 manual approvals, and burns a
-      # full CodeQL Go run (~15 min). Without this guard the cycle
-      # repeats indefinitely.
-      #
-      # Long-term fix is to switch the merge_queue ruleset's
-      # `merge_method` away from MERGE so FF-able PRs land cleanly,
-      # but that's a broader change affecting every staging PR's
-      # commit shape. This guard is the one-line surgical fix that
-      # breaks the cycle without touching merge-queue config.
-      #
-      # Fail-open: if `git diff` errors for any reason, fall through
-      # to the gate check (preserve existing behavior). Only skip
-      # when the diff is DEFINITIVELY empty.
-      - name: Checkout for tree-diff check
-        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
-        with:
-          fetch-depth: 0
-          ref: staging
-      - name: Skip if staging tree == main tree (perpetual-cycle break)
-        id: tree-diff
-        env:
-          HEAD_SHA: ${{ github.event.workflow_run.head_sha || github.sha }}
-        run: |
-          set -eu
-          git fetch origin main --depth=50 || { echo "::warning::git fetch main failed — proceeding (fail-open)"; exit 0; }
-          # Compare staging tip's tree against main's tree. `git diff
-          # --quiet` exits 0 if no differences, 1 if there are.
-          if git diff --quiet origin/main "$HEAD_SHA" -- 2>/dev/null; then
-            {
-              echo "## ⏭ Skipped — no code to promote"
-              echo
-              echo "staging tip (\`${HEAD_SHA:0:8}\`) and \`main\` have identical trees."
-              echo "This is the auto-promote↔auto-sync merge-commit cycle: staging has a"
-              echo "new SHA (a sync-back merge commit) but the underlying file tree is"
-              echo "already on main, so there's no real code to ship."
-              echo
-              echo "Skipping to avoid opening an empty promote PR. Cycle terminates here."
-            } >> "$GITHUB_STEP_SUMMARY"
-            echo "::notice::auto-promote: staging tree == main tree — no code to promote, skipping"
-            echo "skip=true" >> "$GITHUB_OUTPUT"
-          else
-            echo "skip=false" >> "$GITHUB_OUTPUT"
-          fi
-      - name: Check all required gates on this SHA
-        if: steps.tree-diff.outputs.skip != 'true'
-        id: gates
-        env:
-          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HEAD_SHA: ${{ github.event.workflow_run.head_sha || github.sha }}
-          REPO: ${{ github.repository }}
-        run: |
-          set -euo pipefail
-
-          # Required gate workflow files. Use file paths (relative to
-          # .github/workflows/) rather than display names because:
-          #
-          #   1. `gh run list --workflow=<name>` is ambiguous when two
-          #      workflows have the same `name:` — observed 2026-04-28
-          #      with "CodeQL" matching both `codeql.yml` (explicit) and
-          #      GitHub's UI-configured Code-quality default setup
-          #      (internal "codeql"). gh CLI returns "could not resolve
-          #      to a unique workflow" → empty result → gate evaluated
-          #      as missing/none → auto-promote dead-locked despite all
-          #      checks actually passing.
-          #
-          #   2. File paths are the unique identifier for workflows;
-          #      `name:` is just a display string and can collide.
-          #
-          # When adding/removing a gate, update this list AND the
-          # branch-protection required-checks list (which uses check-run
-          # display names, not workflow names; the two are decoupled and
-          # should be kept in sync manually).
-          GATES=(
-            "ci.yml"
-            "e2e-staging-canvas.yml"
-            "e2e-api.yml"
-            "codeql.yml"
-          )
-
-          echo "head_sha=${HEAD_SHA}" >> "$GITHUB_OUTPUT"
-          echo "Checking gates on SHA ${HEAD_SHA}"
-
-          ALL_GREEN=true
-          for gate in "${GATES[@]}"; do
-            # Query the most recent run of this workflow on this SHA.
-            # event=push to avoid picking up PR runs. branch=staging to
-            # guard against someone dispatching the gate on a non-staging
-            # branch at the same SHA.
-            RESULT=$(gh run list \
-              --repo "$REPO" \
-              --workflow "$gate" \
-              --branch staging \
-              --event push \
-              --commit "$HEAD_SHA" \
-              --limit 1 \
-              --json status,conclusion \
-              --jq '.[0] | "\(.status)/\(.conclusion // "none")"' \
-              2>/dev/null || echo "missing/none")
-
-            echo "  $gate → $RESULT"
-
-            # Only completed/success counts. completed/failure or
-            # in_progress/anything or no record at all = abort.
-            if [ "$RESULT" != "completed/success" ]; then
-              ALL_GREEN=false
-            fi
-          done
-
-          echo "all_green=${ALL_GREEN}" >> "$GITHUB_OUTPUT"
-          if [ "$ALL_GREEN" != "true" ]; then
-            echo "::notice::auto-promote: not all gates are green on ${HEAD_SHA} — staying on current main"
-          fi
-
-  promote:
-    needs: check-all-gates-green
-    if: needs.check-all-gates-green.outputs.all_green == 'true'
-    runs-on: ubuntu-latest
-    steps:
-      - name: Check rollout gate
-        env:
-          AUTO_PROMOTE_ENABLED: ${{ vars.AUTO_PROMOTE_ENABLED }}
-          FORCE_INPUT: ${{ github.event.inputs.force }}
-        run: |
-          set -eu
-          # Repo variable AUTO_PROMOTE_ENABLED=true flips this on. While
-          # it's unset, the workflow dry-runs (logs what it would have
-          # done) but doesn't open the promote PR. Set the variable in
-          # Settings → Secrets and variables → Actions → Variables.
-          if [ "${AUTO_PROMOTE_ENABLED:-}" != "true" ] && [ "${FORCE_INPUT:-false}" != "true" ]; then
-            {
-              echo "## ⏸ Auto-promote disabled"
-              echo
-              echo "Repo variable \`AUTO_PROMOTE_ENABLED\` is not set to \`true\`."
-              echo "All gates are green on staging; would have opened a promote PR to \`main\`."
-              echo
-              echo "To enable: Settings → Secrets and variables → Actions → Variables → \`AUTO_PROMOTE_ENABLED=true\`."
-              echo "To test once manually: workflow_dispatch with \`force=true\`."
-            } >> "$GITHUB_STEP_SUMMARY"
-            echo "::notice::auto-promote disabled — dry run only"
-            exit 0
-          fi
-
-      # Mint the App token BEFORE the promote-PR step so the auto-merge
-      # call can use it. GITHUB_TOKEN-initiated merges suppress the
-      # downstream `push` event on main, breaking the
-      # publish-workspace-server-image → canary-verify → redeploy-tenants
-      # chain (issue #2357). Using the App token here means the
-      # merge-queue-landed merge IS able to fire the cascade naturally;
-      # the polling tail below stays as defense-in-depth.
-      - name: Mint App token for promote-PR + downstream dispatch
-        if: ${{ vars.AUTO_PROMOTE_ENABLED == 'true' || github.event.inputs.force == 'true' }}
-        id: app-token
-        uses: actions/create-github-app-token@1b10c78c7865c340bc4f6099eb2f838309f1e8c3 # v3.1.1
-        with:
-          app-id: ${{ secrets.MOLECULE_AI_APP_ID }}
-          private-key: ${{ secrets.MOLECULE_AI_APP_PRIVATE_KEY }}
-
-      - name: Open (or reuse) staging → main promote PR + enable auto-merge
-        if: ${{ vars.AUTO_PROMOTE_ENABLED == 'true' || github.event.inputs.force == 'true' }}
-        env:
-          GH_TOKEN: ${{ steps.app-token.outputs.token }}
-          REPO: ${{ github.repository }}
-          TARGET_SHA: ${{ needs.check-all-gates-green.outputs.head_sha }}
-        run: |
-          set -euo pipefail
-
-          # Look for an existing open promote PR (idempotent on re-run
-          # of the workflow). The PR's head IS the staging branch — the
-          # whole point is "advance main to staging's tip", so we don't
-          # need a per-SHA branch like auto-sync-main-to-staging uses.
-          PR_NUM=$(gh pr list --repo "$REPO" \
-            --base main --head staging --state open \
-            --json number --jq '.[0].number // ""')
-
-          if [ -z "$PR_NUM" ]; then
-            TITLE="staging → main: auto-promote ${TARGET_SHA:0:7}"
-            BODY_FILE=$(mktemp)
-            cat > "$BODY_FILE" <<EOFBODY
-          Automated promotion of \`staging\` (\`${TARGET_SHA:0:8}\`) to \`main\`. All required staging gates green at this SHA: CI, E2E Staging Canvas, E2E API Smoke, CodeQL.
-
-          This PR is auto-generated by \`.github/workflows/auto-promote-staging.yml\` whenever every required gate completes green on the same staging SHA. It exists because main's branch protection requires status checks "set by the expected GitHub apps" — direct \`git push\` from a workflow can't satisfy that, only PR merges through the queue can.
-
-          Merge queue lands this; no human action needed unless gates fail. Reverse-direction sync (the merge commit on main → staging) is handled by \`auto-sync-main-to-staging.yml\`.
-          EOFBODY
-            PR_URL=$(gh pr create --repo "$REPO" \
-              --base main --head staging \
-              --title "$TITLE" \
-              --body-file "$BODY_FILE")
-            PR_NUM=$(echo "$PR_URL" | grep -oE '[0-9]+$' | tail -1)
-            rm -f "$BODY_FILE"
-            echo "::notice::Opened PR #${PR_NUM}"
-          else
-            echo "::notice::Re-using existing promote PR #${PR_NUM}"
-          fi
-
-          # Enable auto-merge — the merge queue picks it up once
-          # required gates are green on the merge_group ref.
-          if ! gh pr merge "$PR_NUM" --repo "$REPO" --auto --merge 2>&1; then
-            echo "::warning::Failed to enable auto-merge on PR #${PR_NUM} — operator may need to merge manually."
-          fi
-
-          {
-            echo "## ✅ Auto-promote PR opened"
-            echo
-            echo "- Source: staging at \`${TARGET_SHA:0:8}\`"
-            echo "- PR: #${PR_NUM}"
-            echo
-            echo "Merge queue lands the PR once required gates are green; no human action needed unless gates fail."
-          } >> "$GITHUB_STEP_SUMMARY"
-
-          # Hand the PR number to the next step so we can dispatch the
-          # tenant-redeploy chain after the merge queue lands the merge.
-          echo "promote_pr_num=${PR_NUM}" >> "$GITHUB_OUTPUT"
-        id: promote_pr
-
-      # The App token minted above (before the promote-PR step) is
-      # also used by the polling tail below. Defense-in-depth: with
-      # the merge-queue-landed merge now using the App token, the
-      # main-branch push event SHOULD fire the publish/canary/redeploy
-      # cascade naturally — but if for any reason it doesn't (e.g. an
-      # unrelated event-suppression edge case), the explicit dispatches
-      # below still wake the chain.
-      - name: Wait for promote merge, then dispatch publish + redeploy (#2357)
-        # Defense-in-depth dispatch. With the auto-merge call above
-        # now using the App token (this commit), the merge-queue-landed
-        # merge SHOULD fire publish-workspace-server-image naturally
-        # via on:push:[main] — App-token-initiated pushes DO trigger
-        # workflow_run cascades, unlike GITHUB_TOKEN-initiated ones
-        # (the documented "no recursion" rule —
-        # https://docs.github.com/en/actions/using-workflows/triggering-a-workflow#triggering-a-workflow-from-a-workflow).
-        #
-        # This explicit dispatch stays as belt-and-suspenders for any
-        # edge case where the natural cascade misfires. If it never
-        # observably fires after this token swap (i.e. the publish
-        # workflow has already started by the time we get here), the
-        # second dispatch is a harmless no-op (publish-workspace-server-image
-        # has its own concurrency group that dedupes).
-        #
-        # See PR for #2357: pre-fix the merge action was via
-        # GITHUB_TOKEN, suppressing the cascade and forcing this tail
-        # to be the SOLE chain trigger. With the auto-merge token swap
-        # the tail becomes redundant in the happy path; keep until
-        # we've observed >=10 successful natural cascades, then drop.
-        if: steps.promote_pr.outputs.promote_pr_num != ''
-        env:
-          GH_TOKEN: ${{ steps.app-token.outputs.token }}
-          REPO: ${{ github.repository }}
-          PR_NUM: ${{ steps.promote_pr.outputs.promote_pr_num }}
-        run: |
-          # Poll for merge — max 30 min (60 × 30s). The merge queue
-          # typically lands within 5-10 min when gates are green. Break
-          # early if the PR is closed without merging (operator action,
-          # gates flipped red post-approval, branch-protection rejection)
-          # so we don't tie up a runner for the full 30 min on a dead PR.
-          MERGED=""
-          STATE=""
-          for _ in $(seq 1 60); do
-            VIEW=$(gh pr view "$PR_NUM" --repo "$REPO" --json mergedAt,state)
-            MERGED=$(echo "$VIEW" | jq -r '.mergedAt // ""')
-            STATE=$(echo "$VIEW" | jq -r '.state // ""')
-            if [ -n "$MERGED" ] && [ "$MERGED" != "null" ]; then
-              echo "::notice::Promote PR #${PR_NUM} merged at ${MERGED}"
-              break
-            fi
-            if [ "$STATE" = "CLOSED" ]; then
-              echo "::warning::Promote PR #${PR_NUM} was closed without merging — skipping deploy dispatch."
-              exit 0
-            fi
-            sleep 30
-          done
-
-          if [ -z "$MERGED" ] || [ "$MERGED" = "null" ]; then
-            echo "::warning::Promote PR #${PR_NUM} didn't merge within 30min — skipping deploy dispatch (manually run \`gh workflow run publish-workspace-server-image.yml --ref main\` once it lands)."
-            exit 0
-          fi
-
-          # Dispatch publish on main using the App token. App-initiated
-          # workflow_dispatch DOES propagate the workflow_run cascade,
-          # unlike GITHUB_TOKEN-initiated dispatch.
-          # publish completes → canary-verify chains via workflow_run →
-          # redeploy-tenants-on-main chains via workflow_run + branches:[main].
-          if gh workflow run publish-workspace-server-image.yml \
-              --repo "$REPO" --ref main 2>&1; then
-            echo "::notice::Dispatched publish-workspace-server-image on ref=main as molecule-ai App — canary-verify and redeploy-tenants-on-main will chain via workflow_run."
-            {
-              echo "## 🚀 Tenant redeploy chain dispatched"
-              echo
-              echo "- publish-workspace-server-image (workflow_dispatch on \`main\`, actor: \`molecule-ai[bot]\`)"
-              echo "- canary-verify will chain on completion"
-              echo "- redeploy-tenants-on-main will chain on canary green"
-            } >> "$GITHUB_STEP_SUMMARY"
-          else
-            echo "::error::Failed to dispatch publish-workspace-server-image. Run manually: gh workflow run publish-workspace-server-image.yml --ref main"
-          fi
-
-          # ALSO dispatch auto-sync-main-to-staging.yml. Same root cause as
-          # publish above (issue #2357): the merge-queue-initiated push to
-          # main is by GITHUB_TOKEN → no `on: push` triggers fire downstream.
-          # Without this dispatch, every staging→main promote leaves staging
-          # one merge commit BEHIND main, which silently dead-locks the NEXT
-          # promote PR as `mergeStateStatus: BEHIND` because main's
-          # branch-protection has `strict: true`. Verified empirically on
-          # 2026-05-02 against PR #2442 (Phase 2 promote): only the explicit
-          # publish-workspace-server-image dispatch fired on the previous
-          # promote SHA 76c604fb, while auto-sync silently no-op'd, leaving
-          # staging behind for ~24h until manually bridged.
-          if gh workflow run auto-sync-main-to-staging.yml \
-              --repo "$REPO" --ref main 2>&1; then
-            echo "::notice::Dispatched auto-sync-main-to-staging on ref=main as molecule-ai App — staging will absorb the new main merge commit via PR + merge queue."
-          else
-            echo "::error::Failed to dispatch auto-sync-main-to-staging. Run manually: gh workflow run auto-sync-main-to-staging.yml --ref main"
-          fi
@@ -1,237 +0,0 @@
-name: Auto-sync main → staging
-
-# Reflects every push to `main` back onto `staging` so the
-# staging-as-superset-of-main invariant holds.
-#
-# Background:
-#
-# `auto-promote-staging.yml` advances main via `git merge --ff-only`
-# + `git push origin main` — that's a clean fast-forward, no merge
-# commit. But manual merges of `staging → main` PRs through the
-# GitHub UI / API create a merge commit on main that staging
-# doesn't have. The next `staging → main` PR then evaluates as
-# "BEHIND" because staging is missing that merge commit, requiring
-# a manual `gh pr update-branch` round-trip.
-#
-# This happened twice on 2026-04-28 (PRs #2202, #2205, both manual
-# bridges). Each time the bridge needed update-branch + a re-CI
-# round before merging. Operationally annoying and avoidable.
-#
-# Architecture:
-#
-# This repo's `staging` branch is protected by a `merge_queue`
-# ruleset (id 15500102) that blocks ALL direct pushes — no bypass
-# even for org admins or the GitHub Actions integration. Direct
-# `git push origin staging` returns GH013. So instead of pushing
-# directly, this workflow:
-#
-#   1. Checks if main is already in staging's ancestry → no-op.
-#   2. Creates an `auto-sync/main-<sha>` branch from staging.
-#   3. Tries `git merge --ff-only origin/main` → if staging hasn't
-#      diverged this is a clean ff.
-#   4. Otherwise `git merge --no-ff origin/main` to absorb main's
-#      tip while keeping staging's history.
-#   5. Pushes the auto-sync branch.
-#   6. Opens a PR (base=staging, head=auto-sync/main-<sha>) and
-#      enables auto-merge so the merge queue lands it.
-#
-# This mirrors the path human PRs take through staging — same
-# rules, same gates, no special-case bypass.
-#
-# Loop safety:
-#
-# `GITHUB_TOKEN`-authored merges (including the merge queue's land
-# of the auto-sync PR) do NOT trigger downstream workflow runs
-# (GitHub Actions safety). So when the auto-sync PR lands on
-# staging, `auto-promote-staging.yml` is NOT triggered by that
-# push. The next developer push to staging triggers auto-promote
-# normally. No loop possible.
-#
-# Concurrency:
-#
-# Two pushes to main in quick succession (e.g., manual UI merge
-# immediately followed by auto-promote-staging's ff-merge) could
-# otherwise open two overlapping auto-sync PRs. The concurrency
-# group serializes runs; the second waits for the first to exit.
-# (The first run exits after opening + auto-merge-queueing the PR,
-# not after the merge actually completes — so multiple PRs can be
-# open simultaneously, but the merge queue handles them serially.)
-
-on:
-  push:
-    branches: [main]
-  # workflow_dispatch lets:
-  #   1. Operators manually backfill a missed sync (e.g. after a manual
-  #      UI merge that the runner missed).
-  #   2. auto-promote-staging.yml's polling tail explicitly invoke us
-  #      after the promote PR lands. This is load-bearing: when the
-  #      merge queue lands a promote-PR merge, the resulting push to
-  #      `main` is "by GITHUB_TOKEN", and per GitHub's no-recursion
-  #      rule (https://docs.github.com/en/actions/using-workflows/triggering-a-workflow#triggering-a-workflow-from-a-workflow)
-  #      that push event does NOT fire any downstream workflows. The
-  #      `on: push` trigger above is silently dead for the very pattern
-  #      we exist to handle. Verified empirically 2026-05-02 against
-  #      SHA 76c604fb (PR #2437 staging→main): only ONE workflow fired
-  #      (publish-workspace-server-image, dispatched explicitly by
-  #      auto-promote's polling tail with an App token). Every other
-  #      `on: push: branches: [main]` workflow — including this one —
-  #      was suppressed. Until the underlying merge call moves to an
-  #      App token, an explicit dispatch is the only reliable path.
-  workflow_dispatch:
-
-permissions:
-  contents: write
-  pull-requests: write
-
-concurrency:
-  group: auto-sync-main-to-staging
-  cancel-in-progress: false
-
-jobs:
-  sync-staging:
-    # ubuntu-latest matches every other workflow in this repo. The
-    # earlier `[self-hosted, macos, arm64]` was a copy-paste artefact
-    # from the molecule-controlplane repo (which IS private and uses a
-    # Mac runner) — molecule-core has no Mac runner registered, so the
-    # job sat unassigned whenever the trigger fired. Verified 2026-05-02:
-    # this is the ONLY workflow in molecule-core/.github/workflows/ with
-    # a non-ubuntu runs-on.
-    runs-on: ubuntu-latest
-    steps:
-      - name: Checkout staging
-        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
-        with:
-          fetch-depth: 0
-          ref: staging
-          token: ${{ secrets.GITHUB_TOKEN }}
-
-      - name: Configure git author
-        run: |
-          git config user.name "github-actions[bot]"
-          git config user.email "41898282+github-actions[bot]@users.noreply.github.com"
-
-      - name: Check if staging already contains main
-        id: check
-        run: |
-          set -euo pipefail
-          git fetch origin main
-          if git merge-base --is-ancestor origin/main HEAD; then
-            echo "needs_sync=false" >> "$GITHUB_OUTPUT"
-            {
-              echo "## ✅ No-op"
-              echo
-              echo "staging already contains \`origin/main\` ($(git rev-parse --short=8 origin/main))."
-            } >> "$GITHUB_STEP_SUMMARY"
-          else
-            echo "needs_sync=true" >> "$GITHUB_OUTPUT"
-            MAIN_SHORT=$(git rev-parse --short=8 origin/main)
-            echo "main_short=${MAIN_SHORT}" >> "$GITHUB_OUTPUT"
-            echo "branch=auto-sync/main-${MAIN_SHORT}" >> "$GITHUB_OUTPUT"
-            echo "::notice::staging is missing main's tip (${MAIN_SHORT}) — opening sync PR"
-          fi
-
-      - name: Create auto-sync branch + merge main
-        if: steps.check.outputs.needs_sync == 'true'
-        id: prep
-        run: |
-          set -euo pipefail
-          BRANCH="${{ steps.check.outputs.branch }}"
-
-          # If a previous auto-sync run already opened a branch for the
-          # same main sha, prefer reusing it (idempotent behavior on
-          # workflow restart). Force-update from latest staging anyway
-          # so it absorbs any staging-side commits that landed since.
-          git checkout -B "$BRANCH"
-
-          if git merge --ff-only origin/main; then
-            echo "did_ff=true" >> "$GITHUB_OUTPUT"
-            echo "::notice::Fast-forwarded ${BRANCH} to origin/main"
-          else
-            echo "did_ff=false" >> "$GITHUB_OUTPUT"
-            if ! git merge --no-ff origin/main -m "chore: sync main → staging (auto)"; then
-              # Hygiene: leave the work tree clean before failing.
-              git merge --abort || true
-              {
-                echo "## ❌ Conflict"
-                echo
-                echo "Auto-merge \`main → staging\` failed with conflicts."
-                echo "A human needs to resolve manually."
-              } >> "$GITHUB_STEP_SUMMARY"
-              exit 1
-            fi
-          fi
-
-      - name: Push auto-sync branch
-        if: steps.check.outputs.needs_sync == 'true'
-        run: |
-          set -euo pipefail
-          # Force-with-lease so a concurrent auto-sync run can't
-          # silently clobber an in-flight branch we just updated. If a
-          # different writer touched the branch, we abort and the next
-          # run picks up the latest state.
-          git push --force-with-lease origin "${{ steps.check.outputs.branch }}"
-
-      - name: Open auto-sync PR + enable auto-merge
-        if: steps.check.outputs.needs_sync == 'true'
-        env:
-          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          BRANCH: ${{ steps.check.outputs.branch }}
-          MAIN_SHORT: ${{ steps.check.outputs.main_short }}
-          DID_FF: ${{ steps.prep.outputs.did_ff }}
-        run: |
-          set -euo pipefail
-
-          # Find existing PR for this branch (idempotent on workflow
-          # restart) before creating a new one.
-          PR_NUM=$(gh pr list --head "$BRANCH" --base staging --state open --json number --jq '.[0].number // ""')
-
-          if [ -z "$PR_NUM" ]; then
-            # Body lives in a temp file to keep the multi-line content
-            # out of the YAML block scalar (un-indented newlines inside
-            # an inline shell string break YAML parsing).
-            BODY_FILE=$(mktemp)
-            if [ "$DID_FF" = "true" ]; then
-              TITLE="chore: sync main → staging (auto, ff to ${MAIN_SHORT})"
-              cat > "$BODY_FILE" <<EOFBODY
-          Automated fast-forward of \`staging\` to \`origin/main\` (\`${MAIN_SHORT}\`). Staging has no in-flight commits that diverge from main. Merge queue lands this; no human action needed.
-
-          This PR is auto-generated by \`.github/workflows/auto-sync-main-to-staging.yml\` on every push to \`main\`. It exists because this repo's \`staging\` branch has a \`merge_queue\` ruleset that blocks direct pushes — even from the GitHub Actions integration.
-          EOFBODY
-            else
-              TITLE="chore: sync main → staging (auto, merge ${MAIN_SHORT})"
-              cat > "$BODY_FILE" <<EOFBODY
-          Automated merge of \`origin/main\` (\`${MAIN_SHORT}\`) into \`staging\`. Staging has commits main doesn't, so this is a non-ff merge that absorbs main's tip. Merge queue lands this.
-
-          This PR is auto-generated by \`.github/workflows/auto-sync-main-to-staging.yml\` on every push to \`main\`.
-          EOFBODY
-            fi
-
-            # gh pr create prints the URL on stdout; extract the PR number.
-            PR_URL=$(gh pr create \
-              --base staging \
-              --head "$BRANCH" \
-              --title "$TITLE" \
-              --body-file "$BODY_FILE")
-            PR_NUM=$(echo "$PR_URL" | grep -oE '[0-9]+$' | tail -1)
-            rm -f "$BODY_FILE"
-            echo "::notice::Opened PR #${PR_NUM}"
-          else
-            echo "::notice::Re-using existing PR #${PR_NUM} for ${BRANCH}"
-          fi
-
-          # Enable auto-merge — the merge queue picks it up once
-          # required gates are green. Use --merge for merge commits
-          # (matches the rest of this repo's PR convention).
-          if ! gh pr merge "$PR_NUM" --auto --merge 2>&1; then
-            echo "::warning::Failed to enable auto-merge on PR #${PR_NUM} — operator may need to merge manually."
-          fi
-
-          {
-            echo "## ✅ Auto-sync PR opened"
-            echo
-            echo "- Branch: \`$BRANCH\`"
-            echo "- PR: #$PR_NUM"
-            echo "- Strategy: $([ "$DID_FF" = "true" ] && echo "ff" || echo "merge commit")"
-            echo
-            echo "Merge queue lands the PR once required gates are green; no human action needed unless gates fail."
-          } >> "$GITHUB_STEP_SUMMARY"
@@ -57,17 +57,42 @@ jobs:
        id: bump
        if: steps.skip.outputs.skip != 'true'
        env:
-          GH_TOKEN: ${{ github.token }}
+          # Gitea-shape token (act_runner forwards GITHUB_TOKEN as a
+          # short-lived per-run secret with read access to this repo).
+          # We hit `/api/v1/repos/.../pulls?state=closed` directly
+          # because `gh pr list` calls Gitea's GraphQL endpoint, which
+          # returns HTTP 405 (issue #75 / post-#66 sweep).
+          GITEA_TOKEN: ${{ github.token }}
+          REPO: ${{ github.repository }}
+          GITEA_API_URL: ${{ github.server_url }}/api/v1
+          PUSH_SHA: ${{ github.sha }}
        run: |
-          # The merged PR for this push commit. `gh pr list --search` finds
-          # closed PRs whose merge commit matches; we take the first.
-          PR=$(gh pr list --state merged --search "${{ github.sha }}" --json number,labels --jq '.[0]' 2>/dev/null || echo "")
+          # Find the merged PR whose merge_commit_sha matches this push.
+          # Gitea's `/repos/{owner}/{repo}/pulls?state=closed` returns
+          # PRs sorted newest-first; we paginate up to 50 and jq-filter
+          # on `merge_commit_sha == PUSH_SHA`. Bounded — auto-tag fires
+          # per push to main, so the matching PR is always among the
+          # most recent closures. 50 is comfortably more than the
+          # ~10-20 staging→main promotes that close in any reasonable
+          # window.
+          set -euo pipefail
+          PRS_JSON=$(curl --fail-with-body -sS \
+            -H "Authorization: token ${GITEA_TOKEN}" \
+            -H "Accept: application/json" \
+            "${GITEA_API_URL}/repos/${REPO}/pulls?state=closed&sort=newest&limit=50" \
+            2>/dev/null || echo "[]")
+          PR=$(printf '%s' "$PRS_JSON" \
+            | jq -c --arg sha "$PUSH_SHA" \
+                '[.[] | select(.merged_at != null and .merge_commit_sha == $sha)] | .[0] // empty')
          if [ -z "$PR" ] || [ "$PR" = "null" ]; then
-            echo "No merged PR found for ${{ github.sha }} — defaulting to patch bump."
+            echo "No merged PR found for ${PUSH_SHA} — defaulting to patch bump."
            echo "kind=patch" >> "$GITHUB_OUTPUT"
            exit 0
          fi
-          LABELS=$(echo "$PR" | jq -r '.labels[].name')
+          # Gitea returns labels under `.labels[].name`, same shape as
+          # GitHub's REST. The previous `gh pr list --json number,labels`
+          # output was identical; jq filter unchanged.
+          LABELS=$(printf '%s' "$PR" | jq -r '.labels[]?.name // empty')
          if echo "$LABELS" | grep -qx 'release:major'; then
            echo "kind=major" >> "$GITHUB_OUTPUT"
          elif echo "$LABELS" | grep -qx 'release:minor'; then
@@ -1,7 +1,7 @@
 name: Block internal-flavored paths

 # Hard CI gate. Internal content (positioning, competitive briefs, sales
-# playbooks, PMM/press drip, draft campaigns) lives in Molecule-AI/internal —
+# playbooks, PMM/press drip, draft campaigns) lives in molecule-ai/internal —
 # this public monorepo must never re-acquire those paths. CEO directive
 # 2026-04-23 after a fleet-wide audit found 79 internal files leaked here.
 #
@@ -135,7 +135,7 @@ jobs:
            echo "::error::Forbidden internal-flavored paths detected:"
            printf "$OFFENDING"
            echo ""
-            echo "These paths belong in Molecule-AI/internal, not this public repo."
+            echo "These paths belong in molecule-ai/internal, not this public repo."
            echo "See docs/internal-content-policy.md for canonical locations."
            echo ""
            echo "If your file is genuinely public-facing (e.g. a blog post"
@@ -0,0 +1,111 @@
+name: branch-protection drift check
+
+# Catches out-of-band edits to branch protection (UI clicks, manual gh
+# api PATCH from a one-off ops session) by comparing live state against
+# tools/branch-protection/apply.sh's desired state every day. Fails the
+# workflow when they drift; the failure is the signal.
+#
+# When it fails: re-run apply.sh to put the live state back to the
+# script's intent, OR update apply.sh to encode the new intent and
+# commit. Either way the script is the source of truth.
+
+on:
+  schedule:
+    # 14:00 UTC daily. Off-hours for most teams; gives a fresh signal
+    # at the start of every working day.
+    - cron: '0 14 * * *'
+  workflow_dispatch:
+  pull_request:
+    branches: [staging, main]
+    paths:
+      - 'tools/branch-protection/**'
+      - '.github/workflows/**'
+      - '.github/workflows/branch-protection-drift.yml'
+
+permissions:
+  contents: read
+
+jobs:
+  drift:
+    name: Branch protection drift
+    runs-on: ubuntu-latest
+    timeout-minutes: 5
+    steps:
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+
+      # Token strategy by trigger:
+      #
+      # - schedule (daily canary): hard-fail when the admin token is
+      #   missing. This is the *only* trigger where silent soft-skip is
+      #   dangerous — a missing secret on the cron run means the drift
+      #   gate has effectively disappeared with no human in the loop to
+      #   notice. Per feedback_schedule_vs_dispatch_secrets_hardening.md
+      #   the rule is "schedule/automated triggers must hard-fail".
+      #
+      # - pull_request (touching tools/branch-protection/**): soft-skip
+      #   with a prominent warning. A PR cannot retroactively drift the
+      #   live state — drift happens *between* PRs (UI clicks, manual
+      #   gh api PATCH) and is the schedule's job to catch. The PR-time
+      #   gate would only catch typos in apply.sh, which the apply.sh
+      #   *_payload unit tests catch better. A human is reviewing the
+      #   PR and will see the warning in the workflow log.
+      #
+      # - workflow_dispatch (operator one-off): soft-skip with warning,
+      #   so an operator can run a diagnostic without configuring the
+      #   secret first.
+      - name: Verify admin token present (hard-fail on schedule only)
+        env:
+          GH_TOKEN_FOR_ADMIN_API: ${{ secrets.GH_TOKEN_FOR_ADMIN_API }}
+        run: |
+          if [[ -n "$GH_TOKEN_FOR_ADMIN_API" ]]; then
+            echo "GH_TOKEN_FOR_ADMIN_API present — drift_check will run with admin scope."
+            exit 0
+          fi
+          if [[ "${{ github.event_name }}" == "schedule" ]]; then
+            echo "::error::GH_TOKEN_FOR_ADMIN_API secret missing on the daily canary." >&2
+            echo "" >&2
+            echo "The schedule run is the SoT for branch-protection drift detection." >&2
+            echo "Without admin scope it silently passes, hiding any out-of-band edits." >&2
+            echo "Set GH_TOKEN_FOR_ADMIN_API at Settings → Secrets and variables → Actions." >&2
+            exit 1
+          fi
+          echo "::warning::GH_TOKEN_FOR_ADMIN_API secret missing — drift_check will be SKIPPED."
+          echo "::warning::PR drift checks need repo-admin scope to read /branches/:b/protection."
+          echo "::warning::This is non-fatal: the daily schedule run is the canonical drift gate."
+          echo "SKIP_DRIFT_CHECK=1" >> "$GITHUB_ENV"
+
+      - name: Run drift check
+        if: env.SKIP_DRIFT_CHECK != '1'
+        env:
+          # Repo-admin scope, needed for /branches/:b/protection.
+          GH_TOKEN: ${{ secrets.GH_TOKEN_FOR_ADMIN_API }}
+        run: bash tools/branch-protection/drift_check.sh
+
+      # Self-test the parity script before running it on the real
+      # workflows — pins the script's classification logic against
+      # synthetic safe/unsafe/missing/unsafe-mix/matrix fixtures so a
+      # regression in the script can't false-pass on the production
+      # workflow audit. Cheap (~0.5s); always runs.
+      - name: Self-test check-name parity script
+        run: bash tools/branch-protection/test_check_name_parity.sh
+
+      # Check-name parity gate (#144 / saved memory
+      # feedback_branch_protection_check_name_parity).
+      #
+      # drift_check.sh asserts the live branch protection matches what
+      # apply.sh would set; check_name_parity.sh closes the orthogonal
+      # gap: it asserts every required check name in apply.sh maps to a
+      # workflow job whose "always emits this status" shape is intact.
+      #
+      # The two checks fail in different scenarios:
+      #
+      #   - drift_check fails → live state was rewritten out-of-band
+      #     (UI click, manual PATCH).
+      #   - check_name_parity fails → an apply.sh required name has no
+      #     emitter, OR the emitting workflow has a top-level paths:
+      #     filter without per-step if-gates (the silent-block shape).
+      #
+      # Cheap (~1s); runs without the admin token because it only reads
+      # apply.sh + .github/workflows/ from the checkout.
+      - name: Run check-name parity gate
+        run: bash tools/branch-protection/check_name_parity.sh
@@ -295,12 +295,16 @@ jobs:
          # See molecule-controlplane#420.
          leaks=()
          for slug in $orgs; do
-            code=$(curl -sS -o /tmp/canary-cleanup.out -w "%{http_code}" \
+            # Tempfile-routed -w + set +e/-e prevents curl-exit-code
+            # pollution of the captured status (lint-curl-status-capture.yml).
+            set +e
+            curl -sS -o /tmp/canary-cleanup.out -w "%{http_code}" \
              -X DELETE "$MOLECULE_CP_URL/cp/admin/tenants/$slug" \
              -H "Authorization: Bearer $ADMIN_TOKEN" \
              -H "Content-Type: application/json" \
-              -d "{\"confirm\":\"$slug\"}" \
-              || echo "000")
+              -d "{\"confirm\":\"$slug\"}" >/tmp/canary-cleanup.code
+            set -e
+            code=$(cat /tmp/canary-cleanup.code 2>/dev/null || echo "000")
            if [ "$code" = "200" ] || [ "$code" = "204" ]; then
              echo "[teardown] deleted $slug (HTTP $code)"
            else
@@ -108,7 +108,7 @@ jobs:
              echo
              echo "One or more canary secrets are unset (\`CANARY_TENANT_URLS\`, \`CANARY_ADMIN_TOKENS\`, \`CANARY_CP_SHARED_SECRET\`)."
              echo "Phase 2 canary fleet has not been stood up yet —"
-              echo "see [canary-tenants.md](https://github.com/Molecule-AI/molecule-controlplane/blob/main/docs/canary-tenants.md)."
+              echo "see [canary-tenants.md](https://git.moleculesai.app/molecule-ai/molecule-controlplane/blob/main/docs/canary-tenants.md)."
              echo
              echo "**Skipped — promote-to-latest will NOT auto-fire.** Dispatch \`promote-latest.yml\` manually when ready."
            } >> "$GITHUB_STEP_SUMMARY"
@@ -87,7 +87,7 @@ jobs:
        run: go mod download
      - if: needs.changes.outputs.platform == 'true'
        run: go build ./cmd/server
-      # CLI (molecli) moved to standalone repo: github.com/Molecule-AI/molecule-cli
+      # CLI (molecli) moved to standalone repo: github.com/molecule-ai/molecule-cli
      - if: needs.changes.outputs.platform == 'true'
        run: go vet ./... || true
      - if: needs.changes.outputs.platform == 'true'
@@ -165,7 +165,7 @@ jobs:
              # Strip the package-import prefix so we can match .coverage-allowlist.txt
              # entries written as paths relative to workspace-server/.
              # Handle both module paths: platform/workspace-server/... and platform/...
-              rel=$(echo "$file" | sed 's|^github.com/Molecule-AI/molecule-monorepo/platform/workspace-server/||; s|^github.com/Molecule-AI/molecule-monorepo/platform/||')
+              rel=$(echo "$file" | sed 's|^github.com/molecule-ai/molecule-monorepo/platform/workspace-server/||; s|^github.com/molecule-ai/molecule-monorepo/platform/||')

              if echo "$ALLOWLIST" | grep -qxF "$rel"; then
                echo "::warning file=workspace-server/$rel::Critical file at ${pct}% coverage (allowlisted, #1823) — fix before expiry."
@@ -235,7 +235,13 @@ jobs:
        run: npx vitest run --coverage
      - name: Upload coverage summary as artifact
        if: needs.changes.outputs.canvas == 'true' && always()
-        uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 # v4.6.2
+        # Pinned to v3 for Gitea act_runner v0.6 compatibility — v4+ uses
+        # the GHES 3.10+ artifact protocol that Gitea 1.22.x does NOT
+        # implement, surfacing as `GHESNotSupportedError: @actions/artifact
+        # v2.0.0+, upload-artifact@v4+ and download-artifact@v4+ are not
+        # currently supported on GHES`. Drop this pin when Gitea ships
+        # the v4 protocol (tracked: post-Gitea-1.23 followup).
+        uses: actions/upload-artifact@c6a366c94c3e0affe28c06c8df20a878f24da3cf # v3.2.2
        with:
          name: canvas-coverage-${{ github.run_id }}
          path: canvas/coverage/
@@ -243,8 +249,8 @@ jobs:
          if-no-files-found: warn

  # MCP Server + SDK removed from CI — now in standalone repos:
-  # - github.com/Molecule-AI/molecule-mcp-server (npm CI)
-  # - github.com/Molecule-AI/molecule-sdk-python (PyPI CI)
+  # - github.com/molecule-ai/molecule-mcp-server (npm CI)
+  # - github.com/molecule-ai/molecule-sdk-python (PyPI CI)

  # e2e-api job moved to .github/workflows/e2e-api.yml (issue #458).
  # It now has workflow-level concurrency (cancel-in-progress: false) so
@@ -272,6 +278,14 @@ jobs:
          find tests/e2e infra/scripts -type f -name '*.sh' -print0 \
            | xargs -0 shellcheck --severity=warning

+      - if: needs.changes.outputs.scripts == 'true'
+        name: Lint cleanup-trap hygiene (RFC #2873)
+        # Asserts every shell E2E test that calls `mktemp` also installs
+        # an EXIT trap. Catches the /tmp-leak class — a missing trap
+        # silently leaks scratch into CI runners (~10-100KB per run).
+        # See tests/e2e/lint_cleanup_traps.sh for the rule + fix pattern.
+        run: bash tests/e2e/lint_cleanup_traps.sh
+
      - if: needs.changes.outputs.scripts == 'true'
        name: Run E2E bash unit tests (no live infra)
        # Pure-bash unit tests for E2E helper libs (lib/*.sh). These pin
@@ -358,6 +372,73 @@ jobs:
      - if: needs.changes.outputs.python == 'true'
        run: python -m pytest --tb=short

-      # SDK + plugin validation moved to standalone repo:
-      # github.com/Molecule-AI/molecule-sdk-python
+      - if: needs.changes.outputs.python == 'true'
+        name: Per-file critical-path coverage (MCP / inbox / auth)
+        # MCP-critical Python files have a per-file floor on top of the
+        # 86% total floor in pytest.ini. Rationale (issue #2790, after
+        # the PR #2766 → PR #2771 cycle): the total floor averages ~6000
+        # lines, so a single MCP file could regress to ~50% with no
+        # complaint as long as other modules compensate. These five
+        # files handle multi-tenant routing + auth + inbox dispatch —
+        # a coverage drop here is the same risk shape as a Go-side
+        # workspace-server token/secrets file dropping below 10%.
+        #
+        # Floor 75% sits below current actuals (80-96%) so this gate is
+        # strictly additive — no existing PR fails. Ratchet plan in
+        # COVERAGE_FLOOR.md.
+        run: |
+          set -e
+          PER_FILE_FLOOR=75
+          CRITICAL_FILES=(
+            "a2a_mcp_server.py"
+            "mcp_cli.py"
+            "a2a_tools.py"
+            "a2a_tools_inbox.py"
+            "inbox.py"
+            "platform_auth.py"
+          )
+
+          # pytest already wrote .coverage; emit a JSON view scoped to
+          # the critical files so jq/python can read the per-file pct
+          # without parsing tabular text. --include uses fnmatch, and
+          # the leading "*" allows the file to live anywhere under the
+          # workspace root (today they sit at workspace/<name>.py).
+          INCLUDES=$(printf '*%s,' "${CRITICAL_FILES[@]}")
+          INCLUDES="${INCLUDES%,}"
+          python -m coverage json -o /tmp/critical-cov.json --include="$INCLUDES"
+
+          FAILED=0
+          for f in "${CRITICAL_FILES[@]}"; do
+            # Match by top-level path key (e.g. "a2a_tools.py", not
+            # "builtin_tools/a2a_tools.py" — different file at 100%).
+            # The keys in coverage.json are paths relative to the run
+            # cwd (workspace/), so the critical-path entry sits at the
+            # bare basename.
+            pct=$(jq -r --arg f "$f" '.files | to_entries | map(select(.key == $f)) | .[0].value.summary.percent_covered // "MISSING"' /tmp/critical-cov.json)
+            if [ "$pct" = "MISSING" ]; then
+              echo "::error file=workspace/$f::No coverage data — file may have moved or test exclusion mis-set."
+              FAILED=$((FAILED+1))
+              continue
+            fi
+            echo "$f: ${pct}%"
+            if awk "BEGIN{exit !($pct < $PER_FILE_FLOOR)}"; then
+              echo "::error file=workspace/$f::${pct}% < ${PER_FILE_FLOOR}% per-file floor (MCP critical path). See COVERAGE_FLOOR.md."
+              FAILED=$((FAILED+1))
+            fi
+          done
+
+          if [ "$FAILED" -gt 0 ]; then
+            echo ""
+            echo "$FAILED MCP critical-path file(s) below the ${PER_FILE_FLOOR}% per-file floor."
+            echo "These paths handle multi-tenant routing, auth tokens, and inbox dispatch."
+            echo "A coverage drop here is the same risk shape as Go-side tokens/secrets files"
+            echo "dropping below 10% (see COVERAGE_FLOOR.md). Either:"
+            echo "  (a) add tests to raise coverage back above ${PER_FILE_FLOOR}%, or"
+            echo "  (b) if this is unavoidable historical debt, file an issue and propose"
+            echo "      adjusting the floor with rationale in COVERAGE_FLOOR.md."
+            exit 1
+          fi
+
+      # SDK + plugin validation moved to standalone repo:
+      # github.com/molecule-ai/molecule-sdk-python

@@ -1,36 +1,92 @@
 name: CodeQL

-# Controls CodeQL scan triggers for this repo.
+# Stub workflow — CodeQL Action is structurally incompatible with Gitea
+# Actions (post-2026-05-06 SCM migration off GitHub).
 #
-# GitHub's "Code quality" default setup (the UI-configured one) is
-# hardcoded to only scan the default branch — on this repo that's
-# `staging`, so PRs promoting staging→main would otherwise never be
-# scanned. This workflow fills that gap by explicitly scanning both
-# branches on push and PR.
+# Why this is a stub, not a real CodeQL run:
 #
-# Runs on ubuntu-latest (GHA-hosted — public repo, free). GHAS is NOT
-# enabled on this repo, so results are not uploaded to the Security
-# tab — the scan fails the PR check on findings, and the SARIF is
-# kept as a workflow artifact for triage.
+# 1. github/codeql-action/init@v4 hits api.github.com endpoints
+#    (CodeQL CLI bundle download + query-pack registry + telemetry)
+#    that Gitea 1.22.x does NOT proxy. The act_runner has
+#    GITHUB_SERVER_URL=https://git.moleculesai.app correctly set
+#    (per saved memory feedback_act_runner_github_server_url and
+#    /config.yaml on the operator host), but the Gitea API surface
+#    simply does not implement the codeql-action bundle endpoints.
+#    Observed in run 1d/3101 (2026-05-07): "::error::404 page not
+#    found" inside the Initialize CodeQL step, before any analysis.
+#
+# 2. PR #35 attempted to mark `continue-on-error: true` at the JOB
+#    level (correct YAML structure). Gitea 1.22.6 does NOT propagate
+#    job-level continue-on-error to the commit-status API — every
+#    matrix leg still posts `failure` to the status surface, which
+#    keeps OVERALL=failure on every push to main + staging and
+#    blocks visual auto-promote signals (#156).
+#
+# 3. Hongming policy decision (2026-05-07, task #156): CodeQL is
+#    ADVISORY, not blocking, on Gitea Actions. We do not block PR
+#    merge or staging→main promotion on CodeQL findings until we
+#    have a Gitea-compatible static-analysis pipeline.
+#
+# What this stub preserves:
+#
+# - Workflow name `CodeQL` (referenced by auto-promote-staging.yml
+#   line 67 as a workflow_run gate — must stay stable).
+# - Job name template `Analyze (${{ matrix.language }})` and the
+#   3-leg matrix (go, javascript-typescript, python). Branch
+#   protection / required-check parity (#144) keys on these
+#   exact context names.
+# - merge_group + push + pull_request + schedule triggers, so the
+#   merge-queue check name still resolves (per saved memory
+#   feedback_branch_protection_check_name_parity).
+#
+# Re-enabling real analysis (future work):
+#
+# - Option A: self-hosted Semgrep / OpenGrep via a custom action
+#   that doesn't hit api.github.com. Tracked behind #156 follow-up.
+# - Option B: Sonatype Nexus IQ or similar, called from a step
+#   that uses the Gitea-issued token only.
+# - Option C: re-host this workflow on a small GitHub mirror used
+#   ONLY for SAST (push-mirrored from Gitea). Acceptable trade-off
+#   if/when payment is restored on a non-suspended GitHub org —
+#   but per saved memory feedback_no_single_source_of_truth, we
+#   should design for multi-vendor backup, not GitHub-only SAST.
+#
+# Until one of those lands, this stub keeps commit-status green so
+# the auto-promote chain isn't permanently red on a tool we cannot
+# actually run.
+#
+# Security policy: ADVISORY. We accept the residual risk of un-scanned
+# pushes during this window. Compensating controls in place:
+#   - secret-scan.yml runs on every push (active, blocks on hits)
+#   - block-internal-paths.yml blocks forbidden file paths
+#   - lint-curl-status-capture.yml catches one specific class of bug
+#   - branch-protection-drift.yml + the merge_group required-checks
+#     parity keep the gate surface stable
+# These are not equivalent to CodeQL coverage. Status of the
+# replacement plan is tracked in #156.

 on:
  push:
    branches: [main, staging]
  pull_request:
    branches: [main, staging]
-  # GitHub merge queue fires `merge_group` for the queue's pre-merge CI run.
-  # Required so CodeQL Analyze checks get a real result on the queued
-  # commit instead of a false-green. Event only fires once merge queue is
-  # enabled on the target branch — safe to add unconditionally.
+  # Required so the matrix legs emit a real result on the queued
+  # commit instead of a false-green when merge queue is enabled.
+  # Per saved memory feedback_branch_protection_check_name_parity:
+  # path-filtered / matrix workflows MUST emit the protected name
+  # via a job that always runs.
  merge_group:
    types: [checks_requested]
  schedule:
-    # Weekly run picks up findings in code that hasn't been touched.
+    # Weekly heartbeat. Cheap on a stub (the no-op job is ~5s) but
+    # keeps the workflow visible in Gitea's Actions UI so the next
+    # operator notices it's a stub instead of a missing surface.
    - cron: '30 1 * * 0'

-# Workflow-level concurrency: only one CodeQL run per branch/PR at a time.
-# `cancel-in-progress: false` queues new runs so a quick follow-up push
-# doesn't nuke a 45-min analysis mid-flight.
+# Workflow-level concurrency: only one stub run per branch/PR at a
+# time. cancel-in-progress: false because a quick follow-up push
+# shouldn't kill an in-flight run — even though the stub is fast,
+# the contract should match a real CodeQL run for when we re-enable.
 concurrency:
  group: codeql-${{ github.ref }}
  cancel-in-progress: false
@@ -38,13 +94,17 @@ concurrency:
 permissions:
  actions: read
  contents: read
-  # No security-events: write — we don't call the upload API.
+  # No security-events: write — we don't call the upload API anyway,
+  # GHAS isn't on Gitea.

 jobs:
  analyze:
+    # Job NAME shape is load-bearing — auto-promote-staging.yml +
+    # branch protection both key on `Analyze (${{ matrix.language }})`.
+    # Do NOT rename without coordinating both surfaces.
    name: Analyze (${{ matrix.language }})
    runs-on: ubuntu-latest
-    timeout-minutes: 45
+    timeout-minutes: 5

    strategy:
      fail-fast: false
@@ -52,77 +112,25 @@ jobs:
        language: [go, javascript-typescript, python]

    steps:
-      - name: Checkout
-        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
-
-      - name: Checkout sibling plugin repo
-        # Same reasoning as publish-workspace-server-image.yml — the Go
-        # module's replace directive needs the plugin source so
-        # CodeQL's "go build" phase can resolve.
-        if: matrix.language == 'go'
-        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
-        with:
-          repository: Molecule-AI/molecule-ai-plugin-github-app-auth
-          path: molecule-ai-plugin-github-app-auth
-          token: ${{ secrets.PLUGIN_REPO_PAT || secrets.GITHUB_TOKEN }}
-
-      # jq is pre-installed on ubuntu-latest — no setup step needed.
-
-      - name: Initialize CodeQL
-        uses: github/codeql-action/init@95e58e9a2cdfd71adc6e0353d5c52f41a045d225 # v4.35.2
-        with:
-          languages: ${{ matrix.language }}
-          # security-extended widens past the default to include the
-          # full security-query set for a public SaaS surface.
-          queries: security-extended
-
-      - name: Autobuild
-        uses: github/codeql-action/autobuild@95e58e9a2cdfd71adc6e0353d5c52f41a045d225 # v4.35.2
-
-      - name: Perform CodeQL Analysis
-        id: analyze
-        uses: github/codeql-action/analyze@95e58e9a2cdfd71adc6e0353d5c52f41a045d225 # v4.35.2
-        with:
-          category: "/language:${{ matrix.language }}"
-          # upload: never — GHAS isn't enabled on this repo, so the
-          # upload API 403s. Write SARIF locally instead.
-          upload: never
-          output: sarif-results/${{ matrix.language }}
-
-      - name: Parse SARIF + fail on findings
-        # The analyze step writes <database>.sarif into the output
-        # directory — database name is the short CodeQL lang id, not
-        # the matrix value (e.g. "javascript-typescript" →
-        # javascript.sarif), so glob rather than hardcode.
-        # Filter to error/warning severity: security-extended emits
-        # "note" rows for informational findings we don't want to fail
-        # the build over.
+      # Single-step stub: log the policy decision + emit success.
+      # Exit 0 explicitly so the commit-status API records `success`
+      # for each of the three matrix legs.
+      - name: CodeQL stub (advisory, non-blocking on Gitea)
        shell: bash
        run: |
          set -euo pipefail
-          dir="sarif-results/${{ matrix.language }}"
-          sarif=$(ls "$dir"/*.sarif 2>/dev/null | head -1 || true)
-          if [ -z "$sarif" ] || [ ! -f "$sarif" ]; then
-            echo "::error::No SARIF file found under $dir"
-            ls -la "$dir" 2>/dev/null || true
-            exit 1
-          fi
-          echo "Parsing $sarif"
-          count=$(jq '[.runs[].results[] | select(.level == "error" or .level == "warning")] | length' "$sarif")
-          echo "CodeQL findings (error+warning) for ${{ matrix.language }}: $count"
-          if [ "$count" -gt 0 ]; then
-            echo "::error::CodeQL found $count issues. Details below; full SARIF in the artifact."
-            jq -r '.runs[].results[] | select(.level == "error" or .level == "warning") | "  - [\(.level)] \(.ruleId // "?"): \(.message.text // "(no message)") @ \(.locations[0].physicalLocation.artifactLocation.uri // "?"):\(.locations[0].physicalLocation.region.startLine // "?")"' "$sarif"
-            exit 1
-          fi
-
-      - name: Upload SARIF artifact
-        # Keep SARIF around on success + failure so triagers can diff.
-        # 14-day retention — longer than default 3, short enough not
-        # to bloat quota.
-        if: always()
-        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
-        with:
-          name: codeql-sarif-${{ matrix.language }}
-          path: sarif-results/${{ matrix.language }}/
-          retention-days: 14
+          cat <<EOF
+          CodeQL is currently ADVISORY on Gitea Actions (post-2026-05-06).
+          Language matrix leg: ${{ matrix.language }}
+          Reason: github/codeql-action/init@v4 calls api.github.com
+                  bundle endpoints that Gitea 1.22.x does not implement.
+                  Observed: "::error::404 page not found" in the Init
+                  CodeQL step on every prior run.
+          Policy: per Hongming decision 2026-05-07 (#156), CodeQL is
+                  non-blocking until a Gitea-compatible SAST pipeline
+                  lands. See workflow file header for replacement
+                  options + compensating controls.
+          Status: emitting success so auto-promote isn't permanently
+                  red on a tool we cannot actually run today.
+          EOF
+          echo "::notice::CodeQL ${{ matrix.language }} — advisory stub, success."
@@ -12,6 +12,59 @@ name: E2E API Smoke Test
 # spending CI cycles. See the in-job comment on the `e2e-api` job for
 # why this is one job (not two-jobs-sharing-name) and the 2026-04-29
 # PR #2264 incident that drove the consolidation.
+#
+# Parallel-safety (Class B Hongming-owned CICD red sweep, 2026-05-08)
+# -------------------------------------------------------------------
+# Same substrate hazard as PR #98 (handlers-postgres-integration). Our
+# Gitea act_runner runs with `container.network: host` (operator host
+# `/opt/molecule/runners/config.yaml`), which means:
+#
+#   * Two concurrent runs both try to bind their `-p 15432:5432` /
+#     `-p 16379:6379` host ports — the second postgres/redis FATALs
+#     with `Address in use` and `docker run` returns exit 125 with
+#     `Conflict. The container name "/molecule-ci-postgres" is already
+#     in use by container ...`. Verified in run a7/2727 on 2026-05-07.
+#   * The fixed container names `molecule-ci-postgres` / `-redis` (the
+#     pre-fix shape) collide on name AS WELL AS port. The cleanup-with-
+#     `docker rm -f` at the start of the second job KILLS the first
+#     job's still-running postgres/redis.
+#
+# Fix shape (mirrors PR #98's bridge-net pattern, adapted because
+# platform-server is a Go binary on the host, not a containerised
+# step):
+#
+#   1. Unique container names per run:
+#         pg-e2e-api-${RUN_ID}-${RUN_ATTEMPT}
+#         redis-e2e-api-${RUN_ID}-${RUN_ATTEMPT}
+#      `${RUN_ID}-${RUN_ATTEMPT}` is unique even across reruns of the
+#      same run_id.
+#   2. Ephemeral host port per run (`-p 0:5432`), then read the actual
+#      bound port via `docker port` and export DATABASE_URL/REDIS_URL
+#      pointing at it. No fixed host-port → no port collision.
+#   3. `127.0.0.1` (NOT `localhost`) in URLs — IPv6 first-resolve was
+#      the original flake fixed in #92 and the script's still IPv6-
+#      enabled.
+#   4. `if: always()` cleanup so containers don't leak when test steps
+#      fail.
+#
+# Issue #94 items #2 + #3 (also fixed here):
+#   * Pre-pull `alpine:latest` so the platform-server's provisioner
+#     (`internal/handlers/container_files.go`) can stand up its
+#     ephemeral token-write helper without a daemon.io round-trip.
+#   * Create `molecule-monorepo-net` bridge network if missing so the
+#     provisioner's container.HostConfig {NetworkMode: ...} attach
+#     succeeds.
+# Item #1 (timeouts) — evidence on recent runs (77/3191, ae/4270, 0e/
+# 2318) shows Postgres ready in 3s, Redis in 1s, Platform in 1s when
+# they DO come up. Timeouts are not the bottleneck; not bumped.
+#
+# Item explicitly NOT fixed here: failing test `Status back online`
+# fails because the platform's langgraph workspace template image
+# (ghcr.io/molecule-ai/workspace-template-langgraph:latest) returns
+# 403 Forbidden post-2026-05-06 GitHub org suspension. That is a
+# template-registry resolution issue (ADR-002 / local-build mode) and
+# belongs in a separate change that touches workspace-server, not
+# this workflow file.

 on:
  push:
@@ -78,11 +131,14 @@ jobs:
    runs-on: ubuntu-latest
    timeout-minutes: 15
    env:
-      DATABASE_URL: postgres://dev:dev@localhost:15432/molecule?sslmode=disable
-      REDIS_URL: redis://localhost:16379
+      # Unique per-run container names so concurrent runs on the host-
+      # network act_runner don't collide on name OR port.
+      # `${RUN_ID}-${RUN_ATTEMPT}` stays unique across reruns of the
+      # same run_id. PORT is set later (after docker port lookup) since
+      # we let Docker assign an ephemeral host port.
+      PG_CONTAINER: pg-e2e-api-${{ github.run_id }}-${{ github.run_attempt }}
+      REDIS_CONTAINER: redis-e2e-api-${{ github.run_id }}-${{ github.run_attempt }}
      PORT: "8080"
-      PG_CONTAINER: molecule-ci-postgres
-      REDIS_CONTAINER: molecule-ci-redis
    steps:
      - name: No-op pass (paths filter excluded this commit)
        if: needs.detect-changes.outputs.api != 'true'
@@ -97,11 +153,53 @@ jobs:
          go-version: 'stable'
          cache: true
          cache-dependency-path: workspace-server/go.sum
+      - name: Pre-pull alpine + ensure provisioner network (Issue #94 items #2 + #3)
+        if: needs.detect-changes.outputs.api == 'true'
+        run: |
+          # Provisioner uses alpine:latest for ephemeral token-write
+          # containers (workspace-server/internal/handlers/container_files.go).
+          # Pre-pull so the first provision in test_api.sh doesn't race
+          # the daemon's pull cache. Idempotent — `docker pull` is a no-op
+          # when the image is already present.
+          docker pull alpine:latest >/dev/null
+          # Provisioner attaches workspace containers to
+          # molecule-monorepo-net (workspace-server/internal/provisioner/
+          # provisioner.go::DefaultNetwork). The bridge already exists on
+          # the operator host's docker daemon — `network create` is
+          # idempotent via `|| true`.
+          docker network create molecule-monorepo-net >/dev/null 2>&1 || true
+          echo "alpine:latest pre-pulled; molecule-monorepo-net ensured."
      - name: Start Postgres (docker)
        if: needs.detect-changes.outputs.api == 'true'
        run: |
+          # Defensive cleanup — only matches THIS run's container name,
+          # so it cannot kill a sibling run's postgres. (Pre-fix the
+          # name was static and this rm hit other runs' containers.)
          docker rm -f "$PG_CONTAINER" 2>/dev/null || true
-          docker run -d --name "$PG_CONTAINER" -e POSTGRES_USER=dev -e POSTGRES_PASSWORD=dev -e POSTGRES_DB=molecule -p 15432:5432 postgres:16
+          # `-p 0:5432` requests an ephemeral host port; we read it back
+          # below and export DATABASE_URL.
+          docker run -d --name "$PG_CONTAINER" \
+            -e POSTGRES_USER=dev -e POSTGRES_PASSWORD=dev -e POSTGRES_DB=molecule \
+            -p 0:5432 postgres:16 >/dev/null
+          # Resolve the host-side port assignment. `docker port` prints
+          # `0.0.0.0:NNNN` (and on host-net runners may also print an
+          # IPv6 line — take the first IPv4 line).
+          PG_PORT=$(docker port "$PG_CONTAINER" 5432/tcp | awk -F: '/^0\.0\.0\.0:/ {print $2; exit}')
+          if [ -z "$PG_PORT" ]; then
+            # Fallback: any first line. Some Docker versions print only
+            # one line.
+            PG_PORT=$(docker port "$PG_CONTAINER" 5432/tcp | head -1 | awk -F: '{print $NF}')
+          fi
+          if [ -z "$PG_PORT" ]; then
+            echo "::error::Could not resolve host port for $PG_CONTAINER"
+            docker port "$PG_CONTAINER" 5432/tcp || true
+            docker logs "$PG_CONTAINER" || true
+            exit 1
+          fi
+          # 127.0.0.1 (NOT localhost) — IPv6 first-resolve flake (#92).
+          echo "PG_PORT=${PG_PORT}" >> "$GITHUB_ENV"
+          echo "DATABASE_URL=postgres://dev:dev@127.0.0.1:${PG_PORT}/molecule?sslmode=disable" >> "$GITHUB_ENV"
+          echo "Postgres host port: ${PG_PORT}"
          for i in $(seq 1 30); do
            if docker exec "$PG_CONTAINER" pg_isready -U dev >/dev/null 2>&1; then
              echo "Postgres ready after ${i}s"
@@ -116,7 +214,20 @@ jobs:
        if: needs.detect-changes.outputs.api == 'true'
        run: |
          docker rm -f "$REDIS_CONTAINER" 2>/dev/null || true
-          docker run -d --name "$REDIS_CONTAINER" -p 16379:6379 redis:7
+          docker run -d --name "$REDIS_CONTAINER" -p 0:6379 redis:7 >/dev/null
+          REDIS_PORT=$(docker port "$REDIS_CONTAINER" 6379/tcp | awk -F: '/^0\.0\.0\.0:/ {print $2; exit}')
+          if [ -z "$REDIS_PORT" ]; then
+            REDIS_PORT=$(docker port "$REDIS_CONTAINER" 6379/tcp | head -1 | awk -F: '{print $NF}')
+          fi
+          if [ -z "$REDIS_PORT" ]; then
+            echo "::error::Could not resolve host port for $REDIS_CONTAINER"
+            docker port "$REDIS_CONTAINER" 6379/tcp || true
+            docker logs "$REDIS_CONTAINER" || true
+            exit 1
+          fi
+          echo "REDIS_PORT=${REDIS_PORT}" >> "$GITHUB_ENV"
+          echo "REDIS_URL=redis://127.0.0.1:${REDIS_PORT}" >> "$GITHUB_ENV"
+          echo "Redis host port: ${REDIS_PORT}"
          for i in $(seq 1 15); do
            if docker exec "$REDIS_CONTAINER" redis-cli ping 2>/dev/null | grep -q PONG; then
              echo "Redis ready after ${i}s"
@@ -135,13 +246,15 @@ jobs:
        if: needs.detect-changes.outputs.api == 'true'
        working-directory: workspace-server
        run: |
+          # DATABASE_URL + REDIS_URL exported by the start-postgres /
+          # start-redis steps point at this run's per-run host ports.
          ./platform-server > platform.log 2>&1 &
          echo $! > platform.pid
      - name: Wait for /health
        if: needs.detect-changes.outputs.api == 'true'
        run: |
          for i in $(seq 1 30); do
-            if curl -sf http://localhost:8080/health > /dev/null; then
+            if curl -sf http://127.0.0.1:8080/health > /dev/null; then
              echo "Platform up after ${i}s"
              exit 0
            fi
@@ -172,6 +285,9 @@ jobs:
      - name: Run poll-mode + since_id cursor E2E (#2339)
        if: needs.detect-changes.outputs.api == 'true'
        run: bash tests/e2e/test_poll_mode_e2e.sh
+      - name: Run poll-mode chat upload E2E (RFC #2891)
+        if: needs.detect-changes.outputs.api == 'true'
+        run: bash tests/e2e/test_poll_mode_chat_upload_e2e.sh
      - name: Dump platform log on failure
        if: failure() && needs.detect-changes.outputs.api == 'true'
        run: cat workspace-server/platform.log || true
@@ -182,6 +298,9 @@ jobs:
            kill "$(cat workspace-server/platform.pid)" 2>/dev/null || true
          fi
      - name: Stop service containers
+        # always() so containers don't leak when test steps fail. The
+        # cleanup is best-effort: if the container is already gone
+        # (e.g. concurrent rerun race), don't fail the job.
        if: always() && needs.detect-changes.outputs.api == 'true'
        run: |
          docker rm -f "$PG_CONTAINER" 2>/dev/null || true
@@ -22,9 +22,9 @@ on:
  # spending CI cycles. See e2e-api.yml for the rationale on why this
  # is a single job rather than two-jobs-sharing-name.
  push:
-    branches: [main, staging]
+    branches: [main]
  pull_request:
-    branches: [main, staging]
+    branches: [main]
  workflow_dispatch:
  schedule:
    # Weekly on Sunday 08:00 UTC — catches Chrome / Playwright / Next.js
@@ -139,7 +139,11 @@ jobs:

      - name: Upload Playwright report on failure
        if: failure() && needs.detect-changes.outputs.canvas == 'true'
-        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
+        # Pinned to v3 for Gitea act_runner v0.6 compatibility — v4+ uses
+        # the GHES 3.10+ artifact protocol that Gitea 1.22.x does NOT
+        # implement (see ci.yml upload step for the canonical error
+        # cite). Drop this pin when Gitea ships the v4 protocol.
+        uses: actions/upload-artifact@c6a366c94c3e0affe28c06c8df20a878f24da3cf # v3.2.2
        with:
          name: playwright-report-staging
          path: canvas/playwright-report-staging/
@@ -147,7 +151,8 @@ jobs:

      - name: Upload screenshots on failure
        if: failure() && needs.detect-changes.outputs.canvas == 'true'
-        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
+        # Pinned to v3 for Gitea act_runner v0.6 compatibility (see above).
+        uses: actions/upload-artifact@c6a366c94c3e0affe28c06c8df20a878f24da3cf # v3.2.2
        with:
          name: playwright-screenshots
          path: canvas/test-results/
@@ -192,12 +197,16 @@ jobs:
          # cleanup miss shouldn't fail-flag the canvas test when the
          # actual smoke check passed; the sweeper is the safety net.
          # See molecule-controlplane#420.
-          code=$(curl -sS -o /tmp/canvas-cleanup.out -w "%{http_code}" \
+          # Tempfile-routed -w + set +e/-e prevents curl-exit-code
+          # pollution of the captured status (lint-curl-status-capture.yml).
+          set +e
+          curl -sS -o /tmp/canvas-cleanup.out -w "%{http_code}" \
            -X DELETE "$MOLECULE_CP_URL/cp/admin/tenants/$slug" \
            -H "Authorization: Bearer $ADMIN_TOKEN" \
            -H "Content-Type: application/json" \
-            -d "{\"confirm\":\"$slug\"}" \
-            || echo "000")
+            -d "{\"confirm\":\"$slug\"}" >/tmp/canvas-cleanup.code
+          set -e
+          code=$(cat /tmp/canvas-cleanup.code 2>/dev/null || echo "000")
          if [ "$code" = "200" ] || [ "$code" = "204" ]; then
            echo "[teardown] deleted $slug (HTTP $code)"
          else
@@ -32,7 +32,7 @@ name: E2E Staging External Runtime

 on:
  push:
-    branches: [staging, main]
+    branches: [main]
    paths:
      - 'workspace-server/internal/handlers/workspace.go'
      - 'workspace-server/internal/handlers/registry.go'
@@ -44,7 +44,7 @@ on:
      - 'tests/e2e/test_staging_external_runtime.sh'
      - '.github/workflows/e2e-staging-external.yml'
  pull_request:
-    branches: [staging, main]
+    branches: [main]
    paths:
      - 'workspace-server/internal/handlers/workspace.go'
      - 'workspace-server/internal/handlers/registry.go'
@@ -159,12 +159,16 @@ jobs:
            # leaked. Sweeper catches the rest within ~45 min.
            leaks=()
            for slug in $orgs; do
-              code=$(curl -sS -o /tmp/external-cleanup.out -w "%{http_code}" \
+              # Tempfile-routed -w + set +e/-e prevents curl-exit-code
+              # pollution of the captured status (lint-curl-status-capture.yml).
+              set +e
+              curl -sS -o /tmp/external-cleanup.out -w "%{http_code}" \
                -X DELETE "$MOLECULE_CP_URL/cp/admin/tenants/$slug" \
                -H "Authorization: Bearer $ADMIN_TOKEN" \
                -H "Content-Type: application/json" \
-                -d "{\"confirm\":\"$slug\"}" \
-                || echo "000")
+                -d "{\"confirm\":\"$slug\"}" >/tmp/external-cleanup.code
+              set -e
+              code=$(cat /tmp/external-cleanup.code 2>/dev/null || echo "000")
              if [ "$code" = "200" ] || [ "$code" = "204" ]; then
                echo "[teardown] deleted $slug (HTTP $code)"
              else
@@ -20,13 +20,12 @@ name: E2E Staging SaaS (full lifecycle)
 #     via the same paths watcher that e2e-api.yml uses)

 on:
-  # Fire on staging push too — previously this only ran on main, which
-  # meant the most thorough end-to-end test caught regressions AFTER
-  # they shipped to staging (and then to the auto-promote PR). Running
-  # on staging push catches them BEFORE the staging→main promotion
-  # opens, so a green canary into auto-promote is more meaningful.
+  # Trunk-based (Phase 3 of internal#81): main is the only branch.
+  # Previously this fired on staging push too because staging was a
+  # superset of main and ran the gate ahead of auto-promote; with no
+  # staging branch, main is where E2E gates the deploy.
  push:
-    branches: [staging, main]
+    branches: [main]
    paths:
      - 'workspace-server/internal/handlers/registry.go'
      - 'workspace-server/internal/handlers/workspace_provision.go'
@@ -36,7 +35,7 @@ on:
      - 'tests/e2e/test_staging_full_saas.sh'
      - '.github/workflows/e2e-staging-saas.yml'
  pull_request:
-    branches: [staging, main]
+    branches: [main]
    paths:
      - 'workspace-server/internal/handlers/registry.go'
      - 'workspace-server/internal/handlers/workspace_provision.go'
@@ -224,12 +223,16 @@ jobs:
          leaks=()
          for slug in $orgs; do
            echo "Safety-net teardown: $slug"
-            code=$(curl -sS -o /tmp/saas-cleanup.out -w "%{http_code}" \
+            # Tempfile-routed -w + set +e/-e prevents curl-exit-code
+            # pollution of the captured status (lint-curl-status-capture.yml).
+            set +e
+            curl -sS -o /tmp/saas-cleanup.out -w "%{http_code}" \
              -X DELETE "$MOLECULE_CP_URL/cp/admin/tenants/$slug" \
              -H "Authorization: Bearer $ADMIN_TOKEN" \
              -H "Content-Type: application/json" \
-              -d "{\"confirm\":\"$slug\"}" \
-              || echo "000")
+              -d "{\"confirm\":\"$slug\"}" >/tmp/saas-cleanup.code
+            set -e
+            code=$(cat /tmp/saas-cleanup.code 2>/dev/null || echo "000")
            if [ "$code" = "200" ] || [ "$code" = "204" ]; then
              echo "[teardown] deleted $slug (HTTP $code)"
            else
@@ -148,12 +148,16 @@ jobs:
          # safety net within ~45 min.
          leaks=()
          for slug in $orgs; do
-            code=$(curl -sS -o /tmp/sanity-cleanup.out -w "%{http_code}" \
+            # Tempfile-routed -w + set +e/-e prevents curl-exit-code
+            # pollution of the captured status (lint-curl-status-capture.yml).
+            set +e
+            curl -sS -o /tmp/sanity-cleanup.out -w "%{http_code}" \
              -X DELETE "$MOLECULE_CP_URL/cp/admin/tenants/$slug" \
              -H "Authorization: Bearer $ADMIN_TOKEN" \
              -H "Content-Type: application/json" \
-              -d "{\"confirm\":\"$slug\"}" \
-              || echo "000")
+              -d "{\"confirm\":\"$slug\"}" >/tmp/sanity-cleanup.code
+            set -e
+            code=$(cat /tmp/sanity-cleanup.code 2>/dev/null || echo "000")
            if [ "$code" = "200" ] || [ "$code" = "204" ]; then
              echo "[teardown] deleted $slug (HTTP $code)"
            else
@@ -0,0 +1,252 @@
+name: Handlers Postgres Integration
+
+# Real-Postgres integration tests for workspace-server/internal/handlers/.
+# Triggered on every PR/push that touches the handlers package.
+#
+# Why this workflow exists
+# ------------------------
+# Strict-sqlmock unit tests pin which SQL statements fire — they're fast
+# and let us iterate without a DB. But sqlmock CANNOT detect bugs that
+# depend on the row state AFTER the SQL runs. The result_preview-lost
+# bug shipped to staging in PR #2854 because every unit test was
+# satisfied with "an UPDATE statement fired" — none verified the row's
+# preview field actually landed. The local-postgres E2E that retrofit
+# self-review caught it took 2 minutes to set up and would have caught
+# the bug at PR-time.
+#
+# Why this workflow does NOT use `services: postgres:` (Class B fix)
+# ------------------------------------------------------------------
+# Our act_runner config has `container.network: host` (operator host
+# /opt/molecule/runners/config.yaml), which act_runner applies to BOTH
+# the job container AND every service container. With host-net, two
+# concurrent runs of this workflow both try to bind 0.0.0.0:5432 — the
+# second postgres FATALs with `could not create any TCP/IP sockets:
+# Address in use`, and Docker auto-removes it (act_runner sets
+# AutoRemove:true on service containers). By the time the migrations
+# step runs `psql`, the postgres container is gone, hence
+# `Connection refused` then `failed to remove container: No such
+# container` at cleanup time.
+#
+# Per-job `container.network` override is silently ignored by
+# act_runner — `--network and --net in the options will be ignored.`
+# appears in the runner log. Documented constraint.
+#
+# So we sidestep `services:` entirely. The job container still uses
+# host-net (inherited from runner config; required for cache server
+# discovery on the bridge IP 172.18.0.17:42631). We launch a sibling
+# postgres on the existing `molecule-monorepo-net` bridge with a
+# UNIQUE name per run — `pg-handlers-${RUN_ID}-${RUN_ATTEMPT}` — and
+# read its bridge IP via `docker inspect`. A host-net job container
+# can reach a bridge-net container directly via the bridge IP (verified
+# manually on operator host 2026-05-08).
+#
+# Trade-offs vs. the original `services:` shape:
+#   + No host-port collision; N parallel runs share the bridge cleanly
+#   + `if: always()` cleanup runs even on test-step failure
+#   - One more step in the workflow (+~3 lines)
+#   - Requires `molecule-monorepo-net` to exist on the operator host
+#     (it does; declared in docker-compose.yml + docker-compose.infra.yml)
+#
+# Class B Hongming-owned CICD red sweep, 2026-05-08.
+#
+# Cost: ~30s job (postgres pull from cache + go build + 4 tests).
+
+on:
+  push:
+    branches: [main, staging]
+  pull_request:
+    branches: [main, staging]
+  merge_group:
+    types: [checks_requested]
+  workflow_dispatch:
+
+concurrency:
+  group: handlers-pg-integ-${{ github.event.pull_request.head.sha || github.sha }}
+  cancel-in-progress: false
+
+jobs:
+  detect-changes:
+    name: detect-changes
+    runs-on: ubuntu-latest
+    outputs:
+      handlers: ${{ steps.filter.outputs.handlers }}
+    steps:
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+      - uses: dorny/paths-filter@fbd0ab8f3e69293af611ebaee6363fc25e6d187d # v4.0.1
+        id: filter
+        with:
+          filters: |
+            handlers:
+              - 'workspace-server/internal/handlers/**'
+              - 'workspace-server/internal/wsauth/**'
+              - 'workspace-server/migrations/**'
+              - '.github/workflows/handlers-postgres-integration.yml'
+
+  # Single-job-with-per-step-if pattern: always runs to satisfy the
+  # required-check name on branch protection; real work gates on the
+  # paths filter. See ci.yml's Platform (Go) for the same shape.
+  integration:
+    name: Handlers Postgres Integration
+    needs: detect-changes
+    runs-on: ubuntu-latest
+    env:
+      # Unique name per run so concurrent jobs don't collide on the
+      # bridge network. ${RUN_ID}-${RUN_ATTEMPT} is unique even across
+      # workflow_dispatch reruns of the same run_id.
+      PG_NAME: pg-handlers-${{ github.run_id }}-${{ github.run_attempt }}
+      # Bridge network already exists on the operator host (declared
+      # in docker-compose.yml + docker-compose.infra.yml).
+      PG_NETWORK: molecule-monorepo-net
+    defaults:
+      run:
+        working-directory: workspace-server
+    steps:
+      - if: needs.detect-changes.outputs.handlers != 'true'
+        working-directory: .
+        run: echo "No handlers/migrations changes — skipping; this job always runs to satisfy the required-check name."
+
+      - if: needs.detect-changes.outputs.handlers == 'true'
+        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+
+      - if: needs.detect-changes.outputs.handlers == 'true'
+        uses: actions/setup-go@40f1582b2485089dde7abd97c1529aa768e1baff # v5
+        with:
+          go-version: 'stable'
+
+      - if: needs.detect-changes.outputs.handlers == 'true'
+        name: Start sibling Postgres on bridge network
+        working-directory: .
+        run: |
+          # Sanity: the bridge network must exist on the operator host.
+          # Hard-fail loud if it doesn't — easier to spot than a silent
+          # auto-create that diverges from the rest of the stack.
+          if ! docker network inspect "${PG_NETWORK}" >/dev/null 2>&1; then
+            echo "::error::Bridge network '${PG_NETWORK}' missing on operator host. Re-run docker-compose.infra.yml or check ops handbook."
+            exit 1
+          fi
+
+          # If a stale container with the same name exists (rerun on
+          # the same run_id), wipe it first.
+          docker rm -f "${PG_NAME}" >/dev/null 2>&1 || true
+
+          docker run -d \
+            --name "${PG_NAME}" \
+            --network "${PG_NETWORK}" \
+            --health-cmd "pg_isready -U postgres" \
+            --health-interval 5s \
+            --health-timeout 5s \
+            --health-retries 10 \
+            -e POSTGRES_PASSWORD=test \
+            -e POSTGRES_DB=molecule \
+            postgres:15-alpine >/dev/null
+
+          # Read back the bridge IP. Always present immediately after
+          # `docker run -d` for bridge networks.
+          PG_HOST=$(docker inspect "${PG_NAME}" \
+            --format "{{(index .NetworkSettings.Networks \"${PG_NETWORK}\").IPAddress}}")
+          if [ -z "${PG_HOST}" ]; then
+            echo "::error::Could not resolve PG_HOST for ${PG_NAME} on ${PG_NETWORK}"
+            docker logs "${PG_NAME}" || true
+            exit 1
+          fi
+          echo "PG_HOST=${PG_HOST}" >> "$GITHUB_ENV"
+          echo "INTEGRATION_DB_URL=postgres://postgres:test@${PG_HOST}:5432/molecule?sslmode=disable" >> "$GITHUB_ENV"
+          echo "Started ${PG_NAME} at ${PG_HOST}:5432"
+
+      - if: needs.detect-changes.outputs.handlers == 'true'
+        name: Apply migrations to Postgres service
+        env:
+          PGPASSWORD: test
+        run: |
+          # Wait for postgres to actually accept connections. Docker's
+          # health-cmd handles container-side readiness, but the wire
+          # to the bridge IP is best-tested with pg_isready directly.
+          for i in {1..15}; do
+            if pg_isready -h "${PG_HOST}" -p 5432 -U postgres -q; then break; fi
+            echo "waiting for postgres at ${PG_HOST}:5432..."; sleep 2
+          done
+
+          # Apply every .up.sql in lexicographic order with
+          # ON_ERROR_STOP=0 — failing migrations are SKIPPED rather than
+          # blocking the suite. This handles the current schema state
+          # where a few historical migrations (e.g. 017_memories_fts_*)
+          # depend on tables that were later renamed/dropped and so
+          # cannot replay from scratch. The migrations that DO succeed
+          # land their tables, which is sufficient for the integration
+          # tests in handlers/.
+          #
+          # Why not maintain a curated allowlist: every new migration
+          # touching a handlers/-tested table would have to update this
+          # workflow. With apply-all-or-skip, a future migration that
+          # adds a column to delegations runs automatically (its base
+          # table 049_delegations.up.sql already succeeded above it in
+          # the order). Operators only need to revisit this if the
+          # migration chain becomes legitimately replayable end-to-end.
+          #
+          # Per-migration result is logged so a failed migration that
+          # SHOULD have been replayable surfaces in the CI log instead
+          # of silently failing.
+          # Apply both *.sql (legacy, lives next to its module) and
+          # *.up.sql (newer up/down convention) in a single
+          # lexicographically-sorted pass. Excluding *.down.sql so the
+          # newest-naming-convention pairs don't undo themselves mid-run.
+          # Pre-#149-followup this loop only globbed *.up.sql, which
+          # silently skipped 001_workspaces.sql + 009_activity_logs.sql
+          # — fine while no integration test depended on those tables,
+          # not fine once a cross-table atomicity test came in.
+          set +e
+          for migration in $(ls migrations/*.sql 2>/dev/null | grep -v '\.down\.sql$' | sort); do
+            if psql -h "${PG_HOST}" -U postgres -d molecule -v ON_ERROR_STOP=1 \
+                  -f "$migration" >/dev/null 2>&1; then
+              echo "✓ $(basename "$migration")"
+            else
+              echo "⊘ $(basename "$migration") (skipped — see comment in workflow)"
+            fi
+          done
+          set -e
+
+          # Sanity: the delegations + workspaces + activity_logs tables
+          # MUST exist for the integration tests to be meaningful. Hard-
+          # fail if any didn't land — that would be a real regression we
+          # want loud.
+          for tbl in delegations workspaces activity_logs pending_uploads; do
+            if ! psql -h "${PG_HOST}" -U postgres -d molecule -tA \
+                -c "SELECT 1 FROM information_schema.tables WHERE table_name = '$tbl'" \
+                | grep -q 1; then
+              echo "::error::$tbl table missing after migration replay — handler integration tests would be meaningless"
+              exit 1
+            fi
+            echo "✓ $tbl table present"
+          done
+
+      - if: needs.detect-changes.outputs.handlers == 'true'
+        name: Run integration tests
+        run: |
+          # INTEGRATION_DB_URL is exported by the start-postgres step;
+          # points at the per-run bridge IP, not 127.0.0.1, so concurrent
+          # workflow runs don't fight over a host-net 5432 port.
+          go test -tags=integration -timeout 5m -v ./internal/handlers/ -run "^TestIntegration_"
+
+      - if: failure() && needs.detect-changes.outputs.handlers == 'true'
+        name: Diagnostic dump on failure
+        env:
+          PGPASSWORD: test
+        run: |
+          echo "::group::postgres container status"
+          docker ps -a --filter "name=${PG_NAME}" --format '{{.Status}} {{.Names}}' || true
+          docker logs "${PG_NAME}" 2>&1 | tail -50 || true
+          echo "::endgroup::"
+          echo "::group::delegations table state"
+          psql -h "${PG_HOST}" -U postgres -d molecule -c "SELECT * FROM delegations LIMIT 50;" || true
+          echo "::endgroup::"
+
+      - if: always() && needs.detect-changes.outputs.handlers == 'true'
+        name: Stop sibling Postgres
+        working-directory: .
+        run: |
+          # always() so containers don't leak when migrations or tests
+          # fail. The cleanup is best-effort: if the container is
+          # already gone (e.g. concurrent rerun race), don't fail the job.
+          docker rm -f "${PG_NAME}" >/dev/null 2>&1 || true
+          echo "Cleaned up ${PG_NAME}"
+
@@ -95,16 +95,57 @@ jobs:
      - if: needs.detect-changes.outputs.run == 'true'
        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2

-      - name: Checkout sibling plugin repo
-        # Dockerfile.tenant copies molecule-ai-plugin-github-app-auth/
-        # at the build-context root (see workspace-server/Dockerfile.tenant
-        # line 19). PLUGIN_REPO_PAT pattern matches publish-workspace-server-image.yml.
+      # github-app-auth sibling-checkout removed 2026-05-07 (#157):
+      # the plugin was dropped + Dockerfile.tenant no longer COPYs it.
+
+      # Pre-clone manifest deps before docker compose builds the tenant
+      # image (Task #173 followup — same pattern as
+      # publish-workspace-server-image.yml's "Pre-clone manifest deps"
+      # step).
+      #
+      # Why pre-clone here too: tests/harness/compose.yml builds tenant-alpha
+      # and tenant-beta from workspace-server/Dockerfile.tenant with
+      # context=../.. (repo root). That Dockerfile expects
+      # .tenant-bundle-deps/{workspace-configs-templates,org-templates,plugins}
+      # to be present at build context root (post-#173 it COPYs from there
+      # instead of running an in-image clone — the in-image clone failed
+      # with "could not read Username for https://git.moleculesai.app"
+      # because there's no auth path inside the build sandbox).
+      #
+      # Without this step harness-replays fails before any replay runs,
+      # with `failed to calculate checksum of ref ...
+      # "/.tenant-bundle-deps/plugins": not found`. Caught by run #892
+      # (main, 2026-05-07T20:28:53Z) and run #964 (staging — same
+      # symptom, different root cause: staging still has the in-image
+      # clone path, hits the auth error directly).
+      #
+      # Token shape matches publish-workspace-server-image.yml: AUTO_SYNC_TOKEN
+      # is the devops-engineer persona PAT, NOT the founder PAT (per
+      # `feedback_per_agent_gitea_identity_default`). clone-manifest.sh
+      # embeds it as basic-auth for the duration of the clones and strips
+      # .git directories — the token never enters the resulting image.
+      - name: Pre-clone manifest deps
        if: needs.detect-changes.outputs.run == 'true'
-        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
-        with:
-          repository: Molecule-AI/molecule-ai-plugin-github-app-auth
-          path: molecule-ai-plugin-github-app-auth
-          token: ${{ secrets.PLUGIN_REPO_PAT || secrets.GITHUB_TOKEN }}
+        env:
+          MOLECULE_GITEA_TOKEN: ${{ secrets.AUTO_SYNC_TOKEN }}
+        run: |
+          set -euo pipefail
+          if [ -z "${MOLECULE_GITEA_TOKEN}" ]; then
+            echo "::error::AUTO_SYNC_TOKEN secret is empty — register the devops-engineer persona PAT in repo Actions secrets"
+            exit 1
+          fi
+          mkdir -p .tenant-bundle-deps
+          bash scripts/clone-manifest.sh \
+            manifest.json \
+            .tenant-bundle-deps/workspace-configs-templates \
+            .tenant-bundle-deps/org-templates \
+            .tenant-bundle-deps/plugins
+          # Sanity-check counts so a silent partial clone fails fast
+          # instead of producing a half-empty image.
+          ws_count=$(find .tenant-bundle-deps/workspace-configs-templates -mindepth 1 -maxdepth 1 -type d | wc -l)
+          org_count=$(find .tenant-bundle-deps/org-templates -mindepth 1 -maxdepth 1 -type d | wc -l)
+          plugins_count=$(find .tenant-bundle-deps/plugins -mindepth 1 -maxdepth 1 -type d | wc -l)
+          echo "Cloned: ws=$ws_count org=$org_count plugins=$plugins_count"

      - name: Install Python deps for replays
        # peer-discovery-404 (and future replays) eval Python against the
@@ -0,0 +1,94 @@
+name: Lint curl status-code capture
+
+# Pins the workflow-bash anti-pattern that produced "HTTP 000000" on the
+# 2026-05-04 redeploy-tenants-on-main run for sha 2b862f6:
+#
+#   HTTP_CODE=$(curl ... -w '%{http_code}' ... || echo "000")
+#
+# When curl exits non-zero (connection reset → 56, --fail-with-body 4xx/5xx
+# → 22), the `-w '%{http_code}'` already wrote a status to stdout — usually
+# "000" for connection failures or the actual code for HTTP errors. The
+# `|| echo "000"` then fires AND appends ANOTHER "000" to the captured
+# stdout, producing values like "000000" or "409000" that fail string
+# comparisons against "200" while looking superficially right.
+#
+# Same class of bug the synth-E2E §7c gate hit twice (PRs #2779/#2783 +
+# #2797). Memory: feedback_curl_status_capture_pollution.md.
+#
+# Fix shape (route -w into a tempfile so curl's exit code can't pollute):
+#
+#   set +e
+#   curl ... -w '%{http_code}' >code.txt 2>/dev/null
+#   set -e
+#   HTTP_CODE=$(cat code.txt 2>/dev/null)
+#   [ -z "$HTTP_CODE" ] && HTTP_CODE="000"
+
+on:
+  pull_request:
+    paths: ['.github/workflows/**']
+  push:
+    branches: [main, staging]
+    paths: ['.github/workflows/**']
+  merge_group:
+    types: [checks_requested]
+
+jobs:
+  scan:
+    name: Scan workflows for curl status-capture pollution
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+      - name: Find curl ... -w '%{http_code}' ... || echo "000" subshells
+        run: |
+          set -uo pipefail
+          # Multi-line aware: look for `$(curl ... -w '%{http_code}' ... || echo "000")`
+          # subshell where the entire command-substitution wraps a curl that
+          # ends with `|| echo "000"`. Must distinguish from the SAFE shape
+          # `$(cat tempfile 2>/dev/null || echo "000")` — `cat` with a missing
+          # tempfile produces empty stdout, no pollution.
+          python3 <<'PY'
+          import os, re, sys, glob
+
+          BAD_FILES = []
+
+          # Match the buggy substitution across newlines: $(curl ... -w '%{http_code}' ... || echo "000")
+          # The `\\n` is the bash line-continuation that lets curl flags span lines.
+          # We collapse continuation lines first, then look for the single-line bad pattern.
+          PATTERN = re.compile(
+              r'\$\(\s*curl\b[^)]*-w\s*[\'"]%\{http_code\}[\'"][^)]*\|\|\s*echo\s+"000"\s*\)',
+              re.DOTALL,
+          )
+
+          # Self-skip: this lint workflow contains the literal anti-pattern in
+          # its own docstring — that's intentional, not a bug.
+          SELF = ".github/workflows/lint-curl-status-capture.yml"
+
+          for f in sorted(glob.glob(".github/workflows/*.yml")):
+              if f == SELF:
+                  continue
+              with open(f) as fh:
+                  content = fh.read()
+              # Collapse bash line-continuations (\\\n + leading whitespace)
+              # into a single logical line so the regex can see the full
+              # curl invocation as one chunk.
+              flat = re.sub(r'\\\s*\n\s*', ' ', content)
+              for m in PATTERN.finditer(flat):
+                  BAD_FILES.append((f, m.group(0)[:120]))
+
+          if not BAD_FILES:
+              print("✓ No curl-status-capture pollution patterns detected")
+              sys.exit(0)
+
+          print(f"::error::Found {len(BAD_FILES)} curl-status-capture pollution site(s):")
+          for f, snippet in BAD_FILES:
+              print(f"::error file={f}::Curl status-capture pollution: '|| echo \"000\"' inside a $(curl ... -w '%{{http_code}}' ...) subshell. On non-2xx or connection failure, curl's -w writes a status, then exits non-zero, then the || echo appends another '000' — producing 'HTTP 000000' or '409000' that fails comparisons silently. Fix: route -w into a tempfile so the exit code can't pollute stdout. See memory feedback_curl_status_capture_pollution.md.")
+              print(f"   matched: {snippet}…")
+          print()
+          print("Fix template:")
+          print('  set +e')
+          print('  curl ... -w \'%{http_code}\' >code.txt 2>/dev/null')
+          print('  set -e')
+          print('  HTTP_CODE=$(cat code.txt 2>/dev/null)')
+          print('  [ -z "$HTTP_CODE" ] && HTTP_CODE="000"')
+          sys.exit(1)
+          PY
@@ -1,14 +1,25 @@
 name: pr-guards

-# Thin caller that delegates to the molecule-ci reusable guard. Today
-# the guard is just "disable auto-merge when a new commit is pushed
-# after auto-merge was enabled" — added 2026-04-27 after PR #2174
-# auto-merged with only its first commit because the second commit
-# was pushed after the merge queue had locked the PR's SHA.
+# PR-time guards. Today the only guard is "disable auto-merge when a
+# new commit is pushed after auto-merge was enabled" — added 2026-04-27
+# after PR #2174 auto-merged with only its first commit because the
+# second commit was pushed after the merge queue had locked the PR's
+# SHA.
 #
-# When more PR-time guards land in molecule-ci, add them here as
-# additional jobs that share the same pull_request:synchronize
-# trigger.
+# Why this is inlined (not delegated to molecule-ci's reusable
+# workflow): the reusable workflow uses `gh pr merge --disable-auto`,
+# which calls GitHub's GraphQL API. Gitea has no GraphQL endpoint and
+# returns HTTP 405 on /api/graphql, so the job failed on every Gitea
+# PR push since the 2026-05-06 migration. Gitea also has no `--auto`
+# merge primitive that this job could be acting on, so the right
+# behaviour on Gitea is "no-op + green status" — not a 405.
+#
+# Inlining (vs. an `if:` on the `uses:` line) keeps the job ALWAYS
+# running, which matters for branch protection: required-check names
+# need a job that emits SUCCESS terminal state, not SKIPPED. See
+# `feedback_branch_protection_check_name_parity` and `feedback_pr_merge_safety_guards`.
+#
+# Issue #88 item 1.

 on:
  pull_request:
@@ -19,4 +30,34 @@ permissions:

 jobs:
  disable-auto-merge-on-push:
-    uses: Molecule-AI/molecule-ci/.github/workflows/disable-auto-merge-on-push.yml@main
+    runs-on: ubuntu-latest
+    steps:
+      # Detect Gitea Actions. act_runner sets GITEA_ACTIONS=true in the
+      # step env on every job. Belt-and-suspenders: also check the repo
+      # url's host, which is independent of any runner-side env config
+      # (covers a future Gitea host where the env var is forgotten).
+      - name: Detect runner host
+        id: host
+        run: |
+          if [[ "${GITEA_ACTIONS:-}" == "true" ]] || [[ "${{ github.server_url }}" == *moleculesai.app* ]] || [[ "${{ github.event.repository.html_url }}" == *moleculesai.app* ]]; then
+            echo "is_gitea=true" >> "$GITHUB_OUTPUT"
+            echo "::notice::Gitea Actions detected — auto-merge gating is not applicable here (Gitea has no --auto merge primitive). Job will no-op."
+          else
+            echo "is_gitea=false" >> "$GITHUB_OUTPUT"
+          fi
+
+      - name: Disable auto-merge (GitHub only)
+        if: steps.host.outputs.is_gitea != 'true'
+        env:
+          GH_TOKEN: ${{ github.token }}
+          PR: ${{ github.event.pull_request.number }}
+          REPO: ${{ github.repository }}
+          NEW_SHA: ${{ github.sha }}
+        run: |
+          set -eu
+          gh pr merge "$PR" --disable-auto -R "$REPO" || true
+          gh pr comment "$PR" -R "$REPO" --body "🔒 Auto-merge disabled — new commit (\`${NEW_SHA:0:7}\`) pushed after auto-merge was enabled. The merge queue locks SHAs at entry, so subsequent pushes can race. Verify the new commit and re-enable with \`gh pr merge --auto\`."
+
+      - name: Gitea no-op
+        if: steps.host.outputs.is_gitea == 'true'
+        run: echo "Gitea Actions — auto-merge gating not applicable; no-op (job intentionally green so branch protection's required-check name lands SUCCESS)."
@@ -25,7 +25,7 @@ name: publish-runtime
 #   3. Publishes to PyPI via the PyPA Trusted Publisher action (OIDC).
 #      No static API token is stored — PyPI verifies the workflow's
 #      OIDC claim against the trusted-publisher config registered for
-#      molecule-ai-workspace-runtime (Molecule-AI/molecule-core,
+#      molecule-ai-workspace-runtime (molecule-ai/molecule-core,
 #      publish-runtime.yml, environment pypi-publish).
 #
 # After publish: the 8 template repos pick up the new version on their
@@ -166,7 +166,7 @@ jobs:

      - name: Publish to PyPI (Trusted Publisher / OIDC)
        # PyPI side is configured: project molecule-ai-workspace-runtime →
-        # publisher Molecule-AI/molecule-core, workflow publish-runtime.yml,
+        # publisher molecule-ai/molecule-core, workflow publish-runtime.yml,
        # environment pypi-publish. The action mints a short-lived OIDC
        # token and exchanges it for a PyPI upload credential — no static
        # API token in this repo's secrets.
@@ -282,42 +282,33 @@ jobs:
          echo "::error::Refusing to fan out cascade against stale or corrupt PyPI surfaces."
          exit 1

-      - name: Fan out repository_dispatch
+      - name: Fan out via push to .runtime-version
        env:
-          # Fine-grained PAT with `actions:write` on the 8 template repos.
-          # GITHUB_TOKEN can't fire dispatches across repos — needs an explicit
-          # token. Stored as a repo secret; rotate per the standard schedule.
-          DISPATCH_TOKEN: ${{ secrets.TEMPLATE_DISPATCH_TOKEN }}
-          # Single source of truth: the publish job's output, which handles
-          # tag/manual-input/auto-bump uniformly. The previous fallback
-          # (`steps.version.outputs.version` from inside the cascade job)
-          # was a dead reference — different job, no shared step scope.
+          # Gitea PAT with write:repository scope on the 8 cascade-active
+          # template repos. Used here for `git push` (NOT for an API
+          # dispatch — Gitea 1.22.6 has no repository_dispatch endpoint;
+          # empirically verified across 6 candidate paths in molecule-
+          # core#20 issuecomment-913). The push trips each template's
+          # existing `on: push: branches: [main]` trigger on
+          # publish-image.yml, which then reads the updated
+          # .runtime-version via its resolve-version job.
+          DISPATCH_TOKEN: ${{ secrets.DISPATCH_TOKEN }}
          RUNTIME_VERSION: ${{ needs.publish.outputs.version }}
        run: |
          set +e   # don't abort on a single repo failure — collect them all
-          # Schedule-vs-dispatch behaviour split (hardened 2026-04-28
-          # after the sweep-cf-orphans soft-skip incident — same class
-          # of bug):
-          #
-          # The earlier "skipping cascade. templates will pick up the
-          # new version on their own next rebuild" message was wrong —
-          # templates only build on this dispatch trigger; without it
-          # they stay pinned to whatever runtime version they last saw.
-          # A silent skip here means "PyPI is current, templates are
-          # not" and the gap is invisible until someone notices a
-          # template still on the old version weeks later.
-          #
-          #   - push                → exit 1 (red CI surfaces the gap)
-          #   - workflow_dispatch   → exit 0 with a warning (operator
-          #                           ran this ad-hoc; let them rerun
-          #                           after fixing the secret)
+
+          # Soft-skip on workflow_dispatch when the token is missing
+          # (operator ad-hoc test); hard-fail on push so unattended
+          # publishes can't silently skip the cascade. Same shape as
+          # the original v1, intentional split per the schedule-vs-
+          # dispatch hardening 2026-04-28.
          if [ -z "$DISPATCH_TOKEN" ]; then
            if [ "${{ github.event_name }}" = "workflow_dispatch" ]; then
-              echo "::warning::TEMPLATE_DISPATCH_TOKEN secret not set — skipping cascade."
+              echo "::warning::DISPATCH_TOKEN secret not set — skipping cascade."
              echo "::warning::set it at Settings → Secrets and Variables → Actions, then rerun. Templates will stay on the prior runtime version until either this token is set or each template is rebuilt manually."
              exit 0
            fi
-            echo "::error::TEMPLATE_DISPATCH_TOKEN secret missing — cascade cannot fan out."
+            echo "::error::DISPATCH_TOKEN secret missing — cascade cannot fan out."
            echo "::error::PyPI was published, but the 8 template repos will NOT pick up the new version until this token is restored and a republish dispatches the cascade."
            echo "::error::set it at Settings → Secrets and Variables → Actions; then re-trigger publish-runtime via workflow_dispatch."
            exit 1
@@ -327,37 +318,119 @@ jobs:
            echo "::error::publish job did not expose a version output — cascade cannot fan out"
            exit 1
          fi
-          # All 9 active workspace template repos. The PR #2536 pruning
-          # ("deprecated, no shipping images") was empirically wrong:
-          # continuous-synth-e2e.yml defaults to langgraph as its primary
-          # canary (line 44), and every excluded template had successful
-          # publish-image runs as of 2026-05-03 — none were dormant.
-          # Symptom of the prune: today's a2a-sdk strict-mode fix
-          # (#2566 / commit e1628c4) cascaded to 4 templates but never
-          # reached langgraph, so the synth-E2E correctly canary'd a fix
-          # that had landed but not deployed. Re-added the 5 templates.
-          # Long-term: derive this list from manifest.json so cascade
-          # scope can't drift from E2E scope — tracked in RFC #388 as a
-          # Phase-1 invariant.
+
+          # All 9 workspace templates declared in manifest.json. The list
+          # MUST stay aligned with manifest.json's workspace_templates —
+          # cascade-list-drift-gate.yml enforces this in CI per the
+          # codex-stuck-on-stale-runtime invariant from PR #2556.
+          # Long-term goal: derive this list from manifest.json so it
+          # can't drift even on a manifest edit (RFC #388 Phase-1).
+          #
+          # Per-template publish-image.yml presence is checked at
+          # cascade-time below: codex doesn't ship one today, so the
+          # cascade soft-skips it with an informational message rather
+          # than dropping it from this list (which would re-introduce
+          # the drift the gate exists to catch).
+          GITEA_URL="${GITEA_URL:-https://git.moleculesai.app}"
          TEMPLATES="claude-code hermes openclaw codex langgraph crewai autogen deepagents gemini-cli"
          FAILED=""
+          SKIPPED=""
+
+          # Configure git identity once. The persona owning DISPATCH_TOKEN
+          # is the same identity that authored this commit on each
+          # template; using a generic "publish-runtime cascade" co-author
+          # trailer in the message keeps the audit trail honest about the
+          # workflow-driven origin.
+          git config --global user.name  "publish-runtime cascade"
+          git config --global user.email "publish-runtime@moleculesai.app"
+
+          WORKDIR="$(mktemp -d)"
          for tpl in $TEMPLATES; do
-            REPO="Molecule-AI/molecule-ai-workspace-template-$tpl"
-            STATUS=$(curl -sS -o /tmp/dispatch.out -w "%{http_code}" \
-              -X POST "https://api.github.com/repos/$REPO/dispatches" \
-              -H "Authorization: Bearer $DISPATCH_TOKEN" \
-              -H "Accept: application/vnd.github+json" \
-              -H "X-GitHub-Api-Version: 2022-11-28" \
-              -d "{\"event_type\":\"runtime-published\",\"client_payload\":{\"runtime_version\":\"$VERSION\"}}")
-            if [ "$STATUS" = "204" ]; then
-              echo "✓ dispatched $tpl ($VERSION)"
-            else
-              echo "::warning::✗ failed to dispatch $tpl: HTTP $STATUS — $(cat /tmp/dispatch.out)"
+            REPO="molecule-ai/molecule-ai-workspace-template-$tpl"
+            CLONE="$WORKDIR/$tpl"
+
+            # Pre-check: skip templates without a publish-image.yml.
+            # The cascade's job is to trip the template's on-push
+            # rebuild — if there's no rebuild workflow, pushing a
+            # .runtime-version commit is just noise on the target
+            # repo. Use the Gitea contents API (no clone required for
+            # the probe). 200 = present; 404 = absent.
+            HTTP=$(curl -sS -o /dev/null -w "%{http_code}" \
+              -H "Authorization: token $DISPATCH_TOKEN" \
+              "$GITEA_URL/api/v1/repos/$REPO/contents/.github/workflows/publish-image.yml")
+            if [ "$HTTP" = "404" ]; then
+              echo "↷ $tpl has no publish-image.yml — soft-skip (informational; manifest still tracks it)"
+              SKIPPED="$SKIPPED $tpl"
+              continue
+            fi
+            if [ "$HTTP" != "200" ]; then
+              echo "::warning::$tpl publish-image.yml probe returned HTTP $HTTP — proceeding anyway, push will surface the real failure if any"
+            fi
+
+            # Use a per-template attempt loop so a transient race (e.g.
+            # human pushing to the same template at the same instant)
+            # doesn't lose the cascade. Bounded retries (3) — beyond
+            # that we surface the failure and let the operator retry.
+            attempt=0
+            success=false
+            while [ $attempt -lt 3 ]; do
+              attempt=$((attempt + 1))
+              rm -rf "$CLONE"
+              if ! git clone --depth=1 \
+                  "https://x-access-token:${DISPATCH_TOKEN}@${GITEA_URL#https://}/$REPO.git" \
+                  "$CLONE" >/tmp/clone.log 2>&1; then
+                echo "::warning::clone $tpl attempt $attempt failed: $(tail -n3 /tmp/clone.log)"
+                sleep 2
+                continue
+              fi
+
+              cd "$CLONE"
+              echo "$VERSION" > .runtime-version
+
+              # Idempotency guard: if the file already matches, this
+              # publish is a re-run for a version already cascaded.
+              # Don't push a no-op commit (would spuriously re-trip the
+              # template's on-push and rebuild for nothing).
+              if git diff --quiet -- .runtime-version; then
+                echo "✓ $tpl already at $VERSION — no commit needed (idempotent)"
+                success=true
+                cd - >/dev/null
+                break
+              fi
+
+              git add .runtime-version
+              git commit -m "chore: pin runtime to $VERSION (publish-runtime cascade)" \
+                -m "Co-Authored-By: publish-runtime cascade <publish-runtime@moleculesai.app>" \
+                >/dev/null
+
+              if git push origin HEAD:main >/tmp/push.log 2>&1; then
+                echo "✓ $tpl pushed $VERSION on attempt $attempt"
+                success=true
+                cd - >/dev/null
+                break
+              fi
+
+              # Likely a non-fast-forward — pull-rebase and retry.
+              # Don't force-push: that would silently overwrite a racing
+              # human/cascade commit.
+              echo "::warning::push $tpl attempt $attempt failed, pull-rebasing: $(tail -n3 /tmp/push.log)"
+              git pull --rebase origin main >/tmp/rebase.log 2>&1 || true
+              cd - >/dev/null
+            done
+
+            if [ "$success" != "true" ]; then
              FAILED="$FAILED $tpl"
            fi
          done
+          rm -rf "$WORKDIR"
+
          if [ -n "$FAILED" ]; then
-            echo "::warning::Cascade incomplete. Failed templates:$FAILED"
-            # Don't fail the whole job — PyPI publish already succeeded;
-            # operators can retry the failed templates manually.
+            echo "::error::Cascade incomplete after 3 retries each. Failed templates:$FAILED"
+            echo "::error::PyPI publish succeeded; failed templates lag the new version. Re-run this workflow_dispatch with the same version to retry only the laggers (idempotent — already-cascaded templates skip)."
+            exit 1
+          fi
+          if [ -n "$SKIPPED" ]; then
+            echo "Cascade complete: pinned $VERSION on cascade-active templates. Soft-skipped (no publish-image.yml):$SKIPPED"
+          else
+            echo "Cascade complete: $VERSION pinned across all manifest workspace_templates."
          fi
@@ -37,6 +37,7 @@ on:
      - 'workspace-server/**'
      - 'canvas/**'
      - 'manifest.json'
+      - 'scripts/**'
      - '.github/workflows/publish-workspace-server-image.yml'
  workflow_dispatch:

@@ -60,8 +61,8 @@ permissions:
  packages: write

 env:
-  IMAGE_NAME: ghcr.io/molecule-ai/platform
-  TENANT_IMAGE_NAME: ghcr.io/molecule-ai/platform-tenant
+  IMAGE_NAME: 153263036946.dkr.ecr.us-east-2.amazonaws.com/molecule-ai/platform
+  TENANT_IMAGE_NAME: 153263036946.dkr.ecr.us-east-2.amazonaws.com/molecule-ai/platform-tenant

 jobs:
  build-and-push:
@@ -70,40 +71,91 @@ jobs:
      - name: Checkout
        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2

-      - name: Checkout sibling plugin repo
-        # workspace-server/Dockerfile expects
-        # ./molecule-ai-plugin-github-app-auth at build-context root because
-        # the Go module has a `replace` directive pointing at /plugin inside
-        # the image. Pre-repo-split the plugin lived in the monorepo; the
-        # 2026-04-18 restructure moved it out but didn't add this clone step
-        # — which is why publish was failing after that restructure.
-        #
-        # Uses a fine-grained PAT (PLUGIN_REPO_PAT) because the plugin repo
-        # is private and the default GITHUB_TOKEN is scoped to THIS repo.
-        # The PAT needs Contents:Read on Molecule-AI/molecule-ai-plugin-
-        # github-app-auth. Falls back to the default token for the (rare)
-        # case where an operator made the plugin repo public.
-        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
-        with:
-          repository: Molecule-AI/molecule-ai-plugin-github-app-auth
-          path: molecule-ai-plugin-github-app-auth
-          token: ${{ secrets.PLUGIN_REPO_PAT || secrets.GITHUB_TOKEN }}
+      # github-app-auth sibling-checkout removed 2026-05-07 (#157):
+      # plugin was dropped + workspace-server/Dockerfile no longer
+      # COPYs it.

-      - name: Log in to GHCR
-        uses: docker/login-action@c94ce9fb468520275223c153574b00df6fe4bcc9 # v3
-        with:
-          registry: ghcr.io
-          username: ${{ github.actor }}
-          password: ${{ secrets.GITHUB_TOKEN }}
-
-      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@4d04d5d9486b7bd6fa91e7baf45bbb4f8b9deedd # v4.0.0
+      # ECR auth + buildx setup are now inline in each build step
+      # below (Task #173, 2026-05-07).
+      #
+      # Why moved inline: aws-actions/configure-aws-credentials@v4 +
+      # aws-actions/amazon-ecr-login@v2 + docker/setup-buildx-action
+      # all left auth state in places that the actual `docker push`
+      # couldn't see on Gitea Actions:
+      #   - The actions wrote to a step-scoped DOCKER_CONFIG path
+      #     that didn't survive into subsequent shell steps.
+      #   - Buildx couldn't bridge the runner container ↔
+      #     operator-host docker daemon auth gap (401 on the
+      #     docker-container driver, "no basic auth credentials"
+      #     with the action-driven login).
+      #
+      # Doing AWS+ECR auth inline (`aws ecr get-login-password |
+      # docker login`) in the same shell step as `docker build` +
+      # `docker push` is the operator-host manual approach, mapped
+      # 1:1 into CI. Auth state is guaranteed to live in the env that
+      # `docker push` actually runs from.
+      #
+      # Post-suspension target is the operator's ECR org
+      # (153263036946.dkr.ecr.us-east-2.amazonaws.com/molecule-ai/*),
+      # which already hosts platform-tenant + workspace-template-* +
+      # runner-base images. AWS creds come from the
+      # AWS_ACCESS_KEY_ID/SECRET secrets bound to the molecule-cp
+      # IAM user. Closes #161.

      - name: Compute tags
        id: tags
        run: |
          echo "sha=${GITHUB_SHA::7}" >> "$GITHUB_OUTPUT"

+      # Pre-clone manifest deps before docker build (Task #173 fix).
+      #
+      # Why pre-clone: post-2026-05-06, every workspace-template-* repo on
+      # Gitea (codex, crewai, deepagents, gemini-cli, langgraph) plus all
+      # 7 org-template-* repos are private. The pre-fix Dockerfile.tenant
+      # ran `git clone` inside an in-image stage, which had no auth path
+      # — every CI build failed with "fatal: could not read Username for
+      # https://git.moleculesai.app". For weeks, every workspace-server
+      # rebuild required a manual operator-host push. Now we clone in the
+      # trusted CI context (where AUTO_SYNC_TOKEN is naturally available)
+      # and Dockerfile.tenant just COPYs from .tenant-bundle-deps/.
+      #
+      # Token shape: AUTO_SYNC_TOKEN is the devops-engineer persona PAT
+      # (see /etc/molecule-bootstrap/agent-secrets.env). Per saved memory
+      # `feedback_per_agent_gitea_identity_default`, every CI surface uses
+      # a per-persona token, never the founder PAT. clone-manifest.sh
+      # embeds it as basic-auth (oauth2:<token>) for the duration of the
+      # clones, then strips .git directories — the token never enters
+      # the resulting image.
+      #
+      # Idempotent: if a re-run finds populated dirs, clone-manifest.sh
+      # skips them; safe to retrigger via path-filter or workflow_dispatch.
+      - name: Pre-clone manifest deps
+        env:
+          MOLECULE_GITEA_TOKEN: ${{ secrets.AUTO_SYNC_TOKEN }}
+        run: |
+          set -euo pipefail
+          if [ -z "${MOLECULE_GITEA_TOKEN}" ]; then
+            echo "::error::AUTO_SYNC_TOKEN secret is empty — register the devops-engineer persona PAT in repo Actions secrets"
+            exit 1
+          fi
+          mkdir -p .tenant-bundle-deps
+          bash scripts/clone-manifest.sh \
+            manifest.json \
+            .tenant-bundle-deps/workspace-configs-templates \
+            .tenant-bundle-deps/org-templates \
+            .tenant-bundle-deps/plugins
+          # Sanity-check counts so a silent partial clone fails fast
+          # instead of producing a half-empty image.
+          ws_count=$(find .tenant-bundle-deps/workspace-configs-templates -mindepth 1 -maxdepth 1 -type d | wc -l)
+          org_count=$(find .tenant-bundle-deps/org-templates -mindepth 1 -maxdepth 1 -type d | wc -l)
+          plugins_count=$(find .tenant-bundle-deps/plugins -mindepth 1 -maxdepth 1 -type d | wc -l)
+          echo "Cloned: ws=$ws_count org=$org_count plugins=$plugins_count"
+          # Counts are derived from manifest.json (9 ws / 7 org / 21
+          # plugins as of 2026-05-07). If manifest.json grows but the
+          # clone step regresses silently, the find above caps at the
+          # actual disk state — but clone-manifest.sh's own EXPECTED vs
+          # CLONED check (line ~95) is the authoritative fail-fast.
+
      # Canary-gated release flow:
      #   - This step always publishes :staging-<sha> + :staging-latest.
      #   - On staging push, staging-CP picks up :staging-latest immediately
@@ -129,58 +181,82 @@ jobs:
      # were running pre-RFC code. Adding the staging trigger above closes
      # that gap. Earlier 2026-04-24 incident: a static :staging-<sha> pin
      # drifted 10 days behind staging — same class of bug, different
-      # mechanism.
-      - name: Build & push platform image to GHCR (staging-<sha> + staging-latest)
-        uses: docker/build-push-action@bcafcacb16a39f128d818304e6c9c0c18556b85f # v7.1.0
-        with:
-          context: .
-          file: ./workspace-server/Dockerfile
-          platforms: linux/amd64
-          push: true
-          tags: |
-            ${{ env.IMAGE_NAME }}:staging-${{ steps.tags.outputs.sha }}
-            ${{ env.IMAGE_NAME }}:staging-latest
-          cache-from: type=gha
-          cache-to: type=gha,mode=max
-          # GIT_SHA bakes into the Go binary via -ldflags so /buildinfo
-          # returns it at runtime — see Dockerfile + buildinfo/buildinfo.go.
-          # This is the same value as the OCI revision label below; passing
-          # it twice is intentional, the OCI label is for registry tooling
-          # while /buildinfo is for the redeploy verification step.
-          build-args: |
-            GIT_SHA=${{ github.sha }}
-          labels: |
-            org.opencontainers.image.source=https://github.com/${{ github.repository }}
-            org.opencontainers.image.revision=${{ github.sha }}
-            org.opencontainers.image.description=Molecule AI platform (Go API server) — pending canary verify
+      # mechanism. ECR repo molecule-ai/platform created 2026-05-07.
+      # Build + push platform image with plain `docker` (no buildx).
+      # GIT_SHA bakes into the Go binary via -ldflags so /buildinfo
+      # returns it at runtime — see Dockerfile + buildinfo/buildinfo.go.
+      # The OCI revision label below carries the same value for registry
+      # tooling; the duplication is intentional.
+      - name: Build & push platform image to ECR (staging-<sha> + staging-latest)
+        env:
+          IMAGE_NAME: ${{ env.IMAGE_NAME }}
+          TAG_SHA: staging-${{ steps.tags.outputs.sha }}
+          TAG_LATEST: staging-latest
+          GIT_SHA: ${{ github.sha }}
+          REPO: ${{ github.repository }}
+          AWS_ACCESS_KEY_ID: ${{ secrets.AWS_ACCESS_KEY_ID }}
+          AWS_SECRET_ACCESS_KEY: ${{ secrets.AWS_SECRET_ACCESS_KEY }}
+          AWS_DEFAULT_REGION: us-east-2
+        run: |
+          set -euo pipefail
+          # ECR auth in-step so config.json is populated in the same
+          # shell env that runs `docker push`. ECR get-login-password
+          # tokens last 12h, plenty for a single-step build+push.
+          ECR_REGISTRY="${IMAGE_NAME%%/*}"
+          aws ecr get-login-password --region us-east-2 | \
+            docker login --username AWS --password-stdin "${ECR_REGISTRY}"
+          docker build \
+            --file ./workspace-server/Dockerfile \
+            --build-arg GIT_SHA="${GIT_SHA}" \
+            --label "org.opencontainers.image.source=https://github.com/${REPO}" \
+            --label "org.opencontainers.image.revision=${GIT_SHA}" \
+            --label "org.opencontainers.image.description=Molecule AI platform (Go API server) — pending canary verify" \
+            --tag "${IMAGE_NAME}:${TAG_SHA}" \
+            --tag "${IMAGE_NAME}:${TAG_LATEST}" \
+            .
+          docker push "${IMAGE_NAME}:${TAG_SHA}"
+          docker push "${IMAGE_NAME}:${TAG_LATEST}"
+
+      # Canvas uses same-origin fetches. The tenant Go platform
+      # reverse-proxies /cp/* to the SaaS CP via its CP_UPSTREAM_URL
+      # env; the tenant's /canvas/viewport, /approvals/pending,
+      # /org/templates etc. live on the tenant platform itself.
+      # Both legs share one origin (the tenant subdomain) so
+      # PLATFORM_URL="" forces canvas to fetch paths as relative,
+      # which land same-origin.
+      #
+      # Self-hosted / private-label deployments override this at
+      # build time with a specific backend (e.g. local dev:
+      # NEXT_PUBLIC_PLATFORM_URL=http://localhost:8080).
+      - name: Build & push tenant image to ECR (staging-<sha> + staging-latest)
+        env:
+          TENANT_IMAGE_NAME: ${{ env.TENANT_IMAGE_NAME }}
+          TAG_SHA: staging-${{ steps.tags.outputs.sha }}
+          TAG_LATEST: staging-latest
+          GIT_SHA: ${{ github.sha }}
+          REPO: ${{ github.repository }}
+          AWS_ACCESS_KEY_ID: ${{ secrets.AWS_ACCESS_KEY_ID }}
+          AWS_SECRET_ACCESS_KEY: ${{ secrets.AWS_SECRET_ACCESS_KEY }}
+          AWS_DEFAULT_REGION: us-east-2
+        run: |
+          set -euo pipefail
+          # Re-login: the platform-image step's docker login wrote to
+          # the same config.json, so this is technically redundant — but
+          # making each push step self-contained keeps the workflow
+          # robust to step reordering / future extraction.
+          ECR_REGISTRY="${TENANT_IMAGE_NAME%%/*}"
+          aws ecr get-login-password --region us-east-2 | \
+            docker login --username AWS --password-stdin "${ECR_REGISTRY}"
+          docker build \
+            --file ./workspace-server/Dockerfile.tenant \
+            --build-arg NEXT_PUBLIC_PLATFORM_URL= \
+            --build-arg GIT_SHA="${GIT_SHA}" \
+            --label "org.opencontainers.image.source=https://github.com/${REPO}" \
+            --label "org.opencontainers.image.revision=${GIT_SHA}" \
+            --label "org.opencontainers.image.description=Molecule AI tenant platform + canvas — pending canary verify" \
+            --tag "${TENANT_IMAGE_NAME}:${TAG_SHA}" \
+            --tag "${TENANT_IMAGE_NAME}:${TAG_LATEST}" \
+            .
+          docker push "${TENANT_IMAGE_NAME}:${TAG_SHA}"
+          docker push "${TENANT_IMAGE_NAME}:${TAG_LATEST}"

-      - name: Build & push tenant image to GHCR (staging-<sha> + staging-latest)
-        uses: docker/build-push-action@bcafcacb16a39f128d818304e6c9c0c18556b85f # v7.1.0
-        with:
-          context: .
-          file: ./workspace-server/Dockerfile.tenant
-          platforms: linux/amd64
-          push: true
-          tags: |
-            ${{ env.TENANT_IMAGE_NAME }}:staging-${{ steps.tags.outputs.sha }}
-            ${{ env.TENANT_IMAGE_NAME }}:staging-latest
-          cache-from: type=gha
-          cache-to: type=gha,mode=max
-          # Canvas uses same-origin fetches. The tenant Go platform
-          # reverse-proxies /cp/* to the SaaS CP via its CP_UPSTREAM_URL
-          # env; the tenant's /canvas/viewport, /approvals/pending,
-          # /org/templates etc. live on the tenant platform itself.
-          # Both legs share one origin (the tenant subdomain) so
-          # PLATFORM_URL="" forces canvas to fetch paths as relative,
-          # which land same-origin.
-          #
-          # Self-hosted / private-label deployments override this at
-          # build time with a specific backend (e.g. local dev:
-          # NEXT_PUBLIC_PLATFORM_URL=http://localhost:8080).
-          build-args: |
-            NEXT_PUBLIC_PLATFORM_URL=
-            GIT_SHA=${{ github.sha }}
-          labels: |
-            org.opencontainers.image.source=https://github.com/${{ github.repository }}
-            org.opencontainers.image.revision=${{ github.sha }}
-            org.opencontainers.image.description=Molecule AI tenant platform + canvas — pending canary verify
@@ -9,7 +9,7 @@ name: redeploy-tenants-on-main
 #
 # This workflow closes the gap by calling the control-plane admin
 # endpoint that performs a canary-first, batched, health-gated rolling
-# redeploy across every live tenant. Implemented in Molecule-AI/
+# redeploy across every live tenant. Implemented in molecule-ai/
 # molecule-controlplane as POST /cp/admin/tenants/redeploy-fleet
 # (feat/tenant-auto-redeploy, landing alongside this workflow).
 #
@@ -146,7 +146,7 @@ jobs:

      - name: Call CP redeploy-fleet
        # CP_ADMIN_API_TOKEN must be set as a repo/org secret on
-        # Molecule-AI/molecule-core, matching the staging/prod CP's
+        # molecule-ai/molecule-core, matching the staging/prod CP's
        # CP_ADMIN_API_TOKEN env. Stored in Railway, mirrored to this
        # repo's secrets for CI.
        env:
@@ -184,12 +184,29 @@ jobs:
          echo "  body: $BODY"

          HTTP_RESPONSE=$(mktemp)
-          HTTP_CODE=$(curl -sS -o "$HTTP_RESPONSE" -w '%{http_code}' \
+          HTTP_CODE_FILE=$(mktemp)
+          # Route -w into its own tempfile so curl's exit code (e.g. 56
+          # on connection-reset, 22 on --fail-with-body 4xx/5xx) can't
+          # pollute the captured stdout. The previous inline-substitution
+          # shape produced "000000" on connection reset (curl wrote
+          # "000" via -w, then the inline echo-fallback appended another
+          # "000") — caught on the 2026-05-04 redeploy of sha 2b862f6.
+          # set +e/-e keeps the non-zero curl exit from tripping the
+          # outer pipeline. See lint-curl-status-capture.yml for the
+          # CI gate that pins this fix shape.
+          set +e
+          curl -sS -o "$HTTP_RESPONSE" -w '%{http_code}' \
            -m 1200 \
            -H "Authorization: Bearer $CP_ADMIN_API_TOKEN" \
            -H "Content-Type: application/json" \
            -X POST "$CP_URL/cp/admin/tenants/redeploy-fleet" \
-            -d "$BODY" || echo "000")
+            -d "$BODY" >"$HTTP_CODE_FILE"
+          set -e
+          # Stderr from curl (e.g. dial errors with -sS) goes to the runner
+          # log so operators can see WHY a connection failed. Stdout is
+          # captured to $HTTP_CODE_FILE because that's where -w writes.
+          HTTP_CODE=$(cat "$HTTP_CODE_FILE" 2>/dev/null || echo "000")
+          [ -z "$HTTP_CODE" ] && HTTP_CODE="000"

          echo "HTTP $HTTP_CODE"
          cat "$HTTP_RESPONSE" | jq . || cat "$HTTP_RESPONSE"
@@ -36,7 +36,7 @@ on:
  workflow_run:
    workflows: ['publish-workspace-server-image']
    types: [completed]
-    branches: [staging]
+    branches: [main]
  workflow_dispatch:
    inputs:
      target_tag:
@@ -97,7 +97,7 @@ jobs:

      - name: Call staging-CP redeploy-fleet
        # CP_STAGING_ADMIN_API_TOKEN must be set as a repo/org secret
-        # on Molecule-AI/molecule-core, matching staging-CP's
+        # on molecule-ai/molecule-core, matching staging-CP's
        # CP_ADMIN_API_TOKEN env var (visible in Railway controlplane
        # / staging environment). Stored separately from the prod
        # CP_ADMIN_API_TOKEN so a leak of one doesn't auth the other.
@@ -146,12 +146,26 @@ jobs:
          echo "  body: $BODY"

          HTTP_RESPONSE=$(mktemp)
-          HTTP_CODE=$(curl -sS -o "$HTTP_RESPONSE" -w '%{http_code}' \
+          HTTP_CODE_FILE=$(mktemp)
+          # Route -w into its own tempfile so curl's exit code (e.g. 56
+          # on connection-reset) can't pollute the captured stdout. The
+          # previous inline-substitution shape produced "000000" on
+          # connection reset — caught on main variant 2026-05-04
+          # redeploying sha 2b862f6. Same fix shape as the synth-E2E
+          # §9c gate (PR #2797). See lint-curl-status-capture.yml for
+          # the CI gate that pins this fix shape.
+          set +e
+          curl -sS -o "$HTTP_RESPONSE" -w '%{http_code}' \
            -m 1200 \
            -H "Authorization: Bearer $CP_STAGING_ADMIN_API_TOKEN" \
            -H "Content-Type: application/json" \
            -X POST "$CP_URL/cp/admin/tenants/redeploy-fleet" \
-            -d "$BODY" || echo "000")
+            -d "$BODY" >"$HTTP_CODE_FILE"
+          set -e
+          # Stderr from curl (-sS shows dial errors etc.) goes to the
+          # runner log so operators can see WHY a connection failed.
+          HTTP_CODE=$(cat "$HTTP_CODE_FILE" 2>/dev/null || echo "000")
+          [ -z "$HTTP_CODE" ] && HTTP_CODE="000"

          echo "HTTP $HTTP_CODE"
          cat "$HTTP_RESPONSE" | jq . || cat "$HTTP_RESPONSE"
@@ -1,105 +0,0 @@
-name: Retarget main PRs to staging
-
-# Mechanical enforcement of SHARED_RULES rule 8 ("Staging-first workflow, no
-# exceptions"). When a bot opens a PR against main, retarget it to staging
-# automatically and leave an explanatory comment. Human CEO-authored PRs (the
-# staging→main promotion PR, etc.) are left alone — they're the authorised
-# exception to the rule.
-#
-# Why an Action instead of only a prompt rule: prompt rules depend on every
-# role's system-prompt.md staying in sync. Today 5 of 8 engineer roles
-# (core-be, core-fe, app-fe, app-qa, devops-engineer) don't have the
-# staging-first section — the bot keeps opening PRs to main. An Action
-# enforces the invariant regardless of prompt drift.
-
-on:
-  pull_request_target:
-    types: [opened, reopened]
-    branches: [main]
-
-permissions:
-  pull-requests: write
-
-jobs:
-  retarget:
-    name: Retarget to staging
-    runs-on: ubuntu-latest
-    # Only fire for bot-authored PRs. Human CEO PRs (staging→main promotion)
-    # are intentional and pass through.
-    #
-    # Head-ref guard: never retarget a PR whose head IS `staging` — those
-    # are the auto-promote staging→main PRs (opened by molecule-ai[bot]
-    # since #2586 switched to an App token, which now passes the bot
-    # filter below). Retargeting head=staging onto base=staging fails
-    # with HTTP 422 "no new commits between base 'staging' and head
-    # 'staging'", which used to surface as a noisy red workflow run on
-    # every auto-promote (caught 2026-05-03 on PR #2588).
-    if: >-
-      github.event.pull_request.head.ref != 'staging'
-      && (
-        github.event.pull_request.user.type == 'Bot'
-        || endsWith(github.event.pull_request.user.login, '[bot]')
-        || github.event.pull_request.user.login == 'app/molecule-ai'
-        || github.event.pull_request.user.login == 'molecule-ai[bot]'
-      )
-    steps:
-      - name: Retarget PR base to staging
-        id: retarget
-        env:
-          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          PR_NUMBER: ${{ github.event.pull_request.number }}
-          PR_AUTHOR: ${{ github.event.pull_request.user.login }}
-        # Issue #1884: when the bot opens a PR against main and there's
-        # already another PR on the same head branch targeting staging,
-        # GitHub's PATCH /pulls returns 422 with
-        # "A pull request already exists for base branch 'staging' …".
-        # The retarget can't proceed — but the right response is to
-        # close the now-redundant main-PR, not to fail the workflow
-        # noisily. Detect that specific 422 and close instead.
-        run: |
-          set +e
-          echo "Retargeting PR #${PR_NUMBER} (author: ${PR_AUTHOR}) from main → staging"
-          PATCH_OUTPUT=$(gh api -X PATCH \
-            "repos/${{ github.repository }}/pulls/${PR_NUMBER}" \
-            -f base=staging \
-            --jq '.base.ref' 2>&1)
-          PATCH_EXIT=$?
-          set -e
-          if [ "$PATCH_EXIT" -eq 0 ]; then
-            echo "::notice::Retargeted PR #${PR_NUMBER} → staging"
-            echo "outcome=retargeted" >> "$GITHUB_OUTPUT"
-            exit 0
-          fi
-          # Specifically match the 422 duplicate-base/head error so
-          # any OTHER PATCH failure (auth, deleted PR, etc.) still
-          # surfaces as a real workflow failure.
-          if echo "$PATCH_OUTPUT" | grep -q "pull request already exists for base branch 'staging'"; then
-            echo "::notice::PR #${PR_NUMBER}: duplicate target-staging PR exists on same head — closing this main-PR as redundant."
-            gh pr close "$PR_NUMBER" \
-              --repo "${{ github.repository }}" \
-              --comment "[retarget-bot] Closing — another PR on the same head branch already targets \`staging\`. This PR is redundant. See issue #1884 for the rationale."
-            echo "outcome=closed-as-duplicate" >> "$GITHUB_OUTPUT"
-            exit 0
-          fi
-          echo "::error::Retarget PATCH failed and was NOT a duplicate-base error:"
-          echo "$PATCH_OUTPUT" >&2
-          exit 1
-
-      - name: Post explainer comment
-        if: steps.retarget.outputs.outcome == 'retargeted'
-        env:
-          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          PR_NUMBER: ${{ github.event.pull_request.number }}
-        run: |
-          gh pr comment "$PR_NUMBER" \
-            --repo "${{ github.repository }}" \
-            --body "$(cat <<'BODY'
-          [retarget-bot] This PR was opened against `main` and has been retargeted to `staging` automatically.
-
-          **Why:** per [SHARED_RULES rule 8](https://github.com/Molecule-AI/molecule-ai-org-template-molecule-dev/blob/main/SHARED_RULES.md), all feature work targets `staging` first; the CEO promotes `staging → main` separately.
-
-          **What changed:** just the base branch — no code change. CI will re-run against `staging`. If you get merge conflicts, rebase on `staging`.
-
-          **If this PR is the CEO's staging→main promotion:** the Action skipped you (only bot-authored PRs are retargeted). If you see this comment on your CEO PR, that's a bug — please tag @HongmingWang-Rabbit.
-          BODY
-          )"
@@ -43,7 +43,20 @@ on:
    types: [checks_requested]

 concurrency:
-  group: ${{ github.workflow }}-${{ github.event.pull_request.head.sha || github.sha }}
+  # Include event_name so a PR sync (event=pull_request) and the
+  # subsequent staging push (event=push) on the SAME merge SHA don't
+  # collide in one group. Without event_name, both runs hashed to
+  # the same key and cancel-in-progress=true cancelled whichever
+  # arrived second — usually the push run, which staging branch-
+  # protection then sees as a CANCELLED required check and refuses
+  # to mark merged. Caught 2026-05-05 across PR #2869's runs (run
+  # ids 25371863455 / 25371811486 / 25371078157 / 25370403142 — every
+  # staging push run cancelled, every matching PR run green).
+  #
+  # Per memory `feedback_concurrency_group_per_sha.md` — same drift
+  # class that broke auto-promote-staging on 2026-04-28. Pin invariant:
+  # event_name + sha is the minimum unique key for these workflows.
+  group: ${{ github.workflow }}-${{ github.event_name }}-${{ github.event.pull_request.head.sha || github.sha }}
  cancel-in-progress: true

 jobs:
@@ -12,7 +12,7 @@ name: Secret scan
 #
 #   jobs:
 #     secret-scan:
-#       uses: Molecule-AI/molecule-core/.github/workflows/secret-scan.yml@staging
+#       uses: molecule-ai/molecule-core/.github/workflows/secret-scan.yml@staging
 #
 # Pin to @staging not @main — staging is the active default branch,
 # main lags via the staging-promotion workflow. Updates ride along
@@ -108,6 +108,14 @@ jobs:
          python3 > stale_slugs.txt <<'PY'
          import json, os
          from datetime import datetime, timezone, timedelta
+          # SSOT for this list lives in the controlplane Go code:
+          # molecule-controlplane/internal/slugs/ephemeral.go
+          # (var EphemeralPrefixes). The redeploy-fleet auto-rollout
+          # also reads from there to SKIP these slugs — without that
+          # filter, fleet redeploy SSM-failed in-flight E2E tenants
+          # whose containers were still booting, breaking the test
+          # that just spun them up (molecule-controlplane#493).
+          # Update both files together.
          EPHEMERAL_PREFIXES = ("e2e-", "rt-e2e-")
          with open("orgs.json") as f:
              data = json.load(f)
@@ -159,12 +167,18 @@ jobs:
            # The DELETE handler requires {"confirm": "<slug>"} matching
            # the URL slug — fat-finger guard. Idempotent: re-issuing
            # picks up via org_purges.last_step.
-            http_code=$(curl -sS -o /tmp/del_resp -w "%{http_code}" \
+            # Tempfile-routed -w + set +e/-e prevents curl-exit-code
+            # pollution of the captured status (lint-curl-status-capture.yml).
+            set +e
+            curl -sS -o /tmp/del_resp -w "%{http_code}" \
              --max-time 60 \
              -X DELETE "$MOLECULE_CP_URL/cp/admin/tenants/$slug" \
              -H "Authorization: Bearer $ADMIN_TOKEN" \
              -H "Content-Type: application/json" \
-              -d "{\"confirm\":\"$slug\"}" || echo "000")
+              -d "{\"confirm\":\"$slug\"}" >/tmp/del_code
+            set -e
+            # Stderr from curl (-sS shows dial errors etc.) goes to runner log.
+            http_code=$(cat /tmp/del_code 2>/dev/null || echo "000")
            if [ "$http_code" = "200" ] || [ "$http_code" = "204" ]; then
              deleted=$((deleted+1))
              echo "  deleted: $slug"
@@ -179,7 +193,47 @@ jobs:
          # sweeper is best-effort. Next hourly tick re-attempts. We
          # only fail loud at the safety-cap gate above.

+      - name: Sweep orphan tunnels
+        # Stale-org cleanup deletes the org (which cascades to tunnel
+        # delete inside the CP). But when that cascade fails partway —
+        # CP transient 5xx after the org row is deleted but before the
+        # CF tunnel delete completes — the tunnel persists with no
+        # matching org row. The reconciler in internal/sweep flags this
+        # as `cf_tunnel kind=orphan`, but nothing automatically reaps it.
+        #
+        # `/cp/admin/orphan-tunnels/cleanup` is the operator-triggered
+        # reaper. Calling it here at the end of every sweep tick
+        # converges the staging CF account to clean even when CP
+        # cascades half-fail.
+        #
+        # PR #492 made the underlying DeleteTunnel actually check
+        # status — pre-fix it silent-succeeded on CF code 1022
+        # ("active connections"), so this step would have been a no-op
+        # against stuck connectors. Post-fix the cleanup invokes
+        # CleanupTunnelConnections + retry, which actually clears the
+        # 1022 case. (#2987)
+        #
+        # Best-effort. Failure here doesn't fail the workflow — next
+        # tick re-attempts. Errors flow to step output for ops review.
+        if: env.DRY_RUN != 'true'
+        run: |
+          set +e
+          curl -sS -o /tmp/cleanup_resp -w "%{http_code}" \
+            --max-time 60 \
+            -X POST "$MOLECULE_CP_URL/cp/admin/orphan-tunnels/cleanup" \
+            -H "Authorization: Bearer $ADMIN_TOKEN" >/tmp/cleanup_code
+          set -e
+          http_code=$(cat /tmp/cleanup_code 2>/dev/null || echo "000")
+          body=$(cat /tmp/cleanup_resp 2>/dev/null | head -c 500)
+          if [ "$http_code" = "200" ]; then
+            count=$(echo "$body" | python3 -c "import sys,json; d=json.loads(sys.stdin.read() or '{}'); print(d.get('deleted_count', 0))" 2>/dev/null || echo "0")
+            failed_n=$(echo "$body" | python3 -c "import sys,json; d=json.loads(sys.stdin.read() or '{}'); print(len(d.get('failed') or {}))" 2>/dev/null || echo "0")
+            echo "Orphan-tunnel sweep: deleted=$count failed=$failed_n"
+          else
+            echo "::warning::orphan-tunnels cleanup returned HTTP $http_code — body: $body"
+          fi
+
      - name: Dry-run summary
        if: env.DRY_RUN == 'true'
        run: |
-          echo "DRY RUN — would have deleted ${{ steps.identify.outputs.count }} org(s). Re-run with dry_run=false to actually delete."
+          echo "DRY RUN — would have deleted ${{ steps.identify.outputs.count }} org(s) AND triggered orphan-tunnels cleanup. Re-run with dry_run=false to actually delete."
@@ -131,6 +131,13 @@ backups/
 # Cloned by publish-workspace-server-image.yml so the Dockerfile's
 # replace-directive path resolves. Lives in its own repo.
 /molecule-ai-plugin-github-app-auth/
+# Tenant-image build context — populated by the workflow's
+# "Pre-clone manifest deps" step. Mirrors the public manifest, holds the
+# same content as the three /<>/ dirs above but namespaced under one
+# parent so the Docker build context is a single COPY-friendly tree.
+# Each entry is a transient working-dir, never source-of-truth, never
+# committed.
+/.tenant-bundle-deps/

 # Internal-flavored content lives in Molecule-AI/internal — NEVER in this
 # public monorepo. Migrated 2026-04-23 (CEO directive). The CI workflow
@@ -22,7 +22,7 @@ development workflow, conventions, and how to get your changes merged.

 ```bash
 # Clone the repo
-git clone https://github.com/Molecule-AI/molecule-core.git
+git clone https://git.moleculesai.app/molecule-ai/molecule-core.git
 cd molecule-core

 # Install git hooks
@@ -57,7 +57,7 @@ See `CLAUDE.md` for a full list of environment variables and their purposes.

 This repo is scoped to **code** (canvas, workspace, workspace-server, related
 infra). Public content (blog posts, marketing copy, OG images, SEO briefs,
-DevRel demos) lives in [`Molecule-AI/docs`](https://github.com/Molecule-AI/docs).
+DevRel demos) lives in [`Molecule-AI/docs`](https://git.moleculesai.app/molecule-ai/docs).
 The `Block forbidden paths` CI gate fails any PR that writes to `marketing/`
 or other removed paths — open against `Molecule-AI/docs` instead.

@@ -110,7 +110,7 @@ causing a render loop when any node position changed.

 1. **Repo-wide:** "Automatically delete head branches" is on. Once a PR merges, the branch is deleted server-side. Any subsequent `git push` to that branch fails with `remote rejected — no such branch`.

-2. **CI:** the `pr-guards` workflow (calling [molecule-ci `disable-auto-merge-on-push`](https://github.com/Molecule-AI/molecule-ci/blob/main/.github/workflows/disable-auto-merge-on-push.yml)) fires on every push to an open PR. If auto-merge was already enabled, it's disabled and a comment is posted. You must explicitly re-enable after verifying the new commit.
+2. **CI:** the `pr-guards` workflow (calling [molecule-ci `disable-auto-merge-on-push`](https://git.moleculesai.app/molecule-ai/molecule-ci/src/branch/main/.github/workflows/disable-auto-merge-on-push.yml)) fires on every push to an open PR. If auto-merge was already enabled, it's disabled and a comment is posted. You must explicitly re-enable after verifying the new commit.

 **Workflow rules that follow from the guards:**
 - Push **all** commits before running `gh pr merge --auto`.
@@ -180,9 +180,9 @@ and run CI manually.
 Code in this repo lands in molecule-core. Some related runtime artifacts
 live in their own repos:

- [`Molecule-AI/molecule-ai-workspace-runtime`](https://github.com/Molecule-AI/molecule-ai-workspace-runtime) — Python adapter SDK (`molecule_runtime`) that runs inside containerized Molecule workspaces. Bridges Claude Code SDK / hermes / langgraph / etc. → A2A queue.
- [`Molecule-AI/molecule-sdk-python`](https://github.com/Molecule-AI/molecule-sdk-python) — `A2AServer` + `RemoteAgentClient` for external agents that register over the public `/registry/register` flow.
- [`Molecule-AI/molecule-mcp-claude-channel`](https://github.com/Molecule-AI/molecule-mcp-claude-channel) — Claude Code channel plugin. Bridges A2A traffic into a running Claude Code session via MCP `notifications/claude/channel`. Polling-based (no tunnel required); install with `claude --channels plugin:molecule@Molecule-AI/molecule-mcp-claude-channel`.
+- [`Molecule-AI/molecule-ai-workspace-runtime`](https://git.moleculesai.app/molecule-ai/molecule-ai-workspace-runtime) — Python adapter SDK (`molecule_runtime`) that runs inside containerized Molecule workspaces. Bridges Claude Code SDK / hermes / langgraph / etc. → A2A queue.
+- [`Molecule-AI/molecule-sdk-python`](https://git.moleculesai.app/molecule-ai/molecule-sdk-python) — `A2AServer` + `RemoteAgentClient` for external agents that register over the public `/registry/register` flow.
+- [`Molecule-AI/molecule-mcp-claude-channel`](https://git.moleculesai.app/molecule-ai/molecule-mcp-claude-channel) — Claude Code channel plugin. Bridges A2A traffic into a running Claude Code session via MCP `notifications/claude/channel`. Polling-based (no tunnel required); install with `claude --channels plugin:molecule@Molecule-AI/molecule-mcp-claude-channel`.

 When extending the **A2A surface** in molecule-core (`workspace-server/internal/handlers/a2a_proxy.go` etc.), consider whether the change has a downstream impact on the runtime SDK or the channel plugin — they're versioned independently but share the wire shape.

@@ -1,7 +1,7 @@
 # Coverage Floor

-CI enforces three coverage gates on `workspace-server` (Go). All defined in
-`.github/workflows/ci.yml` → `platform-build` job.
+CI enforces coverage gates on two surfaces — `workspace-server` (Go) and
+`workspace/` (Python). All defined in `.github/workflows/ci.yml`.

 ## Current floors (2026-04-23)

@@ -76,3 +76,51 @@ This gate makes "no untested critical paths merged" a mechanical property of
 the CI, not a behavioural property of QA agents or individual reviewers —
 which is the only way to make it survive fleet outages, agent rotations, or
 QA process changes.
+
+## Python (workspace/) — added 2026-05-04 from #2790
+
+The Python side has its own gates in the `python-lint` job:
+
+| Gate | Threshold | Where |
+|---|---|---|
+| **Total floor** | `86%` | `workspace/pytest.ini` `--cov-fail-under=86` (issue #1817) |
+| **Critical-path per-file floor** | `75%` | Inline shell step after the pytest run |
+
+### Critical-path Python files
+
+These handle multi-tenant routing, auth tokens, and inbox dispatch. A
+coverage drop here is the same risk shape as a Go-side `tokens*` /
+`secrets*` file regressing below 10%.
+
+- `workspace/a2a_mcp_server.py` — MCP dispatcher (PR #2766 / #2771)
+- `workspace/mcp_cli.py` — molecule-mcp standalone CLI entry
+- `workspace/a2a_tools.py` — workspace-scoped tool implementations
+- `workspace/inbox.py` — multi-workspace inbox + per-workspace cursors
+- `workspace/platform_auth.py` — per-workspace token resolver
+
+### Why 75% (vs 86% total)
+
+The total floor averages ~6000 lines across `workspace/`. A single MCP
+file could drop to ~50% with no CI complaint as long as other modules
+compensate. The per-file floor closes that distribution gap. 75% sits
+below current actuals (80–96% as of 2026-05-04) — strictly additive,
+no existing PR fails.
+
+### Python ratchet plan
+
+| Date | Total | Per-file critical | Notes |
+|---|---|---|---|
+| 2026-05-04 | 86% | 75% | Initial gate (this file). |
+| 2026-06-04 | 86% | 80% | First ratchet — at-floor files must catch up. |
+| 2026-07-04 | 88% | 85% | |
+| 2026-08-04 | 90% | 90% | Target steady-state. |
+
+### Why this Python gate exists
+
+Issue #2790, after the PR #2766 → PR #2771 cycle. PR #2766 added
+multi-workspace routing through `a2a_tools.py` + `a2a_mcp_server.py`,
+shipped to main with green CI, but the dispatcher silently dropped a
+load-bearing kwarg for 4 of 9 tools — caught only by post-merge code
+review. The structural drift gate (`test_dispatcher_schema_drift.py`,
+PR #2791) catches the schema↔dispatcher mismatch class; this floor
+catches the broader "MCP-critical file regressed" class.
@@ -0,0 +1,28 @@
+# Top-level Makefile — convenience wrappers around docker compose.
+#
+# Most molecule-core dev work happens via these shortcuts. CI doesn't
+# use this Makefile; CI calls docker compose / go test directly so the
+# Makefile can evolve without breaking the build.
+
+.PHONY: help dev up down logs build test
+
+help: ## Show this help.
+	@grep -E '^[a-zA-Z_-]+:.*?## ' $(MAKEFILE_LIST) | awk 'BEGIN {FS = ":.*?## "}; {printf "\033[36m%-12s\033[0m %s\n", $$1, $$2}'
+
+dev: ## Start the full stack with air hot-reload for the platform service.
+	docker compose -f docker-compose.yml -f docker-compose.dev.yml up
+
+up: ## Start the full stack in production-shape mode (no air, normal Dockerfile).
+	docker compose up
+
+down: ## Stop the stack and remove containers (volumes preserved).
+	docker compose down
+
+logs: ## Tail logs from all services (Ctrl-C to detach).
+	docker compose logs -f
+
+build: ## Force a fresh build of the platform image (no cache).
+	docker compose build --no-cache platform
+
+test: ## Run Go unit tests in workspace-server/.
+	cd workspace-server && go test -race ./...
@@ -1,7 +1,7 @@
 <div align="center">

 <p>
-  <img src="./docs/assets/branding/molecule-icon.png" alt="Molecule AI Icon Logo" width="160" />
+  <img src="./docs/assets/branding/molecule-icon.svg" alt="Molecule AI" width="160" />
 </p>

 <p>
@@ -39,8 +39,8 @@
  <a href="./docs/agent-runtime/workspace-runtime.md"><strong>Workspace Runtime</strong></a>
 </p>

-[![Deploy on Railway](https://railway.app/button.svg)](https://railway.app/new/template?template=https://github.com/Molecule-AI/molecule-monorepo)
-[![Deploy to Render](https://render.com/images/deploy-to-render-button.svg)](https://render.com/deploy?repo=https://github.com/Molecule-AI/molecule-monorepo)
+[![Deploy on Railway](https://railway.app/button.svg)](https://railway.app/new/template?template=https://git.moleculesai.app/molecule-ai/molecule-core)
+[![Deploy to Render](https://render.com/images/deploy-to-render-button.svg)](https://render.com/deploy?repo=https://git.moleculesai.app/molecule-ai/molecule-core)

 </div>

@@ -53,8 +53,8 @@ Molecule AI is the most powerful way to govern an AI agent organization in produ
 It combines the parts that are usually scattered across demos, internal glue code, and framework-specific tooling into one product:

 - one org-native control plane for teams, roles, hierarchy, and lifecycle
- one runtime layer that lets LangGraph, DeepAgents, Claude Code, CrewAI, AutoGen, and OpenClaw run side by side
- one memory model that keeps recall, sharing, and skill evolution aligned with organizational boundaries
+- one runtime layer that lets **eight** agent runtimes — LangGraph, DeepAgents, Claude Code, CrewAI, AutoGen, **Hermes**, **Gemini CLI**, and OpenClaw — run side by side behind one workspace contract
+- one memory model that keeps recall, sharing, and skill evolution aligned with organizational boundaries (Memory v2 backed by pgvector for semantic recall)
 - one operational surface for observing, pausing, restarting, inspecting, and improving live workspaces

 Most teams can build a workflow, a strong single agent, a coding agent, or a custom multi-agent graph.
@@ -75,7 +75,7 @@ You do not wire collaboration paths by hand. Hierarchy defines the default commu

 ### 3. Runtime choice stops being a dead-end decision

-LangGraph, DeepAgents, Claude Code, CrewAI, AutoGen, and OpenClaw can all plug into the same workspace abstraction. Teams can standardize governance without forcing every group onto one runtime.
+LangGraph, DeepAgents, Claude Code, CrewAI, AutoGen, Hermes, Gemini CLI, and OpenClaw can all plug into the same workspace abstraction. Teams can standardize governance without forcing every group onto one runtime.

 ### 4. Memory is treated like infrastructure

@@ -117,6 +117,8 @@ Molecule AI is not trying to replace the frameworks below. It is the system that
 | **Claude Code** | Shipping on `main` | Real coding workflows, CLI-native continuity | Secure workspace abstraction, A2A delegation, org boundaries, shared control plane |
 | **CrewAI** | Shipping on `main` | Role-based crews | Persistent workspace identity, policy consistency, shared canvas and registry |
 | **AutoGen** | Shipping on `main` | Assistant/tool orchestration | Standardized deployment, hierarchy-aware collaboration, shared ops plane |
+| **Hermes 4** | Shipping on `main` | Hybrid reasoning, native tools, json_schema (NousResearch/hermes-agent) | Option B upstream hook, A2A bridge to OpenAI-compat API, multi-provider provider derivation |
+| **Gemini CLI** | Shipping on `main` | Google Gemini CLI continuity | Workspace lifecycle, A2A, hierarchy-aware collaboration, shared ops plane |
 | **OpenClaw** | Shipping on `main` | CLI-native runtime with its own session model | Workspace lifecycle, templates, activity logs, topology-aware collaboration |
 | **NemoClaw** | WIP on `feat/nemoclaw-t4-docker` | NVIDIA-oriented runtime path | Planned to join the same abstraction once merged; not yet part of `main` |

@@ -182,9 +184,10 @@ The result is not just “an agent that learns.” It is **an organization that

 ## What Ships In `main`

-### Canvas
+### Canvas (v4)

 - Next.js 15 + React Flow + Zustand
+- **warm-paper theme system** — light / dark / follow-system, SSR cookie + nonce'd boot script + ThemeProvider; terminal + code surfaces stay dark unconditionally
 - drag-to-nest team building
 - empty-state deployment + onboarding wizard
 - template palette
@@ -193,8 +196,9 @@ The result is not just “an agent that learns.” It is **an organization that

 ### Platform

- Go/Gin control plane
- workspace CRUD and provisioning
+- Go 1.25 / Gin control plane (80+ HTTP endpoints + Gorilla WebSocket fanout)
+- workspace CRUD and provisioning (pluggable Provisioner — Docker locally, EC2 + SSM in production)
+- **A2A response path is a typed discriminated union (RFC #2967)** — frozen dataclasses + total parser; 100% unit + adversarial fuzz coverage
 - registry and heartbeats
 - browser-safe A2A proxy
 - team expansion/collapse
@@ -204,10 +208,10 @@ The result is not just “an agent that learns.” It is **an organization that

 ### Runtime

- unified `workspace/` image
- adapter-driven execution
+- unified `workspace/` image; thin AMI in production (us-east-2)
+- adapter-driven execution across **8 runtimes** (Claude Code, Hermes, Gemini CLI, LangGraph, DeepAgents, CrewAI, AutoGen, OpenClaw)
 - Agent Card registration
- awareness-backed memory integration
+- awareness-backed memory integration; **Memory v2 backed by pgvector** for semantic recall
 - plugin-mounted shared rules/skills
 - hot-reloadable local skills
 - coordinator-only delegation path
@@ -221,6 +225,21 @@ The result is not just “an agent that learns.” It is **an organization that
 - runtime tiers
 - direct workspace inspection through terminal and files

+### SaaS (via [`molecule-controlplane`](https://git.moleculesai.app/molecule-ai/molecule-controlplane))
+
+- multi-tenant on AWS EC2 + Neon (per-tenant Postgres branch) + Cloudflare Tunnels (per-tenant, no public ports)
+- WorkOS AuthKit + Stripe Checkout + Customer Portal
+- AWS KMS envelope encryption (DB / Redis connection strings); AWS Secrets Manager for tenant bootstrap
+- `tenant_resources` audit table + 30-min boot-event-aware reconciler — every CF / AWS lifecycle event recorded, claim vs live state diffed
+
+### Bring your own Claude Code session (via [`molecule-mcp-claude-channel`](https://git.moleculesai.app/molecule-ai/molecule-mcp-claude-channel))
+
+- Claude Code plugin that bridges Molecule A2A traffic into a local Claude Code session via MCP
+- subscribe to one or more workspaces; peer messages surface as conversation turns; replies route back through Molecule's A2A
+- no tunnel, no public endpoint — the plugin self-registers each watched workspace as `delivery_mode=poll` and long-polls `/activity?since_id=…`
+- multi-tenant friendly: one plugin install can watch workspaces across multiple Molecule tenants (`MOLECULE_PLATFORM_URLS` per-workspace)
+- install via the standard marketplace flow: `/plugin marketplace add Molecule-AI/molecule-mcp-claude-channel` → `/plugin install molecule-channel@molecule-mcp-claude-channel`
+
 ## Built For Teams That Need More Than A Demo

 Molecule AI is especially strong when you need to run:
@@ -233,24 +252,30 @@ Molecule AI is especially strong when you need to run:
 ## Architecture

 ```text
-Canvas (Next.js :3000)  <--HTTP / WS-->  Platform (Go :8080)  <---> Postgres + Redis
-         |                                          |
-         |                                          +--> Docker provisioner / bundles / templates / secrets
+Canvas (Next.js 15, warm-paper :3000)  <--HTTP / WS-->  Platform (Go 1.25 :8080)  <---> Postgres + Redis
+         |                                                           |
+         |                                                           +--> Provisioner: Docker (local) / EC2 + SSM (prod)
+         |                                                           +--> bundles · templates · secrets · KMS
         |
-         +-------------------- shows --------------------> workspaces, teams, tasks, traces, events
+         +------------------------- shows ------------------------> workspaces, teams, tasks, traces, events

-Workspace Runtime (Python image with adapters)
-  - LangGraph / DeepAgents / Claude Code / CrewAI / AutoGen / OpenClaw
-  - Agent Card + A2A server
-  - heartbeat + activity + awareness-backed memory
+Workspace Runtime (Python ≥3.11, image with adapters)
+  - 8 adapters: LangGraph / DeepAgents / Claude Code / CrewAI / AutoGen / Hermes / Gemini CLI / OpenClaw
+  - Agent Card + A2A server (typed-SSOT response path, RFC #2967)
+  - heartbeat + activity + awareness-backed memory (Memory v2 — pgvector semantic recall)
  - skills + plugins + hot reload
+
+SaaS Control Plane (molecule-controlplane, private)
+  - per-tenant EC2 + Neon (Postgres branch) + Cloudflare Tunnel
+  - WorkOS · Stripe · KMS · AWS Secrets Manager
+  - tenant_resources audit + 30-min reconciler
 ```

 ## Quick Start

 ```bash
-git clone https://github.com/Molecule-AI/molecule-monorepo.git
-cd molecule-monorepo
+git clone https://git.moleculesai.app/molecule-ai/molecule-core.git
+cd molecule-core

 cp .env.example .env
 # Defaults boot the stack locally out of the box. See .env.example for
@@ -303,7 +328,11 @@ Then open `http://localhost:3000`:

 ## Current Scope

-The current `main` branch already includes the core platform, canvas, memory model, six production adapters, skill lifecycle, and operational surfaces. Adjacent runtime work such as **NemoClaw** remains branch-level until merged, and this README keeps that distinction explicit on purpose.
+The current `main` branch ships the core platform, Canvas v4 (warm-paper themed), Memory v2 (pgvector semantic recall), the typed-SSOT A2A response path (RFC #2967), **eight production adapters** (Claude Code, Hermes, Gemini CLI, LangGraph, DeepAgents, CrewAI, AutoGen, OpenClaw), skill lifecycle, and operational surfaces.
+
+The companion private repo [`molecule-controlplane`](https://git.moleculesai.app/molecule-ai/molecule-controlplane) provides the SaaS surface — multi-tenant orchestration on EC2 + Neon + Cloudflare Tunnels, KMS envelope encryption, WorkOS auth, Stripe billing, and a `tenant_resources` audit table with a 30-min reconciler.
+
+Adjacent runtime work such as **NemoClaw** remains branch-level until merged, and this README keeps that distinction explicit on purpose.

 ## License

@@ -1,7 +1,7 @@
 <div align="center">

 <p>
-  <img src="./docs/assets/branding/molecule-icon.png" alt="Molecule AI 图案 Logo" width="160" />
+  <img src="./docs/assets/branding/molecule-icon.svg" alt="Molecule AI" width="160" />
 </p>

 <p>
@@ -38,8 +38,8 @@
  <a href="./docs/agent-runtime/workspace-runtime.md"><strong>Workspace Runtime</strong></a>
 </p>

-[![Deploy on Railway](https://railway.app/button.svg)](https://railway.app/new/template?template=https://github.com/Molecule-AI/molecule-core)
-[![Deploy to Render](https://render.com/images/deploy-to-render-button.svg)](https://render.com/deploy?repo=https://github.com/Molecule-AI/molecule-core)
+[![Deploy on Railway](https://railway.app/button.svg)](https://railway.app/new/template?template=https://git.moleculesai.app/molecule-ai/molecule-core)
+[![Deploy to Render](https://render.com/images/deploy-to-render-button.svg)](https://render.com/deploy?repo=https://git.moleculesai.app/molecule-ai/molecule-core)

 </div>

@@ -52,8 +52,8 @@ Molecule AI 是目前最强的 AI Agent 组织治理方案之一，用来把 age
 它把过去分散在 demo、内部胶水代码和各类 framework 私有工具里的关键能力，收敛成一个产品：

 - 一套组织原生 control plane，管理团队、角色、层级和生命周期
- 一套 runtime abstraction，让 LangGraph、DeepAgents、Claude Code、CrewAI、AutoGen、OpenClaw 并存运行
- 一套与组织边界对齐的 memory 模型，把 recall、sharing 和 skill evolution 放进同一体系
+- 一套 runtime abstraction，让 **8 个** agent runtime —— LangGraph、DeepAgents、Claude Code、CrewAI、AutoGen、**Hermes**、**Gemini CLI**、OpenClaw —— 共用一套 workspace 契约
+- 一套与组织边界对齐的 memory 模型，把 recall、sharing 和 skill evolution 放进同一体系（Memory v2 由 pgvector 支撑语义召回）
 - 一套面向线上 workspace 的运维面，统一完成观测、暂停、重启、检查和持续改进

 今天很多团队能做好 workflow、单 agent、coding agent，或者自定义 multi-agent graph 中的一种。
@@ -74,7 +74,7 @@ Molecule AI 填的就是这个空白。

 ### 3. Runtime 选择不再是死路

-LangGraph、DeepAgents、Claude Code、CrewAI、AutoGen、OpenClaw 都可以挂到同一个 workspace abstraction 下。团队可以统一治理方式，而不必统一到底层 runtime。
+LangGraph、DeepAgents、Claude Code、CrewAI、AutoGen、Hermes、Gemini CLI、OpenClaw 都可以挂到同一个 workspace abstraction 下。团队可以统一治理方式，而不必统一到底层 runtime。

 ### 4. Memory 被当成基础设施来做

@@ -116,6 +116,8 @@ Molecule AI 并不是要替代下面这些 framework，而是把它们纳入更
 | **Claude Code** | `main` 已支持 | 真实编码工作流、CLI-native continuity | 安全 workspace 抽象、A2A delegation、组织边界、共享 control plane |
 | **CrewAI** | `main` 已支持 | 角色型 crew 模式清晰 | 持久 workspace 身份、统一策略、共享 Canvas 和 registry |
 | **AutoGen** | `main` 已支持 | assistant/tool orchestration | 统一部署、层级协作、共享运维平面 |
+| **Hermes 4** | `main` 已支持 | 混合推理、原生工具调用、json_schema 输出（NousResearch/hermes-agent） | Option B 上游 hook、A2A 桥接 OpenAI 兼容 API、多 provider 自动派生 |
+| **Gemini CLI** | `main` 已支持 | Google Gemini CLI 持续会话 | workspace 生命周期、A2A、层级感知协作、共享运维平面 |
 | **OpenClaw** | `main` 已支持 | CLI-native runtime，自有 session 模型 | workspace 生命周期、templates、activity logs、拓扑感知协作 |
 | **NemoClaw** | `feat/nemoclaw-t4-docker` 分支 WIP | NVIDIA 方向 runtime 路线 | 计划并入同一抽象层，但当前还不是 `main` 已合并能力 |

@@ -181,9 +183,10 @@ Molecule AI 并不是要替代下面这些 framework，而是把它们纳入更

 ## `main` 分支已经具备什么

-### Canvas
+### Canvas（v4）

 - Next.js 15 + React Flow + Zustand
+- **warm-paper 主题系统** —— light / dark / 跟随系统；SSR cookie + nonce'd boot 脚本 + ThemeProvider；终端与代码面板始终保持深色
 - drag-to-nest 团队构建
 - empty state + onboarding wizard
 - template palette
@@ -192,8 +195,9 @@ Molecule AI 并不是要替代下面这些 framework，而是把它们纳入更

 ### Platform

- Go/Gin control plane
- workspace CRUD 和 provisioning
+- Go 1.25 / Gin control plane（80+ HTTP 端点 + Gorilla WebSocket fanout）
+- workspace CRUD 和 provisioning（可插拔 Provisioner —— 本地 Docker、生产 EC2 + SSM）
+- **A2A 响应路径已收敛为类型化的判别联合（RFC #2967）** —— 冻结 dataclass + 全量 parser；100% 单元测试 + 对抗性 fuzz 覆盖
 - registry 与 heartbeat
 - 浏览器安全的 A2A proxy
 - team expansion/collapse
@@ -203,10 +207,10 @@ Molecule AI 并不是要替代下面这些 framework，而是把它们纳入更

 ### Runtime

- 统一 `workspace/` 镜像
- adapter 驱动执行
+- 统一 `workspace/` 镜像；生产环境采用 thin AMI（us-east-2）
+- adapter 驱动执行，覆盖 **8 个 runtime**（Claude Code、Hermes、Gemini CLI、LangGraph、DeepAgents、CrewAI、AutoGen、OpenClaw）
 - Agent Card 注册
- awareness-backed memory
+- awareness-backed memory；**Memory v2 由 pgvector 支撑**语义召回
 - plugin 挂载共享 rules/skills
 - 本地 skills 热加载
 - coordinator-only delegation 路径
@@ -220,6 +224,21 @@ Molecule AI 并不是要替代下面这些 framework，而是把它们纳入更
 - runtime tiers
 - 终端与文件层面的 workspace 直接排障

+### SaaS（由 [`molecule-controlplane`](https://git.moleculesai.app/molecule-ai/molecule-controlplane) 提供）
+
+- 多租户运行在 AWS EC2 + Neon（每租户一个 Postgres branch）+ Cloudflare Tunnels（每租户一条隧道，对外不开任何端口）
+- WorkOS AuthKit + Stripe Checkout + Customer Portal
+- AWS KMS 信封加密（DB / Redis 连接串）；AWS Secrets Manager 负责租户 bootstrap
+- `tenant_resources` 审计表 + 30 分钟 boot-event-aware reconciler —— 每个 CF / AWS lifecycle 事件都有记录，每 30 分钟比对 claim 与实际状态
+
+### 在 Claude Code 里直接接入（由 [`molecule-mcp-claude-channel`](https://git.moleculesai.app/molecule-ai/molecule-mcp-claude-channel) 提供）
+
+- 把 Molecule A2A 流量桥接到本地 Claude Code 会话的 MCP 插件
+- 订阅一个或多个 workspace；peer 的消息会以 user-turn 出现，回复会经 Molecule A2A 路由出去
+- 无需公网隧道、无需公开端点 —— 插件启动时自动把每个 watched workspace 注册成 `delivery_mode=poll`，长轮询 `/activity?since_id=…`
+- 多租户友好：单次安装即可同时 watch 跨多个 Molecule 租户的 workspace（`MOLECULE_PLATFORM_URLS` 按 workspace 配置）
+- 通过标准 marketplace 流程安装：`/plugin marketplace add Molecule-AI/molecule-mcp-claude-channel` → `/plugin install molecule-channel@molecule-mcp-claude-channel`
+
 ## 适合什么团队

 Molecule AI 特别适合下面这些场景：
@@ -232,23 +251,29 @@ Molecule AI 特别适合下面这些场景：
 ## 架构总览

 ```text
-Canvas (Next.js :3000)  <--HTTP / WS-->  Platform (Go :8080)  <---> Postgres + Redis
-         |                                          |
-         |                                          +--> Docker provisioner / bundles / templates / secrets
+Canvas (Next.js 15, warm-paper :3000)  <--HTTP / WS-->  Platform (Go 1.25 :8080)  <---> Postgres + Redis
+         |                                                           |
+         |                                                           +--> Provisioner: Docker (本地) / EC2 + SSM (生产)
+         |                                                           +--> bundles · templates · secrets · KMS
         |
-         +-------------------- 展示 --------------------> workspaces, teams, tasks, traces, events
+         +------------------------- 展示 ------------------------> workspaces, teams, tasks, traces, events

-Workspace Runtime (Python image with adapters)
-  - LangGraph / DeepAgents / Claude Code / CrewAI / AutoGen / OpenClaw
-  - Agent Card + A2A server
-  - heartbeat + activity + awareness-backed memory
+Workspace Runtime (Python ≥3.11，含 adapter 集合的镜像)
+  - 8 个 adapter: LangGraph / DeepAgents / Claude Code / CrewAI / AutoGen / Hermes / Gemini CLI / OpenClaw
+  - Agent Card + A2A server（typed-SSOT 响应路径，RFC #2967）
+  - heartbeat + activity + awareness-backed memory（Memory v2 —— pgvector 语义召回）
  - skills + plugins + hot reload
+
+SaaS Control Plane (molecule-controlplane，私有)
+  - 每租户 EC2 + Neon (Postgres branch) + Cloudflare Tunnel
+  - WorkOS · Stripe · KMS · AWS Secrets Manager
+  - tenant_resources 审计 + 30 分钟 reconciler
 ```

 ## 快速开始

 ```bash
-git clone https://github.com/Molecule-AI/molecule-core.git
+git clone https://git.moleculesai.app/molecule-ai/molecule-core.git
 cd molecule-core

 cp .env.example .env
@@ -296,7 +321,11 @@ npm run dev

 ## 当前范围说明

-当前 `main` 已经包含核心平台、Canvas、memory model、6 个正式 adapter、skill lifecycle 和主要运维面。像 **NemoClaw** 这样的相邻 runtime 路线仍然属于分支级工作，只有合并后才会进入正式支持列表，这里会明确区分。
+当前 `main` 已经包含核心平台、Canvas v4（warm-paper 主题）、Memory v2（pgvector 语义召回）、typed-SSOT A2A 响应路径（RFC #2967）、**8 个正式 adapter**（Claude Code、Hermes、Gemini CLI、LangGraph、DeepAgents、CrewAI、AutoGen、OpenClaw）、skill lifecycle，以及主要运维面。
+
+配套的私有仓库 [`molecule-controlplane`](https://git.moleculesai.app/molecule-ai/molecule-controlplane) 提供 SaaS 层 —— 多租户编排（EC2 + Neon + Cloudflare Tunnels）、KMS 信封加密、WorkOS 鉴权、Stripe 计费，以及 `tenant_resources` 审计表加 30 分钟 reconciler。
+
+像 **NemoClaw** 这样的相邻 runtime 路线仍然属于分支级工作，只有合并后才会进入正式支持列表，这里会明确区分。

 ## License

@@ -3,6 +3,7 @@ import { cookies, headers } from "next/headers";
 import "./globals.css";
 import { AuthGate } from "@/components/AuthGate";
 import { CookieConsent } from "@/components/CookieConsent";
+import { PurchaseSuccessModal } from "@/components/PurchaseSuccessModal";
 import { ThemeProvider } from "@/lib/theme-provider";
 import {
  THEME_COOKIE,
@@ -86,6 +87,12 @@ export default async function RootLayout({
              vercel preview URL, apex) pass through unchanged. */}
          <AuthGate>{children}</AuthGate>
          <CookieConsent />
+          {/* Demo Mock #1: post-purchase success toast. Mounted at the
+              layout level so it persists across page state transitions
+              (loading → hydrated → error) without being unmounted and
+              losing its open-state. Reads ?purchase_success=1 from the
+              URL on first paint, then strips the param. */}
+          <PurchaseSuccessModal />
        </ThemeProvider>
      </body>
    </html>
@@ -18,7 +18,7 @@
 // quick bounce between signup and either Checkout or the tenant UI.

 import { useEffect, useState } from "react";
-import { fetchSession, redirectToLogin, type Session } from "@/lib/auth";
+import { fetchSession, redirectToLogin, signOut, type Session } from "@/lib/auth";
 import { PLATFORM_URL } from "@/lib/api";
 import { formatCredits, pillTone, bannerKind } from "@/lib/credits";
 import { TermsGate } from "@/components/TermsGate";
@@ -129,7 +129,7 @@ export default function OrgsPage() {
    return <EmptyState banner={justCheckedOut ? <CheckoutBanner /> : null} />;
  }
  return (
-    <Shell>
+    <Shell session={session}>
      {justCheckedOut && <CheckoutBanner />}
      <ul className="space-y-3">
        {orgs.map((o) => (
@@ -160,11 +160,21 @@ function CheckoutBanner() {
  );
 }

-function Shell({ children }: { children: React.ReactNode }) {
+function Shell({
+  children,
+  session,
+}: {
+  children: React.ReactNode;
+  // Optional: when present, the header renders the signed-in email +
+  // a Sign-out button. The empty-state Shell call doesn't have a
+  // session in scope, so accept null and skip the header chrome there.
+  session?: Session | null;
+}) {
  return (
    <main className="min-h-screen bg-surface text-ink">
      <TermsGate>
        <div className="mx-auto max-w-2xl px-6 pt-20 pb-12">
+          {session ? <AccountBar session={session} /> : null}
          <h1 className="text-3xl font-bold text-ink">Your organizations</h1>
          <p className="mt-2 text-ink-mid">
            Each org is an isolated Molecule workspace.
@@ -177,6 +187,40 @@ function Shell({ children }: { children: React.ReactNode }) {
  );
 }

+// AccountBar renders the signed-in email + a Sign-out button at the
+// top of the page. Without this the user has no way to log out — the
+// /cp/auth/signout endpoint exists on the control plane but no UI ever
+// called it. Reported externally on 2026-05-05; this is the fix.
+//
+// Click → calls signOut() which POSTs /cp/auth/signout (clears the
+// WorkOS session cookie + revokes at the provider) then bounces to
+// /cp/auth/login. The signOut helper is best-effort — even on a 5xx
+// or network failure the redirect fires so the user never gets stuck
+// on an authed-looking page after they clicked Sign out.
+function AccountBar({ session }: { session: Session }) {
+  const [signingOut, setSigningOut] = useState(false);
+  return (
+    <div className="mb-6 flex items-center justify-between text-sm text-ink-mid">
+      <span title="Signed-in user">{session.email}</span>
+      <button
+        type="button"
+        disabled={signingOut}
+        onClick={async () => {
+          setSigningOut(true);
+          await signOut();
+          // Redirect happens inside signOut; this line is for tests +
+          // edge cases (jsdom, blocked navigation) where it doesn't.
+          setSigningOut(false);
+        }}
+        className="rounded border border-line bg-surface-card px-3 py-1 text-xs text-ink hover:bg-surface-card disabled:opacity-50"
+        aria-label="Sign out"
+      >
+        {signingOut ? "Signing out…" : "Sign out"}
+      </button>
+    </div>
+  );
+}
+
 // DataResidencyNotice surfaces where workspace data lives so EU-based
 // signups can make an informed choice (GDPR Art. 13 disclosure
 // requirement). Plain text, no icon — the goal is clarity, not
@@ -41,7 +41,7 @@ export default function PricingPage() {
        <p className="mt-2 text-ink-mid">
          We publish the{" "}
          <a
-            href="https://github.com/Molecule-AI/molecule-monorepo"
+            href="https://git.moleculesai.app/molecule-ai/molecule-monorepo"
            className="text-accent underline hover:text-accent"
          >
            full source on GitHub
@@ -1,9 +1,10 @@
 'use client';

-import { useEffect, useMemo, useCallback } from "react";
+import { useEffect, useMemo, useCallback, useRef } from "react";
 import { type Edge, MarkerType } from "@xyflow/react";
 import { api } from "@/lib/api";
 import { useCanvasStore } from "@/store/canvas";
+import { useSocketEvent } from "@/hooks/useSocketEvent";
 import type { ActivityEntry } from "@/types/activity";

 // ── Constants ─────────────────────────────────────────────────────────────────
@@ -11,9 +12,6 @@ import type { ActivityEntry } from "@/types/activity";
 /** 60-minute look-back window for delegation activity */
 export const A2A_WINDOW_MS = 60 * 60 * 1000;

-/** Polling interval — refresh edges every 60 seconds */
-export const A2A_POLL_MS = 60 * 1_000;
-
 /** Threshold for "hot" edges: < 5 minutes → animated + violet stroke */
 export const A2A_HOT_MS = 5 * 60 * 1_000;

@@ -131,6 +129,20 @@ export function buildA2AEdges(
 * `a2aEdges`. Canvas.tsx merges these with topology edges and passes the
 * combined list to ReactFlow.
 *
+ * Update shape (issue #61 Stage 2, replaces the 60s polling loop):
+ *  - On mount (when showA2AEdges): one HTTP fan-out per visible workspace
+ *    (delegation rows, 60-min window). Bootstraps the local row buffer.
+ *  - Steady state: subscribes to ACTIVITY_LOGGED via useSocketEvent.
+ *    Each delegation event from a visible workspace is appended to the
+ *    buffer; edges are re-derived via the existing buildA2AEdges helper.
+ *  - showA2AEdges toggle off: clears edges + buffer.
+ *  - Visible-ID-set change: re-bootstraps so a freshly-shown workspace
+ *    backfills its 60-min history (existing visibleIdsKey selector
+ *    behaviour preserved — that's the 2026-05-04 render-loop fix).
+ *
+ * No interval poll. The singleton ReconnectingSocket already owns
+ * reconnect / backoff / health-check; useSocketEvent inherits those.
+ *
 * Mount this inside CanvasInner (no ReactFlow hook dependency).
 */
 export function A2ATopologyOverlay() {
@@ -157,7 +169,9 @@ export function A2ATopologyOverlay() {
  // the symptom of this re-render storm.
  //
  // The fix is purely the dependency-stability change here; the fetch
-  // logic is unchanged.
+  // logic is unchanged. Post-#61 the polling-driven fetch is gone, but
+  // the visibleIdsKey gate is still required so a peer-discovery write
+  // doesn't trigger a wasteful re-bootstrap.
  const visibleIdsKey = useCanvasStore((s) =>
    s.nodes
      .filter((n) => !n.hidden)
@@ -171,16 +185,42 @@ export function A2ATopologyOverlay() {
    [visibleIdsKey]
  );

-  // Fetch delegation activity for all visible workspaces and rebuild overlay edges.
-  const fetchAndUpdate = useCallback(async () => {
+  // Local rolling buffer of delegation rows. Pruned by A2A_WINDOW_MS on
+  // each rebuild so a long-lived session doesn't accumulate unbounded
+  // history. The buffer's high-water mark is approximately:
+  //    visibleIds.length × bootstrap-fetch-limit (500) + WS arrivals
+  // Real-world ceiling: ~3000 entries at the 60-min boundary, all of
+  // which buildA2AEdges aggregates into at most N² edges.
+  const bufferRef = useRef<ActivityEntry[]>([]);
+  // visibleIdsRef gives the WS handler the latest visible-ID set without
+  // re-subscribing on every render. The bus listener is registered
+  // exactly once per mount; subscriber-side filtering reads from this ref.
+  const visibleIdsRef = useRef(visibleIds);
+  visibleIdsRef.current = visibleIds;
+
+  // Re-derive overlay edges from the current buffer + push to store.
+  // Prunes by A2A_WINDOW_MS first so memory stays bounded across long
+  // sessions and the aggregation cost stays O(window-size).
+  const recomputeAndPush = useCallback(() => {
+    const cutoff = Date.now() - A2A_WINDOW_MS;
+    bufferRef.current = bufferRef.current.filter(
+      (r) => new Date(r.created_at).getTime() > cutoff
+    );
+    setA2AEdges(buildA2AEdges(bufferRef.current));
+  }, [setA2AEdges]);
+
+  // Bootstrap fan-out — one HTTP per visible workspace. Replaces the
+  // 60s polling loop entirely. Race-aware: any WS arrivals that landed
+  // in the buffer DURING the fetch (between the await and resume) are
+  // preserved by id-dedup-with-fetched-first ordering.
+  const bootstrap = useCallback(async () => {
    if (visibleIds.length === 0) {
+      bufferRef.current = [];
      setA2AEdges([]);
      return;
    }
    try {
-      // Fan-out — one request per visible workspace.
-      // Per-request failures are swallowed so one broken workspace doesn't blank the overlay.
-      const allRows = (
+      const fetchedRows = (
        await Promise.all(
          visibleIds.map((id) =>
            api
@@ -192,24 +232,76 @@ export function A2ATopologyOverlay() {
        )
      ).flat();

-      setA2AEdges(buildA2AEdges(allRows));
+      // Merge: fetched rows first, then any in-flight WS arrivals that
+      // accumulated during the await. Dedup by id so rows that appear
+      // in both paths are not double-counted in the aggregation.
+      const merged = [...fetchedRows, ...bufferRef.current];
+      const seen = new Set<string>();
+      bufferRef.current = merged.filter((r) => {
+        if (seen.has(r.id)) return false;
+        seen.add(r.id);
+        return true;
+      });
+      recomputeAndPush();
    } catch {
      // Overlay failure is non-critical — canvas remains functional
    }
-  }, [visibleIds, setA2AEdges]);
+  }, [visibleIds, setA2AEdges, recomputeAndPush]);

  useEffect(() => {
    if (!showA2AEdges) {
-      // Clear edges immediately when toggled off
+      // Clear edges + buffer immediately when toggled off
+      bufferRef.current = [];
      setA2AEdges([]);
      return;
    }
+    void bootstrap();
+  }, [showA2AEdges, bootstrap, setA2AEdges]);

-    // Initial fetch, then poll every 60 s
-    void fetchAndUpdate();
-    const timer = setInterval(() => void fetchAndUpdate(), A2A_POLL_MS);
-    return () => clearInterval(timer);
-  }, [showA2AEdges, fetchAndUpdate, setA2AEdges]);
+  // Live-update path. Filters server-side ACTIVITY_LOGGED events down
+  // to delegation initiations from visible workspaces and appends each
+  // into the rolling buffer, re-deriving edges via buildA2AEdges.
+  //
+  // Only `method === "delegate"` rows count — the same filter
+  // buildA2AEdges applies — so delegate_result rows arriving over the
+  // wire don't double-count.
+  useSocketEvent((msg) => {
+    if (!showA2AEdges) return;
+    if (msg.event !== "ACTIVITY_LOGGED") return;
+
+    const p = (msg.payload || {}) as Record<string, unknown>;
+    if (p.activity_type !== "delegation") return;
+    if (p.method !== "delegate") return;
+
+    const wsId = msg.workspace_id;
+    if (!visibleIdsRef.current.includes(wsId)) return;
+
+    // Synthesise an ActivityEntry from the WS payload so buildA2AEdges
+    // (which the bootstrap path also feeds) handles it identically.
+    const entry: ActivityEntry = {
+      id:
+        (p.id as string) ||
+        `ws-push-${msg.timestamp || Date.now()}-${wsId}`,
+      workspace_id: wsId,
+      activity_type: "delegation",
+      source_id: (p.source_id as string | null) ?? null,
+      target_id: (p.target_id as string | null) ?? null,
+      method: "delegate",
+      summary: (p.summary as string | null) ?? null,
+      request_body: null,
+      response_body: null,
+      duration_ms: (p.duration_ms as number | null) ?? null,
+      status: (p.status as string) || "ok",
+      error_detail: null,
+      created_at:
+        (p.created_at as string) ||
+        msg.timestamp ||
+        new Date().toISOString(),
+    };
+
+    bufferRef.current = [...bufferRef.current, entry];
+    recomputeAndPush();
+  });

  // Pure side-effect — renders nothing
  return null;
@@ -3,6 +3,7 @@
 import { useState, useEffect, useCallback, useRef } from "react";
 import { useCanvasStore } from "@/store/canvas";
 import { api } from "@/lib/api";
+import { useSocketEvent } from "@/hooks/useSocketEvent";
 import { COMM_TYPE_LABELS } from "@/lib/design-tokens";

 interface Communication {
@@ -18,32 +19,71 @@ interface Communication {
  durationMs: number | null;
 }

+/** Workspace-server `ACTIVITY_LOGGED` payload shape. Pulled out so the
+ *  WS handler below has a typed view of the same fields the HTTP
+ *  bootstrap consumes — drift between the two paths is a class of bug
+ *  AgentCommsPanel hit historically. */
+interface ActivityLoggedPayload {
+  id?: string;
+  activity_type?: string;
+  source_id?: string | null;
+  target_id?: string | null;
+  workspace_id?: string;
+  summary?: string | null;
+  status?: string;
+  duration_ms?: number | null;
+  created_at?: string;
+}
+
+/** Fan-out cap for the bootstrap HTTP fetch on mount / on visibility
+ *  re-open. Kept at 3 (carried over from the 2026-05-04 fix) so a
+ *  freshly-mounted overlay on a 15-workspace tenant only spends 3
+ *  round-trips bootstrapping. Live updates after that arrive via the
+ *  WS subscription below — no polling, no fan-out to maintain. */
+const BOOTSTRAP_FAN_OUT_CAP = 3;
+
+/** Cap on the rendered list. Bootstrap + every WS push prepends, the
+ *  list is sliced to this size after each update. Mirrors the prior
+ *  polling-loop behaviour. */
+const COMMS_RENDER_CAP = 20;
+
 /**
 * Overlay showing recent A2A communications between workspaces.
- * Renders as a floating log panel that auto-updates.
+ *
+ * Update shape (issue #61 Stage 1, replaces the 30s polling loop):
+ *  - On mount (when visible): one HTTP bootstrap per online workspace,
+ *    capped at BOOTSTRAP_FAN_OUT_CAP. Yields the initial recent-comms
+ *    window without waiting for live events.
+ *  - Steady state: subscribes to ACTIVITY_LOGGED via useSocketEvent.
+ *    Each event with a matching activity_type from a visible online
+ *    workspace gets synthesised into a Communication and prepended.
+ *  - Visibility re-open: re-bootstraps so the user sees the freshest
+ *    window even if WS was idle while collapsed.
+ *
+ * No interval poll. The singleton ReconnectingSocket in `store/socket.ts`
+ * already owns reconnect/backoff/health-check, and `useSocketEvent`
+ * inherits those guarantees. If WS is genuinely unhealthy, the overlay
+ * shows the bootstrap snapshot until the next visibility re-open or
+ * the next WS reconnect (which fires its own rehydrate burst).
 */
 export function CommunicationOverlay() {
  const [comms, setComms] = useState<Communication[]>([]);
  const [visible, setVisible] = useState(true);
  const selectedNodeId = useCanvasStore((s) => s.selectedNodeId);
  const nodes = useCanvasStore((s) => s.nodes);
+  // nodesRef gives the WS handler current node-name resolution without
+  // re-subscribing on every node-list change. The bus listener is
+  // registered exactly once per mount; subscriber-side filtering reads
+  // the latest value via this ref.
  const nodesRef = useRef(nodes);
  nodesRef.current = nodes;

-  const fetchComms = useCallback(async () => {
+  const bootstrapComms = useCallback(async () => {
    try {
-      // Fan-out cap: each polled workspace = 1 round-trip. The platform
-      // rate limits at 600 req/min/IP; combined with heartbeats + other
-      // canvas polling, every workspace polled here costs ~6 req/min
-      // (1 every 30s × 1 per workspace). Capping at 3 keeps this
-      // overlay's footprint at 18 req/min worst case — well under
-      // budget even with 8+ workspaces visible. Caught 2026-05-04 when
-      // a user with 8+ workspaces (Design Director + 6 sub-agents +
-      // 3 standalones) saw sustained 429s in canvas console.
      const onlineNodes = nodesRef.current.filter((n) => n.data.status === "online");
      const allComms: Communication[] = [];

-      for (const node of onlineNodes.slice(0, 3)) {
+      for (const node of onlineNodes.slice(0, BOOTSTRAP_FAN_OUT_CAP)) {
        try {
          const activities = await api.get<Array<{
            id: string;
@@ -59,8 +99,8 @@ export function CommunicationOverlay() {

          for (const a of activities) {
            if (a.activity_type === "a2a_send" || a.activity_type === "a2a_receive") {
-              const sourceNode = nodes.find((n) => n.id === (a.source_id || a.workspace_id));
-              const targetNode = nodes.find((n) => n.id === (a.target_id || ""));
+              const sourceNode = nodesRef.current.find((n) => n.id === (a.source_id || a.workspace_id));
+              const targetNode = nodesRef.current.find((n) => n.id === (a.target_id || ""));
              allComms.push({
                id: a.id,
                sourceId: a.source_id || a.workspace_id,
@@ -76,11 +116,12 @@ export function CommunicationOverlay() {
            }
          }
        } catch {
-          // Skip workspaces that fail
+          // Per-workspace failures must not blank the panel — the same
+          // robustness the polling version had.
        }
      }

-      // Sort by timestamp, newest first, dedupe
+      // Newest-first with id-dedup, capped at COMMS_RENDER_CAP.
      const seen = new Set<string>();
      const sorted = allComms
        .sort((a, b) => b.timestamp.localeCompare(a.timestamp))
@@ -89,29 +130,78 @@ export function CommunicationOverlay() {
          seen.add(c.id);
          return true;
        })
-        .slice(0, 20);
+        .slice(0, COMMS_RENDER_CAP);

      setComms(sorted);
    } catch {
-      // Silently handle API errors
+      // Bootstrap failure is non-blocking — the WS subscription below
+      // will populate the panel as live events arrive.
    }
  }, []);

+  // Bootstrap once on mount + every time the user re-opens after a
+  // collapse. Closed-panel state intentionally drops live updates so
+  // the panel doesn't churn invisible state — the next open reloads.
  useEffect(() => {
-    // Gate polling on visibility — when the user collapses the overlay
-    // the data isn't being read, so the per-workspace fan-out becomes
-    // pure rate-limit overhead. Pre-fix this overlay polled regardless
-    // of whether the panel was shown, costing ~36 req/min from a
-    // hidden surface.
    if (!visible) return;
-    fetchComms();
-    // 30s cadence (was 10s). At 3-workspace fan-out that's 6 req/min
-    // worst case from this overlay. Combined with heartbeats (~30/min)
-    // and other canvas polling, leaves ample headroom under the 600/
-    // min/IP server-side rate limit even at 8+ workspace tenants.
-    const interval = setInterval(fetchComms, 30000);
-    return () => clearInterval(interval);
-  }, [fetchComms, visible]);
+    bootstrapComms();
+  }, [bootstrapComms, visible]);
+
+  // Live-update path. Filters server-side ACTIVITY_LOGGED events down
+  // to the comm-overlay-relevant subset and prepends each into the
+  // rendered list with the same dedup the bootstrap path uses.
+  //
+  // Scope guard: ignore events for workspaces not in the visible online
+  // set, so a user collapsing one workspace doesn't see its comms
+  // continue to scroll in. Same shape the bootstrap path applies.
+  useSocketEvent((msg) => {
+    if (!visible) return;
+    if (msg.event !== "ACTIVITY_LOGGED") return;
+
+    const p = (msg.payload || {}) as ActivityLoggedPayload;
+    const type = p.activity_type;
+    if (type !== "a2a_send" && type !== "a2a_receive" && type !== "task_update") return;
+
+    const wsId = msg.workspace_id;
+    const onlineSet = new Set(
+      nodesRef.current.filter((n) => n.data.status === "online").map((n) => n.id),
+    );
+    if (!onlineSet.has(wsId)) return;
+
+    const sourceId = p.source_id || wsId;
+    const targetId = p.target_id || "";
+    const sourceNode = nodesRef.current.find((n) => n.id === sourceId);
+    const targetNode = nodesRef.current.find((n) => n.id === targetId);
+
+    const incoming: Communication = {
+      id: p.id || `${msg.timestamp || Date.now()}:${sourceId}:${targetId}`,
+      sourceId,
+      targetId,
+      sourceName: sourceNode?.data.name || "Unknown",
+      targetName: targetNode?.data.name || "Unknown",
+      type: type as Communication["type"],
+      summary: p.summary || "",
+      status: p.status || "ok",
+      timestamp: p.created_at || msg.timestamp || new Date().toISOString(),
+      durationMs: p.duration_ms ?? null,
+    };
+
+    setComms((prev) => {
+      // Prepend, dedup by id, re-cap. Functional setState is necessary
+      // because two ACTIVITY_LOGGED events arriving in the same React
+      // batch would otherwise read a stale `comms` from the closure.
+      const seen = new Set<string>();
+      const merged = [incoming, ...prev]
+        .sort((a, b) => b.timestamp.localeCompare(a.timestamp))
+        .filter((c) => {
+          if (seen.has(c.id)) return false;
+          seen.add(c.id);
+          return true;
+        })
+        .slice(0, COMMS_RENDER_CAP);
+      return merged;
+    });
+  });

  if (!visible || comms.length === 0) {
    return (
@@ -215,16 +215,6 @@ export function ContextMenu() {
    closeContextMenu();
  }, [contextMenu, selectNode, setPanelTab, closeContextMenu]);

-  const handleExpand = useCallback(async () => {
-    if (!contextMenu) return;
-    try {
-      await api.post(`/workspaces/${contextMenu.nodeId}/expand`, {});
-    } catch (e) {
-      showToast("Expand failed", "error");
-    }
-    closeContextMenu();
-  }, [contextMenu, closeContextMenu]);
-
  const setCollapsed = useCanvasStore((s) => s.setCollapsed);
  const handleCollapse = useCallback(async () => {
    if (!contextMenu) return;
@@ -295,7 +285,7 @@ export function ContextMenu() {
          },
          { label: "Zoom to Team", icon: "⊕", action: handleZoomToTeam },
        ]
-      : [{ label: "Expand to Team", icon: "▷", action: handleExpand }]),
+      : []),
    { label: "", icon: "", action: () => {}, divider: true },
    ...(isPaused
      ? [{ label: "Resume", icon: "▶", action: handleResume }]
@@ -48,16 +48,21 @@ export function EmptyState() {
  });

  // "Create blank" bypasses templates entirely — no preflight, no
-  // modal, just POST /workspaces with a default name and tier.
-  // Deliberately NOT routed through useTemplateDeploy because it
-  // has no `template.id` to deploy against.
+  // modal, just POST /workspaces with a default name. Deliberately
+  // NOT routed through useTemplateDeploy because it has no
+  // `template.id` to deploy against.
+  //
+  // tier is omitted so the backend picks a SaaS-aware default
+  // (T4 on SaaS, T3 on self-hosted — see WorkspaceHandler.DefaultTier).
+  // The previous hardcoded `tier: 2` shipped every fresh-tenant agent
+  // at Standard regardless of host, which surprised SaaS users whose
+  // CreateWorkspaceDialog already defaults to T4.
  const createBlank = async () => {
    setBlankCreating(true);
    setBlankError(null);
    try {
      const ws = await api.post<{ id: string }>("/workspaces", {
        name: "My First Agent",
-        tier: 2,
        canvas: firstDeployCoords(),
      });
      handleDeployed(ws.id);
@@ -20,155 +20,6 @@ import * as Dialog from "@radix-ui/react-dialog";

 type Tab = "python" | "curl" | "claude" | "mcp" | "hermes" | "codex" | "openclaw" | "fields";

-// Per-tab help metadata: docs link, where-to-install link, common errors.
-// All URLs verified against repo content (docs/guides/* file paths map to
-// docs.molecule.ai/docs/guides/*; canonical hostname confirmed by existing
-// blog post canonical metadata) or against the snippet text the operator
-// just copied. Never linking to a URL that wasn't already in product —
-// dead links here defeat the purpose of "more comprehensive instructions."
-const TAB_HELP: Record<
-  Tab,
-  {
-    docsUrl?: string;
-    docsLabel?: string;
-    downloadUrl?: string;
-    downloadLabel?: string;
-    commonIssues?: { symptom: string; check: string }[];
-  }
-> = {
-  mcp: {
-    docsUrl: "https://docs.molecule.ai/docs/guides/mcp-server-setup",
-    docsLabel: "MCP server setup guide",
-    downloadUrl: "https://pypi.org/project/molecule-ai-workspace-runtime/",
-    downloadLabel: "molecule-ai-workspace-runtime on PyPI",
-    commonIssues: [
-      {
-        symptom: "Tools not appearing in your agent",
-        check:
-          "Run `claude mcp list` (or your runtime's equivalent) — the molecule entry should be listed. If missing, re-run the `claude mcp add` line.",
-      },
-      {
-        symptom: "ConnectionRefused / DNS error on first call",
-        check:
-          "PLATFORM_URL must include the scheme (https://) and have no trailing slash. Verify with `curl $PLATFORM_URL/healthz`.",
-      },
-    ],
-  },
-  python: {
-    docsUrl:
-      "https://docs.molecule.ai/docs/guides/external-agent-registration",
-    docsLabel: "External agent registration guide",
-    downloadUrl: "https://pypi.org/project/molecule-ai-workspace-runtime/",
-    downloadLabel: "molecule-ai-workspace-runtime on PyPI",
-    commonIssues: [
-      {
-        symptom: "401 from /heartbeat",
-        check:
-          "AUTH_TOKEN expired or wrong workspace_id. Tokens are shown only once at create time — re-create the workspace to get a fresh token.",
-      },
-      {
-        symptom: "AGENT_URL not reachable from platform",
-        check:
-          "Public HTTPS URL required for inbound A2A. Use ngrok or Cloudflare Tunnel if your agent is behind NAT.",
-      },
-    ],
-  },
-  claude: {
-    docsUrl:
-      "https://docs.molecule.ai/docs/guides/external-agent-registration",
-    docsLabel: "External agent registration guide",
-    downloadUrl: "https://claude.com/claude-code",
-    downloadLabel: "Claude Code (claude.com)",
-    commonIssues: [
-      {
-        symptom: "plugin not installed",
-        check:
-          "Run `/plugin marketplace add Molecule-AI/molecule-mcp-claude-channel` then `/plugin install molecule@molecule-mcp-claude-channel` inside Claude Code, then `/reload-plugins`.",
-      },
-      {
-        symptom: "not on the approved channels allowlist",
-        check:
-          "Custom channels need `--dangerously-load-development-channels` on the launch command. Team/Enterprise orgs need admin to set `channelsEnabled` + `allowedChannelPlugins` in claude.ai admin settings.",
-      },
-      {
-        symptom: "Inbound messages not arriving",
-        check:
-          "Check stderr for `molecule channel: connected — watching N workspace(s)`. Verify ~/.claude/channels/molecule/.env has the right PLATFORM_URL + token.",
-      },
-    ],
-  },
-  hermes: {
-    docsUrl:
-      "https://docs.molecule.ai/docs/guides/external-agent-registration",
-    docsLabel: "External agent registration guide",
-    downloadUrl: "https://github.com/NousResearch/hermes-agent",
-    downloadLabel: "hermes-agent (NousResearch)",
-    commonIssues: [
-      {
-        symptom: "Gateway start failure",
-        check:
-          "Tail ~/.hermes/gateway.log. YAML duplicate-key in config.yaml is the most common cause — `gateway:` block must appear exactly once.",
-      },
-      {
-        symptom: "Plugin not discovered after install",
-        check:
-          "Run `pip show hermes-channel-molecule` to confirm install. Some hermes builds need `hermes plugin reload` before the new platform_plugins entry takes effect.",
-      },
-    ],
-  },
-  codex: {
-    docsUrl: "https://docs.molecule.ai/docs/guides/mcp-server-setup",
-    docsLabel: "MCP server setup guide",
-    downloadUrl: "https://github.com/openai/codex",
-    downloadLabel: "openai/codex",
-    commonIssues: [
-      {
-        symptom: "[mcp_servers.molecule] not loaded",
-        check:
-          "Codex must be ≥ 0.57. Check with `codex --version`; upgrade via `npm install -g @openai/codex@latest`.",
-      },
-      {
-        symptom: "TOML parse error after re-running setup",
-        check:
-          "TOML rejects duplicate `[mcp_servers.molecule]` tables. Open ~/.codex/config.toml and remove the old block before pasting the new one.",
-      },
-    ],
-  },
-  openclaw: {
-    docsUrl: "https://docs.molecule.ai/docs/guides/mcp-server-setup",
-    docsLabel: "MCP server setup guide",
-    commonIssues: [
-      {
-        symptom: "Gateway not starting",
-        check:
-          "Tail ~/.openclaw/gateway.log. The loopback bind requires :18789 to be free — check with `lsof -iTCP:18789`.",
-      },
-      {
-        symptom: "openclaw mcp set rejected",
-        check:
-          "The heredoc generates JSON; verify it parsed by running `jq < ~/.openclaw/mcp/molecule.json`. Re-run `openclaw mcp set` if the file is malformed.",
-      },
-    ],
-  },
-  curl: {
-    docsUrl:
-      "https://docs.molecule.ai/docs/guides/external-agent-registration",
-    docsLabel: "External agent registration guide",
-    commonIssues: [
-      {
-        symptom: "401 / 403 on register",
-        check:
-          "WORKSPACE_AUTH_TOKEN must be the value shown at workspace create. Tokens are shown only once.",
-      },
-    ],
-  },
-  fields: {
-    docsUrl:
-      "https://docs.molecule.ai/docs/guides/external-agent-registration",
-    docsLabel: "External agent registration guide",
-  },
-};
-
 export interface ExternalConnectionInfo {
  workspace_id: string;
  platform_url: string;
@@ -452,7 +303,6 @@ export function ExternalConnectModal({ info, onClose }: Props) {
                <Field label="heartbeat_endpoint" value={info.heartbeat_endpoint} onCopy={() => copy(info.heartbeat_endpoint, "hb")} copied={copiedKey === "hb"} />
              </div>
            )}
-            <HelpBlock help={TAB_HELP[tab]} />
          </div>

          <div className="mt-5 flex justify-end gap-2">
@@ -501,70 +351,6 @@ function SnippetBlock({
  );
 }

-// HelpBlock — collapsible "Need help?" section under each tab's snippet.
-// Renders only the keys present in the per-tab help metadata (no empty
-// sections). Closed by default so the snippet stays the visual focus;
-// operators with a working setup never see this. Uses native <details>
-// for keyboard accessibility (Tab + Enter) without extra ARIA wiring.
-function HelpBlock({
-  help,
-}: {
-  help: (typeof TAB_HELP)[Tab] | undefined;
-}) {
-  if (!help) return null;
-  const { docsUrl, docsLabel, downloadUrl, downloadLabel, commonIssues } = help;
-  if (!docsUrl && !downloadUrl && !commonIssues?.length) return null;
-
-  return (
-    <details className="mt-3 border border-line rounded-lg bg-surface text-xs">
-      <summary className="cursor-pointer select-none px-3 py-2 text-ink-mid hover:text-ink">
-        Need help? — install link, docs, common errors
-      </summary>
-      <div className="px-3 pb-3 pt-1 space-y-2">
-        {downloadUrl && (
-          <div>
-            <span className="text-ink-soft">Where to install: </span>
-            <a
-              href={downloadUrl}
-              target="_blank"
-              rel="noopener noreferrer"
-              className="text-accent underline hover:text-accent-strong"
-            >
-              {downloadLabel || downloadUrl}
-            </a>
-          </div>
-        )}
-        {docsUrl && (
-          <div>
-            <span className="text-ink-soft">Documentation: </span>
-            <a
-              href={docsUrl}
-              target="_blank"
-              rel="noopener noreferrer"
-              className="text-accent underline hover:text-accent-strong"
-            >
-              {docsLabel || docsUrl}
-            </a>
-          </div>
-        )}
-        {commonIssues && commonIssues.length > 0 && (
-          <div>
-            <div className="text-ink-soft mb-1">Common errors:</div>
-            <ul className="space-y-1.5 pl-3">
-              {commonIssues.map((issue, i) => (
-                <li key={i}>
-                  <code className="text-warm font-mono">{issue.symptom}</code>
-                  <span className="text-ink-mid"> — {issue.check}</span>
-                </li>
-              ))}
-            </ul>
-          </div>
-        )}
-      </div>
-    </details>
-  );
-}
-
 function Field({
  label,
  value,
@@ -1,29 +1,81 @@
 'use client';

-import { useState, useEffect, useCallback } from "react";
-import { api } from "@/lib/api";
-import { ConfirmDialog } from "@/components/ConfirmDialog";
+/**
+ * MemoryInspectorPanel — Memory v2 redesign.
+ *
+ * Reads the canvas Memory tab from the v2 plugin via the
+ * workspace-server proxy at /v2/{namespaces,memories}, replacing the
+ * v1 LOCAL/TEAM/GLOBAL trio that mapped to the deprecated
+ * shared_context model.
+ *
+ * Surface differences from v1:
+ *   - Namespace dropdown driven by GET /v2/namespaces (workspace /
+ *     team / org / custom — labels rendered server-side).
+ *   - Per-row badges for kind (fact|summary|checkpoint), source
+ *     (agent|runtime|user), pin (📌), TTL countdown, and propagation
+ *     source-workspace if the memory came from a peer.
+ *   - No Edit affordance — v2's plugin contract has no PATCH; the
+ *     model is forget + recommit. Delete (Forget) stays.
+ *
+ * Shipping note: when the plugin isn't wired (MEMORY_PLUGIN_URL
+ * unset), every endpoint returns 503 with a clear hint. The panel
+ * surfaces that as a banner so operators know to set the env var,
+ * rather than rendering a perpetual empty state that looks like
+ * "no memories yet".
+ */
+
+import { useCallback, useEffect, useMemo, useState } from 'react';
+import { api } from '@/lib/api';
+import { ConfirmDialog } from '@/components/ConfirmDialog';

 // ── Types ─────────────────────────────────────────────────────────────────────

-/** Memory entry returned by GET /workspaces/:id/memories */
-export interface MemoryEntry {
-  id: string;
-  workspace_id: string;
-  content: string;
-  scope: "LOCAL" | "TEAM" | "GLOBAL";
-  namespace: string;
-  created_at: string;
-  /**
-   * Semantic similarity score (0–1). Only present when the API is queried
-   * with ?q=<query> and the pgvector backend has been deployed.
-   * Absent on plain list fetches — renders gracefully without a badge.
-   */
-  similarity_score?: number;
+export type NamespaceKind = 'workspace' | 'team' | 'org' | 'custom';
+
+export interface NamespaceView {
+  name: string;
+  kind: NamespaceKind;
+  label: string;
 }

-type Scope = "LOCAL" | "TEAM" | "GLOBAL";
-const SCOPES: Scope[] = ["LOCAL", "TEAM", "GLOBAL"];
+export interface NamespacesResponse {
+  readable: NamespaceView[];
+  writable: NamespaceView[];
+}
+
+export type MemoryKind = 'fact' | 'summary' | 'checkpoint';
+export type MemorySource = 'agent' | 'runtime' | 'user';
+
+export interface MemoryV2 {
+  id: string;
+  namespace: string;
+  content: string;
+  kind: MemoryKind;
+  source: MemorySource;
+  pin: boolean;
+  expires_at?: string | null;
+  created_at: string;
+  /** 0..1 plugin similarity score; only present when ?q= is set. */
+  score?: number | null;
+  // Note: an earlier iteration of this type carried a `source_workspace_id`
+  // field rendered as a "from peer" badge. The propagation contract that
+  // would have populated it ("Reserved for future cross-namespace
+  // propagation semantics" in memory-plugin-v1.yaml) is unimplemented —
+  // nothing in the codebase writes that key. Removed in self-review.
+  // Re-add when propagation gains a concrete shape.
+}
+
+interface MemoriesResponse {
+  memories: MemoryV2[];
+}
+
+// MemoryEntry kept as a back-compat type alias so any other component
+// still importing it doesn't break the build. New consumers should
+// prefer MemoryV2 — the v1 shape (LOCAL/TEAM/GLOBAL scope) is gone.
+//
+// `unknown` is used over `any` so TS still flags accidental field
+// access on the legacy shape.
+export type MemoryEntry = MemoryV2;

 interface Props {
  workspaceId: string;
@@ -31,11 +83,26 @@ interface Props {

 // ── Helpers ───────────────────────────────────────────────────────────────────

-/**
- * Sanitise a memory id for use in an HTML id attribute.
- */
 function sanitizeId(id: string): string {
-  return id.replace(/[^a-zA-Z0-9]/g, "-");
+  return id.replace(/[^a-zA-Z0-9]/g, '-');
+}
+
+/**
+ * Detect a memory-plugin-503 error from the api wrapper's stringified
+ * Error message. Matches on the literal env-var name rather than the
+ * status code, because the api shim renders status codes inside a
+ * larger formatted message and a future status-code reformat would
+ * silently break the detection.
+ *
+ * The substring `MEMORY_PLUGIN_URL` is hard-coded in the handler at
+ * `workspace-server/internal/handlers/memories_v2.go:available()`,
+ * so this is a pinned cross-layer contract — drift is caught by both
+ * the Go test (TestMemoriesV2_PluginUnwired_All503) and the canvas
+ * test (TestMemoryInspectorPanel — plugin unavailable).
+ */
+export function isPluginUnavailableError(err: unknown): boolean {
+  const msg = err instanceof Error ? err.message : '';
+  return msg.includes('MEMORY_PLUGIN_URL');
 }

 function formatRelativeTime(iso: string): string {
@@ -46,6 +113,24 @@ function formatRelativeTime(iso: string): string {
  return new Date(iso).toLocaleDateString();
 }

+/**
+ * Render a TTL countdown like "12h", "3d", or "expired" (when the
+ * stored expires_at is in the past). Non-fatal if expires_at is null
+ * or invalid — falls through to empty string so the badge doesn't
+ * render.
+ */
+export function formatTTL(expiresAt: string | null | undefined): string {
+  if (!expiresAt) return '';
+  const ts = new Date(expiresAt).getTime();
+  if (Number.isNaN(ts)) return '';
+  const diff = ts - Date.now();
+  if (diff <= 0) return 'expired';
+  if (diff < 60_000) return `${Math.floor(diff / 1000)}s`;
+  if (diff < 3_600_000) return `${Math.floor(diff / 60_000)}m`;
+  if (diff < 86_400_000) return `${Math.floor(diff / 3_600_000)}h`;
+  return `${Math.floor(diff / 86_400_000)}d`;
+}
+
 // ── Skeleton rows ──────────────────────────────────────────────────────────────

 function MemorySkeletonRows() {
@@ -70,56 +155,92 @@ function MemorySkeletonRows() {

 // ── Component ─────────────────────────────────────────────────────────────────

+const ALL_NAMESPACES = '__all__';
+
 export function MemoryInspectorPanel({ workspaceId }: Props) {
-  const [activeScope, setActiveScope] = useState<Scope>("LOCAL");
-  const [activeNamespace, setActiveNamespace] = useState("");
-  const [entries, setEntries] = useState<MemoryEntry[]>([]);
+  const [namespaces, setNamespaces] = useState<NamespacesResponse | null>(null);
+  const [activeNamespace, setActiveNamespace] = useState<string>(ALL_NAMESPACES);
+  const [entries, setEntries] = useState<MemoryV2[]>([]);
  const [loading, setLoading] = useState(true);
  const [error, setError] = useState<string | null>(null);

-  // ── Search state (debounced) ────────────────────────────────────────────────
-  const [searchQuery, setSearchQuery] = useState("");
-  const [debouncedQuery, setDebouncedQuery] = useState("");
+  // Plugin-disabled banner (503 from server). Stored separately so we
+  // can keep showing the namespace dropdown empty rather than
+  // hiding the whole panel.
+  const [pluginUnavailable, setPluginUnavailable] = useState(false);
+
+  // Search state (debounced)
+  const [searchQuery, setSearchQuery] = useState('');
+  const [debouncedQuery, setDebouncedQuery] = useState('');

  useEffect(() => {
-    const timer = setTimeout(
-      () => setDebouncedQuery(searchQuery.trim()),
-      300
-    );
+    const timer = setTimeout(() => setDebouncedQuery(searchQuery.trim()), 300);
    return () => clearTimeout(timer);
  }, [searchQuery]);

-  // ── Delete state ─────────────────────────────────────────────────────────────
+  // Delete state
  const [pendingDeleteId, setPendingDeleteId] = useState<string | null>(null);

-  // ── Data loading ────────────────────────────────────────────────────────────
+  // ── Namespace loading ──────────────────────────────────────────────────────
+
+  const loadNamespaces = useCallback(async () => {
+    try {
+      const data = await api.get<NamespacesResponse>(
+        `/workspaces/${workspaceId}/v2/namespaces`,
+      );
+      setNamespaces(data);
+      setPluginUnavailable(false);
+    } catch (e) {
+      // Plugin-unavailable (503) indicates MEMORY_PLUGIN_URL isn't set.
+      // Anything else stays as a generic load failure that the
+      // entries-load path will also flag.
+      if (isPluginUnavailableError(e)) {
+        setPluginUnavailable(true);
+      }
+      setNamespaces({ readable: [], writable: [] });
+    }
+  }, [workspaceId]);
+
+  // ── Entries loading ────────────────────────────────────────────────────────

  const loadEntries = useCallback(async () => {
    setLoading(true);
    setError(null);
    try {
      const params = new URLSearchParams();
-      params.set("scope", activeScope);
-      if (debouncedQuery) params.set("q", debouncedQuery);
-      if (activeNamespace) params.set("namespace", activeNamespace);
+      if (activeNamespace !== ALL_NAMESPACES) {
+        params.set('namespace', activeNamespace);
+      }
+      if (debouncedQuery) params.set('q', debouncedQuery);

-      const url = `/workspaces/${workspaceId}/memories?${params.toString()}`;
-      const data = await api.get<MemoryEntry[]>(url);
+      const url = `/workspaces/${workspaceId}/v2/memories?${params.toString()}`;
+      const data = await api.get<MemoriesResponse>(url);

-      // When a semantic query is active, sort by similarity_score descending.
+      // When a semantic query is active and the plugin returns
+      // scores, sort by score descending so the most-relevant hit
+      // sits at the top. Empty score → push to bottom.
      const sorted = debouncedQuery
-        ? [...data].sort(
-            (a, b) => (b.similarity_score ?? 0) - (a.similarity_score ?? 0)
+        ? [...data.memories].sort(
+            (a, b) => (b.score ?? 0) - (a.score ?? 0),
          )
-        : data;
+        : data.memories;
      setEntries(sorted);
    } catch (e) {
-      setError(e instanceof Error ? e.message : "Failed to load memories");
+      if (isPluginUnavailableError(e)) {
+        setPluginUnavailable(true);
+        setError(null); // surfaced via banner, not row error
+      } else {
+        setError(e instanceof Error ? e.message : 'Failed to load memories');
+      }
      setEntries([]);
    } finally {
      setLoading(false);
    }
-  }, [workspaceId, activeScope, debouncedQuery, activeNamespace]);
+  }, [workspaceId, activeNamespace, debouncedQuery]);
+
+  useEffect(() => {
+    loadNamespaces();
+  }, [loadNamespaces]);

  useEffect(() => {
    loadEntries();
@@ -136,16 +257,35 @@ export function MemoryInspectorPanel({ workspaceId }: Props) {
    setEntries((prev) => prev.filter((e) => e.id !== id));

    try {
-      await api.del(`/workspaces/${workspaceId}/memories/${encodeURIComponent(id)}`);
+      await api.del(`/workspaces/${workspaceId}/v2/memories/${encodeURIComponent(id)}`);
    } catch (e) {
-      setError(e instanceof Error ? e.message : "Delete failed — reloading...");
+      // Reload first (which clears any stale error), THEN set the
+      // delete-failure message — otherwise loadEntries' own
+      // `setError(null)` wipes our error before the user sees it.
+      // Caught by the rollback test in MemoryInspectorPanel.test.tsx.
+      const msg = e instanceof Error ? e.message : 'Delete failed — reloading…';
      await loadEntries();
+      setError(msg);
    }
  }, [pendingDeleteId, workspaceId, loadEntries]);

+  // ── Namespace dropdown options ─────────────────────────────────────────────
+
+  const dropdownOptions = useMemo(() => {
+    const opts: Array<{ value: string; label: string; kind?: NamespaceKind }> = [
+      { value: ALL_NAMESPACES, label: 'All namespaces' },
+    ];
+    if (namespaces) {
+      for (const ns of namespaces.readable) {
+        opts.push({ value: ns.name, label: ns.label, kind: ns.kind });
+      }
+    }
+    return opts;
+  }, [namespaces]);
+
  // ── Render ──────────────────────────────────────────────────────────────────

-  if (loading && entries.length === 0 && !error) {
+  if (loading && entries.length === 0 && !error && !pluginUnavailable) {
    return (
      <div className="flex items-center justify-center h-32">
        <span className="text-xs text-ink-soft">Loading memories…</span>
@@ -155,32 +295,43 @@ export function MemoryInspectorPanel({ workspaceId }: Props) {

  return (
    <div className="flex flex-col h-full">
-      {/* Scope tabs */}
-      <div className="px-4 pt-3 pb-2 border-b border-line/40 shrink-0">
-        <div className="flex items-center gap-1">
-          {SCOPES.map((scope) => (
-            <button
-              type="button"
-              key={scope}
-              onClick={() => setActiveScope(scope)}
-              aria-pressed={activeScope === scope}
-              className={[
-                "px-3 py-1 text-[11px] rounded transition-colors",
-                activeScope === scope
-                  ? "bg-accent-strong text-white"
-                  : "bg-surface-card text-ink-mid hover:bg-surface-card hover:text-ink",
-              ].join(" ")}
-            >
-              {scope}
-            </button>
-          ))}
+      {/* Plugin-unavailable banner */}
+      {pluginUnavailable && (
+        <div
+          role="alert"
+          aria-live="polite"
+          className="mx-4 mt-3 px-3 py-2 bg-amber-950/30 border border-amber-800/40 rounded text-xs text-amber-300 shrink-0"
+          data-testid="plugin-unavailable-banner"
+        >
+          Memory plugin not configured. Set <code>MEMORY_PLUGIN_URL</code> on the
+          workspace-server to enable v2 memory.
        </div>
-      </div>
+      )}

-      {/* Search bar + namespace filter */}
+      {/* Namespace dropdown */}
      <div className="px-4 pt-3 pb-2 border-b border-line/40 shrink-0 space-y-2">
+        <div className="flex items-center gap-2">
+          <label htmlFor="namespace-dropdown" className="text-[10px] text-ink-soft shrink-0">
+            Namespace:
+          </label>
+          <select
+            id="namespace-dropdown"
+            value={activeNamespace}
+            onChange={(e) => setActiveNamespace(e.target.value)}
+            aria-label="Filter by namespace"
+            disabled={pluginUnavailable}
+            className="flex-1 bg-surface-sunken border border-line/60 focus:border-accent/60 rounded px-2 py-1 text-[11px] text-ink focus:outline-none transition-colors min-w-0 disabled:opacity-50 disabled:cursor-not-allowed"
+          >
+            {dropdownOptions.map((opt) => (
+              <option key={opt.value} value={opt.value}>
+                {opt.label}
+              </option>
+            ))}
+          </select>
+        </div>
+
+        {/* Search bar */}
        <div className="relative flex items-center">
-          {/* Magnifying glass icon */}
          <svg
            width="12"
            height="12"
@@ -198,14 +349,15 @@ export function MemoryInspectorPanel({ workspaceId }: Props) {
            onChange={(e) => setSearchQuery(e.target.value)}
            placeholder="Semantic search…"
            aria-label="Search memories"
-            className="w-full bg-surface-sunken border border-line/60 focus:border-accent/60 rounded-lg pl-8 pr-7 py-1.5 text-[11px] text-ink placeholder-zinc-600 focus:outline-none transition-colors"
+            disabled={pluginUnavailable}
+            className="w-full bg-surface-sunken border border-line/60 focus:border-accent/60 rounded-lg pl-8 pr-7 py-1.5 text-[11px] text-ink placeholder-zinc-600 focus:outline-none transition-colors disabled:opacity-50 disabled:cursor-not-allowed"
          />
          {searchQuery && (
            <button
              type="button"
              onClick={() => {
-                setSearchQuery("");
-                setDebouncedQuery("");
+                setSearchQuery('');
+                setDebouncedQuery('');
              }}
              aria-label="Clear search"
              className="absolute right-2 text-ink-soft hover:text-ink transition-colors text-sm leading-none"
@@ -214,37 +366,22 @@ export function MemoryInspectorPanel({ workspaceId }: Props) {
            </button>
          )}
        </div>
-
-        {/* Namespace filter */}
-        <div className="flex items-center gap-2">
-          <label htmlFor="namespace-filter" className="text-[10px] text-ink-soft shrink-0">
-            Namespace:
-          </label>
-          <input
-            id="namespace-filter"
-            type="text"
-            value={activeNamespace}
-            onChange={(e) => setActiveNamespace(e.target.value)}
-            placeholder="all namespaces"
-            aria-label="Filter by namespace"
-            className="flex-1 bg-surface-sunken border border-line/60 focus:border-accent/60 rounded px-2 py-1 text-[11px] text-ink placeholder-zinc-600 focus:outline-none transition-colors min-w-0"
-          />
-        </div>
      </div>

      {/* Toolbar */}
      <div className="px-4 py-2.5 border-b border-line/40 flex items-center justify-between shrink-0">
        <span className="text-[11px] text-ink-soft">
          {debouncedQuery
-            ? `${entries.length} result${entries.length !== 1 ? "s" : ""}`
+            ? `${entries.length} result${entries.length !== 1 ? 's' : ''}`
            : entries.length === 1
-            ? "1 memory"
-            : `${entries.length} memories`}
+              ? '1 memory'
+              : `${entries.length} memories`}
        </span>
        <button
          type="button"
          onClick={loadEntries}
-          className="px-2 py-1 text-[11px] bg-surface-card hover:bg-surface-card text-ink-mid rounded transition-colors"
+          disabled={pluginUnavailable}
+          className="px-2 py-1 text-[11px] bg-surface-card hover:bg-surface-card text-ink-mid rounded transition-colors disabled:opacity-50 disabled:cursor-not-allowed"
          aria-label="Refresh memories"
        >
          ↻ Refresh
@@ -267,40 +404,7 @@ export function MemoryInspectorPanel({ workspaceId }: Props) {
        {loading ? (
          <MemorySkeletonRows />
        ) : entries.length === 0 ? (
-          debouncedQuery ? (
-            <div className="flex flex-col items-center justify-center py-16 gap-3 text-center">
-              <span className="text-4xl text-ink-soft" aria-hidden="true">◇</span>
-              <p className="text-sm font-medium text-ink-mid">
-                No memories match your search
-              </p>
-              <p className="text-[11px] text-ink-soft max-w-[200px] leading-relaxed">
-                Try a different query or{" "}
-                <button
-                  type="button"
-                  onClick={() => {
-                    setSearchQuery("");
-                    setDebouncedQuery("");
-                  }}
-                  className="text-accent hover:text-accent underline transition-colors"
-                >
-                  clear the search
-                </button>
-                .
-              </p>
-            </div>
-          ) : (
-            <div className="flex flex-col items-center justify-center py-16 gap-3 text-center">
-              <span className="text-4xl text-ink-soft" aria-hidden="true">◇</span>
-              <p className="text-sm font-medium text-ink-mid">No {activeScope} memories</p>
-              <p className="text-[11px] text-ink-soft max-w-[200px] leading-relaxed">
-                {activeScope === "LOCAL"
-                  ? "This workspace has not written any local memories yet."
-                  : activeScope === "TEAM"
-                  ? "No team memories shared with this workspace yet."
-                  : "No global memories exist yet."}
-              </p>
-            </div>
-          )
+          <EmptyState query={debouncedQuery} pluginUnavailable={pluginUnavailable} />
        ) : (
          <div className="space-y-1.5">
            {entries.map((entry) => (
@@ -317,9 +421,9 @@ export function MemoryInspectorPanel({ workspaceId }: Props) {
      {/* Delete confirmation dialog */}
      <ConfirmDialog
        open={pendingDeleteId !== null}
-        title="Delete memory"
-        message={`Delete this ${activeScope} memory? This cannot be undone.`}
-        confirmLabel="Delete"
+        title="Forget memory"
+        message="Forget this memory? This cannot be undone."
+        confirmLabel="Forget"
        confirmVariant="danger"
        onConfirm={confirmDelete}
        onCancel={() => setPendingDeleteId(null)}
@@ -328,19 +432,86 @@ export function MemoryInspectorPanel({ workspaceId }: Props) {
  );
 }

+// ── Empty state ─────────────────────────────────────────────────────────────
+
+function EmptyState({
+  query,
+  pluginUnavailable,
+}: {
+  query: string;
+  pluginUnavailable: boolean;
+}) {
+  if (pluginUnavailable) {
+    // The banner already explains the problem; the empty rows just
+    // mirror it so the operator sees both signals.
+    return (
+      <div className="flex flex-col items-center justify-center py-16 gap-3 text-center">
+        <span className="text-4xl text-ink-soft" aria-hidden="true">
+          ◇
+        </span>
+        <p className="text-sm font-medium text-ink-mid">Memory plugin disabled</p>
+        <p className="text-[11px] text-ink-soft max-w-[220px] leading-relaxed">
+          See banner above for the operator-side fix.
+        </p>
+      </div>
+    );
+  }
+  if (query) {
+    return (
+      <div className="flex flex-col items-center justify-center py-16 gap-3 text-center">
+        <span className="text-4xl text-ink-soft" aria-hidden="true">
+          ◇
+        </span>
+        <p className="text-sm font-medium text-ink-mid">No memories match your search</p>
+        <p className="text-[11px] text-ink-soft max-w-[200px] leading-relaxed">
+          Try a different query or clear the search.
+        </p>
+      </div>
+    );
+  }
+  return (
+    <div className="flex flex-col items-center justify-center py-16 gap-3 text-center">
+      <span className="text-4xl text-ink-soft" aria-hidden="true">
+        ◇
+      </span>
+      <p className="text-sm font-medium text-ink-mid">No memories yet</p>
+      <p className="text-[11px] text-ink-soft max-w-[220px] leading-relaxed">
+        Agents commit memories via MCP tools (commit_memory, commit_summary). They
+        appear here once written.
+      </p>
+    </div>
+  );
+}
+
 // ── MemoryEntryRow sub-component ──────────────────────────────────────────────

 interface MemoryEntryRowProps {
-  entry: MemoryEntry;
+  entry: MemoryV2;
  onDelete: () => void;
 }

+const KIND_BADGE_CLASS: Record<MemoryKind, string> = {
+  fact: 'bg-surface-card text-ink-mid',
+  summary: 'bg-blue-950 text-accent',
+  checkpoint: 'bg-violet-950 text-violet-400',
+};
+
+const SOURCE_BADGE_CLASS: Record<MemorySource, string> = {
+  agent: 'bg-surface-card text-ink-mid',
+  runtime: 'bg-amber-950 text-amber-300',
+  user: 'bg-emerald-950 text-emerald-400',
+};
+
 function MemoryEntryRow({ entry, onDelete }: MemoryEntryRowProps) {
  const [expanded, setExpanded] = useState(false);
  const bodyId = `mem-body-${sanitizeId(entry.id)}`;
+  const ttl = formatTTL(entry.expires_at);

  return (
-    <div className="rounded-lg border border-line/60 bg-surface-sunken/50 overflow-hidden">
+    <div
+      className="rounded-lg border border-line/60 bg-surface-sunken/50 overflow-hidden"
+      data-testid={`memory-row-${entry.id}`}
+    >
      {/* Header row */}
      <button
        type="button"
@@ -349,52 +520,89 @@ function MemoryEntryRow({ entry, onDelete }: MemoryEntryRowProps) {
        aria-expanded={expanded}
        aria-controls={bodyId}
      >
-        {/* Scope badge */}
+        {/* Kind badge */}
        <span
          className={[
-            "text-[9px] shrink-0 font-mono px-1 py-0.5 rounded",
-            entry.scope === "LOCAL"
-              ? "bg-surface-card text-ink-mid"
-              : entry.scope === "TEAM"
-              ? "bg-blue-950 text-accent"
-              : "bg-violet-950 text-violet-400",
-          ].join(" ")}
-          title={`Scope: ${entry.scope}`}
+            'text-[9px] shrink-0 font-mono px-1 py-0.5 rounded',
+            KIND_BADGE_CLASS[entry.kind] ?? 'bg-surface-card text-ink-mid',
+          ].join(' ')}
+          title={`Kind: ${entry.kind}`}
+          data-testid="kind-badge"
        >
-          {entry.scope[0]}
+          {entry.kind[0].toUpperCase()}
        </span>

+        {/* Source badge */}
+        <span
+          className={[
+            'text-[9px] shrink-0 font-mono px-1 py-0.5 rounded',
+            SOURCE_BADGE_CLASS[entry.source] ?? 'bg-surface-card text-ink-mid',
+          ].join(' ')}
+          title={`Source: ${entry.source}`}
+          data-testid="source-badge"
+        >
+          {entry.source}
+        </span>
+
+        {/* Pin indicator */}
+        {entry.pin && (
+          <span
+            className="text-[9px] shrink-0"
+            title="Pinned"
+            data-testid="pin-badge"
+            aria-label="Pinned"
+          >
+            📌
+          </span>
+        )}
+
        {/* Namespace tag */}
-        <span className="text-[9px] shrink-0 font-mono text-ink-soft truncate max-w-[80px]" title={entry.namespace}>
+        <span
+          className="text-[9px] shrink-0 font-mono text-ink-soft truncate max-w-[100px]"
+          title={entry.namespace}
+        >
          {entry.namespace}
        </span>

        {/* Content preview */}
        <span className="flex-1 min-w-0 text-[10px] font-mono text-ink-mid truncate text-left">
-          {entry.content.length > 60 ? entry.content.slice(0, 60) + "…" : entry.content}
+          {entry.content.length > 60 ? entry.content.slice(0, 60) + '…' : entry.content}
        </span>

-        {/* Similarity badge */}
-        {entry.similarity_score != null && (
+        {/* Score badge (semantic search only) */}
+        {entry.score != null && (
          <span
            className={[
-              "text-[9px] shrink-0 font-mono tabular-nums",
-              entry.similarity_score >= 0.8
-                ? "text-accent"
-                : "text-ink-mid",
-            ].join(" ")}
-            title={`Similarity: ${(entry.similarity_score * 100).toFixed(1)}%`}
-            data-testid="similarity-badge"
+              'text-[9px] shrink-0 font-mono tabular-nums',
+              entry.score >= 0.8 ? 'text-accent' : 'text-ink-mid',
+            ].join(' ')}
+            title={`Similarity: ${(entry.score * 100).toFixed(1)}%`}
+            data-testid="score-badge"
          >
-            {Math.round(entry.similarity_score * 100)}%
+            {Math.round(entry.score * 100)}%
          </span>
        )}

+        {/* TTL countdown */}
+        {ttl && (
+          <span
+            className={[
+              'text-[9px] shrink-0 font-mono',
+              ttl === 'expired' ? 'text-bad' : 'text-amber-400',
+            ].join(' ')}
+            title={`Expires: ${entry.expires_at}`}
+            data-testid="ttl-badge"
+          >
+            ⌛{ttl}
+          </span>
+        )}
+
+
        <span className="text-[9px] text-ink-soft shrink-0">
          {formatRelativeTime(entry.created_at)}
        </span>
        <span className="text-[9px] text-ink-soft shrink-0" aria-hidden="true">
-          {expanded ? "▼" : "▶"}
+          {expanded ? '▼' : '▶'}
        </span>
      </button>

@@ -412,6 +620,7 @@ function MemoryEntryRow({ entry, onDelete }: MemoryEntryRowProps) {
          <div className="flex items-center justify-between gap-2">
            <span className="text-[9px] text-ink-soft">
              Created: {new Date(entry.created_at).toLocaleString()}
+              {entry.expires_at && ` · Expires: ${new Date(entry.expires_at).toLocaleString()}`}
            </span>
            <button
              type="button"
@@ -419,10 +628,10 @@ function MemoryEntryRow({ entry, onDelete }: MemoryEntryRowProps) {
                e.stopPropagation();
                onDelete();
              }}
-              aria-label="Delete memory"
+              aria-label="Forget memory"
              className="text-[10px] px-2 py-0.5 bg-red-950/40 hover:bg-red-900/50 border border-red-900/30 rounded text-bad transition-colors shrink-0"
            >
-              Delete
+              Forget
            </button>
          </div>
        </div>
@@ -0,0 +1,175 @@
+"use client";
+
+/**
+ * PurchaseSuccessModal — demo-only post-purchase confirmation.
+ *
+ * Mounted on the canvas root (`app/page.tsx`). On first paint it inspects
+ * `?purchase_success=1[&item=<name>]` on the current URL. If present, it
+ * renders a centred modal styled after `ConfirmDialog`, schedules a 5s
+ * auto-dismiss, and rewrites the URL via `history.replaceState` to drop
+ * the params so a refresh after dismiss does NOT re-show the modal.
+ *
+ * Mock for the funding demo — there is no real billing surface behind
+ * this. The marketplace "Purchase" button on the landing page redirects
+ * here with the params; this modal is the only thing the user sees of
+ * the "transaction".
+ *
+ * Styling matches the warm-paper @theme tokens (surface-sunken / line /
+ * ink / good) so it tracks light + dark without per-mode overrides.
+ */
+
+import { useEffect, useRef, useState } from "react";
+import { createPortal } from "react-dom";
+
+const AUTO_DISMISS_MS = 5000;
+
+function readPurchaseParams(): { open: boolean; item: string | null } {
+  if (typeof window === "undefined") return { open: false, item: null };
+  const sp = new URLSearchParams(window.location.search);
+  const flag = sp.get("purchase_success");
+  if (flag !== "1" && flag !== "true") return { open: false, item: null };
+  return { open: true, item: sp.get("item") };
+}
+
+function stripPurchaseParams() {
+  if (typeof window === "undefined") return;
+  const url = new URL(window.location.href);
+  url.searchParams.delete("purchase_success");
+  url.searchParams.delete("item");
+  // replaceState (not pushState) so back-button doesn't return to the
+  // pre-strip URL and re-trigger the modal.
+  window.history.replaceState({}, "", url.toString());
+}
+
+export function PurchaseSuccessModal() {
+  const [open, setOpen] = useState(false);
+  const [item, setItem] = useState<string | null>(null);
+  const [mounted, setMounted] = useState(false);
+  const dialogRef = useRef<HTMLDivElement>(null);
+
+  // Read the URL params once on mount. We don't subscribe to navigation —
+  // this modal is a one-shot for the demo redirect, not a persistent
+  // listener.
+  useEffect(() => {
+    setMounted(true);
+    const { open: shouldOpen, item: itemName } = readPurchaseParams();
+    if (shouldOpen) {
+      setOpen(true);
+      setItem(itemName);
+      // Clean the URL immediately so a refresh after the modal is closed
+      // (or even while it's still open) does NOT re-trigger it.
+      stripPurchaseParams();
+    }
+  }, []);
+
+  // Auto-dismiss timer + Escape handler.
+  useEffect(() => {
+    if (!open) return;
+    const t = window.setTimeout(() => setOpen(false), AUTO_DISMISS_MS);
+    const onKey = (e: KeyboardEvent) => {
+      if (e.key === "Escape") setOpen(false);
+    };
+    window.addEventListener("keydown", onKey);
+    // Focus the close button so keyboard users land on it after redirect.
+    const raf = requestAnimationFrame(() => {
+      dialogRef.current?.querySelector<HTMLButtonElement>("button")?.focus();
+    });
+    return () => {
+      window.clearTimeout(t);
+      window.removeEventListener("keydown", onKey);
+      cancelAnimationFrame(raf);
+    };
+  }, [open]);
+
+  if (!open || !mounted) return null;
+
+  const itemLabel = item ? decodeURIComponent(item) : "Your new agent";
+
+  return createPortal(
+    <div
+      className="fixed inset-0 z-[9999] flex items-center justify-center"
+      data-testid="purchase-success-modal"
+    >
+      {/* Backdrop — click closes, matches ConfirmDialog backdrop. */}
+      <div
+        className="absolute inset-0 bg-black/60 backdrop-blur-sm"
+        onClick={() => setOpen(false)}
+        aria-hidden="true"
+      />
+
+      <div
+        ref={dialogRef}
+        role="dialog"
+        aria-modal="true"
+        aria-labelledby="purchase-success-title"
+        className="relative bg-surface-sunken border border-line rounded-xl shadow-2xl shadow-black/50 max-w-[420px] w-full mx-4 overflow-hidden"
+      >
+        <div className="px-6 pt-6 pb-4">
+          <div className="flex items-start gap-4">
+            {/* Success glyph — uses --color-good so it tracks the theme.
+                Inline SVG over an emoji so it stays readable + on-brand
+                in both light and dark. */}
+            <div
+              className="flex h-10 w-10 flex-shrink-0 items-center justify-center rounded-full"
+              style={{
+                background:
+                  "color-mix(in srgb, var(--color-good) 15%, transparent)",
+                color: "var(--color-good)",
+              }}
+            >
+              <svg
+                width="22"
+                height="22"
+                viewBox="0 0 24 24"
+                fill="none"
+                aria-hidden="true"
+              >
+                <circle
+                  cx="12"
+                  cy="12"
+                  r="10"
+                  stroke="currentColor"
+                  strokeWidth="1.5"
+                />
+                <path
+                  d="M7.5 12.5L10.5 15.5L16.5 9.5"
+                  stroke="currentColor"
+                  strokeWidth="1.8"
+                  strokeLinecap="round"
+                  strokeLinejoin="round"
+                />
+              </svg>
+            </div>
+            <div className="flex-1">
+              <h3
+                id="purchase-success-title"
+                className="text-base font-semibold text-ink"
+              >
+                Purchase successful
+              </h3>
+              <p className="mt-1.5 text-[13px] leading-relaxed text-ink-mid">
+                <span className="font-medium text-ink">{itemLabel}</span> has
+                been added to your workspace. Provisioning starts in the
+                background — you can keep working while it spins up.
+              </p>
+            </div>
+          </div>
+        </div>
+
+        <div className="flex items-center justify-between gap-3 px-6 py-3 border-t border-line bg-surface/50">
+          <span className="font-mono text-[10.5px] uppercase tracking-[0.12em] text-ink-soft">
+            auto-dismiss · {AUTO_DISMISS_MS / 1000}s
+          </span>
+          <button
+            type="button"
+            onClick={() => setOpen(false)}
+            className="px-3.5 py-1.5 text-[13px] rounded-lg bg-accent hover:bg-accent-strong text-white transition-colors focus:outline-none focus-visible:ring-2 focus-visible:ring-offset-2 focus-visible:ring-offset-surface-sunken focus-visible:ring-accent/60"
+          >
+            Close
+          </button>
+        </div>
+      </div>
+    </div>,
+    document.body,
+  );
+}
@@ -283,11 +283,11 @@ export function SidePanel() {
        {panelTab === "skills" && <SkillsTab key={selectedNodeId} workspaceId={selectedNodeId} data={node.data} />}
        {panelTab === "activity" && <ActivityTab key={selectedNodeId} workspaceId={selectedNodeId} />}
        {panelTab === "chat" && <ChatTab key={selectedNodeId} workspaceId={selectedNodeId} data={node.data} />}
-        {panelTab === "terminal" && <TerminalTab key={selectedNodeId} workspaceId={selectedNodeId} />}
+        {panelTab === "terminal" && <TerminalTab key={selectedNodeId} workspaceId={selectedNodeId} data={node.data} />}
        {panelTab === "config" && <ConfigTab key={selectedNodeId} workspaceId={selectedNodeId} />}
        {panelTab === "schedule" && <ScheduleTab key={selectedNodeId} workspaceId={selectedNodeId} />}
        {panelTab === "channels" && <ChannelsTab key={selectedNodeId} workspaceId={selectedNodeId} />}
-        {panelTab === "files" && <FilesTab key={selectedNodeId} workspaceId={selectedNodeId} />}
+        {panelTab === "files" && <FilesTab key={selectedNodeId} workspaceId={selectedNodeId} data={node.data} />}
        {panelTab === "memory" && <MemoryInspectorPanel key={selectedNodeId} workspaceId={selectedNodeId} />}
        {panelTab === "traces" && <TracesTab key={selectedNodeId} workspaceId={selectedNodeId} />}
        {panelTab === "events" && <EventsTab key={selectedNodeId} workspaceId={selectedNodeId} />}
@@ -316,7 +316,7 @@ export function Toolbar() {
            <div className="space-y-2">
              <HelpRow shortcut="⌘K" text="Search workspaces and jump straight into Details or Chat." />
              <HelpRow shortcut="Palette" text="Open the template palette to deploy a new workspace." />
-              <HelpRow shortcut="Right-click" text="Use node actions for expand, duplicate, export, restart, or delete." />
+              <HelpRow shortcut="Right-click" text="Use node actions for duplicate, export, restart, or delete." />
              <HelpRow shortcut="Chat" text="If a task is still running, the chat tab resumes that session automatically." />
              <HelpRow shortcut="Config" text="Use the Config tab for skills, model, secrets, and runtime settings." />
              <HelpRow shortcut="Dbl-click / Z" text="Zoom canvas to fit a team node and all its sub-workspaces." />
@@ -3,6 +3,7 @@
 import { useCallback, useMemo } from "react";
 import { Handle, NodeResizer, Position, type NodeProps, type Node } from "@xyflow/react";
 import { useCanvasStore, type WorkspaceNodeData } from "@/store/canvas";
+import { getConfigurationError, getConfigurationStatus } from "@/store/canvas-topology";
 import { showToast } from "@/components/Toaster";
 import { Tooltip } from "@/components/Tooltip";
 import { STATUS_CONFIG, TIER_CONFIG } from "@/lib/design-tokens";
@@ -35,8 +36,28 @@ function EjectIcon(props: React.SVGProps<SVGSVGElement>) {
 }

 export function WorkspaceNode({ id, data }: NodeProps<Node<WorkspaceNodeData>>) {
-  const statusCfg = STATUS_CONFIG[data.status] || STATUS_CONFIG.offline;
+  // Configuration-status overlay (PR #2756 / #467 chain). When the
+  // workspace is reachable but adapter.setup() failed (typically a
+  // missing/rotated LLM credential), the agent_card carries
+  // configuration_status: "not_configured". Surface this as a distinct
+  // tile state so the operator sees a useful error instead of an
+  // ambiguous "online but silent" workspace.
+  //
+  // The override only applies when the underlying status is "online" —
+  // a workspace that's actually offline / failed / provisioning gets
+  // its own treatment. "online + not_configured" is the gap PR #2756
+  // introduced; everything else was already covered.
+  const isMisconfigured =
+    data.status === "online" &&
+    getConfigurationStatus(data.agentCard) === "not_configured";
+  const configurationError = getConfigurationError(data.agentCard);
+  const effectiveStatus = isMisconfigured ? "not_configured" : data.status;
+  const statusCfg = STATUS_CONFIG[effectiveStatus] || STATUS_CONFIG.offline;
  const tierCfg = TIER_CONFIG[data.tier] || { label: `T${data.tier}`, color: "text-ink-mid bg-surface-card border border-line" };
+  const tooltipExtra = isMisconfigured && configurationError
+    ? `Agent not configured: ${configurationError}`
+    : null;
+  void tooltipExtra; // wired in via aria-label below; reserved here for future tooltip surface.
  // Org-deploy context — four derived flags off one store subscription.
  // Drives the shimmer while provisioning, the dimmed/non-draggable
  // treatment on locked descendants, and the Cancel pill on the root.
@@ -75,7 +96,12 @@ export function WorkspaceNode({ id, data }: NodeProps<Node<WorkspaceNodeData>>)
    <div
      role="button"
      tabIndex={0}
-      aria-label={`${data.name} workspace — ${data.status}`}
+      aria-label={
+        isMisconfigured && configurationError
+          ? `${data.name} workspace — agent not configured: ${configurationError}`
+          : `${data.name} workspace — ${data.status}`
+      }
+      title={isMisconfigured && configurationError ? `Agent not configured: ${configurationError}` : undefined}
      aria-pressed={isSelected}
      onClick={(e) => {
        e.stopPropagation();
@@ -283,11 +309,12 @@ export function WorkspaceNode({ id, data }: NodeProps<Node<WorkspaceNodeData>>)

        {/* Bottom row: status / active tasks */}
        <div className="flex items-center justify-between mt-0.5">
-          {data.status !== "online" ? (
+          {effectiveStatus !== "online" ? (
            <div className={`text-[10px] uppercase tracking-widest font-medium ${
-              data.status === "failed" ? "text-bad" :
-              data.status === "degraded" ? "text-warm" :
-              data.status === "provisioning" ? "text-accent" :
+              effectiveStatus === "failed" ? "text-bad" :
+              effectiveStatus === "degraded" ? "text-warm" :
+              effectiveStatus === "not_configured" ? "text-warm" :
+              effectiveStatus === "provisioning" ? "text-accent" :
              "text-ink-mid"
            }`}>
              {statusCfg.label}
@@ -313,6 +340,19 @@ export function WorkspaceNode({ id, data }: NodeProps<Node<WorkspaceNodeData>>)
            {data.lastSampleError}
          </div>
        )}
+
+        {/* Configuration error preview — same visual as the degraded
+         *  error preview but keyed off the agent_card's configuration_status.
+         *  Tells the operator which env var is missing so they can fix it
+         *  without having to dig into the workspace logs. */}
+        {isMisconfigured && configurationError && (
+          <div
+            className="text-[10px] text-warm truncate mt-1 bg-warm/10 px-1.5 py-0.5 rounded border border-warm/40"
+            title={configurationError}
+          >
+            {configurationError}
+          </div>
+        )}
      </div>

      <Handle
@@ -41,6 +41,10 @@ vi.mock("@/store/canvas", () => ({
 // ── Imports (after mocks) ─────────────────────────────────────────────────────

 import { api } from "@/lib/api";
+import {
+  emitSocketEvent,
+  _resetSocketEventListenersForTests,
+} from "@/store/socket-events";
 import {
  buildA2AEdges,
  formatA2ARelativeTime,
@@ -342,6 +346,151 @@ describe("A2ATopologyOverlay component", () => {
    expect(mockGet.mock.calls.length).toBe(callsAfterMount);
  });

+  // ── #61 Stage 2: ACTIVITY_LOGGED subscription tests ────────────────────────
+  //
+  // Pin the post-#61 behaviour: WS push for delegation contributes to
+  // the overlay's edge buffer with NO additional HTTP fetch. Same shape
+  // as Stage 1 (CommunicationOverlay).
+
+  describe("#61 stage 2 — ACTIVITY_LOGGED subscription", () => {
+    beforeEach(() => {
+      _resetSocketEventListenersForTests();
+    });
+    afterEach(() => {
+      _resetSocketEventListenersForTests();
+    });
+
+    function emitDelegation(overrides: {
+      workspaceId?: string;
+      sourceId?: string;
+      targetId?: string;
+      method?: string;
+      activityType?: string;
+    } = {}) {
+      // Use Date.now() (real time, fake-timer-frozen) rather than the
+      // hardcoded NOW constant — buildA2AEdges prunes by Date.now() -
+      // A2A_WINDOW_MS, so a row dated against the wrong epoch silently
+      // falls outside the window and the test fails for a confusing
+      // reason ("edges array empty" vs "filter dropped my row").
+      const realNow = Date.now();
+      emitSocketEvent({
+        event: "ACTIVITY_LOGGED",
+        workspace_id: overrides.workspaceId ?? "ws-a",
+        timestamp: new Date(realNow).toISOString(),
+        payload: {
+          id: `act-${Math.random().toString(36).slice(2)}`,
+          activity_type: overrides.activityType ?? "delegation",
+          method: overrides.method ?? "delegate",
+          source_id: overrides.sourceId ?? "ws-a",
+          target_id: overrides.targetId ?? "ws-b",
+          status: "ok",
+          created_at: new Date(realNow - 30_000).toISOString(),
+        },
+      });
+    }
+
+    it("does NOT poll on a 60s interval after bootstrap (post-#61)", async () => {
+      // eslint-disable-next-line @typescript-eslint/no-explicit-any
+      mockGet.mockResolvedValue([] as any);
+      render(<A2ATopologyOverlay />);
+      await act(async () => { await Promise.resolve(); });
+      const callsAfterBootstrap = mockGet.mock.calls.length;
+      expect(callsAfterBootstrap).toBe(2); // ws-a + ws-b
+
+      // Pre-#61: a 60s clock tick would fire a fresh fan-out (2 more
+      // calls). Post-#61: no interval, no extra calls.
+      await act(async () => {
+        vi.advanceTimersByTime(120_000);
+      });
+      expect(mockGet.mock.calls.length).toBe(callsAfterBootstrap);
+    });
+
+    it("WS push for a delegation event from a visible workspace updates edges with NO HTTP call", async () => {
+      // eslint-disable-next-line @typescript-eslint/no-explicit-any
+      mockGet.mockResolvedValue([] as any);
+      render(<A2ATopologyOverlay />);
+      await act(async () => { await Promise.resolve(); await Promise.resolve(); });
+      mockGet.mockClear();
+      mockStoreState.setA2AEdges.mockClear();
+
+      await act(async () => {
+        emitDelegation({ sourceId: "ws-a", targetId: "ws-b" });
+      });
+
+      // Edges-set called with at least one a2a edge for the new push.
+      const calls = mockStoreState.setA2AEdges.mock.calls;
+      expect(calls.length).toBeGreaterThanOrEqual(1);
+      const lastCall = calls[calls.length - 1][0] as Array<{ id: string }>;
+      expect(lastCall.some((e) => e.id === "a2a-ws-a-ws-b")).toBe(true);
+
+      // Critical: no HTTP fetch fired during the WS path.
+      expect(mockGet).not.toHaveBeenCalled();
+    });
+
+    it("WS push for a non-delegation activity_type is ignored", async () => {
+      // eslint-disable-next-line @typescript-eslint/no-explicit-any
+      mockGet.mockResolvedValue([] as any);
+      render(<A2ATopologyOverlay />);
+      await act(async () => { await Promise.resolve(); });
+      mockStoreState.setA2AEdges.mockClear();
+
+      await act(async () => {
+        emitDelegation({ activityType: "a2a_send" });
+      });
+
+      // setA2AEdges must not be called by the WS handler — the only
+      // setA2AEdges calls in this test came from the initial bootstrap.
+      expect(mockStoreState.setA2AEdges).not.toHaveBeenCalled();
+    });
+
+    it("WS push for a delegate_result row is ignored (mirrors buildA2AEdges filter)", async () => {
+      // eslint-disable-next-line @typescript-eslint/no-explicit-any
+      mockGet.mockResolvedValue([] as any);
+      render(<A2ATopologyOverlay />);
+      await act(async () => { await Promise.resolve(); });
+      mockStoreState.setA2AEdges.mockClear();
+
+      await act(async () => {
+        emitDelegation({ method: "delegate_result" });
+      });
+
+      // delegate_result rows do not contribute to the edge count — they
+      // are completion signals, not initiations.
+      expect(mockStoreState.setA2AEdges).not.toHaveBeenCalled();
+    });
+
+    it("WS push from a hidden workspace is ignored", async () => {
+      // eslint-disable-next-line @typescript-eslint/no-explicit-any
+      mockGet.mockResolvedValue([] as any);
+      render(<A2ATopologyOverlay />);
+      await act(async () => { await Promise.resolve(); });
+      mockStoreState.setA2AEdges.mockClear();
+
+      await act(async () => {
+        emitDelegation({ workspaceId: "ws-hidden" });
+      });
+
+      expect(mockStoreState.setA2AEdges).not.toHaveBeenCalled();
+    });
+
+    it("WS push while showA2AEdges is false is ignored", async () => {
+      mockStoreState.showA2AEdges = false;
+      // eslint-disable-next-line @typescript-eslint/no-explicit-any
+      mockGet.mockResolvedValue([] as any);
+      render(<A2ATopologyOverlay />);
+      // The mount path with showA2AEdges=false calls setA2AEdges([])
+      // once — clear that to isolate the WS path.
+      mockStoreState.setA2AEdges.mockClear();
+
+      await act(async () => {
+        emitDelegation();
+      });
+
+      expect(mockStoreState.setA2AEdges).not.toHaveBeenCalled();
+      expect(mockGet).not.toHaveBeenCalled();
+    });
+  });
+
  it("re-fetches when the visible ID set actually changes", async () => {
    // eslint-disable-next-line @typescript-eslint/no-explicit-any
    mockGet.mockResolvedValue([] as any);
@@ -36,6 +36,10 @@ vi.mock("@/hooks/useWorkspaceName", () => ({
  useWorkspaceName: () => () => "Test WS",
 }));

+import {
+  emitSocketEvent,
+  _resetSocketEventListenersForTests,
+} from "@/store/socket-events";
 import { ActivityTab } from "../tabs/ActivityTab";

 // ── Fixtures ──────────────────────────────────────────────────────────────────
@@ -358,6 +362,191 @@ describe("ActivityTab — refresh button", () => {
  });
 });

+// ── Suite 6.5: ACTIVITY_LOGGED subscription (#61 stage 3) ─────────────────────
+//
+// Pin the post-#61 behaviour: WS push extends the rendered list with NO
+// additional HTTP fetch. The 5s polling loop is gone; live updates
+// arrive over the WebSocket bus.
+
+describe("ActivityTab — #61 stage 3: ACTIVITY_LOGGED subscription", () => {
+  beforeEach(() => {
+    vi.clearAllMocks();
+    mockGet.mockResolvedValue([]);
+    _resetSocketEventListenersForTests();
+  });
+  afterEach(() => {
+    cleanup();
+    _resetSocketEventListenersForTests();
+  });
+
+  function emitActivity(overrides: {
+    workspaceId?: string;
+    activityType?: string;
+    summary?: string;
+    id?: string;
+  } = {}) {
+    const realNow = Date.now();
+    emitSocketEvent({
+      event: "ACTIVITY_LOGGED",
+      workspace_id: overrides.workspaceId ?? "ws-1",
+      timestamp: new Date(realNow).toISOString(),
+      payload: {
+        id: overrides.id ?? `act-${Math.random().toString(36).slice(2)}`,
+        activity_type: overrides.activityType ?? "agent_log",
+        source_id: null,
+        target_id: null,
+        method: null,
+        summary: overrides.summary ?? "live-pushed",
+        status: "ok",
+        created_at: new Date(realNow - 5_000).toISOString(),
+      },
+    });
+  }
+
+  it("WS push for matching workspace prepends to the list with NO HTTP call", async () => {
+    render(<ActivityTab workspaceId="ws-1" />);
+    await waitFor(() => {
+      expect(screen.getByText(/0 activities|no activity/i)).toBeTruthy();
+    });
+    mockGet.mockClear();
+
+    await act(async () => {
+      emitActivity({ summary: "live-row-from-bus" });
+    });
+
+    await waitFor(() => {
+      expect(screen.getByText(/live-row-from-bus/)).toBeTruthy();
+    });
+    expect(mockGet).not.toHaveBeenCalled();
+  });
+
+  it("WS push for a different workspace is ignored", async () => {
+    render(<ActivityTab workspaceId="ws-1" />);
+    await waitFor(() => screen.getByText(/no activity/i));
+
+    await act(async () => {
+      emitActivity({
+        workspaceId: "ws-other",
+        summary: "should-not-render-other-ws",
+      });
+    });
+
+    expect(screen.queryByText(/should-not-render-other-ws/)).toBeNull();
+  });
+
+  it("WS push respects the active filter — non-matching activity_type is ignored", async () => {
+    render(<ActivityTab workspaceId="ws-1" />);
+    await waitFor(() => screen.getByText(/no activity/i));
+
+    // Apply "Tasks" filter.
+    clickButton(/tasks/i);
+    await waitFor(() => {
+      expect(
+        screen.getByRole("button", { name: /tasks/i }).getAttribute("aria-pressed"),
+      ).toBe("true");
+    });
+
+    // Push an a2a_send (does NOT match task_update filter).
+    await act(async () => {
+      emitActivity({
+        activityType: "a2a_send",
+        summary: "should-not-render-filter-mismatch",
+      });
+    });
+
+    expect(
+      screen.queryByText(/should-not-render-filter-mismatch/),
+    ).toBeNull();
+  });
+
+  it("WS push respects the active filter — matching activity_type is rendered", async () => {
+    render(<ActivityTab workspaceId="ws-1" />);
+    await waitFor(() => screen.getByText(/no activity/i));
+
+    clickButton(/tasks/i);
+    await waitFor(() => {
+      expect(
+        screen.getByRole("button", { name: /tasks/i }).getAttribute("aria-pressed"),
+      ).toBe("true");
+    });
+
+    await act(async () => {
+      emitActivity({
+        activityType: "task_update",
+        summary: "task-filter-match",
+      });
+    });
+
+    await waitFor(() => {
+      expect(screen.getByText(/task-filter-match/)).toBeTruthy();
+    });
+  });
+
+  it("WS push while autoRefresh is paused is ignored", async () => {
+    render(<ActivityTab workspaceId="ws-1" />);
+    await waitFor(() => screen.getByText(/no activity/i));
+
+    // Toggle Live → Paused.
+    clickButton(/live/i);
+    await waitFor(() => {
+      expect(screen.getByText(/Paused/)).toBeTruthy();
+    });
+
+    await act(async () => {
+      emitActivity({ summary: "should-not-render-paused" });
+    });
+
+    expect(screen.queryByText(/should-not-render-paused/)).toBeNull();
+  });
+
+  it("WS push for a row already in the list is deduped (no double-render)", async () => {
+    // Bootstrap with one row — same id as the WS push to trigger dedup.
+    mockGet.mockResolvedValueOnce([
+      makeEntry({ id: "shared-id", summary: "bootstrap-summary" }),
+    ]);
+    render(<ActivityTab workspaceId="ws-1" />);
+    await waitFor(() => {
+      expect(screen.getByText(/bootstrap-summary/)).toBeTruthy();
+    });
+    mockGet.mockClear();
+
+    // Push a row with the SAME id but a different summary — must not
+    // render the new summary; original row stays.
+    await act(async () => {
+      emitActivity({
+        id: "shared-id",
+        summary: "should-not-replace-existing",
+      });
+    });
+
+    expect(screen.queryByText(/should-not-replace-existing/)).toBeNull();
+    // Also verify count didn't grow.
+    expect(screen.getByText(/1 activities/)).toBeTruthy();
+  });
+
+  it("does NOT poll on a 5s interval after mount (post-#61)", async () => {
+    vi.useFakeTimers();
+    try {
+      render(<ActivityTab workspaceId="ws-1" />);
+      // Drain the mount-time bootstrap promise.
+      await act(async () => {
+        await Promise.resolve();
+        await Promise.resolve();
+      });
+      const callsAfterBootstrap = mockGet.mock.calls.length;
+      expect(callsAfterBootstrap).toBeGreaterThanOrEqual(1);
+
+      // Pre-#61: a 30s clock advance fires 6 more polls. Post-#61: 0.
+      await act(async () => {
+        vi.advanceTimersByTime(30_000);
+      });
+      expect(mockGet.mock.calls.length).toBe(callsAfterBootstrap);
+    } finally {
+      vi.useRealTimers();
+    }
+  });
+});
+
 // ── Suite 7: Activity count ───────────────────────────────────────────────────

 describe("ActivityTab — activity count", () => {
@@ -1,18 +1,28 @@
 // @vitest-environment jsdom
 /**
- * CommunicationOverlay tests — pin the rate-limit fix shipped 2026-05-04.
+ * CommunicationOverlay tests — pin both the 2026-05-04 fan-out cap fix
+ * AND the 2026-05-07 polling → ACTIVITY_LOGGED-subscriber refactor
+ * (issue #61 stage 1).
 *
- * The overlay polls /workspaces/:id/activity?limit=5 for each online
- * workspace. Pre-fix it (a) polled regardless of visibility and (b)
- * fanned out to 6 workspaces every 10s. With 8+ workspaces a user
- * triggered sustained 429s (server-side rate limit is 600 req/min/IP).
+ * The overlay used to poll /workspaces/:id/activity?limit=5 on a 30s
+ * interval per online workspace (capped at 3). Post-#61: it bootstraps
+ * once on mount via the same HTTP path (cap of 3 retained), then
+ * subscribes to ACTIVITY_LOGGED via the global socket bus for live
+ * updates. No interval poll.
 *
 * These tests pin:
- *  1. Fan-out cap of 3 — even with 6 online nodes, only 3 fetches
- *  2. Visibility gate — when collapsed, no polling
+ *  1. Bootstrap fan-out cap of 3 — even with 6 online nodes, only 3
+ *     HTTP fetches on mount.
+ *  2. Visibility gate — when collapsed, no HTTP fetches; re-open
+ *     re-bootstraps.
+ *  3. NO interval polling — advancing the clock past 30s does not fire
+ *     additional HTTP calls.
+ *  4. WS push extends the rendered list without firing any HTTP call.
+ *  5. WS push for an offline workspace is ignored.
+ *  6. WS push for a non-comm activity_type is ignored.
 *
- * If a future refactor pushes either dial back up, CI fails before
- * the regression hits a paying tenant.
+ * If a future refactor regresses any of these, CI fails before the
+ * regression hits a paying tenant.
 */
 import { describe, it, expect, vi, beforeEach, afterEach } from "vitest";
 import { render, cleanup, act, fireEvent } from "@testing-library/react";
@@ -23,7 +33,7 @@ vi.mock("@/lib/api", () => ({
  api: { get: vi.fn() },
 }));

-// Six online nodes — enough to verify the cap of 3.
+// Six online nodes — enough to verify the bootstrap cap of 3.
 const mockStoreState = {
  selectedNodeId: null as string | null,
  nodes: [
@@ -56,6 +66,10 @@ vi.mock("@/lib/design-tokens", () => ({
 // ── Imports (after mocks) ─────────────────────────────────────────────────────

 import { api } from "@/lib/api";
+import {
+  emitSocketEvent,
+  _resetSocketEventListenersForTests,
+} from "@/store/socket-events";
 import { CommunicationOverlay } from "../CommunicationOverlay";

 const mockGet = vi.mocked(api.get);
@@ -66,30 +80,34 @@ beforeEach(() => {
  vi.useFakeTimers();
  mockGet.mockReset();
  mockGet.mockResolvedValue([]);
+  // Drop any subscribers the previous test left on the singleton bus —
+  // each render adds one via useSocketEvent.
+  _resetSocketEventListenersForTests();
 });

 afterEach(() => {
  cleanup();
  vi.useRealTimers();
+  _resetSocketEventListenersForTests();
 });

 // ── Tests ─────────────────────────────────────────────────────────────────────

-describe("CommunicationOverlay — fan-out cap", () => {
-  it("polls at most 3 of 6 online workspaces (rate-limit floor)", async () => {
+describe("CommunicationOverlay — bootstrap fan-out cap", () => {
+  it("bootstraps at most 3 of 6 online workspaces (rate-limit floor preserved post-#61)", async () => {
    await act(async () => {
      render(<CommunicationOverlay />);
    });
-    // Mount fires the first poll synchronously (no interval tick yet).
-    // Pre-fix: 6 calls. Post-fix: 3.
+    // Mount fires the bootstrap synchronously — pre-#61 this was the
+    // first poll cycle; post-#61 it's the only HTTP fetch (live updates
+    // arrive via WS push). 6 nodes → 3 fetches.
    expect(mockGet).toHaveBeenCalledTimes(3);
-    // Verify the calls are for the FIRST 3 online nodes (slice order).
    expect(mockGet).toHaveBeenCalledWith("/workspaces/ws-1/activity?limit=5");
    expect(mockGet).toHaveBeenCalledWith("/workspaces/ws-2/activity?limit=5");
    expect(mockGet).toHaveBeenCalledWith("/workspaces/ws-3/activity?limit=5");
  });

-  it("never polls offline workspaces", async () => {
+  it("never bootstraps offline workspaces", async () => {
    await act(async () => {
      render(<CommunicationOverlay />);
    });
@@ -99,40 +117,39 @@ describe("CommunicationOverlay — fan-out cap", () => {
  });
 });

-describe("CommunicationOverlay — cadence", () => {
-  it("uses 30s interval cadence (was 10s pre-fix)", async () => {
+describe("CommunicationOverlay — no interval polling (post-#61)", () => {
+  // The pre-#61 implementation re-fetched every 30s per workspace.
+  // Post-#61 the only HTTP path is the bootstrap on mount + on
+  // visibility-toggle. This test pins the absence of any interval
+  // poll: a 60s clock advance must not produce a second round of
+  // fetches.
+  it("does NOT poll on a 30s interval after bootstrap", async () => {
    await act(async () => {
      render(<CommunicationOverlay />);
    });
-    expect(mockGet).toHaveBeenCalledTimes(3); // initial mount poll
+    expect(mockGet).toHaveBeenCalledTimes(3); // initial bootstrap
+    mockGet.mockClear();

-    // Advance 10s — pre-fix this would fire another poll. Post-fix: silent.
+    // Advance 60s — well past any plausible cadence the prior version
+    // could have used.
    await act(async () => {
-      vi.advanceTimersByTime(10_000);
+      vi.advanceTimersByTime(60_000);
    });
-    expect(mockGet).toHaveBeenCalledTimes(3);
-
-    // Advance to 30s — interval fires.
-    await act(async () => {
-      vi.advanceTimersByTime(20_000);
-    });
-    expect(mockGet).toHaveBeenCalledTimes(6); // +3 from second tick
+    expect(mockGet).not.toHaveBeenCalled();
  });
 });

 describe("CommunicationOverlay — visibility gate", () => {
-  // The visibility gate is the dial that drops collapsed-panel polling
-  // to ZERO. The cadence test above can't catch its removal — if a
-  // refactor dropped `if (!visible) return`, the cadence test would
-  // still pass because the effect would still fire every 30s.
+  // The visibility gate now does two things post-#61:
+  //   - while closed, the WS handler short-circuits (no setComms churn)
+  //   - re-opening triggers a fresh bootstrap so the list reflects
+  //     anything that happened while the panel was collapsed
  //
  // Direct probe: render with comms-returning mock so the panel
  // actually renders (close button only exists in the expanded panel,
  // not the collapsed button-state). Click close, advance the clock,
  // assert no further fetches.
-  it("stops polling after the user collapses the panel", async () => {
-    // Mock returns one a2a_send so comms.length > 0 → panel renders →
-    // close button accessible.
+  it("stops fetching while collapsed and re-bootstraps on re-open", async () => {
    mockGet.mockResolvedValue([
      {
        id: "act-1",
@@ -150,29 +167,202 @@ describe("CommunicationOverlay — visibility gate", () => {
    const { getByLabelText } = await act(async () => {
      return render(<CommunicationOverlay />);
    });
-    // Drain pending microtasks (resolves the await in fetchComms) so
-    // setComms lands and the panel renders. Don't advance time — that
-    // would fire the next interval tick and pollute the assertion.
+    // Drain pending microtasks (resolves the await in bootstrap) so
+    // setComms lands and the panel renders. Don't advance time — it's
+    // not load-bearing for the gate test, but matches the pattern used
+    // pre-#61 for stability.
    await act(async () => {
      await Promise.resolve();
      await Promise.resolve();
      await Promise.resolve();
    });
-    // Initial mount polled 3 workspaces.
-    expect(mockGet).toHaveBeenCalledTimes(3);
+    expect(mockGet).toHaveBeenCalledTimes(3); // initial bootstrap
    mockGet.mockClear();

-    // Click the close button. Synchronous getByLabelText avoids
-    // findBy's internal setTimeout (deadlocks under useFakeTimers).
+    // Click close. While closed, no fetches and no WS-driven updates.
+    const closeBtn = getByLabelText("Close communications panel");
+    await act(async () => {
+      fireEvent.click(closeBtn);
+    });
+    await act(async () => {
+      vi.advanceTimersByTime(60_000);
+    });
+    expect(mockGet).not.toHaveBeenCalled();
+
+    // Re-open via the collapsed button. Must trigger a fresh bootstrap.
+    const openBtn = getByLabelText("Show communications panel");
+    await act(async () => {
+      fireEvent.click(openBtn);
+    });
+    await act(async () => {
+      await Promise.resolve();
+      await Promise.resolve();
+    });
+    expect(mockGet).toHaveBeenCalledTimes(3); // re-bootstrap on re-open
+  });
+});
+
+describe("CommunicationOverlay — WS subscription (#61 stage 1 core)", () => {
+  // The load-bearing post-#61 behaviour. Every test in this block must
+  // verify (a) the WS push DID update the rendered comms list, and
+  // (b) NO additional HTTP call was fired — the whole point of the
+  // refactor is to remove the polling-driven HTTP traffic.
+  function emitActivityLogged(overrides: Partial<{
+    workspaceId: string;
+    payload: Record<string, unknown>;
+  }> = {}) {
+    emitSocketEvent({
+      event: "ACTIVITY_LOGGED",
+      workspace_id: overrides.workspaceId ?? "ws-1",
+      timestamp: new Date().toISOString(),
+      payload: {
+        id: `act-${Math.random().toString(36).slice(2)}`,
+        activity_type: "a2a_send",
+        source_id: "ws-1",
+        target_id: "ws-2",
+        summary: "live push",
+        status: "ok",
+        duration_ms: 42,
+        created_at: new Date().toISOString(),
+        ...overrides.payload,
+      },
+    });
+  }
+
+  it("WS push for a comm activity_type extends the rendered list with NO additional HTTP call", async () => {
+    const { container } = await act(async () => {
+      return render(<CommunicationOverlay />);
+    });
+    expect(mockGet).toHaveBeenCalledTimes(3); // bootstrap
+    mockGet.mockClear();
+
+    await act(async () => {
+      emitActivityLogged({ payload: { summary: "hello" } });
+    });
+    await act(async () => {
+      await Promise.resolve();
+    });
+
+    // Two pins:
+    //   1. comms list reflects the live push (look for the summary text)
+    //   2. zero HTTP fetches fired during the WS path
+    expect(container.textContent).toContain("hello");
+    expect(mockGet).not.toHaveBeenCalled();
+  });
+
+  it("WS push for an offline workspace is ignored", async () => {
+    const { container } = await act(async () => {
+      return render(<CommunicationOverlay />);
+    });
+    mockGet.mockClear();
+
+    await act(async () => {
+      emitActivityLogged({
+        workspaceId: "ws-offline",
+        payload: { source_id: "ws-offline", summary: "should-not-render" },
+      });
+    });
+    await act(async () => {
+      await Promise.resolve();
+    });
+
+    expect(container.textContent).not.toContain("should-not-render");
+    expect(mockGet).not.toHaveBeenCalled();
+  });
+
+  it("WS push for a non-comm activity_type is ignored (e.g. delegation)", async () => {
+    const { container } = await act(async () => {
+      return render(<CommunicationOverlay />);
+    });
+    mockGet.mockClear();
+
+    await act(async () => {
+      emitActivityLogged({
+        payload: {
+          activity_type: "delegation",
+          summary: "should-not-render-delegation",
+        },
+      });
+    });
+    await act(async () => {
+      await Promise.resolve();
+    });
+
+    expect(container.textContent).not.toContain("should-not-render-delegation");
+    expect(mockGet).not.toHaveBeenCalled();
+  });
+
+  it("WS push while the panel is collapsed is ignored (no churn on hidden state)", async () => {
+    // Bootstrap with one comm so the panel renders → close button
+    // accessible. Then collapse, emit a WS push, re-open: the rendered
+    // list must come from the re-bootstrap, NOT from the WS-push that
+    // arrived during the closed state. Also: nothing visible while
+    // closed (the collapsed button shows only the count, not summaries).
+    mockGet.mockResolvedValue([
+      {
+        id: "act-bootstrap",
+        workspace_id: "ws-1",
+        activity_type: "a2a_send",
+        source_id: "ws-1",
+        target_id: "ws-2",
+        summary: "bootstrap-summary",
+        status: "ok",
+        duration_ms: 1,
+        created_at: new Date().toISOString(),
+      },
+    ]);
+    const { getByLabelText, container } = await act(async () => {
+      return render(<CommunicationOverlay />);
+    });
+    await act(async () => {
+      await Promise.resolve();
+      await Promise.resolve();
+    });
+
+    // Collapse.
    const closeBtn = getByLabelText("Close communications panel");
    await act(async () => {
      fireEvent.click(closeBtn);
    });

-    // Advance well past the 30s cadence — gate should suppress the tick.
+    // Bootstrap mock returns nothing on the re-open path so we can
+    // distinguish "WS push leaked through the gate" from "re-bootstrap
+    // refilled the list."
+    mockGet.mockReset();
+    mockGet.mockResolvedValue([]);
+
    await act(async () => {
-      vi.advanceTimersByTime(60_000);
+      emitActivityLogged({
+        payload: { summary: "leaked-while-closed" },
+      });
    });
+    await act(async () => {
+      await Promise.resolve();
+    });
+
+    // Closed state: rendered DOM must not show any push-derived text.
+    expect(container.textContent).not.toContain("leaked-while-closed");
+  });
+
+  it("non-ACTIVITY_LOGGED events are ignored (e.g. WORKSPACE_OFFLINE)", async () => {
+    const { container } = await act(async () => {
+      return render(<CommunicationOverlay />);
+    });
+    mockGet.mockClear();
+
+    await act(async () => {
+      emitSocketEvent({
+        event: "WORKSPACE_OFFLINE",
+        workspace_id: "ws-1",
+        timestamp: new Date().toISOString(),
+        payload: { summary: "should-not-render-event" },
+      });
+    });
+    await act(async () => {
+      await Promise.resolve();
+    });
+
+    expect(container.textContent).not.toContain("should-not-render-event");
    expect(mockGet).not.toHaveBeenCalled();
  });
 });
@@ -228,4 +228,38 @@ describe("ContextMenu — keyboard accessibility", () => {
    );
    expect(closeContextMenu).toHaveBeenCalled();
  });
+
+  // The "Expand to Team" right-click action was removed in Phase 2 of
+  // RFC #2857 — every workspace can already have children via the
+  // regular CreateWorkspace flow with parent_id, so a separate
+  // backend bulk-create handler (which was non-idempotent and leaked
+  // EC2s on every duplicate call) was deleted in PR #2856 and the
+  // canvas affordance is gone with it.
+  it("'Expand to Team' menu item is gone (childless workspace)", () => {
+    // Default mockStore.nodes = [] → no children → workspace is childless.
+    render(<ContextMenu />);
+    const items = screen.getAllByRole("menuitem");
+    const labels = items.map((el) => el.textContent?.trim() ?? "");
+    // Literal absence — vitest's toContain uses Object.is/===, so the
+    // earlier `.not.toContain(expect.stringMatching(...))` shape passed
+    // for ANY string array (asymmetric matchers only work with toEqual /
+    // arrayContaining). Pin the production string verbatim.
+    expect(labels.some((l) => l.includes("Expand to Team"))).toBe(false);
+    // Sanity: childless menu still has the regular actions.
+    expect(labels.some((l) => l.includes("Delete"))).toBe(true);
+    expect(labels.some((l) => l.includes("Restart"))).toBe(true);
+  });
+
+  it("'Collapse Team' is still present when the workspace HAS children", () => {
+    // Mark a child belonging to ws-1 so hasChildren() returns true.
+    mockStore.nodes = [{ id: "child-1", data: { parentId: "ws-1" } }];
+    render(<ContextMenu />);
+    const items = screen.getAllByRole("menuitem");
+    const labels = items.map((el) => el.textContent?.trim() ?? "");
+    expect(labels.some((l) => /Collapse Team|Expand Team/.test(l))).toBe(true);
+    expect(labels.some((l) => l.includes("Arrange Children"))).toBe(true);
+    expect(labels.some((l) => l.includes("Zoom to Team"))).toBe(true);
+    // Cleanup for other tests.
+    mockStore.nodes = [];
+  });
 });
@@ -1,16 +1,29 @@
 // @vitest-environment jsdom
 /**
- * MemoryInspectorPanel tests — issue #909
+ * MemoryInspectorPanel — v2 redesign tests.
 *
- * Covers: loading, empty state, scope tabs, namespace filter,
- * entry list, expand, delete flow, optimistic updates, Refresh, semantic search.
+ * Coverage targets every behavior the panel surfaces:
+ *   - Initial load wires GET /v2/namespaces + GET /v2/memories
+ *   - Plugin-unavailable banner (503) renders + disables interactions
+ *   - Generic error renders in the error banner
+ *   - Namespace dropdown populates from /v2/namespaces.readable; "All
+ *     namespaces" is the default
+ *   - Selecting a namespace re-fetches with ?namespace=...
+ *   - Search input debounces + scopes the request to ?q=
+ *   - Search results sort by score descending
+ *   - Empty-state copy differs by query / plugin-state / no-data
+ *   - Per-row badges render (kind / source / pin / TTL / score /
+ *     score) and TTL countdown handles past/future/null
+ *   - Delete (Forget) flow: optimistic removal, confirmation dialog,
+ *     server failure rolls back via reload
+ *   - formatTTL helper covers s/m/h/d/expired/null/invalid branches
 */
-import { describe, it, expect, vi, beforeEach, afterEach } from "vitest";
-import { render, screen, fireEvent, waitFor, cleanup, act } from "@testing-library/react";
+import { describe, it, expect, vi, beforeEach, afterEach } from 'vitest';
+import { render, screen, fireEvent, waitFor, cleanup } from '@testing-library/react';

 // ── Mocks ─────────────────────────────────────────────────────────────────────

-vi.mock("@/lib/api", () => ({
+vi.mock('@/lib/api', () => ({
  api: {
    get: vi.fn(),
    post: vi.fn(),
@@ -18,7 +31,7 @@ vi.mock("@/lib/api", () => ({
  },
 }));

-vi.mock("@/components/ConfirmDialog", () => ({
+vi.mock('@/components/ConfirmDialog', () => ({
  ConfirmDialog: ({
    open,
    title,
@@ -33,435 +46,473 @@ vi.mock("@/components/ConfirmDialog", () => ({
    confirmVariant?: string;
    onConfirm: () => void;
    onCancel: () => void;
-    singleButton?: boolean;
  }) =>
    open ? (
      <div data-testid="confirm-dialog">
        <p data-testid="dialog-title">{title}</p>
        <p data-testid="dialog-message">{message}</p>
-        <button onClick={onConfirm}>Confirm Delete</button>
-        <button onClick={onCancel}>Cancel Delete</button>
+        <button onClick={onConfirm}>Confirm</button>
+        <button onClick={onCancel}>Cancel</button>
      </div>
    ) : null,
 }));

-import { api } from "@/lib/api";
-import { MemoryInspectorPanel } from "../MemoryInspectorPanel";
-
-// ── Typed mock helpers ────────────────────────────────────────────────────────
+import { api } from '@/lib/api';
+import {
+  MemoryInspectorPanel,
+  formatTTL,
+  isPluginUnavailableError,
+  type MemoryV2,
+  type NamespacesResponse,
+} from '../MemoryInspectorPanel';

 const mockGet = vi.mocked(api.get);
 const mockDel = vi.mocked(api.del);

-// ── Sample fixtures ───────────────────────────────────────────────────────────
+// ── Fixtures ──────────────────────────────────────────────────────────────────

-const NOW = "2026-04-17T12:00:00.000Z";
-
-const MEMORY_A: import("../MemoryInspectorPanel").MemoryEntry = {
-  id: "mem-a",
-  workspace_id: "ws-1",
-  content: "Remember to review PRs before merging",
-  scope: "LOCAL",
-  namespace: "general",
-  created_at: NOW,
+const NS_RESPONSE: NamespacesResponse = {
+  readable: [
+    { name: 'workspace:ws-1', kind: 'workspace', label: 'Workspace (ws-1)' },
+    { name: 'team:t-1', kind: 'team', label: 'Team (t-1)' },
+  ],
+  writable: [{ name: 'workspace:ws-1', kind: 'workspace', label: 'Workspace (ws-1)' }],
 };

-const MEMORY_B: import("../MemoryInspectorPanel").MemoryEntry = {
-  id: "mem-b",
-  workspace_id: "ws-1",
-  content: "Team knowledge: deploy happens on Fridays",
-  scope: "TEAM",
-  namespace: "procedures",
-  created_at: NOW,
+const MEM_BASIC: MemoryV2 = {
+  id: 'mem-a',
+  namespace: 'workspace:ws-1',
+  content: 'Remember the standup is at 10am',
+  kind: 'fact',
+  source: 'agent',
+  pin: false,
+  created_at: '2026-04-17T12:00:00.000Z',
 };

-const TWO_MEMORIES = [MEMORY_A, MEMORY_B];
+const MEM_PINNED: MemoryV2 = {
+  id: 'mem-pinned',
+  namespace: 'team:t-1',
+  content: 'Team retro every Friday',
+  kind: 'summary',
+  source: 'user',
+  pin: true,
+  expires_at: new Date(Date.now() + 86_400_000).toISOString(),
+  created_at: '2026-04-17T12:00:00.000Z',
+};
+
+const MEM_RUNTIME_CHECKPOINT: MemoryV2 = {
+  id: 'mem-checkpoint',
+  namespace: 'team:t-1',
+  content: 'Runtime checkpoint',
+  kind: 'checkpoint',
+  source: 'runtime',
+  pin: false,
+  created_at: '2026-04-17T12:00:00.000Z',
+};
+
+const MEM_EXPIRED: MemoryV2 = {
+  id: 'mem-expired',
+  namespace: 'workspace:ws-1',
+  content: 'Stale memory',
+  kind: 'fact',
+  source: 'agent',
+  pin: false,
+  expires_at: new Date(Date.now() - 1000).toISOString(),
+  created_at: '2026-04-17T12:00:00.000Z',
+};

 // ── Setup / teardown ──────────────────────────────────────────────────────────

 beforeEach(() => {
-  vi.clearAllMocks();
+  mockGet.mockReset();
+  mockDel.mockReset();
 });

 afterEach(() => {
  cleanup();
 });

-// ── Helper: flush microtasks + React state updates ─────────────────────────────
-async function flushUpdates(): Promise<void> {
-  await act(async () => {});
+// Helper: stub a basic two-call flow (namespaces + memories).
+function stubFetch(memories: MemoryV2[], namespaces: NamespacesResponse = NS_RESPONSE) {
+  mockGet.mockImplementation(((url: string) => {
+    if (url.includes('/v2/namespaces')) {
+      return Promise.resolve(namespaces);
+    }
+    return Promise.resolve({ memories });
+  }) as typeof api.get);
 }

-// ── Loading & empty state ─────────────────────────────────────────────────────
+// ── isPluginUnavailableError helper ─────────────────────────────────────────

-describe("MemoryInspectorPanel — loading and empty state", () => {
-  it("shows loading indicator before data arrives", () => {
-    // eslint-disable-next-line @typescript-eslint/no-explicit-any
-    mockGet.mockReturnValue(new Promise(() => {}) as any);
-    render(<MemoryInspectorPanel workspaceId="ws-1" />);
-    expect(screen.getByText(/loading memories/i)).toBeTruthy();
-  });
-
-  it("renders empty state when API returns []", async () => {
-    // eslint-disable-next-line @typescript-eslint/no-explicit-any
-    mockGet.mockResolvedValue([] as any);
-    render(<MemoryInspectorPanel workspaceId="ws-1" />);
-    await flushUpdates();
-    expect(screen.getByText("No LOCAL memories")).toBeTruthy();
-  });
-
-  it("fetches from the correct workspace memories endpoint with scope=LOCAL", async () => {
-    // eslint-disable-next-line @typescript-eslint/no-explicit-any
-    mockGet.mockResolvedValue([] as any);
-    render(<MemoryInspectorPanel workspaceId="ws-abc-123" />);
-    await flushUpdates();
-    expect(mockGet).toHaveBeenCalledWith(
-      "/workspaces/ws-abc-123/memories?scope=LOCAL"
-    );
-  });
-
-  it("shows error banner when fetch throws", async () => {
-    mockGet.mockRejectedValue(new Error("Network error"));
-    render(<MemoryInspectorPanel workspaceId="ws-1" />);
-    await flushUpdates();
-    expect(screen.getByText("Network error")).toBeTruthy();
-  });
-});
-
-// ── Scope tabs ────────────────────────────────────────────────────────────────
-
-describe("MemoryInspectorPanel — scope tabs", () => {
-  it("renders LOCAL, TEAM, GLOBAL tabs", async () => {
-    // eslint-disable-next-line @typescript-eslint/no-explicit-any
-    mockGet.mockResolvedValue([] as any);
-    render(<MemoryInspectorPanel workspaceId="ws-1" />);
-    await flushUpdates();
-    expect(screen.getByRole("button", { name: "LOCAL" })).toBeTruthy();
-    expect(screen.getByRole("button", { name: "TEAM" })).toBeTruthy();
-    expect(screen.getByRole("button", { name: "GLOBAL" })).toBeTruthy();
-  });
-
-  it("LOCAL is active by default", async () => {
-    // eslint-disable-next-line @typescript-eslint/no-explicit-any
-    mockGet.mockResolvedValue([] as any);
-    render(<MemoryInspectorPanel workspaceId="ws-1" />);
-    await flushUpdates();
-    expect(screen.getByRole("button", { name: "LOCAL" }).getAttribute("aria-pressed")).toBe("true");
-  });
-
-  it("clicking TEAM tab re-fetches with scope=TEAM", async () => {
-    // eslint-disable-next-line @typescript-eslint/no-explicit-any
-    mockGet.mockResolvedValue([] as any);
-    render(<MemoryInspectorPanel workspaceId="ws-1" />);
-    await flushUpdates();
-
-    mockGet.mockClear();
-    fireEvent.click(screen.getByRole("button", { name: "TEAM" }));
-    await flushUpdates();
-    expect(mockGet).toHaveBeenCalledWith(
-      "/workspaces/ws-1/memories?scope=TEAM"
-    );
-  });
-
-  it("clicking GLOBAL tab re-fetches with scope=GLOBAL", async () => {
-    // eslint-disable-next-line @typescript-eslint/no-explicit-any
-    mockGet.mockResolvedValue([] as any);
-    render(<MemoryInspectorPanel workspaceId="ws-1" />);
-    await flushUpdates();
-
-    mockGet.mockClear();
-    fireEvent.click(screen.getByRole("button", { name: "GLOBAL" }));
-    await flushUpdates();
-    expect(mockGet).toHaveBeenCalledWith(
-      "/workspaces/ws-1/memories?scope=GLOBAL"
-    );
-  });
-
-  it("shows scope-specific empty state when switching tabs", async () => {
-    // eslint-disable-next-line @typescript-eslint/no-explicit-any
-    mockGet.mockResolvedValue([] as any);
-    render(<MemoryInspectorPanel workspaceId="ws-1" />);
-    await flushUpdates();
-
-    fireEvent.click(screen.getByRole("button", { name: "TEAM" }));
-    await flushUpdates();
-    expect(screen.getByText("No TEAM memories")).toBeTruthy();
-  });
-});
-
-// ── Namespace filter ──────────────────────────────────────────────────────────
-
-describe("MemoryInspectorPanel — namespace filter", () => {
-  it("renders namespace filter input", async () => {
-    // eslint-disable-next-line @typescript-eslint/no-explicit-any
-    mockGet.mockResolvedValue([] as any);
-    render(<MemoryInspectorPanel workspaceId="ws-1" />);
-    await flushUpdates();
-    expect(screen.getByLabelText("Filter by namespace")).toBeTruthy();
-  });
-
-  it("includes namespace param in API call when set", async () => {
-    vi.useFakeTimers();
-    try {
-      // eslint-disable-next-line @typescript-eslint/no-explicit-any
-      mockGet.mockResolvedValue([] as any);
-      render(<MemoryInspectorPanel workspaceId="ws-1" />);
-      await flushUpdates();
-
-      mockGet.mockClear();
-      fireEvent.change(screen.getByLabelText("Filter by namespace"), {
-        target: { value: "facts" },
-      });
-      // Advance past the 300ms debounce
-      act(() => { vi.advanceTimersByTime(350); });
-      await flushUpdates();
-
-      expect(mockGet).toHaveBeenCalledWith(
-        "/workspaces/ws-1/memories?scope=LOCAL&namespace=facts"
-      );
-    } finally {
-      vi.useRealTimers();
-    }
-  });
-});
-
-// ── Entry list ───────────────────────────────────────────────────────────────
-
-describe("MemoryInspectorPanel — entry list", () => {
-  beforeEach(() => {
-    // eslint-disable-next-line @typescript-eslint/no-explicit-any
-    mockGet.mockResolvedValue(TWO_MEMORIES as any);
-  });
-
-  it("renders a row for every memory", async () => {
-    render(<MemoryInspectorPanel workspaceId="ws-1" />);
-    await flushUpdates();
-    expect(screen.getByText(/Remember to review PRs before merging/)).toBeTruthy();
-    expect(screen.getByText(/Team knowledge: deploy happens on Fridays/)).toBeTruthy();
-  });
-
-  it("displays memory count in toolbar", async () => {
-    render(<MemoryInspectorPanel workspaceId="ws-1" />);
-    await flushUpdates();
-    expect(screen.getByText("2 memories")).toBeTruthy();
-  });
-
-  it("displays scope badge for each entry", async () => {
-    render(<MemoryInspectorPanel workspaceId="ws-1" />);
-    await flushUpdates();
-    expect(screen.getByTitle("Scope: LOCAL")).toBeTruthy();
-    expect(screen.getByTitle("Scope: TEAM")).toBeTruthy();
-  });
-
-  it("entries are collapsed by default (pre region not visible)", async () => {
-    render(<MemoryInspectorPanel workspaceId="ws-1" />);
-    await flushUpdates();
-    // Expanded region (pre tag) should not exist in DOM yet
-    expect(screen.queryByRole("region")).toBeNull();
-  });
-});
-
-// ── Expand / collapse ─────────────────────────────────────────────────────────
-
-describe("MemoryInspectorPanel — expand/collapse", () => {
-  beforeEach(() => {
-    // eslint-disable-next-line @typescript-eslint/no-explicit-any
-    mockGet.mockResolvedValue(TWO_MEMORIES as any);
-  });
-
-  it("clicking a row header expands it and shows the full content in a pre tag", async () => {
-    render(<MemoryInspectorPanel workspaceId="ws-1" />);
-    await flushUpdates();
-
-    fireEvent.click(
-      screen.getByText(/Remember to review PRs before merging/).closest("button")!
-    );
-    await flushUpdates();
-    // After expand, a region with the full content <pre> should appear
-    expect(screen.getByRole("region")).toBeTruthy();
-  });
-
-  it("clicking the header again collapses the row (pre region removed)", async () => {
-    render(<MemoryInspectorPanel workspaceId="ws-1" />);
-    await flushUpdates();
-
-    const headerBtn = screen
-      .getByText(/Remember to review PRs before merging/)
-      .closest("button")!;
-    fireEvent.click(headerBtn); // expand
-    await flushUpdates();
-    expect(screen.getByRole("region")).toBeTruthy();
-
-    fireEvent.click(headerBtn); // collapse
-    await flushUpdates();
-    // After collapse, the region (pre) is removed from the DOM
-    expect(screen.queryByRole("region")).toBeNull();
-  });
-});
-
-// ── Delete flow ───────────────────────────────────────────────────────────────
-
-describe("MemoryInspectorPanel — delete flow", () => {
-  beforeEach(() => {
-    // eslint-disable-next-line @typescript-eslint/no-explicit-any
-    mockGet.mockResolvedValue(TWO_MEMORIES as any);
-    // eslint-disable-next-line @typescript-eslint/no-explicit-any
-    mockDel.mockResolvedValue({ status: "deleted" } as any);
-  });
-
-  /** Helper: expand memory-A and click its Delete button */
-  async function openDeleteForMemoryA() {
-    render(<MemoryInspectorPanel workspaceId="ws-1" />);
-    await flushUpdates();
-    fireEvent.click(
-      screen.getByText(/Remember to review PRs before merging/).closest("button")!
-    );
-    await flushUpdates();
-    fireEvent.click(screen.getByRole("button", { name: "Delete memory" }));
-    await flushUpdates();
-  }
-
-  it("opens ConfirmDialog when Delete is clicked", async () => {
-    await openDeleteForMemoryA();
-    expect(screen.getByTestId("confirm-dialog")).toBeTruthy();
-    expect(screen.getByTestId("dialog-title").textContent).toBe("Delete memory");
-  });
-
-  it("calls api.del with the correct URL-encoded path on confirm", async () => {
-    await openDeleteForMemoryA();
-    fireEvent.click(screen.getByText("Confirm Delete"));
-    await flushUpdates();
-    expect(mockDel).toHaveBeenCalledWith("/workspaces/ws-1/memories/mem-a");
-  });
-
-  it("removes the entry optimistically after confirm", async () => {
-    await openDeleteForMemoryA();
-    fireEvent.click(screen.getByText("Confirm Delete"));
-    await flushUpdates();
-    expect(screen.queryByText(/Remember to review PRs before merging/)).toBeNull();
-    // Sibling entry unaffected
-    expect(screen.getByText(/Team knowledge: deploy happens on Fridays/)).toBeTruthy();
-  });
-
-  it("closes ConfirmDialog without deleting when Cancel is clicked", async () => {
-    await openDeleteForMemoryA();
-    fireEvent.click(screen.getByText("Cancel Delete"));
-    await flushUpdates();
-    expect(screen.queryByTestId("confirm-dialog")).toBeNull();
-    expect(mockDel).not.toHaveBeenCalled();
-    // Sibling memory entry (MEMORY_B) is still in the list
-    expect(screen.getByText(/Team knowledge: deploy happens on Fridays/)).toBeTruthy();
-  });
-});
-
-// ── Refresh ───────────────────────────────────────────────────────────────────
-
-describe("MemoryInspectorPanel — Refresh button", () => {
-  it("re-fetches entries when Refresh is clicked", async () => {
-    // eslint-disable-next-line @typescript-eslint/no-explicit-any
-    mockGet.mockResolvedValue([] as any);
-    render(<MemoryInspectorPanel workspaceId="ws-1" />);
-    await flushUpdates();
-    expect(screen.getByText("No LOCAL memories")).toBeTruthy();
-
-    expect(mockGet).toHaveBeenCalledTimes(1);
-    fireEvent.click(screen.getByRole("button", { name: "Refresh memories" }));
-    await flushUpdates();
-    expect(mockGet).toHaveBeenCalledTimes(2);
-  });
-});
-
-// ── role=alert a11y ──────────────────────────────────────────────────────────
-
-describe("MemoryInspectorPanel — error elements have role=alert", () => {
-  it("fetch error banner has role='alert'", async () => {
-    mockGet.mockRejectedValue(new Error("Network error"));
-    render(<MemoryInspectorPanel workspaceId="ws-1" />);
-    await flushUpdates();
-    const alert = screen.getByRole("alert");
-    expect(alert).toBeTruthy();
-    expect(alert.textContent).toContain("Network error");
-  });
-});
-
-// ── Semantic search ──────────────────────────────────────────────────────────
-
-describe("MemoryInspectorPanel — semantic search", () => {
-  afterEach(() => {
-    vi.useRealTimers();
-  });
-
-  it("debounces search input by 300ms before calling API", async () => {
-    vi.useFakeTimers();
-    // eslint-disable-next-line @typescript-eslint/no-explicit-any
-    mockGet.mockResolvedValue([] as any);
-    render(<MemoryInspectorPanel workspaceId="ws-1" />);
-    await flushUpdates();
-
-    mockGet.mockClear();
-
-    fireEvent.change(screen.getByLabelText("Search memories"), {
-      target: { value: "deploy" },
-    });
-
-    // 200ms — debounce has NOT fired yet
-    act(() => { vi.advanceTimersByTime(200); });
-    await flushUpdates();
-    expect(mockGet).not.toHaveBeenCalled();
-
-    // 350ms total — debounce fires
-    act(() => { vi.advanceTimersByTime(150); });
-    await flushUpdates();
-
-    expect(mockGet).toHaveBeenCalledWith(
-      "/workspaces/ws-1/memories?scope=LOCAL&q=deploy"
-    );
-  });
-
-  it("renders similarity-badge when entry has similarity_score", async () => {
-    // eslint-disable-next-line @typescript-eslint/no-explicit-any
-    mockGet.mockResolvedValue([{ ...MEMORY_A, similarity_score: 0.87 }] as any);
-    render(<MemoryInspectorPanel workspaceId="ws-1" />);
-    await flushUpdates();
-    const badge = document.querySelector('[data-testid="similarity-badge"]');
-    expect(badge).toBeTruthy();
-    expect(badge?.textContent).toBe("87%");
-  });
-
-  it("does not render similarity-badge when entry has no similarity_score", async () => {
-    // eslint-disable-next-line @typescript-eslint/no-explicit-any
-    mockGet.mockResolvedValue([MEMORY_A] as any);
-    render(<MemoryInspectorPanel workspaceId="ws-1" />);
-    await flushUpdates();
+describe('isPluginUnavailableError', () => {
+  it('matches the literal env var contract from the server handler', () => {
    expect(
-      document.querySelector('[data-testid="similarity-badge"]')
-    ).toBeNull();
+      isPluginUnavailableError(
+        new Error('API GET /workspaces/x/v2/memories: 503 {"error":"memory plugin is not configured (set MEMORY_PLUGIN_URL)"}'),
+      ),
+    ).toBe(true);
  });

-  it("clear button resets query immediately and re-fetches without ?q=", async () => {
-    vi.useFakeTimers();
-    // eslint-disable-next-line @typescript-eslint/no-explicit-any
-    mockGet.mockResolvedValue([] as any);
-    render(<MemoryInspectorPanel workspaceId="ws-1" />);
-    await flushUpdates();
+  it('does not false-match on generic 503 errors that don\'t mention the env var', () => {
+    expect(isPluginUnavailableError(new Error('API GET /foo: 503 something else'))).toBe(false);
+  });

-    fireEvent.change(screen.getByLabelText("Search memories"), {
-      target: { value: "deploy" },
+  it('does not false-match on plain 4xx errors', () => {
+    expect(isPluginUnavailableError(new Error('API GET /foo: 401 unauthorized'))).toBe(false);
+  });
+
+  it('returns false for non-Error inputs', () => {
+    expect(isPluginUnavailableError(null)).toBe(false);
+    expect(isPluginUnavailableError(undefined)).toBe(false);
+    expect(isPluginUnavailableError('a string')).toBe(false);
+    expect(isPluginUnavailableError({ message: 'MEMORY_PLUGIN_URL' })).toBe(false);
+  });
+});
+
+// ── formatTTL helper ─────────────────────────────────────────────────────────
+
+describe('formatTTL', () => {
+  it('returns empty string for null/undefined/empty', () => {
+    expect(formatTTL(null)).toBe('');
+    expect(formatTTL(undefined)).toBe('');
+    expect(formatTTL('')).toBe('');
+  });
+
+  it('returns empty for invalid date strings', () => {
+    expect(formatTTL('not-a-date')).toBe('');
+  });
+
+  it('returns "expired" for past timestamps', () => {
+    const past = new Date(Date.now() - 5000).toISOString();
+    expect(formatTTL(past)).toBe('expired');
+  });
+
+  it('formats <60s as seconds', () => {
+    const future = new Date(Date.now() + 30_000).toISOString();
+    expect(formatTTL(future)).toMatch(/^\d{1,2}s$/);
+  });
+
+  it('formats <60m as minutes', () => {
+    const future = new Date(Date.now() + 30 * 60_000).toISOString();
+    expect(formatTTL(future)).toMatch(/^\d{1,2}m$/);
+  });
+
+  it('formats <24h as hours', () => {
+    const future = new Date(Date.now() + 5 * 3_600_000).toISOString();
+    expect(formatTTL(future)).toMatch(/^\d{1,2}h$/);
+  });
+
+  it('formats >24h as days', () => {
+    const future = new Date(Date.now() + 3 * 86_400_000).toISOString();
+    expect(formatTTL(future)).toMatch(/^\d{1,2}d$/);
+  });
+});
+
+// ── Initial load + dropdown ─────────────────────────────────────────────────
+
+describe('MemoryInspectorPanel — initial load', () => {
+  it('fetches namespaces and memories on mount', async () => {
+    stubFetch([MEM_BASIC]);
+    render(<MemoryInspectorPanel workspaceId="ws-1" />);
+
+    await waitFor(() => {
+      const calls = mockGet.mock.calls.map((c) => c[0]);
+      expect(calls.some((u) => u.includes('/v2/namespaces'))).toBe(true);
+      expect(calls.some((u) => u.includes('/v2/memories'))).toBe(true);
+    });
+  });
+
+  it('renders the row contents from the memories response', async () => {
+    stubFetch([MEM_BASIC]);
+    render(<MemoryInspectorPanel workspaceId="ws-1" />);
+    await waitFor(() => {
+      expect(screen.getByText(/Remember the standup is at 10am/)).toBeTruthy();
+    });
+  });
+
+  it('populates the namespace dropdown with readable entries + "All namespaces"', async () => {
+    stubFetch([]);
+    render(<MemoryInspectorPanel workspaceId="ws-1" />);
+    await waitFor(() => screen.getByLabelText('Filter by namespace'));
+    const select = screen.getByLabelText('Filter by namespace') as HTMLSelectElement;
+    const optionLabels = Array.from(select.options).map((o) => o.textContent ?? '');
+    expect(optionLabels[0]).toContain('All namespaces');
+    expect(optionLabels.join('|')).toContain('Workspace (ws-1)');
+    expect(optionLabels.join('|')).toContain('Team (t-1)');
+  });
+
+  it('selecting a namespace re-fetches with ?namespace=', async () => {
+    stubFetch([MEM_BASIC]);
+    render(<MemoryInspectorPanel workspaceId="ws-1" />);
+    await waitFor(() => screen.getByLabelText('Filter by namespace'));
+
+    const select = screen.getByLabelText('Filter by namespace') as HTMLSelectElement;
+    fireEvent.change(select, { target: { value: 'team:t-1' } });
+
+    await waitFor(() => {
+      const calls = mockGet.mock.calls.map((c) => c[0] as string);
+      expect(calls.some((u) => u.includes('namespace=team%3At-1'))).toBe(true);
+    });
+  });
+});
+
+// ── Plugin unavailable (503) ────────────────────────────────────────────────
+
+describe('MemoryInspectorPanel — plugin unavailable', () => {
+  it('renders the operator-hint banner and disables search input', async () => {
+    mockGet.mockRejectedValue(new Error('HTTP 503: memory plugin is not configured (set MEMORY_PLUGIN_URL)'));
+    render(<MemoryInspectorPanel workspaceId="ws-1" />);
+    await waitFor(() => screen.getByTestId('plugin-unavailable-banner'));
+
+    const searchInput = screen.getByLabelText('Search memories') as HTMLInputElement;
+    expect(searchInput.disabled).toBe(true);
+  });
+
+  it('shows the empty-state explaining plugin disabled', async () => {
+    mockGet.mockRejectedValue(new Error('API GET /workspaces/x/v2/memories: 503 {"error":"memory plugin is not configured (set MEMORY_PLUGIN_URL)"}'));
+    render(<MemoryInspectorPanel workspaceId="ws-1" />);
+    await waitFor(() => screen.getByText(/Memory plugin disabled/i));
+  });
+});
+
+// ── Generic error (non-503) ─────────────────────────────────────────────────
+
+describe('MemoryInspectorPanel — generic errors', () => {
+  it('surfaces a non-503 error in the error banner', async () => {
+    mockGet.mockImplementation(((url: string) => {
+      if (url.includes('/v2/namespaces')) {
+        return Promise.resolve(NS_RESPONSE);
+      }
+      return Promise.reject(new Error('upstream timeout'));
+    }) as typeof api.get);
+
+    render(<MemoryInspectorPanel workspaceId="ws-1" />);
+    await waitFor(() => {
+      // Error banner has role=alert
+      const alerts = screen.getAllByRole('alert');
+      const found = alerts.some((a) => a.textContent?.includes('upstream timeout'));
+      expect(found).toBe(true);
+    });
+  });
+});
+
+// ── Search ──────────────────────────────────────────────────────────────────
+
+describe('MemoryInspectorPanel — search', () => {
+  it('eventually fires query with ?q= after debounce', async () => {
+    stubFetch([MEM_BASIC]);
+    render(<MemoryInspectorPanel workspaceId="ws-1" />);
+    await waitFor(() => screen.getByLabelText('Search memories'));
+
+    fireEvent.change(screen.getByLabelText('Search memories'), {
+      target: { value: 'standup' },
    });

-    act(() => { vi.advanceTimersByTime(350); });
-    await flushUpdates();
-
-    expect(mockGet).toHaveBeenCalledWith(
-      "/workspaces/ws-1/memories?scope=LOCAL&q=deploy"
+    await waitFor(
+      () => {
+        const calls = mockGet.mock.calls.map((c) => c[0] as string);
+        expect(calls.some((u) => u.includes('q=standup'))).toBe(true);
+      },
+      { timeout: 1500 },
    );
-    mockGet.mockClear();
+  });

-    fireEvent.click(screen.getByRole("button", { name: "Clear search" }));
-    await flushUpdates();
+  it('sorts results by score descending when query active', async () => {
+    const lowScore: MemoryV2 = { ...MEM_BASIC, id: 'low', score: 0.2, content: 'low' };
+    const highScore: MemoryV2 = { ...MEM_BASIC, id: 'high', score: 0.95, content: 'high' };
+    // Plugin returns in arbitrary order; component sorts.
+    mockGet.mockImplementation(((url: string) => {
+      if (url.includes('/v2/namespaces')) return Promise.resolve(NS_RESPONSE);
+      return Promise.resolve({ memories: [lowScore, highScore] });
+    }) as typeof api.get);

-    expect(mockGet).toHaveBeenCalledWith(
-      "/workspaces/ws-1/memories?scope=LOCAL"
+    render(<MemoryInspectorPanel workspaceId="ws-1" />);
+    await waitFor(() => screen.getByLabelText('Search memories'));
+    fireEvent.change(screen.getByLabelText('Search memories'), {
+      target: { value: 'something' },
+    });
+
+    await waitFor(
+      () => {
+        const rows = screen.getAllByTestId(/^memory-row-/);
+        // First row should be the high-score one
+        expect(rows[0].getAttribute('data-testid')).toBe('memory-row-high');
+      },
+      { timeout: 1500 },
+    );
+  });
+
+  it('clear-button resets the query', async () => {
+    stubFetch([MEM_BASIC]);
+    render(<MemoryInspectorPanel workspaceId="ws-1" />);
+    await waitFor(() => screen.getByLabelText('Search memories'));
+
+    fireEvent.change(screen.getByLabelText('Search memories'), {
+      target: { value: 'foo' },
+    });
+    fireEvent.click(screen.getByLabelText('Clear search'));
+    expect((screen.getByLabelText('Search memories') as HTMLInputElement).value).toBe('');
+  });
+
+  it('renders no-results empty-state when search has no matches', async () => {
+    stubFetch([]);
+    render(<MemoryInspectorPanel workspaceId="ws-1" />);
+    await waitFor(() => screen.getByLabelText('Search memories'));
+    fireEvent.change(screen.getByLabelText('Search memories'), {
+      target: { value: 'nothing' },
+    });
+    await waitFor(
+      () => {
+        expect(screen.getByText(/No memories match your search/i)).toBeTruthy();
+      },
+      { timeout: 1500 },
    );
  });
 });
+
+// ── Per-row badges ───────────────────────────────────────────────────────────
+
+describe('MemoryInspectorPanel — row badges', () => {
+  it('renders kind, source, pin, TTL badges per shape', async () => {
+    stubFetch([MEM_PINNED, MEM_RUNTIME_CHECKPOINT]);
+    render(<MemoryInspectorPanel workspaceId="ws-1" />);
+
+    await waitFor(() => {
+      // Pinned memory: kind=summary, source=user, pin=true, TTL>0
+      const pinnedRow = screen.getByTestId('memory-row-mem-pinned');
+      expect(pinnedRow.querySelector('[data-testid="kind-badge"]')?.textContent).toBe('S');
+      expect(pinnedRow.querySelector('[data-testid="source-badge"]')?.textContent).toBe('user');
+      expect(pinnedRow.querySelector('[data-testid="pin-badge"]')).toBeTruthy();
+      expect(pinnedRow.querySelector('[data-testid="ttl-badge"]')?.textContent).toMatch(/^⌛\d+[hd]$/);
+
+      // Checkpoint memory: kind=checkpoint, source=runtime, no pin, no TTL
+      const propRow = screen.getByTestId('memory-row-mem-checkpoint');
+      expect(propRow.querySelector('[data-testid="kind-badge"]')?.textContent).toBe('C');
+      expect(propRow.querySelector('[data-testid="source-badge"]')?.textContent).toBe('runtime');
+      expect(propRow.querySelector('[data-testid="pin-badge"]')).toBeNull();
+      expect(propRow.querySelector('[data-testid="ttl-badge"]')).toBeNull();
+    });
+  });
+
+  it('TTL badge shows "expired" for past expires_at', async () => {
+    stubFetch([MEM_EXPIRED]);
+    render(<MemoryInspectorPanel workspaceId="ws-1" />);
+    await waitFor(() => {
+      const row = screen.getByTestId('memory-row-mem-expired');
+      expect(row.querySelector('[data-testid="ttl-badge"]')?.textContent).toBe('⌛expired');
+    });
+  });
+
+  it('expanding a row shows full content + Forget button', async () => {
+    stubFetch([MEM_BASIC]);
+    render(<MemoryInspectorPanel workspaceId="ws-1" />);
+    await waitFor(() => screen.getByTestId('memory-row-mem-a'));
+
+    const row = screen.getByTestId('memory-row-mem-a');
+    const headerButton = row.querySelector('button');
+    expect(headerButton).toBeTruthy();
+    fireEvent.click(headerButton!);
+
+    await waitFor(() => {
+      expect(screen.getByLabelText('Forget memory')).toBeTruthy();
+    });
+  });
+});
+
+// ── Delete (Forget) flow ──────────────────────────────────────────────────────
+
+describe('MemoryInspectorPanel — forget flow', () => {
+  it('opens the confirm dialog on Forget click and removes optimistically on confirm', async () => {
+    stubFetch([MEM_BASIC]);
+    mockDel.mockResolvedValue({ status: 'deleted' });
+    render(<MemoryInspectorPanel workspaceId="ws-1" />);
+
+    // Expand row, click Forget
+    await waitFor(() => screen.getByTestId('memory-row-mem-a'));
+    const row = screen.getByTestId('memory-row-mem-a');
+    fireEvent.click(row.querySelector('button')!);
+    await waitFor(() => screen.getByLabelText('Forget memory'));
+    fireEvent.click(screen.getByLabelText('Forget memory'));
+
+    // Dialog appears with v2-shaped copy (Forget, not Delete)
+    expect(screen.getByTestId('dialog-title').textContent).toBe('Forget memory');
+    fireEvent.click(screen.getByText('Confirm'));
+
+    // Optimistic removal happens immediately
+    await waitFor(() => {
+      expect(screen.queryByTestId('memory-row-mem-a')).toBeNull();
+    });
+    // DELETE called with the right path
+    await waitFor(() => {
+      const delPaths = mockDel.mock.calls.map((c) => c[0] as string);
+      expect(delPaths.some((p) => p.includes('/v2/memories/mem-a'))).toBe(true);
+    });
+  });
+
+  it('cancelling the dialog leaves the row in place', async () => {
+    stubFetch([MEM_BASIC]);
+    render(<MemoryInspectorPanel workspaceId="ws-1" />);
+    await waitFor(() => screen.getByTestId('memory-row-mem-a'));
+
+    fireEvent.click(screen.getByTestId('memory-row-mem-a').querySelector('button')!);
+    await waitFor(() => screen.getByLabelText('Forget memory'));
+    fireEvent.click(screen.getByLabelText('Forget memory'));
+    fireEvent.click(screen.getByText('Cancel'));
+
+    expect(screen.queryByTestId('memory-row-mem-a')).toBeTruthy();
+    expect(mockDel).not.toHaveBeenCalled();
+  });
+
+  it('rolls back on server failure by reloading entries', async () => {
+    stubFetch([MEM_BASIC]);
+    mockDel.mockRejectedValue(new Error('upstream 502'));
+
+    render(<MemoryInspectorPanel workspaceId="ws-1" />);
+    await waitFor(() => screen.getByTestId('memory-row-mem-a'));
+    fireEvent.click(screen.getByTestId('memory-row-mem-a').querySelector('button')!);
+    await waitFor(() => screen.getByLabelText('Forget memory'));
+    fireEvent.click(screen.getByLabelText('Forget memory'));
+    fireEvent.click(screen.getByText('Confirm'));
+
+    // After failure, error banner surfaces + reload re-fetches memories
+    await waitFor(() => {
+      const alerts = screen.getAllByRole('alert');
+      const found = alerts.some((a) => a.textContent?.includes('upstream 502'));
+      expect(found).toBe(true);
+    });
+  });
+});
+
+// ── Empty state when no memories at all ────────────────────────────────────
+
+describe('MemoryInspectorPanel — empty state', () => {
+  it('renders the "no memories yet" empty state when not searching', async () => {
+    stubFetch([]);
+    render(<MemoryInspectorPanel workspaceId="ws-1" />);
+    await waitFor(() => {
+      expect(screen.getByText('No memories yet')).toBeTruthy();
+    });
+  });
+});
+
+// ── Refresh ─────────────────────────────────────────────────────────────────
+
+describe('MemoryInspectorPanel — refresh', () => {
+  it('Refresh button refetches memories', async () => {
+    stubFetch([MEM_BASIC]);
+    render(<MemoryInspectorPanel workspaceId="ws-1" />);
+    await waitFor(() => screen.getByLabelText('Refresh memories'));
+
+    const before = mockGet.mock.calls.filter((c) =>
+      (c[0] as string).includes('/v2/memories'),
+    ).length;
+    fireEvent.click(screen.getByLabelText('Refresh memories'));
+
+    await waitFor(() => {
+      const after = mockGet.mock.calls.filter((c) =>
+        (c[0] as string).includes('/v2/memories'),
+      ).length;
+      expect(after).toBe(before + 1);
+    });
+  });
+});
@@ -1,8 +1,9 @@
 "use client";

-import { useState, useEffect, useCallback } from "react";
+import { useState, useEffect, useCallback, useRef } from "react";
 import { api } from "@/lib/api";
 import { ConversationTraceModal } from "@/components/ConversationTraceModal";
+import { useSocketEvent } from "@/hooks/useSocketEvent";
 import { type ActivityEntry } from "@/types/activity";
 import { useWorkspaceName } from "@/hooks/useWorkspaceName";
 import { inferA2AErrorHint } from "./chat/a2aErrorHint";
@@ -48,6 +49,15 @@ export function ActivityTab({ workspaceId }: Props) {
  const [traceOpen, setTraceOpen] = useState(false);
  const resolveName = useWorkspaceName();

+  // Refs let the WS handler read the latest filter / autoRefresh
+  // selection without re-subscribing on every state change. The bus
+  // listener is registered exactly once per mount via useSocketEvent's
+  // ref-internal pattern; subscriber-side filtering reads from these.
+  const filterRef = useRef(filter);
+  filterRef.current = filter;
+  const autoRefreshRef = useRef(autoRefresh);
+  autoRefreshRef.current = autoRefresh;
+
  const loadActivities = useCallback(async () => {
    try {
      const typeParam = filter !== "all" ? `?type=${filter}` : "";
@@ -66,11 +76,58 @@ export function ActivityTab({ workspaceId }: Props) {
    loadActivities();
  }, [loadActivities]);

-  useEffect(() => {
-    if (!autoRefresh) return;
-    const interval = setInterval(loadActivities, 5000);
-    return () => clearInterval(interval);
-  }, [loadActivities, autoRefresh]);
+  // Live-update path (issue #61 stage 3, replaces the 5s setInterval).
+  // ACTIVITY_LOGGED events from this workspace prepend to the rendered
+  // list — dedup by id so a server-side update + a poll reply don't
+  // double-render the same row.
+  //
+  // Honours the user's autoRefresh toggle: when paused, live updates
+  // are dropped until the user re-enables Live (or hits Refresh, which
+  // re-bootstraps via loadActivities).
+  //
+  // Filter awareness: matches the server-side `?type=<filter>`
+  // semantics so the panel doesn't show rows the user excluded.
+  useSocketEvent((msg) => {
+    if (!autoRefreshRef.current) return;
+    if (msg.event !== "ACTIVITY_LOGGED") return;
+    if (msg.workspace_id !== workspaceId) return;
+
+    const p = (msg.payload || {}) as Record<string, unknown>;
+    const activityType = (p.activity_type as string) || "";
+
+    const f = filterRef.current;
+    if (f !== "all" && activityType !== f) return;
+
+    const entry: ActivityEntry = {
+      id:
+        (p.id as string) ||
+        `ws-push-${msg.timestamp || Date.now()}-${msg.workspace_id}`,
+      workspace_id: msg.workspace_id,
+      activity_type: activityType,
+      source_id: (p.source_id as string | null) ?? null,
+      target_id: (p.target_id as string | null) ?? null,
+      method: (p.method as string | null) ?? null,
+      summary: (p.summary as string | null) ?? null,
+      request_body: (p.request_body as Record<string, unknown> | null) ?? null,
+      response_body:
+        (p.response_body as Record<string, unknown> | null) ?? null,
+      duration_ms: (p.duration_ms as number | null) ?? null,
+      status: (p.status as string) || "ok",
+      error_detail: (p.error_detail as string | null) ?? null,
+      created_at:
+        (p.created_at as string) ||
+        msg.timestamp ||
+        new Date().toISOString(),
+    };
+
+    setActivities((prev) => {
+      // Dedup by id — a row that arrived via the bootstrap fetch and
+      // also fires ACTIVITY_LOGGED from a delayed server-side hook
+      // must render exactly once.
+      if (prev.some((e) => e.id === entry.id)) return prev;
+      return [entry, ...prev];
+    });
+  });

  return (
    <div className="flex flex-col h-full">
@@ -1,18 +1,18 @@
 "use client";

-import { useState, useRef, useEffect, useCallback } from "react";
+import { useState, useRef, useEffect, useCallback, useLayoutEffect } from "react";
 import ReactMarkdown from "react-markdown";
 import remarkGfm from "remark-gfm";
 import { api } from "@/lib/api";
 import { useCanvasStore, type WorkspaceNodeData } from "@/store/canvas";
 import { useSocketEvent } from "@/hooks/useSocketEvent";
 import { type ChatMessage, type ChatAttachment, createMessage, appendMessageDeduped } from "./chat/types";
-import { uploadChatFiles, downloadChatFile } from "./chat/uploads";
-import { AttachmentChip, PendingAttachmentPill } from "./chat/AttachmentViews";
+import { uploadChatFiles, downloadChatFile, isPlatformAttachment } from "./chat/uploads";
+import { PendingAttachmentPill } from "./chat/AttachmentViews";
+import { AttachmentPreview } from "./chat/AttachmentPreview";
 import { extractFilesFromTask } from "./chat/message-parser";
 import { AgentCommsPanel } from "./chat/AgentCommsPanel";
 import { appendActivityLine } from "./chat/activityLog";
-import { activityRowToMessages, type ActivityRowForHydration } from "./chat/historyHydration";
 import { runtimeDisplayName } from "@/lib/runtime-names";
 import { ConfirmDialog } from "@/components/ConfirmDialog";

@@ -49,38 +49,12 @@ interface A2AResponse {
  };
 }

-/** Detect activity-log rows that the workspace's own runtime fired
- *  against itself but were misclassified as canvas-source. The proper
- *  fix is the X-Workspace-ID header from `self_source_headers()` in
- *  workspace/platform_auth.py, which makes the platform record
- *  source_id = workspace_id. But three failure modes still leak a
- *  self-message into "My Chat":
- *
- *    1. Historical rows already in the DB with source_id=NULL.
- *    2. Workspace containers running pre-fix heartbeat.py / main.py
- *       (the fix only takes effect after an image rebuild + redeploy).
- *    3. Future internal triggers added without the helper.
- *
- *  This client-side filter recognises the heartbeat trigger by its
- *  exact prefix — the heartbeat assembles
- *
- *    "Delegation results are ready. Review them and take appropriate
- *     action:\n" + summary_lines + report_instruction
- *
- *  in workspace/heartbeat.py. The prefix is template-fixed so a
- *  string match is reliable. If the heartbeat copy ever changes,
- *  update this constant in the same commit.
- *
- *  This is a backstop, not the primary defence — the X-Workspace-ID
- *  header is. Filtering content is fragile to copy edits, so keep
- *  the list narrow. */
-const INTERNAL_SELF_MESSAGE_PREFIXES = [
-  "Delegation results are ready. Review them and take appropriate action",
-];
-
-function isInternalSelfMessage(text: string): boolean {
-  return INTERNAL_SELF_MESSAGE_PREFIXES.some((p) => text.startsWith(p));
-}
+// Internal-self-message filtering moved server-side in RFC #2945
+// PR-C/D — the platform's /chat-history endpoint applies the
+// IsInternalSelfMessage predicate before returning rows, so the
+// client no longer needs the local backstop on the history path.
+// The proper fix is still X-Workspace-ID header (source_id=workspace_id);
+// the platform-side prefix filter handles the residual cases.

 // extractReplyText pulls the agent's text reply out of an A2A response.
 // Concatenates ALL text parts (joined with "\n") rather than returning
@@ -124,29 +98,68 @@ function extractReplyText(resp: A2AResponse): string {
 // doesn't). Single source of truth for file-part parsing across
 // live chat, activity log replay, and any future consumers.

+/** Initial chat history page size. The newest N messages are rendered
+ *  on first paint; older history is fetched on demand via loadOlder()
+ *  when the user scrolls the top sentinel into view. */
+const INITIAL_HISTORY_LIMIT = 10;
+/** Subsequent older-history batch size. Larger than INITIAL so a long
+ *  scroll-back doesn't fan out into many round-trips. */
+const OLDER_HISTORY_BATCH = 20;
+
 /**
- * Load chat history from the activity_logs database via the platform API.
- * Uses source=canvas to only get user-initiated messages (not agent-to-agent).
+ * Load chat history from the platform's typed /chat-history endpoint.
+ *
+ * Server-side rendering of activity_logs rows into ChatMessage shape
+ * lives in workspace-server/internal/messagestore/postgres_store.go
+ * (RFC #2945 PR-C/D). The server already applies the canvas-source
+ * filter, the internal-self-message predicate, the role decision
+ * (status=error vs agent-error prefix → system), and the v0/v1
+ * file-shape extraction. Canvas just renders what it receives.
+ *
+ * Wire shape (mirrors ChatMessage exactly, no per-row mapping needed):
+ *
+ *   GET /workspaces/:id/chat-history?limit=N&before_ts=T
+ *   200 → {"messages": ChatMessage[], "reached_end": boolean}
+ *
+ * Pagination:
+ *  - Pass `limit` to bound the page size (newest-first from server).
+ *  - Pass `beforeTs` (RFC3339) to fetch rows STRICTLY OLDER than that
+ *    timestamp. Combined with limit, this yields the next-older page
+ *    when scrolling backward through history.
+ *
+ * `reachedEnd` is propagated from the server. The server computes it
+ * by comparing rowCount vs limit so a partial last page is correctly
+ * detected even when the row→bubble fan-out is non-1:1 (each row
+ * produces 1-2 bubbles).
 */
-async function loadMessagesFromDB(workspaceId: string): Promise<{ messages: ChatMessage[]; error: string | null }> {
+async function loadMessagesFromDB(
+  workspaceId: string,
+  limit: number,
+  beforeTs?: string,
+): Promise<{ messages: ChatMessage[]; error: string | null; reachedEnd: boolean }> {
  try {
-    const activities = await api.get<ActivityRowForHydration[]>(
-      `/workspaces/${workspaceId}/activity?type=a2a_receive&source=canvas&limit=50`,
+    const params = new URLSearchParams({ limit: String(limit) });
+    if (beforeTs) params.set("before_ts", beforeTs);
+    const resp = await api.get<{ messages: ChatMessage[]; reached_end: boolean }>(
+      `/workspaces/${workspaceId}/chat-history?${params.toString()}`,
    );

-    const messages: ChatMessage[] = [];
-    // Activities are newest-first, reverse for chronological order.
-    // Per-row mapping lives in chat/historyHydration.ts so it can be
-    // unit-tested without spinning up the full ChatTab component
-    // (regression cover for the timestamp-collapse bug).
-    for (const a of [...activities].reverse()) {
-      messages.push(...activityRowToMessages(a, isInternalSelfMessage));
-    }
-    return { messages, error: null };
+    // Server emits oldest-first within the page (RFC #2945 PR-C-2
+    // post-fix: server reverses row-aware before returning so the
+    // wire is display-ready). Canvas appends/prepends without
+    // reordering — this avoids the pair-flip bug a naive flat
+    // reverse causes when each row produces a (user, agent) pair
+    // with the same timestamp.
+    return {
+      messages: resp.messages ?? [],
+      error: null,
+      reachedEnd: resp.reached_end,
+    };
  } catch (err) {
    return {
      messages: [],
      error: err instanceof Error ? err.message : "Failed to load chat history",
+      reachedEnd: true,
    };
  }
 }
@@ -256,6 +269,60 @@ function MyChatPanel({ workspaceId, data }: Props) {
  const [error, setError] = useState<string | null>(null);
  const [confirmRestart, setConfirmRestart] = useState(false);
  const bottomRef = useRef<HTMLDivElement>(null);
+  // First-mount scroll-to-bottom needs `behavior: "instant"` — long
+  // conversations smooth-animate for ~300ms which any concurrent
+  // re-render can interrupt, leaving the user stuck mid-conversation
+  // when the chat tab opens. Subsequent appends (new agent messages)
+  // keep `smooth` for the visual "landing" feel. Flipped the first
+  // time messages.length goes positive, so a workspace switch (which
+  // remounts ChatTab) gets a fresh instant jump too.
+  const hasInitialScrollRef = useRef(false);
+  // Lazy-load older history on scroll-up.
+  // - containerRef = the scrollable messages viewport
+  // - topRef       = sentinel above the messages list; IO observes it
+  //                  and triggers loadOlder() when it enters view
+  // - hasMore      = false once a fetch returns < limit rows; stops IO
+  // - loadingOlder = drives the "Loading older messages…" UI label
+  // - inflightRef  = synchronous guard against double-entry of loadOlder
+  //                  when the IO callback fires twice in the same
+  //                  microtask (state-based guard would be stale until
+  //                  the next React commit)
+  // - scrollAnchorRef = saves distance-from-bottom before a prepend
+  //                  so the useLayoutEffect below can restore the
+  //                  user's exact viewport position. Without this,
+  //                  prepending older messages would jump the scroll
+  //                  position by the height of the new content.
+  // - oldestMessageRef / hasMoreRef = let the loadOlder closure read
+  //                  the latest values without taking them as deps —
+  //                  every live agent push mutates `messages`, and
+  //                  having loadOlder depend on `messages` would tear
+  //                  down + re-arm the IntersectionObserver on every
+  //                  push. Refs decouple the observer lifecycle from
+  //                  message-list updates.
+  const containerRef = useRef<HTMLDivElement>(null);
+  const topRef = useRef<HTMLDivElement>(null);
+  const [hasMore, setHasMore] = useState(true);
+  const [loadingOlder, setLoadingOlder] = useState(false);
+  const inflightRef = useRef(false);
+  // The scroll anchor includes the first-message id as it was BEFORE
+  // the prepend — see useLayoutEffect below for why. Without this tag,
+  // a live agent push that appends WHILE loadOlder is in flight would
+  // run useLayoutEffect against the append (anchor still set), the
+  // "restore" math would scroll the user to a stale offset, AND the
+  // append's normal scroll-to-bottom would be swallowed.
+  const scrollAnchorRef = useRef<
+    { savedDistanceFromBottom: number; expectFirstIdNotEqual: string | null } | null
+  >(null);
+  const oldestMessageRef = useRef<ChatMessage | null>(null);
+  const hasMoreRef = useRef(true);
+  // Monotonic token bumped on workspace switch + on every loadOlder
+  // entry. Each fetch's .then() captures its own token; if the token
+  // has moved, the resolved messages belong to a stale workspace or a
+  // superseded fetch and we silently drop them. Without this guard, a
+  // workspace switch mid-fetch would have the in-flight promise
+  // resolve into the new workspace's setMessages — the user sees
+  // someone else's history briefly.
+  const fetchTokenRef = useRef(0);
  // Files the user has picked but not yet sent. Cleared on send
  // (upload success) or by the × on each pill.
  const [pendingFiles, setPendingFiles] = useState<File[]>([]);
@@ -294,17 +361,144 @@ function MyChatPanel({ workspaceId, data }: Props) {
    sendInFlightRef.current = false;
  }, []);

-  // Load chat history from database on mount
-  useEffect(() => {
+  // Initial-load fetch — used by the mount effect and the "Retry"
+  // button below. Single source of truth so the two paths can't drift
+  // (e.g. INITIAL_HISTORY_LIMIT bumped in the effect but not the
+  // retry, leading to inconsistent first-paint sizes).
+  const loadInitial = useCallback(() => {
    setLoading(true);
    setLoadError(null);
-    loadMessagesFromDB(workspaceId).then(({ messages: msgs, error: fetchErr }) => {
-      setMessages(msgs);
-      setLoadError(fetchErr);
-      setLoading(false);
-    });
+    setHasMore(true);
+    // Bump the token; any in-flight fetch from the previous workspace
+    // (or a previous retry) will see token != myToken in its .then()
+    // and silently bail — the late response can't clobber the new
+    // workspace's state.
+    fetchTokenRef.current += 1;
+    const myToken = fetchTokenRef.current;
+    loadMessagesFromDB(workspaceId, INITIAL_HISTORY_LIMIT).then(
+      ({ messages: msgs, error: fetchErr, reachedEnd }) => {
+        if (fetchTokenRef.current !== myToken) return;
+        setMessages(msgs);
+        setLoadError(fetchErr);
+        setHasMore(!reachedEnd);
+        setLoading(false);
+      },
+    );
  }, [workspaceId]);

+  // Load chat history on mount / workspace switch.
+  // Initial load is bounded to INITIAL_HISTORY_LIMIT (newest 10) — the
+  // rest streams in as the user scrolls up via loadOlder() below. Pre-
+  // 2026-05-05 this fetched the newest 50 in one shot; on a long-running
+  // workspace that meant 50× message-bubble paint + DOM cost on every
+  // tab-open even when the user only wanted to read the last few.
+  useEffect(() => {
+    loadInitial();
+  }, [loadInitial]);
+
+  // Mirror the latest oldest-message + hasMore into refs so loadOlder
+  // can read them without taking `messages` as a dep. Every live push
+  // through agentMessages would otherwise recreate loadOlder and tear
+  // down the IO observer.
+  useEffect(() => {
+    oldestMessageRef.current = messages[0] ?? null;
+  }, [messages]);
+  useEffect(() => {
+    hasMoreRef.current = hasMore;
+  }, [hasMore]);
+
+  // Fetch the next-older batch and prepend. Stable identity (deps =
+  // [workspaceId]) so the IntersectionObserver effect below doesn't
+  // re-arm on every messages update.
+  const loadOlder = useCallback(async () => {
+    // inflightRef is the load-bearing guard — synchronous, set BEFORE
+    // any await, so two IO callbacks dispatched in the same microtask
+    // can't both pass. The state checks are defensive secondary
+    // gates for the slow-scroll case.
+    if (inflightRef.current || !hasMoreRef.current) return;
+    const oldest = oldestMessageRef.current;
+    if (!oldest) return;
+    const container = containerRef.current;
+    if (!container) return;
+    inflightRef.current = true;
+    // Capture the user's distance-from-bottom BEFORE we prepend so the
+    // useLayoutEffect can restore it after the new DOM lands. The
+    // expectFirstIdNotEqual tag is what the layout effect checks
+    // against `messages[0].id` to disambiguate prepend (id changed) vs
+    // append (id unchanged → live message landed mid-fetch). Without
+    // it, an agent push during loadOlder runs the "restore" against a
+    // stale anchor — user gets yanked + the append's bottom-pin is
+    // swallowed.
+    scrollAnchorRef.current = {
+      savedDistanceFromBottom: container.scrollHeight - container.scrollTop,
+      expectFirstIdNotEqual: oldest.id,
+    };
+    fetchTokenRef.current += 1;
+    const myToken = fetchTokenRef.current;
+    setLoadingOlder(true);
+    try {
+      const { messages: older, reachedEnd } = await loadMessagesFromDB(
+        workspaceId,
+        OLDER_HISTORY_BATCH,
+        oldest.timestamp,
+      );
+      // Workspace switched (or another loadOlder bumped the token)
+      // mid-fetch — drop these results, they belong to a stale tab.
+      if (fetchTokenRef.current !== myToken) {
+        scrollAnchorRef.current = null;
+        return;
+      }
+      if (older.length > 0) {
+        setMessages((prev) => [...older, ...prev]);
+      } else {
+        // Nothing came back — clear the anchor so the next paint doesn't
+        // try to "restore" against a no-op prepend.
+        scrollAnchorRef.current = null;
+      }
+      setHasMore(!reachedEnd);
+    } finally {
+      setLoadingOlder(false);
+      inflightRef.current = false;
+    }
+  }, [workspaceId]);
+
+  // IntersectionObserver on the top sentinel. Fires loadOlder() the
+  // moment the user scrolls within 200px of the top. AbortController
+  // unwires cleanly on workspace switch / unmount; root is the
+  // scrollable container so we observe only what's visible inside it.
+  //
+  // Dependencies:
+  //  - loadOlder    — stable per workspaceId (refs decouple it from
+  //                   message updates), so this dep is here for the
+  //                   workspace-switch case only
+  //  - hasMore      — re-run when older history runs out so we
+  //                   disconnect cleanly
+  //  - hasMessages  — load-bearing: the sentinel JSX is gated on
+  //                   `messages.length > 0`, so topRef.current is null
+  //                   on the empty-messages render. We re-arm exactly
+  //                   once when messages first land. NOT depending on
+  //                   `messages.length` (or `messages`) directly so
+  //                   each subsequent message append doesn't tear down
+  //                   + re-arm the observer.
+  const hasMessages = messages.length > 0;
+  useEffect(() => {
+    const top = topRef.current;
+    const container = containerRef.current;
+    if (!top || !container) return;
+    if (!hasMore) return; // stop observing when no older history exists
+    const ac = new AbortController();
+    const io = new IntersectionObserver(
+      (entries) => {
+        if (ac.signal.aborted) return;
+        if (entries[0]?.isIntersecting) loadOlder();
+      },
+      { root: container, rootMargin: "200px 0px 0px 0px", threshold: 0 },
+    );
+    io.observe(top);
+    ac.signal.addEventListener("abort", () => io.disconnect());
+    return () => ac.abort();
+  }, [loadOlder, hasMore, hasMessages]);
+
  // Agent reachability
  useEffect(() => {
    const reachable = data.status === "online" || data.status === "degraded";
@@ -316,7 +510,41 @@ function MyChatPanel({ workspaceId, data }: Props) {
    currentTaskRef.current = data.currentTask;
  }, [data.currentTask]);

-  useEffect(() => {
+  // Scroll behavior across messages updates:
+  //  - Prepend (loadOlder landed)  → restore the user's saved
+  //    distance-from-bottom so their reading position is unchanged.
+  //  - Append / initial            → pin to latest bubble.
+  // useLayoutEffect (not useEffect) so scroll restoration runs BEFORE
+  // paint — otherwise the user sees the page jump for one frame.
+  useLayoutEffect(() => {
+    const container = containerRef.current;
+    const anchor = scrollAnchorRef.current;
+    // Only honor the anchor when this messages-update is the prepend
+    // we expected. messages[0].id is the test:
+    //   - prepend  → messages[0] is one of the older rows → id !== expectFirstIdNotEqual
+    //   - append   → messages[0] unchanged → id === expectFirstIdNotEqual → fall through
+    // Without this check, an agent push that lands mid-loadOlder would
+    // run the restore against the append's update, yank the user's
+    // scroll, AND swallow the append's bottom-pin.
+    if (
+      anchor &&
+      container &&
+      messages.length > 0 &&
+      messages[0].id !== anchor.expectFirstIdNotEqual
+    ) {
+      container.scrollTop = container.scrollHeight - anchor.savedDistanceFromBottom;
+      scrollAnchorRef.current = null;
+      return;
+    }
+    // Instant on first arrival of messages — smooth-scroll on a long
+    // conversation gets interrupted by concurrent renders and leaves
+    // the user stuck in the middle. After the first jump, subsequent
+    // appends animate as before.
+    if (!hasInitialScrollRef.current && messages.length > 0) {
+      hasInitialScrollRef.current = true;
+      bottomRef.current?.scrollIntoView({ behavior: "instant" as ScrollBehavior });
+      return;
+    }
    bottomRef.current?.scrollIntoView({ behavior: "smooth" });
  }, [messages]);

@@ -735,7 +963,7 @@ function MyChatPanel({ workspaceId, data }: Props) {
        </div>
      )}
      {/* Messages */}
-      <div className="flex-1 overflow-y-auto p-3 space-y-3">
+      <div ref={containerRef} className="flex-1 overflow-y-auto p-3 space-y-3">
        {loading && (
          <div className="text-xs text-ink-soft text-center py-4">Loading chat history...</div>
        )}
@@ -748,15 +976,7 @@ function MyChatPanel({ workspaceId, data }: Props) {
              Failed to load chat history: {loadError}
            </p>
            <button
-              onClick={() => {
-                setLoading(true);
-                setLoadError(null);
-                loadMessagesFromDB(workspaceId).then(({ messages: msgs, error: fetchErr }) => {
-                  setMessages(msgs);
-                  setLoadError(fetchErr);
-                  setLoading(false);
-                });
-              }}
+              onClick={loadInitial}
              className="text-[10px] px-2 py-0.5 rounded bg-red-800/40 text-bad hover:bg-red-700/50 transition-colors"
            >
              Retry
@@ -768,6 +988,24 @@ function MyChatPanel({ workspaceId, data }: Props) {
            No messages yet. Send a message to start chatting with this agent.
          </div>
        )}
+        {/* Top sentinel for lazy-loading older history. The IO observer
+            in the effect above watches this; entering view triggers the
+            next-older batch fetch. Sits ABOVE messages.map so it's the
+            first thing the user reaches when scrolling up.
+
+            Only mounted when there might be more history (hasMore) so a
+            short conversation doesn't pay an idle observer. The
+            "Loading older messages…" line replaces the sentinel during
+            the fetch so the user sees feedback for the scroll-up
+            gesture. Once we hit the end, we drop the sentinel entirely
+            instead of showing a "no more messages" footer — the user's
+            scroll resting against the top of the conversation IS the
+            signal. */}
+        {hasMore && messages.length > 0 && (
+          <div ref={topRef} className="text-xs text-ink-soft text-center py-1">
+            {loadingOlder ? "Loading older messages…" : " "}
+          </div>
+        )}
        {messages.map((msg) => (
          <div key={msg.id} className={`flex ${msg.role === "user" ? "justify-end" : "justify-start"}`}>
            <div
@@ -806,14 +1044,85 @@ function MyChatPanel({ workspaceId, data }: Props) {
                      : "dark:prose-invert dark:[--tw-prose-invert-body:theme(colors.zinc.100)] dark:[--tw-prose-invert-headings:theme(colors.white)] dark:[--tw-prose-invert-bold:theme(colors.white)] dark:[--tw-prose-invert-code:theme(colors.zinc.100)]"
                  }`}
                >
-                  <ReactMarkdown remarkPlugins={[remarkGfm]}>{msg.content}</ReactMarkdown>
+                  <ReactMarkdown
+                    remarkPlugins={[remarkGfm]}
+                    components={{
+                      // Default ReactMarkdown renders `<a href="...">`
+                      // with no target and no scheme handling, so:
+                      //
+                      //   1. http/https links navigate the canvas tab
+                      //      itself away — user loses canvas state.
+                      //   2. workspace://, file://, and bare /workspace/
+                      //      paths from agent-authored markdown produce
+                      //      an unhandled-protocol click → browser ends
+                      //      up at about:blank with no download (the
+                      //      reported bug from 2026-05-05).
+                      //
+                      // Override: external URLs open in a new tab with
+                      // rel="noopener noreferrer"; in-container paths
+                      // route through downloadChatFile so the browser
+                      // gets a real Blob with proper auth headers.
+                      a: ({ href, children, ...rest }) => {
+                        const url = String(href ?? "");
+                        // Use the SSOT helper isPlatformAttachment so
+                        // the markdown link override and the chip
+                        // download path agree on which schemes need
+                        // auth-routed download. Pre-fix this list was
+                        // duplicated and missed `platform-pending:`,
+                        // producing about:blank for poll-mode uploads.
+                        if (isPlatformAttachment(url)) {
+                          return (
+                            <a
+                              href={url}
+                              {...rest}
+                              onClick={(e) => {
+                                e.preventDefault();
+                                // Construct a synthetic ChatAttachment
+                                // and route through the same
+                                // authenticated download path the
+                                // download chips use. Filename is the
+                                // last path segment so Save-As prefills
+                                // sensibly.
+                                const name = url.split(/[\\/]/).pop() || "download";
+                                downloadChatFile(workspaceId, {
+                                  uri: url,
+                                  name,
+                                }).catch((err) => {
+                                  setError(
+                                    err instanceof Error
+                                      ? `Download failed: ${err.message}`
+                                      : "Download failed",
+                                  );
+                                });
+                              }}
+                            >
+                              {children}
+                            </a>
+                          );
+                        }
+                        // External (http(s) / mailto / unknown scheme):
+                        // open in new tab so canvas state survives.
+                        return (
+                          <a
+                            href={url}
+                            target="_blank"
+                            rel="noopener noreferrer"
+                            {...rest}
+                          >
+                            {children}
+                          </a>
+                        );
+                      },
+                    }}
+                  >{msg.content}</ReactMarkdown>
                </div>
              )}
              {msg.attachments && msg.attachments.length > 0 && (
                <div className={`flex flex-wrap gap-1 ${msg.content ? "mt-1.5" : ""}`}>
                  {msg.attachments.map((att, i) => (
-                    <AttachmentChip
+                    <AttachmentPreview
                      key={`${msg.id}-${i}`}
+                      workspaceId={workspaceId}
                      attachment={att}
                      onDownload={downloadAttachment}
                      tone={msg.role === "user" ? "user" : "agent"}
@@ -912,7 +1221,22 @@ function MyChatPanel({ workspaceId, data }: Props) {
            value={input}
            onChange={(e) => setInput(e.target.value)}
            onKeyDown={(e) => {
-              if (e.key === "Enter" && !e.shiftKey) {
+              // IME-safe send: while a CJK / Japanese / Korean IME is
+              // composing, Enter accepts the candidate selection — not a
+              // newline, not a send. `e.nativeEvent.isComposing` is the
+              // standard signal (modern WebKit/Blink/Gecko); the keyCode
+              // 229 fallback covers older Safari / WebKit-based mobile
+              // browsers that delay setting isComposing on the
+              // composition-end Enter. Reported 2026-05-05: typing
+              // Chinese with the system IME, pressing Enter to commit
+              // a candidate would inadvertently send the half-typed
+              // message.
+              if (
+                e.key === "Enter" &&
+                !e.shiftKey &&
+                !e.nativeEvent.isComposing &&
+                e.keyCode !== 229
+              ) {
                e.preventDefault();
                sendMessage();
              }
@@ -6,6 +6,7 @@ import { useCanvasStore } from "@/store/canvas";
 import { type ConfigData, DEFAULT_CONFIG, TextInput, NumberInput, Toggle, TagList, Section } from "./config/form-inputs";
 import { parseYaml, toYaml } from "./config/yaml-utils";
 import { SecretsSection } from "./config/secrets-section";
+import { ExternalConnectionSection } from "./ExternalConnectionSection";
 import {
  ProviderModelSelector,
  buildProviderCatalog,
@@ -20,20 +21,39 @@ interface Props {
 // --- Agent Card Section ---

 function AgentCardSection({ workspaceId }: { workspaceId: string }) {
-  const [card, setCard] = useState<Record<string, unknown> | null>(null);
-  const [loading, setLoading] = useState(true);
+  // Initial card value comes from the canvas store — node.data.agentCard
+  // is hydrated by the platform stream when the workspace appears in the
+  // graph, so reading it here avoids a duplicate `GET /workspaces/${id}`
+  // (the parent ConfigTab.loadConfig already fetches workspace metadata,
+  // and refetching here adds a serialised RTT to the panel-open path —
+  // contributed to the ~20s detail-panel load reported in core#11).
+  // Local state still tracks the edited/saved value so the editor flow
+  // is unchanged.
+  const storeCard = useCanvasStore((s) => {
+    // Defensive against test mocks that omit `nodes` (some test files
+    // stub the store with a minimal shape). In production `nodes` is
+    // always an array — empty or not — so the optional chaining only
+    // matters for the test path.
+    const node = s.nodes?.find?.((n) => n.id === workspaceId);
+    return (node?.data.agentCard as
+      | Record<string, unknown>
+      | null
+      | undefined) ?? null;
+  });
+  const [card, setCard] = useState<Record<string, unknown> | null>(storeCard);
  const [editing, setEditing] = useState(false);
  const [draft, setDraft] = useState("");
  const [saving, setSaving] = useState(false);
  const [error, setError] = useState<string | null>(null);
  const [success, setSuccess] = useState(false);

+  // If the store updates while this section is mounted (another tab
+  // pushed an update via the platform event stream), reflect that —
+  // unless the user is mid-edit, in which case we don't clobber their
+  // unsaved draft.
  useEffect(() => {
-    api.get<Record<string, unknown>>(`/workspaces/${workspaceId}`)
-      .then((ws) => setCard((ws.agent_card as Record<string, unknown>) || null))
-      .catch(() => {})
-      .finally(() => setLoading(false));
-  }, [workspaceId]);
+    if (!editing) setCard(storeCard);
+  }, [storeCard, editing]);

  const handleSave = async () => {
    setError(null);
@@ -52,9 +72,7 @@ function AgentCardSection({ workspaceId }: { workspaceId: string }) {

  return (
    <Section title="Agent Card" defaultOpen={false}>
-      {loading ? (
-        <div className="text-[10px] text-ink-soft">Loading...</div>
-      ) : editing ? (
+      {editing ? (
        <div className="space-y-2">
          <textarea
            aria-label="Agent card JSON editor"
@@ -220,47 +238,72 @@ export function ConfigTab({ workspaceId }: Props) {
    setLoading(true);
    setError(null);

-    // ALWAYS load workspace metadata first (runtime + model). These are the
-    // source of truth regardless of whether the runtime uses our config.yaml
-    // template. Without this the form falls back to empty/default values on
-    // a hermes workspace (which doesn't use our template), creating the
-    // appearance that the saved runtime is unset — and worse, clicking Save
-    // would silently flip `runtime` from `hermes` back to the dropdown
-    // default `LangGraph`. See GH #1894.
-    let wsMetadataRuntime = "";
-    let wsMetadataModel = "";
-    let wsMetadataTier: number | null = null;
-    try {
-      const ws = await api.get<{ runtime?: string; tier?: number }>(`/workspaces/${workspaceId}`);
-      wsMetadataRuntime = (ws.runtime || "").trim();
-      if (typeof ws.tier === "number") wsMetadataTier = ws.tier;
-    } catch { /* fall back to config.yaml */ }
-    try {
-      const m = await api.get<{ model?: string }>(`/workspaces/${workspaceId}/model`);
-      wsMetadataModel = (m.model || "").trim();
-    } catch { /* non-fatal */ }
+    // Load workspace metadata (runtime + model + provider) in parallel.
+    // These are independent GETs against three workspace-server endpoints
+    // and used to be awaited serially — for SaaS workspaces each call
+    // round-trips through an EIC SSH tunnel, so the previous serial
+    // pattern stacked 3-5s of tunnel-setup latency per call (core#11).
+    // Promise.all overlaps them; the per-call cost stays the same but
+    // wall time drops to max() instead of sum().
+    //
+    // Each leg has its own .catch handler that yields a sentinel value,
+    // matching the previous semantics:
+    //   - /workspaces/${id}: required source-of-truth for runtime+tier;
+    //     fall back to YAML if the GET fails (rare, network-class only).
+    //   - /workspaces/${id}/model: non-fatal; empty model lets the form
+    //     fall through to YAML runtime_config.model.
+    //   - /workspaces/${id}/provider: non-fatal; old workspace-servers
+    //     return 404, in which case provider="" and Save skips the PUT.
+    //
+    // See GH #1894 for the workspace-row-as-source-of-truth rationale
+    // that motivated splitting from a single config.yaml read.
+    const [wsRes, modelRes, providerRes] = await Promise.all([
+      api.get<{ runtime?: string; tier?: number }>(`/workspaces/${workspaceId}`)
+        .catch(() => ({} as { runtime?: string; tier?: number })),
+      api.get<{ model?: string }>(`/workspaces/${workspaceId}/model`)
+        .catch(() => ({} as { model?: string })),
+      api.get<{ provider?: string }>(`/workspaces/${workspaceId}/provider`)
+        .catch(() => null),
+    ]);
+    const wsMetadataRuntime = (wsRes.runtime || "").trim();
+    const wsMetadataModel = (modelRes.model || "").trim();
+    const wsMetadataTier: number | null =
+      typeof wsRes.tier === "number" ? wsRes.tier : null;
+    if (providerRes !== null) {
+      const loadedProvider = (providerRes.provider || "").trim();
+      setProvider(loadedProvider);
+      setOriginalProvider(loadedProvider);
+    } else {
+      setProvider("");
+      setOriginalProvider("");
+    }
    // originalModel is set further down once the YAML has been parsed —
    // we want it to reflect what the form ACTUALLY rendered, which may
    // be the YAML's runtime_config.model fallback when MODEL_PROVIDER
    // is empty. Setting it here from wsMetadataModel alone would be
    // wrong for hermes/pre-#240 workspaces.

-    // Load explicit provider override (Option B PR-5). Endpoint returns
-    // {provider: "", source: "default"} when no override is set, so the
-    // empty string is the legitimate "auto-derive" signal — don't treat
-    // it as a load error. Non-fatal: an older workspace-server that
-    // predates PR-2 returns 404 here; the form falls back to "" and
-    // Save just won't PUT the provider field.
-    try {
-      const p = await api.get<{ provider?: string }>(`/workspaces/${workspaceId}/provider`);
-      const loadedProvider = (p.provider || "").trim();
-      setProvider(loadedProvider);
-      setOriginalProvider(loadedProvider);
-    } catch {
-      setProvider("");
-      setOriginalProvider("");
+    // Skip the config.yaml fetch entirely for runtimes that manage
+    // their own config (external, hermes, etc.) — they don't have a
+    // platform-side template, so the GET would 404. The catch block
+    // below handles 404 gracefully, but issuing the request adds
+    // browser-console noise + a wasted RTT on every open of the
+    // Config tab for the affected workspaces. Reported on
+    // production reno-stars 2026-05-05 (workspace runtime=external,
+    // 404 on /files/config.yaml visible in the console even though
+    // the form rendered correctly).
+    if (RUNTIMES_WITH_OWN_CONFIG.has(wsMetadataRuntime)) {
+      setConfig({
+        ...DEFAULT_CONFIG,
+        runtime: wsMetadataRuntime,
+        model: wsMetadataModel,
+        ...(wsMetadataModel ? { runtime_config: { model: wsMetadataModel } } : {}),
+        ...(wsMetadataTier !== null ? { tier: wsMetadataTier } : {}),
+      } as ConfigData);
+      setOriginalModel(wsMetadataModel);
+      setLoading(false);
+      return;
    }
-
    try {
      const res = await api.get<{ content: string }>(`/workspaces/${workspaceId}/files/config.yaml`);
      const parsed = parseYaml(res.content);
@@ -886,11 +929,24 @@ export function ConfigTab({ workspaceId }: Props) {
            </Section>
          )}

-          <Section title="Skills & Tools" defaultOpen={false}>
-            <TagList label="Skills" values={config.skills || []} onChange={(v) => update("skills", v)} placeholder="e.g. code-review" />
-            <TagList label="Tools" values={config.tools || []} onChange={(v) => update("tools", v)} placeholder="e.g. web_search, filesystem" />
-            <TagList label="Prompt Files" values={config.prompt_files || []} onChange={(v) => update("prompt_files", v)} placeholder="e.g. system-prompt.md" />
-            <TagList label="Shared Context" values={config.shared_context || []} onChange={(v) => update("shared_context", v)} placeholder="e.g. architecture.md" />
+          {/* Skills + Tools used to live here as TagList inputs. They were
+              redundant with their dedicated tabs:
+              - Skills → managed via SkillsTab (per-workspace skill folders)
+              - Tools  → managed via the Plugins tab (install/uninstall)
+              Editing them here only set the config.yaml field; the
+              actual install/load happened elsewhere. Removed to stop
+              showing the misnamed list-input affordance. */}
+
+          <Section title="Prompt Files" defaultOpen={false}>
+            <p className="text-[10px] text-ink-soft px-1 pb-1">
+              Markdown files that compose this workspace&apos;s system prompt.
+              Loaded in order at boot from the workspace config dir
+              (e.g. <code className="font-mono">system-prompt.md</code>,{' '}
+              <code className="font-mono">CLAUDE.md</code>,{' '}
+              <code className="font-mono">AGENTS.md</code>). Edit the file
+              contents directly via the Files tab.
+            </p>
+            <TagList label="Files (load order)" values={config.prompt_files || []} onChange={(v) => update("prompt_files", v)} placeholder="e.g. system-prompt.md" />
          </Section>

          <Section title="A2A Protocol" defaultOpen={false}>
@@ -947,6 +1003,9 @@ export function ConfigTab({ workspaceId }: Props) {
            : "This runtime manages its own config outside the platform template."}
        </div>
      )}
+      {!error && config.runtime === "external" && (
+        <ExternalConnectionSection workspaceId={workspaceId} />
+      )}
      {success && (
        <div className="mx-3 mb-2 px-3 py-1.5 bg-green-900/30 border border-green-800 rounded text-xs text-good">Saved</div>
      )}
@@ -0,0 +1,146 @@
+'use client';
+
+// ExternalConnectionSection — credential lifecycle controls for runtime=external
+// workspaces. Surfaced inside ConfigTab when the workspace's runtime is
+// "external"; ignored for hermes/claude-code/etc. (those have their own
+// restart-mints-token path).
+//
+// Two affordances:
+//
+//   1. "Show connection info" (read-only)
+//        Fetches GET /workspaces/:id/external/connection. Returns the
+//        connect block (PLATFORM_URL, WORKSPACE_ID, all 7 snippets) WITH
+//        auth_token="". The modal masks the token field and labels it
+//        "rotate to reveal a new token — current token is unrecoverable".
+//
+//   2. "Rotate credentials" (destructive)
+//        POST /workspaces/:id/external/rotate. Revokes any prior live
+//        tokens, mints a fresh one, returns the same connect block with
+//        auth_token populated. Old credentials stop working IMMEDIATELY —
+//        the previously-paired agent will fail auth on its next heartbeat.
+//        Confirm dialog explains this before firing.
+//
+// Reuses the existing ExternalConnectModal so the snippet UX is the
+// same as on Create — operators don't have to learn a second modal.
+
+import { useState } from "react";
+import * as Dialog from "@radix-ui/react-dialog";
+
+import { api } from "@/lib/api";
+import {
+  ExternalConnectModal,
+  type ExternalConnectionInfo,
+} from "../ExternalConnectModal";
+
+interface Props {
+  workspaceId: string;
+}
+
+export function ExternalConnectionSection({ workspaceId }: Props) {
+  const [info, setInfo] = useState<ExternalConnectionInfo | null>(null);
+  const [busy, setBusy] = useState<"show" | "rotate" | null>(null);
+  const [confirmRotate, setConfirmRotate] = useState(false);
+  const [error, setError] = useState<string | null>(null);
+
+  async function showConnection() {
+    setError(null);
+    setBusy("show");
+    try {
+      const resp = await api.get<{ connection: ExternalConnectionInfo }>(
+        `/workspaces/${workspaceId}/external/connection`,
+      );
+      setInfo(resp.connection);
+    } catch (e) {
+      setError(e instanceof Error ? e.message : String(e));
+    } finally {
+      setBusy(null);
+    }
+  }
+
+  async function doRotate() {
+    setError(null);
+    setBusy("rotate");
+    setConfirmRotate(false);
+    try {
+      const resp = await api.post<{ connection: ExternalConnectionInfo }>(
+        `/workspaces/${workspaceId}/external/rotate`,
+        {},
+      );
+      setInfo(resp.connection);
+    } catch (e) {
+      setError(e instanceof Error ? e.message : String(e));
+    } finally {
+      setBusy(null);
+    }
+  }
+
+  return (
+    <div className="mx-3 mt-3 p-3 bg-surface-sunken/50 border border-line rounded">
+      <h3 className="text-xs text-ink-mid font-medium mb-1">External Connection</h3>
+      <p className="text-[10px] text-ink-soft mb-2">
+        This workspace runs an external agent. Use these controls to
+        re-show the setup snippets or rotate the workspace token.
+      </p>
+
+      <div className="flex gap-2 flex-wrap">
+        <button
+          type="button"
+          onClick={showConnection}
+          disabled={busy !== null}
+          className="px-3 py-1.5 bg-surface-card hover:bg-surface-card text-xs rounded text-ink-mid disabled:opacity-30 transition-colors focus:outline-none focus-visible:ring-2 focus-visible:ring-accent/60"
+        >
+          {busy === "show" ? "Loading…" : "Show connection info"}
+        </button>
+        <button
+          type="button"
+          onClick={() => setConfirmRotate(true)}
+          disabled={busy !== null}
+          className="px-3 py-1.5 bg-red-900/30 hover:bg-red-900/50 border border-red-800/60 text-xs rounded text-bad disabled:opacity-30 transition-colors focus:outline-none focus-visible:ring-2 focus-visible:ring-red-600/60"
+        >
+          {busy === "rotate" ? "Rotating…" : "Rotate credentials"}
+        </button>
+      </div>
+
+      {error && (
+        <div className="mt-2 px-2 py-1 bg-red-900/30 border border-red-800 rounded text-[10px] text-bad">
+          {error}
+        </div>
+      )}
+
+      <Dialog.Root open={confirmRotate} onOpenChange={setConfirmRotate}>
+        <Dialog.Portal>
+          <Dialog.Overlay className="fixed inset-0 bg-black/60 z-50" />
+          <Dialog.Content className="fixed left-1/2 top-1/2 z-50 w-[min(440px,92vw)] -translate-x-1/2 -translate-y-1/2 rounded-xl bg-surface-sunken border border-line p-5 shadow-2xl">
+            <Dialog.Title className="text-sm font-medium text-ink mb-2">
+              Rotate workspace credentials?
+            </Dialog.Title>
+            <Dialog.Description className="text-xs text-ink-mid mb-4 leading-relaxed">
+              This will mint a new <code className="font-mono">workspace_auth_token</code> and{' '}
+              <strong>immediately invalidate the current one</strong>. Your external
+              agent will start failing authentication on its next heartbeat
+              until you redeploy it with the new token.
+            </Dialog.Description>
+            <div className="flex justify-end gap-2">
+              <button
+                type="button"
+                onClick={() => setConfirmRotate(false)}
+                className="px-3 py-1.5 bg-surface-card text-xs rounded text-ink-mid"
+              >
+                Cancel
+              </button>
+              <button
+                type="button"
+                onClick={doRotate}
+                className="px-3 py-1.5 bg-red-700 hover:bg-red-600 text-xs rounded text-white"
+              >
+                Rotate
+              </button>
+            </div>
+          </Dialog.Content>
+        </Dialog.Portal>
+      </Dialog.Root>
+
+      <ExternalConnectModal info={info} onClose={() => setInfo(null)} />
+    </div>
+  );
+}
@@ -2,9 +2,11 @@

 import { useState, useEffect, useRef, useMemo } from "react";
 import { showToast } from "../Toaster";
+import type { WorkspaceNodeData } from "@/store/canvas";
 import { FilesToolbar } from "./FilesTab/FilesToolbar";
 import { FileTree } from "./FilesTab/FileTree";
 import { FileEditor } from "./FilesTab/FileEditor";
+import { NotAvailablePanel } from "./FilesTab/NotAvailablePanel";
 import { useFilesApi } from "./FilesTab/useFilesApi";
 import { buildTree } from "./FilesTab/tree";

@@ -14,9 +16,40 @@ export type { TreeNode } from "./FilesTab/tree";

 interface Props {
  workspaceId: string;
+  /** Workspace metadata from the canvas store. Optional for back-compat
+   *  with any caller that still mounts <FilesTab workspaceId=.../> without
+   *  threading data through (legacy tests). When present, runtime gates
+   *  the early-return below. Mirrors TerminalTab's prop shape (#2830). */
+  data?: WorkspaceNodeData;
 }

-export function FilesTab({ workspaceId }: Props) {
+/** Runtimes whose filesystem the platform doesn't own. The canvas can't
+ *  list/read/write files on these — the agent runs on the user's own
+ *  hardware (mac laptop, mac mini, hermes-on-home-server) and reaches
+ *  the platform via the heartbeat-based polling Phase 30 layer.
+ *
+ *  Keep narrow — only add a runtime here when its provisioner genuinely
+ *  has no platform-owned filesystem. Otherwise the user loses access to
+ *  a real surface (e.g. claude-code SaaS workspaces have files served
+ *  by ListFiles via EIC; they belong on the rendering path, not here). */
+const RUNTIMES_WITHOUT_FILES = new Set(["external"]);
+
+export function FilesTab({ workspaceId, data }: Props) {
+  // Early-return for runtimes whose filesystem is not platform-owned.
+  // Skips the whole useFilesApi hook + tree render below — without this,
+  // mounting the tab for an external workspace would issue a GET that
+  // the platform can technically answer (it reads its own DB row, not
+  // the user's machine), but every result row is fictional. Showing
+  // "0 files / No config files yet" reads as a bug. The placeholder
+  // makes the absence intentional and points the user at the right
+  // surface (Chat).
+  if (data && RUNTIMES_WITHOUT_FILES.has(data.runtime)) {
+    return <NotAvailablePanel runtime={data.runtime} />;
+  }
+  return <PlatformOwnedFilesTab workspaceId={workspaceId} />;
+}
+
+function PlatformOwnedFilesTab({ workspaceId }: { workspaceId: string }) {
  const [root, setRoot] = useState("/configs");
  const [selectedFile, setSelectedFile] = useState<string | null>(null);
  const [fileContent, setFileContent] = useState("");
@@ -45,11 +78,36 @@ export function FilesTab({ workspaceId }: Props) {
    readFile,
    writeFile,
    deleteFile,
+    downloadFileByPath,
    downloadAllFiles,
    uploadFiles,
+    uploadDataTransferItems,
    deleteAllFiles,
  } = useFilesApi(workspaceId, root);

+  // PR-D: track whether the user is currently dragging files OVER
+  // the root area (not over a specific subdir row). Used to show
+  // the "Drop to upload to root" highlight on the tree column.
+  const [rootDragHover, setRootDragHover] = useState(false);
+
+  const handleDropToTarget = (
+    targetDir: string,
+    items: DataTransferItemList,
+  ) => {
+    // canDelete is the gate proxy — same constraint as the toolbar
+    // Upload button (today only /configs is writable from the canvas
+    // surface). Without this check, dropping on /home would post
+    // through /workspaces/<id>/files/<path>, which the backend would
+    // reject only after an HTTP round-trip. Fail fast.
+    if (root !== "/configs") {
+      setError(
+        `Upload only allowed in /configs (current root: ${root}). Switch root or use Upload button.`,
+      );
+      return;
+    }
+    void uploadDataTransferItems(items, targetDir);
+  };
+
  const tree = useMemo(() => buildTree(files), [files]);

  const openFile = async (path: string) => {
@@ -190,8 +248,46 @@ export function FilesTab({ workspaceId }: Props) {
      )}

      <div className="flex flex-1 min-h-0">
-        {/* File tree */}
-        <div className="w-[180px] border-r border-line/40 overflow-y-auto shrink-0">
+        {/* File tree column. PR-D: outer div is the drop zone for
+            "drop on root" — when the user drags into the column area
+            (not over a specific subdir row), the drop targets the
+            current root directory. Subdirectory rows in <FileTree>
+            stop propagation on their own drop event so a drop on
+            /configs/skills doesn't ALSO fire root-area drop. */}
+        <div
+          className={`w-[180px] border-r border-line/40 overflow-y-auto shrink-0 transition-colors ${
+            rootDragHover ? "bg-accent/10 outline outline-1 outline-accent/40 -outline-offset-2" : ""
+          }`}
+          onDragOver={(e) => {
+            // Only highlight + accept the drop when uploads are
+            // actually allowed for the current root. Without this
+            // check the user gets a misleading drag affordance,
+            // drops, then sees the toolbar's "switch root" toast —
+            // bad UX.
+            if (root !== "/configs") return;
+            e.preventDefault();
+            e.dataTransfer.dropEffect = "copy";
+          }}
+          onDragEnter={(e) => {
+            if (root !== "/configs") return;
+            e.preventDefault();
+            setRootDragHover(true);
+          }}
+          onDragLeave={(e) => {
+            const next = e.relatedTarget as Node | null;
+            if (!next || !(e.currentTarget as HTMLElement).contains(next)) {
+              setRootDragHover(false);
+            }
+          }}
+          onDrop={(e) => {
+            if (root !== "/configs") return;
+            e.preventDefault();
+            setRootDragHover(false);
+            if (e.dataTransfer.items?.length) {
+              handleDropToTarget("", e.dataTransfer.items);
+            }
+          }}
+        >
          {/* New file input */}
          {showNewFile && (
            <div className="px-2 py-1 border-b border-line/40">
@@ -209,14 +305,27 @@ export function FilesTab({ workspaceId }: Props) {

          {files.length === 0 ? (
            <div className="px-3 py-4 text-[10px] text-ink-soft text-center">
-              No config files yet
+              {rootDragHover
+                ? "Drop to upload to root"
+                : root === "/configs"
+                  ? "No config files yet — drag files here to upload"
+                  : "No config files yet"}
            </div>
          ) : (
            <FileTree
              nodes={tree}
              selectedPath={selectedFile}
              onSelect={openFile}
+              // Delete is currently gated to /configs to match the
+              // toolbar's New / Upload / Clear affordances. Context
+              // menu and inline ✕ both honour the gate. PR-A made the
+              // backend EIC delete work on all roots — keeping the
+              // canvas gate conservative until we want to expose
+              // /home /workspace deletion intentionally.
              onDelete={root === "/configs" ? setConfirmDelete : () => {}}
+              onDownload={downloadFileByPath}
+              canDelete={root === "/configs"}
+              onDropToTarget={handleDropToTarget}
              expandedDirs={expandedDirs}
              onToggleDir={toggleDir}
              loadingDir={loadingDir}
@@ -1,41 +1,129 @@
 "use client";

+import { useState } from "react";
 import { type TreeNode, getIcon } from "./tree";
+import { FileTreeContextMenu, type MenuItem } from "./FileTreeContextMenu";

 interface TreeCallbacks {
  selectedPath: string | null;
  onSelect: (path: string) => void;
  onDelete: (path: string) => void;
+  /** PR-C: right-click → Download. Files only — directories ignore. */
+  onDownload: (path: string) => void;
+  /** Whether the active root permits delete. Wire into the Delete
+   *  context-menu item's `disabled` flag so the user gets the same
+   *  affordance as the toolbar (which gates Clear/New on /configs). */
+  canDelete: boolean;
+  /** PR-D: drop files/folders from the OS onto this row. targetDir
+   *  is the directory path (relative to the active root) under which
+   *  the dropped contents should land; "" means root. */
+  onDropToTarget?: (targetDir: string, items: DataTransferItemList) => void;
  expandedDirs: Set<string>;
  onToggleDir: (path: string) => void;
  loadingDir: string | null;
 }

+/**
+ * FileTree renders the workspace tree + owns the right-click context
+ * menu (PR-C) and the drop-target hover state (PR-D). Lifting the
+ * menu state here (vs each row) means only one menu open at a time —
+ * opening a new row's menu auto-closes the prior one. Same UX as
+ * VSCode / Theia.
+ */
 export function FileTree({
  nodes,
  selectedPath,
  onSelect,
  onDelete,
+  onDownload,
+  canDelete,
+  onDropToTarget,
  expandedDirs,
  onToggleDir,
  loadingDir,
  depth = 0,
 }: TreeCallbacks & { nodes: TreeNode[]; depth?: number }) {
+  const [menu, setMenu] = useState<{
+    x: number;
+    y: number;
+    items: MenuItem[];
+  } | null>(null);
+  // PR-D: hover-target highlight state for drag-drop. Lifted next to
+  // the menu state so both shared-across-rows interactions live in
+  // one place.
+  const [hoverDir, setHoverDir] = useState<string | null>(null);
+
+  const openContextMenu = (e: React.MouseEvent, node: TreeNode) => {
+    e.preventDefault();
+    // Items composed per-row so the available actions reflect the
+    // node type (files get Open + Download; directories get Delete
+    // only since "open a directory in the editor" doesn't apply
+    // and "Export folder" is the toolbar's job).
+    const items: MenuItem[] = [];
+    if (!node.isDir) {
+      items.push({
+        id: "open",
+        label: "Open",
+        icon: "⤴",
+        onClick: () => onSelect(node.path),
+      });
+      items.push({
+        id: "download",
+        label: "Download",
+        icon: "↓",
+        onClick: () => onDownload(node.path),
+      });
+    }
+    items.push({
+      id: "delete",
+      label: "Delete",
+      icon: "✕",
+      destructive: true,
+      disabled: !canDelete,
+      onClick: () => onDelete(node.path),
+    });
+    setMenu({ x: e.clientX, y: e.clientY, items });
+  };
+
+  // Single state lifted to the top-level tree; nested <FileTree>s
+  // (rendered for expanded directories below) do NOT instantiate
+  // their own menus or drop-targets — they call back via prop
+  // drilling. This keeps "only one menu open" + "only one drop
+  // target highlighted" as structural invariants rather than
+  // render-order coincidences.
+  const childCallbacks: TreeCallbacks = {
+    selectedPath,
+    onSelect,
+    onDelete,
+    onDownload,
+    canDelete,
+    onDropToTarget,
+    expandedDirs,
+    onToggleDir,
+    loadingDir,
+  };
+
  return (
    <div>
      {nodes.map((node) => (
        <TreeItem
          key={`${node.path}:${node.isDir ? "dir" : "file"}`}
          node={node}
-          selectedPath={selectedPath}
-          onSelect={onSelect}
-          onDelete={onDelete}
-          expandedDirs={expandedDirs}
-          onToggleDir={onToggleDir}
-          loadingDir={loadingDir}
+          openContextMenu={openContextMenu}
+          hoverDir={hoverDir}
+          setHoverDir={setHoverDir}
          depth={depth}
+          {...childCallbacks}
        />
      ))}
+      {menu && (
+        <FileTreeContextMenu
+          x={menu.x}
+          y={menu.y}
+          items={menu.items}
+          onClose={() => setMenu(null)}
+        />
+      )}
    </div>
  );
 }
@@ -45,22 +133,81 @@ function TreeItem({
  selectedPath,
  onSelect,
  onDelete,
+  onDownload,
+  canDelete,
+  onDropToTarget,
  expandedDirs,
  onToggleDir,
  loadingDir,
  depth,
-}: TreeCallbacks & { node: TreeNode; depth: number }) {
+  openContextMenu,
+  hoverDir,
+  setHoverDir,
+}: TreeCallbacks & {
+  node: TreeNode;
+  depth: number;
+  openContextMenu: (e: React.MouseEvent, node: TreeNode) => void;
+  hoverDir: string | null;
+  setHoverDir: (p: string | null) => void;
+}) {
  const isSelected = selectedPath === node.path;
  const expanded = expandedDirs.has(node.path);
  const isLoading = loadingDir === node.path;
+  const isDropTarget = node.isDir && hoverDir === node.path;
+
+  // PR-D drag handlers — only directory rows are valid drop targets
+  // (dropping a file ON another file is ambiguous; treat it as
+  // dropping in the parent dir, which the root area handles). When a
+  // drag enters a directory row, mark it the hover target. When the
+  // cursor leaves to a non-child element, clear it. drop fires the
+  // upload callback with the row's path.
+  const dragProps = node.isDir && onDropToTarget
+    ? {
+        onDragOver: (e: React.DragEvent) => {
+          // preventDefault is REQUIRED to opt this element into the
+          // drop target list — without it, browsers refuse to fire
+          // the drop event regardless of the drop handler.
+          e.preventDefault();
+          e.dataTransfer.dropEffect = "copy";
+        },
+        onDragEnter: (e: React.DragEvent) => {
+          e.preventDefault();
+          setHoverDir(node.path);
+        },
+        onDragLeave: (e: React.DragEvent) => {
+          // Only clear hover when leaving to an element OUTSIDE this
+          // row — bare leave-events fire for every child crossed
+          // (the icon, the label, the ✕ button). Without the
+          // contains() check the highlight flickers.
+          const next = e.relatedTarget as Node | null;
+          if (!next || !(e.currentTarget as HTMLElement).contains(next)) {
+            setHoverDir(null);
+          }
+        },
+        onDrop: (e: React.DragEvent) => {
+          e.preventDefault();
+          e.stopPropagation();
+          setHoverDir(null);
+          if (e.dataTransfer.items?.length) {
+            onDropToTarget(node.path, e.dataTransfer.items);
+          }
+        },
+      }
+    : {};

  if (node.isDir) {
    return (
      <div>
        <div
-          className="group w-full flex items-center gap-1 px-2 py-0.5 text-left hover:bg-surface-card/40 transition-colors cursor-pointer"
+          className={`group w-full flex items-center gap-1 px-2 py-0.5 text-left transition-colors cursor-pointer ${
+            isDropTarget
+              ? "bg-accent/20 outline outline-1 outline-accent/60"
+              : "hover:bg-surface-card/40"
+          }`}
          style={{ paddingLeft: `${depth * 12 + 8}px` }}
          onClick={() => onToggleDir(node.path)}
+          onContextMenu={(e) => openContextMenu(e, node)}
+          {...dragProps}
        >
          <span className="text-[9px] text-ink-soft w-3">{isLoading ? "…" : expanded ? "▼" : "▶"}</span>
          <span className="text-[10px]">📁</span>
@@ -82,6 +229,9 @@ function TreeItem({
            selectedPath={selectedPath}
            onSelect={onSelect}
            onDelete={onDelete}
+            onDownload={onDownload}
+            canDelete={canDelete}
+            onDropToTarget={onDropToTarget}
            expandedDirs={expandedDirs}
            onToggleDir={onToggleDir}
            loadingDir={loadingDir}
@@ -99,6 +249,7 @@ function TreeItem({
      }`}
      style={{ paddingLeft: `${depth * 12 + 20}px` }}
      onClick={() => onSelect(node.path)}
+      onContextMenu={(e) => openContextMenu(e, node)}
    >
      <span className="text-[9px]">{getIcon(node.name, false)}</span>
      <span className="text-[10px] flex-1 truncate font-mono">{node.name}</span>
@@ -0,0 +1,141 @@
+"use client";
+
+import { useEffect, useRef } from "react";
+
+/**
+ * FileTreeContextMenu — VSCode-style right-click menu for a single
+ * file-tree row. Pops at the cursor's viewport coords; dismisses on
+ * outside-click, Esc, blur, or scroll.
+ *
+ * Why a custom component (no library): the menu is one of several
+ * "small popovers" in canvas; pulling in a dnd / popover lib for one
+ * surface adds 10x the bytes of this implementation. The patterns
+ * (outside-click + Esc + portal-free fixed position) match the
+ * ContextMenu used in canvas/Toolbar so the keyboard-nav muscle
+ * memory is uniform.
+ *
+ * Items are rendered from a `MenuItem[]` so callers can add/remove
+ * actions without touching this component (e.g. PR-D will add an
+ * "Upload to this folder" item for directory rows).
+ *
+ * Accessibility:
+ * - role="menu" + role="menuitem" so screen readers announce the
+ *   surface as a menu, not a generic div.
+ * - First item gets autofocus so keyboard users can ↓/↑/Enter without
+ *   reaching for the mouse.
+ * - Esc + outside-click + Tab dismisses; behaves like every other
+ *   menu the user has touched on the canvas.
+ */
+export interface MenuItem {
+  /** Stable identifier for testing + analytics. */
+  id: string;
+  label: string;
+  /** Optional left icon glyph; not load-bearing. */
+  icon?: string;
+  /** Destructive (rendered in red) — for Delete-class actions. */
+  destructive?: boolean;
+  /** Item-specific click handler. The menu auto-closes after onClick
+   *  fires so handlers don't have to call onClose themselves. */
+  onClick: () => void;
+  /** Disabled items render but don't fire onClick (useful for
+   *  Delete-on-non-/configs case where the caller wants to surface
+   *  the item but explain it's gated). Currently unused — placeholder
+   *  for future options. */
+  disabled?: boolean;
+}
+
+interface Props {
+  /** Viewport-coordinate position of the cursor that opened the menu. */
+  x: number;
+  y: number;
+  items: MenuItem[];
+  onClose: () => void;
+}
+
+export function FileTreeContextMenu({ x, y, items, onClose }: Props) {
+  const ref = useRef<HTMLDivElement>(null);
+  // First item gets initial focus for keyboard ↓/↑/Enter nav.
+  const firstItemRef = useRef<HTMLButtonElement>(null);
+
+  useEffect(() => {
+    firstItemRef.current?.focus();
+  }, []);
+
+  // Outside-click + Esc dismiss. Per memory
+  // (feedback_abort_controller_for_rerendered_listeners), use an
+  // AbortController so re-mounts (caller toggles the menu) don't leak
+  // listeners.
+  useEffect(() => {
+    const ctrl = new AbortController();
+    const onPointerDown = (e: MouseEvent) => {
+      if (ref.current && !ref.current.contains(e.target as Node)) onClose();
+    };
+    const onKeyDown = (e: KeyboardEvent) => {
+      if (e.key === "Escape") {
+        e.preventDefault();
+        onClose();
+      } else if (e.key === "ArrowDown" || e.key === "ArrowUp") {
+        // Roving focus across .menuitem buttons. Doing this with
+        // tabindex management because Tab / Shift+Tab leave the menu
+        // (which is the right thing — the user is escaping the menu).
+        e.preventDefault();
+        const buttons = ref.current?.querySelectorAll<HTMLButtonElement>(
+          "[role='menuitem']:not([disabled])",
+        );
+        if (!buttons || buttons.length === 0) return;
+        const arr = Array.from(buttons);
+        const cur = arr.indexOf(document.activeElement as HTMLButtonElement);
+        const next =
+          e.key === "ArrowDown"
+            ? (cur + 1) % arr.length
+            : (cur - 1 + arr.length) % arr.length;
+        arr[next].focus();
+      }
+    };
+    // `mousedown` (not `click`) so the menu dismisses BEFORE the
+    // tree-row's click handler would fire — otherwise clicking
+    // outside also selects a different row, which is not what the
+    // user expected when "outside-click closes the menu".
+    document.addEventListener("mousedown", onPointerDown, { signal: ctrl.signal });
+    document.addEventListener("keydown", onKeyDown, { signal: ctrl.signal });
+    // Scroll inside any ancestor also dismisses — the fixed-position
+    // menu would otherwise stay anchored to viewport coords while the
+    // row it points at scrolled away. Use capture so we catch scroll
+    // on inner panels (FileTree's overflow-y-auto wrapper).
+    document.addEventListener("scroll", onClose, { signal: ctrl.signal, capture: true });
+    return () => ctrl.abort();
+  }, [onClose]);
+
+  return (
+    <div
+      ref={ref}
+      role="menu"
+      aria-label="File actions"
+      className="fixed z-[1000] min-w-[140px] py-1 bg-surface-elevated border border-line/60 rounded-md shadow-xl shadow-black/30 text-[11px]"
+      style={{ left: x, top: y }}
+    >
+      {items.map((item, i) => (
+        <button
+          key={item.id}
+          ref={i === 0 ? firstItemRef : undefined}
+          type="button"
+          role="menuitem"
+          disabled={item.disabled}
+          onClick={() => {
+            if (item.disabled) return;
+            item.onClick();
+            onClose();
+          }}
+          className={
+            item.destructive
+              ? "w-full text-left px-3 py-1 text-bad hover:bg-red-900/30 focus:bg-red-900/30 focus:outline-none disabled:opacity-40 disabled:pointer-events-none transition-colors"
+              : "w-full text-left px-3 py-1 text-ink-mid hover:bg-surface-card hover:text-ink focus:bg-surface-card focus:text-ink focus:outline-none disabled:opacity-40 disabled:pointer-events-none transition-colors"
+          }
+        >
+          {item.icon && <span className="inline-block w-4 mr-1.5 text-ink-soft">{item.icon}</span>}
+          {item.label}
+        </button>
+      ))}
+    </div>
+  );
+}
@@ -0,0 +1,58 @@
+"use client";
+
+/**
+ * NotAvailablePanel — full-tab placeholder for runtimes whose filesystem
+ * the platform doesn't own (today: runtime === "external").
+ *
+ * Pre-fix the FilesTab tried to GET /workspaces/<id>/files for these
+ * workspaces. The platform answered with [] (no rows in workspace_files
+ * for an external workspace by definition), but the canvas rendered
+ * "0 files / No config files yet" which reads identically to the SaaS
+ * empty-listing bug fixed in PR-A. Showing an explicit placeholder
+ * makes the absence intentional and routes the user toward the
+ * supported surface (Chat) for these workspaces.
+ *
+ * Mirrors the same affordance TerminalTab adopted for runtimes without
+ * a TTY in PR #2830 — uniform "feature-not-applicable" UX across tabs.
+ */
+export function NotAvailablePanel({ runtime }: { runtime: string }) {
+  return (
+    <div className="flex flex-col items-center justify-center h-full p-8 text-center bg-surface-sunken/30">
+      {/* Folder-with-slash icon. Custom inline SVG so we don't depend
+          on an icon set being present at canvas build-time (matches
+          TerminalTab's NotAvailablePanel pattern). */}
+      <svg
+        width="72"
+        height="72"
+        viewBox="0 0 72 72"
+        fill="none"
+        aria-hidden="true"
+        className="text-ink-soft mb-4"
+      >
+        {/* Folder body */}
+        <path
+          d="M10 22 L10 56 a4 4 0 0 0 4 4 L58 60 a4 4 0 0 0 4 -4 L62 26 a4 4 0 0 0 -4 -4 L34 22 L28 16 L14 16 a4 4 0 0 0 -4 4 Z"
+          stroke="currentColor"
+          strokeWidth="2.5"
+          strokeLinejoin="round"
+          fill="none"
+          opacity="0.6"
+        />
+        {/* Diagonal cancel slash */}
+        <path
+          d="M14 14 L58 58"
+          stroke="currentColor"
+          strokeWidth="3"
+          strokeLinecap="round"
+        />
+      </svg>
+      <h3 className="text-sm font-medium text-ink mb-1.5">Files not available</h3>
+      <p className="text-[11px] text-ink-soft max-w-xs leading-relaxed">
+        This workspace runs the{" "}
+        <span className="font-mono text-ink-mid">{runtime}</span> runtime,
+        whose filesystem isn't owned by the platform. Use the Chat tab to
+        interact with the agent directly.
+      </p>
+    </div>
+  );
+}
@@ -0,0 +1,136 @@
+// @vitest-environment jsdom
+//
+// Pins the right-click context menu added in PR-C of issue #2999.
+// VSCode-style affordance: Open / Download / Delete on file rows,
+// Delete on directory rows. Delete is gated by `canDelete` (parent
+// only enables on /configs root, matching the toolbar's gate).
+//
+// Pinned branches:
+//   1. Right-click on a file row opens the menu at the click coords
+//      with Open + Download + Delete items.
+//   2. Right-click on a directory row opens the menu with Delete
+//      only (no Open/Download — directories don't have one-click
+//      semantics in this surface).
+//   3. Clicking Download fires the onDownload callback with the
+//      row's path.
+//   4. Clicking Delete fires onDelete with the row's path (when
+//      canDelete=true).
+//   5. Delete is disabled in the rendered menu when canDelete=false
+//      and clicking it does NOT fire onDelete (gate is real).
+//   6. Esc dismisses the menu.
+//   7. Click outside the menu dismisses it.
+
+import { describe, it, expect, vi, afterEach } from "vitest";
+import { render, screen, cleanup, fireEvent, act } from "@testing-library/react";
+import React from "react";
+import { FileTree } from "../FileTree";
+import type { TreeNode } from "../tree";
+
+afterEach(cleanup);
+
+const file: TreeNode = { name: "config.yaml", path: "config.yaml", isDir: false, children: [], size: 0 };
+const dir: TreeNode = {
+  name: "skills",
+  path: "skills",
+  isDir: true,
+  children: [],
+  size: 0,
+};
+
+function renderTree(props: Partial<React.ComponentProps<typeof FileTree>> = {}) {
+  const defaults = {
+    nodes: [file, dir],
+    selectedPath: null,
+    onSelect: vi.fn(),
+    onDelete: vi.fn(),
+    onDownload: vi.fn(),
+    canDelete: true,
+    expandedDirs: new Set<string>(),
+    onToggleDir: vi.fn(),
+    loadingDir: null,
+  };
+  const merged = { ...defaults, ...props };
+  return { ...render(<FileTree {...merged} />), props: merged };
+}
+
+describe("FileTree right-click context menu", () => {
+  it("right-click on a file row opens menu with Open/Download/Delete", () => {
+    renderTree();
+    fireEvent.contextMenu(screen.getByText("config.yaml"), {
+      clientX: 50,
+      clientY: 100,
+    });
+    expect(screen.getByRole("menu")).not.toBeNull();
+    expect(screen.getByRole("menuitem", { name: /Open/i })).not.toBeNull();
+    expect(screen.getByRole("menuitem", { name: /Download/i })).not.toBeNull();
+    expect(screen.getByRole("menuitem", { name: /Delete/i })).not.toBeNull();
+  });
+
+  it("right-click on a directory row opens menu with Delete only (no Open/Download)", () => {
+    renderTree();
+    fireEvent.contextMenu(screen.getByText("skills"), { clientX: 60, clientY: 120 });
+    expect(screen.getByRole("menu")).not.toBeNull();
+    expect(screen.queryByRole("menuitem", { name: /Open/i })).toBeNull();
+    expect(screen.queryByRole("menuitem", { name: /Download/i })).toBeNull();
+    expect(screen.getByRole("menuitem", { name: /Delete/i })).not.toBeNull();
+  });
+
+  it("clicking Download fires onDownload with the row's path", () => {
+    const { props } = renderTree();
+    fireEvent.contextMenu(screen.getByText("config.yaml"), { clientX: 0, clientY: 0 });
+    fireEvent.click(screen.getByRole("menuitem", { name: /Download/i }));
+    expect(props.onDownload).toHaveBeenCalledWith("config.yaml");
+    // Menu auto-closes after click.
+    expect(screen.queryByRole("menu")).toBeNull();
+  });
+
+  it("clicking Delete fires onDelete with the row's path when canDelete=true", () => {
+    const { props } = renderTree({ canDelete: true });
+    fireEvent.contextMenu(screen.getByText("config.yaml"), { clientX: 0, clientY: 0 });
+    fireEvent.click(screen.getByRole("menuitem", { name: /Delete/i }));
+    expect(props.onDelete).toHaveBeenCalledWith("config.yaml");
+  });
+
+  it("Delete is disabled when canDelete=false; clicking does not fire onDelete", () => {
+    const { props } = renderTree({ canDelete: false });
+    fireEvent.contextMenu(screen.getByText("config.yaml"), { clientX: 0, clientY: 0 });
+    const del = screen.getByRole("menuitem", { name: /Delete/i }) as HTMLButtonElement;
+    expect(del.disabled).toBe(true);
+    fireEvent.click(del);
+    expect(props.onDelete).not.toHaveBeenCalled();
+    // Menu stays open on disabled click — same as VSCode (the user
+    // can read the disabled-state hint without losing the menu).
+    expect(screen.getByRole("menu")).not.toBeNull();
+  });
+
+  it("Esc dismisses the menu", () => {
+    renderTree();
+    fireEvent.contextMenu(screen.getByText("config.yaml"), { clientX: 0, clientY: 0 });
+    expect(screen.getByRole("menu")).not.toBeNull();
+    act(() => {
+      fireEvent.keyDown(document, { key: "Escape" });
+    });
+    expect(screen.queryByRole("menu")).toBeNull();
+  });
+
+  it("click outside the menu dismisses it", () => {
+    renderTree();
+    fireEvent.contextMenu(screen.getByText("config.yaml"), { clientX: 0, clientY: 0 });
+    expect(screen.getByRole("menu")).not.toBeNull();
+    // mousedown on document.body — outside the menu.
+    act(() => {
+      fireEvent.mouseDown(document.body);
+    });
+    expect(screen.queryByRole("menu")).toBeNull();
+  });
+
+  it("opening a second context menu replaces the first (only one open at a time)", () => {
+    renderTree();
+    fireEvent.contextMenu(screen.getByText("config.yaml"), { clientX: 10, clientY: 10 });
+    fireEvent.contextMenu(screen.getByText("skills"), { clientX: 20, clientY: 20 });
+    // Only one menu in the DOM. The second open replaced the first
+    // because the menu state is lifted to the FileTree, not per-row.
+    const menus = screen.getAllByRole("menu");
+    expect(menus.length).toBe(1);
+  });
+});
@@ -0,0 +1,212 @@
+// @vitest-environment jsdom
+//
+// Pins the drag-drop upload added in PR-D of issue #2999.
+// Two layers of coverage:
+//
+//  1. The pure walker (collectFileEntries / walkEntry) — pins the
+//     recursion shape against silent folder truncation. Browsers
+//     return up to ~100 entries per readEntries() call; if the loop
+//     stops early, large folder uploads silently drop files. We
+//     simulate a multi-batch reader to discriminate.
+//
+//  2. FileTree directory-row drop handlers — pins that dragover/drop
+//     events fire onDropToTarget with the directory's path + the
+//     drop's DataTransferItemList.
+
+import { describe, it, expect, vi, afterEach } from "vitest";
+import { render, screen, cleanup, fireEvent } from "@testing-library/react";
+import React from "react";
+import { FileTree } from "../FileTree";
+import type { TreeNode } from "../tree";
+import { __testables } from "../useFilesApi";
+
+afterEach(cleanup);
+
+// ---- Walker tests ----
+
+/**
+ * Build a fake FileSystemEntry tree we can hand to walkEntry. The
+ * shape mimics what webkitGetAsEntry returns from a real OS drag —
+ * directory entries expose createReader, file entries expose file().
+ */
+function fakeFileEntry(name: string, content = "x"): {
+  isFile: true;
+  isDirectory: false;
+  name: string;
+  fullPath: string;
+  file: (cb: (f: File) => void) => void;
+} {
+  return {
+    isFile: true,
+    isDirectory: false,
+    name,
+    fullPath: "/" + name,
+    file: (cb) => cb(new File([content], name, { type: "text/plain" })),
+  };
+}
+
+function fakeDirEntry(
+  name: string,
+  childBatches: ReturnType<typeof fakeFileEntry>[][],
+): {
+  isFile: false;
+  isDirectory: true;
+  name: string;
+  fullPath: string;
+  createReader: () => { readEntries: (cb: (entries: unknown[]) => void) => void };
+} {
+  let i = 0;
+  return {
+    isFile: false,
+    isDirectory: true,
+    name,
+    fullPath: "/" + name,
+    createReader: () => ({
+      readEntries: (cb) => {
+        // Mimic browser semantics: emit one batch per call, then
+        // an empty array to signal end-of-stream. A walker that
+        // calls readEntries only once would silently truncate at
+        // the first batch.
+        if (i < childBatches.length) {
+          cb(childBatches[i++]);
+        } else {
+          cb([]);
+        }
+      },
+    }),
+  };
+}
+
+describe("walkEntry — folder-recursion drop walker", () => {
+  it("collects a single dropped file", async () => {
+    const out: { file: File; relativePath: string }[] = [];
+    await __testables.walkEntry(fakeFileEntry("README.md") as never, "", out);
+    expect(out.length).toBe(1);
+    expect(out[0].relativePath).toBe("README.md");
+    expect(out[0].file.name).toBe("README.md");
+  });
+
+  it("walks a folder and preserves the relative path under the folder name", async () => {
+    const out: { file: File; relativePath: string }[] = [];
+    const folder = fakeDirEntry("skills", [
+      [fakeFileEntry("a.md"), fakeFileEntry("b.md")],
+    ]);
+    await __testables.walkEntry(folder as never, "", out);
+    expect(out.map((e) => e.relativePath).sort()).toEqual([
+      "skills/a.md",
+      "skills/b.md",
+    ]);
+  });
+
+  it("loops readEntries until empty so a multi-batch folder isn't truncated", async () => {
+    // Browsers limit each readEntries() call to ~100 entries. Our
+    // walker MUST call it again until an empty batch is returned.
+    // Fake reader emits two batches of 2 + an implicit empty → 4
+    // total. A buggy walker that only takes the first batch would
+    // see only 2.
+    const out: { file: File; relativePath: string }[] = [];
+    const folder = fakeDirEntry("big", [
+      [fakeFileEntry("1.txt"), fakeFileEntry("2.txt")],
+      [fakeFileEntry("3.txt"), fakeFileEntry("4.txt")],
+    ]);
+    await __testables.walkEntry(folder as never, "", out);
+    expect(out.length).toBe(4);
+  });
+
+  it("walks nested directories and accumulates the full path", async () => {
+    const out: { file: File; relativePath: string }[] = [];
+    const inner = fakeDirEntry("web-search", [[fakeFileEntry("SKILL.md")]]);
+    // Outer dir whose first batch contains a sub-dir entry.
+    const outer = {
+      isFile: false,
+      isDirectory: true,
+      name: "skills",
+      fullPath: "/skills",
+      createReader: () => {
+        let i = 0;
+        return {
+          readEntries: (cb: (entries: unknown[]) => void) => {
+            if (i++ === 0) cb([inner]);
+            else cb([]);
+          },
+        };
+      },
+    };
+    await __testables.walkEntry(outer as never, "", out);
+    expect(out.length).toBe(1);
+    expect(out[0].relativePath).toBe("skills/web-search/SKILL.md");
+  });
+});
+
+// ---- FileTree drag-drop wiring ----
+
+const file: TreeNode = { name: "config.yaml", path: "config.yaml", isDir: false, children: [], size: 0 };
+const skillsDir: TreeNode = { name: "skills", path: "skills", isDir: true, children: [], size: 0 };
+
+function renderTree(props: Partial<React.ComponentProps<typeof FileTree>> = {}) {
+  // PR-D test defaults must include PR-C's onDownload + canDelete now
+  // that they're required on the TreeCallbacks shape (the rebase
+  // surfaced this — the merged tree depends on both feature sets).
+  const defaults: React.ComponentProps<typeof FileTree> = {
+    nodes: [file, skillsDir],
+    selectedPath: null,
+    onSelect: vi.fn(),
+    onDelete: vi.fn(),
+    onDownload: vi.fn(),
+    canDelete: true,
+    onDropToTarget: vi.fn(),
+    expandedDirs: new Set<string>(),
+    onToggleDir: vi.fn(),
+    loadingDir: null,
+  };
+  const merged = { ...defaults, ...props };
+  return { ...render(<FileTree {...merged} />), props: merged };
+}
+
+describe("FileTree directory-row drag-drop", () => {
+  it("dragover on a directory row preventDefault's so the drop will fire", () => {
+    renderTree();
+    const row = screen.getByText("skills");
+    const dragOver = new Event("dragover", { bubbles: true, cancelable: true });
+    Object.defineProperty(dragOver, "dataTransfer", {
+      value: { dropEffect: "" },
+    });
+    row.parentElement!.dispatchEvent(dragOver);
+    // preventDefault registers via the React handler — without it
+    // the drop event would never fire, so this assertion is the
+    // load-bearing one.
+    expect(dragOver.defaultPrevented).toBe(true);
+  });
+
+  it("drop on a directory row fires onDropToTarget with that path + the items list", () => {
+    const { props } = renderTree();
+    const row = screen.getByText("skills").parentElement!;
+    const fakeItems = { length: 1, 0: { kind: "file" } } as unknown as DataTransferItemList;
+    fireEvent.drop(row, { dataTransfer: { items: fakeItems } });
+    expect(props.onDropToTarget).toHaveBeenCalledWith("skills", fakeItems);
+  });
+
+  it("drop on a FILE row does NOT fire onDropToTarget (only directories are valid targets)", () => {
+    const { props } = renderTree();
+    const fileRow = screen.getByText("config.yaml").parentElement!;
+    const fakeItems = { length: 1, 0: { kind: "file" } } as unknown as DataTransferItemList;
+    fireEvent.drop(fileRow, { dataTransfer: { items: fakeItems } });
+    expect(props.onDropToTarget).not.toHaveBeenCalled();
+  });
+
+  it("drop with no DataTransferItems does NOT fire onDropToTarget", () => {
+    const { props } = renderTree();
+    const row = screen.getByText("skills").parentElement!;
+    fireEvent.drop(row, { dataTransfer: { items: { length: 0 } } });
+    expect(props.onDropToTarget).not.toHaveBeenCalled();
+  });
+
+  it("dragenter sets the drop-target highlight on the directory row", () => {
+    renderTree();
+    const row = screen.getByText("skills").parentElement!;
+    fireEvent.dragEnter(row, { dataTransfer: {} });
+    // Highlight class is the discriminator — without dragenter
+    // wiring the row stays in its hover-only style.
+    expect(row.className).toMatch(/bg-accent|outline-accent/);
+  });
+});
@@ -90,6 +90,43 @@ export function useFilesApi(workspaceId: string, root: string) {
    [workspaceId]
  );

+  /**
+   * Fetch a file's content from the server and trigger a browser
+   * download. Used by the right-click "Download" context-menu item
+   * (PR-C of issue #2999) — distinct from `handleDownloadFile` in
+   * FilesTab which downloads the CURRENTLY-OPEN-IN-EDITOR file from
+   * the in-memory `editContent` buffer (so unsaved edits round-trip
+   * to disk). This helper downloads the on-server content, suitable
+   * for arbitrary tree rows the user hasn't opened.
+   */
+  const downloadFileByPath = useCallback(
+    async (path: string) => {
+      try {
+        const res = await api.get<{ content: string }>(
+          `/workspaces/${workspaceId}/files/${path}?root=${encodeURIComponent(root)}`,
+        );
+        // text/plain is correct for the canvas's text-only file
+        // surface (config.yaml, prompts, skill markdown). Binary
+        // files would need an Accept-arraybuffer path; the API
+        // returns string today so this matches the wire shape.
+        const blob = new Blob([res.content], { type: "text/plain" });
+        const url = URL.createObjectURL(blob);
+        const a = document.createElement("a");
+        a.href = url;
+        a.download = path.split("/").pop() || "file";
+        a.click();
+        URL.revokeObjectURL(url);
+        showToast(`Downloaded ${a.download}`, "success");
+      } catch (e) {
+        showToast(
+          `Download failed: ${e instanceof Error ? e.message : "unknown error"}`,
+          "error",
+        );
+      }
+    },
+    [workspaceId, root],
+  );
+
  const downloadAllFiles = useCallback(async () => {
    const fileEntries = files.filter((f) => !f.dir);
    const results = await Promise.allSettled(
@@ -114,16 +151,20 @@ export function useFilesApi(workspaceId: string, root: string) {
  }, [files, workspaceId]);

  const uploadFiles = useCallback(
-    async (fileList: FileList) => {
+    async (fileList: FileList, targetDir = "") => {
      let uploaded = 0;
      for (const file of Array.from(fileList)) {
        const path = file.webkitRelativePath || file.name;
        const parts = path.split("/");
+        // For folder picker: webkitRelativePath is "<picked-folder>/a/b.txt"
+        // — strip the picked-folder prefix so files land flat under the
+        // workspace's target dir, not under a redundant outer folder.
        const relPath = parts.length > 1 ? parts.slice(1).join("/") : parts[0];
+        const finalPath = targetDir ? `${targetDir}/${relPath}` : relPath;
        if (file.size > 1_000_000) continue;
        try {
          const content = await file.text();
-          await api.put(`/workspaces/${workspaceId}/files/${relPath}`, { content });
+          await api.put(`/workspaces/${workspaceId}/files/${finalPath}`, { content });
          uploaded++;
        } catch {
          /* skip binary */
@@ -131,7 +172,7 @@ export function useFilesApi(workspaceId: string, root: string) {
      }
      if (uploaded > 0) {
        useCanvasStore.getState().updateNodeData(workspaceId, { needsRestart: true });
-        showToast(`Uploaded ${uploaded} files`, "success");
+        showToast(`Uploaded ${uploaded} files${targetDir ? ` to ${targetDir}` : ""}`, "success");
        loadFiles();
      }
      return uploaded;
@@ -139,6 +180,58 @@ export function useFilesApi(workspaceId: string, root: string) {
    [workspaceId, loadFiles]
  );

+  /**
+   * Upload files dragged from the OS via the HTML5 DataTransferItemList
+   * API. Unlike the folder-picker path (uploadFiles), this preserves
+   * the dropped folder structure under `targetDir` — drag a "skills/"
+   * folder onto the /configs/skills row and you get
+   * /configs/skills/skills/* (the OUTER folder name is preserved
+   * because the user explicitly chose to drop a NAMED folder, unlike
+   * the folder-picker which always wraps the picked dir).
+   *
+   * Walks FileSystemDirectoryEntry recursively via webkitGetAsEntry.
+   * VSCode/JupyterLab use the same primitive — there's no other
+   * portable browser API for "drag a folder from OS". `webkit*`
+   * naming is a Chromium relic; Firefox + Safari implement the same
+   * surface.
+   *
+   * Returns the number of files uploaded so the caller can show a
+   * tally / fail toast.
+   */
+  const uploadDataTransferItems = useCallback(
+    async (items: DataTransferItemList, targetDir = "") => {
+      const fileEntries = collectFileEntries(items);
+      let uploaded = 0;
+      for (const { file, relativePath } of await fileEntries) {
+        if (file.size > 1_000_000) continue;
+        const finalPath = targetDir
+          ? `${targetDir}/${relativePath}`
+          : relativePath;
+        try {
+          const content = await file.text();
+          await api.put(`/workspaces/${workspaceId}/files/${finalPath}`, {
+            content,
+          });
+          uploaded++;
+        } catch {
+          /* skip binary */
+        }
+      }
+      if (uploaded > 0) {
+        useCanvasStore
+          .getState()
+          .updateNodeData(workspaceId, { needsRestart: true });
+        showToast(
+          `Uploaded ${uploaded} file${uploaded === 1 ? "" : "s"}${targetDir ? ` to ${targetDir}` : ""}`,
+          "success",
+        );
+        loadFiles();
+      }
+      return uploaded;
+    },
+    [workspaceId, loadFiles],
+  );
+
  const deleteAllFiles = useCallback(async () => {
    let deleted = 0;
    for (const f of files) {
@@ -165,8 +258,98 @@ export function useFilesApi(workspaceId: string, root: string) {
    readFile,
    writeFile,
    deleteFile,
+    downloadFileByPath,
    downloadAllFiles,
    uploadFiles,
+    uploadDataTransferItems,
    deleteAllFiles,
  };
 }
+
+// ----- DataTransfer entry walker (PR-D) ---------------------------------
+
+/**
+ * Minimal subset of the FileSystem Entry API surface we use. The DOM
+ * lib types this as FileSystemEntry / FileSystemFileEntry /
+ * FileSystemDirectoryEntry but the relevant methods are callback-
+ * based. Keep the shape narrow + explicit so the recursion below
+ * type-checks without pulling in the full DOM lib types.
+ */
+interface FSEntry {
+  isFile: boolean;
+  isDirectory: boolean;
+  name: string;
+  fullPath: string;
+  file?(success: (f: File) => void, fail?: (e: unknown) => void): void;
+  createReader?(): { readEntries(success: (entries: FSEntry[]) => void): void };
+}
+
+interface CollectedEntry {
+  file: File;
+  /** Path relative to the dropped root (e.g. "skills/web-search/SKILL.md"
+   *  for a dropped "skills/" folder containing web-search/SKILL.md). */
+  relativePath: string;
+}
+
+/**
+ * Walk a DataTransferItemList, returning every file entry as a flat
+ * array keyed by the path relative to the originally-dropped item.
+ * Folders dropped from the OS expand recursively; loose files
+ * passthrough with name as the relative path.
+ *
+ * Skips items where webkitGetAsEntry() returns null — that's how
+ * the browser signals a non-file payload (e.g. a dragged URL or
+ * text snippet).
+ */
+async function collectFileEntries(
+  items: DataTransferItemList,
+): Promise<CollectedEntry[]> {
+  const out: CollectedEntry[] = [];
+  for (let i = 0; i < items.length; i++) {
+    const item = items[i];
+    if (item.kind !== "file") continue;
+    // webkitGetAsEntry is the standardised name; older Firefox used
+    // getAsEntry. Both Chromium + Firefox + Safari ship the webkit-
+    // prefixed variant today. There's no non-prefixed alternative.
+    const entry = (item as DataTransferItem & {
+      webkitGetAsEntry?: () => FSEntry | null;
+    }).webkitGetAsEntry?.();
+    if (!entry) continue;
+    await walkEntry(entry, "", out);
+  }
+  return out;
+}
+
+async function walkEntry(
+  entry: FSEntry,
+  prefix: string,
+  out: CollectedEntry[],
+): Promise<void> {
+  const name = entry.name;
+  const relPath = prefix ? `${prefix}/${name}` : name;
+  if (entry.isFile && entry.file) {
+    const file = await new Promise<File>((resolve, reject) => {
+      entry.file!(resolve, reject);
+    });
+    out.push({ file, relativePath: relPath });
+    return;
+  }
+  if (entry.isDirectory && entry.createReader) {
+    const reader = entry.createReader();
+    // readEntries returns up to ~100 at a time on Chromium; loop
+    // until empty so large folders aren't truncated.
+    let batch: FSEntry[] = [];
+    do {
+      batch = await new Promise<FSEntry[]>((resolve) =>
+        reader.readEntries(resolve),
+      );
+      for (const child of batch) {
+        await walkEntry(child, relPath, out);
+      }
+    } while (batch.length > 0);
+  }
+}
+
+// Exported for direct testing — the recursion + readEntries batching
+// is the part most likely to silently truncate a real folder upload.
+export const __testables = { collectFileEntries, walkEntry };
@@ -10,6 +10,7 @@ interface Props {
 interface MemoryEntry {
  key: string;
  value: unknown;
+  version?: number;
  expires_at: string | null;
  updated_at: string;
 }
@@ -28,6 +29,10 @@ export function MemoryTab({ workspaceId }: Props) {
  const [newValue, setNewValue] = useState("");
  const [newTTL, setNewTTL] = useState("");
  const [error, setError] = useState<string | null>(null);
+  const [editingKey, setEditingKey] = useState<string | null>(null);
+  const [editValue, setEditValue] = useState("");
+  const [editTTL, setEditTTL] = useState("");
+  const [editError, setEditError] = useState<string | null>(null);

  const awarenessUrl = useMemo(() => {
    try {
@@ -109,6 +114,69 @@ export function MemoryTab({ workspaceId }: Props) {
    }
  };

+  const beginEdit = (entry: MemoryEntry) => {
+    setEditError(null);
+    setEditingKey(entry.key);
+    // Stringify objects/arrays as pretty JSON; render plain strings raw so the
+    // editor doesn't surprise users with surrounding quotes.
+    setEditValue(
+      typeof entry.value === "string"
+        ? entry.value
+        : JSON.stringify(entry.value, null, 2),
+    );
+    if (entry.expires_at) {
+      const remainingMs = new Date(entry.expires_at).getTime() - Date.now();
+      const ttl = Math.max(0, Math.floor(remainingMs / 1000));
+      setEditTTL(ttl > 0 ? String(ttl) : "");
+    } else {
+      setEditTTL("");
+    }
+  };
+
+  const cancelEdit = () => {
+    setEditingKey(null);
+    setEditValue("");
+    setEditTTL("");
+    setEditError(null);
+  };
+
+  const handleEditSave = async (entry: MemoryEntry) => {
+    setEditError(null);
+
+    let parsedValue: unknown;
+    try {
+      parsedValue = JSON.parse(editValue);
+    } catch {
+      parsedValue = editValue;
+    }
+
+    // if_match_version closes the silent-overwrite hole when two writers
+    // race. The handler returns 409 with the current version on mismatch
+    // — surface that as a retry hint and reload to pick up the new state.
+    const body: Record<string, unknown> = { key: entry.key, value: parsedValue };
+    if (typeof entry.version === "number") {
+      body.if_match_version = entry.version;
+    }
+    if (editTTL) {
+      const ttl = parseInt(editTTL);
+      if (!Number.isNaN(ttl) && ttl > 0) body.ttl_seconds = ttl;
+    }
+
+    try {
+      await api.post(`/workspaces/${workspaceId}/memory`, body);
+      cancelEdit();
+      loadMemory();
+    } catch (e) {
+      const message = e instanceof Error ? e.message : "Failed to save";
+      if (message.includes("409") || /if_match_version mismatch/i.test(message)) {
+        setEditError("This entry changed since you opened it. Reloading.");
+        loadMemory();
+      } else {
+        setEditError(message);
+      }
+    }
+  };
+
  const openAwareness = () => {
    window.open(awarenessUrl, "_blank", "noopener,noreferrer");
  };
@@ -308,24 +376,71 @@ export function MemoryTab({ workspaceId }: Props) {

                  {expanded === entry.key && (
                    <div className="px-3 pb-2 space-y-2">
-                      <pre className="text-[10px] text-ink-mid bg-surface-sunken rounded p-2 overflow-x-auto max-h-40">
-                        {JSON.stringify(entry.value, null, 2)}
-                      </pre>
+                      {editingKey === entry.key ? (
+                        <div className="space-y-2">
+                          <textarea
+                            value={editValue}
+                            onChange={(e) => setEditValue(e.target.value)}
+                            rows={4}
+                            aria-label={`Edit value for ${entry.key}`}
+                            className="w-full bg-surface-sunken border border-line rounded px-2 py-1 text-xs font-mono text-ink focus:outline-none focus:border-accent resize-none"
+                          />
+                          <input
+                            value={editTTL}
+                            onChange={(e) => setEditTTL(e.target.value)}
+                            placeholder="TTL in seconds (blank = no expiry)"
+                            aria-label={`Edit TTL for ${entry.key}`}
+                            className="w-full bg-surface-sunken border border-line rounded px-2 py-1 text-xs text-ink focus:outline-none focus:border-accent"
+                          />
+                          {editError && (
+                            <div role="alert" className="text-[10px] text-bad">
+                              {editError}
+                            </div>
+                          )}
+                          <div className="flex gap-2">
+                            <button
+                              type="button"
+                              onClick={() => handleEditSave(entry)}
+                              className="px-3 py-1 bg-accent hover:bg-accent-strong text-xs rounded text-white"
+                            >
+                              Save
+                            </button>
+                            <button
+                              type="button"
+                              onClick={cancelEdit}
+                              className="px-3 py-1 bg-surface-card hover:bg-surface-elevated text-xs rounded text-ink-mid"
+                            >
+                              Cancel
+                            </button>
+                          </div>
+                        </div>
+                      ) : (
+                        <pre className="text-[10px] text-ink-mid bg-surface-sunken rounded p-2 overflow-x-auto max-h-40">
+                          {JSON.stringify(entry.value, null, 2)}
+                        </pre>
+                      )}
                      <div className="flex items-center justify-between">
                        <span className="text-[9px] text-ink-soft">
                          Updated: {new Date(entry.updated_at).toLocaleString()}
                        </span>
-                        <button
-                          type="button"
-                          onClick={() => handleDelete(entry.key)}
-                          // hover:text-bad on top of text-bad was a no-op.
-                          // Switch to a hover bg + focus-visible ring so
-                          // the destructive button visibly responds and
-                          // keyboard users see focus.
-                          className="text-[10px] text-bad hover:bg-red-950/40 rounded px-1 transition-colors focus:outline-none focus-visible:ring-2 focus-visible:ring-red-500/60"
-                        >
-                          Delete
-                        </button>
+                        <div className="flex items-center gap-2">
+                          {editingKey !== entry.key && (
+                            <button
+                              type="button"
+                              onClick={() => beginEdit(entry)}
+                              className="text-[10px] text-ink-mid hover:bg-surface-elevated rounded px-1 transition-colors focus:outline-none focus-visible:ring-2 focus-visible:ring-accent/60"
+                            >
+                              Edit
+                            </button>
+                          )}
+                          <button
+                            type="button"
+                            onClick={() => handleDelete(entry.key)}
+                            className="text-[10px] text-bad hover:bg-red-950/40 rounded px-1 transition-colors focus:outline-none focus-visible:ring-2 focus-visible:ring-red-500/60"
+                          >
+                            Delete
+                          </button>
+                        </div>
                      </div>
                    </div>
                  )}
@@ -297,10 +297,49 @@ export function SkillsTab({ workspaceId, data }: Props) {
    }
  };

+  // Compact-empty pattern: when the workspace has zero plugins
+  // installed AND the registry isn't open, collapse the whole
+  // "Plugins" section into a single inline pill rather than rendering
+  // the full panel chrome. Reported on production 2026-05-05 (#2971):
+  // the empty state's panel-with-zero-list-rows layout gives the user
+  // a lot of vertical real estate for content that's just "0
+  // installed + Install button". The compact form keeps that
+  // affordance without the chrome.
+  //
+  // Expanded/full layout still fires when installed.length > 0 OR
+  // when the user opens the registry (clicked "+ Install Plugin").
+  // Once a plugin is installed the section auto-expands to surface
+  // the list.
+  const compactEmpty = installed.length === 0 && !showRegistry && installedLoaded;
+
+  if (compactEmpty) {
+    return (
+      <div className="p-4 space-y-4">
+        <div
+          className="flex items-center justify-between gap-2 rounded-full border border-line/60 bg-surface-sunken/70 px-3 py-1.5"
+          aria-label="Plugins (none installed)"
+        >
+          <div className="flex items-center gap-2">
+            <span className="text-[10px] uppercase tracking-[0.2em] text-ink-soft">Plugins</span>
+            <span className="text-[11px] text-ink-mid">0 installed</span>
+          </div>
+          <button
+            onClick={() => setShowRegistry(true)}
+            className="rounded-full border border-violet-700/50 bg-violet-950/30 px-3 py-0.5 text-[10px] text-violet-200 hover:bg-violet-900/40 transition-colors"
+            aria-expanded="false"
+            aria-controls="plugins-section"
+          >
+            + Install Plugin
+          </button>
+        </div>
+      </div>
+    );
+  }
+
  return (
    <div className="p-4 space-y-4">
      {/* Plugins section */}
-      <div className="rounded-xl border border-line bg-surface-sunken/70 p-3">
+      <div id="plugins-section" className="rounded-xl border border-line bg-surface-sunken/70 p-3">
        <div className="flex items-center justify-between gap-3">
          <div>
            <div className="text-[10px] uppercase tracking-[0.22em] text-ink-soft">Plugins</div>
@@ -311,6 +350,8 @@ export function SkillsTab({ workspaceId, data }: Props) {
          <button
            onClick={() => setShowRegistry(!showRegistry)}
            className="rounded-full border border-violet-700/50 bg-violet-950/30 px-3 py-1 text-[10px] text-violet-200 hover:bg-violet-900/40 transition-colors"
+            aria-expanded={showRegistry}
+            aria-controls="plugins-registry"
          >
            {showRegistry ? "Hide Registry" : "+ Install Plugin"}
          </button>
@@ -1,16 +1,105 @@
 "use client";

 import { useEffect, useRef, useState, useCallback } from "react";
+import type { WorkspaceNodeData } from "@/store/canvas";

 interface Props {
  workspaceId: string;
+  /** Workspace metadata from the canvas store. Optional for back-compat
+   *  with any caller that still mounts <TerminalTab workspaceId=... />
+   *  without threading data through (e.g. tests). When present, the
+   *  runtime field gates the early-return below. */
+  data?: WorkspaceNodeData;
 }

 import { deriveWsBaseUrl } from "@/lib/ws-url";

 const WS_URL = deriveWsBaseUrl();

-export function TerminalTab({ workspaceId }: Props) {
+/**
+ * NotAvailablePanel — full-tab placeholder with a big terminal-off icon
+ * for runtimes that don't expose a TTY (e.g. external workspaces, where
+ * the platform doesn't own the process). Pre-fix the tab tried to open
+ * a WebSocket against /ws/terminal/<id> for these workspaces, the server
+ * 404'd, and the user saw "Connection failed" — which reads as a bug,
+ * not as "this runtime intentionally has no shell". This banner makes
+ * the absence intentional.
+ */
+function NotAvailablePanel({ runtime }: { runtime: string }) {
+  return (
+    <div className="flex flex-col items-center justify-center h-full p-8 text-center bg-surface-sunken/30">
+      {/* Big terminal-off icon — bracket "[_]" with a slash through it.
+          Custom inline SVG so we don't depend on an icon set being
+          present at canvas build-time. */}
+      <svg
+        width="72"
+        height="72"
+        viewBox="0 0 72 72"
+        fill="none"
+        aria-hidden="true"
+        className="text-ink-soft mb-4"
+      >
+        <rect
+          x="10"
+          y="14"
+          width="52"
+          height="44"
+          rx="4"
+          stroke="currentColor"
+          strokeWidth="2.5"
+          fill="none"
+          opacity="0.6"
+        />
+        <path
+          d="M22 30 L30 36 L22 42"
+          stroke="currentColor"
+          strokeWidth="2.5"
+          strokeLinecap="round"
+          strokeLinejoin="round"
+          opacity="0.7"
+        />
+        <path
+          d="M34 44 L44 44"
+          stroke="currentColor"
+          strokeWidth="2.5"
+          strokeLinecap="round"
+          opacity="0.7"
+        />
+        {/* Diagonal cancel slash */}
+        <path
+          d="M14 14 L58 58"
+          stroke="currentColor"
+          strokeWidth="3"
+          strokeLinecap="round"
+        />
+      </svg>
+      <h3 className="text-sm font-medium text-ink mb-1.5">Terminal not available</h3>
+      <p className="text-[11px] text-ink-soft max-w-xs leading-relaxed">
+        This workspace runs the{" "}
+        <span className="font-mono text-ink-mid">{runtime}</span> runtime,
+        which doesn't expose a shell. Use the Chat tab to interact with the
+        agent directly.
+      </p>
+    </div>
+  );
+}
+
+/** Runtimes that don't expose a TTY. Keep narrow — only add a runtime
+ *  here when its provisioner genuinely has no shell endpoint, otherwise
+ *  the user loses access to a real debugging surface. */
+const RUNTIMES_WITHOUT_TERMINAL = new Set(["external"]);
+
+export function TerminalTab({ workspaceId, data }: Props) {
+  // Early-return for runtimes that have no shell. Skips the entire
+  // xterm + WebSocket dance below — without this, mounting the tab
+  // for an external workspace pops the WS, gets a 404 from the
+  // workspace-server (no /ws/terminal/<id> route registered for it),
+  // and shows "Connection failed" with a Reconnect button — confusing
+  // because the workspace IS healthy, just doesn't have a TTY.
+  if (data && RUNTIMES_WITHOUT_TERMINAL.has(data.runtime)) {
+    return <NotAvailablePanel runtime={data.runtime} />;
+  }
+
  const containerRef = useRef<HTMLDivElement>(null);
  const termRef = useRef<{ dispose: () => void } | null>(null);
  const wsRef = useRef<WebSocket | null>(null);
@@ -0,0 +1,141 @@
+// @vitest-environment jsdom
+//
+// Pins two regressions reported on production 2026-05-05:
+//
+// 1. IME composition + Enter key: typing Chinese (or any CJK / IME-
+//    composed text) and pressing Enter to commit the candidate
+//    selection used to send the half-typed message. The fix checks
+//    `event.nativeEvent.isComposing` (and a `keyCode === 229`
+//    fallback for older WebKit) before treating Enter as send.
+//
+// 2. Markdown link clicks: the agent's ReactMarkdown-rendered links
+//    used to:
+//       - http/https → navigate canvas tab away (user lost canvas state)
+//       - workspace://path / file:///workspace/... / /workspace/... →
+//         browser hit about:blank (unhandled protocol).
+//    Fix: external links get target="_blank" + noopener; in-container
+//    paths route through downloadChatFile (same auth path as chips).
+
+import { describe, it, expect, vi, afterEach, beforeEach } from "vitest";
+import { render, screen, cleanup, fireEvent, waitFor } from "@testing-library/react";
+import React from "react";
+
+afterEach(cleanup);
+
+// Mock the api module so render doesn't try to talk to a real CP.
+const apiGet = vi.fn((_path: string): Promise<unknown> => Promise.resolve([]));
+const apiPost = vi.fn((_path: string, _body: unknown): Promise<unknown> => Promise.resolve({}));
+vi.mock("@/lib/api", () => ({
+  api: {
+    get: (path: string) => apiGet(path),
+    post: (path: string, body: unknown) => apiPost(path, body),
+    del: vi.fn(),
+    patch: vi.fn(),
+    put: vi.fn(),
+  },
+}));
+
+vi.mock("@/store/canvas", () => ({
+  useCanvasStore: vi.fn((selector?: (s: unknown) => unknown) =>
+    selector ? selector({ agentMessages: {}, consumeAgentMessages: () => [] }) : {},
+  ),
+}));
+
+// Capture the downloadChatFile call so the markdown-link test can
+// assert in-container paths route through the authenticated download
+// path rather than the browser's bare anchor click.
+const downloadChatFileMock = vi.fn((_workspaceId: string, _att: { uri: string; name: string }) => Promise.resolve());
+vi.mock("../chat/uploads", async () => {
+  const actual = await vi.importActual<typeof import("../chat/uploads")>("../chat/uploads");
+  return {
+    ...actual,
+    downloadChatFile: (workspaceId: string, att: { uri: string; name: string }) =>
+      downloadChatFileMock(workspaceId, att),
+  };
+});
+
+beforeEach(() => {
+  apiGet.mockClear();
+  apiPost.mockClear();
+  downloadChatFileMock.mockClear();
+  // jsdom doesn't implement scrollIntoView; ChatTab calls it after
+  // every render with a new message.
+  Element.prototype.scrollIntoView = vi.fn();
+  // Stub IntersectionObserver — the lazy-history sentinel uses it.
+  class FakeIO {
+    observe() {}
+    unobserve() {}
+    disconnect() {}
+  }
+  (window as unknown as { IntersectionObserver: unknown }).IntersectionObserver = FakeIO;
+  (globalThis as unknown as { IntersectionObserver: unknown }).IntersectionObserver = FakeIO;
+});
+
+import { ChatTab } from "../ChatTab";
+
+const minimalData = {
+  status: "online" as const,
+  runtime: "claude-code",
+  currentTask: null,
+} as unknown as Parameters<typeof ChatTab>[0]["data"];
+
+describe("ChatTab — IME-safe Enter key", () => {
+  it("does NOT send the message when Enter fires during IME composition (isComposing)", async () => {
+    render(<ChatTab workspaceId="ws-ime" data={minimalData} />);
+
+    // Find the textarea by its aria-label.
+    const textarea = await screen.findByLabelText(/Message to agent/i);
+    fireEvent.change(textarea, { target: { value: "你好" } });
+
+    // Simulate the Enter that commits an IME selection: isComposing=true.
+    fireEvent.keyDown(textarea, { key: "Enter", isComposing: true });
+
+    // sendMessage POSTs via api.post; assert it was NOT called.
+    await waitFor(() => {
+      expect(apiPost).not.toHaveBeenCalled();
+    });
+    // And the input is preserved — ChatTab clears it only on actual send.
+    expect((textarea as HTMLTextAreaElement).value).toBe("你好");
+  });
+
+  it("does NOT send when keyCode is 229 (older Safari IME fallback)", async () => {
+    render(<ChatTab workspaceId="ws-ime2" data={minimalData} />);
+    const textarea = await screen.findByLabelText(/Message to agent/i);
+    fireEvent.change(textarea, { target: { value: "한국어" } });
+
+    // keyCode 229 is the older-Safari signal that an IME is composing.
+    // Some mobile WebKit-based browsers delay setting isComposing on
+    // the composition-end Enter; the keyCode fallback covers that.
+    fireEvent.keyDown(textarea, { key: "Enter", keyCode: 229 });
+
+    await waitFor(() => {
+      expect(apiPost).not.toHaveBeenCalled();
+    });
+  });
+
+  it("DOES send on a non-composing Enter (the happy path stays intact)", async () => {
+    render(<ChatTab workspaceId="ws-ok" data={minimalData} />);
+    const textarea = await screen.findByLabelText(/Message to agent/i);
+    fireEvent.change(textarea, { target: { value: "hello world" } });
+
+    fireEvent.keyDown(textarea, { key: "Enter" /* no isComposing, no 229 */ });
+
+    // The api.post for /a2a fires inside sendMessage. waitFor since
+    // the call goes through several effects.
+    await waitFor(() => {
+      expect(apiPost).toHaveBeenCalled();
+    });
+  });
+
+  it("Shift+Enter inserts newline regardless (no send)", async () => {
+    render(<ChatTab workspaceId="ws-shift" data={minimalData} />);
+    const textarea = await screen.findByLabelText(/Message to agent/i);
+    fireEvent.change(textarea, { target: { value: "line 1" } });
+
+    fireEvent.keyDown(textarea, { key: "Enter", shiftKey: true });
+
+    await waitFor(() => {
+      expect(apiPost).not.toHaveBeenCalled();
+    });
+  });
+});
@@ -0,0 +1,339 @@
+// @vitest-environment jsdom
+//
+// Pins the lazy-loading chat-history pagination.
+//
+// PR-C-2 (RFC #2945): canvas was migrated from /activity?type=a2a_receive
+// to /chat-history. Server now returns typed ChatMessage[] in
+// display-ready oldest-first order. These tests guard the canvas-side
+// pagination invariants against the new endpoint surface.
+//
+// Pinned branches:
+//   1. Initial fetch carries `limit=10` and NO before_ts (newest-first
+//      slice). Pre-fix this was limit=50.
+//   2. Server returning fewer than `limit` rows clears `hasMore` so the
+//      top sentinel is removed and the IO observer disconnects — no
+//      "Loading older messages…" spinner on a short conversation.
+//   3. Server returning exactly `limit` rows on the first batch keeps
+//      hasMore=true so the sentinel mounts (verified indirectly by
+//      asserting the rendered bubble count matches the full page).
+//   4. The retry button after a failed initial load uses the same
+//      INITIAL_HISTORY_LIMIT (10), not the legacy 50.
+//   5. before_ts cursor is the OLDEST timestamp from the current page,
+//      passed verbatim to walk backward.
+//   6. Inflight guard rejects duplicate IO triggers while a loadOlder
+//      fetch is in flight.
+
+import { describe, it, expect, vi, afterEach, beforeEach } from "vitest";
+import { render, screen, cleanup, waitFor, fireEvent } from "@testing-library/react";
+import React from "react";
+
+afterEach(cleanup);
+
+// Both ChatTab sub-panels (MyChat + AgentComms) mount simultaneously so
+// keyboard tab order and aria-controls land on a real DOM. MyChat's
+// loadMessagesFromDB hits /chat-history; AgentComms's polling hits a
+// different URL. Route the mock by URL so each gets a sensible default
+// and only MyChat's calls land in the assertion array.
+const myChatHistoryCalls: string[] = [];
+let myChatNextResponse:
+  | { ok: true; messages: unknown[]; reachedEnd?: boolean }
+  | { ok: false; err: Error } = { ok: true, messages: [] };
+
+const apiGet = vi.fn((path: string): Promise<unknown> => {
+  if (path.includes("/chat-history")) {
+    myChatHistoryCalls.push(path);
+    if (myChatNextResponse.ok) {
+      const reached_end =
+        myChatNextResponse.reachedEnd !== undefined
+          ? myChatNextResponse.reachedEnd
+          : myChatNextResponse.messages.length < 10;
+      return Promise.resolve({
+        messages: myChatNextResponse.messages,
+        reached_end,
+      });
+    }
+    return Promise.reject(myChatNextResponse.err);
+  }
+  // AgentComms / heartbeat / anything else — empty array safe default.
+  return Promise.resolve([]);
+});
+const apiPost = vi.fn();
+vi.mock("@/lib/api", () => ({
+  api: {
+    get: (path: string) => apiGet(path),
+    post: (path: string, body: unknown) => apiPost(path, body),
+    del: vi.fn(),
+    patch: vi.fn(),
+    put: vi.fn(),
+  },
+}));
+
+vi.mock("@/store/canvas", () => ({
+  useCanvasStore: vi.fn((selector?: (s: unknown) => unknown) =>
+    selector ? selector({ agentMessages: {}, consumeAgentMessages: () => [] }) : {},
+  ),
+}));
+
+// Capture IntersectionObserver instances so tests can drive callbacks
+// directly (jsdom has no layout, so nothing crosses thresholds on its
+// own) AND assert observer-instance count to pin the perf invariant
+// that live-message churn doesn't tear down + re-arm the observer.
+type IOInstance = {
+  callback: IntersectionObserverCallback;
+  observed: Element[];
+  disconnected: boolean;
+};
+const ioInstances: IOInstance[] = [];
+
+beforeEach(() => {
+  apiGet.mockClear();
+  apiPost.mockReset();
+  myChatHistoryCalls.length = 0;
+  myChatNextResponse = { ok: true, messages: [] };
+  ioInstances.length = 0;
+  class FakeIO {
+    private inst: IOInstance;
+    constructor(cb: IntersectionObserverCallback) {
+      this.inst = { callback: cb, observed: [], disconnected: false };
+      ioInstances.push(this.inst);
+    }
+    observe(el: Element) {
+      this.inst.observed.push(el);
+    }
+    unobserve() {}
+    disconnect() {
+      this.inst.disconnected = true;
+    }
+  }
+  (window as unknown as { IntersectionObserver: unknown }).IntersectionObserver = FakeIO;
+  (globalThis as unknown as { IntersectionObserver: unknown }).IntersectionObserver = FakeIO;
+  Element.prototype.scrollIntoView = vi.fn();
+});
+
+function triggerIntersection(instanceIdx = -1) {
+  const inst = ioInstances.at(instanceIdx);
+  if (!inst) throw new Error(`no IO instance at ${instanceIdx}`);
+  inst.callback(
+    [{ isIntersecting: true, target: inst.observed[0] } as IntersectionObserverEntry],
+    inst as unknown as IntersectionObserver,
+  );
+}
+
+import { ChatTab } from "../ChatTab";
+
+// makeMessagePair returns a (user, agent) pair sharing a timestamp,
+// matching the wire shape /chat-history emits per activity_logs row.
+// Server-side reverseRowChunks ensures the wire is oldest-first across
+// rows but [user, agent] within each row.
+function makeMessagePair(seq: number): unknown[] {
+  // Zero-pad seq into the minute slot so seq=10 produces a valid
+  // timestamp (00:10:00Z, not 00:010:00Z).
+  const mm = String(seq).padStart(2, "0");
+  const ts = `2026-05-05T00:${mm}:00Z`;
+  return [
+    { id: `u-${seq}`, role: "user", content: `user msg ${seq}`, timestamp: ts },
+    { id: `a-${seq}`, role: "agent", content: `agent reply ${seq}`, timestamp: ts },
+  ];
+}
+
+// pageOldestFirst builds a wire-shape page (oldest-first within page)
+// of `count` row-pairs starting at seq=`start`. Mirrors the server's
+// post-reverseRowChunks emission order.
+function pageOldestFirst(start: number, count: number): unknown[] {
+  const out: unknown[] = [];
+  for (let i = 0; i < count; i++) {
+    out.push(...makeMessagePair(start + i));
+  }
+  return out;
+}
+
+const minimalData = {
+  status: "online" as const,
+  runtime: "claude-code",
+  currentTask: null,
+} as unknown as Parameters<typeof ChatTab>[0]["data"];
+
+describe("ChatTab lazy history pagination", () => {
+  it("initial fetch carries limit=10 (not the legacy 50) and hits /chat-history", async () => {
+    myChatNextResponse = { ok: true, messages: makeMessagePair(1) };
+    render(<ChatTab workspaceId="ws-1" data={minimalData} />);
+    await waitFor(() => expect(myChatHistoryCalls.length).toBe(1));
+    const url = myChatHistoryCalls[0];
+    expect(url).toContain("/chat-history");
+    expect(url).toContain("limit=10");
+    expect(url).not.toContain("limit=50");
+    // before_ts should NOT be set on the initial fetch — that's the
+    // newest-first slice the user lands on.
+    expect(url).not.toContain("before_ts");
+    // /chat-history filters source-canvas server-side; client should
+    // NOT pass type/source params (they belonged to /activity).
+    expect(url).not.toContain("type=a2a_receive");
+    expect(url).not.toContain("source=canvas");
+  });
+
+  it("hides the top sentinel when initial fetch returns fewer than the limit", async () => {
+    // 3 < 10 → server says "no more older history exists"; sentinel
+    // should NOT mount and the "Loading older messages…" line should
+    // never appear.
+    myChatNextResponse = { ok: true, messages: pageOldestFirst(1, 3) };
+    render(<ChatTab workspaceId="ws-2" data={minimalData} />);
+    await waitFor(() => expect(myChatHistoryCalls.length).toBe(1));
+    await waitFor(() => {
+      expect(screen.queryByText(/Loading chat history/i)).toBeNull();
+    });
+    expect(screen.queryByText(/Loading older messages/i)).toBeNull();
+  });
+
+  it("renders all messages when initial fetch returns exactly the limit", async () => {
+    // limit=10 row-pairs → 20 ChatMessages. reachedEnd should be FALSE
+    // so the sentinel mounts. Verified by bubble counts.
+    myChatNextResponse = {
+      ok: true,
+      messages: pageOldestFirst(1, 10),
+      reachedEnd: false,
+    };
+    render(<ChatTab workspaceId="ws-3" data={minimalData} />);
+    await waitFor(() => expect(myChatHistoryCalls.length).toBe(1));
+    await waitFor(() => {
+      expect(screen.queryByText(/Loading chat history/i)).toBeNull();
+    });
+    expect(screen.getAllByText(/user msg/).length).toBe(10);
+    expect(screen.getAllByText(/agent reply/).length).toBe(10);
+  });
+
+  it("retry-after-failure uses limit=10, not the legacy 50", async () => {
+    myChatNextResponse = { ok: false, err: new Error("network down") };
+    render(<ChatTab workspaceId="ws-4" data={minimalData} />);
+    const retry = await screen.findByText(/Retry/);
+    myChatNextResponse = { ok: true, messages: makeMessagePair(1) };
+    fireEvent.click(retry);
+    await waitFor(() => expect(myChatHistoryCalls.length).toBe(2));
+    const retryUrl = myChatHistoryCalls[1];
+    expect(retryUrl).toContain("/chat-history");
+    expect(retryUrl).toContain("limit=10");
+    expect(retryUrl).not.toContain("limit=50");
+  });
+
+  it("loadOlder fetches limit=20 with before_ts=oldest.timestamp", async () => {
+    // Initial page = 10 row-pairs in oldest-first order (seq 1..10).
+    // The oldest (and so the cursor for loadOlder) is seq=1's
+    // timestamp 2026-05-05T00:01:00Z.
+    myChatNextResponse = {
+      ok: true,
+      messages: pageOldestFirst(1, 10),
+      reachedEnd: false,
+    };
+    render(<ChatTab workspaceId="ws-load-older" data={minimalData} />);
+    await waitFor(() => expect(myChatHistoryCalls.length).toBe(1));
+    await waitFor(() => expect(ioInstances.length).toBeGreaterThan(0));
+
+    // Stage older-batch response, then fire IO callback.
+    myChatNextResponse = {
+      ok: true,
+      messages: pageOldestFirst(0, 1),
+      reachedEnd: true,
+    };
+    triggerIntersection();
+
+    await waitFor(() => expect(myChatHistoryCalls.length).toBe(2));
+    const olderUrl = myChatHistoryCalls[1];
+    expect(olderUrl).toContain("/chat-history");
+    expect(olderUrl).toContain("limit=20");
+    expect(olderUrl).toContain("before_ts=");
+    expect(decodeURIComponent(olderUrl)).toContain("before_ts=2026-05-05T00:01:00Z");
+  });
+
+  it("inflight guard rejects a second IO trigger while first loadOlder is in flight", async () => {
+    myChatNextResponse = {
+      ok: true,
+      messages: pageOldestFirst(1, 10),
+      reachedEnd: false,
+    };
+    render(<ChatTab workspaceId="ws-inflight" data={minimalData} />);
+    await waitFor(() => expect(myChatHistoryCalls.length).toBe(1));
+    await waitFor(() => expect(ioInstances.length).toBeGreaterThan(0));
+
+    // Hold the next loadOlder fetch open with a manual deferred so we
+    // can fire the second trigger while the first is in-flight.
+    let release!: (resp: unknown) => void;
+    const deferred = new Promise<unknown>((res) => {
+      release = res;
+    });
+    apiGet.mockImplementationOnce((path: string): Promise<unknown> => {
+      myChatHistoryCalls.push(path);
+      return deferred;
+    });
+
+    triggerIntersection(); // start loadOlder #1
+    await waitFor(() => expect(myChatHistoryCalls.length).toBe(2));
+
+    // Second IO trigger lands while #1 is still pending.
+    triggerIntersection();
+    triggerIntersection();
+    triggerIntersection();
+    // Without the inflight guard, each of these would have started a
+    // new fetch. With the guard, none of them do — call count stays 2.
+    await new Promise((r) => setTimeout(r, 10));
+    expect(myChatHistoryCalls.length).toBe(2);
+
+    // Release the first fetch with a valid wire response shape.
+    release({ messages: [], reached_end: true });
+    await waitFor(() => expect(myChatHistoryCalls.length).toBe(2));
+  });
+
+  it("empty older response clears the scroll anchor and unmounts the sentinel", async () => {
+    myChatNextResponse = {
+      ok: true,
+      messages: pageOldestFirst(1, 10),
+      reachedEnd: false,
+    };
+    render(<ChatTab workspaceId="ws-anchor" data={minimalData} />);
+    await waitFor(() => expect(myChatHistoryCalls.length).toBe(1));
+    await waitFor(() => expect(ioInstances.length).toBeGreaterThan(0));
+
+    myChatNextResponse = {
+      ok: true,
+      messages: [],
+      reachedEnd: true,
+    };
+    triggerIntersection();
+    await waitFor(() => expect(myChatHistoryCalls.length).toBe(2));
+
+    await waitFor(() => {
+      expect(screen.queryByText(/Loading older messages/i)).toBeNull();
+    });
+  });
+
+  it("IntersectionObserver does not churn when older messages prepend", async () => {
+    myChatNextResponse = {
+      ok: true,
+      messages: pageOldestFirst(1, 10),
+      reachedEnd: false,
+    };
+    render(<ChatTab workspaceId="ws-stable-io" data={minimalData} />);
+    await waitFor(() => expect(myChatHistoryCalls.length).toBe(1));
+    await waitFor(() => expect(ioInstances.length).toBeGreaterThan(0));
+
+    const observerBefore = ioInstances.at(-1);
+    expect(observerBefore).toBeDefined();
+    expect(observerBefore!.disconnected).toBe(false);
+
+    // Trigger three older-batch prepends. Each batch returns the full
+    // OLDER_HISTORY_BATCH (20 row-pairs = 40 messages) so reachedEnd
+    // stays false and the sentinel keeps mounting.
+    for (let batch = 0; batch < 3; batch++) {
+      myChatNextResponse = {
+        ok: true,
+        messages: pageOldestFirst(-(batch + 1) * 20, 20),
+        reachedEnd: false,
+      };
+      const callsBefore = myChatHistoryCalls.length;
+      triggerIntersection();
+      await waitFor(() => expect(myChatHistoryCalls.length).toBe(callsBefore + 1));
+    }
+
+    // The original observer is still the live one — no churn.
+    expect(observerBefore!.disconnected).toBe(false);
+    expect(ioInstances.at(-1)).toBe(observerBefore);
+  });
+});
@@ -0,0 +1,125 @@
+// @vitest-environment jsdom
+//
+// Regression tests for the ConfigTab section restructure (user feedback
+// 2026-05-04: "Skills and Tools are having their own tab as plugin, and
+// Prompt Files are in the file system which can be directly edited. Am
+// I missing something?" + "Tools should be merged into plugin then, and
+// for prompt files... should be in another section than in skill& tools").
+//
+// What this pins:
+//   1. The "Skills & Tools" section title is gone.
+//   2. Editable Skills + Tools tag inputs are gone (managed elsewhere).
+//   3. A dedicated "Prompt Files" section exists with explanatory text.
+//
+// If a future PR re-adds the Skills/Tools tag inputs to ConfigTab, this
+// suite catches it.
+
+import { describe, it, expect, vi, afterEach, beforeEach } from "vitest";
+import { render, screen, cleanup, waitFor, fireEvent } from "@testing-library/react";
+import React from "react";
+
+afterEach(cleanup);
+
+const apiGet = vi.fn();
+vi.mock("@/lib/api", () => ({
+  api: {
+    get: (path: string) => apiGet(path),
+    patch: vi.fn(),
+    put: vi.fn(),
+    post: vi.fn(),
+    del: vi.fn(),
+  },
+}));
+
+const storeUpdateNodeData = vi.fn();
+const storeRestartWorkspace = vi.fn();
+vi.mock("@/store/canvas", () => ({
+  useCanvasStore: Object.assign(
+    (selector: (s: unknown) => unknown) =>
+      selector({ restartWorkspace: storeRestartWorkspace, updateNodeData: storeUpdateNodeData }),
+    {
+      getState: () => ({
+        restartWorkspace: storeRestartWorkspace,
+        updateNodeData: storeUpdateNodeData,
+      }),
+    },
+  ),
+}));
+
+vi.mock("../AgentCardSection", () => ({
+  AgentCardSection: () => <div data-testid="agent-card-stub" />,
+}));
+
+import { ConfigTab } from "../ConfigTab";
+
+beforeEach(() => {
+  apiGet.mockReset();
+  apiGet.mockImplementation((path: string) => {
+    if (path === `/workspaces/ws-test`) {
+      return Promise.resolve({ runtime: "claude-code" });
+    }
+    if (path === `/workspaces/ws-test/model`) {
+      return Promise.resolve({ model: "claude-opus-4-7" });
+    }
+    if (path === `/workspaces/ws-test/provider`) {
+      return Promise.resolve({ provider: "anthropic-oauth", source: "default" });
+    }
+    if (path === `/workspaces/ws-test/files/config.yaml`) {
+      return Promise.resolve({ content: "name: test\nruntime: claude-code\n" });
+    }
+    if (path === "/templates") {
+      return Promise.resolve([
+        { id: "claude-code", name: "Claude Code", runtime: "claude-code", providers: [] },
+      ]);
+    }
+    return Promise.reject(new Error(`unmocked api.get: ${path}`));
+  });
+});
+
+describe("ConfigTab section restructure", () => {
+  it("does not render a 'Skills & Tools' section title", async () => {
+    render(<ConfigTab workspaceId="ws-test" />);
+    await waitFor(() => expect(apiGet).toHaveBeenCalled());
+    // Section button uses the title as its accessible name; should be absent.
+    expect(screen.queryByRole("button", { name: /Skills\s*&\s*Tools/i })).toBeNull();
+  });
+
+  it("does not render an editable Skills tag input", async () => {
+    render(<ConfigTab workspaceId="ws-test" />);
+    await waitFor(() => expect(apiGet).toHaveBeenCalled());
+    // TagList renders its label; check no input labelled "Skills" in the form.
+    // (Skills are managed via the dedicated Skills tab.)
+    const skillsLabels = screen
+      .queryAllByText(/^Skills$/)
+      .filter((el) => el.tagName.toLowerCase() === "label");
+    expect(skillsLabels).toHaveLength(0);
+  });
+
+  it("does not render an editable Tools tag input", async () => {
+    render(<ConfigTab workspaceId="ws-test" />);
+    await waitFor(() => expect(apiGet).toHaveBeenCalled());
+    // Tools are managed via the Plugins tab — install a plugin → its tools
+    // become available. No reason to type tool names here.
+    const toolsLabels = screen
+      .queryAllByText(/^Tools$/)
+      .filter((el) => el.tagName.toLowerCase() === "label");
+    expect(toolsLabels).toHaveLength(0);
+  });
+
+  it("renders a dedicated 'Prompt Files' section with explanatory copy", async () => {
+    render(<ConfigTab workspaceId="ws-test" />);
+    await waitFor(() => expect(apiGet).toHaveBeenCalled());
+    // Section is collapsed by default — find + expand first.
+    const sectionButton = screen.getByRole("button", { name: /Prompt Files/i });
+    expect(sectionButton).toBeTruthy();
+    fireEvent.click(sectionButton);
+    // Explanatory copy mentions system-prompt.md (split across <code> tags
+    // so use textContent on any element rather than the default text matcher).
+    await waitFor(() => {
+      const matches = screen.queryAllByText((_, el) =>
+        (el?.textContent || "").includes("system-prompt.md"),
+      );
+      expect(matches.length).toBeGreaterThan(0);
+    });
+  });
+});
@@ -0,0 +1,156 @@
+// @vitest-environment jsdom
+//
+// ExternalConnectionSection — coverage for the credential-rotate +
+// re-show-instructions UI on the Config tab.
+//
+// What this pins:
+//   1. "Show connection info" → GET /external/connection, opens modal
+//      with auth_token=""
+//   2. "Rotate credentials" → confirm dialog → POST /external/rotate,
+//      opens modal with the returned auth_token
+//   3. Confirm dialog cancels without firing the POST
+//   4. API failure surfaces an error chip (no silent loss)
+
+import { describe, it, expect, vi, afterEach, beforeEach } from "vitest";
+import {
+  render,
+  screen,
+  cleanup,
+  fireEvent,
+  waitFor,
+} from "@testing-library/react";
+import React from "react";
+
+afterEach(cleanup);
+
+const apiGet = vi.fn();
+const apiPost = vi.fn();
+vi.mock("@/lib/api", () => ({
+  api: {
+    get: (path: string) => apiGet(path),
+    post: (path: string, body?: unknown) => apiPost(path, body),
+    patch: vi.fn(),
+    put: vi.fn(),
+    del: vi.fn(),
+  },
+}));
+
+import { ExternalConnectionSection } from "../ExternalConnectionSection";
+
+beforeEach(() => {
+  apiGet.mockReset();
+  apiPost.mockReset();
+});
+
+const SAMPLE_INFO = {
+  workspace_id: "ws-test",
+  platform_url: "https://platform.example.test",
+  auth_token: "",
+  registry_endpoint: "https://platform.example.test/registry/register",
+  heartbeat_endpoint: "https://platform.example.test/registry/heartbeat",
+  // The modal stamps these snippets server-side; for the test we
+  // bake workspace_id into one so the rendered DOM contains a
+  // findable token after the modal mounts.
+  curl_register_template: "# curl ws=ws-test",
+  python_snippet: "# py ws=ws-test",
+  claude_code_channel_snippet: "# claude ws=ws-test",
+  universal_mcp_snippet: "# mcp ws=ws-test",
+  hermes_channel_snippet: "# hermes ws=ws-test",
+  codex_snippet: "# codex ws=ws-test",
+  openclaw_snippet: "# openclaw ws=ws-test",
+};
+
+describe("ExternalConnectionSection", () => {
+  it("renders both action buttons", () => {
+    render(<ExternalConnectionSection workspaceId="ws-test" />);
+    expect(screen.getByRole("button", { name: /show connection info/i })).toBeTruthy();
+    expect(screen.getByRole("button", { name: /rotate credentials/i })).toBeTruthy();
+  });
+
+  it("'Show connection info' calls GET /external/connection and opens modal with blank token", async () => {
+    apiGet.mockResolvedValue({ connection: { ...SAMPLE_INFO, auth_token: "" } });
+    render(<ExternalConnectionSection workspaceId="ws-test" />);
+
+    fireEvent.click(screen.getByRole("button", { name: /show connection info/i }));
+
+    await waitFor(() =>
+      expect(apiGet).toHaveBeenCalledWith("/workspaces/ws-test/external/connection"),
+    );
+    // The ExternalConnectModal renders the workspace_id field in its
+    // copy-block. document.body covers Radix's portal mount point.
+    await waitFor(() => {
+      expect(document.body.textContent || "").toContain("ws-test");
+    });
+  });
+
+  it("'Rotate credentials' opens confirm dialog before firing POST", async () => {
+    render(<ExternalConnectionSection workspaceId="ws-test" />);
+    fireEvent.click(screen.getByRole("button", { name: /rotate credentials/i }));
+
+    // Confirm dialog appears with the destructive copy.
+    await waitFor(() => {
+      expect(
+        screen.getByText(/Rotate workspace credentials\?/i),
+      ).toBeTruthy();
+    });
+    expect(screen.getByText(/immediately invalidate the current one/i)).toBeTruthy();
+
+    // POST must NOT have fired yet — only on confirm.
+    expect(apiPost).not.toHaveBeenCalled();
+  });
+
+  it("Cancel in confirm dialog dismisses without rotating", async () => {
+    render(<ExternalConnectionSection workspaceId="ws-test" />);
+    fireEvent.click(screen.getByRole("button", { name: /rotate credentials/i }));
+
+    await waitFor(() =>
+      expect(screen.getByText(/Rotate workspace credentials\?/i)).toBeTruthy(),
+    );
+    fireEvent.click(screen.getByRole("button", { name: /^cancel$/i }));
+
+    await waitFor(() =>
+      expect(screen.queryByText(/Rotate workspace credentials\?/i)).toBeNull(),
+    );
+    expect(apiPost).not.toHaveBeenCalled();
+  });
+
+  it("Confirm in dialog POSTs to /external/rotate and opens modal with returned token", async () => {
+    apiPost.mockResolvedValue({
+      connection: { ...SAMPLE_INFO, auth_token: "fresh-tok-123" },
+    });
+    render(<ExternalConnectionSection workspaceId="ws-test" />);
+
+    fireEvent.click(screen.getByRole("button", { name: /rotate credentials/i }));
+    await waitFor(() =>
+      expect(screen.getByText(/Rotate workspace credentials\?/i)).toBeTruthy(),
+    );
+    // Click the dialog's Rotate button (NOT the section's — the section's
+    // "Rotate credentials" stays mounted; the dialog's "Rotate" is the
+    // commit button. getAllByRole returns both; pick the one inside the
+    // dialog by name "Rotate" exact-match).
+    const rotateBtns = screen.getAllByRole("button", { name: /^rotate$/i });
+    expect(rotateBtns.length).toBeGreaterThanOrEqual(1);
+    fireEvent.click(rotateBtns[rotateBtns.length - 1]);
+
+    await waitFor(() =>
+      expect(apiPost).toHaveBeenCalledWith(
+        "/workspaces/ws-test/external/rotate",
+        {},
+      ),
+    );
+  });
+
+  it("Surfaces API errors as a visible chip, not silent loss", async () => {
+    apiGet.mockRejectedValue(new Error("forbidden"));
+    render(<ExternalConnectionSection workspaceId="ws-test" />);
+
+    fireEvent.click(screen.getByRole("button", { name: /show connection info/i }));
+
+    await waitFor(() => {
+      const matches = screen.queryAllByText((_, el) =>
+        (el?.textContent || "").toLowerCase().includes("forbidden"),
+      );
+      expect(matches.length).toBeGreaterThan(0);
+    });
+  });
+});
@@ -0,0 +1,119 @@
+// @vitest-environment jsdom
+//
+// Pins the "Files not available" early-return for runtimes whose
+// filesystem the platform doesn't own (today: runtime === "external").
+//
+// Pre-fix: FilesTab issued a GET /workspaces/<id>/files for every
+// workspace. The platform's response for an external workspace is
+// always [] (no rows in workspace_files), but the canvas rendered
+// "0 files / No config files yet" — visually identical to the SaaS
+// empty-listing bug fixed in PR-A. The placeholder makes the absence
+// intentional.
+//
+// Pinned branches:
+//   1. external runtime → "Files not available" banner renders,
+//      runtime name surfaces in the body so user knows WHY.
+//   2. external runtime → useFilesApi is NOT invoked. Verified by
+//      asserting the mocked api.get was never called.
+//   3. claude-code (or any other runtime) → no banner, normal mount
+//      proceeds (`/configs` toolbar visible). Pre-fix regression cover.
+//   4. data prop omitted (legacy callers) → no early-return, falls
+//      through to normal mount.
+
+import { describe, it, expect, vi, beforeEach, afterEach } from "vitest";
+import { render, screen, cleanup, waitFor } from "@testing-library/react";
+import React from "react";
+
+afterEach(cleanup);
+
+// Mock the api module so the normal-mount branches don't try to
+// fetch against a real backend — and so we can assert the
+// external-runtime branch never fires a request.
+const apiCalls: string[] = [];
+vi.mock("@/lib/api", () => ({
+  api: {
+    get: vi.fn((path: string) => {
+      apiCalls.push(path);
+      return Promise.resolve([]);
+    }),
+    put: vi.fn(() => Promise.resolve()),
+    del: vi.fn(() => Promise.resolve()),
+  },
+}));
+
+// useCanvasStore is referenced by useFilesApi for the needsRestart
+// flag. The Toaster import inside FilesTab also pulls the store
+// indirectly. Stub minimally to satisfy the import chain.
+vi.mock("@/store/canvas", async () => {
+  const actual = await vi.importActual<typeof import("@/store/canvas")>(
+    "@/store/canvas",
+  );
+  return {
+    ...actual,
+    useCanvasStore: {
+      getState: () => ({
+        updateNodeData: vi.fn(),
+      }),
+    },
+  };
+});
+
+vi.mock("../Toaster", () => ({
+  showToast: vi.fn(),
+}));
+
+beforeEach(() => {
+  apiCalls.length = 0;
+});
+
+import { FilesTab } from "../FilesTab";
+
+const externalData = { runtime: "external", status: "online" } as unknown as Parameters<
+  typeof FilesTab
+>[0]["data"];
+
+const claudeData = { runtime: "claude-code", status: "online" } as unknown as Parameters<
+  typeof FilesTab
+>[0]["data"];
+
+describe("FilesTab not-available early-return for runtimes without platform-owned filesystem", () => {
+  it("external runtime renders the not-available banner with runtime name", () => {
+    render(<FilesTab workspaceId="ws-ext" data={externalData} />);
+    expect(screen.getByText(/Files not available/i)).not.toBeNull();
+    // Runtime name must surface so the user understands WHY — without
+    // it the placeholder reads as a generic error.
+    expect(screen.getByText(/external/)).not.toBeNull();
+    // Chat tab is the recommended alternative — flagged in copy so the
+    // user knows where to go next instead of bouncing tabs.
+    expect(screen.getByText(/Chat tab/i)).not.toBeNull();
+  });
+
+  it("external runtime does NOT issue any /files API call", async () => {
+    render(<FilesTab workspaceId="ws-ext" data={externalData} />);
+    // Tolerate one microtask boundary in case useEffect schedules.
+    await new Promise((r) => setTimeout(r, 0));
+    const filesCalls = apiCalls.filter((p) => p.includes("/files"));
+    expect(filesCalls).toEqual([]);
+  });
+
+  it("claude-code runtime does NOT render the banner (normal mount)", async () => {
+    render(<FilesTab workspaceId="ws-claude" data={claudeData} />);
+    // The normal-mount path renders the FilesToolbar with the root
+    // selector. Wait for it (useEffect → loadFiles → setLoading false).
+    await waitFor(() => {
+      expect(screen.queryByText(/Files not available/i)).toBeNull();
+    });
+    // Toolbar's root selector confirms we're on the platform-owned
+    // rendering path, not the placeholder.
+    expect(screen.getByLabelText(/File root directory/i)).not.toBeNull();
+  });
+
+  it("data prop omitted falls through to normal mount (back-compat)", async () => {
+    render(<FilesTab workspaceId="ws-no-data" />);
+    await waitFor(() => {
+      expect(screen.queryByText(/Files not available/i)).toBeNull();
+    });
+    // Without data we can't gate on runtime — must mount normally.
+    expect(screen.getByLabelText(/File root directory/i)).not.toBeNull();
+  });
+});
@@ -0,0 +1,141 @@
+// @vitest-environment jsdom
+//
+// Pins the compact-when-empty layout for the SkillsTab Plugins section
+// (issue #2971, reported on production 2026-05-05).
+//
+// Three states matter for layout:
+//   1. installed.length === 0 + registry closed + load completed → COMPACT pill
+//   2. installed.length > 0  → FULL panel + installed list
+//   3. registry open (showRegistry=true) → FULL panel + registry browser
+//
+// The compact-empty path is the new behavior; the other two were
+// pre-existing. This test pins all three so a future refactor that
+// over-collapses (showing compact when plugins are installed) or
+// over-expands (showing full panel on empty load) fails loudly.
+
+import { describe, it, expect, vi, beforeEach, afterEach } from "vitest";
+import { render, screen, cleanup, fireEvent, waitFor } from "@testing-library/react";
+import React from "react";
+
+afterEach(cleanup);
+
+const apiGet = vi.fn();
+vi.mock("@/lib/api", () => ({
+  api: {
+    get: (path: string, opts?: unknown) => apiGet(path, opts),
+    post: vi.fn(() => Promise.resolve({})),
+    del: vi.fn(),
+    patch: vi.fn(),
+    put: vi.fn(),
+  },
+}));
+
+beforeEach(() => {
+  apiGet.mockReset();
+  Element.prototype.scrollIntoView = vi.fn();
+});
+
+import { SkillsTab } from "../SkillsTab";
+
+const minimalData = {
+  status: "online" as const,
+  runtime: "claude-code",
+  currentTask: "",
+  agentCard: undefined,
+} as unknown as Parameters<typeof SkillsTab>[0]["data"];
+
+describe("SkillsTab Plugins compact-empty layout", () => {
+  it("renders compact pill when installed.length === 0 and registry closed", async () => {
+    // Both fetches return empty arrays — workspace is fresh, no plugins.
+    apiGet.mockImplementation((path: string) => {
+      if (path.endsWith("/plugins") || path === "/plugins" || path === "/plugins/sources") {
+        return Promise.resolve([]);
+      }
+      return Promise.resolve([]);
+    });
+    render(<SkillsTab workspaceId="ws-fresh" data={minimalData} />);
+
+    // Wait for the installedLoaded gate to flip — without that the
+    // component renders a "loading" state, not the compact pill.
+    await waitFor(() => {
+      expect(screen.getByLabelText(/Plugins \(none installed\)/i)).toBeTruthy();
+    });
+
+    // Compact assertions: the rounded-xl panel chrome MUST NOT be in
+    // the DOM (we'd see two "Plugins" labels — one in the header,
+    // one in the pill — if the layout regressed to "always full
+    // panel"). The compact form has exactly one "Plugins" label.
+    const labels = screen.getAllByText("Plugins");
+    expect(labels).toHaveLength(1);
+
+    // The full-panel chrome's id="plugins-section" should NOT be
+    // rendered when we're in compact mode.
+    expect(document.getElementById("plugins-section")).toBeNull();
+  });
+
+  it("renders full panel when installed.length > 0", async () => {
+    apiGet.mockImplementation((path: string) => {
+      if (path.endsWith("/plugins")) {
+        return Promise.resolve([
+          { name: "memory-postgres", version: "1.0.0", description: "memory backend", supported_on_runtime: true },
+        ]);
+      }
+      return Promise.resolve([]);
+    });
+    render(<SkillsTab workspaceId="ws-installed" data={minimalData} />);
+
+    await waitFor(() => {
+      expect(screen.getByText(/1 installed/i)).toBeTruthy();
+    });
+
+    // Full-panel chrome MUST be present — id pin.
+    expect(document.getElementById("plugins-section")).not.toBeNull();
+    // Compact pill ariaLabel MUST NOT be present.
+    expect(screen.queryByLabelText(/Plugins \(none installed\)/i)).toBeNull();
+  });
+
+  it("expands to full panel when user clicks + Install Plugin from compact pill", async () => {
+    apiGet.mockImplementation(() => Promise.resolve([]));
+    render(<SkillsTab workspaceId="ws-expand" data={minimalData} />);
+
+    // Start compact — wait for the compact pill to settle so we click
+    // the right button (initial render before installedLoaded flips
+    // doesn't have either layout, and the post-load compact pill is
+    // what we want to interact with).
+    await waitFor(() => {
+      expect(screen.getByLabelText(/Plugins \(none installed\)/i)).toBeTruthy();
+    });
+    const installBtn = screen.getByRole("button", { name: /\+ Install Plugin/i });
+    expect(installBtn.getAttribute("aria-expanded")).toBe("false");
+
+    fireEvent.click(installBtn);
+
+    // After click, registry opens → full panel renders. The compact
+    // pill's aria-label should be gone; the full-panel id should
+    // appear. Generous waitFor — a registry fetch may also fire in
+    // the React effect chain, and we want to assert the compact →
+    // full transition without racing it.
+    await waitFor(
+      () => {
+        expect(document.getElementById("plugins-section")).not.toBeNull();
+      },
+      { timeout: 3000 },
+    );
+    expect(screen.queryByLabelText(/Plugins \(none installed\)/i)).toBeNull();
+  });
+
+  it("does NOT collapse to compact while initial load is pending (avoid flash)", () => {
+    // Returning a never-resolving promise means installedLoaded stays
+    // false. The compact pill MUST NOT render in this state — that
+    // would flash compact → full as the load completes, which looks
+    // janky. The component shows a loading shell instead (the
+    // existing pre-fix behavior).
+    apiGet.mockImplementation(() => new Promise(() => {}));
+    render(<SkillsTab workspaceId="ws-loading" data={minimalData} />);
+
+    // Synchronous assertion — no waitFor — since we want to confirm
+    // the compact pill is NOT rendered before any network round-trip
+    // finishes.
+    expect(screen.queryByLabelText(/Plugins \(none installed\)/i)).toBeNull();
+  });
+});
@@ -0,0 +1,107 @@
+// @vitest-environment jsdom
+//
+// Pins the "Terminal not available" early-return added 2026-05-05.
+//
+// Pre-fix: TerminalTab tried to open /ws/terminal/<id> for every
+// workspace including external runtimes (which have no shell endpoint).
+// The server returned 404, status flipped to "error", user saw
+// "Connection failed" with a Reconnect button — reading as a bug
+// when really the runtime intentionally has no TTY. Now: when
+// data.runtime is in RUNTIMES_WITHOUT_TERMINAL, render a banner +
+// big icon instead of mounting xterm/WS.
+//
+// Pinned branches:
+//   1. external runtime → "Terminal not available" banner renders,
+//      runtime name surfaces in the body so the user knows WHY.
+//   2. external runtime → xterm + WebSocket are NOT initialised.
+//      Verified by checking the global WebSocket constructor isn't
+//      called.
+//   3. claude-code (or any other runtime) → no banner, normal mount
+//      proceeds. Pre-fix regression cover.
+//   4. data prop omitted (back-compat with any caller that doesn't
+//      thread it through) → no early-return, falls through to normal
+//      mount. Tested via the absence of the banner.
+
+import { describe, it, expect, vi, beforeEach, afterEach } from "vitest";
+import { render, screen, cleanup } from "@testing-library/react";
+import React from "react";
+
+afterEach(cleanup);
+
+// xterm + addon-fit are dynamically imported by TerminalTab. Stub them
+// so the tests don't pull a 200KB+ dependency just to verify the
+// not-available banner. The stubs only matter for the non-banner
+// branches; the banner returns BEFORE the dynamic import.
+vi.mock("xterm", () => ({
+  Terminal: vi.fn().mockImplementation(() => ({
+    loadAddon: vi.fn(),
+    open: vi.fn(),
+    onData: vi.fn(),
+    write: vi.fn(),
+    dispose: vi.fn(),
+    onResize: vi.fn(),
+    cols: 80,
+    rows: 24,
+  })),
+}));
+vi.mock("@xterm/addon-fit", () => ({
+  FitAddon: vi.fn().mockImplementation(() => ({
+    fit: vi.fn(),
+  })),
+}));
+
+// Track WebSocket constructor calls — this is the load-bearing
+// assertion for "external doesn't even try to connect".
+let wsConstructed = 0;
+beforeEach(() => {
+  wsConstructed = 0;
+  (globalThis as unknown as { WebSocket: unknown }).WebSocket = vi
+    .fn()
+    .mockImplementation(() => {
+      wsConstructed++;
+      return {
+        addEventListener: vi.fn(),
+        removeEventListener: vi.fn(),
+        send: vi.fn(),
+        close: vi.fn(),
+        readyState: 0,
+      };
+    });
+});
+
+import { TerminalTab } from "../TerminalTab";
+
+const externalData = { runtime: "external", status: "online" } as unknown as Parameters<
+  typeof TerminalTab
+>[0]["data"];
+
+const claudeData = { runtime: "claude-code", status: "online" } as unknown as Parameters<
+  typeof TerminalTab
+>[0]["data"];
+
+describe("TerminalTab not-available early-return for runtimes without TTY", () => {
+  it("external runtime renders the not-available banner with runtime name", () => {
+    render(<TerminalTab workspaceId="ws-ext" data={externalData} />);
+    expect(screen.getByText(/Terminal not available/i)).not.toBeNull();
+    // Runtime name surfaces so user knows WHY there's no terminal.
+    expect(screen.getByText(/external/)).not.toBeNull();
+  });
+
+  it("external runtime does NOT open a WebSocket", async () => {
+    render(<TerminalTab workspaceId="ws-ext" data={externalData} />);
+    // Wait a tick for any deferred init (there shouldn't be any, but
+    // tolerate a microtask boundary).
+    await new Promise((r) => setTimeout(r, 0));
+    expect(wsConstructed).toBe(0);
+  });
+
+  it("claude-code runtime does NOT render the banner (normal mount)", () => {
+    render(<TerminalTab workspaceId="ws-claude" data={claudeData} />);
+    expect(screen.queryByText(/Terminal not available/i)).toBeNull();
+  });
+
+  it("data prop omitted falls through to normal mount (back-compat)", () => {
+    render(<TerminalTab workspaceId="ws-no-data" />);
+    expect(screen.queryByText(/Terminal not available/i)).toBeNull();
+  });
+});
@@ -1,6 +1,6 @@
 "use client";

-import { useState, useEffect, useMemo, useRef } from "react";
+import { useState, useEffect, useLayoutEffect, useMemo, useRef, useCallback } from "react";
 import ReactMarkdown from "react-markdown";
 import remarkGfm from "remark-gfm";
 import { api } from "@/lib/api";
@@ -184,13 +184,23 @@ function unwrapErrorText(raw: string | null): string {
 export function AgentCommsPanel({ workspaceId }: { workspaceId: string }) {
  const [messages, setMessages] = useState<CommMessage[]>([]);
  const [loading, setLoading] = useState(true);
+  const [loadError, setLoadError] = useState<string | null>(null);
  // Dedup by timestamp+type+peer to handle API load + WebSocket race
  const seenKeys = useRef(new Set<string>());
  const bottomRef = useRef<HTMLDivElement>(null);
+  // Mirrors the my-chat scroll behaviour from ChatTab (PR #2903) —
+  // smooth-scroll on a long history gets interrupted by concurrent
+  // renders and lands the panel mid-conversation. Switch the first
+  // arrival to instant; subsequent appends animate.
+  const hasInitialScrollRef = useRef(false);

-  // Load history
-  useEffect(() => {
+  // Load history. Extracted so the error-state retry button can
+  // re-invoke without remount. ChatTab uses the same shape
+  // (loadInitial → loadError state → retry button).
+  const loadInitial = useCallback(() => {
    setLoading(true);
+    setLoadError(null);
+    seenKeys.current.clear();
    api.get<ActivityEntry[]>(`/workspaces/${workspaceId}/activity?source=agent&limit=50`)
      .then((entries) => {
        const filtered = (entries ?? [])
@@ -234,10 +244,15 @@ export function AgentCommsPanel({ workspaceId }: { workspaceId: string }) {
        // the .then body) — the panel just sat on the empty state
        // with zero signal.
        console.warn("AgentCommsPanel: load activity failed", err);
+        setLoadError(err instanceof Error ? err.message : String(err));
        setLoading(false);
      });
  }, [workspaceId]);

+  useEffect(() => {
+    loadInitial();
+  }, [loadInitial]);
+
  // Live updates routed through the global ReconnectingSocket. The
  // previous pattern of `new WebSocket(WS_URL)` per panel had no
  // onclose / no reconnect, so any drop (idle timeout, browser
@@ -358,7 +373,18 @@ export function AgentCommsPanel({ workspaceId }: { workspaceId: string }) {
    } catch { /* ignore */ }
  });

-  useEffect(() => {
+  // useLayoutEffect (not useEffect) so the scroll runs BEFORE paint —
+  // otherwise the user sees the panel jump for one frame on every
+  // append. Mirrors ChatTab's MyChatPanel scroll block.
+  useLayoutEffect(() => {
+    if (!hasInitialScrollRef.current && messages.length > 0) {
+      // Instant on first arrival — smooth-scroll on a long history
+      // gets interrupted by concurrent renders and lands the panel
+      // mid-conversation (the chat-opens-in-middle bug class).
+      hasInitialScrollRef.current = true;
+      bottomRef.current?.scrollIntoView({ behavior: "instant" as ScrollBehavior });
+      return;
+    }
    bottomRef.current?.scrollIntoView({ behavior: "smooth" });
  }, [messages]);

@@ -366,6 +392,27 @@ export function AgentCommsPanel({ workspaceId }: { workspaceId: string }) {
    return <div className="text-xs text-ink-soft text-center py-8">Loading agent communications...</div>;
  }

+  if (loadError !== null && messages.length === 0) {
+    // Mirrors ChatTab my-chat error UI — surfaces the load failure
+    // with a retry button instead of silently rendering empty state.
+    return (
+      <div
+        role="alert"
+        className="mx-2 mt-2 rounded-lg border border-red-800/50 bg-red-950/30 px-3 py-2.5"
+      >
+        <p className="text-[11px] text-bad mb-1.5">
+          Failed to load agent communications: {loadError}
+        </p>
+        <button
+          onClick={loadInitial}
+          className="text-[10px] px-2 py-0.5 rounded bg-red-800/40 text-bad hover:bg-red-700/50 transition-colors"
+        >
+          Retry
+        </button>
+      </div>
+    );
+  }
+
  if (messages.length === 0) {
    return (
      <div className="text-xs text-ink-soft text-center py-8">
@@ -0,0 +1,124 @@
+"use client";
+
+// AttachmentAudio — inline native HTML5 <audio controls> player for
+// chat attachments (RFC #2991, PR-2).
+//
+// Same auth + Blob-URL pattern as AttachmentImage / AttachmentVideo.
+// Native audio control bar handles play/pause/scrub/volume/download,
+// and there's no fullscreen UI to worry about (audio doesn't need
+// AttachmentLightbox).
+
+import { useState, useEffect, useRef } from "react";
+import type { ChatAttachment } from "./types";
+import { isPlatformAttachment, resolveAttachmentHref } from "./uploads";
+import { AttachmentChip } from "./AttachmentViews";
+
+interface Props {
+  workspaceId: string;
+  attachment: ChatAttachment;
+  onDownload: (a: ChatAttachment) => void;
+  tone: "user" | "agent";
+}
+
+type FetchState =
+  | { kind: "idle" }
+  | { kind: "loading" }
+  | { kind: "ready"; src: string }
+  | { kind: "error" };
+
+export function AttachmentAudio({ workspaceId, attachment, onDownload, tone }: Props) {
+  const [state, setState] = useState<FetchState>({ kind: "idle" });
+  const blobUrlRef = useRef<string | null>(null);
+
+  useEffect(() => {
+    let cancelled = false;
+    setState({ kind: "loading" });
+
+    if (!isPlatformAttachment(attachment.uri)) {
+      const href = resolveAttachmentHref(workspaceId, attachment.uri);
+      if (!cancelled) setState({ kind: "ready", src: href });
+      return;
+    }
+
+    void (async () => {
+      try {
+        const href = resolveAttachmentHref(workspaceId, attachment.uri);
+        const headers: Record<string, string> = {};
+        const adminToken = process.env.NEXT_PUBLIC_ADMIN_TOKEN;
+        if (adminToken) headers["Authorization"] = `Bearer ${adminToken}`;
+        const slug = getTenantSlug();
+        if (slug) headers["X-Molecule-Org-Slug"] = slug;
+        const res = await fetch(href, {
+          headers,
+          credentials: "include",
+          signal: AbortSignal.timeout(60_000),
+        });
+        if (!res.ok) {
+          if (!cancelled) setState({ kind: "error" });
+          return;
+        }
+        const blob = await res.blob();
+        const url = URL.createObjectURL(blob);
+        blobUrlRef.current = url;
+        if (cancelled) {
+          URL.revokeObjectURL(url);
+          return;
+        }
+        setState({ kind: "ready", src: url });
+      } catch {
+        if (!cancelled) setState({ kind: "error" });
+      }
+    })();
+
+    return () => {
+      cancelled = true;
+      if (blobUrlRef.current) {
+        URL.revokeObjectURL(blobUrlRef.current);
+        blobUrlRef.current = null;
+      }
+    };
+  }, [workspaceId, attachment.uri]);
+
+  if (state.kind === "error") {
+    return <AttachmentChip attachment={attachment} onDownload={onDownload} tone={tone} />;
+  }
+  if (state.kind === "idle" || state.kind === "loading") {
+    return (
+      <div
+        className="rounded-md border border-line/50 bg-surface-card/40 animate-pulse"
+        style={{ width: 280, height: 40 }}
+        aria-label={`Loading ${attachment.name}`}
+      />
+    );
+  }
+
+  return (
+    <div
+      className={`inline-flex flex-col gap-1 rounded-md border px-2 py-1 ${
+        tone === "user" ? "border-blue-400/30 bg-accent-strong/10" : "border-line/50 bg-surface-card/40"
+      }`}
+    >
+      {/* Filename label so the user knows what they're hearing
+          before pressing play. Short, single-line, truncated. */}
+      <span className="text-[10px] text-ink-mid truncate max-w-[280px]" title={attachment.name}>
+        {attachment.name}
+      </span>
+      <audio
+        controls
+        preload="metadata"
+        src={state.src}
+        style={{ width: 280, height: 32 }}
+        onError={() => setState({ kind: "error" })}
+      >
+        {attachment.name}
+      </audio>
+    </div>
+  );
+}
+
+function getTenantSlug(): string | null {
+  if (typeof window === "undefined") return null;
+  const host = window.location.hostname;
+  const m = host.match(/^([^.]+)\.moleculesai\.app$/);
+  return m ? m[1] : null;
+}
@@ -0,0 +1,198 @@
+"use client";
+
+// AttachmentImage — inline image thumbnail + click-to-fullscreen.
+// First "specialized renderer" landing under RFC #2991 PR-1.
+//
+// Auth model
+// ----------
+//
+// The Critical UX/Security trade-off (per RFC's hostile-self-review
+// item #2): the bytes live behind workspace auth. A bare
+// <img src="https://reno-stars.../chat/download?path=…"> WILL NOT
+// include our cookie + Origin headers when the browser loads it —
+// even for same-origin canvas-server, the auth chain (cookie + token
+// + X-Molecule-Org-Slug header) is JS-injected, not browser-default.
+//
+// Solution: same auth path the chip download uses. Fetch the bytes
+// with the JS auth headers, wrap in a Blob, hand the browser an
+// ObjectURL. The image renders from local memory; no second request,
+// no auth leakage, no CORS pain.
+//
+// That same blob URL is what the lightbox shows on click — single
+// fetch, cached for the lifetime of the message bubble.
+//
+// Failure modes
+// -------------
+//
+// - Fetch fails (404, 403, network) → fall back to AttachmentChip
+//   (the existing file-pill download flow). The user still gets a
+//   working download; we just lose the inline preview.
+// - Decoded as non-image (server returned wrong Content-Type, or
+//   bytes are corrupt) → onError handler swaps to AttachmentChip.
+// - Bytes too large — no enforcement here; the server caps at 25MB
+//   per file (chat_files.go), which is too big for a thumbnail but
+//   acceptable for a chat-attached image. If we hit pain we can
+//   downscale via canvas, but defer that to v2.
+
+import { useState, useEffect, useRef } from "react";
+import type { ChatAttachment } from "./types";
+import { isPlatformAttachment, resolveAttachmentHref } from "./uploads";
+import { AttachmentLightbox } from "./AttachmentLightbox";
+import { AttachmentChip } from "./AttachmentViews";
+
+interface Props {
+  workspaceId: string;
+  attachment: ChatAttachment;
+  onDownload: (a: ChatAttachment) => void;
+  tone: "user" | "agent";
+}
+
+type FetchState =
+  | { kind: "idle" }
+  | { kind: "loading" }
+  | { kind: "ready"; blobUrl: string }
+  | { kind: "error" };
+
+export function AttachmentImage({ workspaceId, attachment, onDownload, tone }: Props) {
+  const [state, setState] = useState<FetchState>({ kind: "idle" });
+  const [open, setOpen] = useState(false);
+  // Track whether we created the ObjectURL so cleanup runs on the
+  // exact value we minted (state could change between effect setup
+  // and effect cleanup if a new fetch fires).
+  const blobUrlRef = useRef<string | null>(null);
+
+  useEffect(() => {
+    let cancelled = false;
+    setState({ kind: "loading" });
+
+    // For non-platform URIs (http/https external image hosts) we can
+    // skip the auth fetch — browser loads them directly. We bail out
+    // of the auth-fetch flow and use the raw URL via resolveAttachmentHref.
+    if (!isPlatformAttachment(attachment.uri)) {
+      const href = resolveAttachmentHref(workspaceId, attachment.uri);
+      if (!cancelled) setState({ kind: "ready", blobUrl: href });
+      return;
+    }
+
+    // Platform-auth path: identical to downloadChatFile but we keep
+    // the blob (don't trigger a Save-As). Use the same headers it does
+    // by going through it indirectly — no, downloadChatFile triggers a
+    // Save-As. Need a separate fetch.
+    void (async () => {
+      try {
+        const href = resolveAttachmentHref(workspaceId, attachment.uri);
+        const headers: Record<string, string> = {};
+        // Read the same env var downloadChatFile reads — single source
+        // of truth would be cleaner; refactor opportunity for PR-2 if
+        // we add the same path to AttachmentVideo.
+        const adminToken = process.env.NEXT_PUBLIC_ADMIN_TOKEN;
+        if (adminToken) headers["Authorization"] = `Bearer ${adminToken}`;
+        const slug = getTenantSlug();
+        if (slug) headers["X-Molecule-Org-Slug"] = slug;
+        const res = await fetch(href, {
+          headers,
+          credentials: "include",
+          signal: AbortSignal.timeout(30_000),
+        });
+        if (!res.ok) {
+          if (!cancelled) setState({ kind: "error" });
+          return;
+        }
+        const blob = await res.blob();
+        const url = URL.createObjectURL(blob);
+        blobUrlRef.current = url;
+        if (cancelled) {
+          URL.revokeObjectURL(url);
+          return;
+        }
+        setState({ kind: "ready", blobUrl: url });
+      } catch {
+        if (!cancelled) setState({ kind: "error" });
+      }
+    })();
+
+    return () => {
+      cancelled = true;
+      // Free the ObjectURL when the bubble unmounts — keeps memory
+      // bounded across long chat histories.
+      if (blobUrlRef.current) {
+        URL.revokeObjectURL(blobUrlRef.current);
+        blobUrlRef.current = null;
+      }
+    };
+  }, [workspaceId, attachment.uri]);
+
+  // Failure → render the existing file chip. Maintains the download
+  // affordance even if preview fails; the user never gets stuck.
+  if (state.kind === "error") {
+    return <AttachmentChip attachment={attachment} onDownload={onDownload} tone={tone} />;
+  }
+
+  // Loading → small placeholder pill so the bubble doesn't reflow
+  // when the image lands. Sized to roughly the thumbnail's aspect
+  // ratio guess (a 240x180 box) so the layout is stable.
+  if (state.kind === "loading" || state.kind === "idle") {
+    return (
+      <div
+        className="rounded-md border border-line/50 bg-surface-card/40 animate-pulse"
+        style={{ width: 240, height: 180 }}
+        aria-label={`Loading ${attachment.name}`}
+      />
+    );
+  }
+
+  // Ready → inline thumbnail with click handler. The img has its
+  // own onError so a corrupt blob (server returned the right size
+  // but invalid bytes) falls through to the chip too.
+  return (
+    <>
+      <button
+        type="button"
+        onClick={() => setOpen(true)}
+        title={`Preview ${attachment.name}`}
+        className={`group relative inline-block max-w-full rounded-lg overflow-hidden border focus:outline-none focus-visible:ring-2 focus-visible:ring-accent/60 ${
+          tone === "user" ? "border-blue-400/30" : "border-line/50"
+        }`}
+        aria-label={`Open ${attachment.name} preview`}
+      >
+        <img
+          src={state.blobUrl}
+          alt={attachment.name}
+          // Cap thumbnail so a tall portrait image doesn't blow up
+          // the message bubble. The lightbox shows the full size.
+          style={{ maxWidth: 240, maxHeight: 180, display: "block" }}
+          onError={() => setState({ kind: "error" })}
+        />
+        {/* Tiny filename label on hover — same affordance as Slack/
+            Discord. Helps when several images land in one bubble. */}
+        <div className="absolute bottom-0 inset-x-0 bg-black/60 text-white text-[10px] px-1.5 py-0.5 truncate opacity-0 group-hover:opacity-100 transition-opacity">
+          {attachment.name}
+        </div>
+      </button>
+      <AttachmentLightbox
+        open={open}
+        onClose={() => setOpen(false)}
+        ariaLabel={`Preview of ${attachment.name}`}
+      >
+        <img
+          src={state.blobUrl}
+          alt={attachment.name}
+          className="max-w-[95vw] max-h-[90vh] object-contain"
+        />
+      </AttachmentLightbox>
+    </>
+  );
+}
+
+// Internal helper — duplicated from uploads.ts (it's not exported
+// there). Kept local so this component doesn't reach into private
+// surface; if AttachmentVideo / AttachmentPDF in PR-2/PR-3 also need
+// it, lift to an exported helper at that point (the third-caller
+// rule).
+function getTenantSlug(): string | null {
+  if (typeof window === "undefined") return null;
+  const host = window.location.hostname;
+  // Tenant subdomain shape: <slug>.moleculesai.app
+  const m = host.match(/^([^.]+)\.moleculesai\.app$/);
+  return m ? m[1] : null;
+}
@@ -0,0 +1,122 @@
+"use client";
+
+// AttachmentLightbox — shared fullscreen modal for image / PDF /
+// (future) any-fullscreen-renderable kind. Owns:
+//   - Backdrop + centered viewport
+//   - Esc to close
+//   - Click-outside to close
+//   - Focus trap (focus enters the modal on open, restored on close)
+//   - prefers-reduced-motion respect (no animation)
+//
+// Per RFC #2991 Phase 2: this is the third-caller justification for
+// the abstraction (image, PDF, future video-fullscreen all want the
+// same modal contract). Not invented for a single caller.
+//
+// Design choices:
+//
+// 1. Portals — we don't use ReactDOM.createPortal because the canvas
+//    chat surface already renders at a high z-index and the modal's
+//    fixed-position layout reaches the viewport regardless. Saves a
+//    portal mount in the common case + avoids the SSR warning (canvas
+//    is "use client" but the parent shell is server-rendered).
+//
+// 2. Focus trap — inline implementation (not a 3rd-party dep). The
+//    chat lightbox needs to trap focus only across two interactive
+//    elements (close button + content), so a 100-line manual trap
+//    beats pulling in focus-trap-react for ~12KB.
+//
+// 3. Escape key — listened on `document` (not on the modal element)
+//    because the user can be focused anywhere when they hit Esc,
+//    including outside the modal if focus restoration ever fails.
+//    The cleanup runs on unmount so leaked listeners don't persist.
+
+import { useEffect, useRef, useCallback, type ReactNode } from "react";
+
+interface Props {
+  /** Render the lightbox when true. Caller controls open state. */
+  open: boolean;
+  /** Caller's handler for "close" — Esc, click-outside, X button. */
+  onClose: () => void;
+  /** Accessible label for the modal — voiced by screen readers when
+   *  the dialog opens. The caller knows what's inside (image alt
+   *  text, PDF filename) and supplies it. */
+  ariaLabel: string;
+  /** The thing being shown in fullscreen — <img>, <embed>, etc.
+   *  Caller is responsible for sizing it to fit the viewport (we
+   *  give it max-w-full max-h-full via CSS). */
+  children: ReactNode;
+}
+
+export function AttachmentLightbox({ open, onClose, ariaLabel, children }: Props) {
+  const closeButtonRef = useRef<HTMLButtonElement>(null);
+  const previousFocusRef = useRef<HTMLElement | null>(null);
+
+  // Focus enters the close button on open + restores to whatever
+  // had focus when the modal closes. Without this, the user's
+  // focus is left wherever they clicked (often the chip) and Tab
+  // walks them back through the chat surface — disorienting.
+  useEffect(() => {
+    if (!open) return;
+    previousFocusRef.current = document.activeElement as HTMLElement | null;
+    closeButtonRef.current?.focus();
+    return () => {
+      previousFocusRef.current?.focus?.();
+    };
+  }, [open]);
+
+  // Esc closes; bound on document so the user can press Esc
+  // regardless of where focus actually is.
+  useEffect(() => {
+    if (!open) return;
+    const onKey = (e: KeyboardEvent) => {
+      if (e.key === "Escape") {
+        e.preventDefault();
+        onClose();
+      }
+    };
+    document.addEventListener("keydown", onKey);
+    return () => document.removeEventListener("keydown", onKey);
+  }, [open, onClose]);
+
+  // Click on the backdrop (NOT the content) closes. Content's own
+  // onClick stops propagation so the user can interact (e.g. native
+  // PDF viewer controls) without dismissing the modal.
+  const onBackdropClick = useCallback(
+    (e: React.MouseEvent) => {
+      if (e.target === e.currentTarget) onClose();
+    },
+    [onClose],
+  );
+
+  if (!open) return null;
+
+  return (
+    <div
+      role="dialog"
+      aria-modal="true"
+      aria-label={ariaLabel}
+      className="fixed inset-0 z-50 flex items-center justify-center bg-black/85 motion-reduce:transition-none transition-opacity"
+      onClick={onBackdropClick}
+    >
+      {/* Close button — top-right, large hit area, keyboard-focusable.
+          ariaLabel includes "Close" so SR users hear what action it
+          performs, not just the X glyph. */}
+      <button
+        ref={closeButtonRef}
+        onClick={onClose}
+        aria-label="Close preview"
+        className="absolute top-4 right-4 rounded-full bg-white/10 hover:bg-white/20 text-white p-2 focus:outline-none focus-visible:ring-2 focus-visible:ring-white"
+      >
+        <svg width="20" height="20" viewBox="0 0 24 24" fill="none" aria-hidden="true">
+          <path d="M5 5l14 14M19 5l-14 14" stroke="currentColor" strokeWidth="2" strokeLinecap="round" />
+        </svg>
+      </button>
+      <div
+        className="max-w-[95vw] max-h-[90vh] flex items-center justify-center"
+        onClick={(e) => e.stopPropagation()}
+      >
+        {children}
+      </div>
+    </div>
+  );
+}
@@ -0,0 +1,197 @@
+"use client";
+
+// AttachmentPDF — inline PDF preview using the browser's native viewer
+// (RFC #2991, PR-3).
+//
+// Why browser-native (not PDF.js / pdfjs-dist):
+//
+//   - Chrome / Edge / Firefox / Safari (desktop) all ship a built-in
+//     PDF viewer. <embed src="…blob"> renders correctly; user gets
+//     scroll, zoom, search, print for free.
+//   - PDF.js adds ~3 MB to the canvas bundle. For an MVP that
+//     specifically targets desktop chat, the browser viewer is good
+//     enough. v2 can wire pdfjs-dist if Safari mobile coverage
+//     becomes a real ask (its built-in viewer is preview-only).
+//
+// Auth model: identical to AttachmentImage / Video / Audio — fetch
+// bytes with JS-injected auth headers, wrap in Blob, hand the
+// browser an ObjectURL. <embed src="blob:…#toolbar=0"> would
+// suppress the toolbar; we keep it on so the user gets standard
+// PDF affordances.
+//
+// Fullscreen: AttachmentLightbox hosts the PDF at viewport size on
+// click. Same shared modal as image — third caller justifies the
+// abstraction (per RFC #2991 design).
+//
+// Failure modes:
+//
+//   - Fetch fail → AttachmentChip fallback (download still works)
+//   - Browser refuses to render the PDF (Safari mobile, plugin
+//     disabled, corrupt bytes) → <embed onError> swap to chip.
+//     Note: <embed> doesn't fire onError reliably across browsers.
+//     Defensive fallback: if blob load triggers no onLoad after a
+//     timeout, swap to chip. Implemented as a 3-second watchdog.
+
+import { useState, useEffect, useRef } from "react";
+import type { ChatAttachment } from "./types";
+import { isPlatformAttachment, resolveAttachmentHref } from "./uploads";
+import { AttachmentLightbox } from "./AttachmentLightbox";
+import { AttachmentChip } from "./AttachmentViews";
+
+interface Props {
+  workspaceId: string;
+  attachment: ChatAttachment;
+  onDownload: (a: ChatAttachment) => void;
+  tone: "user" | "agent";
+}
+
+type FetchState =
+  | { kind: "idle" }
+  | { kind: "loading" }
+  | { kind: "ready"; blobUrl: string }
+  | { kind: "error" };
+
+export function AttachmentPDF({ workspaceId, attachment, onDownload, tone }: Props) {
+  const [state, setState] = useState<FetchState>({ kind: "idle" });
+  const [open, setOpen] = useState(false);
+  const blobUrlRef = useRef<string | null>(null);
+
+  useEffect(() => {
+    let cancelled = false;
+    setState({ kind: "loading" });
+
+    if (!isPlatformAttachment(attachment.uri)) {
+      const href = resolveAttachmentHref(workspaceId, attachment.uri);
+      if (!cancelled) setState({ kind: "ready", blobUrl: href });
+      return;
+    }
+
+    void (async () => {
+      try {
+        const href = resolveAttachmentHref(workspaceId, attachment.uri);
+        const headers: Record<string, string> = {};
+        const adminToken = process.env.NEXT_PUBLIC_ADMIN_TOKEN;
+        if (adminToken) headers["Authorization"] = `Bearer ${adminToken}`;
+        const slug = getTenantSlug();
+        if (slug) headers["X-Molecule-Org-Slug"] = slug;
+        const res = await fetch(href, {
+          headers,
+          credentials: "include",
+          signal: AbortSignal.timeout(60_000),
+        });
+        if (!res.ok) {
+          if (!cancelled) setState({ kind: "error" });
+          return;
+        }
+        const blob = await res.blob();
+        const url = URL.createObjectURL(blob);
+        blobUrlRef.current = url;
+        if (cancelled) {
+          URL.revokeObjectURL(url);
+          return;
+        }
+        setState({ kind: "ready", blobUrl: url });
+      } catch {
+        if (!cancelled) setState({ kind: "error" });
+      }
+    })();
+
+    return () => {
+      cancelled = true;
+      if (blobUrlRef.current) {
+        URL.revokeObjectURL(blobUrlRef.current);
+        blobUrlRef.current = null;
+      }
+    };
+  }, [workspaceId, attachment.uri]);
+
+  if (state.kind === "error") {
+    return <AttachmentChip attachment={attachment} onDownload={onDownload} tone={tone} />;
+  }
+  if (state.kind === "idle" || state.kind === "loading") {
+    return (
+      <div
+        className="rounded-md border border-line/50 bg-surface-card/40 animate-pulse flex items-center gap-1.5 px-2 py-1 text-[10px] text-ink-mid"
+        style={{ width: 240 }}
+        aria-label={`Loading ${attachment.name}`}
+      >
+        <PdfGlyph />
+        Loading {attachment.name}…
+      </div>
+    );
+  }
+
+  // PDF preview chip — clicking it opens the full embed in the
+  // shared lightbox. We don't inline-embed in the bubble because
+  // even a small embed renders at 600×400 minimum on most browsers
+  // (the PDF viewer's natural scale), which would dominate every
+  // chat bubble. Slack/Linear/Notion all gate PDF preview behind a
+  // click for the same reason.
+  return (
+    <>
+      <button
+        type="button"
+        onClick={() => setOpen(true)}
+        title={`Preview ${attachment.name}`}
+        className={`inline-flex items-center gap-1.5 rounded-md border px-2 py-1 text-[10px] hover:bg-surface-card/70 focus:outline-none focus-visible:ring-2 focus-visible:ring-accent/60 ${
+          tone === "user"
+            ? "border-blue-400/30 bg-accent-strong/10 text-blue-100"
+            : "border-line/50 bg-surface-card/40 text-ink"
+        }`}
+        aria-label={`Open ${attachment.name} preview`}
+      >
+        <PdfGlyph />
+        <span className="truncate max-w-[200px]">{attachment.name}</span>
+        <span className="opacity-60 shrink-0">PDF</span>
+      </button>
+      <AttachmentLightbox
+        open={open}
+        onClose={() => setOpen(false)}
+        ariaLabel={`Preview of ${attachment.name}`}
+      >
+        <embed
+          src={state.blobUrl}
+          type="application/pdf"
+          // The lightbox's content slot caps at 95vw / 90vh, so size
+          // 100% within that and let the user scroll inside the PDF
+          // viewer.
+          style={{ width: "95vw", height: "90vh" }}
+          aria-label={attachment.name}
+        />
+      </AttachmentLightbox>
+    </>
+  );
+}
+
+function PdfGlyph() {
+  return (
+    <svg
+      width="11"
+      height="11"
+      viewBox="0 0 16 16"
+      fill="none"
+      aria-hidden="true"
+      className="shrink-0 opacity-70"
+    >
+      <path
+        d="M4 2h5l3 3v9a1 1 0 0 1-1 1H4a1 1 0 0 1-1-1V3a1 1 0 0 1 1-1Z"
+        stroke="currentColor"
+        strokeWidth="1.3"
+      />
+      <path d="M9 2v3h3" stroke="currentColor" strokeWidth="1.3" />
+      <path
+        d="M5.5 9.5h1m1 0h1m-3 2h2"
+        stroke="currentColor"
+        strokeWidth="1.1"
+        strokeLinecap="round"
+      />
+    </svg>
+  );
+}
+
+function getTenantSlug(): string | null {
+  if (typeof window === "undefined") return null;
+  const host = window.location.hostname;
+  const m = host.match(/^([^.]+)\.moleculesai\.app$/);
+  return m ? m[1] : null;
+}
@@ -0,0 +1,90 @@
+"use client";
+
+// AttachmentPreview — the SSOT dispatch point for chat-attachment
+// rendering (RFC #2991, PR-1).
+//
+// Replaces the previous direct-AttachmentChip usage in ChatTab so
+// every attachment routes through the same preview-kind taxonomy.
+// Adding a new renderer (PDF, video, audio, text) in PR-2/PR-3 is a
+// one-arm extension to the switch below — no touch-points scattered
+// across ChatTab.tsx, AgentCommsPanel.tsx, or other chat consumers.
+//
+// Per the RFC's Phase 2: this is the only file that should directly
+// import any kind-specific component. ChatTab and other callers
+// import only AttachmentPreview — no leaking of the kind taxonomy
+// into the consumer surface.
+
+import type { ChatAttachment } from "./types";
+import { getAttachmentPreviewKind } from "./preview-kind";
+import { AttachmentImage } from "./AttachmentImage";
+import { AttachmentVideo } from "./AttachmentVideo";
+import { AttachmentAudio } from "./AttachmentAudio";
+import { AttachmentPDF } from "./AttachmentPDF";
+import { AttachmentTextPreview } from "./AttachmentTextPreview";
+import { AttachmentChip } from "./AttachmentViews";
+
+interface Props {
+  workspaceId: string;
+  attachment: ChatAttachment;
+  /** Caller's download handler — used for the kind=file fallback
+   *  and as the kind-specific renderers' fallback when their own
+   *  preview fails (e.g. image fetch errored). */
+  onDownload: (a: ChatAttachment) => void;
+  /** Tone follows the message bubble's role — used for visual
+   *  variant only. */
+  tone: "user" | "agent";
+}
+
+export function AttachmentPreview({ workspaceId, attachment, onDownload, tone }: Props) {
+  const kind = getAttachmentPreviewKind(attachment.mimeType, attachment.uri, attachment.name);
+  switch (kind) {
+    case "image":
+      return (
+        <AttachmentImage
+          workspaceId={workspaceId}
+          attachment={attachment}
+          onDownload={onDownload}
+          tone={tone}
+        />
+      );
+    case "video":
+      return (
+        <AttachmentVideo
+          workspaceId={workspaceId}
+          attachment={attachment}
+          onDownload={onDownload}
+          tone={tone}
+        />
+      );
+    case "audio":
+      return (
+        <AttachmentAudio
+          workspaceId={workspaceId}
+          attachment={attachment}
+          onDownload={onDownload}
+          tone={tone}
+        />
+      );
+    case "pdf":
+      return (
+        <AttachmentPDF
+          workspaceId={workspaceId}
+          attachment={attachment}
+          onDownload={onDownload}
+          tone={tone}
+        />
+      );
+    case "text":
+      return (
+        <AttachmentTextPreview
+          workspaceId={workspaceId}
+          attachment={attachment}
+          onDownload={onDownload}
+          tone={tone}
+        />
+      );
+    case "file":
+    default:
+      return <AttachmentChip attachment={attachment} onDownload={onDownload} tone={tone} />;
+  }
+}
@@ -0,0 +1,190 @@
+"use client";
+
+// AttachmentTextPreview — inline preview for text/code/JSON/YAML/etc
+// (RFC #2991, PR-3).
+//
+// Shape: render first N lines (~10) in monospace inside the bubble.
+// Click "Show more" to expand fully; the lightbox is reserved for
+// image/PDF where viewport-size matters. For text, the bubble itself
+// can host the full content.
+//
+// Why no syntax highlighting (yet):
+//
+//   - Pulling in shiki / highlight.js / prism adds 200-500KB to the
+//     bundle for a feature that's nice-to-have. MVP uses plain
+//     <pre><code>.
+//   - Future: lazy-load shiki on first text-attachment render. v2
+//     if the user reports the gap.
+//
+// Auth: same fetch+text() pattern as image/video/audio, but we read
+// the text directly instead of building a Blob URL — no <img>/<video>
+// element to feed.
+//
+// Memory: text files are usually small. We cap the preview at 256 KB
+// fetched (large logs would otherwise crash the bubble). If the file
+// exceeds the cap, we show what we got + a "truncated" note + a chip
+// to download the full file.
+
+import { useState, useEffect } from "react";
+import type { ChatAttachment } from "./types";
+import { isPlatformAttachment, resolveAttachmentHref } from "./uploads";
+import { AttachmentChip } from "./AttachmentViews";
+
+interface Props {
+  workspaceId: string;
+  attachment: ChatAttachment;
+  onDownload: (a: ChatAttachment) => void;
+  tone: "user" | "agent";
+}
+
+type FetchState =
+  | { kind: "idle" }
+  | { kind: "loading" }
+  | { kind: "ready"; text: string; truncated: boolean }
+  | { kind: "error" };
+
+const PREVIEW_LINE_COUNT = 10;
+const MAX_FETCH_BYTES = 256 * 1024; // 256 KB
+
+export function AttachmentTextPreview({ workspaceId, attachment, onDownload, tone }: Props) {
+  const [state, setState] = useState<FetchState>({ kind: "idle" });
+  const [expanded, setExpanded] = useState(false);
+
+  useEffect(() => {
+    let cancelled = false;
+    setState({ kind: "loading" });
+
+    void (async () => {
+      try {
+        const href = resolveAttachmentHref(workspaceId, attachment.uri);
+        const headers: Record<string, string> = {};
+        if (isPlatformAttachment(attachment.uri)) {
+          const adminToken = process.env.NEXT_PUBLIC_ADMIN_TOKEN;
+          if (adminToken) headers["Authorization"] = `Bearer ${adminToken}`;
+          const slug = getTenantSlug();
+          if (slug) headers["X-Molecule-Org-Slug"] = slug;
+        }
+        const res = await fetch(href, {
+          headers,
+          credentials: "include",
+          signal: AbortSignal.timeout(30_000),
+        });
+        if (!res.ok) {
+          if (!cancelled) setState({ kind: "error" });
+          return;
+        }
+        // Read up to MAX_FETCH_BYTES. Use the standard ReadableStream
+        // path so we don't materialise a 100MB log into memory.
+        const reader = res.body?.getReader();
+        if (!reader) {
+          // Fallback: small text file, just .text() it.
+          const text = await res.text();
+          if (cancelled) return;
+          setState({
+            kind: "ready",
+            text: text.slice(0, MAX_FETCH_BYTES),
+            truncated: text.length > MAX_FETCH_BYTES,
+          });
+          return;
+        }
+        let received = 0;
+        const chunks: BlobPart[] = [];
+        while (received < MAX_FETCH_BYTES) {
+          const { value, done } = await reader.read();
+          if (done) break;
+          // Copy into a fresh ArrayBuffer-backed view — TS in lib.dom
+          // 2026 narrows BlobPart away from SharedArrayBuffer-backed
+          // Uint8Arrays. Blob() accepts the copy fine at runtime.
+          const copy = new Uint8Array(value.byteLength);
+          copy.set(value);
+          chunks.push(copy.buffer);
+          received += value.byteLength;
+        }
+        // If we hit the cap but the stream isn't done, mark truncated.
+        const truncated = received >= MAX_FETCH_BYTES;
+        if (truncated) reader.cancel();
+        const blob = new Blob(chunks);
+        const text = await blob.text();
+        if (cancelled) return;
+        setState({ kind: "ready", text, truncated });
+      } catch {
+        if (!cancelled) setState({ kind: "error" });
+      }
+    })();
+
+    return () => {
+      cancelled = true;
+    };
+  }, [workspaceId, attachment.uri]);
+
+  if (state.kind === "error") {
+    return <AttachmentChip attachment={attachment} onDownload={onDownload} tone={tone} />;
+  }
+  if (state.kind === "idle" || state.kind === "loading") {
+    return (
+      <div
+        className="rounded-md border border-line/50 bg-surface-card/40 animate-pulse"
+        style={{ width: 320, height: 80 }}
+        aria-label={`Loading ${attachment.name}`}
+      />
+    );
+  }
+
+  const lines = state.text.split("\n");
+  const preview = expanded ? state.text : lines.slice(0, PREVIEW_LINE_COUNT).join("\n");
+  const showExpandButton = !expanded && lines.length > PREVIEW_LINE_COUNT;
+
+  return (
+    <div
+      className={`inline-block max-w-full rounded-md border ${
+        tone === "user" ? "border-blue-400/30 bg-accent-strong/10" : "border-line/50 bg-surface-card/40"
+      }`}
+    >
+      <div className="flex items-center justify-between px-2 py-1 border-b border-line/40 text-[10px] text-ink-mid">
+        <span className="truncate max-w-[220px]" title={attachment.name}>
+          {attachment.name}
+        </span>
+        <button
+          type="button"
+          onClick={() => onDownload(attachment)}
+          className="text-ink-soft hover:text-ink"
+          title={`Download ${attachment.name}`}
+          aria-label={`Download ${attachment.name}`}
+        >
+          ⬇
+        </button>
+      </div>
+      <pre className="overflow-x-auto px-2 py-1.5 text-[10px] leading-snug text-ink whitespace-pre font-mono max-w-[480px] max-h-[300px]">
+        <code>{preview}</code>
+      </pre>
+      {showExpandButton && (
+        <button
+          type="button"
+          onClick={() => setExpanded(true)}
+          className="block w-full text-center text-[10px] text-ink-mid hover:text-ink py-1 border-t border-line/40"
+        >
+          Show all {lines.length} lines
+        </button>
+      )}
+      {state.truncated && (
+        <div className="px-2 py-1 text-[10px] text-warm border-t border-line/40">
+          Preview truncated at {Math.round(MAX_FETCH_BYTES / 1024)} KB —{" "}
+          <button
+            type="button"
+            onClick={() => onDownload(attachment)}
+            className="underline"
+          >
+            download full file
+          </button>
+        </div>
+      )}
+    </div>
+  );
+}
+
+function getTenantSlug(): string | null {
+  if (typeof window === "undefined") return null;
+  const host = window.location.hostname;
+  const m = host.match(/^([^.]+)\.moleculesai\.app$/);
+  return m ? m[1] : null;
+}
@@ -0,0 +1,157 @@
+"use client";
+
+// AttachmentVideo — inline native HTML5 <video controls> player for
+// chat attachments (RFC #2991, PR-2).
+//
+// Why HTML5-native (vs custom JS player):
+//
+//   - Browser vendors ship hardware-accelerated decoders, captions,
+//     and fullscreen UI. We get all of it for free.
+//   - Native fullscreen via the <video> element's built-in button
+//     (no AttachmentLightbox needed for video — the browser does it).
+//   - Mobile-friendly: iOS / Android Safari + Chrome handle the
+//     pinch + scrub UX the user already knows.
+//
+// Auth model — identical to AttachmentImage:
+// platform-auth URIs need our cookie/token, so we fetch the bytes,
+// wrap in a Blob, hand the browser an ObjectURL via <video src=>.
+// External (http/https) URIs skip the fetch and use the raw URL.
+//
+// Memory caveat: a Blob holds the entire video in JS memory until
+// the bubble unmounts. For multi-hundred-MB videos this is bad. The
+// server caps single-file uploads at 25MB (chat_files.go), so we're
+// bounded; if larger files become a real shape, switch to streaming
+// via MediaSource or just `<video src=…>` with a credentials-aware
+// fetch via service worker. v2 if measured-needed.
+
+import { useState, useEffect, useRef } from "react";
+import type { ChatAttachment } from "./types";
+import { isPlatformAttachment, resolveAttachmentHref } from "./uploads";
+import { AttachmentChip } from "./AttachmentViews";
+
+interface Props {
+  workspaceId: string;
+  attachment: ChatAttachment;
+  onDownload: (a: ChatAttachment) => void;
+  tone: "user" | "agent";
+}
+
+type FetchState =
+  | { kind: "idle" }
+  | { kind: "loading" }
+  | { kind: "ready"; src: string }
+  | { kind: "error" };
+
+export function AttachmentVideo({ workspaceId, attachment, onDownload, tone }: Props) {
+  const [state, setState] = useState<FetchState>({ kind: "idle" });
+  const blobUrlRef = useRef<string | null>(null);
+
+  useEffect(() => {
+    let cancelled = false;
+    setState({ kind: "loading" });
+
+    if (!isPlatformAttachment(attachment.uri)) {
+      // External video (http/https) — let the browser stream it
+      // natively without the JS-blob detour.
+      const href = resolveAttachmentHref(workspaceId, attachment.uri);
+      if (!cancelled) setState({ kind: "ready", src: href });
+      return;
+    }
+
+    void (async () => {
+      try {
+        const href = resolveAttachmentHref(workspaceId, attachment.uri);
+        const headers: Record<string, string> = {};
+        const adminToken = process.env.NEXT_PUBLIC_ADMIN_TOKEN;
+        if (adminToken) headers["Authorization"] = `Bearer ${adminToken}`;
+        const slug = getTenantSlug();
+        if (slug) headers["X-Molecule-Org-Slug"] = slug;
+        const res = await fetch(href, {
+          headers,
+          credentials: "include",
+          // Videos are larger than images on average; give the request
+          // more headroom. The server's per-request body cap (50MB) is
+          // still the actual ceiling.
+          signal: AbortSignal.timeout(120_000),
+        });
+        if (!res.ok) {
+          if (!cancelled) setState({ kind: "error" });
+          return;
+        }
+        const blob = await res.blob();
+        const url = URL.createObjectURL(blob);
+        blobUrlRef.current = url;
+        if (cancelled) {
+          URL.revokeObjectURL(url);
+          return;
+        }
+        setState({ kind: "ready", src: url });
+      } catch {
+        if (!cancelled) setState({ kind: "error" });
+      }
+    })();
+
+    return () => {
+      cancelled = true;
+      if (blobUrlRef.current) {
+        URL.revokeObjectURL(blobUrlRef.current);
+        blobUrlRef.current = null;
+      }
+    };
+  }, [workspaceId, attachment.uri]);
+
+  if (state.kind === "error") {
+    return <AttachmentChip attachment={attachment} onDownload={onDownload} tone={tone} />;
+  }
+  if (state.kind === "idle" || state.kind === "loading") {
+    return (
+      <div
+        className="rounded-md border border-line/50 bg-surface-card/40 animate-pulse"
+        style={{ width: 320, height: 180 }}
+        aria-label={`Loading ${attachment.name}`}
+      />
+    );
+  }
+
+  return (
+    <div
+      className={`inline-block rounded-lg overflow-hidden border ${
+        tone === "user" ? "border-blue-400/30" : "border-line/50"
+      }`}
+    >
+      <video
+        controls
+        // preload="metadata" so the browser fetches just enough to
+        // show duration + first frame thumbnail without streaming
+        // the whole file before the user clicks play.
+        preload="metadata"
+        // playsInline keeps mobile Safari from auto-fullscreening
+        // on play; the user can still hit the native fullscreen
+        // button (or PiP on Chrome) if they want.
+        playsInline
+        // Native fullscreen via the <video> control bar; no
+        // AttachmentLightbox needed for video.
+        src={state.src}
+        // Cap thumbnail / inline display so the bubble doesn't blow
+        // up vertical layout for tall portrait clips. The native
+        // fullscreen button uses the original aspect ratio.
+        style={{ maxWidth: 320, maxHeight: 240, display: "block" }}
+        // Bytes that aren't actually a valid video (corrupt blob,
+        // wrong Content-Type) fail load → swap to chip.
+        onError={() => setState({ kind: "error" })}
+      >
+        <track kind="captions" />
+        {attachment.name}
+      </video>
+    </div>
+  );
+}
+
+// Internal helper — same shape as AttachmentImage's. Lifted to a
+// shared util in PR-2.5 if a third caller needs it (PDF, audio).
+function getTenantSlug(): string | null {
+  if (typeof window === "undefined") return null;
+  const host = window.location.hostname;
+  const m = host.match(/^([^.]+)\.moleculesai\.app$/);
+  return m ? m[1] : null;
+}
@@ -0,0 +1,115 @@
+// @vitest-environment jsdom
+import { describe, it, expect, vi, beforeEach, afterEach } from "vitest";
+import { render, screen, fireEvent, waitFor } from "@testing-library/react";
+
+// API mock — tests can override per case via apiGetMock.mockImplementationOnce.
+const apiGetMock = vi.fn<(url: string) => Promise<unknown>>();
+vi.mock("@/lib/api", () => ({
+  api: {
+    get: (url: string) => apiGetMock(url),
+  },
+}));
+
+// useSocketEvent — no-op for these render tests; live updates aren't
+// what we're verifying here.
+vi.mock("@/hooks/useSocketEvent", () => ({
+  useSocketEvent: () => {},
+}));
+
+// Canvas store — peer name resolution.
+vi.mock("@/store/canvas", () => ({
+  useCanvasStore: {
+    getState: () => ({
+      nodes: [
+        { id: "ws-self", data: { name: "Self" } },
+        { id: "ws-peer", data: { name: "Peer Agent" } },
+      ],
+    }),
+  },
+}));
+
+// Toaster shim — AgentCommsPanel imports showToast.
+vi.mock("../../Toaster", () => ({
+  showToast: vi.fn(),
+}));
+
+import { AgentCommsPanel } from "../AgentCommsPanel";
+
+// jsdom doesn't implement scrollIntoView. Tests that observe the call
+// install a spy here; tests that don't care still need a no-op stub
+// so the component doesn't throw.
+const scrollSpy = vi.fn<(opts?: ScrollIntoViewOptions | boolean) => void>();
+beforeEach(() => {
+  apiGetMock.mockReset();
+  scrollSpy.mockReset();
+  Element.prototype.scrollIntoView = scrollSpy as unknown as Element["scrollIntoView"];
+});
+
+afterEach(() => {
+  vi.clearAllMocks();
+});
+
+describe("AgentCommsPanel — initial-state parity with ChatTab my-chat", () => {
+  it("shows loading text while history fetch is in flight", () => {
+    apiGetMock.mockReturnValueOnce(new Promise(() => { /* never resolves */ }));
+    render(<AgentCommsPanel workspaceId="ws-self" />);
+    expect(screen.getByText("Loading agent communications...")).toBeDefined();
+  });
+
+  it("renders error UI with a Retry button when the history fetch rejects", async () => {
+    apiGetMock.mockRejectedValueOnce(new Error("network down"));
+    render(<AgentCommsPanel workspaceId="ws-self" />);
+
+    // Wait for the error state to render — loading→error transition is async.
+    const alert = await waitFor(() => screen.getByRole("alert"));
+    expect(alert.textContent).toMatch(/Failed to load agent communications/);
+    expect(alert.textContent).toMatch(/network down/);
+
+    // Retry button must be present and trigger a refetch.
+    const retry = screen.getByRole("button", { name: "Retry" });
+    apiGetMock.mockResolvedValueOnce([]); // success on retry
+    fireEvent.click(retry);
+
+    // Two calls total: initial load + retry. Pin via mock call count.
+    await waitFor(() => expect(apiGetMock.mock.calls.length).toBe(2));
+  });
+
+  it("falls back to empty-state copy when load succeeds with zero rows", async () => {
+    apiGetMock.mockResolvedValueOnce([]);
+    render(<AgentCommsPanel workspaceId="ws-self" />);
+    await waitFor(() =>
+      expect(screen.getByText("No agent-to-agent communications yet.")).toBeDefined(),
+    );
+  });
+
+  it("scrollIntoView is called with behavior=instant on the first message arrival", async () => {
+    apiGetMock.mockResolvedValueOnce([
+      {
+        id: "act-1",
+        activity_type: "a2a_send",
+        source_id: "ws-self",
+        target_id: "ws-peer",
+        method: "message/send",
+        summary: "Delegating",
+        request_body: { message: { parts: [{ text: "hi" }] } },
+        response_body: null,
+        status: "ok",
+        created_at: "2026-04-25T18:00:00Z",
+      },
+    ]);
+    render(<AgentCommsPanel workspaceId="ws-self" />);
+
+    // useLayoutEffect is what makes the first call instant — wait for
+    // the panel to render at least one message.
+    await waitFor(() => expect(scrollSpy.mock.calls.length).toBeGreaterThan(0));
+
+    // The pinned contract: SOME call uses behavior: "instant" — the
+    // first-arrival case. Subsequent appends use "smooth", but those
+    // can't fire here (no live update yet).
+    const sawInstant = scrollSpy.mock.calls.some((args) => {
+      const opts = args[0];
+      return typeof opts === "object" && opts !== null && "behavior" in opts && opts.behavior === "instant";
+    });
+    expect(sawInstant).toBe(true);
+  });
+});
@@ -0,0 +1,317 @@
+// @vitest-environment jsdom
+//
+// AttachmentPreview component tests — pin the dispatch contract:
+// each kind goes to its dedicated renderer; kind=file falls back to
+// the chip; failure modes don't strand the user without a download.
+//
+// Per RFC #2991 Phase 4: every test must be able to fail. No
+// asserting-the-mock; we render the real component and inspect what
+// the DOM actually shows.
+
+import { describe, it, expect, vi, afterEach, beforeEach } from "vitest";
+import { render, screen, fireEvent, cleanup, waitFor, act } from "@testing-library/react";
+import React from "react";
+
+afterEach(cleanup);
+
+// Mock the auth-token env var so AttachmentImage's fetch doesn't
+// hit a real network. The fetch is itself mocked below.
+vi.stubEnv("NEXT_PUBLIC_ADMIN_TOKEN", "test-token");
+
+// Mock fetch so the AttachmentImage path can return a synthetic blob.
+// Tests override per-case to simulate success / 404 / network fail.
+const fetchMock = vi.fn();
+beforeEach(() => {
+  fetchMock.mockReset();
+  vi.stubGlobal("fetch", fetchMock);
+  // jsdom doesn't implement URL.createObjectURL — stub.
+  global.URL.createObjectURL = vi.fn(() => "blob:test-url");
+  global.URL.revokeObjectURL = vi.fn();
+});
+
+import { AttachmentPreview } from "../AttachmentPreview";
+import type { ChatAttachment } from "../types";
+
+const onDownload = vi.fn();
+
+function preview(att: ChatAttachment) {
+  return render(
+    <AttachmentPreview
+      workspaceId="ws-1"
+      attachment={att}
+      onDownload={onDownload}
+      tone="agent"
+    />,
+  );
+}
+
+describe("AttachmentPreview dispatch", () => {
+  it("kind=file → renders the AttachmentChip download button (existing fallback)", () => {
+    preview({ uri: "workspace:/workspace/tmp/foo.zip", name: "foo.zip", mimeType: "application/zip" });
+    // The chip's button title is `Download <name>`. Pre-fix this was
+    // the only render path; now it's the kind=file fallback.
+    expect(screen.getByTitle(/Download foo\.zip/i)).toBeTruthy();
+  });
+
+  it("kind=image (mime) → renders the AttachmentImage path (loading placeholder until fetch resolves)", async () => {
+    // never-resolving fetch → component sits in loading state. Pin
+    // the loading placeholder shape.
+    fetchMock.mockReturnValue(new Promise(() => {}));
+    preview({ uri: "workspace:/workspace/tmp/photo.png", name: "photo.png", mimeType: "image/png" });
+    expect(await screen.findByLabelText(/Loading photo\.png/i)).toBeTruthy();
+    // The chip download button must NOT be in the DOM during the
+    // image path's loading state — proves dispatch routed correctly.
+    expect(screen.queryByTitle(/Download photo\.png/i)).toBeNull();
+  });
+
+  it("kind=image (extension fallback when mime is empty) → image path", async () => {
+    fetchMock.mockReturnValue(new Promise(() => {}));
+    preview({ uri: "workspace:/workspace/screenshot.jpg", name: "screenshot.jpg" /* no mime */ });
+    expect(await screen.findByLabelText(/Loading screenshot\.jpg/i)).toBeTruthy();
+  });
+
+  it("kind=image fetch fails (404) → falls back to AttachmentChip so the user can still download", async () => {
+    fetchMock.mockResolvedValue({ ok: false, status: 404 });
+    preview({ uri: "workspace:/workspace/tmp/missing.png", name: "missing.png", mimeType: "image/png" });
+    // The fallback chip shows up on error.
+    await waitFor(() => {
+      expect(screen.getByTitle(/Download missing\.png/i)).toBeTruthy();
+    });
+  });
+
+  it("kind=image fetch network error → falls back to chip", async () => {
+    fetchMock.mockRejectedValue(new Error("network down"));
+    preview({ uri: "workspace:/workspace/tmp/x.png", name: "x.png", mimeType: "image/png" });
+    await waitFor(() => {
+      expect(screen.getByTitle(/Download x\.png/i)).toBeTruthy();
+    });
+  });
+
+  it("kind=image success → renders <img> + clicking opens the lightbox", async () => {
+    fetchMock.mockResolvedValue({
+      ok: true,
+      blob: async () => new Blob(["fake-png-bytes"], { type: "image/png" }),
+    });
+    preview({ uri: "workspace:/workspace/tmp/ok.png", name: "ok.png", mimeType: "image/png" });
+
+    // Image element shows up after the fetch resolves.
+    const img = await screen.findByAltText(/ok\.png/);
+    expect(img).toBeTruthy();
+    expect((img as HTMLImageElement).src).toBe("blob:test-url");
+
+    // Lightbox closed initially — the dialog must not be in the DOM.
+    expect(screen.queryByRole("dialog")).toBeNull();
+
+    // Click the thumbnail button (the surrounding <button>) → lightbox opens.
+    const button = screen.getByLabelText(/Open ok\.png preview/i);
+    fireEvent.click(button);
+
+    expect(await screen.findByRole("dialog")).toBeTruthy();
+    expect(screen.getByLabelText(/Close preview/i)).toBeTruthy();
+  });
+
+  it("kind=image lightbox closes on Esc keypress", async () => {
+    fetchMock.mockResolvedValue({
+      ok: true,
+      blob: async () => new Blob(["b"], { type: "image/png" }),
+    });
+    preview({ uri: "workspace:/workspace/tmp/x.png", name: "x.png", mimeType: "image/png" });
+    await screen.findByAltText(/x\.png/);
+    fireEvent.click(screen.getByLabelText(/Open x\.png preview/i));
+    expect(await screen.findByRole("dialog")).toBeTruthy();
+
+    // Esc on document — lightbox listens there per design (not on
+    // the modal element) so the user can press Esc anywhere.
+    act(() => {
+      const event = new KeyboardEvent("keydown", { key: "Escape", bubbles: true });
+      document.dispatchEvent(event);
+    });
+    await waitFor(() => {
+      expect(screen.queryByRole("dialog")).toBeNull();
+    });
+  });
+
+  it("kind=image lightbox closes on backdrop click but not on inner content click", async () => {
+    fetchMock.mockResolvedValue({
+      ok: true,
+      blob: async () => new Blob(["b"], { type: "image/png" }),
+    });
+    preview({ uri: "workspace:/workspace/tmp/x.png", name: "x.png", mimeType: "image/png" });
+    await screen.findByAltText(/x\.png/);
+    fireEvent.click(screen.getByLabelText(/Open x\.png preview/i));
+    const dialog = await screen.findByRole("dialog");
+
+    // Click on the inner content (the lightbox image) — must NOT close.
+    const lightboxImg = dialog.querySelector("img");
+    if (!lightboxImg) throw new Error("lightbox img missing");
+    fireEvent.click(lightboxImg);
+    expect(screen.queryByRole("dialog")).toBeTruthy();
+
+    // Click on the backdrop (the dialog itself) — closes.
+    fireEvent.click(dialog);
+    await waitFor(() => {
+      expect(screen.queryByRole("dialog")).toBeNull();
+    });
+  });
+
+  // ─── PR-2: video / audio dispatch ───────────────────────────────
+
+  it("kind=video → renders <video controls> after fetch resolves", async () => {
+    fetchMock.mockResolvedValue({
+      ok: true,
+      blob: async () => new Blob(["fake-mp4"], { type: "video/mp4" }),
+    });
+    preview({ uri: "workspace:/workspace/clip.mp4", name: "clip.mp4", mimeType: "video/mp4" });
+    // Loading placeholder first.
+    expect(await screen.findByLabelText(/Loading clip\.mp4/i)).toBeTruthy();
+    // After the blob resolves, a <video> element with controls=true
+    // is in the DOM. Use a tag query — there's no built-in role for
+    // <video>, but the element is unambiguous in the bubble.
+    await waitFor(() => {
+      const v = document.querySelector("video");
+      expect(v).not.toBeNull();
+      // controls attribute pinned — without it the user can't play.
+      expect(v?.hasAttribute("controls")).toBe(true);
+      // src is the blob URL we minted.
+      expect((v as HTMLVideoElement).src).toBe("blob:test-url");
+    });
+    // Chip MUST NOT render — proves dispatch routed to video, not file.
+    expect(screen.queryByTitle(/Download clip\.mp4/i)).toBeNull();
+  });
+
+  it("kind=video fetch fails → falls back to AttachmentChip", async () => {
+    fetchMock.mockResolvedValue({ ok: false, status: 404 });
+    preview({ uri: "workspace:/workspace/missing.mp4", name: "missing.mp4", mimeType: "video/mp4" });
+    await waitFor(() => {
+      expect(screen.getByTitle(/Download missing\.mp4/i)).toBeTruthy();
+    });
+  });
+
+  it("kind=video by extension fallback (no mime) → video path", async () => {
+    fetchMock.mockReturnValue(new Promise(() => {}));
+    preview({ uri: "workspace:/workspace/recording.webm", name: "recording.webm" });
+    expect(await screen.findByLabelText(/Loading recording\.webm/i)).toBeTruthy();
+  });
+
+  it("kind=audio → renders <audio controls> with filename label", async () => {
+    fetchMock.mockResolvedValue({
+      ok: true,
+      blob: async () => new Blob(["fake-mp3"], { type: "audio/mpeg" }),
+    });
+    preview({ uri: "workspace:/workspace/song.mp3", name: "song.mp3", mimeType: "audio/mpeg" });
+    await waitFor(() => {
+      const a = document.querySelector("audio");
+      expect(a).not.toBeNull();
+      expect(a?.hasAttribute("controls")).toBe(true);
+      expect((a as HTMLAudioElement).src).toBe("blob:test-url");
+    });
+    // Filename label pinned: helps the user know what they're hearing
+    // BEFORE pressing play. Multiple matches — `<span>` text and the
+    // `<audio>`'s fallback `{name}` text node — so getAllByText.
+    expect(screen.getAllByText("song.mp3").length).toBeGreaterThan(0);
+  });
+
+  it("kind=audio fetch fails → falls back to chip", async () => {
+    fetchMock.mockResolvedValue({ ok: false, status: 403 });
+    preview({ uri: "workspace:/workspace/locked.wav", name: "locked.wav", mimeType: "audio/wav" });
+    await waitFor(() => {
+      expect(screen.getByTitle(/Download locked\.wav/i)).toBeTruthy();
+    });
+  });
+
+  // ─── PR-3: PDF / text dispatch ─────────────────────────────────────
+
+  it("kind=pdf → renders the PDF preview chip (click opens lightbox)", async () => {
+    fetchMock.mockResolvedValue({
+      ok: true,
+      blob: async () => new Blob(["%PDF-1.4..."], { type: "application/pdf" }),
+    });
+    preview({ uri: "workspace:/workspace/doc.pdf", name: "doc.pdf", mimeType: "application/pdf" });
+
+    // Loading placeholder first.
+    expect(await screen.findByLabelText(/Loading doc\.pdf/i)).toBeTruthy();
+
+    // After fetch, preview chip with "PDF" tag rendered.
+    await waitFor(() => {
+      // The button title is "Preview doc.pdf"; alongside is a "PDF" tag.
+      expect(screen.getByLabelText(/Open doc\.pdf preview/i)).toBeTruthy();
+    });
+
+    // Click → lightbox opens with <embed> inside.
+    fireEvent.click(screen.getByLabelText(/Open doc\.pdf preview/i));
+    const dialog = await screen.findByRole("dialog");
+    expect(dialog).toBeTruthy();
+    expect(dialog.querySelector("embed[type='application/pdf']")).not.toBeNull();
+  });
+
+  it("kind=pdf fetch fails → falls back to chip", async () => {
+    fetchMock.mockResolvedValue({ ok: false, status: 404 });
+    preview({ uri: "workspace:/workspace/missing.pdf", name: "missing.pdf", mimeType: "application/pdf" });
+    await waitFor(() => {
+      expect(screen.getByTitle(/Download missing\.pdf/i)).toBeTruthy();
+    });
+  });
+
+  it("kind=text (text/plain) → renders inline <pre><code> preview", async () => {
+    const body = "line1\nline2\nline3";
+    fetchMock.mockResolvedValue({
+      ok: true,
+      body: null,
+      text: async () => body,
+    });
+    preview({ uri: "workspace:/workspace/log.txt", name: "log.txt", mimeType: "text/plain" });
+
+    // testing-library normalizes whitespace by default. The <pre>
+    // contains the literal text node, so query the DOM directly.
+    await waitFor(() => {
+      const code = document.querySelector("pre code");
+      expect(code).not.toBeNull();
+      expect(code?.textContent).toBe("line1\nline2\nline3");
+    });
+  });
+
+  it("kind=text long content → shows 'Show all N lines' button when >10 lines", async () => {
+    // 25 lines, default preview shows 10. Button labels with full count.
+    const body = Array.from({ length: 25 }, (_, i) => `line ${i + 1}`).join("\n");
+    fetchMock.mockResolvedValue({
+      ok: true,
+      body: null,
+      text: async () => body,
+    });
+    preview({ uri: "workspace:/workspace/big.txt", name: "big.txt", mimeType: "text/plain" });
+
+    await waitFor(() => {
+      expect(screen.getByRole("button", { name: /Show all 25 lines/i })).toBeTruthy();
+    });
+    // Pre-expand: only first 10 lines in <code>; line 11+ absent.
+    let code = document.querySelector("pre code");
+    expect(code?.textContent?.includes("line 10")).toBe(true);
+    expect(code?.textContent?.includes("line 11")).toBe(false);
+
+    // After clicking expand, all 25 lines present.
+    fireEvent.click(screen.getByRole("button", { name: /Show all 25 lines/i }));
+    await waitFor(() => {
+      code = document.querySelector("pre code");
+      expect(code?.textContent?.includes("line 25")).toBe(true);
+    });
+  });
+
+  it("kind=text fetch fails → chip fallback", async () => {
+    fetchMock.mockResolvedValue({ ok: false, status: 404 });
+    preview({ uri: "workspace:/workspace/missing.json", name: "missing.json", mimeType: "application/json" });
+    await waitFor(() => {
+      expect(screen.getByTitle(/Download missing\.json/i)).toBeTruthy();
+    });
+  });
+
+  // ─── universal-fallback regression ─────────────────────────────────
+
+  it("kind=file is the universal fallback for unknown MIME (regression: don't try to preview a zip)", () => {
+    // Critical safety: agent could attach a misnamed file. Pre-fix
+    // the chip path was unconditional; we want unknown MIME to
+    // STILL go to the chip even though the extension matches an
+    // image kind.
+    preview({ uri: "workspace:/workspace/tmp/x.docx", name: "x.docx", mimeType: "application/vnd.zip-disguised-as-doc" });
+    expect(screen.getByTitle(/Download x\.docx/i)).toBeTruthy();
+  });
+});
@@ -0,0 +1,112 @@
+// preview-kind unit tests — exhaustive table of MIME / extension
+// combinations. The kind helper is a pure function; this is the
+// regression line for "what renders as what" across the entire chat
+// surface.
+
+import { describe, it, expect } from "vitest";
+import { getAttachmentPreviewKind } from "../preview-kind";
+
+describe("getAttachmentPreviewKind", () => {
+  describe("strict MIME match", () => {
+    const cases: Array<[string, ReturnType<typeof getAttachmentPreviewKind>]> = [
+      // images
+      ["image/png", "image"],
+      ["image/jpeg", "image"],
+      ["image/gif", "image"],
+      ["image/webp", "image"],
+      ["image/svg+xml", "image"],
+      ["image/avif", "image"],
+      ["IMAGE/PNG", "image"], // case-insensitive
+      ["  image/png  ", "image"], // trim
+      // video
+      ["video/mp4", "video"],
+      ["video/webm", "video"],
+      ["video/quicktime", "video"],
+      // audio
+      ["audio/mpeg", "audio"],
+      ["audio/wav", "audio"],
+      ["audio/ogg", "audio"],
+      // pdf
+      ["application/pdf", "pdf"],
+      // text family
+      ["text/plain", "text"],
+      ["text/markdown", "text"],
+      ["text/html", "text"],
+      ["text/css", "text"],
+      ["text/javascript", "text"],
+      ["text/csv", "text"],
+      ["application/json", "text"],
+      ["application/yaml", "text"],
+      ["application/x-yaml", "text"],
+      ["application/javascript", "text"],
+      ["application/typescript", "text"],
+      // unknown / non-renderable → file
+      ["application/zip", "file"],
+      ["application/octet-stream", "file"],
+      ["application/x-tar", "file"],
+      ["application/vnd.ms-excel", "file"],
+      ["weird/unknown-thing", "file"],
+    ];
+    for (const [mime, expected] of cases) {
+      it(`mimeType=${JSON.stringify(mime)} → ${expected}`, () => {
+        expect(getAttachmentPreviewKind(mime)).toBe(expected);
+      });
+    }
+  });
+
+  describe("extension fallback when MIME is missing or generic", () => {
+    const cases: Array<[string | undefined, string | undefined, string | undefined, ReturnType<typeof getAttachmentPreviewKind>]> = [
+      // [mime, uri, name, expected]
+      [undefined, "workspace:/tmp/screenshot.png", "screenshot.png", "image"],
+      ["", "workspace:/tmp/photo.JPG", "photo.JPG", "image"],
+      ["application/octet-stream", "workspace:/tmp/clip.mp4", "clip.mp4", "video"],
+      [undefined, "workspace:/foo/song.mp3", "song.mp3", "audio"],
+      [undefined, "workspace:/docs/report.pdf", "report.pdf", "pdf"],
+      [undefined, "workspace:/code/main.py", "main.py", "text"],
+      [undefined, "workspace:/data/notes.md", "notes.md", "text"],
+      // No extension → file
+      [undefined, "workspace:/tmp/Dockerfile", "Dockerfile", "file"],
+      // Trailing dot → file
+      [undefined, "workspace:/tmp/weird.", "weird.", "file"],
+      // URL with query string + fragment → strip before parsing
+      [undefined, "https://example.com/foo.png?download=1#anchor", "", "image"],
+      // Unknown extension → file
+      [undefined, "workspace:/tmp/something.xyz", "something.xyz", "file"],
+      // Empty
+      [undefined, "", "", "file"],
+      [undefined, undefined, undefined, "file"],
+    ];
+    for (const [mime, uri, name, expected] of cases) {
+      it(`mime=${mime ?? "<undef>"} uri=${uri} name=${name} → ${expected}`, () => {
+        expect(getAttachmentPreviewKind(mime, uri, name)).toBe(expected);
+      });
+    }
+  });
+
+  describe("MIME wins over extension", () => {
+    it("explicit mime=application/zip + extension=.png → file (don't render zip as image)", () => {
+      // Critical safety: agent might attach a .png-named file that's
+      // actually a zip. The strict-MIME branch wins and we render
+      // the chip, not an <img> that 404s on broken bytes.
+      expect(getAttachmentPreviewKind("application/zip", "x.png", "x.png")).toBe("file");
+    });
+
+    it("explicit mime=text/plain + extension=.png → text", () => {
+      expect(getAttachmentPreviewKind("text/plain", "log.png", "log.png")).toBe("text");
+    });
+  });
+
+  describe("regression: hostile-reviewer cases", () => {
+    it("does NOT misclassify image/svg+xml as text (svg is image even though it has XML)", () => {
+      expect(getAttachmentPreviewKind("image/svg+xml")).toBe("image");
+    });
+
+    it("application/octet-stream + extension=.docx → file (no renderer, don't try)", () => {
+      expect(getAttachmentPreviewKind("application/octet-stream", "f.docx", "f.docx")).toBe("file");
+    });
+
+    it("non-canonical MIME application/json works", () => {
+      expect(getAttachmentPreviewKind("application/json")).toBe("text");
+    });
+  });
+});
@@ -1,5 +1,5 @@
 import { describe, it, expect } from "vitest";
-import { resolveAttachmentHref } from "../uploads";
+import { isPlatformAttachment, resolveAttachmentHref } from "../uploads";

 describe("resolveAttachmentHref — URI scheme normalisation", () => {
  const wsId = "aaaaaaaa-bbbb-cccc-dddd-eeeeeeeeeeee";
@@ -39,3 +39,128 @@ describe("resolveAttachmentHref — URI scheme normalisation", () => {
    expect(resolveAttachmentHref(wsId, "s3://bucket/key")).toBe("s3://bucket/key");
  });
 });
+
+// #2973 follow-up to #2968: cover the platform-pending: scheme branch
+// (poll-mode chat uploads) + the isPlatformAttachment SSOT helper that
+// the chip-download and markdown-link paths both consume.
+//
+// Pre-fix the platform-pending: URI fell through to the raw URI →
+// browser saw an unhandled-protocol click → about:blank. The fix
+// resolves it to the platform pending-uploads endpoint with auth
+// headers attached.
+describe("resolveAttachmentHref — platform-pending: scheme (poll-mode uploads)", () => {
+  // Use a chat workspace ID that DIFFERS from the one in the URI, so
+  // tests can verify which one the resolver uses. The forward-across-
+  // workspace case is real production behavior — files dragged into one
+  // workspace's chat can be referenced from another.
+  const chatWs = "chat-ws-aaaaaaaa";
+  const sourceWs = "source-ws-bbbbbbbb";
+
+  it("resolves a well-formed platform-pending: URI to /pending-uploads/<file>/content", () => {
+    const url = resolveAttachmentHref(
+      chatWs,
+      `platform-pending:${sourceWs}/file-12345`,
+    );
+    expect(url).toContain(`/workspaces/${sourceWs}/pending-uploads/file-12345/content`);
+  });
+
+  it("uses the URI's wsid, NOT the chat workspace_id (cross-workspace forwarding)", () => {
+    // The two ids differ — this is the case PR #2968's commit
+    // explicitly calls out. A regression that flipped this would
+    // silently mis-route the download to the WRONG workspace's
+    // pending-uploads store, returning 404 (or worse, leaking).
+    const url = resolveAttachmentHref(
+      chatWs,
+      `platform-pending:${sourceWs}/file-xyz`,
+    );
+    expect(url).toContain(`/workspaces/${sourceWs}/`);
+    expect(url).not.toContain(`/workspaces/${chatWs}/`);
+  });
+
+  it("falls back to raw URI when platform-pending: is missing the slash", () => {
+    // Defensive: a URI that drifted from the expected wsid/fileid shape
+    // returns raw rather than producing a broken /pending-uploads//
+    // path. Pinned to detect a regression where a future "helpful"
+    // change synthesizes empty wsid/fileID.
+    expect(resolveAttachmentHref(chatWs, "platform-pending:no-slash")).toBe(
+      "platform-pending:no-slash",
+    );
+  });
+
+  it("falls back to raw URI when platform-pending: has empty fileID", () => {
+    expect(resolveAttachmentHref(chatWs, "platform-pending:abc/")).toBe(
+      "platform-pending:abc/",
+    );
+  });
+
+  it("falls back to raw URI when platform-pending: has empty wsid", () => {
+    expect(resolveAttachmentHref(chatWs, "platform-pending:/file-xyz")).toBe(
+      "platform-pending:/file-xyz",
+    );
+  });
+
+  it("regression: exact production repro from #2968 (reno-stars)", () => {
+    // From the original PR #2968 body: the chat's markdown-link
+    // override fell through on this exact shape and the browser
+    // navigated to about:blank. Pin the post-fix output so a future
+    // refactor can't reintroduce the original bug.
+    const url = resolveAttachmentHref(
+      "chat-ws",
+      "platform-pending:d76977b1-uuid/bb0dcaf3-uuid",
+    );
+    expect(url).toContain("/workspaces/d76977b1-uuid/pending-uploads/bb0dcaf3-uuid/content");
+    expect(url).not.toContain("chat-ws");
+  });
+});
+
+describe("isPlatformAttachment", () => {
+  it("returns true for platform-pending: URIs", () => {
+    expect(isPlatformAttachment("platform-pending:abc/file")).toBe(true);
+  });
+
+  it("returns true even for malformed platform-pending: URIs", () => {
+    // The helper is a SHAPE check — caller routes through
+    // downloadChatFile and downloadChatFile handles the malformed case
+    // downstream. Pinning so a future helper that "validates" the
+    // wsid/fileID shape doesn't silently break the auth-attached
+    // download flow for in-flight URIs.
+    expect(isPlatformAttachment("platform-pending:no-slash")).toBe(true);
+  });
+
+  it("returns true for workspace:<allowed-root> URIs", () => {
+    expect(isPlatformAttachment("workspace:/configs/foo")).toBe(true);
+    expect(isPlatformAttachment("workspace:/workspace/x.pdf")).toBe(true);
+  });
+
+  it("returns true for file:///<allowed-root> URIs", () => {
+    expect(isPlatformAttachment("file:///workspace/x")).toBe(true);
+  });
+
+  it("returns true for absolute paths under allowed roots", () => {
+    expect(isPlatformAttachment("/home/user/x")).toBe(true);
+    expect(isPlatformAttachment("/configs/y")).toBe(true);
+  });
+
+  it("returns FALSE for bare HTTPS URLs to other origins", () => {
+    // Auth-leak class regression: a helper that always returned true
+    // would attach workspace tokens to third-party requests. Pin
+    // the negative case explicitly.
+    expect(isPlatformAttachment("https://example.com/file")).toBe(false);
+    expect(isPlatformAttachment("http://example.com/file")).toBe(false);
+  });
+
+  it("returns FALSE for non-allowlisted root paths", () => {
+    expect(isPlatformAttachment("/etc/passwd")).toBe(false);
+    expect(isPlatformAttachment("/var/log/x")).toBe(false);
+    expect(isPlatformAttachment("/tmp/x")).toBe(false);
+  });
+
+  it("returns FALSE for empty string", () => {
+    expect(isPlatformAttachment("")).toBe(false);
+  });
+
+  it("returns FALSE for unrecognised schemes", () => {
+    expect(isPlatformAttachment("s3://bucket/key")).toBe(false);
+    expect(isPlatformAttachment("ftp://server/file")).toBe(false);
+  });
+});
@@ -0,0 +1,154 @@
+// preview-kind.ts — single source of truth for "what renderer should
+// this attachment use" (RFC #2991, PR-1).
+//
+// Per the RFC's Phase 2 design, MIME type is the dispatch axis. The
+// wire shape (ChatAttachment.mimeType) already carries it end-to-end
+// from the server's chat_files.go through agent_message_writer.go to
+// the canvas hydrater — we just need to map it to a render kind.
+//
+// Why a separate file from AttachmentPreview.tsx: the kind helper is
+// a pure function that's easier to unit-test in isolation than a
+// React component, and unit tests across MIME families are the
+// regression line for new types added later.
+
+/** The render-kind taxonomy. Each kind has a dedicated component:
+ *
+ *    image  → AttachmentImage (inline thumbnail + click → lightbox)
+ *    video  → AttachmentVideo (HTML5 <video controls>, native fullscreen)
+ *    audio  → AttachmentAudio (HTML5 <audio controls>)
+ *    pdf    → AttachmentPDF (browser-native <embed>, fullscreen modal)
+ *    text   → AttachmentTextPreview (monospace, first N lines, expand)
+ *    file   → AttachmentChip (existing fallback — generic file pill)
+ *
+ * NB: `text` includes JSON, YAML, source code, plain text — anything
+ * that renders sensibly as preformatted ASCII without a specialized
+ * viewer. PR-1 ships only `image` + `file`; PR-2 adds video/audio;
+ * PR-3 adds pdf + text. All routed through this same dispatch table
+ * so adding a new kind is a one-line registration. */
+export type AttachmentPreviewKind = "image" | "video" | "audio" | "pdf" | "text" | "file";
+
+/** Maps a MIME type to the render kind. Falls back to "file" for
+ *  any MIME we don't have a renderer for (current behavior — the
+ *  attachment chip is the universal fallback).
+ *
+ *  Filename-based fallback: when mimeType is missing or generic
+ *  (application/octet-stream), inspect the URI's extension. The
+ *  workspace-server's chat_files.go derives Content-Type from the
+ *  file extension, but agent-emitted attachments may not always
+ *  set mimeType, and the canvas should still preview a file named
+ *  `screenshot.png` even if the wire shape lacks the MIME.
+ *
+ *  Strict MIME match always wins; extension fallback only applies
+ *  to empty / generic. Unknown extension → "file". */
+export function getAttachmentPreviewKind(
+  mimeType: string | undefined,
+  uri?: string,
+  name?: string,
+): AttachmentPreviewKind {
+  const mime = (mimeType ?? "").toLowerCase().trim();
+
+  // Strict MIME match (preferred — set by server's Content-Type
+  // detection or by the agent's explicit mimeType field).
+  if (mime.startsWith("image/")) return "image";
+  if (mime.startsWith("video/")) return "video";
+  if (mime.startsWith("audio/")) return "audio";
+  if (mime === "application/pdf") return "pdf";
+  if (
+    mime.startsWith("text/") ||
+    mime === "application/json" ||
+    mime === "application/yaml" ||
+    mime === "application/x-yaml" ||
+    mime === "application/javascript" ||
+    mime === "application/typescript"
+  ) {
+    return "text";
+  }
+
+  // Extension-based fallback — only when MIME is missing or
+  // application/octet-stream (the server's "I don't know" default).
+  // Skip when MIME is set to something specific we just don't have
+  // a renderer for (e.g. application/zip → file is correct).
+  const looksGeneric = mime === "" || mime === "application/octet-stream";
+  if (looksGeneric) {
+    const ext = extractExtension(uri, name);
+    if (ext) {
+      const kind = EXTENSION_KIND.get(ext);
+      if (kind) return kind;
+    }
+  }
+
+  return "file";
+}
+
+// Extension → kind table for the fallback branch. Keep this list
+// short and curated — every entry is a UX commitment to render
+// inline, and a wrong inference (e.g. .doc rendered as text) is
+// worse than the generic file chip.
+const EXTENSION_KIND: ReadonlyMap<string, AttachmentPreviewKind> = new Map([
+  // Images
+  ["png", "image"],
+  ["jpg", "image"],
+  ["jpeg", "image"],
+  ["gif", "image"],
+  ["webp", "image"],
+  ["svg", "image"],
+  ["avif", "image"],
+  ["bmp", "image"],
+  // Video
+  ["mp4", "video"],
+  ["webm", "video"],
+  ["mov", "video"],
+  ["mkv", "video"],
+  // Audio
+  ["mp3", "audio"],
+  ["wav", "audio"],
+  ["ogg", "audio"],
+  ["m4a", "audio"],
+  ["flac", "audio"],
+  // PDF
+  ["pdf", "pdf"],
+  // Text-ish (rendered as preformatted ASCII)
+  ["txt", "text"],
+  ["md", "text"],
+  ["json", "text"],
+  ["yaml", "text"],
+  ["yml", "text"],
+  ["js", "text"],
+  ["ts", "text"],
+  ["tsx", "text"],
+  ["jsx", "text"],
+  ["py", "text"],
+  ["go", "text"],
+  ["rs", "text"],
+  ["java", "text"],
+  ["c", "text"],
+  ["cpp", "text"],
+  ["h", "text"],
+  ["hpp", "text"],
+  ["sh", "text"],
+  ["bash", "text"],
+  ["html", "text"],
+  ["css", "text"],
+  ["sql", "text"],
+  ["toml", "text"],
+  ["ini", "text"],
+  ["xml", "text"],
+  ["csv", "text"],
+  ["log", "text"],
+]);
+
+/** Extracts the lowercased extension from a uri or name, without
+ *  the leading dot. Returns "" when no extension is present. */
+function extractExtension(uri: string | undefined, name: string | undefined): string {
+  // Prefer name (always a leaf path); fall back to uri's last
+  // segment. Strip query string + fragment so a URI like
+  // "https://example.com/foo.png?download=1" still parses as png.
+  const candidate = name || uri || "";
+  if (!candidate) return "";
+  let leaf = candidate.split(/[\\/]/).pop() || "";
+  // Drop ?query and #fragment.
+  leaf = leaf.split(/[?#]/)[0];
+  const dot = leaf.lastIndexOf(".");
+  if (dot < 0 || dot === leaf.length - 1) return "";
+  return leaf.slice(dot + 1).toLowerCase();
+}
@@ -44,6 +44,8 @@ export async function uploadChatFiles(
 *    - `workspace:<abs-path>` (our canonical form)
 *    - `file:///workspace/...` (some agents emit this)
 *    - `/workspace/...` (bare absolute path inside the container)
+ *    - `platform-pending:<wsid>/<file_id>` (poll-mode upload, staged
+ *      on platform side; resolves to /pending-uploads/<file_id>/content)
 *  Everything that looks like an allowed-root container path is
 *  rewritten to the authenticated /chat/download endpoint. HTTP(S)
 *  URIs pass through unchanged so we can also render links to
@@ -53,6 +55,35 @@ export function resolveAttachmentHref(
  workspaceId: string,
  uri: string,
 ): string {
+  // platform-pending: agents-emitted URI that lives in the platform-side
+  // staging layer (poll-mode chat uploads, see workspace-server's
+  // chat_files.go ~line 690 + pendinguploads.Storage). The wire shape
+  // is `platform-pending:<workspace_id>/<file_id>`. Resolving it
+  // requires hitting GET /workspaces/<wsid>/pending-uploads/<file_id>/content
+  // which streams the bytes with full workspace auth. Without this
+  // case the browser sees an unhandled-protocol click → about:blank,
+  // which was the user-visible bug from 2026-05-05 (reno-stars).
+  if (uri.startsWith("platform-pending:")) {
+    const rest = uri.slice("platform-pending:".length);
+    const slash = rest.indexOf("/");
+    // Defensive: if the URI doesn't have the expected wsid/fileid
+    // shape, fall through to raw-URI handling so the consumer can
+    // still try to render it (rather than producing a broken /pending-
+    // uploads/// path).
+    if (slash > 0) {
+      const wsid = rest.slice(0, slash);
+      const fileID = rest.slice(slash + 1);
+      if (wsid && fileID) {
+        // Use the URI's own workspace_id (the bytes live in THAT
+        // workspace's pending-uploads store), not the chat's
+        // workspace_id — these CAN differ when a user drags a file
+        // into one workspace's chat that gets forwarded to another
+        // (cross-workspace delegation, agent forwarding).
+        return `${PLATFORM_URL}/workspaces/${wsid}/pending-uploads/${fileID}/content`;
+      }
+    }
+    return uri;
+  }
  const containerPath = normalizeWorkspaceUri(uri);
  if (containerPath) {
    return `${PLATFORM_URL}/workspaces/${workspaceId}/chat/download?path=${encodeURIComponent(containerPath)}`;
@@ -60,6 +91,14 @@ export function resolveAttachmentHref(
  return uri;
 }

+/** Returns true when the URI points at a platform-side resource that
+ *  requires our auth headers — caller should route through
+ *  downloadChatFile rather than letting the browser navigate. */
+export function isPlatformAttachment(uri: string): boolean {
+  if (uri.startsWith("platform-pending:")) return true;
+  return normalizeWorkspaceUri(uri) !== null;
+}
+
 /** Extracts the absolute container path from a workspace-scoped URI,
 *  or null if the URI isn't a container path. The matching roots
 *  mirror the server's `allowedRoots` allowlist. */
@@ -96,8 +135,7 @@ export async function downloadChatFile(
  attachment: ChatAttachment,
 ): Promise<void> {
  const href = resolveAttachmentHref(workspaceId, attachment.uri);
-  const isContainerPath = normalizeWorkspaceUri(attachment.uri) !== null;
-  if (!isContainerPath) {
+  if (!isPlatformAttachment(attachment.uri)) {
    // External URL — let the browser navigate. Opens in new tab so
    // the canvas context survives a navigation. `href` here is the
    // raw URI (http(s), or anything else the agent sent back).
@@ -22,7 +22,6 @@ export interface ConfigData {
  // task_budget maps to output_config.task_budget.total (requires beta header task-budgets-2026-03-13)
  task_budget?: number;
  prompt_files: string[];
-  shared_context: string[];
  skills: string[];
  tools: string[];
  a2a: { port: number; streaming: boolean; push_notifications: boolean };
@@ -40,7 +39,6 @@ export const DEFAULT_CONFIG: ConfigData = {
  effort: "",
  task_budget: 0,
  prompt_files: [],
-  shared_context: [],
  skills: [],
  tools: [],
  a2a: { port: 8000, streaming: true, push_notifications: true },
@@ -120,7 +120,6 @@ export function toYaml(config: ConfigData): string {
  if (config.effort) { lines.push(""); simple("effort", config.effort); }
  if (config.task_budget && config.task_budget > 0) { simple("task_budget", config.task_budget); }
  if (config.prompt_files?.length) { lines.push(""); list("prompt_files", config.prompt_files); }
-  if (config.shared_context?.length) { lines.push(""); list("shared_context", config.shared_context); }
  lines.push(""); list("skills", config.skills);
  if (config.tools?.length) { list("tools", config.tools); }
  lines.push(""); obj("a2a", config.a2a as unknown as Record<string, unknown>);
@@ -2,7 +2,7 @@
 * @vitest-environment jsdom
 */
 import { describe, it, expect, vi, afterEach } from "vitest";
-import { fetchSession, redirectToLogin } from "../auth";
+import { fetchSession, redirectToLogin, signOut } from "../auth";

 afterEach(() => {
  vi.unstubAllGlobals();
@@ -110,3 +110,157 @@ describe("redirectToLogin", () => {
    expect((window.location as unknown as { href: string }).href).toBe(signupHref);
  });
 });
+
+describe("signOut", () => {
+  // Helper — most tests need the same window.location stub.
+  function stubLocation(): void {
+    Object.defineProperty(window, "location", {
+      writable: true,
+      value: {
+        href: "https://acme.moleculesai.app/orgs",
+        pathname: "/orgs",
+        hostname: "acme.moleculesai.app",
+        protocol: "https:",
+      },
+    });
+  }
+
+  it("POSTs to /cp/auth/signout with credentials:include", async () => {
+    stubLocation();
+    const fetchMock = vi.fn().mockResolvedValue({
+      ok: true,
+      status: 200,
+      json: async () => ({ ok: true, logout_url: "" }),
+    });
+    vi.stubGlobal("fetch", fetchMock);
+
+    await signOut();
+
+    expect(fetchMock).toHaveBeenCalledTimes(1);
+    expect(fetchMock).toHaveBeenCalledWith(
+      expect.stringContaining("/cp/auth/signout"),
+      expect.objectContaining({ method: "POST", credentials: "include" }),
+    );
+  });
+
+  it("navigates to provider logout_url when the response includes one", async () => {
+    // The hosted-logout path is what actually breaks the SSO re-auth
+    // loop reported on PR #2913. Without this, AuthKit's browser
+    // cookie keeps the user signed in via SSO and any subsequent
+    // /cp/auth/login silently re-auths.
+    stubLocation();
+    const hostedLogout =
+      "https://api.workos.com/user_management/sessions/logout?session_id=cookie&return_to=https%3A%2F%2Fapp.moleculesai.app%2Forgs";
+    vi.stubGlobal(
+      "fetch",
+      vi.fn().mockResolvedValue({
+        ok: true,
+        status: 200,
+        json: async () => ({ ok: true, logout_url: hostedLogout }),
+      }),
+    );
+
+    await signOut();
+
+    const after = (window.location as unknown as { href: string }).href;
+    expect(after).toBe(hostedLogout);
+  });
+
+  it("falls back to /cp/auth/login when logout_url is empty (DisabledProvider / dev)", async () => {
+    // DisabledProvider returns "" — the local /cp/auth/login redirect
+    // works in dev/test where there's no SSO session to escape.
+    stubLocation();
+    vi.stubGlobal(
+      "fetch",
+      vi.fn().mockResolvedValue({
+        ok: true,
+        status: 200,
+        json: async () => ({ ok: true, logout_url: "" }),
+      }),
+    );
+
+    await signOut();
+
+    const after = (window.location as unknown as { href: string }).href;
+    // Tenant subdomain (acme.moleculesai.app) → auth origin is app.moleculesai.app.
+    expect(after).toBe("https://app.moleculesai.app/cp/auth/login");
+  });
+
+  it("redirects even when the POST fails so the user isn't stuck on an authed page", async () => {
+    // Critical UX invariant: clicking 'Sign out' MUST navigate away from
+    // the authenticated app, even if the network is down or the cookie
+    // is already invalid. Anything else looks like the button is
+    // broken — the precise complaint that triggered this fix.
+    stubLocation();
+    vi.stubGlobal("fetch", vi.fn().mockRejectedValue(new Error("network down")));
+
+    await signOut();
+
+    const after = (window.location as unknown as { href: string }).href;
+    expect(after).toBe("https://app.moleculesai.app/cp/auth/login");
+  });
+
+  it("redirects on 401 (session already invalid) just like 200", async () => {
+    // A user with an already-invalid cookie should still see the
+    // logout flow complete — no error, no stuck-on-app dead end.
+    // Note: 401 means res.ok=false → we don't read .json() at all,
+    // so a missing body is fine.
+    stubLocation();
+    vi.stubGlobal(
+      "fetch",
+      vi.fn().mockResolvedValue({
+        ok: false,
+        status: 401,
+        json: async () => ({}),
+      }),
+    );
+
+    await signOut();
+
+    const after = (window.location as unknown as { href: string }).href;
+    expect(after).toBe("https://app.moleculesai.app/cp/auth/login");
+  });
+
+  it("falls back to /cp/auth/login when the response body is malformed", async () => {
+    // Defensive parsing: a body that isn't valid JSON, or doesn't
+    // have logout_url, or has logout_url as the wrong type — none of
+    // these should strand the user on the authed page. Fallback path
+    // takes over.
+    stubLocation();
+    vi.stubGlobal(
+      "fetch",
+      vi.fn().mockResolvedValue({
+        ok: true,
+        status: 200,
+        json: async () => {
+          throw new Error("not json");
+        },
+      }),
+    );
+
+    await signOut();
+
+    const after = (window.location as unknown as { href: string }).href;
+    expect(after).toBe("https://app.moleculesai.app/cp/auth/login");
+  });
+
+  it("falls back to /cp/auth/login when logout_url is the wrong type", async () => {
+    // Even valid JSON should be type-checked: a non-string logout_url
+    // (e.g. server-side bug, version drift) must not crash or open-
+    // redirect the user.
+    stubLocation();
+    vi.stubGlobal(
+      "fetch",
+      vi.fn().mockResolvedValue({
+        ok: true,
+        status: 200,
+        json: async () => ({ ok: true, logout_url: 42 }),
+      }),
+    );
+
+    await signOut();
+
+    const after = (window.location as unknown as { href: string }).href;
+    expect(after).toBe("https://app.moleculesai.app/cp/auth/login");
+  });
+});
@@ -67,3 +67,80 @@ export function redirectToLogin(screenHint: "sign-up" | "sign-in" = "sign-in"):
  const dest = `${authOrigin}${AUTH_BASE}/${path}?return_to=${encodeURIComponent(returnTo)}`;
  window.location.href = dest;
 }
+
+/**
+ * signOut posts to /cp/auth/signout to clear the WorkOS session cookie
+ * + revoke at the provider, then navigates the browser to the
+ * provider-supplied hosted logout URL (so the provider's BROWSER-side
+ * SSO cookie is cleared too — without this, AuthKit silently re-auths
+ * via SSO on the next /cp/auth/login and the user is "still signed
+ * in" after pressing Sign out).
+ *
+ * Two-layer flow:
+ *  1. POST /cp/auth/signout → CP clears OUR session cookie + revokes
+ *     session_id at the provider API. Response includes
+ *     `logout_url` — the AuthKit hosted URL the BROWSER must navigate
+ *     to so the provider's own browser cookie is cleared.
+ *  2. window.location.href = <logout_url> → AuthKit clears its
+ *     session, then redirects the browser to the configured
+ *     return_to (defaults to APP_URL/orgs).
+ *
+ * Best-effort by design: a 5xx, network failure, missing logout_url
+ * (DisabledProvider, dev), or stale cookie still results in the
+ * browser navigating away — leaving the user on a logged-in-looking
+ * page after they clicked "Sign out" is the worst possible UX. The
+ * fallback path navigates to /cp/auth/login on the auth origin, which
+ * works correctly in environments without a hosted logout flow (dev,
+ * tests, DisabledProvider).
+ *
+ * Throws nothing — callers can disable the button optimistically or
+ * await this and trust it returns. On a redirect-blocked test
+ * environment (jsdom under vitest) we still exit cleanly so unit tests
+ * can spy on the fetch call.
+ */
+export async function signOut(): Promise<void> {
+  let logoutURL: string | undefined;
+  // Fire-and-tolerate the POST. credentials:include is mandatory cross-
+  // origin so the SaaS canvas (acme.moleculesai.app) can hit
+  // app.moleculesai.app/cp/auth/signout with the session cookie.
+  try {
+    const res = await fetch(`${getAuthOrigin()}${AUTH_BASE}/signout`, {
+      method: "POST",
+      credentials: "include",
+    });
+    if (res.ok) {
+      // Body shape: {"ok": true, "logout_url": "..."}. logout_url is
+      // empty for DisabledProvider (dev/local) — we fall back to
+      // /cp/auth/login below. Defensive parsing: a malformed body
+      // shouldn't strand the user on the authed page.
+      const body: unknown = await res.json().catch(() => null);
+      if (
+        body &&
+        typeof body === "object" &&
+        "logout_url" in body &&
+        typeof (body as { logout_url: unknown }).logout_url === "string" &&
+        (body as { logout_url: string }).logout_url
+      ) {
+        logoutURL = (body as { logout_url: string }).logout_url;
+      }
+    }
+  } catch {
+    // Ignore — we still redirect below.
+  }
+  if (typeof window === "undefined") return;
+  if (logoutURL) {
+    // Hosted logout: AuthKit clears its SSO cookie + redirects to
+    // return_to (configured server-side). This is the path that
+    // actually breaks the SSO re-auth loop.
+    window.location.href = logoutURL;
+    return;
+  }
+  // Fallback: no hosted logout (dev, DisabledProvider, network
+  // failure). Land on the login screen rather than the current URL:
+  // returning to a tenant URL after signout would just re-redirect
+  // through /cp/auth/login due to AuthGate. Send the user straight
+  // there with no return_to so they don't loop back into the org they
+  // just left.
+  const authOrigin = getAuthOrigin();
+  window.location.href = `${authOrigin}${AUTH_BASE}/login`;
+}
@@ -5,6 +5,13 @@ export const STATUS_CONFIG: Record<string, { dot: string; glow: string; label: s
  degraded: { dot: "bg-amber-400", glow: "shadow-amber-400/50", label: "Degraded", bar: "from-amber-500/20 to-transparent" },
  failed: { dot: "bg-red-400", glow: "shadow-red-400/50", label: "Failed", bar: "from-red-500/20 to-transparent" },
  provisioning: { dot: "bg-sky-400 motion-safe:animate-pulse", glow: "shadow-sky-400/50", label: "Starting", bar: "from-sky-500/20 to-transparent" },
+  // not_configured: derived state from agent_card.configuration_status (PR #2756 chain).
+  // Workspace is reachable (heartbeating, /agent-card serves) but adapter.setup()
+  // failed — typically a missing/rotated LLM credential. Amber to differentiate from
+  // online (green) and failed (red) — the workspace itself is healthy, just needs
+  // configuration. Hover renders agent_card.configuration_error in the tooltip so
+  // the operator sees the exact env var to set.
+  not_configured: { dot: "bg-amber-300", glow: "shadow-amber-300/50", label: "Not configured", bar: "from-amber-400/20 to-transparent" },
 };

 export function statusDotClass(status: string): string {
@@ -0,0 +1,103 @@
+import { describe, it, expect } from "vitest";
+import {
+  getConfigurationStatus,
+  getConfigurationError,
+} from "../canvas-topology";
+
+// Tests for the getConfigurationStatus / getConfigurationError helpers
+// (issue #467 / PR #2756 chain). Surfacing the workspace's
+// `agent_card.configuration_status` is the user-visible payoff of
+// PR #2756's decoupling — without it, a misconfigured workspace looks
+// identical to a healthy one in the canvas tile.
+
+describe("getConfigurationStatus", () => {
+  it("returns null when agentCard is null", () => {
+    expect(getConfigurationStatus(null)).toBe(null);
+  });
+
+  it("returns null when agentCard has no configuration_status", () => {
+    expect(getConfigurationStatus({ name: "x" })).toBe(null);
+  });
+
+  it("returns 'ready' when agent reports configuration ok", () => {
+    expect(
+      getConfigurationStatus({ configuration_status: "ready" }),
+    ).toBe("ready");
+  });
+
+  it("returns 'not_configured' when agent reports setup failed", () => {
+    expect(
+      getConfigurationStatus({ configuration_status: "not_configured" }),
+    ).toBe("not_configured");
+  });
+
+  it("ignores unknown values defensively", () => {
+    // A future agent reporting a status string we don't yet recognise
+    // shouldn't crash the canvas — we treat it as 'no info' (null).
+    expect(
+      getConfigurationStatus({ configuration_status: "starting" }),
+    ).toBe(null);
+    expect(
+      getConfigurationStatus({ configuration_status: 42 }),
+    ).toBe(null);
+    expect(
+      getConfigurationStatus({ configuration_status: null }),
+    ).toBe(null);
+  });
+});
+
+describe("getConfigurationError", () => {
+  it("returns null when agentCard is null", () => {
+    expect(getConfigurationError(null)).toBe(null);
+  });
+
+  it("returns null when status is 'ready' even if error string present", () => {
+    // Defensive: if the agent somehow ships configuration_status=ready
+    // alongside a stale configuration_error from a previous boot, we
+    // trust the live status flag and don't surface the stale error.
+    expect(
+      getConfigurationError({
+        configuration_status: "ready",
+        configuration_error: "stale: was unset",
+      }),
+    ).toBe(null);
+  });
+
+  it("returns the error string when status is 'not_configured'", () => {
+    expect(
+      getConfigurationError({
+        configuration_status: "not_configured",
+        configuration_error:
+          "RuntimeError: Neither OPENAI_API_KEY nor MINIMAX_API_KEY is set",
+      }),
+    ).toBe(
+      "RuntimeError: Neither OPENAI_API_KEY nor MINIMAX_API_KEY is set",
+    );
+  });
+
+  it("returns null when status is 'not_configured' but error is missing", () => {
+    expect(
+      getConfigurationError({ configuration_status: "not_configured" }),
+    ).toBe(null);
+  });
+
+  it("returns null when error is empty string", () => {
+    // Empty string isn't actionable for the operator — treat same as
+    // missing.
+    expect(
+      getConfigurationError({
+        configuration_status: "not_configured",
+        configuration_error: "",
+      }),
+    ).toBe(null);
+  });
+
+  it("returns null when error is non-string", () => {
+    expect(
+      getConfigurationError({
+        configuration_status: "not_configured",
+        configuration_error: { reason: "object" },
+      }),
+    ).toBe(null);
+  });
+});
@@ -564,3 +564,42 @@ export function extractSkillNames(agentCard: Record<string, unknown> | null): st
    .map((skill: Record<string, unknown>) => String(skill.name || skill.id || ""))
    .filter(Boolean);
 }
+
+/**
+ * Returns the configuration status reported by the workspace, or null
+ * when the agent card doesn't carry one (older runtime, or pre-PR #2756
+ * worker).
+ *
+ * Pairs with molecule-core PR #2756: when adapter.setup() fails, the
+ * runtime mounts a not-configured handler AND advertises the failure
+ * via agent_card.configuration_status = "not_configured" +
+ * configuration_error = "<reason>". Canvas reads both to render a
+ * "needs config" tile instead of a confused "online but silent" state.
+ *
+ * Returns null (not undefined) so callers can distinguish "no info"
+ * from explicit values via a strict equality check.
+ */
+export function getConfigurationStatus(
+  agentCard: Record<string, unknown> | null,
+): "ready" | "not_configured" | null {
+  if (!agentCard) return null;
+  const raw = agentCard.configuration_status;
+  if (raw === "ready" || raw === "not_configured") return raw;
+  return null;
+}
+
+/**
+ * Returns the configuration error string from the agent card when
+ * configuration_status is "not_configured", or null otherwise.
+ *
+ * Already redacted server-side via secret_redactor (PR #2778) — safe to
+ * render in the UI verbatim.
+ */
+export function getConfigurationError(
+  agentCard: Record<string, unknown> | null,
+): string | null {
+  if (!agentCard) return null;
+  if (getConfigurationStatus(agentCard) !== "not_configured") return null;
+  const raw = agentCard.configuration_error;
+  return typeof raw === "string" && raw.length > 0 ? raw : null;
+}
@@ -7,6 +7,32 @@ export default defineConfig({
  test: {
    environment: 'node',
    exclude: ['e2e/**', 'node_modules/**', '**/dist/**'],
+    // CI-conditional test timeout (issue #96).
+    //
+    // Vitest's 5000ms default is too tight for the first test in any
+    // file under our CI shape: `npx vitest run --coverage` on the
+    // self-hosted Gitea Actions Docker runner. The cold-start cost
+    // (v8 coverage instrumentation init + JSDOM bootstrap + module-
+    // graph import for @/components/* and @/lib/* + first React
+    // render) consistently consumes 5-7 seconds for the first
+    // synchronous test in heavyweight component files
+    // (ActivityTab.test.tsx, CreateWorkspaceDialog.test.tsx,
+    // ConfigTab.provider.test.tsx) — even though every subsequent
+    // test in the same file completes in 100-1500ms.
+    //
+    // Empirically the worst observed first-test was 6453ms in a
+    // single file (CreateWorkspaceDialog). 30000ms gives ~5x
+    // headroom over that on CI; we still keep 5000ms locally so
+    // genuine waitFor races / hung promises stay sensitive in dev.
+    //
+    // Same vitest pattern documented at:
+    //   https://vitest.dev/config/testtimeout
+    //   https://vitest.dev/guide/coverage#profiling-test-performance
+    //
+    // Per-test duration is still emitted to the CI log; if a test
+    // ever silently approaches 25-30s under this raised ceiling that
+    // will surface as a duration regression and we revisit.
+    testTimeout: process.env.CI ? 30000 : 5000,
    // Coverage is instrumented but NOT yet a CI gate — first land
    // observability so we can see the baseline, then dial in
    // thresholds + a hard gate in a follow-up PR (#1815). Today's
@@ -0,0 +1,43 @@
+# docker-compose.dev.yml — overlay over docker-compose.yml for local dev
+# with air-driven live reload of the platform (workspace-server) service.
+#
+# Usage:
+#   docker compose -f docker-compose.yml -f docker-compose.dev.yml up
+#   (or `make dev` shorthand from repo root)
+#
+# What this overlay changes vs docker-compose.yml alone:
+#   - Platform service uses workspace-server/Dockerfile.dev (air on top of
+#     golang:1.25-alpine) instead of the multi-stage prod Dockerfile.
+#   - Platform service bind-mounts the host's workspace-server/ source
+#     into /app/workspace-server so air sees source edits live.
+#   - Other services (postgres, redis, langfuse, etc.) inherit unchanged
+#     from docker-compose.yml.
+#
+# What stays the same:
+#   - All env vars, volumes, depends_on, healthchecks from docker-compose.yml.
+#   - Network topology + ports.
+#   - Postgres/Redis as service containers (no in-process replacements).
+
+services:
+  platform:
+    build:
+      context: .
+      dockerfile: workspace-server/Dockerfile.dev
+    # Rebind source: edits under host's workspace-server/ propagate live.
+    # The named volume on go-build-cache speeds up first build per container.
+    volumes:
+      - ./workspace-server:/app/workspace-server
+      - go-build-cache:/root/.cache/go-build
+      - go-mod-cache:/go/pkg/mod
+    # Air signals the running binary on rebuild; ensure shell stops cleanly.
+    init: true
+    # Mark the service as dev-mode so the platform can short-circuit any
+    # behavior that's incompatible with hot-reload (e.g. background
+    # cron-style watchers that don't survive process restart). No-op
+    # today; reserved for future flag use.
+    environment:
+      MOLECULE_DEV_HOT_RELOAD: "1"
+
+volumes:
+  go-build-cache:
+  go-mod-cache:
@@ -212,8 +212,8 @@ services:
    #   docker compose pull canvas && docker compose up -d canvas
    # First-time local setup or testing unreleased changes — build from source:
    #   docker compose build canvas && docker compose up -d canvas
-    # Note: GHCR images are private — `docker login ghcr.io` required before pull.
-    image: ghcr.io/molecule-ai/canvas:latest
+    # Note: ECR images require AWS auth — `aws ecr get-login-password --region us-east-2 | docker login --username AWS --password-stdin 153263036946.dkr.ecr.us-east-2.amazonaws.com` before pull.
+    image: 153263036946.dkr.ecr.us-east-2.amazonaws.com/molecule-ai/canvas:latest
    build:
      context: ./canvas
      dockerfile: Dockerfile
@@ -0,0 +1,74 @@
+# ADR-002: Local-build mode signalled by `MOLECULE_IMAGE_REGISTRY` presence
+
+* Status: Accepted (2026-05-07)
+* Issue: #63 (closes Task #194)
+* Decision: Hongming (CTO) + Claude Opus 4.7 (implementation)
+
+## Context
+
+Pre-2026-05-06, every Molecule deployment — both production tenants and OSS contributor laptops — pulled workspace-template-* container images from `ghcr.io/molecule-ai/`. Production tenants additionally set `MOLECULE_IMAGE_REGISTRY` to an AWS ECR mirror via Railway env / EC2 user-data, but the OSS default was the upstream GHCR org.
+
+On 2026-05-06 the `Molecule-AI` GitHub org was suspended (saved memory: `feedback_github_botring_fingerprint`). GHCR now returns **403 Forbidden** for every `molecule-ai/workspace-template-*` manifest. OSS contributors who clone `molecule-core` and run `go run ./workspace-server/cmd/server` cannot provision a workspace — every first provision fails with:
+
+```
+docker image "ghcr.io/molecule-ai/workspace-template-claude-code:latest" not found after pull attempt
+```
+
+Production tenants are unaffected (their `MOLECULE_IMAGE_REGISTRY` points at ECR, which we still control), but OSS onboarding is broken. Workspace template repos are intentionally separate from `molecule-core` (each runtime is OSS-shape and forkable), and they are mirrored to Gitea (`https://git.moleculesai.app/molecule-ai/molecule-ai-workspace-template-<runtime>`) — but the provisioner has no path that consumes Gitea source directly.
+
+## Decision
+
+When `MOLECULE_IMAGE_REGISTRY` is **unset** (or empty), the provisioner switches to a **local-build mode** that:
+
+1. Looks up the workspace-template repo's HEAD sha on Gitea via a single API call.
+2. Checks whether a SHA-pinned local image (`molecule-local/workspace-template-<runtime>:<sha12>`) already exists; if so, reuses it.
+3. Otherwise shallow-clones the repo into `~/.cache/molecule/workspace-template-build/<runtime>/<sha12>/` and runs `docker build --platform=linux/amd64 -t <tag> .`.
+4. Hands the SHA-pinned tag to Docker for ContainerCreate, bypassing the registry-pull path entirely.
+
+When `MOLECULE_IMAGE_REGISTRY` is **set**, behavior is unchanged: pull the image from that registry. Existing prod tenants and self-hosters who mirror to a private registry are not affected.
+
+## Consequences
+
+### Positive
+
+* **Zero-config OSS onboarding** — `git clone molecule-core && go run ./workspace-server/cmd/server` boots end-to-end without any registry credentials.
+* **Production tenants protected** — same env var, same semantics in SaaS-mode. Migration is a no-op.
+* **No new env var** — extending an existing var's semantics ("where to pull, OR build locally if absent") rather than introducing `MOLECULE_LOCAL_BUILD=1` keeps the surface small.
+* **SHA-pinned cache** — repeat builds are O(API-call); only template-repo HEAD changes invalidate.
+* **Production-parity image** — amd64 emulation on Apple Silicon honours `feedback_local_must_mimic_production`. The provisioner's existing `defaultImagePlatform()` already forces amd64 for parity; building amd64 locally lets that decision stay consistent.
+
+### Negative
+
+* **Conflates two concerns** — `MOLECULE_IMAGE_REGISTRY` now signals BOTH "where to pull" AND "build locally if absent." A future operator who unsets it expecting a hard error will instead get a slow first-provision. Documented in the runbook.
+* **First-provision is slow on Apple Silicon** — 5–10 min via QEMU emulation on the cold path. Mitigated by SHA-cache (subsequent runs are <1s lookup + 0s build).
+* **Coverage gap** — only 4 of 9 runtimes are mirrored to Gitea today (`claude-code`, `hermes`, `langgraph`, `autogen`). The other 5 fail with an actionable "not mirrored" error. Mirroring those repos is a separate task.
+* **Implicit trust boundary** — operator running `go run` implicitly trusts `molecule-ai/molecule-ai-workspace-template-*` repos on Gitea. This is the same trust they would extend to the GHCR images today; not a new attack surface.
+
+## Alternatives considered
+
+1. **New env var `MOLECULE_LOCAL_BUILD=1`** — explicit, but requires OSS contributors to know it exists. Violates the zero-config goal.
+2. **Push pre-built images to a Gitea container registry, mirror tag from upstream** — operationally cleaner but: (a) Gitea's container-registry add-on isn't deployed on the operator host, (b) defeats the OSS-contributor goal of "hack on the source, see your changes," since they'd still pull a stale image.
+3. **Embed Dockerfiles in molecule-core itself, drop the standalone template repos** — would work but breaks the OSS-shape principle; templates are intentionally separable, anyone-can-fork artifacts.
+4. **Build native arch on Apple Silicon (arm64) and drop the platform pin in local-mode** — fast, but creates `linux/arm64` images that diverge from the amd64-only prod runtime. Local-vs-prod debug behavior would diverge. Rejected per `feedback_local_must_mimic_production`.
+
+## Security review
+
+* **Gitea repo URL allowlist** — runtime name must be in the `knownRuntimes` allowlist (defence-in-depth against a future code path that lets cfg.Runtime carry untrusted input). Repo prefix is hardcoded to `https://git.moleculesai.app/molecule-ai/molecule-ai-workspace-template-`; forks can override via `MOLECULE_LOCAL_TEMPLATE_REPO_PREFIX` (opt-in, default off).
+* **Token handling** — clones are anonymous over HTTPS by default (templates are public). `MOLECULE_GITEA_TOKEN`, if set, is passed via URL userinfo for the clone and as `Authorization: token` for the API call. The token is **masked in every log line** via `maskTokenInURL` / `maskTokenInString` and never appears in the cache dir path.
+* **No silent fallback** — if Gitea is unreachable or the runtime isn't mirrored, we return a clear error mentioning the repo URL and the missing runtime. We **never** fall back to GHCR/ECR (that would be a confusing bug for an OSS contributor who happened to have stale ECR creds in their docker config).
+* **Build-arg injection** — `docker build` is invoked with NO `--build-arg` from external input. Dockerfile is consumed as-is.
+* **Cache poisoning** — cache key is the Gitea HEAD sha + Dockerfile content; a force-push to the template repo's main branch regenerates the key on next run. Cache dir is per-user (`$HOME/.cache`), so cross-user attacks aren't relevant in single-user dev mode.
+
+## Versioning + back-compat
+
+* Existing prod tenants set `MOLECULE_IMAGE_REGISTRY=<ECR url>` → unchanged behavior.
+* Existing local installs that set the var → unchanged behavior.
+* Existing local installs that don't set it → switch to local-build path. Migration: none required (additive); first provision will take 5–10 min instead of failing.
+* No deprecations.
+
+## References
+
+* Issue #63 — feat(workspace-server): local-dev provisioner builds from Gitea source
+* Saved memory `feedback_local_must_mimic_production` — local docker must mimic prod, no bypasses
+* Saved memory `reference_post_suspension_pipeline` — full post-2026-05-06 stack shape
+* Saved memory `feedback_github_botring_fingerprint` — what got the org suspended
@@ -27,11 +27,11 @@ prompt_files:
 # AGENTS.md-style example:
 #   prompt_files: [AGENTS.md]

-# Files to share with direct children (1-level inheritance)
-# Children fetch these at startup via GET /workspaces/:id/shared-context
-shared_context:
-  - architecture.md
-  - conventions.md
+# NOTE: `shared_context` (parent → child file injection at boot) was removed.
+# To share knowledge across a team, use memory v2's team:<id> namespace via
+# the recall_memory MCP tool — the agent pulls it on demand instead of
+# paying for it at every boot. For large blob-shaped artefacts, see RFC
+# #2789 (platform-owned shared file storage).

 # Skills to load -- folder names under skills/
 skills:
@@ -123,7 +123,6 @@ env:
 | `runtime` | No | Adapter to use: `langgraph` (default), `claude-code`, `crewai`, `autogen`, `deepagents`, `openclaw`. See [Agent Runtime Adapters](./cli-runtime.md). |
 | `model` | Yes | LangChain-compatible provider string (e.g. `anthropic:claude-sonnet-4-6`). Overridden by `MODEL_PROVIDER` env var if set. |
 | `prompt_files` | No | Ordered list of markdown files to load as system prompt. Defaults to `["system-prompt.md"]` if omitted. `MEMORY.md` and `USER.md` are auto-appended when present so frozen memory snapshots do not need to be duplicated here. Supports any agent framework's file structure (OpenClaw, Claude Code, etc.) |
-| `shared_context` | No | Files from this workspace's config dir to share with direct children. Children fetch these at startup and inject into their system prompt as `## Parent Context`. 1-level inheritance only (grandchildren don't see grandparent's context). |
 | `skills` | Yes | List of skill folder names to load from `skills/` |
 | `tools` | No | Built-in tools from workspace-template |
 | `memory` | No | Memory backend config (defaults to filesystem) |
@@ -157,7 +156,6 @@ The file watcher monitors the entire config directory. When `config.yaml` change
 | `name`, `description`, `version` | Yes | Rebuild Agent Card with new metadata |
 | `a2a` | **No** | Port and protocol changes require container restart |
 | `delegation` | Yes | Retry/timeout defaults take effect on next delegation call |
-| `shared_context` | Yes | Children fetch on next prompt rebuild; no restart needed |
 | `sub_workspaces` | **No** | Team structure changes go through `POST /workspaces/:id/expand` |

 See [Skills — Live Reload](./skills.md#live-reload) for the full file watcher flow.
@@ -24,21 +24,19 @@ When you receive a task, break it into sub-tasks and delegate to your team.
 Always review work before reporting completion to the caller.
 ```

-### 2. Parent Context (if child workspace)
+### 2. Team-shared knowledge (on demand)

-If this workspace was created via team expansion (has a `PARENT_ID` env var), it fetches its parent's shared context files at startup via `GET /workspaces/{parent_id}/shared-context`. The parent declares which files to share in its `config.yaml`:
+Team-scoped knowledge is no longer injected at boot. The previous
+`shared_context` field + `GET /workspaces/{parent_id}/shared-context`
+fetch was removed; agents now pull team-shared knowledge on demand via
+memory v2's `team:<id>` namespace using the `recall_memory` MCP tool.

-```yaml
-shared_context:
-  - architecture.md
-  - conventions.md
-```
-
-These files are injected as a `## Parent Context` section, with each file rendered under a `### {filename}` heading. This gives children the parent's project knowledge (architecture, conventions, API schemas) without exposing the parent's system prompt or full config.
-
-**1-level inheritance only:** A grandchild sees its direct parent's shared context, not its grandparent's. This mirrors the L2 Team Memory scope.
-
-**Graceful degradation:** If the parent is offline or the endpoint returns an error, the child starts normally without parent context.
+This shifts cost from "every boot, always" to "only when the agent
+asks", and lets team members write to the shared store from anywhere
+that can resolve the namespace (canvas Memory tab, agent
+`commit_memory`, admin import). For large blob-shaped artefacts (full
+architecture docs, brand assets, PDFs) see RFC #2789 (platform-owned
+shared file storage).

 ### 3. Skill Instructions

--- a/Show More
+++ b/Show More