Merge pull request 'fix(org-import): reconcile mode + audit-event emission' (#137) from fix/org-import-reconcile-and-audit into main

Closes the additive-import zombie bug — re-running /org/import with a
tree shape that reparents same-named roles left the prior workspace
online because lookupExistingChild's dedupe is parent-scoped (different
parent_id → "different" workspace). Caught 2026-05-08 after a dev-tree
re-import left 8 orphans co-existing with the new tree on canvas until
manual cascade-delete.

Three layers in this PR:

- mode="reconcile" on /org/import — after the import loop, online
  workspaces whose name matches an imported name but whose id isn't in
  the result set are cascade-deleted. Default mode "" / "merge"
  preserves existing additive behavior. Empty-set guards prevent
  accidental "delete everything" if either array comes up empty.

- WorkspaceHandler.CascadeDelete extracted as a callable helper from
  the existing Delete HTTP handler so OrgImport's reconcile path shares
  the same teardown sequence (#73 race guard, container stop, volume
  removal, token revocation, schedule disable, event broadcast). The
  HTTP Delete handler still inlines the same logic; deduplication
  tracked as tech-debt follow-up.

- emitOrgEvent(structure_events) records org.import.started +
  org.import.completed with mode, created/skipped/reconcile_removed
  counts, duration_ms, error. Replaces the lost-on-restart stdout-only
  log shape for an audit-trail surface that's queryable by SQL. Closes
  the "what happened at 20:13?" debugging gap that motivated this fix.

Verified live against the local platform: cascade-delete on an old
tree's removed root cleared 8 surviving orphans; mode="reconcile" with
a freshly-INSERTed fake orphan removed exactly the fake; idempotent
re-run of reconcile is a no-op (0 removed, no errors); structure_events
captures every started+completed pair with full payload.

7 new unit tests (walkOrgWorkspaceNames flat/nested/spawning:false/
empty-name; emitOrgEvent success + DB-error-swallow; errString). Full
handler suite green.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

.github/scripts/lint_secret_pattern_drift.py

@@ -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",
     ),
 ]
 

.github/workflows/auto-promote-on-e2e.yml

@@ -1,467 +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
-          # for the same SHA via Gitea's commit-status API.
-          #
-          # GitHub-era this was `gh run list --workflow=X --commit=SHA
-          # --json status,conclusion` returning either `[]` (no run on
-          # this SHA) or `[{status, conclusion}]` (the run's state).
-          # Gitea has NO workflow-runs API at all — `/api/v1/repos/.../
-          # actions/runs` returns 404 (verified 2026-05-07, issue #75).
-          # However Gitea Actions DOES emit a commit status per workflow
-          # job, with `context = "<Workflow Name> / <Job Name> (<event>)"`,
-          # which is exactly what we need: each E2E run leg becomes one
-          # status row on the SHA, and the aggregate state encodes the
-          # run's outcome.
-          #
-          # Mapping:
-          #   0 matched contexts          → "none/none"      (E2E paths-
-          #                                                    filtered
-          #                                                    out — same
-          #                                                    semantic
-          #                                                    as before)
-          #   any context = pending       → "in_progress/none" (defer)
-          #   any context = error|failure → "completed/failure" (abort)
-          #   all contexts = success      → "completed/success" (proceed)
-          #
-          # The "completed/cancelled" and "completed/timed_out" buckets
-          # don't have direct Gitea analogs (Gitea statuses are
-          # success / failure / error / pending / warning). Per-SHA
-          # concurrency cancellation surfaces as `error` on Gitea, which
-          # we map to "completed/failure" rather than "completed/cancelled"
-          # — losing the soft-defer semantic of the cancelled bucket on
-          # this fleet. Tradeoff: the staleness alarm (auto-promote-stale-
-          # alarm.yml) still catches a stuck :latest within 4h, and a
-          # legitimate cancel is rare enough that aborting + manual
-          # re-dispatch is acceptable. If we measure cancel frequency
-          # > 1/week, revisit by reading the run-step-summary text via
-          # a follow-up script.
-          #
-          # Network or auth blips collapse to "none/none" via the curl
-          # `|| true` fallback, matching the pre-Gitea behaviour where
-          # an empty list also degenerated to none/none.
-          GITEA_API_URL="${GITHUB_SERVER_URL:-https://git.moleculesai.app}/api/v1"
-          STATUSES_JSON=$(curl --fail-with-body -sS \
-            -H "Authorization: token ${GH_TOKEN}" \
-            -H "Accept: application/json" \
-            "${GITEA_API_URL}/repos/${REPO}/commits/${SHA}/statuses?limit=100" \
-            2>/dev/null || echo "[]")
-          RESULT=$(printf '%s' "$STATUSES_JSON" | jq -r '
-            # Filter to E2E Staging SaaS (full lifecycle) statuses.
-            # Match by leading workflow-name prefix so the "<job>
-            # (<event>)" tail is irrelevant. Gitea emits the workflow
-            # name verbatim from the YAML `name:` field.
-            [.[] | select(.context | startswith("E2E Staging SaaS (full lifecycle) /"))] as $rows
-            | if ($rows | length) == 0 then
-                "none/none"
-              elif any($rows[]; .status == "pending") then
-                "in_progress/none"
-              elif any($rows[]; .status == "failure" or .status == "error") then
-                "completed/failure"
-              elif all($rows[]; .status == "success") then
-                "completed/success"
-              else
-                # Mixed / unknown — fall through to *) bucket below.
-                "completed/" + ($rows[0].status // "unknown")
-              end
-          ' 2>/dev/null || echo "none/none")
-
-          echo "E2E Staging SaaS for ${SHA:0:7}: $RESULT"
-
-          case "$RESULT" in
-            completed/success)
-              echo "proceed=true" >> "$GITHUB_OUTPUT"
-              echo "::notice::E2E green for this SHA — proceeding with promote"
-              ;;
-            completed/failure|completed/timed_out)
-              echo "proceed=false" >> "$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."
-              } >> "$GITHUB_STEP_SUMMARY"
-              exit 1
-              ;;
-            completed/cancelled)
-              # GitHub-era only: cancelled ≠ failure. Gitea statuses
-              # don't expose a "cancelled" state — a per-SHA concurrency
-              # cancellation surfaces as `failure` or `error` on Gitea
-              # and is now handled by the failure branch above. This
-              # arm is kept for backwards compatibility / dual-host
-              # operation (if we ever add a non-Gitea fallback) but
-              # under the post-#75 flow it's unreachable.
-              echo "proceed=false" >> "$GITHUB_OUTPUT"
-              {
-                echo "## ⏭ Auto-promote deferred — E2E Staging SaaS was cancelled"
-                echo
-                echo "E2E Staging SaaS for \`${SHA:0:7}\`: \`$RESULT\`"
-                echo "Likely per-SHA concurrency (newer push superseded this E2E run)."
-                echo "The newer SHA's E2E will fire its own promote when it lands."
-                echo "If you need this specific SHA promoted, manually dispatch."
-              } >> "$GITHUB_STEP_SUMMARY"
-              ;;
-            in_progress/*|queued/*|requested/*|waiting/*|pending/*)
-              echo "proceed=false" >> "$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"

.github/workflows/auto-promote-staging.yml

@@ -1,492 +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 schedules Gitea auto-merge so the PR lands
-# automatically once approval + status checks are satisfied.
-#
-# ============================================================
-# What this workflow does
-# ============================================================
-#
-# 1. On a workflow_run completion event for one of the staging gate
-#    workflows (CI, E2E Staging Canvas, E2E API Smoke, CodeQL),
-#    checks if the combined status on the staging head SHA is green.
-# 2. If green, opens (or re-uses) a PR `head: staging → base: main`
-#    via Gitea REST `POST /api/v1/repos/.../pulls`.
-# 3. Schedules auto-merge via `POST /api/v1/repos/.../pulls/{index}/merge`
-#    with `merge_when_checks_succeed: true`. Gitea waits for the
-#    approval requirement on `main` (`required_approvals: 1`) and
-#    the status-check gates, then merges.
-# 4. The merge commit lands on `main` and fires
-#    `publish-workspace-server-image.yml` naturally via its
-#    `on: push: branches: [main]` trigger — no explicit dispatch
-#    needed (see "Why no workflow_dispatch tail" below).
-#
-# `auto-sync-main-to-staging.yml` is the reverse-direction
-# counterpart (main → staging, fast-forward push). Together they
-# keep the staging-superset-of-main invariant tight.
-#
-# ============================================================
-# Why Gitea REST (and not `gh pr create`)
-# ============================================================
-#
-# Pre-2026-05-06 this workflow used `gh pr create`, `gh pr merge --auto`,
-# `gh run list`, and `gh workflow run` against GitHub. After the
-# GitHub→Gitea cutover those calls fail because:
-#
-#   - `gh pr create / merge / view / list` route to GitHub GraphQL
-#     (`/api/graphql`). Gitea does not expose a GraphQL endpoint;
-#     every call returns `HTTP 405 Method Not Allowed` — same root
-#     cause as #65 (auto-sync) which PR #66 fixed by dropping `gh`
-#     entirely.
-#   - `gh run list --workflow=...` GitHub-shape; Gitea has the
-#     simpler `GET /repos/.../commits/{ref}/status` combined-status
-#     endpoint instead.
-#   - `gh workflow run X.yml` calls `POST /repos/.../actions/workflows/{id}/dispatches`,
-#     which does NOT exist on Gitea 1.22.6 (verified via swagger.v1.json).
-#
-# So this workflow uses direct `curl` calls to Gitea REST. No `gh`
-# CLI dependency, no GraphQL, no missing-endpoint footgun.
-#
-# ============================================================
-# Why no workflow_dispatch tail (was load-bearing on GitHub, dead on Gitea)
-# ============================================================
-#
-# The GitHub-era version had a 60-line polling step that waited for
-# the promote PR to merge, then explicitly dispatched
-# `publish-workspace-server-image.yml` on `--ref main`. That step
-# existed because GitHub's GITHUB_TOKEN-initiated merges suppress
-# downstream `on: push` workflows (the documented "no recursion" rule
-# — https://docs.github.com/en/actions/using-workflows/triggering-a-workflow#triggering-a-workflow-from-a-workflow).
-# The explicit dispatch was the workaround.
-#
-# Gitea Actions does NOT have this no-recursion rule. PR #66's auto-
-# sync merge to main fired `auto-promote-staging` on the next push
-# trigger naturally. So the cascade fires on the natural push event;
-# the explicit dispatch is dead code. (And even if we wanted to
-# preserve it, Gitea has no `workflow_dispatch` REST endpoint.)
-#
-# Removed in this rewrite. If we ever observe the cascade misfire,
-# operator can push an empty commit to `main` to wake it.
-#
-# ============================================================
-# Why open a PR (and not direct push)
-# ============================================================
-#
-# `main` branch protection has `enable_push: false` with NO
-# `push_whitelist_usernames`. Direct push is impossible for any
-# persona, including admins. PR-mediated merge is the only path,
-# which is intentional: prod state mutations (and staging→main IS a
-# prod mutation, since the next deploy fans out to tenants) require
-# Hongming's approval per `feedback_prod_apply_needs_hongming_chat_go`.
-#
-# The auto-merge schedule preserves this gate: `merge_when_checks_succeed`
-# does NOT bypass `required_approvals: 1`. Gitea waits for BOTH
-# approval AND green checks before merging. Hongming reviews via the
-# canvas/chat-handle of the PR notification, approves, and Gitea
-# auto-merges within seconds.
-#
-# ============================================================
-# Identity + token (anti-bot-ring per saved-memory
-# `feedback_per_agent_gitea_identity_default`)
-# ============================================================
-#
-# This workflow uses `secrets.AUTO_SYNC_TOKEN` — a personal access
-# token issued to the `devops-engineer` Gitea persona. NOT the
-# founder PAT. The bot-ring fingerprint that triggered the GitHub
-# org suspension on 2026-05-06 was characterised by founder PAT
-# acting as CI at machine speed.
-#
-# Token scope: `push: true` (read+write) on this repo. The persona
-# can: open PRs, comment on PRs, schedule auto-merge. The persona
-# CANNOT bypass main's branch protection (`required_approvals: 1`
-# still applies — only Hongming's review unblocks merge).
-#
-# Authorship: the PR is opened by `devops-engineer`; the merge
-# commit credits Hongming-as-approver and `devops-engineer` as
-# the merger.
-#
-# ============================================================
-# Failure modes & operational notes
-# ============================================================
-#
-# A — staging gates not all green at trigger time:
-#     - The combined-status check returns `state: pending|failure`.
-#       Workflow exits 0 with a step-summary "not all green; staying
-#       on current main". Re-fires on the next gate completion.
-#
-# B — Gitea PR-create returns non-201 (e.g. 422 already-exists):
-#     - Idempotent: the workflow first GETs the existing open
-#       staging→main PR. If found, reuse it; if not, POST a new one.
-#       422 should never surface; if it does (race), step summary
-#       captures the body and the next workflow_run picks up.
-#
-# C — `merge_when_checks_succeed` schedule fails:
-#     - 422 with "Pull request is not mergeable" if there are
-#       conflicts or stale base. Step summary surfaces it; operator
-#       (or `auto-sync-main-to-staging`) needs to bring staging up
-#       to date with main first. Workflow exits 1 to surface red.
-#
-# D — `AUTO_SYNC_TOKEN` rotated / wrong scope:
-#     - 401/403 on first REST call. Step summary surfaces it.
-#       Re-issue the token from `~/.molecule-ai/personas/` on the
-#       operator host and update the repo Actions secret.
-#
-# ============================================================
-# Loop safety
-# ============================================================
-#
-# When the promote PR merges to main, `auto-sync-main-to-staging.yml`
-# fires (on:push:main) and pushes the merge commit back to staging.
-# That push to staging is by `devops-engineer`, NOT this workflow's
-# token, and triggers the staging gate workflows. When they all
-# complete, we end up back here — but the tree-diff guard catches
-# it: staging tree == main tree (the merge commit changes nothing),
-# so we skip and the cycle terminates.
-
-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: read
-  pull-requests: 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. Would race the GET-or-POST PR step.
-#   2. Would both call merge-schedule (idempotent — fine on Gitea).
-# cancel-in-progress: false because the second run on a fresh staging
-# tip should NOT kill the first which has already opened the PR.
-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 pre-cutover on GitHub). On Gitea the cycle shape
-      # is different (auto-sync uses fast-forward, no merge commit),
-      # but the tree-diff guard is cheap insurance and protects against
-      # any future merge-style regression.
-      - 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 (cycle-break safety)
-        id: tree-diff
-        env:
-          HEAD_SHA: ${{ github.event.workflow_run.head_sha || github.sha }}
-        run: |
-          set -eu
-          git fetch origin main --depth=50 || { echo "::warning::git fetch main failed — proceeding (fail-open)"; exit 0; }
-          if git diff --quiet origin/main "$HEAD_SHA" -- 2>/dev/null; then
-            {
-              echo "## Skipped — no code to promote"
-              echo
-              echo "staging tip (\`${HEAD_SHA:0:8}\`) and \`main\` have identical trees."
-              echo "Skipping to avoid opening an empty promote PR."
-            } >> "$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 combined status on staging head
-        if: steps.tree-diff.outputs.skip != 'true'
-        id: gates
-        env:
-          GITEA_TOKEN: ${{ secrets.AUTO_SYNC_TOKEN }}
-          HEAD_SHA: ${{ github.event.workflow_run.head_sha || github.sha }}
-          REPO: ${{ github.repository }}
-          GITEA_HOST: ${{ vars.GITEA_HOST || 'https://git.moleculesai.app' }}
-        run: |
-          set -euo pipefail
-
-          # Gitea-native combined-status endpoint aggregates every
-          # check context attached to a SHA. This is structurally
-          # cleaner than the GitHub-era per-workflow `gh run list`
-          # loop because:
-          #
-          #   1. There's no risk of "workflow name collision" (the
-          #      GitHub-era code had to switch from `--workflow=NAME`
-          #      to `--workflow=FILE.YML` to disambiguate "CodeQL"
-          #      between the explicit workflow and GitHub's UI-
-          #      configured default setup; Gitea has no such
-          #      duplicate-name surface).
-          #   2. Gitea's combined state already encodes the AND
-          #      across all contexts: success only if EVERY context
-          #      is success. Pending or failure on any context
-          #      produces non-success state.
-          #
-          # See https://docs.gitea.com/api/1.22 for the schema —
-          # `state` is one of: success, pending, failure, error.
-
-          echo "head_sha=${HEAD_SHA}" >> "$GITHUB_OUTPUT"
-          echo "Checking combined status on SHA ${HEAD_SHA}"
-
-          # `set +o pipefail` for the http-code capture pattern; restore
-          # immediately. Pattern hardened per `feedback_curl_status_capture_pollution`.
-          BODY_FILE=$(mktemp)
-          set +e
-          STATUS=$(curl -sS \
-            -H "Authorization: token ${GITEA_TOKEN}" \
-            -H "Accept: application/json" \
-            -o "${BODY_FILE}" \
-            -w "%{http_code}" \
-            "${GITEA_HOST}/api/v1/repos/${REPO}/commits/${HEAD_SHA}/status")
-          CURL_RC=$?
-          set -e
-
-          if [ "${CURL_RC}" -ne 0 ] || [ "${STATUS}" != "200" ]; then
-            echo "::error::combined-status fetch failed: curl=${CURL_RC} http=${STATUS}"
-            cat "${BODY_FILE}" | head -c 500 || true
-            rm -f "${BODY_FILE}"
-            echo "all_green=false" >> "$GITHUB_OUTPUT"
-            exit 0
-          fi
-
-          STATE=$(jq -r '.state // "missing"' < "${BODY_FILE}")
-          TOTAL=$(jq -r '.total_count // 0' < "${BODY_FILE}")
-          rm -f "${BODY_FILE}"
-
-          echo "Combined status: state=${STATE} total_count=${TOTAL}"
-
-          if [ "${STATE}" = "success" ] && [ "${TOTAL}" -gt 0 ]; then
-            echo "all_green=true" >> "$GITHUB_OUTPUT"
-            echo "::notice::All gates green on ${HEAD_SHA} (${TOTAL} contexts)"
-          else
-            echo "all_green=false" >> "$GITHUB_OUTPUT"
-            {
-              echo "## Not promoting — combined status not green"
-              echo
-              echo "- SHA: \`${HEAD_SHA:0:8}\`"
-              echo "- Combined state: \`${STATE}\`"
-              echo "- Context count: ${TOTAL}"
-              echo
-              echo "Will re-fire on the next gate completion. Investigate any red gate via the Actions UI."
-            } >> "$GITHUB_STEP_SUMMARY"
-            echo "::notice::auto-promote: combined status is ${STATE} 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: ${{ github.event.inputs.force }}
-        run: |
-          set -eu
-          # Repo variable AUTO_PROMOTE_ENABLED=true flips this on. While
-          # it's unset, the workflow dry-runs (logs what it would have
-          # done) but doesn't open the promote PR. Set the variable in
-          # Settings → Actions → Variables.
-          if [ "${AUTO_PROMOTE_ENABLED:-}" != "true" ] && [ "${FORCE_INPUT:-false}" != "true" ]; then
-            {
-              echo "## Auto-promote disabled"
-              echo
-              echo "Repo variable \`AUTO_PROMOTE_ENABLED\` is not set to \`true\`."
-              echo "All gates are green on staging; would have opened a promote PR to \`main\`."
-              echo
-              echo "To enable: Settings → Actions → Variables → \`AUTO_PROMOTE_ENABLED=true\`."
-              echo "To test once manually: workflow_dispatch with \`force=true\`."
-            } >> "$GITHUB_STEP_SUMMARY"
-            echo "::notice::auto-promote disabled — dry run only"
-            exit 0
-          fi
-
-      - name: Open or reuse promote PR + schedule auto-merge
-        if: ${{ vars.AUTO_PROMOTE_ENABLED == 'true' || github.event.inputs.force == 'true' }}
-        env:
-          GITEA_TOKEN: ${{ secrets.AUTO_SYNC_TOKEN }}
-          REPO: ${{ github.repository }}
-          TARGET_SHA: ${{ needs.check-all-gates-green.outputs.head_sha }}
-          GITEA_HOST: ${{ vars.GITEA_HOST || 'https://git.moleculesai.app' }}
-        run: |
-          set -euo pipefail
-
-          API="${GITEA_HOST}/api/v1/repos/${REPO}"
-          AUTH=(-H "Authorization: token ${GITEA_TOKEN}" -H "Accept: application/json")
-
-          # http_status_get RESULT_VAR URL
-          # Sets RESULT_VAR to "<http_code>:<body_file>". Curl status
-          # capture pattern per `feedback_curl_status_capture_pollution`:
-          # http_code goes to its own tempfile-equivalent (-w), body to
-          # another tempfile, set +e/-e bracket protects pipeline state.
-          http_get() {
-            local body_file="$1"; shift
-            local url="$1"; shift
-            set +e
-            local code
-            code=$(curl -sS "${AUTH[@]}" -o "${body_file}" -w "%{http_code}" "${url}")
-            local rc=$?
-            set -e
-            if [ "${rc}" -ne 0 ]; then
-              echo "::error::curl GET failed (rc=${rc}) on ${url}"
-              return 99
-            fi
-            echo "${code}"
-          }
-          http_post_json() {
-            local body_file="$1"; shift
-            local data="$1"; shift
-            local url="$1"; shift
-            set +e
-            local code
-            code=$(curl -sS "${AUTH[@]}" -H "Content-Type: application/json" \
-              -X POST -d "${data}" -o "${body_file}" -w "%{http_code}" "${url}")
-            local rc=$?
-            set -e
-            if [ "${rc}" -ne 0 ]; then
-              echo "::error::curl POST failed (rc=${rc}) on ${url}"
-              return 99
-            fi
-            echo "${code}"
-          }
-
-          # Step 1: look for an existing open staging→main promote PR
-          # (idempotent on workflow re-run). Gitea doesn't have a
-          # head/base filter on the list endpoint that's as ergonomic
-          # as gh's, but the dedicated `/pulls/{base}/{head}` lookup
-          # works.
-          BODY=$(mktemp)
-          STATUS=$(http_get "${BODY}" "${API}/pulls/main/staging") || true
-
-          PR_NUM=""
-          if [ "${STATUS}" = "200" ]; then
-            STATE=$(jq -r '.state // "missing"' < "${BODY}")
-            if [ "${STATE}" = "open" ]; then
-              PR_NUM=$(jq -r '.number // ""' < "${BODY}")
-              echo "::notice::Re-using existing open promote PR #${PR_NUM}"
-            fi
-          fi
-          rm -f "${BODY}"
-
-          # Step 2: if no open PR, create one.
-          if [ -z "${PR_NUM}" ]; then
-            TITLE="staging → main: auto-promote ${TARGET_SHA:0:7}"
-            BODY_TEXT=$(cat <<EOFBODY
-          Automated promotion of \`staging\` (\`${TARGET_SHA:0:8}\`) to \`main\`. All required staging gates are green at this SHA (combined status reported success).
-
-          This PR is auto-generated by \`.github/workflows/auto-promote-staging.yml\` whenever every required gate completes green on the same staging SHA.
-
-          **Approval gate:** \`main\` branch protection requires 1 approval before this can land. Once approved, Gitea will auto-merge (the workflow scheduled \`merge_when_checks_succeed: true\` immediately after open).
-
-          The reverse-direction sync (the merge commit on \`main\` → \`staging\`) is handled automatically by \`auto-sync-main-to-staging.yml\` after this PR lands.
-
-          ---
-          - Source: staging at \`${TARGET_SHA}\`
-          - Opened by: \`devops-engineer\` persona (anti-bot-ring; never founder PAT)
-          - Refs: #65, #73, #195
-          EOFBODY
-          )
-            REQ=$(jq -n \
-              --arg title "${TITLE}" \
-              --arg body "${BODY_TEXT}" \
-              --arg base "main" \
-              --arg head "staging" \
-              '{title:$title, body:$body, base:$base, head:$head}')
-
-            BODY=$(mktemp)
-            STATUS=$(http_post_json "${BODY}" "${REQ}" "${API}/pulls")
-
-            if [ "${STATUS}" = "201" ]; then
-              PR_NUM=$(jq -r '.number // ""' < "${BODY}")
-              echo "::notice::Opened promote PR #${PR_NUM}"
-            else
-              echo "::error::Failed to create promote PR: HTTP ${STATUS}"
-              jq -r '.message // .' < "${BODY}" | head -c 500
-              rm -f "${BODY}"
-              exit 1
-            fi
-            rm -f "${BODY}"
-          fi
-
-          # Step 3: schedule auto-merge. merge_when_checks_succeed
-          # tells Gitea to wait for both:
-          #   - all required status checks to pass
-          #   - the required-approvals gate (1 approval on main)
-          # before merging. On approval+green, Gitea merges within
-          # seconds. On any check failing or approval being denied,
-          # the schedule stays armed but doesn't fire.
-          #
-          # Idempotent: re-arming on an already-armed PR is a no-op.
-          REQ=$(jq -n '{Do:"merge", merge_when_checks_succeed:true}')
-          BODY=$(mktemp)
-          STATUS=$(http_post_json "${BODY}" "${REQ}" "${API}/pulls/${PR_NUM}/merge")
-
-          # Gitea returns:
-          #   - 200/204 on successful immediate merge (gates already green AND approved)
-          #   - 405 "Please try again later" when scheduled successfully but waiting
-          #   - 422 on "Pull request is not mergeable" (conflict, stale base, etc.)
-          #
-          # 405 here is benign — Gitea's way of saying "scheduled, not merging now".
-          # We treat 200/204/405 as success, anything else as failure.
-          case "${STATUS}" in
-            200|204)
-              MERGE_OUTCOME="merged-immediately"
-              echo "::notice::Promote PR #${PR_NUM} merged immediately (gates+approval already green)"
-              ;;
-            405)
-              MERGE_OUTCOME="auto-merge-scheduled"
-              echo "::notice::Promote PR #${PR_NUM}: auto-merge scheduled (Gitea will land on approval+green)"
-              ;;
-            422)
-              MERGE_OUTCOME="not-mergeable"
-              echo "::warning::Promote PR #${PR_NUM}: not mergeable (conflict, stale base, or already merging)."
-              jq -r '.message // .' < "${BODY}" | head -c 500
-              ;;
-            *)
-              echo "::error::Unexpected status ${STATUS} on merge schedule"
-              jq -r '.message // .' < "${BODY}" | head -c 500
-              rm -f "${BODY}"
-              exit 1
-              ;;
-          esac
-          rm -f "${BODY}"
-
-          {
-            echo "## Auto-promote PR opened"
-            echo
-            echo "- Source: staging at \`${TARGET_SHA:0:8}\`"
-            echo "- PR: #${PR_NUM}"
-            echo "- Outcome: \`${MERGE_OUTCOME}\`"
-            echo
-            if [ "${MERGE_OUTCOME}" = "auto-merge-scheduled" ]; then
-              echo "Gitea will auto-merge once Hongming approves and all checks are green. No human action needed beyond approval."
-            elif [ "${MERGE_OUTCOME}" = "merged-immediately" ]; then
-              echo "Merged immediately. \`publish-workspace-server-image.yml\` will fire naturally on the resulting \`main\` push."
-            else
-              echo "PR is not auto-merging. Operator may need to bring staging up to date with main, then re-trigger this workflow via workflow_dispatch."
-            fi
-          } >> "$GITHUB_STEP_SUMMARY"

.github/workflows/auto-promote-stale-alarm.yml

@@ -1,83 +0,0 @@
-name: auto-promote-stale-alarm
-
-# Hourly cron + on-demand alarm for the silent-block failure mode that
-# motivated issue #2975:
-#   - The auto-promote-staging.yml workflow opened a PR + armed
-#     auto-merge, but main's branch protection requires a human review
-#     (reviewDecision=REVIEW_REQUIRED). The PR sat BLOCKED with no
-#     surface-up-the-stack for 12+ hours, holding 25 commits hostage
-#     including the Memory v2 redesign and a reno-stars data-loss fix.
-#
-# This workflow runs `scripts/check-stale-promote-pr.sh` against the
-# repo's open auto-promote PRs (base=main head=staging). When a PR has
-# been BLOCKED on REVIEW_REQUIRED for >4h, it:
-#   1. Emits a workflow-level warning (visible in run summary + the
-#      Actions UI feed).
-#   2. Posts a comment on the PR (idempotent — one alarm per PR).
-#
-# The detection logic lives in scripts/check-stale-promote-pr.sh so
-# it's unit-testable with stubbed `gh` (see test-check-stale-promote-pr.sh).
-# This file is the schedule + invocation surface only — SSOT for the
-# detector itself.
-
-on:
-  schedule:
-    # Hourly. Cheap (one `gh pr list` + jq), and 1h granularity is
-    # plenty for a 4h staleness threshold — operators see the alarm
-    # within at most 1h of crossing the threshold.
-    - cron: "27 * * * *"  # at :27 to dodge the cron herd at :00
-  workflow_dispatch:
-    inputs:
-      stale_hours:
-        description: "Hours after which a BLOCKED+REVIEW_REQUIRED PR is stale (default 4)"
-        required: false
-        default: "4"
-      post_comment:
-        description: "Post a comment on stale PRs (default true)"
-        required: false
-        default: "true"
-
-permissions:
-  contents: read
-  pull-requests: write  # post comments on stale PRs
-
-# Serialize so the on-demand and scheduled runs don't double-comment
-# the same PR. cancel-in-progress=false because the script is idempotent
-# (existing comment marker prevents dupes), but a scheduled run firing
-# while a manual one runs would just re-list the same PR set.
-concurrency:
-  group: auto-promote-stale-alarm
-  cancel-in-progress: false
-
-jobs:
-  scan:
-    runs-on: ubuntu-latest
-    steps:
-      - name: Checkout (need scripts/ only)
-        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
-        with:
-          sparse-checkout: |
-            scripts/check-stale-promote-pr.sh
-          sparse-checkout-cone-mode: false
-      - name: Run stale-PR detector
-        env:
-          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          GITHUB_REPOSITORY: ${{ github.repository }}
-          STALE_HOURS: ${{ inputs.stale_hours || '4' }}
-          POST_COMMENT: ${{ inputs.post_comment || 'true' }}
-        run: |
-          # The script's exit code reflects the count of stale PRs.
-          # We don't want a stale finding to fail the workflow run —
-          # the warning + comment are the signal, the green/red is
-          # noise. So convert any non-zero exit to a workflow notice
-          # and exit 0.
-          set +e
-          bash scripts/check-stale-promote-pr.sh
-          rc=$?
-          set -e
-          if [ "$rc" -ne 0 ]; then
-            echo "::notice::Stale PR detector found $rc PR(s) needing attention. See warnings above + comments on the PRs."
-          fi
-          # Always succeed — operator-facing surface is the warning,
-          # not the workflow status.
-          exit 0

.github/workflows/auto-sync-canary.yml

@@ -1,404 +0,0 @@
-name: Auto-sync canary — AUTO_SYNC_TOKEN rotation drift
-
-# Synthetic health check for the AUTO_SYNC_TOKEN secret consumed by
-# auto-sync-main-to-staging.yml (PR #66) and publish-workspace-server-image.yml.
-#
-# ============================================================
-# Why this workflow exists
-# ============================================================
-#
-# PR #66 fixed auto-sync (replaced GitHub-era `gh pr create` — which
-# 405s on Gitea's GraphQL endpoint — with a direct git push from the
-# `devops-engineer` persona's `AUTO_SYNC_TOKEN`). Hostile self-review
-# weakest spot #3 of that PR:
-#
-#   "Token rotation silently breaks auto-sync. If AUTO_SYNC_TOKEN is
-#    rotated without updating the repo secret, every push to main
-#    fails red on the auto-sync push step. The workflow surfaces the
-#    failure mode in the step summary (failure mode B in the header),
-#    but there's no proactive monitoring."
-#
-# Detection latency under the status quo: rotation is only caught on
-# the next push to `main`. During quiet periods (no main push for
-# many hours) the staging-superset-of-main invariant silently breaks.
-#
-# This workflow closes the gap: every 6 hours, it fires the auth
-# surface that auto-sync depends on and emits a red workflow status
-# if AUTO_SYNC_TOKEN has drifted out of validity.
-#
-# ============================================================
-# What this checks (Option B — read-only verify)
-# ============================================================
-#
-# 1. `GET /api/v1/user` against Gitea with the token → validates the
-#    token authenticates AND resolves to `devops-engineer` (catches
-#    the case where the token was regenerated under a different
-#    persona by mistake).
-# 2. `GET /api/v1/repos/molecule-ai/molecule-core` with the token →
-#    validates the token has `read:repository` scope on this repo
-#    (the v2 scope contract — see saved memory
-#    `reference_persona_token_v2_scope`).
-# 3. `git push --dry-run` of the current staging SHA back to
-#    `refs/heads/staging` via `https://oauth2:<token>@<gitea>/...`
-#    → validates the EXACT HTTPS basic-auth path that
-#    `actions/checkout` + `git push origin staging` use inside
-#    auto-sync-main-to-staging.yml. NOP by construction (push the
-#    current tip to itself = "Everything up-to-date"); auth is
-#    checked at the smart-protocol handshake BEFORE the empty-diff
-#    computation, so bad token → exit 128 with "Authentication
-#    failed". `git ls-remote` is NOT used here because Gitea
-#    falls back to anonymous read on public repos and would
-#    silently green-light a rotated token.
-#
-# Each step exits non-zero with an actionable error message if it
-# fails. The workflow status itself is the operator-facing surface.
-#
-# ============================================================
-# What this does NOT check (intentional)
-# ============================================================
-#
-# - **Branch-protection authz** (failure mode C in auto-sync header):
-#   would require an actual write to staging. Already monitored by
-#   `branch-protection-drift.yml` daily. Don't duplicate.
-# - **Conflict resolution** (failure mode A): a real conflict is data-
-#   driven, not auth-driven; can't synthesise it without polluting
-#   staging. Already surfaces immediately on the next main push.
-# - **Concurrency** (failure mode D): handled by workflow concurrency
-#   group on auto-sync, not a credential issue.
-#
-# ============================================================
-# Why Option B (read-only) and not the alternatives
-# ============================================================
-#
-# Considered + rejected (see issue #72 for full write-up):
-#
-# - **Option A — full auto-sync on schedule**: every run creates a
-#   no-op merge commit on staging when main hasn't advanced. 4 noise
-#   commits/day. And races the real `push:` trigger when main has
-#   advanced. Rejected.
-#
-# - **Option C — push to dedicated `auto-sync-canary` branch**: would
-#   exercise authz too, but adds branch noise on Gitea AND requires
-#   maintaining a second branch protection (or expanding staging's
-#   whitelist to a junk branch). Authz already covered by
-#   `branch-protection-drift.yml`. Rejected.
-#
-# Prior art for the chosen Option B shape:
-#   - Cloudflare's `/user/tokens/verify` endpoint (read-only auth
-#     probe explicitly designed for credential canaries).
-#   - AWS Secrets Manager rotation Lambda's `testSecret` step (auth
-#     probe before promoting AWSPENDING → AWSCURRENT).
-#   - HashiCorp Vault's `vault token lookup` for renewal canaries.
-#
-# ============================================================
-# Operator runbook — what to do when this workflow goes RED
-# ============================================================
-#
-# 1. **Identify which step failed**:
-#    - Step "Verify token authenticates as devops-engineer" red →
-#      token is invalid OR resolves to wrong persona.
-#    - Step "Verify token has repo read scope" red → token valid but
-#      stripped of `read:repository` scope (or repo perms changed).
-#    - Step "Verify git HTTPS auth path via no-op dry-run push to
-#      staging" red → token rotated/revoked OR Gitea git-HTTPS
-#      surface is broken (rare). Auth check happens on the
-#      smart-protocol handshake, separate from the API path.
-#
-# 2. **Re-issue the token** on the operator host:
-#    ```
-#    ssh root@5.78.80.188 'docker exec --user git molecule-gitea-1 \
-#      gitea admin user generate-access-token \
-#      --username devops-engineer \
-#      --token-name persona-devops-engineer-vN \
-#      --scopes "read:repository,write:repository,read:user,read:organization,read:issue,write:issue,read:notification,read:misc"'
-#    ```
-#    Update `/etc/molecule-bootstrap/agent-secrets.env` in place
-#    (per `feedback_unified_credentials_file`). The previous token
-#    file lands at `.bak.<date>`.
-#
-# 3. **Update the repo Actions secret** at:
-#    Settings → Secrets and variables → Actions → AUTO_SYNC_TOKEN
-#    Paste the new token. (Don't echo it in chat — but per
-#    `feedback_passwords_in_chat_are_burned`, a paste in a 1:1
-#    Claude session is within trust boundary.)
-#
-# 4. **Re-run this canary** via workflow_dispatch. Confirm GREEN.
-#
-# 5. **Backfill any missed main → staging syncs** by re-running
-#    `auto-sync-main-to-staging.yml` from its workflow_dispatch
-#    surface, OR by pushing an empty commit to main (if you'd
-#    rather force a real trigger).
-#
-# ============================================================
-# Security notes
-# ============================================================
-#
-# - Token usage: read-only (`GET /api/v1/user`, `GET /api/v1/repos/...`,
-#   `git ls-remote`). No write paths. Same blast-radius profile as
-#   `actions/checkout` on a public repo.
-# - The token NEVER appears in logs: every `curl` uses a header
-#   variable, never inline; the `git ls-remote` URL builds the
-#   `oauth2:$TOKEN@host` form into a single env var that's not
-#   echoed. GitHub Actions secret-masking covers anything that does
-#   slip through.
-# - No new token introduced — same `AUTO_SYNC_TOKEN` the workflow
-#   under monitor uses. Per least-privilege we deliberately do NOT
-#   broaden scope for the canary.
-
-on:
-  schedule:
-    # Every 6 hours at :17 (offsets the cron herd at :00). Justification
-    # from issue #72: cheap to run (~5s wall-clock, no quota), 3h average
-    # detection latency, 6h max. 1h would be 24× the runs for marginal
-    # benefit; daily would be 6× longer latency and worse than status
-    # quo on a quiet-main day.
-    - cron: '17 */6 * * *'
-  workflow_dispatch:
-
-# No concurrency group needed — the canary is read-only and idempotent.
-# Two parallel runs (e.g. operator dispatch during a scheduled tick) are
-# harmless: same result, doubled HTTPS calls, no shared state.
-
-permissions:
-  contents: read
-
-jobs:
-  verify-token:
-    name: Verify AUTO_SYNC_TOKEN validity
-    runs-on: ubuntu-latest
-    # 2 min surfaces hangs (Gitea API stall, DNS issue) within one
-    # cron interval. Realistic worst case is ~10s: 2 curls + 1 git
-    # ls-remote, each capped by the explicit timeouts below.
-    timeout-minutes: 2
-
-    env:
-      # Pinned in env so individual steps can read it without
-      # repeating the secret reference. GitHub masks the value in
-      # logs automatically.
-      AUTO_SYNC_TOKEN: ${{ secrets.AUTO_SYNC_TOKEN }}
-      # MUST stay in sync with auto-sync-main-to-staging.yml's
-      # `git config user.name "devops-engineer"` line. Renaming the
-      # devops-engineer persona requires updating both files (and
-      # the staging branch protection's `push_whitelist_usernames`).
-      EXPECTED_PERSONA: devops-engineer
-      GITEA_HOST: git.moleculesai.app
-      REPO_PATH: molecule-ai/molecule-core
-
-    steps:
-      - name: Verify AUTO_SYNC_TOKEN secret is configured
-        # Schedule-vs-dispatch behaviour split, per
-        # `feedback_schedule_vs_dispatch_secrets_hardening`:
-        #
-        #   - schedule: hard-fail when the secret is missing. The
-        #     whole point of the canary is to surface drift; soft-
-        #     skipping on missing-secret would make the canary
-        #     itself drift-invisible (sweep-cf-orphans #2088 lesson).
-        #   - workflow_dispatch: hard-fail too — there's no scenario
-        #     where an operator wants this canary to silently no-op.
-        #     The workflow has no other ad-hoc utility; if you ran
-        #     it, you wanted the answer.
-        run: |
-          if [ -z "${AUTO_SYNC_TOKEN}" ]; then
-            echo "::error::AUTO_SYNC_TOKEN secret is not set on this repo." >&2
-            echo "::error::Set it at Settings → Secrets and variables → Actions." >&2
-            echo "::error::Without it, auto-sync-main-to-staging.yml will fail every push to main." >&2
-            exit 1
-          fi
-          echo "AUTO_SYNC_TOKEN is configured (value masked)."
-
-      - name: Verify token authenticates as ${{ env.EXPECTED_PERSONA }}
-        # Calls Gitea's `/api/v1/user` — the canonical
-        # auth-probe-with-no-side-effects endpoint (mirrors
-        # Cloudflare's /user/tokens/verify).
-        #
-        # Failure surfaces:
-        #   - HTTP 401: token invalid (rotated, revoked, or never
-        #     correctly registered).
-        #   - HTTP 200 but username != devops-engineer: token was
-        #     regenerated under the wrong persona — this would let
-        #     auth pass but commit attribution would be wrong, and
-        #     branch-protection authz would fail because only
-        #     `devops-engineer` is whitelisted.
-        run: |
-          set -euo pipefail
-          response_file="$(mktemp)"
-          code_file="$(mktemp)"
-          # `--max-time 30`: full call ceiling. `--connect-timeout 10`:
-          # DNS + TCP. `-w "%{http_code}"` routed to a tempfile so curl's
-          # exit code can't pollute the captured status — see
-          # feedback_curl_status_capture_pollution + the
-          # `lint-curl-status-capture.yml` gate that rejects the unsafe
-          # `$(curl ... || echo "000")` shape.
-          set +e
-          curl -sS -o "$response_file" \
-            --max-time 30 --connect-timeout 10 \
-            -w "%{http_code}" \
-            -H "Authorization: token ${AUTO_SYNC_TOKEN}" \
-            -H "Accept: application/json" \
-            "https://${GITEA_HOST}/api/v1/user" >"$code_file" 2>/dev/null
-          set -e
-          status=$(cat "$code_file" 2>/dev/null || true)
-          [ -z "$status" ] && status="000"
-
-          if [ "$status" != "200" ]; then
-            echo "::error::Token rotation suspected: GET /api/v1/user returned HTTP $status (expected 200)." >&2
-            echo "::error::Likely cause: AUTO_SYNC_TOKEN has been rotated/revoked on Gitea but the repo Actions secret was not updated." >&2
-            echo "::error::Runbook: see header comment of this workflow file." >&2
-            # Print response body but redact anything that looks like a token.
-            sed -E 's/[A-Fa-f0-9]{32,}/<redacted>/g' "$response_file" >&2 || true
-            exit 1
-          fi
-
-          username=$(python3 -c "import json,sys; print(json.load(open(sys.argv[1])).get('login',''))" "$response_file")
-          if [ "$username" != "${EXPECTED_PERSONA}" ]; then
-            echo "::error::Token resolves to user '$username', expected '${EXPECTED_PERSONA}'." >&2
-            echo "::error::AUTO_SYNC_TOKEN must be the devops-engineer persona PAT (not founder PAT, not another persona)." >&2
-            echo "::error::Auto-sync push will fail because only 'devops-engineer' is whitelisted on staging branch protection." >&2
-            exit 1
-          fi
-          echo "Token authenticates as: $username ✓"
-
-      - name: Verify token has repo read scope
-        # `GET /api/v1/repos/<owner>/<repo>` requires `read:repository`
-        # on the persona's v2 scope contract. If the scope was
-        # narrowed/dropped on rotation we catch it here, before the
-        # next main push reveals it via a checkout failure.
-        run: |
-          set -euo pipefail
-          response_file="$(mktemp)"
-          code_file="$(mktemp)"
-          # See first probe step for the rationale on the tempfile-routed
-          # `-w "%{http_code}"` pattern — the unsafe `|| echo "000"` shape
-          # is rejected by lint-curl-status-capture.yml.
-          set +e
-          curl -sS -o "$response_file" \
-            --max-time 30 --connect-timeout 10 \
-            -w "%{http_code}" \
-            -H "Authorization: token ${AUTO_SYNC_TOKEN}" \
-            -H "Accept: application/json" \
-            "https://${GITEA_HOST}/api/v1/repos/${REPO_PATH}" >"$code_file" 2>/dev/null
-          set -e
-          status=$(cat "$code_file" 2>/dev/null || true)
-          [ -z "$status" ] && status="000"
-
-          if [ "$status" != "200" ]; then
-            echo "::error::Token lacks read:repository scope on ${REPO_PATH}: HTTP $status." >&2
-            echo "::error::Auto-sync's actions/checkout step will fail with this token." >&2
-            echo "::error::Re-issue with v2 scope contract: read:repository,write:repository,read:user,read:organization,read:issue,write:issue,read:notification,read:misc" >&2
-            sed -E 's/[A-Fa-f0-9]{32,}/<redacted>/g' "$response_file" >&2 || true
-            exit 1
-          fi
-          echo "Token has read:repository on ${REPO_PATH} ✓"
-
-      - name: Verify git HTTPS auth path via no-op dry-run push to staging
-        # Final probe: exercise the EXACT auth path that
-        # `actions/checkout` + `git push origin staging` use in
-        # auto-sync-main-to-staging.yml. Gitea's API and git-HTTPS
-        # surfaces share the token-lookup code path internally but
-        # the wire-level error shapes differ — historically (#173)
-        # the API path was healthy while git-HTTPS rejected, so
-        # checking only the API would have given false-green.
-        #
-        # IMPORTANT: `git ls-remote` on a public repo (which
-        # molecule-core is) succeeds even with a junk token because
-        # Gitea falls back to anonymous-read. `ls-remote` therefore
-        # CANNOT validate auth on this surface. We use
-        # `git push --dry-run` instead — push is auth-gated even on
-        # public repos.
-        #
-        # NOP shape: read the current staging SHA via authenticated
-        # ls-remote (the SHA itself is public; auth is incidental
-        # here, used only to colocate the discovery in one step), then
-        # `git push --dry-run <SHA>:refs/heads/staging`. Pushing the
-        # current tip back to itself is "Everything up-to-date" with
-        # exit 0 when auth succeeds. With a bad token Gitea returns
-        # HTTP 401 in the smart-protocol handshake and git exits 128
-        # with "Authentication failed".
-        #
-        # The dry-run never reaches Gitea's pre-receive hook (which
-        # is where branch-protection authz runs), so this probe does
-        # not validate failure mode C. That's intentional —
-        # branch-protection-drift.yml owns authz monitoring; this
-        # canary owns auth.
-        env:
-          # Don't hang waiting for password prompt if auth fails on a
-          # terminal-attached run. (In Actions there's no terminal,
-          # but the env-var hardens against an interactive runner
-          # config.)
-          GIT_TERMINAL_PROMPT: "0"
-        run: |
-          set -euo pipefail
-          # Token is in $AUTO_SYNC_TOKEN (job-level env). Compose the
-          # URL as a local var that's never echoed.
-          url="https://oauth2:${AUTO_SYNC_TOKEN}@${GITEA_HOST}/${REPO_PATH}"
-
-          # Step a: read current staging SHA. ~1KB; auth-gated only
-          # on private repos but always works on public — used here
-          # only to discover the SHA, not to validate auth.
-          staging_ref=$(timeout 30s git ls-remote --refs "$url" refs/heads/staging 2>&1) || {
-            redacted=$(echo "$staging_ref" | sed -E "s|oauth2:[^@]+@|oauth2:<redacted>@|g")
-            echo "::error::ls-remote against staging failed (network/DNS issue):" >&2
-            echo "$redacted" >&2
-            exit 1
-          }
-          if ! echo "$staging_ref" | grep -qE '^[0-9a-f]{40}[[:space:]]+refs/heads/staging$'; then
-            echo "::error::ls-remote returned unexpected shape:" >&2
-            echo "$staging_ref" | sed -E "s|oauth2:[^@]+@|oauth2:<redacted>@|g" >&2
-            exit 1
-          fi
-          staging_sha=$(echo "$staging_ref" | awk '{print $1}')
-
-          # Step b: spin up an ephemeral local repo. `git push` always
-          # requires a local repo even when pushing a remote SHA that
-          # isn't in the local object DB (the protocol negotiates and
-          # discovers we don't need to send any objects). We don't use
-          # `actions/checkout` for this — it would clone the whole
-          # repo (~hundreds of MB) for what's essentially `git init`.
-          tmp_repo="$(mktemp -d)"
-          trap 'rm -rf "$tmp_repo"' EXIT
-          git -C "$tmp_repo" init -q
-          # Author config required for any git operation; values are
-          # arbitrary because nothing gets committed here.
-          git -C "$tmp_repo" config user.email canary@auto-sync.local
-          git -C "$tmp_repo" config user.name auto-sync-canary
-
-          # Step c: dry-run push the current staging SHA back to
-          # staging. NOP by construction — the remote tip equals the
-          # SHA we're pushing, so "Everything up-to-date" is the
-          # success path.
-          #
-          # Authentication is checked at the smart-protocol handshake,
-          # BEFORE the dry-run can compute an empty diff. Bad token
-          # → "Authentication failed", exit 128. Good token → exit 0.
-          set +e
-          push_out=$(timeout 30s git -C "$tmp_repo" push --dry-run "$url" "${staging_sha}:refs/heads/staging" 2>&1)
-          push_rc=$?
-          set -e
-
-          if [ "$push_rc" -ne 0 ]; then
-            redacted=$(echo "$push_out" | sed -E "s|oauth2:[^@]+@|oauth2:<redacted>@|g")
-            echo "::error::Token rotation suspected: git push --dry-run against staging failed via the AUTO_SYNC_TOKEN HTTPS auth path (exit $push_rc)." >&2
-            echo "::error::This is the EXACT auth path that actions/checkout + git push use in auto-sync-main-to-staging.yml." >&2
-            echo "::error::Likely cause: AUTO_SYNC_TOKEN was rotated/revoked on Gitea but the repo Actions secret was not updated. Runbook: see header." >&2
-            echo "$redacted" >&2
-            exit 1
-          fi
-
-          echo "git HTTPS auth path: NOP push --dry-run to staging → ${staging_sha:0:8} ✓"
-
-      - name: Summarise canary result
-        # Everything passed — surface a green summary. (Failures
-        # already wrote ::error:: lines and exited above; if we got
-        # here, all three probes passed.)
-        run: |
-          {
-            echo "## Auto-sync canary: GREEN"
-            echo ""
-            echo "AUTO_SYNC_TOKEN is healthy:"
-            echo "- Authenticates as \`${EXPECTED_PERSONA}\` ✓"
-            echo "- Has \`read:repository\` scope on \`${REPO_PATH}\` ✓"
-            echo "- Git HTTPS auth path: no-op dry-run push to \`refs/heads/staging\` succeeds ✓"
-            echo ""
-            echo "Auto-sync main → staging will succeed on the next push to main."
-            echo "If this canary ever goes RED, see the runbook in this workflow's header."
-          } >> "$GITHUB_STEP_SUMMARY"

.github/workflows/auto-sync-main-to-staging.yml

@@ -1,255 +0,0 @@
-name: Auto-sync main → staging
-
-# Reflects every push to `main` back onto `staging` so the
-# staging-as-superset-of-main invariant holds.
-#
-# ============================================================
-# What this workflow does
-# ============================================================
-#
-# On every push to `main`:
-#   1. Checks if staging already contains main → no-op.
-#   2. Fetches both branches, merges main into staging in the
-#      runner workspace (fast-forward if possible, else
-#      `--no-ff` merge commit).
-#   3. Pushes staging directly to origin via the
-#      `devops-engineer` persona's `AUTO_SYNC_TOKEN`.
-#
-# Authoritative path: a single `git push origin staging` from
-# inside this workflow is the SSOT for advancing staging after
-# a main push. No PR, no merge queue, no human approval —
-# staging is mechanically maintained as a superset of main.
-#
-# `auto-promote-staging.yml` is the reverse-direction
-# counterpart (staging → main, gated on green CI). Together
-# they keep the staging-superset-of-main invariant tight.
-#
-# ============================================================
-# Why direct push (and not "open a PR")
-# ============================================================
-#
-# Pre-2026-05-06 the canonical SCM was GitHub.com, where:
-#   - The `staging` branch had a `merge_queue` ruleset that
-#     blocked ALL direct pushes (no bypass even for org
-#     admins or the GitHub Actions integration).
-#   - Therefore this workflow opened a PR via `gh pr create`
-#     and let auto-merge land it through the queue.
-#
-# Post-2026-05-06 the canonical SCM is Gitea
-# (`git.moleculesai.app/molecule-ai/molecule-core`). Gitea:
-#   - Has no `merge_queue` concept.
-#   - Allows direct push to protected branches via per-user
-#     `push_whitelist_usernames` on the branch protection.
-#   - Does not expose a GraphQL endpoint, so `gh pr create`
-#     returns `HTTP 405 Method Not Allowed
-#     (https://git.moleculesai.app/api/graphql)` — the
-#     pre-suspension architecture cannot work on Gitea.
-#
-# The molecule-ai/molecule-core staging branch protection
-# (verified via `GET /api/v1/repos/.../branch_protections`)
-# whitelists `devops-engineer` for direct push. So the
-# correct Gitea-shape architecture is: authenticate as
-# `devops-engineer`, merge locally, push staging directly.
-#
-# This is structurally simpler than the GitHub-era PR dance
-# and removes the dependence on `gh` CLI / GraphQL entirely.
-#
-# ============================================================
-# Identity + token (anti-bot-ring per saved-memory
-# `feedback_per_agent_gitea_identity_default`)
-# ============================================================
-#
-# This workflow uses `secrets.AUTO_SYNC_TOKEN`, which is a
-# personal access token issued to the `devops-engineer`
-# persona on Gitea — NOT the founder PAT. The bot-ring
-# fingerprint that triggered the GitHub org suspension on
-# 2026-05-06 was characterised by founder PAT acting as CI
-# at machine speed; per-persona identities split the
-# attribution honestly.
-#
-# Token scope on Gitea: repo write. Push target restricted
-# to `staging` (this workflow is the only writer; main is
-# untouched). Compromise blast radius: bounded to staging
-# branch + this repo's read surface.
-#
-# Commits are authored by the persona email
-# `devops-engineer@agents.moleculesai.app` so commit history
-# reflects which automation produced the merge.
-#
-# ============================================================
-# Failure modes & operational notes
-# ============================================================
-#
-# A — staging has commits main doesn't, and the merge
-#     conflicts:
-#     - The `--no-ff` merge step exits non-zero. Workflow
-#       fails red. Operator (devops-engineer or human)
-#       resolves manually:
-#         git fetch origin
-#         git checkout staging
-#         git merge --no-ff origin/main
-#         # resolve conflicts
-#         git push origin staging
-#     - Step summary surfaces the conflict so the failed run
-#       is self-explanatory.
-#
-# B — `AUTO_SYNC_TOKEN` rotated / wrong scope:
-#     - `git push` step exits non-zero with `HTTP 401` /
-#       `403`. Step summary surfaces the failed push.
-#     - Re-issue the token from `~/.molecule-ai/personas/`
-#       on the operator host and update the repo Actions
-#       secret. Re-run the workflow.
-#
-# C — staging branch protection no longer whitelists
-#     `devops-engineer`:
-#     - `git push` exits non-zero with a Gitea protected-
-#       branch rejection. Step summary surfaces it.
-#     - Re-add `devops-engineer` to
-#       `push_whitelist_usernames` on the staging
-#       protection (Settings → Branches → staging).
-#
-# D — concurrent push to main while a sync is in flight:
-#     - The `concurrency` group below serialises runs.
-#       The second waits for the first; if main advances
-#       again while we're syncing, the second run picks
-#       up the new tip on its own fetch.
-#
-# ============================================================
-# Loop safety
-# ============================================================
-#
-# The push to staging from this workflow does NOT itself
-# fire a `push: branches: [main]` event (different branch),
-# so there's no risk of self-recursion. `auto-promote-staging.yml`
-# fires on `workflow_run` of CI etc. — it sees the new
-# staging tip on its next gate-completion event, NOT on this
-# push directly. No loop.
-
-on:
-  push:
-    branches: [main]
-  # workflow_dispatch lets operators manually backfill a
-  # missed sync (e.g. if AUTO_SYNC_TOKEN was rotated and a
-  # main push slipped through while the secret was stale).
-  workflow_dispatch:
-
-permissions:
-  contents: write
-
-concurrency:
-  group: auto-sync-main-to-staging
-  cancel-in-progress: false
-
-jobs:
-  sync-staging:
-    runs-on: ubuntu-latest
-    steps:
-      - name: Checkout staging (with devops-engineer push token)
-        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
-        with:
-          fetch-depth: 0
-          ref: staging
-          # AUTO_SYNC_TOKEN authenticates as the
-          # `devops-engineer` Gitea persona — the only
-          # identity whitelisted for direct push to
-          # staging. See header comment for context.
-          token: ${{ secrets.AUTO_SYNC_TOKEN }}
-
-      - name: Configure git author
-        run: |
-          # Per-persona identity, NOT founder PAT.
-          # `feedback_per_agent_gitea_identity_default`.
-          git config user.name "devops-engineer"
-          git config user.email "devops-engineer@agents.moleculesai.app"
-
-      - name: Check if staging already contains main
-        id: check
-        run: |
-          set -euo pipefail
-          git fetch origin main
-          if git merge-base --is-ancestor origin/main HEAD; then
-            echo "needs_sync=false" >> "$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 "::notice::staging is missing main's tip (${MAIN_SHORT}) — merging in-runner and pushing"
-          fi
-
-      - name: Merge main into staging (in-runner)
-        if: steps.check.outputs.needs_sync == 'true'
-        id: merge
-        run: |
-          set -euo pipefail
-          # Already on staging from checkout. Try fast-forward
-          # first (cleanest history); fall back to merge commit
-          # if staging has commits main doesn't.
-          if git merge --ff-only origin/main; then
-            echo "did_ff=true" >> "$GITHUB_OUTPUT"
-            echo "::notice::Fast-forwarded staging to origin/main"
-          else
-            echo "did_ff=false" >> "$GITHUB_OUTPUT"
-            if ! git merge --no-ff origin/main \
-                -m "chore: sync main → staging (auto, ${{ steps.check.outputs.main_short }})"; then
-              # Hygiene: leave the work tree clean before failing.
-              git merge --abort || true
-              {
-                echo "## Conflict"
-                echo
-                echo "Auto-merge \`main → staging\` failed with conflicts."
-                echo "A human (or devops-engineer persona) needs to resolve manually:"
-                echo
-                echo '```'
-                echo "git fetch origin"
-                echo "git checkout staging"
-                echo "git merge --no-ff origin/main"
-                echo "# resolve conflicts"
-                echo "git push origin staging"
-                echo '```'
-              } >> "$GITHUB_STEP_SUMMARY"
-              exit 1
-            fi
-          fi
-
-      - name: Push staging to origin
-        if: steps.check.outputs.needs_sync == 'true'
-        run: |
-          set -euo pipefail
-          # Direct push to staging. devops-engineer persona is
-          # whitelisted for direct push on the staging branch
-          # protection (Settings → Branches → staging).
-          #
-          # No --force / --force-with-lease: a fast-forward or
-          # legitimate merge commit on top of current staging
-          # is the only thing we'd ever push. If origin/staging
-          # advanced under us (concurrent merge), the push
-          # legitimately rejects and the next run picks up the
-          # new state.
-          if ! git push origin staging; then
-            {
-              echo "## Push rejected"
-              echo
-              echo "Direct push to \`staging\` failed. Likely causes:"
-              echo "- \`AUTO_SYNC_TOKEN\` rotated / wrong scope (HTTP 401/403)"
-              echo "- \`devops-engineer\` no longer in"
-              echo "  \`push_whitelist_usernames\` on the staging"
-              echo "  branch protection (HTTP 422)"
-              echo "- staging advanced concurrently — re-running this"
-              echo "  workflow on the new main tip will pick it up"
-            } >> "$GITHUB_STEP_SUMMARY"
-            exit 1
-          fi
-
-          {
-            echo "## Auto-sync succeeded"
-            echo
-            echo "- staging advanced to: \`$(git rev-parse --short=8 HEAD)\`"
-            echo "- main tip: \`${{ steps.check.outputs.main_short }}\`"
-            echo "- Strategy: $([ "${{ steps.merge.outputs.did_ff }}" = "true" ] && echo "fast-forward" || echo "merge commit")"
-            echo "- Pushed by: \`devops-engineer\` (per-agent persona, anti-bot-ring)"
-          } >> "$GITHUB_STEP_SUMMARY"

.github/workflows/canary-staging.yml

@@ -20,6 +20,19 @@ on:
     # a few minutes under load — that's fine for a canary.
     - cron: '*/30 * * * *'
   workflow_dispatch:
+    inputs:
+      keep_on_failure:
+        description: >-
+          Skip teardown when the canary fails (debugging only). The
+          tenant org + EC2 + CF tunnel + DNS stay alive so an operator
+          can SSM into the workspace EC2 and capture docker logs of the
+          failing claude-code container. REMEMBER to manually delete
+          via DELETE /cp/admin/tenants/<slug> when done so the org
+          doesn't accumulate cost. Only honored on workflow_dispatch;
+          cron runs always tear down (we don't want unattended cron
+          to leak resources).
+        type: boolean
+        default: false
 
 # Serialise with the full-SaaS workflow so they don't contend for the
 # same org-create quota on staging. Different group key from
@@ -80,6 +93,14 @@ jobs:
       # is "Token Plan only" but cheap-per-token and fast.
       E2E_MODEL_SLUG: MiniMax-M2.7-highspeed
       E2E_RUN_ID: "canary-${{ github.run_id }}"
+      # Debug-only: when an operator dispatches with keep_on_failure=true,
+      # the canary script's E2E_KEEP_ORG=1 path skips teardown so the
+      # tenant org + EC2 stay alive for SSM-based log capture. Cron runs
+      # never set this (the input only exists on workflow_dispatch) so
+      # unattended cron always tears down. See molecule-core#129
+      # failure mode #1 — capturing the actual exception requires
+      # docker logs from the live container.
+      E2E_KEEP_ORG: ${{ github.event.inputs.keep_on_failure == 'true' && '1' || '0' }}
 
     steps:
       - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
@@ -137,27 +158,28 @@ jobs:
         id: canary
         run: bash tests/e2e/test_staging_full_saas.sh
 
-      # Alerting: open an issue only after THREE consecutive failures so
-      # transient flakes (Cloudflare DNS hiccup, AWS API blip) don't spam
-      # the issue list. If an issue is already open, we still comment on
-      # every failure so ops sees the streak. Auto-close on next green.
+      # Alerting: open a sticky issue on the FIRST failure; comment on
+      # subsequent failures; auto-close on next green. Comment-on-existing
+      # de-duplicates so a single open issue accumulates the streak —
+      # ops sees one issue with N comments rather than N issues.
       #
-      # Threshold rationale: canary fires every 30 min, so 3 failures =
-      # ~90 min of consecutive red — well past any single-run flake but
-      # still tight enough that a real outage gets surfaced before the
-      # next deploy window.
+      # Why no consecutive-failures threshold (e.g., wait 3 runs before
+      # filing): the prior threshold check used
+      # `github.rest.actions.listWorkflowRuns()` which Gitea 1.22.6 does
+      # not expose (returns 404). On Gitea Actions the threshold call
+      # ALWAYS failed, breaking the entire alerting step and going days
+      # silent on real regressions (38h+ chronic red on 2026-05-07/08
+      # before this fix; tracked in molecule-core#129). Filing on first
+      # failure is also better UX — we want to know about the first red,
+      # not wait 90 min for it to "count." Real flakes get one issue +
+      # a quick close-on-green; persistent reds accumulate comments.
       - name: Open issue on failure
         if: failure()
         uses: actions/github-script@3a2844b7e9c422d3c10d287c895573f7108da1b3 # v9.0.0
-        env:
-          # Inject the workflow path explicitly — context.workflow is
-          # the *name*, not the file path the actions API needs.
-          WORKFLOW_PATH: '.github/workflows/canary-staging.yml'
-          CONSECUTIVE_THRESHOLD: '3'
         with:
           script: |
             const title = '🔴 Canary failing: staging SaaS smoke';
-            const runURL = `https://github.com/${context.repo.owner}/${context.repo.repo}/actions/runs/${context.runId}`;
+            const runURL = `${context.serverUrl}/${context.repo.owner}/${context.repo.repo}/actions/runs/${context.runId}`;
 
             // Find an existing open canary issue (stable title match).
             // If one exists, this isn't a "first failure" — comment and exit.
@@ -177,32 +199,12 @@ jobs:
               return;
             }
 
-            // No open issue yet — check the last N-1 runs' conclusions.
-            // We open the issue only if the last (THRESHOLD-1) runs ALSO
-            // failed (so this is the 3rd consecutive red).
-            const threshold = parseInt(process.env.CONSECUTIVE_THRESHOLD, 10);
-            const { data: runs } = await github.rest.actions.listWorkflowRuns({
-              owner: context.repo.owner, repo: context.repo.repo,
-              workflow_id: process.env.WORKFLOW_PATH,
-              status: 'completed',
-              per_page: threshold,
-              // Skip the current in-progress run; it isn't 'completed' yet.
-            });
-            // listWorkflowRuns returns recent first. We need (threshold-1)
-            // prior failures (current run is the threshold-th).
-            const priorFailures = (runs.workflow_runs || [])
-              .slice(0, threshold - 1)
-              .filter(r => r.id !== context.runId)
-              .filter(r => r.conclusion === 'failure')
-              .length;
-            if (priorFailures < threshold - 1) {
-              core.info(`Below threshold: ${priorFailures + 1}/${threshold} consecutive failures — not filing yet`);
-              return;
-            }
-
+            // No open issue yet — file one on this first failure. The
+            // comment-on-existing branch above means subsequent failures
+            // accumulate as comments on this same issue, so we don't
+            // spam new issues per run.
             const body =
-              `Canary run failed at ${new Date().toISOString()}, ` +
-              `${threshold} consecutive runs red.\n\n` +
+              `Canary run failed at ${new Date().toISOString()}.\n\n` +
               `Run: ${runURL}\n\n` +
               `This issue auto-closes on the next green canary run. ` +
               `Consecutive failures add a comment here rather than a new issue.`;
@@ -211,7 +213,7 @@ jobs:
               title, body,
               labels: ['canary-staging', 'bug'],
             });
-            core.info(`Opened canary failure issue (${threshold} consecutive reds)`);
+            core.info('Opened canary failure issue (first red)');
 
       - name: Auto-close canary issue on success
         if: success()

.github/workflows/canary-verify.yml

@@ -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"

.github/workflows/e2e-api.yml

@@ -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 ${i}s"
               exit 0
             fi
@@ -185,6 +298,9 @@ jobs:
             kill "$(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

.github/workflows/e2e-staging-canvas.yml

@@ -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

.github/workflows/e2e-staging-external.yml

@@ -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'

.github/workflows/e2e-staging-saas.yml

@@ -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'

.github/workflows/handlers-postgres-integration.yml

@@ -14,12 +14,42 @@ name: Handlers Postgres Integration
 # self-review caught it took 2 minutes to set up and would have caught
 # the bug at PR-time.
 #
-# This job spins a Postgres service container, applies the migration,
-# and runs `go test -tags=integration` against a live DB. Required
-# check on staging branch protection — backend handler PRs cannot
-# merge without a real-DB regression gate.
+# 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.
 #
-# Cost: ~30s job (postgres pull from GH cache + go build + 4 tests).
+# 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:
@@ -59,20 +89,14 @@ jobs:
     name: Handlers Postgres Integration
     needs: detect-changes
     runs-on: ubuntu-latest
-    services:
-      postgres:
-        image: postgres:15-alpine
-        env:
-          POSTGRES_PASSWORD: test
-          POSTGRES_DB: molecule
-        ports:
-          - 5432:5432
-        # GHA spins this with --health-cmd built in for postgres images.
-        options: >-
-          --health-cmd pg_isready
-          --health-interval 5s
-          --health-timeout 5s
-          --health-retries 10
+    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
@@ -89,16 +113,57 @@ jobs:
         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 ${PG_HOST}:5432"
+
       - if: needs.detect-changes.outputs.handlers == 'true'
         name: Apply migrations to Postgres service
         env:
           PGPASSWORD: test
         run: |
-          # Wait for postgres to actually accept connections (the
-          # GHA --health-cmd is best-effort but psql can still race).
+          # Wait for postgres to actually accept connections. Docker's
+          # health-cmd handles container-side readiness, but the wire
+          # to the bridge IP is best-tested with pg_isready directly.
           for i in {1..15}; do
-            if pg_isready -h localhost -p 5432 -U postgres -q; then break; fi
-            echo "waiting for postgres..."; sleep 2
+            if pg_isready -h "${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
@@ -131,7 +196,7 @@ jobs:
           # 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 localhost -U postgres -d molecule -v ON_ERROR_STOP=1 \
+            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
@@ -145,7 +210,7 @@ jobs:
           # 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 localhost -U postgres -d molecule -tA \
+            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"
@@ -156,16 +221,32 @@ jobs:
 
       - if: needs.detect-changes.outputs.handlers == 'true'
         name: Run integration tests
-        env:
-          INTEGRATION_DB_URL: postgres://postgres:test@localhost:5432/molecule?sslmode=disable
         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: needs.detect-changes.outputs.handlers == 'true' && failure()
+      - if: failure() && needs.detect-changes.outputs.handlers == 'true'
         name: Diagnostic dump on failure
         env:
           PGPASSWORD: test
         run: |
-          echo "::group::delegations table state"
-          psql -h localhost -U postgres -d molecule -c "SELECT * FROM delegations LIMIT 50;" || true
+          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}"
+

.github/workflows/harness-replays.yml

@@ -119,6 +119,17 @@ jobs:
       # symptom, different root cause: staging still has the in-image
       # clone path, hits the auth error directly).
       #
+      # 2026-05-08 sub-finding (#192): the clone step ALSO fails when
+      # any referenced workspace-template repo is private and the
+      # AUTO_SYNC_TOKEN bearer (devops-engineer persona) lacks read
+      # access. Root cause: 5 of 9 workspace-template repos
+      # (openclaw, codex, crewai, deepagents, gemini-cli) had been
+      # marked private with no team grant. Resolution: flipped them
+      # to public per `feedback_oss_first_repo_visibility_default`
+      # (the OSS surface should be public). Layer-3 (customer-private +
+      # marketplace third-party repos) tracked separately in
+      # internal#102.
+      #
       # 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

.github/workflows/pr-guards.yml

@@ -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)."

.github/workflows/publish-runtime.yml

@@ -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 — $(cat /tmp/dispatch.out)"
+            CLONE="$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

.github/workflows/redeploy-tenants-on-staging.yml

@@ -36,7 +36,7 @@ on:
   workflow_run:
     workflows: ['publish-workspace-server-image']
     types: [completed]
-    branches: [staging]
+    branches: [main]
   workflow_dispatch:
     inputs:
       target_tag:

.github/workflows/retarget-main-to-staging.yml

@@ -1,276 +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
-# PRs, etc.) are left alone — they're the authorised exception
-# to the rule.
-#
-# ============================================================
-# What this workflow does
-# ============================================================
-#
-# On `pull_request_target` opened/reopened against `main`:
-#   1. If the PR head is `staging`, skip (the auto-promote PRs
-#      MUST stay base=main).
-#   2. If the PR author is a bot, retarget the PR base to
-#      `staging` via Gitea REST `PATCH /pulls/{N}` body
-#      `{"base":"staging"}`.
-#   3. If the retarget returns 422 "pull request already exists
-#      for base branch 'staging'" (issue #1884 case: another PR
-#      on the same head already targets staging), close the
-#      now-redundant main-PR via Gitea REST instead of failing
-#      red.
-#   4. Post an explainer comment on the retargeted PR via
-#      Gitea REST `POST /issues/{N}/comments`.
-#
-# ============================================================
-# Why Gitea REST (and not `gh api / gh pr close / gh pr comment`)
-# ============================================================
-#
-# Pre-2026-05-06 this workflow used `gh api -X PATCH "repos/{owner}/{repo}/pulls/{N}" -f base=staging`
-# plus `gh pr close` and `gh pr comment`. After the GitHub→Gitea
-# cutover those calls fail because:
-#
-#   - `gh` CLI defaults to `api.github.com`. Even with `GH_HOST`
-#     pointing at Gitea, `gh pr close / comment` route through
-#     GraphQL (`/api/graphql`) which Gitea does not expose.
-#     Empirical: every `gh pr *` call returns
-#     `HTTP 405 Method Not Allowed (https://git.moleculesai.app/api/graphql)`
-#     — same root cause as #65 (auto-sync, fixed in PR #66) and
-#     #73/#195 (auto-promote, fixed in PR #78).
-#   - `gh api -X PATCH /pulls/{N}` happens to use a REST path
-#     that Gitea also has, but the `gh` host-resolution layer
-#     and pagination/retry logic don't always hit Gitea cleanly,
-#     and the cost of switching to direct `curl` is one extra
-#     line of code.
-#
-# So this workflow uses direct `curl` calls to Gitea REST. No
-# `gh` CLI dependency, no GraphQL, no flaky host-resolution.
-#
-# ============================================================
-# Identity + token (anti-bot-ring per saved-memory
-# `feedback_per_agent_gitea_identity_default`)
-# ============================================================
-#
-# Pre-fix this workflow used the per-job ephemeral
-# `secrets.GITHUB_TOKEN`. On Gitea Actions that token has
-# narrow scope and unpredictable cross-PR write capability.
-#
-# Post-fix: `secrets.AUTO_SYNC_TOKEN` (the `devops-engineer`
-# Gitea persona). Same persona used by `auto-sync-main-to-staging.yml`
-# (PR #66) and `auto-promote-staging.yml` (PR #78). Token scope:
-# `push: true` repo write, sufficient for PR-edit + close + comment.
-#
-# Why this token does NOT need branch-protection bypass:
-# patching a PR's base ref is a PR-level operation that does not
-# require push perms on either branch (the PR's own commits stay
-# put; only the metadata changes).
-#
-# ============================================================
-# Failure modes & operational notes
-# ============================================================
-#
-# A — PATCH base→staging returns 422 "pull request already exists"
-#     (issue #1884 case):
-#     - Detected by string-match on response body. Workflow
-#       falls through to closing the now-redundant main-PR
-#       (Gitea REST `PATCH /pulls/{N}` with `state: closed`)
-#       and posts an explanation comment. Step summary surfaces.
-#
-# B — `AUTO_SYNC_TOKEN` rotated / wrong scope:
-#     - First REST call returns 401/403. Step summary surfaces.
-#       Re-issue token from `~/.molecule-ai/personas/` on the
-#       operator host and update repo Actions secret.
-#
-# C — PR was deleted between trigger and run:
-#     - REST call returns 404. Workflow exits 0 with a notice
-#       (the rule was already enforced or the PR is gone).
-#
-# D — author is not actually a bot but the filter mis-fires:
-#     - Filter is conservative: only triggers on
-#       `user.type == 'Bot'`, `login` ends with `[bot]`, or
-#       known bot logins (`molecule-ai[bot]`, `app/molecule-ai`).
-#       Human PRs slip through unaffected. If a NEW bot login
-#       starts shipping main-PRs, add it to the filter.
-
-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
-    # `devops-engineer` since PR #78 / #195 fix). Retargeting
-    # head=staging onto base=staging fails with HTTP 422 "no new
-    # commits between base 'staging' and head 'staging'", which
-    # would surface as a noisy red workflow run on every
-    # auto-promote (caught 2026-05-03 on the GitHub-era 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]'
-        || github.event.pull_request.user.login == 'devops-engineer'
-      )
-    steps:
-      - name: Retarget PR base to staging via Gitea REST
-        id: retarget
-        env:
-          GITEA_TOKEN: ${{ secrets.AUTO_SYNC_TOKEN }}
-          GITEA_HOST: ${{ vars.GITEA_HOST || 'https://git.moleculesai.app' }}
-          REPO: ${{ github.repository }}
-          PR_NUMBER: ${{ github.event.pull_request.number }}
-          PR_AUTHOR: ${{ github.event.pull_request.user.login }}
-        # Issue #1884 case: when the bot opens a PR against main
-        # and there's already another PR on the same head branch
-        # targeting staging, Gitea's PATCH returns 422 with a
-        # body mentioning "pull request already exists for base
-        # branch 'staging'" (the Gitea message wording is
-        # slightly different from GitHub's; the substring match
-        # below covers both for forward/back compat).
-        # The retarget can't proceed — but the right response is
-        # to close the now-redundant main-PR, not to fail the
-        # workflow noisily. Detect that specific 422 and close
-        # instead.
-        run: |
-          set -euo pipefail
-
-          API="${GITEA_HOST}/api/v1/repos/${REPO}"
-          AUTH=(-H "Authorization: token ${GITEA_TOKEN}" -H "Accept: application/json")
-
-          echo "Retargeting PR #${PR_NUMBER} (author: ${PR_AUTHOR}) from main → staging"
-
-          # Curl-status-capture pattern per `feedback_curl_status_capture_pollution`:
-          # http_code via -w to its own scalar, body to a tempfile, set +e/-e
-          # bracket so curl's non-zero-on-4xx doesn't pollute the script's exit chain.
-          BODY_FILE=$(mktemp)
-          REQ='{"base":"staging"}'
-
-          set +e
-          STATUS=$(curl -sS "${AUTH[@]}" -H "Content-Type: application/json" \
-            -X PATCH -d "${REQ}" \
-            -o "${BODY_FILE}" -w "%{http_code}" \
-            "${API}/pulls/${PR_NUMBER}")
-          CURL_RC=$?
-          set -e
-
-          if [ "${CURL_RC}" -ne 0 ]; then
-            echo "::error::curl PATCH failed (rc=${CURL_RC})"
-            rm -f "${BODY_FILE}"
-            exit 1
-          fi
-
-          if [ "${STATUS}" = "201" ] || [ "${STATUS}" = "200" ]; then
-            NEW_BASE=$(jq -r '.base.ref // "?"' < "${BODY_FILE}")
-            rm -f "${BODY_FILE}"
-            if [ "${NEW_BASE}" = "staging" ]; then
-              echo "::notice::Retargeted PR #${PR_NUMBER} → staging"
-              echo "outcome=retargeted" >> "$GITHUB_OUTPUT"
-              exit 0
-            fi
-            echo "::error::PATCH returned ${STATUS} but base.ref is '${NEW_BASE}', not 'staging'"
-            exit 1
-          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.
-          BODY=$(cat "${BODY_FILE}" || true)
-          rm -f "${BODY_FILE}"
-
-          if [ "${STATUS}" = "422" ] && echo "${BODY}" | grep -qE "(pull request already exists for base branch 'staging'|already exists.*base.*staging)"; then
-            echo "::notice::PR #${PR_NUMBER}: duplicate target-staging PR exists on same head — closing this main-PR as redundant."
-
-            # Close the now-redundant main-PR via Gitea REST
-            # (PATCH state=closed). Post comment explaining
-            # rationale BEFORE close so the comment lands on the
-            # PR (commenting on a closed PR works on Gitea, but
-            # historically caused notification ordering surprises).
-
-            CLOSE_BODY_FILE=$(mktemp)
-            CMT_REQ=$(jq -n '{body:"[retarget-bot] Closing — another PR on the same head branch already targets `staging`. This PR is redundant. See issue #1884 for the rationale."}')
-            set +e
-            CMT_STATUS=$(curl -sS "${AUTH[@]}" -H "Content-Type: application/json" \
-              -X POST -d "${CMT_REQ}" \
-              -o "${CLOSE_BODY_FILE}" -w "%{http_code}" \
-              "${API}/issues/${PR_NUMBER}/comments")
-            set -e
-            if [ "${CMT_STATUS}" != "201" ]; then
-              echo "::warning::dup-close comment POST returned ${CMT_STATUS}; continuing to close anyway"
-              cat "${CLOSE_BODY_FILE}" | head -c 300 || true
-            fi
-            rm -f "${CLOSE_BODY_FILE}"
-
-            CLOSE_REQ='{"state":"closed"}'
-            CLOSE_RESP=$(mktemp)
-            set +e
-            CL_STATUS=$(curl -sS "${AUTH[@]}" -H "Content-Type: application/json" \
-              -X PATCH -d "${CLOSE_REQ}" \
-              -o "${CLOSE_RESP}" -w "%{http_code}" \
-              "${API}/pulls/${PR_NUMBER}")
-            set -e
-            if [ "${CL_STATUS}" = "201" ] || [ "${CL_STATUS}" = "200" ]; then
-              echo "::notice::Closed PR #${PR_NUMBER} as redundant"
-              echo "outcome=closed-as-duplicate" >> "$GITHUB_OUTPUT"
-              rm -f "${CLOSE_RESP}"
-              exit 0
-            fi
-            echo "::error::Failed to close redundant PR: HTTP ${CL_STATUS}"
-            cat "${CLOSE_RESP}" | head -c 300 || true
-            rm -f "${CLOSE_RESP}"
-            exit 1
-          fi
-
-          echo "::error::Retarget PATCH failed and was NOT a duplicate-base error: HTTP ${STATUS}"
-          echo "${BODY}" | head -c 500 >&2
-          exit 1
-
-      - name: Post explainer comment
-        if: steps.retarget.outputs.outcome == 'retargeted'
-        env:
-          GITEA_TOKEN: ${{ secrets.AUTO_SYNC_TOKEN }}
-          GITEA_HOST: ${{ vars.GITEA_HOST || 'https://git.moleculesai.app' }}
-          REPO: ${{ github.repository }}
-          PR_NUMBER: ${{ github.event.pull_request.number }}
-        run: |
-          set -euo pipefail
-
-          API="${GITEA_HOST}/api/v1/repos/${REPO}"
-          AUTH=(-H "Authorization: token ${GITEA_TOKEN}" -H "Accept: application/json")
-
-          # PR comments live on the issue endpoint in Gitea
-          # (PRs ARE issues — same endpoint, different sub-resources
-          # for diffs/files/etc.). The body uses jq to safely
-          # encode the multi-line markdown without shell-quote
-          # nightmares.
-          REQ=$(jq -n '{body:"[retarget-bot] This PR was opened against `main` and has been retargeted to `staging` automatically.\n\n**Why:** per [SHARED_RULES rule 8](https://git.moleculesai.app/molecule-ai/molecule-ai-org-template-molecule-dev/src/branch/main/SHARED_RULES.md), all feature work targets `staging` first; the CEO promotes `staging → main` separately.\n\n**What changed:** just the base branch — no code change. CI will re-run against `staging`. If you get merge conflicts, rebase on `staging`.\n\n**If this PR is the CEO`s staging→main promotion:** the Action skipped you (only bot-authored PRs are retargeted, head=staging is also exempted). If you see this comment on your CEO PR, that`s a bug — please tag @hongmingwang."}')
-
-          BODY_FILE=$(mktemp)
-          set +e
-          STATUS=$(curl -sS "${AUTH[@]}" -H "Content-Type: application/json" \
-            -X POST -d "${REQ}" \
-            -o "${BODY_FILE}" -w "%{http_code}" \
-            "${API}/issues/${PR_NUMBER}/comments")
-          set -e
-
-          if [ "${STATUS}" = "201" ]; then
-            echo "::notice::Posted explainer comment on PR #${PR_NUMBER}"
-          else
-            echo "::warning::Failed to post explainer (HTTP ${STATUS}) — retarget itself succeeded"
-            cat "${BODY_FILE}" | head -c 300 || true
-          fi
-          rm -f "${BODY_FILE}"

CONTRIBUTING.md

@@ -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.
 

Makefile

@@ -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 ./...

README.md

@@ -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
 

README.zh-CN.md

@@ -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
 

canvas/.dockerignore

@@ -0,0 +1,10 @@
+# Excluded from `docker build` context. Without this, the COPY . . step in
+# canvas/Dockerfile clobbers the freshly-installed node_modules with the
+# host's (potentially broken / wrong-arch) copy — the @tailwindcss/oxide
+# native binary disagreed and broke `next build`.
+node_modules
+.next
+.git
+*.log
+.env*
+!.env.example

canvas/Dockerfile

@@ -1,7 +1,11 @@
 FROM node:22-alpine AS builder
 WORKDIR /app
 COPY package.json package-lock.json* ./
-RUN npm install
+# `npm ci` (not `install`) for lockfile-exact reproducibility.
+# `--include=optional` ensures the platform-specific @tailwindcss/oxide
+# native binary lands — without it, postcss fails with "Cannot read
+# properties of undefined (reading 'All')" at build time.
+RUN npm ci --include=optional
 COPY . .
 ARG NEXT_PUBLIC_PLATFORM_URL=http://localhost:8080
 ARG NEXT_PUBLIC_WS_URL=ws://localhost:8080/ws

canvas/src/app/pricing/page.tsx

@@ -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

canvas/src/components/tabs/ChatTab.tsx

@@ -13,7 +13,6 @@ 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";
 
@@ -50,38 +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
@@ -134,8 +107,19 @@ const INITIAL_HISTORY_LIMIT = 10;
 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).
@@ -143,10 +127,10 @@ const OLDER_HISTORY_BATCH = 20;
  *    timestamp. Combined with limit, this yields the next-older page
  *    when scrolling backward through history.
  *
- * `reachedEnd` is true when the server returned fewer rows than asked
- * for — caller uses this to disable further older-batch fetches.
- * (Counts row-level returns, not chat-bubble count: each row may
- * produce 1-2 bubbles.)
+ * `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,
@@ -154,25 +138,23 @@ async function loadMessagesFromDB(
   beforeTs?: string,
 ): Promise<{ messages: ChatMessage[]; error: string | null; reachedEnd: boolean }> {
   try {
-    const params = new URLSearchParams({
-      type: "a2a_receive",
-      source: "canvas",
-      limit: String(limit),
-    });
+    const params = new URLSearchParams({ limit: String(limit) });
     if (beforeTs) params.set("before_ts", beforeTs);
-    const activities = await api.get<ActivityRowForHydration[]>(
-      `/workspaces/${workspaceId}/activity?${params.toString()}`,
+    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, reachedEnd: activities.length < limit };
+    // 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: [],

canvas/src/components/tabs/ConfigTab.tsx

@@ -21,20 +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);
@@ -53,9 +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"
@@ -221,47 +238,51 @@ 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

canvas/src/components/tabs/__tests__/ChatTab.lazyHistory.test.tsx

@@ -1,13 +1,11 @@
 // @vitest-environment jsdom
 //
-// Pins the lazy-loading chat-history pagination added 2026-05-05.
+// Pins the lazy-loading chat-history pagination.
 //
-// Pre-fix: ChatTab fetched the newest 50 messages on every mount and
-// scrolled to bottom, paying full DOM cost up-front even when the user
-// only wanted to read the last few bubbles. Post-fix: initial load is
-// bounded to 10 newest, and an IntersectionObserver on a top sentinel
-// triggers loadOlder() (batch of 20 with `before_ts` cursor) when the
-// user scrolls up.
+// 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
@@ -20,11 +18,10 @@
 //      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.
-//
-// IntersectionObserver / scroll-anchor restoration is exercised by the
-// E2E synth-canary suite — pinning it in jsdom would require mocking
-// the observer and faking layout, which is brittler than trusting a
-// live-DOM canary against the staging tenant.
+//   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";
@@ -33,24 +30,31 @@ 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. Both fire
-// /activity GETs on mount: MyChat's hits `type=a2a_receive&source=canvas`,
-// AgentComms's hits a different filter. Route the mock by URL so each
-// gets a sensible default and only MyChat's call is what the assertions
-// scrutinise.
-const myChatActivityCalls: string[] = [];
-let myChatNextResponse: { ok: true; rows: unknown[] } | { ok: false; err: Error } = {
-  ok: true,
-  rows: [],
-};
+// 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("type=a2a_receive") && path.includes("source=canvas")) {
-    myChatActivityCalls.push(path);
-    if (myChatNextResponse.ok) return Promise.resolve(myChatNextResponse.rows);
+  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 is a safe
-  // default that won't blow up the corresponding component's .then().
+  // AgentComms / heartbeat / anything else — empty array safe default.
   return Promise.resolve([]);
 });
 const apiPost = vi.fn();
@@ -84,8 +88,8 @@ const ioInstances: IOInstance[] = [];
 beforeEach(() => {
   apiGet.mockClear();
   apiPost.mockReset();
-  myChatActivityCalls.length = 0;
-  myChatNextResponse = { ok: true, rows: [] };
+  myChatHistoryCalls.length = 0;
+  myChatNextResponse = { ok: true, messages: [] };
   ioInstances.length = 0;
   class FakeIO {
     private inst: IOInstance;
@@ -101,20 +105,12 @@ beforeEach(() => {
       this.inst.disconnected = true;
     }
   }
-  // Install on every reachable global — different bundlers / module
-  // graphs can resolve `IntersectionObserver` via `window`, `globalThis`,
-  // or the bare global. Without all three, jsdom's own (pre-existing)
-  // stub silently wins and ioInstances stays empty.
   (window as unknown as { IntersectionObserver: unknown }).IntersectionObserver = FakeIO;
   (globalThis as unknown as { IntersectionObserver: unknown }).IntersectionObserver = FakeIO;
-  // jsdom doesn't implement scrollIntoView; ChatTab calls it after every
-  // messages update.
   Element.prototype.scrollIntoView = vi.fn();
 });
 
 function triggerIntersection(instanceIdx = -1) {
-  // -1 → the latest observer (the live one). Tests targeting an old
-  // (disconnected) instance pass a positive index.
   const inst = ioInstances.at(instanceIdx);
   if (!inst) throw new Error(`no IO instance at ${instanceIdx}`);
   inst.callback(
@@ -125,25 +121,30 @@ function triggerIntersection(instanceIdx = -1) {
 
 import { ChatTab } from "../ChatTab";
 
-function makeActivityRow(seq: number): Record<string, unknown> {
-  // Zero-pad seq into the minute slot so "seq=10" doesn't produce
-  // the invalid timestamp "00:010:00Z" (caught by the loadOlder URL
-  // assertion below — first version of the helper used `0${seq}` and
-  // the test failed on `before_ts` having an extra digit).
+// makeMessagePair returns a (user, agent) pair sharing a timestamp,
+// matching the wire shape /chat-history emits per activity_logs row.
+// Server-side reverseRowChunks ensures the wire is oldest-first across
+// rows but [user, agent] within each row.
+function makeMessagePair(seq: number): unknown[] {
+  // Zero-pad seq into the minute slot so seq=10 produces a valid
+  // timestamp (00:10:00Z, not 00:010:00Z).
   const mm = String(seq).padStart(2, "0");
-  return {
-    activity_type: "a2a_receive",
-    status: "ok",
-    created_at: `2026-05-05T00:${mm}:00Z`,
-    request_body: { params: { message: { parts: [{ kind: "text", text: `user msg ${seq}` }] } } },
-    response_body: { result: `agent reply ${seq}` },
-  };
+  const ts = `2026-05-05T00:${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 },
+  ];
 }
 
-// Server returns newest-first; the helper builds a server-shape page
-// so the order in the rendered messages array matches production.
-function newestFirstPage(start: number, count: number): unknown[] {
-  return Array.from({ length: count }, (_, i) => makeActivityRow(start + count - 1 - i));
+// 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 = {
@@ -153,28 +154,30 @@ const minimalData = {
 } as unknown as Parameters<typeof ChatTab>[0]["data"];
 
 describe("ChatTab lazy history pagination", () => {
-  it("initial fetch carries limit=10 (not the legacy 50)", async () => {
-    myChatNextResponse = { ok: true, rows: [makeActivityRow(1)] };
+  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(myChatActivityCalls.length).toBe(1));
-    const url = myChatActivityCalls[0];
+    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 (it can't, since the sentinel is what triggers it).
-    myChatNextResponse = {
-      ok: true,
-      rows: [makeActivityRow(1), makeActivityRow(2), makeActivityRow(3)],
-    };
+    // never appear.
+    myChatNextResponse = { ok: true, messages: pageOldestFirst(1, 3) };
     render(<ChatTab workspaceId="ws-2" data={minimalData} />);
-    await waitFor(() => expect(myChatActivityCalls.length).toBe(1));
+    await waitFor(() => expect(myChatHistoryCalls.length).toBe(1));
     await waitFor(() => {
       expect(screen.queryByText(/Loading chat history/i)).toBeNull();
     });
@@ -182,15 +185,15 @@ describe("ChatTab lazy history pagination", () => {
   });
 
   it("renders all messages when initial fetch returns exactly the limit", async () => {
-    // 10 == limit → server might have more older rows; sentinel SHOULD
-    // mount so the IO observer can fire loadOlder() on scroll-up. We
-    // verify by checking the rendered bubble count — if hasMore stayed
-    // true the sentinel render path doesn't crash and all 10 rows
-    // produced their pair of bubbles.
-    const fullPage = Array.from({ length: 10 }, (_, i) => makeActivityRow(i + 1));
-    myChatNextResponse = { ok: true, rows: fullPage };
+    // 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(myChatActivityCalls.length).toBe(1));
+    await waitFor(() => expect(myChatHistoryCalls.length).toBe(1));
     await waitFor(() => {
       expect(screen.queryByText(/Loading chat history/i)).toBeNull();
     });
@@ -202,54 +205,67 @@ describe("ChatTab lazy history pagination", () => {
     myChatNextResponse = { ok: false, err: new Error("network down") };
     render(<ChatTab workspaceId="ws-4" data={minimalData} />);
     const retry = await screen.findByText(/Retry/);
-    myChatNextResponse = { ok: true, rows: [makeActivityRow(1)] };
+    myChatNextResponse = { ok: true, messages: makeMessagePair(1) };
     fireEvent.click(retry);
-    await waitFor(() => expect(myChatActivityCalls.length).toBe(2));
-    const retryUrl = myChatActivityCalls[1];
+    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 rows in newest-first order (seq 10..1). After
-    // the component reverses to oldest-first for display, messages[0]
-    // is built from seq=1 — the oldest — and its timestamp is what
-    // before_ts should carry.
-    myChatNextResponse = { ok: true, rows: newestFirstPage(1, 10) };
+    // 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(myChatActivityCalls.length).toBe(1));
+    await waitFor(() => expect(myChatHistoryCalls.length).toBe(1));
     await waitFor(() => expect(ioInstances.length).toBeGreaterThan(0));
 
-    // Stage the older-batch response, then fire the IO callback.
-    myChatNextResponse = { ok: true, rows: newestFirstPage(0, 1) };
+    // Stage older-batch response, then fire IO callback.
+    myChatNextResponse = {
+      ok: true,
+      messages: pageOldestFirst(0, 1),
+      reachedEnd: true,
+    };
     triggerIntersection();
 
-    await waitFor(() => expect(myChatActivityCalls.length).toBe(2));
-    const olderUrl = myChatActivityCalls[1];
+    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, rows: newestFirstPage(1, 10) };
+    myChatNextResponse = {
+      ok: true,
+      messages: pageOldestFirst(1, 10),
+      reachedEnd: false,
+    };
     render(<ChatTab workspaceId="ws-inflight" data={minimalData} />);
-    await waitFor(() => expect(myChatActivityCalls.length).toBe(1));
+    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!: (rows: unknown[]) => void;
-    const deferred = new Promise<unknown[]>((res) => {
+    let release!: (resp: unknown) => void;
+    const deferred = new Promise<unknown>((res) => {
       release = res;
     });
     apiGet.mockImplementationOnce((path: string): Promise<unknown> => {
-      myChatActivityCalls.push(path);
+      myChatHistoryCalls.push(path);
       return deferred;
     });
 
     triggerIntersection(); // start loadOlder #1
-    await waitFor(() => expect(myChatActivityCalls.length).toBe(2));
+    await waitFor(() => expect(myChatHistoryCalls.length).toBe(2));
 
     // Second IO trigger lands while #1 is still pending.
     triggerIntersection();
@@ -258,79 +274,62 @@ describe("ChatTab lazy history pagination", () => {
     // 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(myChatActivityCalls.length).toBe(2);
+    expect(myChatHistoryCalls.length).toBe(2);
 
-    // Release the first fetch. Inflight clears in the finally block;
-    // a subsequent IO trigger is permitted again (verified by checking
-    // we can fire a follow-up after release without hanging the test).
-    release([]);
-    await waitFor(() => expect(myChatActivityCalls.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 () => {
-    // The bug we're pinning: if loadOlder returns 0 rows, the
-    // scrollAnchorRef must be cleared so the next paint doesn't try to
-    // restore against a no-op prepend (which would fight the natural
-    // bottom-pin for any subsequent live message). hasMore flipping to
-    // false is the same flag-flip path; sentinel disappearing is the
-    // observable proxy.
-    myChatNextResponse = { ok: true, rows: newestFirstPage(1, 10) };
+    myChatNextResponse = {
+      ok: true,
+      messages: pageOldestFirst(1, 10),
+      reachedEnd: false,
+    };
     render(<ChatTab workspaceId="ws-anchor" data={minimalData} />);
-    await waitFor(() => expect(myChatActivityCalls.length).toBe(1));
+    await waitFor(() => expect(myChatHistoryCalls.length).toBe(1));
     await waitFor(() => expect(ioInstances.length).toBeGreaterThan(0));
 
-    myChatNextResponse = { ok: true, rows: [] }; // empty → reachedEnd
+    myChatNextResponse = {
+      ok: true,
+      messages: [],
+      reachedEnd: true,
+    };
     triggerIntersection();
-    await waitFor(() => expect(myChatActivityCalls.length).toBe(2));
+    await waitFor(() => expect(myChatHistoryCalls.length).toBe(2));
 
-    // After reachedEnd the sentinel unmounts (hasMore=false). We can't
-    // peek scrollAnchorRef directly, but we can assert the consequence:
-    // scrollIntoView (the bottom-pin for live appends) is not blocked
-    // by a stale anchor. Trigger a re-render via an unrelated state
-    // change… in practice the safest assertion here is that the
-    // sentinel disappeared (proving the empty response propagated to
-    // hasMore correctly, which is the same flag-flip path as anchor
-    // clearing).
     await waitFor(() => {
       expect(screen.queryByText(/Loading older messages/i)).toBeNull();
     });
   });
 
   it("IntersectionObserver does not churn when older messages prepend", async () => {
-    // Whole-PR perf invariant: prepending older history (the load-bearing
-    // user gesture) must NOT tear down + re-arm the IO observer.
-    // Triggering loadOlder is the cleanest way to drive a messages
-    // mutation from inside the test, since live agent push goes through
-    // a Zustand store that's harder to drive reliably from jsdom.
-    //
-    // Pre-fix, loadOlder depended on `messages`, so every prepend
-    // recreated loadOlder → re-ran the IO effect → new observer. Each
-    // call to triggerIntersection() produced a fresh disconnected
-    // observer + a new live one. Post-fix, the observer survives.
-    myChatNextResponse = { ok: true, rows: newestFirstPage(1, 10) };
+    myChatNextResponse = {
+      ok: true,
+      messages: pageOldestFirst(1, 10),
+      reachedEnd: false,
+    };
     render(<ChatTab workspaceId="ws-stable-io" data={minimalData} />);
-    await waitFor(() => expect(myChatActivityCalls.length).toBe(1));
+    await waitFor(() => expect(myChatHistoryCalls.length).toBe(1));
     await waitFor(() => expect(ioInstances.length).toBeGreaterThan(0));
 
-    // Snapshot the observer instance after first paint stabilises.
     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 rows) so reachedEnd stays false and the
-    // sentinel keeps mounting. Pre-fix, each prepend mutated `messages`
-    // → recreated loadOlder → re-ran the IO effect → new observer.
+    // 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,
-        rows: newestFirstPage(-(batch + 1) * 20, 20),
+        messages: pageOldestFirst(-(batch + 1) * 20, 20),
+        reachedEnd: false,
       };
-      const callsBefore = myChatActivityCalls.length;
+      const callsBefore = myChatHistoryCalls.length;
       triggerIntersection();
-      await waitFor(() =>
-        expect(myChatActivityCalls.length).toBe(callsBefore + 1),
-      );
+      await waitFor(() => expect(myChatHistoryCalls.length).toBe(callsBefore + 1));
     }
 
     // The original observer is still the live one — no churn.

canvas/src/components/tabs/chat/AttachmentAudio.tsx

@@ -9,6 +9,7 @@
 // AttachmentLightbox).
 
 import { useState, useEffect, useRef } from "react";
+import { platformAuthHeaders } from "@/lib/api";
 import type { ChatAttachment } from "./types";
 import { isPlatformAttachment, resolveAttachmentHref } from "./uploads";
 import { AttachmentChip } from "./AttachmentViews";
@@ -43,13 +44,8 @@ export function AttachmentAudio({ workspaceId, attachment, onDownload, tone }: P
     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,
+          headers: platformAuthHeaders(),
           credentials: "include",
           signal: AbortSignal.timeout(60_000),
         });
@@ -116,9 +112,5 @@ export function AttachmentAudio({ workspaceId, attachment, onDownload, tone }: P
   );
 }
 
-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;
-}
+// Local getTenantSlug() removed — auth-header construction now goes
+// through platformAuthHeaders() from @/lib/api (#178).

canvas/src/components/tabs/chat/AttachmentImage.tsx

@@ -35,6 +35,7 @@
 //   downscale via canvas, but defer that to v2.
 
 import { useState, useEffect, useRef } from "react";
+import { platformAuthHeaders } from "@/lib/api";
 import type { ChatAttachment } from "./types";
 import { isPlatformAttachment, resolveAttachmentHref } from "./uploads";
 import { AttachmentLightbox } from "./AttachmentLightbox";
@@ -75,22 +76,14 @@ export function AttachmentImage({ workspaceId, attachment, onDownload, tone }: P
     }
 
     // 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.
+    // the blob (don't trigger a Save-As). Auth headers come from the
+    // shared `platformAuthHeaders()` helper — one source of truth for
+    // every authenticated raw fetch in the canvas (#178).
     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,
+          headers: platformAuthHeaders(),
           credentials: "include",
           signal: AbortSignal.timeout(30_000),
         });
@@ -184,15 +177,7 @@ export function AttachmentImage({ workspaceId, attachment, onDownload, tone }: P
   );
 }
 
-// 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;
-}
+// Local getTenantSlug() removed — auth-header construction now goes
+// through platformAuthHeaders() from @/lib/api which uses the canonical
+// getTenantSlug() from @/lib/tenant. This eliminates the duplicate
+// hostname-regex + the duplicate bearer-token-attach pattern (#178).

canvas/src/components/tabs/chat/AttachmentPDF.tsx

@@ -33,6 +33,7 @@
 //     timeout, swap to chip. Implemented as a 3-second watchdog.
 
 import { useState, useEffect, useRef } from "react";
+import { platformAuthHeaders } from "@/lib/api";
 import type { ChatAttachment } from "./types";
 import { isPlatformAttachment, resolveAttachmentHref } from "./uploads";
 import { AttachmentLightbox } from "./AttachmentLightbox";
@@ -69,13 +70,8 @@ export function AttachmentPDF({ workspaceId, attachment, onDownload, tone }: Pro
     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,
+          headers: platformAuthHeaders(),
           credentials: "include",
           signal: AbortSignal.timeout(60_000),
         });
@@ -189,9 +185,5 @@ function PdfGlyph() {
   );
 }
 
-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;
-}
+// Local getTenantSlug() removed — auth-header construction now goes
+// through platformAuthHeaders() from @/lib/api (#178).

canvas/src/components/tabs/chat/AttachmentTextPreview.tsx

@@ -26,6 +26,7 @@
 // to download the full file.
 
 import { useState, useEffect } from "react";
+import { platformAuthHeaders } from "@/lib/api";
 import type { ChatAttachment } from "./types";
 import { isPlatformAttachment, resolveAttachmentHref } from "./uploads";
 import { AttachmentChip } from "./AttachmentViews";
@@ -57,13 +58,13 @@ export function AttachmentTextPreview({ workspaceId, attachment, onDownload, ton
     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;
-        }
+        // Only attach platform auth headers for in-platform URIs —
+        // off-platform URLs (HTTP/HTTPS attachments) MUST NOT receive
+        // our bearer token (it would leak the admin token to a third
+        // party). The branch is preserved with the new shared helper.
+        const headers: Record<string, string> = isPlatformAttachment(attachment.uri)
+          ? platformAuthHeaders()
+          : {};
         const res = await fetch(href, {
           headers,
           credentials: "include",
@@ -182,9 +183,5 @@ export function AttachmentTextPreview({ workspaceId, attachment, onDownload, ton
   );
 }
 
-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;
-}
+// Local getTenantSlug() removed — auth-header construction now goes
+// through platformAuthHeaders() from @/lib/api (#178).

canvas/src/components/tabs/chat/AttachmentVideo.tsx

@@ -25,6 +25,7 @@
 // fetch via service worker. v2 if measured-needed.
 
 import { useState, useEffect, useRef } from "react";
+import { platformAuthHeaders } from "@/lib/api";
 import type { ChatAttachment } from "./types";
 import { isPlatformAttachment, resolveAttachmentHref } from "./uploads";
 import { AttachmentChip } from "./AttachmentViews";
@@ -61,13 +62,8 @@ export function AttachmentVideo({ workspaceId, attachment, onDownload, tone }: P
     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,
+          headers: platformAuthHeaders(),
           credentials: "include",
           // Videos are larger than images on average; give the request
           // more headroom. The server's per-request body cap (50MB) is
@@ -147,11 +143,5 @@ export function AttachmentVideo({ workspaceId, attachment, onDownload, tone }: P
   );
 }
 
-// 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;
-}
+// Local getTenantSlug() removed — auth-header construction now goes
+// through platformAuthHeaders() from @/lib/api (#178).

canvas/src/components/tabs/chat/uploads.ts

@@ -1,12 +1,16 @@
-import { PLATFORM_URL } from "@/lib/api";
-import { getTenantSlug } from "@/lib/tenant";
+import { PLATFORM_URL, platformAuthHeaders } from "@/lib/api";
 import type { ChatAttachment } from "./types";
 
 /** Chat attachments are intentionally uploaded via a direct fetch()
  *  instead of the `api.post` helper — `api.post` JSON-stringifies the
- *  body, which would 500 on a Blob. Mirrors the header plumbing
- *  (tenant slug, admin token, credentials) so SaaS + self-hosted
- *  callers work the same way. */
+ *  body, which would 500 on a Blob. Auth headers (tenant slug, admin
+ *  token, credentials) come from `platformAuthHeaders()` — the same
+ *  helper `request()` uses, so a missing bearer surfaces as a single
+ *  fix site instead of N copies. We deliberately do NOT set
+ *  Content-Type so the browser writes the multipart boundary into the
+ *  header; setting it manually would yield a multipart body the server
+ *  can't parse. See lib/api.ts platformAuthHeaders() for the full
+ *  rationale on why this pair must stay matched. */
 export async function uploadChatFiles(
   workspaceId: string,
   files: File[],
@@ -16,18 +20,12 @@ export async function uploadChatFiles(
   const form = new FormData();
   for (const f of files) form.append("files", f, f.name);
 
-  const headers: Record<string, string> = {};
-  const slug = getTenantSlug();
-  if (slug) headers["X-Molecule-Org-Slug"] = slug;
-  const adminToken = process.env.NEXT_PUBLIC_ADMIN_TOKEN;
-  if (adminToken) headers["Authorization"] = `Bearer ${adminToken}`;
-
   // Uploads legitimately take a while on cold cache (tar write +
   // docker cp into the container). 60s is comfortable for the 25MB/
   // 50MB caps the server enforces.
   const res = await fetch(`${PLATFORM_URL}/workspaces/${workspaceId}/chat/uploads`, {
     method: "POST",
-    headers,
+    headers: platformAuthHeaders(),
     body: form,
     credentials: "include",
     signal: AbortSignal.timeout(60_000),
@@ -143,14 +141,8 @@ export async function downloadChatFile(
     return;
   }
 
-  const headers: Record<string, string> = {};
-  const slug = getTenantSlug();
-  if (slug) headers["X-Molecule-Org-Slug"] = slug;
-  const adminToken = process.env.NEXT_PUBLIC_ADMIN_TOKEN;
-  if (adminToken) headers["Authorization"] = `Bearer ${adminToken}`;
-
   const res = await fetch(href, {
-    headers,
+    headers: platformAuthHeaders(),
     credentials: "include",
     signal: AbortSignal.timeout(60_000),
   });

canvas/src/lib/__tests__/platform-auth-headers.test.ts

@@ -0,0 +1,97 @@
+// @vitest-environment jsdom
+import { describe, it, expect, beforeEach, afterEach, vi } from "vitest";
+
+// Tests for platformAuthHeaders — the shared helper extracted in #178
+// to consolidate the bearer-token-attach + tenant-slug-attach pattern
+// that was previously duplicated across 7 raw-fetch callsites in the
+// canvas (uploads + 5 Attachment* components + the api.ts request()
+// function).
+//
+// What we pin here:
+//  - Returns a fresh object each call (so callers can mutate without
+//    leaking into each other).
+//  - Empty result on a non-tenant host with no admin token (the
+//    localhost / self-hosted shape).
+//  - Bearer attached when NEXT_PUBLIC_ADMIN_TOKEN is set.
+//  - X-Molecule-Org-Slug attached when window.location.hostname is a
+//    tenant subdomain (<slug>.moleculesai.app).
+//  - Both attached when both apply (the production SaaS shape).
+//
+// Why jsdom: getTenantSlug() reads window.location.hostname. Node-only
+// environment yields no window and getTenantSlug returns null
+// unconditionally — wouldn't exercise the slug branch.
+
+import { platformAuthHeaders } from "../api";
+
+describe("platformAuthHeaders", () => {
+  let originalAdminToken: string | undefined;
+
+  beforeEach(() => {
+    originalAdminToken = process.env.NEXT_PUBLIC_ADMIN_TOKEN;
+    delete process.env.NEXT_PUBLIC_ADMIN_TOKEN;
+  });
+
+  afterEach(() => {
+    if (originalAdminToken === undefined) delete process.env.NEXT_PUBLIC_ADMIN_TOKEN;
+    else process.env.NEXT_PUBLIC_ADMIN_TOKEN = originalAdminToken;
+    // jsdom resets hostname between tests via the @vitest-environment
+    // pragma's per-test isolation. No explicit reset needed.
+  });
+
+  it("returns an empty object on a non-tenant host with no admin token", () => {
+    // jsdom default hostname is "localhost" — not a tenant slug, so
+    // getTenantSlug() returns null and no X-Molecule-Org-Slug is added.
+    const headers = platformAuthHeaders();
+    expect(headers).toEqual({});
+  });
+
+  it("attaches Authorization when NEXT_PUBLIC_ADMIN_TOKEN is set", () => {
+    process.env.NEXT_PUBLIC_ADMIN_TOKEN = "local-dev-admin";
+    const headers = platformAuthHeaders();
+    expect(headers).toEqual({ Authorization: "Bearer local-dev-admin" });
+  });
+
+  it("does NOT attach Authorization when NEXT_PUBLIC_ADMIN_TOKEN is empty string", () => {
+    // Empty-string env is the JS-side shape of `KEY=` in .env.
+    // Treating it as unset matches the matched-pair guard in
+    // next.config.ts (admin-token-pair.test.ts) — symmetric semantics.
+    process.env.NEXT_PUBLIC_ADMIN_TOKEN = "";
+    const headers = platformAuthHeaders();
+    expect(headers).toEqual({});
+  });
+
+  it("attaches X-Molecule-Org-Slug on a tenant subdomain", () => {
+    Object.defineProperty(window, "location", {
+      value: { hostname: "reno-stars.moleculesai.app" },
+      writable: true,
+    });
+    const headers = platformAuthHeaders();
+    expect(headers).toEqual({ "X-Molecule-Org-Slug": "reno-stars" });
+  });
+
+  it("attaches both when both apply (production SaaS shape)", () => {
+    Object.defineProperty(window, "location", {
+      value: { hostname: "reno-stars.moleculesai.app" },
+      writable: true,
+    });
+    process.env.NEXT_PUBLIC_ADMIN_TOKEN = "tenant-bearer";
+    const headers = platformAuthHeaders();
+    // Pin exact-equality on the full shape — substring/contains
+    // assertions would also pass for an extra-header bug.
+    expect(headers).toEqual({
+      "X-Molecule-Org-Slug": "reno-stars",
+      Authorization: "Bearer tenant-bearer",
+    });
+  });
+
+  it("returns a fresh object each call (callers can mutate safely)", () => {
+    process.env.NEXT_PUBLIC_ADMIN_TOKEN = "tok";
+    const a = platformAuthHeaders();
+    const b = platformAuthHeaders();
+    expect(a).not.toBe(b); // distinct refs
+    expect(a).toEqual(b); // same content
+    a["Content-Type"] = "application/json";
+    // Mutation on `a` does not leak into `b`.
+    expect(b["Content-Type"]).toBeUndefined();
+  });
+});

canvas/src/lib/api.ts

@@ -21,6 +21,45 @@ export interface RequestOptions {
   timeoutMs?: number;
 }
 
+/**
+ * Build the platform auth header set used by every authenticated fetch
+ * from the canvas. Returns a fresh object so callers can mutate (e.g.
+ * append `Content-Type` for JSON requests, omit it for FormData).
+ *
+ * SaaS cross-origin shape:
+ *  - `X-Molecule-Org-Slug` — derived from `window.location.hostname`
+ *    by `getTenantSlug()`. Control plane uses it for fly-replay
+ *    routing. Empty on localhost / non-tenant hosts — safe to omit.
+ *  - `Authorization: Bearer <token>` — `NEXT_PUBLIC_ADMIN_TOKEN` baked
+ *    into the canvas build (see canvas/Dockerfile L8/L11). Required by
+ *    the workspace-server when `ADMIN_TOKEN` is set on the server side
+ *    (Tier-2b AdminAuth gate, wsauth_middleware.go ~L245). Empty when
+ *    no admin token was provisioned — the Tier-1 session-cookie path
+ *    handles that case via `credentials:"include"`.
+ *
+ * Why a shared helper: the two-line "read env, attach bearer; read
+ * slug, attach header" pattern was duplicated across `request()` and
+ * 7 raw-fetch callsites (chat uploads/download + 5 Attachment*
+ * components) before this consolidation. A new poller or raw fetch
+ * that forgets one of the two headers silently 401s against
+ * workspace-server when ADMIN_TOKEN is set — the exact bug shape
+ * called out in #178 / closes the post-#176 self-review gap.
+ *
+ * Callers that want JSON Content-Type should spread this and add it
+ * themselves; FormData callers should NOT add Content-Type (the
+ * browser sets the multipart boundary). Centralizing the auth pair
+ * but leaving Content-Type up to the caller is the minimum viable
+ * shared shape.
+ */
+export function platformAuthHeaders(): Record<string, string> {
+  const headers: Record<string, string> = {};
+  const slug = getTenantSlug();
+  if (slug) headers["X-Molecule-Org-Slug"] = slug;
+  const adminToken = process.env.NEXT_PUBLIC_ADMIN_TOKEN;
+  if (adminToken) headers["Authorization"] = `Bearer ${adminToken}`;
+  return headers;
+}
+
 async function request<T>(
   method: string,
   path: string,
@@ -28,17 +67,16 @@ async function request<T>(
   retryCount = 0,
   options?: RequestOptions,
 ): Promise<T> {
-  // SaaS cross-origin shape:
-  //  - X-Molecule-Org-Slug: derived from window.location.hostname by
-  //    getTenantSlug(). Control plane uses it for fly-replay routing.
-  //    Empty on localhost / non-tenant hosts — safe to omit.
-  //  - credentials:"include": sends the session cookie cross-origin.
-  //    Cookie's Domain=.moleculesai.app attribute + cp's CORS allow this.
-  const headers: Record<string, string> = { "Content-Type": "application/json" };
+  // JSON-bodied request — Content-Type is JSON. Auth pair comes from
+  // the shared helper; see its doc comment for the SaaS-shape rationale.
+  const headers: Record<string, string> = {
+    "Content-Type": "application/json",
+    ...platformAuthHeaders(),
+  };
+  // Re-read slug locally for the 401 handler below — `headers` already
+  // has it, but the 401 branch needs the bare value to gate the
+  // session-probe + redirect logic on tenant context.
   const slug = getTenantSlug();
-  if (slug) headers["X-Molecule-Org-Slug"] = slug;
-  const adminToken = process.env.NEXT_PUBLIC_ADMIN_TOKEN;
-  if (adminToken) headers["Authorization"] = `Bearer ${adminToken}`;
 
   const res = await fetch(`${PLATFORM_URL}${path}`, {
     method,

canvas/vitest.config.ts

@@ -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

docker-compose.dev.yml

@@ -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:

docker-compose.yml

@@ -13,6 +13,7 @@ services:
       - pgdata:/var/lib/postgresql/data
     networks:
       - molecule-monorepo-net
+    restart: unless-stopped
     healthcheck:
       test: ["CMD-SHELL", "pg_isready -U ${POSTGRES_USER:-dev}"]
       interval: 2s
@@ -50,6 +51,7 @@ services:
       - redisdata:/data
     networks:
       - molecule-monorepo-net
+    restart: unless-stopped
     healthcheck:
       test: ["CMD", "redis-cli", "ping"]
       interval: 2s
@@ -126,6 +128,10 @@ services:
       REDIS_URL: redis://redis:6379
       PORT: "${PLATFORM_PORT:-8080}"
       PLATFORM_URL: "http://platform:${PLATFORM_PORT:-8080}"
+      # Container network namespace is already isolated; "all interfaces"
+      # inside the container = the bridge interface only. The fail-open
+      # default (127.0.0.1) would block host-to-container access.
+      BIND_ADDR: "${BIND_ADDR:-0.0.0.0}"
       # Default MOLECULE_ENV=development so the WorkspaceAuth / AdminAuth
       # middleware fail-open path activates when ADMIN_TOKEN is unset —
       # otherwise the canvas (which runs without a bearer in pure local
@@ -195,12 +201,28 @@ services:
       # App private key — read-only bind-mount. The host-side path is
       # gitignored per .gitignore rules (/.secrets/ + *.pem).
       - ./.secrets/github-app.pem:/secrets/github-app.pem:ro
+      # Per-role persona credentials (molecule-core#242 local surface).
+      # Sourced at workspace creation time by org_import.go::loadPersonaEnvFile
+      # when a workspace.yaml carries `role: <name>`. The host-side dir is
+      # populated by the operator-host bootstrap kit (28 dev-tree personas);
+      # /etc/molecule-bootstrap/personas is the in-container path the
+      # platform expects (matches the prod tenant-EC2 path so the same code
+      # works in both modes).
+      #
+      # Read-only mount — workspace-server only reads, never writes here.
+      # If the host dir is empty/missing the platform's loadPersonaEnvFile
+      # silently no-ops per its existing semantics, so this mount is safe
+      # even on a fresh machine that hasn't run the bootstrap kit yet.
+      - ${MOLECULE_PERSONA_ROOT_HOST:-${HOME}/.molecule-ai/personas}:/etc/molecule-bootstrap/personas:ro
     ports:
       - "${PLATFORM_PUBLISH_PORT:-8080}:${PLATFORM_PORT:-8080}"
     networks:
       - molecule-monorepo-net
+    restart: unless-stopped
     healthcheck:
-      test: ["CMD-SHELL", "wget --no-verbose --tries=1 --spider http://localhost:${PLATFORM_PORT:-8080}/health || exit 1"]
+      # Plain GET — `--spider` would issue HEAD, which returns 404 because
+      # /health is registered as GET only.
+      test: ["CMD-SHELL", "wget -qO /dev/null --tries=1 http://localhost:${PLATFORM_PORT:-8080}/health || exit 1"]
       interval: 5s
       timeout: 5s
       retries: 10
@@ -212,8 +234,8 @@ 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
@@ -238,7 +260,7 @@ services:
     networks:
       - molecule-monorepo-net
     healthcheck:
-      test: ["CMD-SHELL", "wget --no-verbose --tries=1 --spider http://127.0.0.1:${CANVAS_PORT:-3000} || exit 1"]
+      test: ["CMD-SHELL", "wget -qO /dev/null --tries=1 http://127.0.0.1:${CANVAS_PORT:-3000} || exit 1"]
       interval: 10s
       timeout: 5s
       retries: 10

docs/architecture/backends.md

@@ -2,7 +2,7 @@
 
 **Status:** living document — update when you ship a feature that touches one backend.
 **Owner:** workspace-server + controlplane teams.
-**Last audit:** 2026-05-05 (Claude agent — `provisionWorkspaceAuto` / `StopWorkspaceAuto` / `HasProvisioner` SoT pattern landed in PRs #2811 + #2824).
+**Last audit:** 2026-05-07 (plugin install/uninstall closed for EC2 backend via EIC SSH push to the bind-mounted `/configs/plugins/<name>/`, mirroring the Files API PR #1702 pattern).
 
 ## Why this exists
 
@@ -54,7 +54,7 @@ For "do we have any backend?", use `HasProvisioner()`, never bare `h.provisioner
 | **Files API** | | | | |
 | List / Read / Write / Replace / Delete | `container_files.go`, `template_import.go` | `docker exec` + tar `CopyToContainer` | SSH via EIC tunnel (PR #1702) | ✅ parity as of 2026-04-22 (previously docker-only) |
 | **Plugins** | | | | |
-| Install / uninstall / list | `plugins_install.go` | `deliverToContainer()` + volume rm | **gap — no live plugin delivery** | 🔴 **docker-only** |
+| Install / uninstall / list | `plugins_install.go` + `plugins_install_eic.go` | `deliverToContainer()` → exec+`CopyToContainer` on local container | `instance_id` set → EIC SSH push of the staged tarball into the EC2's bind-mounted `/configs/plugins/<name>/` (per `workspaceFilePathPrefix`), `chown 1000:1000`, restart | ✅ parity |
 | **Terminal (WebSocket)** | | | | |
 | Dispatch | `terminal.go:90-105` | `instance_id=""` → `handleLocalConnect` → `docker attach` | `instance_id` set → `handleRemoteConnect` → EIC SSH + `docker exec` | ✅ parity (different implementations, same UX) |
 | **A2A proxy** | | | | |

docs/architecture/canary-release.md

@@ -4,7 +4,7 @@ How a workspace-server code change reaches the prod tenant fleet — and how to
 
 > **⚠️ State note (2026-04-22):** this doc describes the **intended design**. As of this write, the canary fleet described below is **not actually running** — no canary tenants are provisioned, `CANARY_TENANT_URLS` / `CANARY_ADMIN_TOKENS` / `CANARY_CP_SHARED_SECRET` are empty in repo secrets, and `canary-verify.yml` fails every run.
 >
-> Current merges gate on manual `promote-latest.yml` dispatches, not canary. See [molecule-controlplane/docs/canary-tenants.md](https://github.com/Molecule-AI/molecule-controlplane/blob/main/docs/canary-tenants.md) for the Phase 1 code work that's already shipped + the Phase 2 plan for actually standing up the fleet + a "should we even do this now?" decision framework.
+> Current merges gate on manual `promote-latest.yml` dispatches, not canary. See [molecule-controlplane/docs/canary-tenants.md](https://git.moleculesai.app/molecule-ai/molecule-controlplane/src/branch/main/docs/canary-tenants.md) for the Phase 1 code work that's already shipped + the Phase 2 plan for actually standing up the fleet + a "should we even do this now?" decision framework.
 >
 > **Account-specific identifiers (AWS account ID, IAM role name) referenced below in the original design have been redacted from this public doc.** The actual values — if they exist — are in `Molecule-AI/internal/runbooks/canary-fleet.md`. If you're implementing Phase 2, start there.
 >

docs/architecture/molecule-technical-doc.md

@@ -1,7 +1,7 @@
 # Molecule AI — Comprehensive Technical Documentation
 
 > Definitive technical reference for the Molecule AI Agent Team platform.
-> Based on a full non-invasive scan of the [molecule-monorepo](https://github.com/Molecule-AI/molecule-monorepo) repository.
+> Based on a full non-invasive scan of the [molecule-monorepo](https://git.moleculesai.app/molecule-ai/molecule-monorepo) repository.
 
 ---
 
@@ -1149,11 +1149,11 @@ Molecule AI's workspace abstraction is **runtime-agnostic by design**. A workspa
 
 ## Links
 
-- **GitHub**: https://github.com/Molecule-AI/molecule-monorepo
-- **Architecture Docs**: https://github.com/Molecule-AI/molecule-monorepo/tree/main/docs/architecture
-- **API Protocol**: https://github.com/Molecule-AI/molecule-monorepo/tree/main/docs/api-protocol
-- **Agent Runtime**: https://github.com/Molecule-AI/molecule-monorepo/tree/main/docs/agent-runtime
-- **Product Docs**: https://github.com/Molecule-AI/molecule-monorepo/tree/main/docs/product
+- **GitHub**: https://git.moleculesai.app/molecule-ai/molecule-monorepo
+- **Architecture Docs**: https://git.moleculesai.app/molecule-ai/molecule-monorepo/src/branch/main/docs/architecture
+- **API Protocol**: https://git.moleculesai.app/molecule-ai/molecule-monorepo/src/branch/main/docs/api-protocol
+- **Agent Runtime**: https://git.moleculesai.app/molecule-ai/molecule-monorepo/src/branch/main/docs/agent-runtime
+- **Product Docs**: https://git.moleculesai.app/molecule-ai/molecule-monorepo/src/branch/main/docs/product
 
 ---
 

docs/architecture/secrets-key-custody.md

@@ -79,7 +79,7 @@ For SOC2 / ISO 27001 / customer security questionnaires:
 
 ## Pointers
 
-- KMS envelope code: [`molecule-controlplane/internal/crypto/kms.go`](https://github.com/Molecule-AI/molecule-controlplane/blob/main/internal/crypto/kms.go)
-- Static-key fallback: [`molecule-controlplane/internal/crypto/aes.go`](https://github.com/Molecule-AI/molecule-controlplane/blob/main/internal/crypto/aes.go)
+- KMS envelope code: [`molecule-controlplane/internal/crypto/kms.go`](https://git.moleculesai.app/molecule-ai/molecule-controlplane/src/branch/main/internal/crypto/kms.go)
+- Static-key fallback: [`molecule-controlplane/internal/crypto/aes.go`](https://git.moleculesai.app/molecule-ai/molecule-controlplane/src/branch/main/internal/crypto/aes.go)
 - Tenant secrets handler: [`workspace-server/internal/crypto/aes.go`](../../workspace-server/internal/crypto/aes.go)
 - Tenant secrets schema: [database-schema.md](./database-schema.md#workspace_secrets)

docs/assets/branding/molecule-icon.svg

@@ -0,0 +1,28 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 64 64">
+  <style>
+    .bg { fill: #0a1120; }
+    .accent { fill: #7fe8d6; }
+    .accent-stroke { stroke: #7fe8d6; }
+    @media (prefers-color-scheme: light) {
+      .bg { fill: #f5f7fa; }
+      .accent { fill: #1a8a72; }
+      .accent-stroke { stroke: #1a8a72; }
+    }
+  </style>
+  <rect class="bg" width="64" height="64" rx="14"/>
+  <g class="accent-stroke" stroke-width="2.4" stroke-linecap="round" fill="none">
+    <line x1="32" y1="32" x2="12" y2="14"/>
+    <line x1="32" y1="32" x2="52" y2="18"/>
+    <line x1="32" y1="32" x2="10" y2="40"/>
+    <line x1="32" y1="32" x2="54" y2="44"/>
+    <line x1="32" y1="32" x2="32" y2="56"/>
+  </g>
+  <g class="accent">
+    <circle cx="32" cy="32" r="6.5"/>
+    <circle cx="12" cy="14" r="3.5"/>
+    <circle cx="52" cy="18" r="3.5"/>
+    <circle cx="10" cy="40" r="3.5"/>
+    <circle cx="54" cy="44" r="3.5"/>
+    <circle cx="32" cy="56" r="3.5"/>
+  </g>
+</svg>

docs/assets/branding/molecule-logo.svg

@@ -0,0 +1,17 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 64 64" role="img" aria-label="Molecule AI">
+  <g stroke="#7fe8d6" stroke-width="2.6" stroke-linecap="round" fill="none">
+    <line x1="32" y1="32" x2="12" y2="14"/>
+    <line x1="32" y1="32" x2="52" y2="18"/>
+    <line x1="32" y1="32" x2="10" y2="40"/>
+    <line x1="32" y1="32" x2="54" y2="44"/>
+    <line x1="32" y1="32" x2="32" y2="56"/>
+  </g>
+  <g fill="#7fe8d6">
+    <circle cx="32" cy="32" r="7"/>
+    <circle cx="12" cy="14" r="3.6"/>
+    <circle cx="52" cy="18" r="3.6"/>
+    <circle cx="10" cy="40" r="3.6"/>
+    <circle cx="54" cy="44" r="3.6"/>
+    <circle cx="32" cy="56" r="3.6"/>
+  </g>
+</svg>

docs/blog/2026-04-17-deploy-anywhere/index.md

@@ -10,7 +10,7 @@ tags: [platform, fly.io, deployment, infrastructure]
 
 Your infrastructure choice just got decoupled from your agent platform choice. Molecule AI now ships three production-ready workspace backends — `docker`, `flyio`, and `controlplane` — and switching between them takes a single environment variable. Your agent code, model choices, and workspace topology stay exactly the same.
 
-This post covers what shipped in [PR #501](https://github.com/Molecule-AI/molecule-core/pull/501) (Fly Machines provisioner) and [PR #503](https://github.com/Molecule-AI/molecule-core/pull/503) (control plane provisioner), and which backend fits your situation.
+This post covers what shipped in [PR #501](https://git.moleculesai.app/molecule-ai/molecule-core/pull/501) (Fly Machines provisioner) and [PR #503](https://git.moleculesai.app/molecule-ai/molecule-core/pull/503) (control plane provisioner), and which backend fits your situation.
 
 ## Before: One Deployment Model for Every Use Case
 
@@ -107,4 +107,4 @@ No changes to agent code, tool definitions, or orchestration logic. Swap `CONTAI
 
 ---
 
-*[PR #501](https://github.com/Molecule-AI/molecule-core/pull/501) (Fly Machines provisioner) and [PR #503](https://github.com/Molecule-AI/molecule-core/pull/503) (control plane provisioner) are both merged to `main`. Molecule AI is open source — contributions welcome.*
+*[PR #501](https://git.moleculesai.app/molecule-ai/molecule-core/pull/501) (Fly Machines provisioner) and [PR #503](https://git.moleculesai.app/molecule-ai/molecule-core/pull/503) (control plane provisioner) are both merged to `main`. Molecule AI is open source — contributions welcome.*

docs/blog/2026-04-20-chrome-devtools-mcp-seo/index.md

@@ -299,8 +299,8 @@ Or use the Canvas UI: Workspace → Config → MCP Servers → Add browser MCP s
 
 **Try it free** — Molecule AI is open source and self-hostable. Get a workspace running in under 5 minutes.
 
-→ [Get started on GitHub →](https://github.com/Molecule-AI/molecule-core)
+→ [Get started on GitHub →](https://git.moleculesai.app/molecule-ai/molecule-core)
 
 ---
 
-*Have a browser automation use case you want to see covered? Open a discussion on [GitHub Discussions](https://github.com/Molecule-AI/molecule-core/discussions) — or file an issue with the `enhancement` label.*
+*Have a browser automation use case you want to see covered? File an issue with the `enhancement` label on the [molecule-core issue tracker](https://git.moleculesai.app/molecule-ai/molecule-core/issues).*

docs/blog/2026-04-20-remote-workspaces/index.md

@@ -148,7 +148,7 @@ Then follow the [quick-start guide](/docs/guides/remote-workspaces.md).
 Or run the annotated example directly:
 
 ```bash
-git clone https://github.com/Molecule-AI/molecule-sdk-python
+git clone https://git.moleculesai.app/molecule-ai/molecule-sdk-python
 cd molecule-sdk-python/examples/remote-agent
 # Create workspace with runtime:external, grab the ID, then:
 WORKSPACE_ID=<your-id> PLATFORM_URL=https://acme.moleculesai.app python3 run.py
@@ -160,6 +160,6 @@ The agent appears on the canvas within seconds.
 
 → [Remote Workspaces Guide →](/docs/guides/remote-workspaces.md)
 → [External Agent Registration Reference →](/docs/guides/external-agent-registration.md)
-→ [molecule-sdk-python →](https://github.com/Molecule-AI/molecule-sdk-python)
+→ [molecule-sdk-python →](https://git.moleculesai.app/molecule-ai/molecule-sdk-python)
 
 *Phase 30 shipped in PRs #1075–#1083 and #1085–#1100 on `molecule-core`.*

docs/blog/2026-04-20-secure-by-design/index.md

@@ -27,7 +27,7 @@ The biggest user-facing change: every Molecule AI org can now mint named, revoca
 
 → [User guide: Organization API Keys](/docs/guides/org-api-keys.md)
 → [Architecture: Org API Keys](/docs/architecture/org-api-keys.md)
-→ PRs: [#1105](https://github.com/Molecule-AI/molecule-core/pull/1105), [#1107](https://github.com/Molecule-AI/molecule-core/pull/1107), [#1109](https://github.com/Molecule-AI/molecule-core/pull/1109), [#1110](https://github.com/Molecule-AI/molecule-core/pull/1110)
+→ PRs: [#1105](https://git.moleculesai.app/molecule-ai/molecule-core/pull/1105), [#1107](https://git.moleculesai.app/molecule-ai/molecule-core/pull/1107), [#1109](https://git.moleculesai.app/molecule-ai/molecule-core/pull/1109), [#1110](https://git.moleculesai.app/molecule-ai/molecule-core/pull/1110)
 
 ---
 
@@ -48,7 +48,7 @@ AdminAuth now accepts a session-verification tier that runs **before** the beare
 **Self-hosted / local dev:** `CP_UPSTREAM_URL` is unset → this feature is disabled, behaviour is unchanged.
 
 → [Guide: Same-Origin Canvas Fetches & Session Auth](/docs/guides/same-origin-canvas-fetches.md)
-→ PRs: [#1099](https://github.com/Molecule-AI/molecule-core/pull/1099), [#1100](https://github.com/Molecule-AI/molecule-core/pull/1100)
+→ PRs: [#1099](https://git.moleculesai.app/molecule-ai/molecule-core/pull/1099), [#1100](https://git.moleculesai.app/molecule-ai/molecule-core/pull/1100)
 
 ---
 
@@ -87,7 +87,7 @@ The proxy is **fail-closed**: only an explicit allowlist of paths (`/cp/auth/`,
 This is also the structural fix for the lateral-movement risk that session auth introduced: without the allowlist, a tenant-authed browser user could have proxied `/cp/admin/*` requests upstream and exploited the fact that those endpoints accept WorkOS session cookies. The allowlist makes that impossible by construction.
 
 → [Guide: Same-Origin Canvas Fetches & Session Auth](/docs/guides/same-origin-canvas-fetches.md)
-→ PR: [#1095](https://github.com/Molecule-AI/molecule-core/pull/1095)
+→ PR: [#1095](https://git.moleculesai.app/molecule-ai/molecule-core/pull/1095)
 
 ---
 
@@ -99,7 +99,7 @@ The waitlist itself is a Canvas-administered list with email hashing in audit lo
 
 This is the operational surface that makes the above security work matter: the beta is invitation-only, credentials are scoped, and every admin action is auditable.
 
-→ Control plane PRs [#145](https://github.com/Molecule-AI/molecule-controlplane/pull/145), [#148](https://github.com/Molecule-AI/molecule-controlplane/pull/148), [#150](https://github.com/Molecule-AI/molecule-controlplane/pull/150)
+→ Control plane PRs [#145](https://git.moleculesai.app/molecule-ai/molecule-controlplane/pull/145), [#148](https://git.moleculesai.app/molecule-ai/molecule-controlplane/pull/148), [#150](https://git.moleculesai.app/molecule-ai/molecule-controlplane/pull/150)
 
 ---
 

docs/blog/2026-04-21-discord-adapter/index.md

@@ -12,7 +12,7 @@ Your team is in Discord. Your AI agents are in Molecule AI. Until today, those t
 
 That's now one webhook URL.
 
-Molecule AI workspaces can now connect to Discord. Here's what shipped in [PR #656](https://github.com/Molecule-AI/molecule-core/pull/656).
+Molecule AI workspaces can now connect to Discord. Here's what shipped in [PR #656](https://git.moleculesai.app/molecule-ai/molecule-core/pull/656).
 
 ---
 
@@ -70,7 +70,7 @@ For inbound slash commands, point your Discord app's **Interactions Endpoint URL
 
 ## Security: Webhook Tokens Don't Appear in Logs
 
-Webhook URLs contain a token (`/webhooks/{id}/{token}`). If that token leaks into server logs, it's a rotation event. The Discord adapter is explicit about this: HTTP request errors are logged without the URL, and the adapter returns a generic error message. This was hardened in [PR #659](https://github.com/Molecule-AI/molecule-core/pull/659).
+Webhook URLs contain a token (`/webhooks/{id}/{token}`). If that token leaks into server logs, it's a rotation event. The Discord adapter is explicit about this: HTTP request errors are logged without the URL, and the adapter returns a generic error message. This was hardened in [PR #659](https://git.moleculesai.app/molecule-ai/molecule-core/pull/659).
 
 ---
 
@@ -97,4 +97,4 @@ Documentation: [Social Channels guide](/docs/agent-runtime/social-channels#disco
 
 ---
 
-*Discord adapter shipped in [PR #656](https://github.com/Molecule-AI/molecule-core/pull/656). Security hardening in [PR #659](https://github.com/Molecule-AI/molecule-core/pull/659). Molecule AI is open source — contributions welcome.*
+*Discord adapter shipped in [PR #656](https://git.moleculesai.app/molecule-ai/molecule-core/pull/656). Security hardening in [PR #659](https://git.moleculesai.app/molecule-ai/molecule-core/pull/659). Molecule AI is open source — contributions welcome.*

docs/blog/2026-04-22-a2a-v1-agent-platform/index.md

@@ -133,4 +133,4 @@ With protocol-native A2A, you get:
 
 Molecule AI's external agent registration is production-ready. Documentation is live at [External Agent Registration Guide](https://docs.molecule.ai/docs/guides/external-agent-registration). The npm package for the MCP server is available at [`@molecule-ai/mcp-server`](https://www.npmjs.com/package/@molecule-ai/mcp-server).
 
-Read the full [A2A v1.0 protocol spec](https://github.com/Molecule-AI/molecule-core/blob/main/docs/api-protocol/a2a-protocol.md) on GitHub.
+Read the full [A2A v1.0 protocol spec](https://git.moleculesai.app/molecule-ai/molecule-core/src/branch/main/docs/api-protocol/a2a-protocol.md) on GitHub.

docs/blog/2026-04-22-remote-workspaces/index.md

@@ -45,7 +45,7 @@ canonicalUrl: "https://docs.molecule.ai/blog/remote-workspaces"
   " proficiencyLevel": "Expert",
   "genre": ["technical documentation", "product announcement"],
   "sameAs": [
-    "https://github.com/Molecule-AI/molecule-core",
+    "https://git.moleculesai.app/molecule-ai/molecule-core",
     "https://molecule.ai"
   ]
 }
@@ -270,7 +270,7 @@ Configure it in your project's `.mcp.json` and any AI agent (Claude Code, Cursor
 
 → [External Agent Registration Guide](/docs/guides/external-agent-registration) — full step-by-step with Python and Node.js reference implementations
 
-→ [GitHub: molecule-core](https://github.com/Molecule-AI/molecule-core) — source and issues
+→ [GitHub: molecule-core](https://git.moleculesai.app/molecule-ai/molecule-core) — source and issues
 
 → [Phase 30 Launch Thread on X](https://x.com) — follow for updates
 

docs/blog/a2a-v1-production-reference-2026-04-24.md

@@ -170,4 +170,4 @@ The `staging` branch is now on `a2a-sdk` 1.0.0. The `main` branch still carries
 
 If you're running `a2a-sdk` 0.3.x and planning the 1.0.0 migration, this post is the reference. The four breaking changes are well-contained, the migration is a single PR, and the eight smoke scenarios above will tell you whether the upgrade is clean before you merge.
 
-Questions? The [A2A protocol spec](https://github.com/google-a2a/a2a-specification) is the authoritative source. For Molecule AI's production A2A implementation, see [External Agent Registration](https://docs.molecule.ai/docs/guides/external-agent-registration) or open an issue in the [molecule-core](https://github.com/Molecule-AI/molecule-core) repo.
+Questions? The [A2A protocol spec](https://github.com/google-a2a/a2a-specification) is the authoritative source. For Molecule AI's production A2A implementation, see [External Agent Registration](https://docs.molecule.ai/docs/guides/external-agent-registration) or open an issue in the [molecule-core](https://git.moleculesai.app/molecule-ai/molecule-core) repo.

docs/engineering/postmortem-2026-04-23-boot-event-401.md

@@ -3,8 +3,8 @@
 **Date:** 2026-04-23
 **Severity:** High — every new SaaS tenant blocked
 **Detection path:** E2E Staging SaaS run 24848425822 failed at "tenant provisioning"; investigation of CP Railway logs surfaced the auth mismatch.
-**Status:** Fix pushed on [molecule-controlplane#238](https://github.com/Molecule-AI/molecule-controlplane/pull/238).
-**Related:** [issue #239](https://github.com/Molecule-AI/molecule-controlplane/issues/239) (Cloudflare DNS record quota), [testing-strategy.md](../engineering/testing-strategy.md)
+**Status:** Fix pushed on [molecule-controlplane#238](https://git.moleculesai.app/molecule-ai/molecule-controlplane/pull/238).
+**Related:** [issue #239](https://git.moleculesai.app/molecule-ai/molecule-controlplane/issues/239) (Cloudflare DNS record quota), [testing-strategy.md](../engineering/testing-strategy.md)
 
 ## Summary
 
@@ -35,7 +35,7 @@ The flow was:
 
 ### The commit that introduced the bug
 
-[molecule-controlplane#235](https://github.com/Molecule-AI/molecule-controlplane/pull/235) — "fix(provision): wait for tenant boot-event before falling back to canary". Merged 2026-04-22.
+[molecule-controlplane#235](https://git.moleculesai.app/molecule-ai/molecule-controlplane/pull/235) — "fix(provision): wait for tenant boot-event before falling back to canary". Merged 2026-04-22.
 
 Before #235, readiness was determined via a canary probe through Cloudflare's edge — which didn't need CP-side auth, so the INSERT ordering didn't matter. #235 made boot-events the primary readiness signal but didn't move the INSERT earlier. The race was latent before but became load-bearing after.
 
@@ -90,7 +90,7 @@ bootReady, _ := provisioner.WaitForTenantReady(ctx, h.db, org.ID, 4*time.Minute)
 h.db.ExecContext(ctx, `UPDATE org_instances SET status = 'running' WHERE org_id = $1`, org.ID)
 ```
 
-See [molecule-controlplane#238](https://github.com/Molecule-AI/molecule-controlplane/pull/238) for the full diff.
+See [molecule-controlplane#238](https://git.moleculesai.app/molecule-ai/molecule-controlplane/pull/238) for the full diff.
 
 ## Lessons
 
@@ -122,9 +122,9 @@ Early investigation blamed the hermes provider 401 bug (a separate, known issue
 
 ## Follow-ups
 
-- [ ] Land [molecule-controlplane#238](https://github.com/Molecule-AI/molecule-controlplane/pull/238)
+- [ ] Land [molecule-controlplane#238](https://git.moleculesai.app/molecule-ai/molecule-controlplane/pull/238)
 - [ ] Redeploy staging-api, verify E2E goes green
 - [ ] Add CP integration test suite (see lesson #2)
 - [ ] Wire E2E failure → notification (see lesson #3)
 - [ ] Add invariant comment in `provisionTenant` (see lesson #4)
-- [ ] Cloudflare DNS quota cleanup — [molecule-controlplane#239](https://github.com/Molecule-AI/molecule-controlplane/issues/239)
+- [ ] Cloudflare DNS quota cleanup — [molecule-controlplane#239](https://git.moleculesai.app/molecule-ai/molecule-controlplane/issues/239)

docs/engineering/pr-hygiene.md

@@ -138,5 +138,5 @@ If you see any of these, don't try to "clean it up in place" — **cherry-pick o
 
 ## Related
 
-- [Issue #1822](https://github.com/Molecule-AI/molecule-core/issues/1822) — backend parity drift tracker (example of docs that have to stay current)
+- [Issue #1822](https://git.moleculesai.app/molecule-ai/molecule-core/issues/1822) — backend parity drift tracker (example of docs that have to stay current)
 - [Postmortem: CP boot-event 401](./postmortem-2026-04-23-boot-event-401.md) — caught before shipping because a reviewer could read the diff

docs/engineering/testing-strategy.md

@@ -103,9 +103,9 @@ A bad test:
 
 ## Related
 
-- [Issue #1821](https://github.com/Molecule-AI/molecule-core/issues/1821) — policy tracking issue
-- [Issue #1815](https://github.com/Molecule-AI/molecule-core/issues/1815) — Canvas coverage instrumentation
-- [Issue #1818](https://github.com/Molecule-AI/molecule-core/issues/1818) — Python pytest-cov
-- [Issue #1814](https://github.com/Molecule-AI/molecule-core/issues/1814) — workspace_provision_test.go unblock
-- [Issue #1816](https://github.com/Molecule-AI/molecule-core/issues/1816) — tokens.go coverage
-- [Issue #1819](https://github.com/Molecule-AI/molecule-core/issues/1819) — wsauth_middleware coverage
+- [Issue #1821](https://git.moleculesai.app/molecule-ai/molecule-core/issues/1821) — policy tracking issue
+- [Issue #1815](https://git.moleculesai.app/molecule-ai/molecule-core/issues/1815) — Canvas coverage instrumentation
+- [Issue #1818](https://git.moleculesai.app/molecule-ai/molecule-core/issues/1818) — Python pytest-cov
+- [Issue #1814](https://git.moleculesai.app/molecule-ai/molecule-core/issues/1814) — workspace_provision_test.go unblock
+- [Issue #1816](https://git.moleculesai.app/molecule-ai/molecule-core/issues/1816) — tokens.go coverage
+- [Issue #1819](https://git.moleculesai.app/molecule-ai/molecule-core/issues/1819) — wsauth_middleware coverage

docs/guides/external-workspace-quickstart.md

@@ -153,7 +153,7 @@ The `id` field is your workspace ID — remember it.
 |---|---|
 | "Failed to send message — agent may be unreachable" | The tenant couldn't POST to your URL. Verify `curl https://<your-tunnel>/health` returns 200 from another machine. |
 | Response takes > 30s | Canvas times out around 30s. Keep initial implementations simple. For long-running work, return a placeholder and use [polling mode](#next-step-polling-mode-preview) (once available). |
-| Agent duplicated in chat | Known canvas bug where WebSocket + HTTP responses both render. Fixed in [PR #1517](https://github.com/Molecule-AI/molecule-core/pull/1517). |
+| Agent duplicated in chat | Known canvas bug where WebSocket + HTTP responses both render. Fixed in [PR #1517](https://git.moleculesai.app/molecule-ai/molecule-core/pull/1517). |
 | Agent replies but canvas shows "Agent unreachable" | Check the tenant can reach your URL. Cloudflare quick tunnels rotate — the URL in your canvas may point at a dead tunnel after restart. |
 | Getting 404 when POSTing to tenant | Add `X-Molecule-Org-Id` header. The tenant's security layer 404s unmatched origin requests by design. |
 
@@ -215,7 +215,7 @@ Push mode (this guide) works today but requires an inbound-reachable URL — whi
 
 Your agent makes only outbound HTTPS calls to the platform, pulling messages from an inbox queue and posting replies back. Works behind any NAT/firewall, tolerates offline laptops, no tunnel needed.
 
-See the [design doc](https://github.com/Molecule-AI/internal/blob/main/product/external-workspaces-polling.md) (internal) and [implementation tracking issue](https://github.com/Molecule-AI/molecule-core/issues?q=polling+mode) once opened.
+See the [design doc](https://git.moleculesai.app/molecule-ai/internal/src/branch/main/product/external-workspaces-polling.md) (internal) and the implementation tracking issue (search `polling+mode` on the [molecule-core issue tracker](https://git.moleculesai.app/molecule-ai/molecule-core/issues)).
 
 ---
 
@@ -255,7 +255,7 @@ If all four pass and canvas still shows your agent as unreachable, see the [remo
 ## Feedback
 
 This is a new path. Tell us what broke:
-- Open an issue: https://github.com/Molecule-AI/molecule-core/issues/new?labels=external-workspace
+- Open an issue: https://git.moleculesai.app/molecule-ai/molecule-core/issues/new?labels=external-workspace
 - Join #external-workspaces on our Slack
 - Submit a PR improving this doc if something tripped you up — the faster we can make the quickstart, the more developers we bring in
 

docs/guides/remote-workspaces.md

@@ -143,5 +143,5 @@ The agent appears on the canvas with a **purple REMOTE badge** within seconds. F
 ## Next Steps
 
 - **[External Agent Registration Guide →](/docs/guides/external-agent-registration)** — full endpoint reference, Python + Node.js examples, troubleshooting
-- **[molecule-sdk-python →](https://github.com/Molecule-AI/molecule-sdk-python)** — SDK source, `RemoteAgentClient` API docs
-- **[SDK Examples →](https://github.com/Molecule-AI/molecule-sdk-python/tree/main/examples/remote-agent)** — `run.py` demo script, annotated walkthrough
+- **[molecule-sdk-python →](https://git.moleculesai.app/molecule-ai/molecule-sdk-python)** — SDK source, `RemoteAgentClient` API docs
+- **[SDK Examples →](https://git.moleculesai.app/molecule-ai/molecule-sdk-python/src/branch/main/examples/remote-agent)** — `run.py` demo script, annotated walkthrough

docs/guides/skill-catalog.md

@@ -61,7 +61,7 @@ molecule skills install arxiv-research --from community
 
 Community skills are reviewed by the Molecule AI team before being
 listed. Submit a skill for review by opening a PR against
-[`molecule-ai/skills`](https://github.com/Molecule-AI/skills).
+[`molecule-ai/skills`](https://git.moleculesai.app/molecule-ai/skills).
 
 ## Installing via config.yaml
 
@@ -151,7 +151,7 @@ molecule skills bundle my-custom-skill --output ./org-templates/my-role/
 ```
 
 **Publishing to the community:** Open a PR against
-[`molecule-ai/skills`](https://github.com/Molecule-AI/skills) with a
+[`molecule-ai/skills`](https://git.moleculesai.app/molecule-ai/skills) with a
 complete skill package. Community skills are reviewed for security and
 correctness before listing.
 

docs/integrations/runtime-native-mcp-status.md

@@ -99,10 +99,10 @@ fork needed in production.
   `resolve_platform_id` for plugin-platform-safe deserialization, and
   `self.adapters[adapter.platform]` keying fix (caught by real-subprocess
   test before merge — see below).
-- **Plugin package**: [Molecule-AI/hermes-platform-molecule-a2a](https://github.com/Molecule-AI/hermes-platform-molecule-a2a)
+- **Plugin package**: [Molecule-AI/hermes-platform-molecule-a2a](https://git.moleculesai.app/molecule-ai/hermes-platform-molecule-a2a)
   v0.1.0 — public, MIT-licensed. 11 unit tests + 8 in-process E2E
   + 4 real-subprocess E2E checkpoints all green.
-- **Workspace template patch**: [Molecule-AI/molecule-ai-workspace-template-hermes#32](https://github.com/Molecule-AI/molecule-ai-workspace-template-hermes/pull/32)
+- **Workspace template patch**: [Molecule-AI/molecule-ai-workspace-template-hermes#32](https://git.moleculesai.app/molecule-ai/molecule-ai-workspace-template-hermes/pull/32)
   — Dockerfile installs the patched fork + plugin into the hermes
   installer's venv; start.sh seeds `platforms.molecule-a2a` config
   stanza. Pre-demo deliberately install-only; adapter.py rewrite to
@@ -157,9 +157,9 @@ intermediate shim earns its complexity.
 ## Codex (OpenAI Codex CLI)
 
 **Status:** Template SHIPPED. Repo live at
-[`Molecule-AI/molecule-ai-workspace-template-codex`](https://github.com/Molecule-AI/molecule-ai-workspace-template-codex)
+[`Molecule-AI/molecule-ai-workspace-template-codex`](https://git.moleculesai.app/molecule-ai/molecule-ai-workspace-template-codex)
 (14 files, 1411 LOC, 12/12 tests). molecule-core registration in
-[PR #2512](https://github.com/Molecule-AI/molecule-core/pull/2512).
+[PR #2512](https://git.moleculesai.app/molecule-ai/molecule-core/pull/2512).
 E2E with real A2A traffic remains.
 
 **Path:** Persistent `codex app-server` stdio JSON-RPC client

docs/memory-plugins/CHANGELOG.md

@@ -101,7 +101,7 @@ incident-shaped.
 ## [v1.0.0] — initial release (RFC #2728, PRs #2729-#2742)
 
 Initial plugin contract + 11-PR rollout. See
-[issue #2728](https://github.com/Molecule-AI/molecule-core/issues/2728)
+[issue #2728](https://git.moleculesai.app/molecule-ai/molecule-core/issues/2728)
 for the full RFC.
 
 Endpoints: `/v1/health`, `/v1/namespaces/{name}` (PUT/PATCH/DELETE),

docs/plugins/agentskills-compat.md

@@ -160,11 +160,11 @@ not expose.
 | `molecule-skill-update-docs` | `[claude_code]` | `[claude_code, hermes]` |
 
 Companion PRs:
-- [molecule-ai-plugin-ecc#2](https://github.com/Molecule-AI/molecule-ai-plugin-ecc/pull/2)
-- [molecule-ai-plugin-superpowers#2](https://github.com/Molecule-AI/molecule-ai-plugin-superpowers/pull/2)
-- [molecule-ai-plugin-molecule-dev#2](https://github.com/Molecule-AI/molecule-ai-plugin-molecule-dev/pull/2)
-- [molecule-ai-plugin-molecule-skill-cron-learnings#2](https://github.com/Molecule-AI/molecule-ai-plugin-molecule-skill-cron-learnings/pull/2)
-- [molecule-ai-plugin-molecule-skill-update-docs#2](https://github.com/Molecule-AI/molecule-ai-plugin-molecule-skill-update-docs/pull/2)
+- [molecule-ai-plugin-ecc#2](https://git.moleculesai.app/molecule-ai/molecule-ai-plugin-ecc/pull/2)
+- [molecule-ai-plugin-superpowers#2](https://git.moleculesai.app/molecule-ai/molecule-ai-plugin-superpowers/pull/2)
+- [molecule-ai-plugin-molecule-dev#2](https://git.moleculesai.app/molecule-ai/molecule-ai-plugin-molecule-dev/pull/2)
+- [molecule-ai-plugin-molecule-skill-cron-learnings#2](https://git.moleculesai.app/molecule-ai/molecule-ai-plugin-molecule-skill-cron-learnings/pull/2)
+- [molecule-ai-plugin-molecule-skill-update-docs#2](https://git.moleculesai.app/molecule-ai/molecule-ai-plugin-molecule-skill-update-docs/pull/2)
 
 Security note: Security Auditor was offline at time of change. Self-assessed
 as non-security-impacting — adding `hermes` to a string list in `plugin.yaml`

docs/quickstart.md

@@ -17,7 +17,7 @@ This path is aligned to the current repository and current UI. It gets you from
 ## The one-command path
 
 ```bash
-git clone https://github.com/Molecule-AI/molecule-monorepo.git
+git clone https://git.moleculesai.app/molecule-ai/molecule-monorepo.git
 cd molecule-monorepo
 ./scripts/dev-start.sh
 ```
@@ -42,7 +42,7 @@ If you'd rather run each component yourself — useful when you're iterating on
 ### Step 1: Clone the repository
 
 ```bash
-git clone https://github.com/Molecule-AI/molecule-monorepo.git
+git clone https://git.moleculesai.app/molecule-ai/molecule-monorepo.git
 cd molecule-monorepo
 ```
 

docs/runbooks/handlers-postgres-integration-port-collision.md

@@ -0,0 +1,137 @@
+# Runbook — Handlers Postgres Integration port-collision substrate
+
+**Status:** Resolved 2026-05-08 (PR for class B Hongming-owned CICD red sweep).
+
+## Symptom
+
+`Handlers Postgres Integration` workflow fails on staging push and PRs.
+Step `Apply migrations to Postgres service` shows:
+
+```
+psql: error: connection to server at "127.0.0.1", port 5432 failed: Connection refused
+```
+
+Job-cleanup step further down logs:
+
+```
+Cleaning up services for job Handlers Postgres Integration
+failed to remove container: Error response from daemon: No such container: <id>
+```
+
+…confirming the postgres service container was already gone before
+cleanup ran.
+
+## Root cause
+
+Our Gitea act_runner (operator host `5.78.80.188`,
+`/opt/molecule/runners/config.yaml`) sets:
+
+```yaml
+container:
+  network: host
+```
+
+…which act_runner applies to BOTH the job container AND every
+`services:` container in a workflow. Multiple workflow instances
+running concurrently across the 16 parallel runners each try to bind
+postgres on `0.0.0.0:5432`. The first wins; subsequent instances exit
+immediately with:
+
+```
+LOG:  could not bind IPv4 address "0.0.0.0": Address in use
+HINT: Is another postmaster already running on port 5432?
+FATAL: could not create any TCP/IP sockets
+```
+
+act_runner sets `AutoRemove:true` on service containers, so Docker
+garbage-collects them as soon as they exit. By the time the migrations
+step runs `pg_isready` / `psql`, the container is gone and connection
+refused.
+
+Reproduction (operator host):
+
+```bash
+docker run --rm -d --name pg-A --network host \
+  -e POSTGRES_PASSWORD=test postgres:15-alpine
+docker run -d --name pg-B --network host \
+  -e POSTGRES_PASSWORD=test postgres:15-alpine
+docker logs pg-B   # FATAL: could not create any TCP/IP sockets
+```
+
+## Why per-job override doesn't work
+
+The natural fix — per-job `container.network` override — is silently
+ignored by act_runner. The runner log emits:
+
+```
+--network and --net in the options will be ignored.
+```
+
+This is a documented act_runner constraint: container network is a
+runner-wide setting, not per-job. Source: gitea/act_runner config docs
++ vegardit/docker-gitea-act-runner issue #7.
+
+Flipping the global `container.network` to `bridge` would break every
+other workflow in the repo (cache server discovery,
+`molecule-monorepo-net` peer access during integration tests, etc.) —
+unacceptable blast radius for a per-test bug.
+
+## Fix shape
+
+`handlers-postgres-integration.yml` no longer uses `services: postgres:`.
+It launches a sibling postgres container manually on the existing
+`molecule-monorepo-net` bridge network with a per-run unique name:
+
+```yaml
+env:
+  PG_NAME: pg-handlers-${{ github.run_id }}-${{ github.run_attempt }}
+  PG_NETWORK: molecule-monorepo-net
+
+steps:
+  - name: Start sibling Postgres on bridge network
+    run: |
+      docker run -d --name "${PG_NAME}" --network "${PG_NETWORK}" \
+        ...
+        postgres:15-alpine
+      PG_HOST=$(docker inspect "${PG_NAME}" \
+        --format "{{(index .NetworkSettings.Networks \"${PG_NETWORK}\").IPAddress}}")
+      echo "PG_HOST=${PG_HOST}" >> "$GITHUB_ENV"
+
+  # … migrations + tests use ${PG_HOST}, not 127.0.0.1 …
+
+  - if: always() && …
+    name: Stop sibling Postgres
+    run: docker rm -f "${PG_NAME}" || true
+```
+
+The host-net job container can reach a bridge-net container via the
+bridge IP directly (verified manually, 2026-05-08). Two parallel runs
+use different names + different bridge IPs — no collision.
+
+## Future-proofing
+
+Other workflows that hit the same shape (any `services:` with a
+fixed-port image) will exhibit the same failure mode under
+host-network runner config. Translate using this same pattern:
+
+1. Drop the `services:` block.
+2. Use `${{ github.run_id }}-${{ github.run_attempt }}` for unique
+   container name.
+3. Launch on `molecule-monorepo-net` (already trusted bridge in
+   `docker-compose.infra.yml`).
+4. Read back the bridge IP via `docker inspect` and export as a step env.
+5. `if: always()` cleanup step at the end.
+
+If the count of such workflows grows, factor into a composite action
+(`./.github/actions/sibling-postgres`) so the substrate logic lives
+in one place.
+
+## Related
+
+- Issue #88 (closed by #92): localhost → 127.0.0.1 fix that unmasked
+  this collision; the IPv6 fix is correct, port collision is the new
+  layer.
+- Issue #94 created `molecule-monorepo-net` + `alpine:latest` as
+  prereqs.
+- Saved memory `feedback_act_runner_github_server_url` documents
+  another act_runner-vs-GHA divergence (server URL).

docs/tutorials/chrome-devtools-mcp-quickstart.md

@@ -198,7 +198,7 @@ Lighthouse audit against staging.yourapp.com:
   FCP: 2.4s | LCP: 5.2s | CLS: 0.18 | TBT: 620ms
 
 Performance regression detected — opening GitHub issue.
-Issue: https://github.com/Molecule-AI/molecule-core/issues/1527
+Issue: https://git.moleculesai.app/molecule-ai/molecule-core/issues/1527
 Label: performance-regression | Assignees: @your-team
 ```
 

docs/tutorials/fly-machines-provisioner.md

@@ -85,8 +85,8 @@ Fly Machines start in milliseconds and run in 35+ regions. Provisioning agent wo
 
 ## Related
 
-- PR #501: [feat(platform): Fly Machines provisioner](https://github.com/Molecule-AI/molecule-core/pull/501)
-- PR #481: [feat(ci): deploy to Fly after image push](https://github.com/Molecule-AI/molecule-core/pull/481)
+- PR #501: [feat(platform): Fly Machines provisioner](https://git.moleculesai.app/molecule-ai/molecule-core/pull/501)
+- PR #481: [feat(ci): deploy to Fly after image push](https://git.moleculesai.app/molecule-ai/molecule-core/pull/481)
 - [Fly Machines API docs](https://fly.io/docs/machines/api/)
 - [Platform API reference](../api-reference.md)
-- Issue [#525](https://github.com/Molecule-AI/molecule-core/issues/525)
+- Issue [#525](https://git.moleculesai.app/molecule-ai/molecule-core/issues/525)

docs/tutorials/gemini-cli-runtime.md

@@ -61,6 +61,6 @@ The real power surfaces when you mix runtimes on the same Molecule AI tenant. Yo
 
 ## Related
 
-- PR #379: [feat(adapters): add gemini-cli runtime adapter](https://github.com/Molecule-AI/molecule-core/pull/379)
+- PR #379: [feat(adapters): add gemini-cli runtime adapter](https://git.moleculesai.app/molecule-ai/molecule-core/pull/379)
 - [Multi-provider Hermes docs](../architecture/hermes.md)
 - [Workspace runtimes reference](../reference/runtimes.md)

docs/tutorials/google-adk-runtime.md

@@ -68,7 +68,7 @@ ADK workspaces participate in the same A2A network as Claude Code, Gemini CLI, H
 
 ## Related
 
-- PR #550: [feat(adapters): add google-adk runtime adapter](https://github.com/Molecule-AI/molecule-core/pull/550)
+- PR #550: [feat(adapters): add google-adk runtime adapter](https://git.moleculesai.app/molecule-ai/molecule-core/pull/550)
 - [Google ADK (adk-python)](https://github.com/google/adk-python)
 - [Gemini CLI runtime tutorial](./gemini-cli-runtime.md)
 - [Platform API reference](../api-reference.md)

docs/tutorials/hermes-multi-provider-dispatch.md

@@ -176,9 +176,9 @@ What is on the roadmap for Phase 2d (not yet shipped):
 
 ## Related
 
-- PR #240: [Phase 2a — native Anthropic dispatch](https://github.com/Molecule-AI/molecule-core/pull/240)
-- PR #255: [Phase 2b — native Gemini dispatch](https://github.com/Molecule-AI/molecule-core/pull/255)
-- PR #267: [Phase 2c — multi-turn history on all paths](https://github.com/Molecule-AI/molecule-core/pull/267)
+- PR #240: [Phase 2a — native Anthropic dispatch](https://git.moleculesai.app/molecule-ai/molecule-core/pull/240)
+- PR #255: [Phase 2b — native Gemini dispatch](https://git.moleculesai.app/molecule-ai/molecule-core/pull/255)
+- PR #267: [Phase 2c — multi-turn history on all paths](https://git.moleculesai.app/molecule-ai/molecule-core/pull/267)
 - [Hermes adapter design](../adapters/hermes-adapter-design.md)
 - [Platform API reference](../api-reference.md)
-- Issue [#513](https://github.com/Molecule-AI/molecule-core/issues/513)
+- Issue [#513](https://git.moleculesai.app/molecule-ai/molecule-core/issues/513)

docs/tutorials/lark-feishu-channel.md

@@ -90,6 +90,6 @@ Molecule AI canvas without code changes.
 
 ## Related
 
-- PR #480: [feat(channels): Lark / Feishu channel adapter](https://github.com/Molecule-AI/molecule-core/pull/480)
+- PR #480: [feat(channels): Lark / Feishu channel adapter](https://git.moleculesai.app/molecule-ai/molecule-core/pull/480)
 - [Social channels architecture](../agent-runtime/social-channels.md)
 - [Channel adapter reference](../api-reference.md#channels)

docs/workspace-runtime-package.md

@@ -98,14 +98,14 @@ Each of the 8 adapter template repos contains:
 
 | Adapter | Repo |
 |---------|------|
-| claude-code | https://github.com/Molecule-AI/molecule-ai-workspace-template-claude-code |
-| langgraph | https://github.com/Molecule-AI/molecule-ai-workspace-template-langgraph |
-| crewai | https://github.com/Molecule-AI/molecule-ai-workspace-template-crewai |
-| autogen | https://github.com/Molecule-AI/molecule-ai-workspace-template-autogen |
-| deepagents | https://github.com/Molecule-AI/molecule-ai-workspace-template-deepagents |
-| hermes | https://github.com/Molecule-AI/molecule-ai-workspace-template-hermes |
-| gemini-cli | https://github.com/Molecule-AI/molecule-ai-workspace-template-gemini-cli |
-| openclaw | https://github.com/Molecule-AI/molecule-ai-workspace-template-openclaw |
+| claude-code | https://git.moleculesai.app/molecule-ai/molecule-ai-workspace-template-claude-code |
+| langgraph | https://git.moleculesai.app/molecule-ai/molecule-ai-workspace-template-langgraph |
+| crewai | https://git.moleculesai.app/molecule-ai/molecule-ai-workspace-template-crewai |
+| autogen | https://git.moleculesai.app/molecule-ai/molecule-ai-workspace-template-autogen |
+| deepagents | https://git.moleculesai.app/molecule-ai/molecule-ai-workspace-template-deepagents |
+| hermes | https://git.moleculesai.app/molecule-ai/molecule-ai-workspace-template-hermes |
+| gemini-cli | https://git.moleculesai.app/molecule-ai/molecule-ai-workspace-template-gemini-cli |
+| openclaw | https://git.moleculesai.app/molecule-ai/molecule-ai-workspace-template-openclaw |
 
 ## Adapter discovery (ADAPTER_MODULE)
 
@@ -244,7 +244,7 @@ correctness before pushing a `runtime-v*` tag.
 ## Writing a new adapter
 
 Use the GitHub template repo
-[`Molecule-AI/molecule-ai-workspace-template-starter`](https://github.com/Molecule-AI/molecule-ai-workspace-template-starter)
+[`molecule-ai/molecule-ai-workspace-template-starter`](https://git.moleculesai.app/molecule-ai/molecule-ai-workspace-template-starter) (note: the starter repo did not survive the 2026-05-06 GitHub-org-suspension migration; recreation tracked at internal#41)
 — it ships with the canonical Dockerfile + adapter.py skeleton + config.yaml
 schema + the `repository_dispatch: [runtime-published]` cascade receiver
 already wired up. No follow-up setup PR required.
@@ -256,7 +256,7 @@ gh repo create Molecule-AI/molecule-ai-workspace-template-<runtime> \
   --public \
   --description "Molecule AI workspace template: <runtime>"
 
-git clone https://github.com/Molecule-AI/molecule-ai-workspace-template-<runtime>
+git clone https://git.moleculesai.app/molecule-ai/molecule-ai-workspace-template-<runtime>.git
 cd molecule-ai-workspace-template-<runtime>
 ```
 
@@ -286,7 +286,7 @@ After `git push`:
 If the canonical shape changes (e.g. `config.yaml` schema gets a new field,
 the `BaseAdapter` interface adds a method, the reusable CI workflow
 signature changes), update the
-[starter](https://github.com/Molecule-AI/molecule-ai-workspace-template-starter)
+[starter](https://git.moleculesai.app/molecule-ai/molecule-ai-workspace-template-starter) (recreation pending — see note above)
 **first**. Existing templates can either migrate at their own pace or be
 touched in a coordinated cleanup PR. Either way, future templates pick up
 the new shape from day one.

manifest.json

@@ -1,47 +1,46 @@
 {
-  "_comment": "Pin refs to release tags for reproducible builds. 'main' is OK while all repos are internal.",
+  "_comment": "OSS surface registry — every repo listed here MUST be public on git.moleculesai.app. Layer-3 customer/private templates are NOT registered here; they are handled at provision-time via the per-tenant credential resolver (see internal#102 RFC). 'main' refs are pinned to tags before broad rollout.",
   "version": 1,
   "plugins": [
-    {"name": "browser-automation", "repo": "Molecule-AI/molecule-ai-plugin-browser-automation", "ref": "main"},
-    {"name": "ecc", "repo": "Molecule-AI/molecule-ai-plugin-ecc", "ref": "main"},
-    {"name": "gh-identity", "repo": "Molecule-AI/molecule-ai-plugin-gh-identity", "ref": "main"},
-    {"name": "molecule-audit", "repo": "Molecule-AI/molecule-ai-plugin-molecule-audit", "ref": "main"},
-    {"name": "molecule-audit-trail", "repo": "Molecule-AI/molecule-ai-plugin-molecule-audit-trail", "ref": "main"},
-    {"name": "molecule-careful-bash", "repo": "Molecule-AI/molecule-ai-plugin-molecule-careful-bash", "ref": "main"},
-    {"name": "molecule-compliance", "repo": "Molecule-AI/molecule-ai-plugin-molecule-compliance", "ref": "main"},
-    {"name": "molecule-dev", "repo": "Molecule-AI/molecule-ai-plugin-molecule-dev", "ref": "main"},
-    {"name": "molecule-freeze-scope", "repo": "Molecule-AI/molecule-ai-plugin-molecule-freeze-scope", "ref": "main"},
-    {"name": "molecule-hitl", "repo": "Molecule-AI/molecule-ai-plugin-molecule-hitl", "ref": "main"},
-    {"name": "molecule-prompt-watchdog", "repo": "Molecule-AI/molecule-ai-plugin-molecule-prompt-watchdog", "ref": "main"},
-    {"name": "molecule-security-scan", "repo": "Molecule-AI/molecule-ai-plugin-molecule-security-scan", "ref": "main"},
-    {"name": "molecule-session-context", "repo": "Molecule-AI/molecule-ai-plugin-molecule-session-context", "ref": "main"},
-    {"name": "molecule-skill-code-review", "repo": "Molecule-AI/molecule-ai-plugin-molecule-skill-code-review", "ref": "main"},
-    {"name": "molecule-skill-cron-learnings", "repo": "Molecule-AI/molecule-ai-plugin-molecule-skill-cron-learnings", "ref": "main"},
-    {"name": "molecule-skill-cross-vendor-review", "repo": "Molecule-AI/molecule-ai-plugin-molecule-skill-cross-vendor-review", "ref": "main"},
-    {"name": "molecule-skill-llm-judge", "repo": "Molecule-AI/molecule-ai-plugin-molecule-skill-llm-judge", "ref": "main"},
-    {"name": "molecule-skill-update-docs", "repo": "Molecule-AI/molecule-ai-plugin-molecule-skill-update-docs", "ref": "main"},
-    {"name": "molecule-workflow-retro", "repo": "Molecule-AI/molecule-ai-plugin-molecule-workflow-retro", "ref": "main"},
-    {"name": "molecule-workflow-triage", "repo": "Molecule-AI/molecule-ai-plugin-molecule-workflow-triage", "ref": "main"},
-    {"name": "superpowers", "repo": "Molecule-AI/molecule-ai-plugin-superpowers", "ref": "main"}
+    {"name": "browser-automation", "repo": "molecule-ai/molecule-ai-plugin-browser-automation", "ref": "main"},
+    {"name": "ecc", "repo": "molecule-ai/molecule-ai-plugin-ecc", "ref": "main"},
+    {"name": "gh-identity", "repo": "molecule-ai/molecule-ai-plugin-gh-identity", "ref": "main"},
+    {"name": "molecule-audit", "repo": "molecule-ai/molecule-ai-plugin-molecule-audit", "ref": "main"},
+    {"name": "molecule-audit-trail", "repo": "molecule-ai/molecule-ai-plugin-molecule-audit-trail", "ref": "main"},
+    {"name": "molecule-careful-bash", "repo": "molecule-ai/molecule-ai-plugin-molecule-careful-bash", "ref": "main"},
+    {"name": "molecule-compliance", "repo": "molecule-ai/molecule-ai-plugin-molecule-compliance", "ref": "main"},
+    {"name": "molecule-dev", "repo": "molecule-ai/molecule-ai-plugin-molecule-dev", "ref": "main"},
+    {"name": "molecule-freeze-scope", "repo": "molecule-ai/molecule-ai-plugin-molecule-freeze-scope", "ref": "main"},
+    {"name": "molecule-hitl", "repo": "molecule-ai/molecule-ai-plugin-molecule-hitl", "ref": "main"},
+    {"name": "molecule-prompt-watchdog", "repo": "molecule-ai/molecule-ai-plugin-molecule-prompt-watchdog", "ref": "main"},
+    {"name": "molecule-security-scan", "repo": "molecule-ai/molecule-ai-plugin-molecule-security-scan", "ref": "main"},
+    {"name": "molecule-session-context", "repo": "molecule-ai/molecule-ai-plugin-molecule-session-context", "ref": "main"},
+    {"name": "molecule-skill-code-review", "repo": "molecule-ai/molecule-ai-plugin-molecule-skill-code-review", "ref": "main"},
+    {"name": "molecule-skill-cron-learnings", "repo": "molecule-ai/molecule-ai-plugin-molecule-skill-cron-learnings", "ref": "main"},
+    {"name": "molecule-skill-cross-vendor-review", "repo": "molecule-ai/molecule-ai-plugin-molecule-skill-cross-vendor-review", "ref": "main"},
+    {"name": "molecule-skill-llm-judge", "repo": "molecule-ai/molecule-ai-plugin-molecule-skill-llm-judge", "ref": "main"},
+    {"name": "molecule-skill-update-docs", "repo": "molecule-ai/molecule-ai-plugin-molecule-skill-update-docs", "ref": "main"},
+    {"name": "molecule-workflow-retro", "repo": "molecule-ai/molecule-ai-plugin-molecule-workflow-retro", "ref": "main"},
+    {"name": "molecule-workflow-triage", "repo": "molecule-ai/molecule-ai-plugin-molecule-workflow-triage", "ref": "main"},
+    {"name": "superpowers", "repo": "molecule-ai/molecule-ai-plugin-superpowers", "ref": "main"}
   ],
   "workspace_templates": [
-    {"name": "claude-code-default", "repo": "Molecule-AI/molecule-ai-workspace-template-claude-code", "ref": "main"},
-    {"name": "hermes", "repo": "Molecule-AI/molecule-ai-workspace-template-hermes", "ref": "main"},
-    {"name": "openclaw", "repo": "Molecule-AI/molecule-ai-workspace-template-openclaw", "ref": "main"},
-    {"name": "codex", "repo": "Molecule-AI/molecule-ai-workspace-template-codex", "ref": "main"},
-    {"name": "langgraph", "repo": "Molecule-AI/molecule-ai-workspace-template-langgraph", "ref": "main"},
-    {"name": "crewai", "repo": "Molecule-AI/molecule-ai-workspace-template-crewai", "ref": "main"},
-    {"name": "autogen", "repo": "Molecule-AI/molecule-ai-workspace-template-autogen", "ref": "main"},
-    {"name": "deepagents", "repo": "Molecule-AI/molecule-ai-workspace-template-deepagents", "ref": "main"},
-    {"name": "gemini-cli", "repo": "Molecule-AI/molecule-ai-workspace-template-gemini-cli", "ref": "main"}
+    {"name": "claude-code-default", "repo": "molecule-ai/molecule-ai-workspace-template-claude-code", "ref": "main"},
+    {"name": "hermes", "repo": "molecule-ai/molecule-ai-workspace-template-hermes", "ref": "main"},
+    {"name": "openclaw", "repo": "molecule-ai/molecule-ai-workspace-template-openclaw", "ref": "main"},
+    {"name": "codex", "repo": "molecule-ai/molecule-ai-workspace-template-codex", "ref": "main"},
+    {"name": "langgraph", "repo": "molecule-ai/molecule-ai-workspace-template-langgraph", "ref": "main"},
+    {"name": "crewai", "repo": "molecule-ai/molecule-ai-workspace-template-crewai", "ref": "main"},
+    {"name": "autogen", "repo": "molecule-ai/molecule-ai-workspace-template-autogen", "ref": "main"},
+    {"name": "deepagents", "repo": "molecule-ai/molecule-ai-workspace-template-deepagents", "ref": "main"},
+    {"name": "gemini-cli", "repo": "molecule-ai/molecule-ai-workspace-template-gemini-cli", "ref": "main"}
   ],
   "org_templates": [
-    {"name": "molecule-dev", "repo": "Molecule-AI/molecule-ai-org-template-molecule-dev", "ref": "main"},
-    {"name": "free-beats-all", "repo": "Molecule-AI/molecule-ai-org-template-free-beats-all", "ref": "main"},
-    {"name": "medo-smoke", "repo": "Molecule-AI/molecule-ai-org-template-medo-smoke", "ref": "main"},
-    {"name": "molecule-worker-gemini", "repo": "Molecule-AI/molecule-ai-org-template-molecule-worker-gemini", "ref": "main"},
-    {"name": "reno-stars", "repo": "Molecule-AI/molecule-ai-org-template-reno-stars", "ref": "main"},
-    {"name": "ux-ab-lab", "repo": "Molecule-AI/molecule-ai-org-template-ux-ab-lab", "ref": "main"},
-    {"name": "mock-bigorg", "repo": "Molecule-AI/molecule-ai-org-template-mock-bigorg", "ref": "main"}
+    {"name": "molecule-dev", "repo": "molecule-ai/molecule-ai-org-template-molecule-dev", "ref": "main"},
+    {"name": "free-beats-all", "repo": "molecule-ai/molecule-ai-org-template-free-beats-all", "ref": "main"},
+    {"name": "medo-smoke", "repo": "molecule-ai/molecule-ai-org-template-medo-smoke", "ref": "main"},
+    {"name": "molecule-worker-gemini", "repo": "molecule-ai/molecule-ai-org-template-molecule-worker-gemini", "ref": "main"},
+    {"name": "ux-ab-lab", "repo": "molecule-ai/molecule-ai-org-template-ux-ab-lab", "ref": "main"},
+    {"name": "mock-bigorg", "repo": "molecule-ai/molecule-ai-org-template-mock-bigorg", "ref": "main"}
   ]
 }

scripts/README.md

@@ -11,7 +11,7 @@ There are three related scripts; pick the right one:
 |---|---|---|
 | `measure-coordinator-task-bounds.sh` | **Canonical** v1 harness for the RFC #2251 / Issue 4 reproduction. Provisions a PM coordinator + Researcher child via `claude-code-default` + `langgraph` templates, sends a synthesis-heavy A2A kickoff, observes elapsed time + activity trace. | OSS-shape platform — localhost or any `/workspaces`-shaped endpoint. Has tenant/admin-token guards for non-localhost runs. |
 | `measure-coordinator-task-bounds-runner.sh` | Generalised runner for the same measurement contract but with **arbitrary template + secret + model combinations** (Hermes/MiniMax, etc.). Useful for cross-runtime variants without modifying the canonical harness. | Same as above (local or SaaS via `MODE=saas`). |
-| `measure-coordinator-task-bounds.sh` (in [molecule-controlplane](https://github.com/Molecule-AI/molecule-controlplane)) | **Production-shape** variant that bootstraps a real staging tenant via `POST /cp/admin/orgs`, then runs the same measurement against `<slug>.staging.moleculesai.app`. | Staging controlplane only — refuses to run against production. |
+| `measure-coordinator-task-bounds.sh` (in [molecule-controlplane](https://git.moleculesai.app/molecule-ai/molecule-controlplane)) | **Production-shape** variant that bootstraps a real staging tenant via `POST /cp/admin/orgs`, then runs the same measurement against `<slug>.staging.moleculesai.app`. | Staging controlplane only — refuses to run against production. |
 
 See `reference_harness_pair_pattern` (auto-memory) for when to use which
 and the cross-repo design rationale.

scripts/build_runtime_package.py

@@ -278,7 +278,7 @@ include = ["molecule_runtime*"]
 README_TEMPLATE = """\
 # molecule-ai-workspace-runtime
 
-Shared workspace runtime for [Molecule AI](https://github.com/Molecule-AI/molecule-core)
+Shared workspace runtime for [Molecule AI](https://git.moleculesai.app/molecule-ai/molecule-core)
 agent adapters. Installed by every workspace template image
 (`workspace-template-claude-code`, `-langgraph`, `-hermes`, etc.) to provide
 A2A delegation, heartbeat, memory, plugin loading, and skill management.
@@ -376,7 +376,7 @@ hold:
    non-plugin-sourced server, which Claude Code rejects with
    `channel_enable requires a marketplace plugin`. Until the
    official `moleculesai/claude-code-plugin` marketplace lands
-   (tracking [#2936](https://github.com/Molecule-AI/molecule-core/issues/2936)),
+   (tracking [#2936](https://git.moleculesai.app/molecule-ai/molecule-core/issues/2936)),
    operators who want push must scaffold their own local marketplace
    under
    `~/.claude/marketplaces/molecule-local/` containing a
@@ -389,14 +389,14 @@ hold:
 Symptom of any condition failing: messages arrive but only via the
 poll path (every ~1–60s), not real-time. There's currently no
 diagnostic surfaced — `molecule-mcp doctor` (tracking
-[#2937](https://github.com/Molecule-AI/molecule-core/issues/2937)) is
+[#2937](https://git.moleculesai.app/molecule-ai/molecule-core/issues/2937)) is
 planned.
 
 If you don't need real-time push, the default poll path works
 universally with no extra setup; both modes converge on the same
 `inbox_pop` ack so messages never duplicate.
 
-See [`docs/workspace-runtime-package.md`](https://github.com/Molecule-AI/molecule-core/blob/main/docs/workspace-runtime-package.md)
+See [`docs/workspace-runtime-package.md`](https://git.moleculesai.app/molecule-ai/molecule-core/src/branch/main/docs/workspace-runtime-package.md)
 for the publish flow and architecture.
 """
 

scripts/clone-manifest.sh

@@ -8,27 +8,24 @@
 # Requires: git, jq (lighter than python3 — ~2MB vs ~50MB in Alpine)
 #
 # Auth (optional):
-#   When MOLECULE_GITEA_TOKEN is set, embed it as the basic-auth password so
-#   private Gitea repos clone successfully. When unset, clone anonymously
-#   (works only for repos that are public on git.moleculesai.app).
+#   Post-2026-05-08 (#192): every repo in manifest.json is public on
+#   git.moleculesai.app. Anonymous clone works for the entire registered
+#   set. The OSS-surface contract is recorded in manifest.json's _comment
+#   — Layer-3 customer/private templates (e.g. reno-stars) are NOT in the
+#   manifest; they are handled at provision-time via the per-tenant
+#   credential resolver (internal#102 RFC).
 #
-#   This is the path the publish-workspace-server-image.yml workflow uses:
-#   it injects AUTO_SYNC_TOKEN (devops-engineer persona PAT, repo:read on
-#   the molecule-ai org) so the in-CI pre-clone step succeeds for ALL
-#   manifest entries — including the 5 private workspace-template-* repos
-#   (codex, crewai, deepagents, gemini-cli, langgraph) and all 7
-#   org-template-* repos.
+#   MOLECULE_GITEA_TOKEN is therefore optional today. Kept supported for
+#   two reasons: (a) historical CI configs that still inject
+#   AUTO_SYNC_TOKEN remain harmless, (b) reserved for the case where a
+#   private internal-only template is later registered via a ci-readonly
+#   team grant — review must explicitly sign off on that, since it
+#   violates the public-OSS-surface contract.
 #
-#   The token never enters the Docker image: this script runs in the
-#   trusted CI context BEFORE `docker buildx build`, populates
+#   The token (when set) never enters the Docker image: this script runs
+#   in the trusted CI context BEFORE `docker buildx build`, populates
 #   .tenant-bundle-deps/, then `Dockerfile.tenant` COPYs from there with
 #   the .git directories already stripped (see line ~67 below).
-#
-#   For backward compatibility — and so a fresh clone works without
-#   secrets when (eventually) the workspace-template-* repos flip public —
-#   the unset path remains a plain anonymous HTTPS clone. That path will
-#   FAIL with "could not read Username" on private repos today; CI MUST
-#   set MOLECULE_GITEA_TOKEN.
 
 set -euo pipefail
 
@@ -68,22 +65,19 @@ clone_category() {
             continue
         fi
 
-        # Post-2026-05-06 GitHub-org-suspension: clone from Gitea instead.
-        # manifest.json paths still read "Molecule-AI/..." (the historic
-        # github.com slug); Gitea lowercases the org part to "molecule-ai/".
-        # Lowercase the org segment on the fly so we don't need to rewrite
-        # every manifest entry.
-        repo_gitea="$(echo "$repo" | awk -F/ '{ printf "%s", tolower($1); for (i=2; i<=NF; i++) printf "/%s", $i; print "" }')"
-
         # Build the clone URL. When MOLECULE_GITEA_TOKEN is set (CI path)
         # embed it as basic-auth so private repos succeed. The username
         # part ("oauth2") is conventional and ignored by Gitea — only the
         # token-as-password is verified.
+        #
+        # manifest.json was migrated to lowercase org slugs on
+        # 2026-05-07 (post-suspension reconciliation), so we use $repo
+        # verbatim — no on-the-fly tolower transform needed.
         if [ -n "${MOLECULE_GITEA_TOKEN:-}" ]; then
-            clone_url="https://oauth2:${MOLECULE_GITEA_TOKEN}@git.moleculesai.app/${repo_gitea}.git"
-            display_url="https://oauth2:***@git.moleculesai.app/${repo_gitea}.git"
+            clone_url="https://oauth2:${MOLECULE_GITEA_TOKEN}@git.moleculesai.app/${repo}.git"
+            display_url="https://oauth2:***@git.moleculesai.app/${repo}.git"
         else
-            clone_url="https://git.moleculesai.app/${repo_gitea}.git"
+            clone_url="https://git.moleculesai.app/${repo}.git"
             display_url="$clone_url"
         fi
 

scripts/demo-freeze.sh

@@ -10,11 +10,11 @@
 #           → PyPI auto-bumps molecule-ai-workspace-runtime patch version
 #           → repository_dispatch fans out to 8 workspace-template-* repos
 #           → each template repo rebuilds and re-tags
-#             ghcr.io/molecule-ai/workspace-template-<runtime>:latest
+#             153263036946.dkr.ecr.us-east-2.amazonaws.com/molecule-ai/workspace-template-<runtime>:latest
 #
 #   PATH 2: any merge to a workspace-template-* repo's main branch
 #           → that repo's publish-image.yml fires
-#           → ghcr.io/molecule-ai/workspace-template-<runtime>:latest
+#           → 153263036946.dkr.ecr.us-east-2.amazonaws.com/molecule-ai/workspace-template-<runtime>:latest
 #             gets re-tagged
 #
 #   provisioner.go:296 RuntimeImages[runtime] reads `:latest` at every

scripts/refresh-workspace-images.sh

@@ -51,7 +51,7 @@ log "pulling latest images for: ${RUNTIMES[*]}"
 PULLED=()
 FAILED=()
 for rt in "${RUNTIMES[@]}"; do
-  IMG="ghcr.io/molecule-ai/workspace-template-$rt:latest"
+  IMG="153263036946.dkr.ecr.us-east-2.amazonaws.com/molecule-ai/workspace-template-$rt:latest"
   if docker pull "$IMG" >/dev/null 2>&1; then
     log "  ✓ $rt"
     PULLED+=("$rt")

scripts/rollback-latest.sh

@@ -1,9 +1,10 @@
 #!/bin/bash
-# rollback-latest.sh — moves the :latest tag on ghcr.io/molecule-ai/platform
-# (and the matching tenant image) back to a prior :staging-<sha> digest
-# without rebuilding anything. Prod tenants auto-pull :latest every 5
-# min, so this is the fast path when a canary-verified image turns out
-# to have a runtime regression that canary didn't catch.
+# rollback-latest.sh — moves the :latest tag on the platform image
+# (and the matching tenant image) on AWS ECR back to a prior
+# :staging-<sha> digest without rebuilding anything. Prod tenants
+# auto-pull :latest every 5 min, so this is the fast path when a
+# canary-verified image turns out to have a runtime regression that
+# canary didn't catch.
 #
 # Usage:
 #   scripts/rollback-latest.sh <sha>
@@ -12,12 +13,14 @@
 # Prereqs:
 #   - crane on $PATH (brew install crane OR download from
 #     https://github.com/google/go-containerregistry/releases)
-#   - GHCR token exported as GITHUB_TOKEN with write:packages scope
+#   - aws CLI authenticated for region us-east-2 with ECR pull/push
+#     access to the molecule-ai/platform + platform-tenant repositories.
+#     `aws sts get-caller-identity` should succeed.
 #
 # What it does (per image — platform + tenant):
-#   crane digest ghcr.io/…:<sha>         # verify the target sha exists
-#   crane tag    ghcr.io/…:<sha> latest  # retag remotely, single API call
-#   crane digest ghcr.io/…:latest        # confirm the move
+#   crane digest <ecr>:<sha>         # verify the target sha exists
+#   crane tag    <ecr>:<sha> latest  # retag remotely, single API call
+#   crane digest <ecr>:latest        # confirm the move
 #
 # Exit codes: 0 = both retagged, 1 = tag missing / crane error, 2 = bad args.
 
@@ -30,21 +33,23 @@ if [ "${1:-}" = "" ]; then
 fi
 
 TARGET_SHA="$1"
-PLATFORM=ghcr.io/molecule-ai/platform
-TENANT=ghcr.io/molecule-ai/platform-tenant
+ECR_HOST=153263036946.dkr.ecr.us-east-2.amazonaws.com
+PLATFORM=$ECR_HOST/molecule-ai/platform
+TENANT=$ECR_HOST/molecule-ai/platform-tenant
 
 if ! command -v crane >/dev/null; then
   echo "ERROR: crane not installed. brew install crane" >&2
   exit 1
 fi
-if [ -z "${GITHUB_TOKEN:-}" ]; then
-  echo "ERROR: GITHUB_TOKEN unset. export it with write:packages scope." >&2
+if ! command -v aws >/dev/null; then
+  echo "ERROR: aws CLI not installed. brew install awscli" >&2
   exit 1
 fi
 
-# Log in once. crane stores creds in a config file keyed by registry;
-# re-running is cheap.
-printf '%s\n' "$GITHUB_TOKEN" | crane auth login ghcr.io -u "${GITHUB_ACTOR:-$(whoami)}" --password-stdin >/dev/null
+# Log in once. ECR auth is via short-lived password from `aws ecr
+# get-login-password`. crane stores creds in a config file keyed by
+# registry; re-running is cheap.
+aws ecr get-login-password --region us-east-2 | crane auth login "$ECR_HOST" -u AWS --password-stdin >/dev/null
 
 roll() {
   local image="$1"

tests/e2e/STAGING_SAAS_E2E.md

@@ -105,5 +105,5 @@ Hard per-workflow timeouts (15–40 min) cap runaway cost. Three teardown layers
 
 ## Known gaps (tracked elsewhere)
 
-- [#1369](https://github.com/Molecule-AI/molecule-core/issues/1369): SaaS canvas Files / Terminal / Peers tabs — architecturally broken; whitelisted in the spec
+- [#1369](https://git.moleculesai.app/molecule-ai/molecule-core/issues/1369): SaaS canvas Files / Terminal / Peers tabs — architecturally broken; whitelisted in the spec
 - LLM-driven delegation (autonomous `delegate_task` tool use) — probabilistic, not in v1; proxy mechanics covered

tests/harness/README.md

@@ -1,5 +1,7 @@
 # Production-shape local harness
 
+<!-- Retrigger Harness Replays after Class G #168 + clone-manifest fix (#42). -->
+
 The harness brings up the SaaS tenant topology on localhost using the
 same `Dockerfile.tenant` image that ships to production. Tests target
 the cf-proxy on `http://localhost:8080` and pass the tenant identity

tests/harness/cf-proxy/Dockerfile

@@ -0,0 +1,14 @@
+# cf-proxy harness image — nginx + the harness's tenant-routing config baked
+# in at build time.
+#
+# Why bake (not bind-mount): on Gitea Actions / act_runner, the runner is a
+# container talking to the OUTER docker daemon over the host socket; runc
+# resolves bind-mount source paths on the outer host filesystem, where the
+# repo at `/workspace/.../tests/harness/cf-proxy/nginx.conf` is invisible.
+# Compose `configs:` (with `file:`) falls back to bind mounts when swarm is
+# not active, so it hits the same gap. A build-time COPY uploads the file
+# as part of the docker build context — the daemon receives the tarball
+# directly and never bind-mounts. See issue #88 item 2.
+FROM nginx:1.27-alpine
+
+COPY nginx.conf /etc/nginx/nginx.conf

tests/harness/compose.yml

@@ -167,15 +167,26 @@ services:
   # Production shape: same single CF tunnel front-doors every tenant
   # subdomain — the Host header carries the tenant identity, not the
   # routing destination. Local cf-proxy mirrors this exactly.
+  #
+  # nginx.conf delivery: built into a custom image via cf-proxy/Dockerfile
+  # (a thin nginx:1.27-alpine + COPY). NOT a bind mount and NOT a
+  # compose `configs:` block, both of which break under Gitea's
+  # act_runner: the runner talks to the OUTER docker daemon over the
+  # host socket, and runc resolves bind sources on the outer host
+  # filesystem, where `/workspace/.../tests/harness/cf-proxy/nginx.conf`
+  # is invisible. Compose `configs:` falls back to bind mounts without
+  # swarm, so it hits the same gap. A build context, by contrast, is
+  # uploaded to the daemon as a tarball at build time — no bind. See
+  # issue #88 item 2.
   cf-proxy:
-    image: nginx:1.27-alpine
+    build:
+      context: ./cf-proxy
+      dockerfile: Dockerfile
     depends_on:
       tenant-alpha:
         condition: service_healthy
       tenant-beta:
         condition: service_healthy
-    volumes:
-      - ./cf-proxy/nginx.conf:/etc/nginx/nginx.conf:ro
     # Bind to 127.0.0.1 only — hardcoded ADMIN_TOKENs make 0.0.0.0
     # exposure unsafe even on a local network.
     ports:

tools/check-template-parity.sh

@@ -18,7 +18,7 @@
 #
 # Or inline via curl:
 #
-#     bash <(curl -fsSL https://raw.githubusercontent.com/Molecule-AI/molecule-core/main/tools/check-template-parity.sh) \
+#     bash <(curl -fsSL https://git.moleculesai.app/molecule-ai/molecule-core/raw/branch/main/tools/check-template-parity.sh) \
 #          install.sh start.sh
 #
 # Exit codes:

workspace-server/.air.toml

@@ -0,0 +1,49 @@
+# air.toml — live-reload config for local docker-compose dev mode.
+#
+# Active when the platform service runs from workspace-server/Dockerfile.dev
+# (selected via docker-compose.dev.yml overlay). In production, the regular
+# Dockerfile builds a static binary; air is dev-only.
+#
+# Reference: https://github.com/air-verse/air
+
+root = "."
+testdata_dir = "testdata"
+tmp_dir = "tmp"
+
+[build]
+  # Same build invocation as Dockerfile's builder stage minus the
+  # CGO_ENABLED=0 toggle (CGO ok in dev for richer race detector output).
+  cmd = "go build -o ./tmp/server ./cmd/server"
+  bin = "tmp/server"
+  full_bin = ""
+  args_bin = []
+  # Watch every .go and .yaml file under workspace-server/.
+  include_ext = ["go", "yaml", "tmpl"]
+  # Don't watch tests, build artifacts, vendored deps, or migration .sql
+  # (migrations need a clean DB anyway — handled by docker-compose down/up).
+  exclude_dir = ["assets", "tmp", "vendor", "testdata", "node_modules"]
+  exclude_file = []
+  # _test.go and *_mock.go shouldn't trigger a rebuild — saves cycles.
+  exclude_regex = ["_test\\.go$", "_mock\\.go$"]
+  exclude_unchanged = true
+  follow_symlink = false
+  log = "build-errors.log"
+  # Kill running binary 1s before starting new one.
+  kill_delay = "1s"
+  send_interrupt = true
+  stop_on_error = true
+  # Debounce: wait this long after last change before triggering rebuild.
+  delay = 500
+
+[log]
+  time = false
+
+[color]
+  main = "magenta"
+  watcher = "cyan"
+  build = "yellow"
+  runner = "green"
+
+[misc]
+  # Don't keep the tmp/ dir around between runs.
+  clean_on_exit = true

workspace-server/.gitignore

@@ -1,2 +1,5 @@
 # The compiled binary, not the cmd/server package.
 /server
+
+# air live-reload build cache (Dockerfile.dev + docker-compose.dev.yml).
+/tmp/

workspace-server/Dockerfile.dev

@@ -0,0 +1,44 @@
+# Dockerfile.dev — local-development image with air-driven live reload.
+#
+# Selected by docker-compose.dev.yml (overlay over docker-compose.yml).
+# Production stays on workspace-server/Dockerfile (static binary, no air).
+#
+# Workflow:
+#   1. docker compose -f docker-compose.yml -f docker-compose.dev.yml up
+#   2. Edit any .go file under workspace-server/
+#   3. air detects, rebuilds, kills old binary, starts new one (~3-5s)
+#   4. No `docker compose up --build` needed
+#
+# Templates + plugins are NOT pre-cloned here — air-mode assumes the
+# developer's filesystem has the workspace-configs-templates/ + plugins/
+# dirs available, mounted at runtime via docker-compose.dev.yml.
+
+FROM golang:1.25-alpine
+
+# air + git (for go mod) + ca-certs (for TLS) + tzdata (for time-zone DB)
+# + docker-cli + docker-cli-buildx so the platform binary can shell out to
+# /var/run/docker.sock (bind-mounted from host) for local-build provisioning.
+# docker-cli alone is insufficient: alpine's docker-cli enables BuildKit by
+# default but ships without buildx, producing
+# `ERROR: BuildKit is enabled but the buildx component is missing or broken`
+# on every `docker build`. docker-cli-buildx provides the buildx subcommand.
+RUN apk add --no-cache git ca-certificates tzdata wget docker-cli docker-cli-buildx \
+ && go install github.com/air-verse/air@latest
+
+WORKDIR /app/workspace-server
+
+# Pre-fetch deps so the first `air` rebuild on a fresh container is fast.
+# These are bind-mount-overridden at runtime, so the COPY here is just
+# to warm the module cache.
+COPY workspace-server/go.mod workspace-server/go.sum ./
+RUN go mod download
+
+# Source is bind-mounted at runtime (see docker-compose.dev.yml volumes
+# block) so the Dockerfile doesn't need to COPY it. air watches the
+# bind-mounted dir for changes.
+
+ENV CGO_ENABLED=0
+ENV GOFLAGS="-buildvcs=false"
+
+# Run air with the .air.toml in the bind-mounted source dir.
+CMD ["air", "-c", ".air.toml"]

workspace-server/cmd/server/main.go

@@ -249,6 +249,19 @@ func main() {
 		})
 	}
 
+	// CP-mode orphan sweeper — SaaS counterpart to the Docker sweeper
+	// above. Re-issues cpProv.Stop for any workspace at status='removed'
+	// with a non-NULL instance_id, healing the deprovision split-write
+	// race documented in #2989: tenant marks status='removed' BEFORE
+	// calling CP DELETE, so a transient CP failure leaves the EC2
+	// running with no retry path. cpProv.Stop is idempotent against
+	// already-terminated instances; on success we clear instance_id.
+	if cpProv != nil {
+		go supervised.RunWithRecover(ctx, "cp-orphan-sweeper", func(c context.Context) {
+			registry.StartCPOrphanSweeper(c, cpProv)
+		})
+	}
+
 	// Pending-uploads GC sweep — deletes acked rows past their retention
 	// window plus unacked rows past expires_at. Without this the
 	// pending_uploads table grows unbounded; even with the 24h hard TTL,

workspace-server/internal/handlers/eic_tunnel_pool.go

@@ -0,0 +1,457 @@
+package handlers
+
+// eic_tunnel_pool.go — refcounted pool for EIC SSH tunnels keyed on
+// instanceID. Reuses one tunnel across N file ops, amortising the
+// ssh-keygen + SendSSHPublicKey + open-tunnel + waitForPort cost
+// (~3-5s) over multiple cats/finds (~50-200ms each).
+//
+// Origin: core#11 — canvas detail-panel config + filesystem load
+// took ~20s. ConfigTab fans out 4 GETs serially; the slowest is
+// /files/config.yaml which dispatches to readFileViaEIC. Without a
+// pool, every readFileViaEIC + listFilesViaEIC + writeFileViaEIC +
+// deleteFileViaEIC pays the full setup cost even when fired
+// back-to-back on the same workspace EC2.
+//
+// The pool keeps one eicSSHSession alive per instanceID for up to
+// poolTTL. SendSSHPublicKey grants a 60s key validity, so poolTTL
+// must stay strictly below that to avoid serving requests on a
+// just-expired key. We default to 50s with a 10s safety margin.
+//
+// Concurrency model:
+//
+//   - Single mutex guards the entries map.
+//   - Slow path (tunnel setup) runs OUTSIDE the lock, gated by an
+//     "intent" placeholder so concurrent acquires for the same
+//     instanceID don't both build a tunnel — the loser drops its
+//     setup and uses the winner's.
+//   - Refcount on each entry; eviction blocked while refcount > 0.
+//   - Janitor goroutine sweeps every poolJanitorInterval, drops
+//     entries where refcount == 0 && expiresAt < now.
+//
+// Test injection:
+//
+//   - poolSetupTunnel is a package-level var so tests can swap the
+//     slow path for a counting stub. Production wires it to
+//     realWithEICTunnel-style setup.
+//   - withEICTunnel (the public, single-shot API) is also a var
+//     (already, see template_files_eic.go). It's rebound here to
+//     pooledWithEICTunnel which routes through globalEICTunnelPool.
+//   - Tests that need single-shot behaviour can set poolTTL = 0,
+//     which makes pooledWithEICTunnel fall through to the underlying
+//     setup directly (no pool entry kept).
+
+import (
+	"context"
+	"fmt"
+	"sync"
+	"time"
+)
+
+// poolTTL is the maximum age of a pooled tunnel. Must be strictly
+// less than the SendSSHPublicKey grant window (60s) so we never
+// serve a request through a key that's about to expire mid-op.
+//
+// Configurable via init-time wiring (see initEICTunnelPool); not a
+// const so tests can pin TTL=0 (disable pooling) or TTL=50ms (drive
+// eviction tests).
+var poolTTL = 50 * time.Second
+
+// poolJanitorInterval is how often the janitor goroutine sweeps for
+// expired idle entries. Tighter than poolTTL so eviction is timely;
+// loose enough that the goroutine doesn't burn CPU.
+var poolJanitorInterval = 10 * time.Second
+
+// poolMaxEntries caps simultaneous instanceIDs the pool tracks.
+// Beyond this, new acquires evict the LRU entry. Defends against a
+// pathological caller (e.g. a sweep over hundreds of workspace
+// EC2s) from leaking unbounded tunnel processes. 32 is a generous
+// ceiling for the canvas use case (one human navigates ≤ ~5
+// workspaces at a time).
+var poolMaxEntries = 32
+
+// poolSetupTunnel is the slow-path tunnel constructor. Wrapped in a
+// var so tests can inject a counter stub. Returns a session and a
+// cleanup function (closes the open-tunnel subprocess + scrubs the
+// ephemeral keydir). nil session + non-nil err means setup failed
+// and there is nothing to clean up.
+//
+// Production wiring lives in eic_tunnel_pool_setup.go (a thin shim
+// over the existing realWithEICTunnel logic).
+var poolSetupTunnel = func(ctx context.Context, instanceID string) (
+	sess eicSSHSession, cleanup func(), err error) {
+	return setupRealEICTunnel(ctx, instanceID)
+}
+
+// pooledTunnel is one entry in the pool. session is shared by N
+// concurrent fn calls; cleanup runs once when refcount returns to
+// zero AND the entry is past expiresAt or evicted.
+//
+// lastUsed tracks the most recent acquire time for LRU bookkeeping
+// (overflow eviction). expiresAt is set at construction and not
+// extended on use — a tunnel cannot live past poolTTL even if it's
+// hot, because the underlying SendSSHPublicKey grant expires.
+type pooledTunnel struct {
+	session   eicSSHSession
+	cleanup   func()
+	expiresAt time.Time
+	lastUsed  time.Time
+	refcount  int
+	poisoned  bool // true if a fn returned a tunnel-fatal error; do not reuse
+}
+
+// eicTunnelPool is the package-level pool. Single instance lives
+// in globalEICTunnelPool; constructor runs lazily on first acquire.
+type eicTunnelPool struct {
+	mu      sync.Mutex
+	entries map[string]*pooledTunnel
+	// pendingSetups guards concurrent setup for the same instanceID.
+	// First acquirer takes the slot; later ones wait on the channel.
+	pendingSetups map[string]chan struct{}
+	stopJanitor   chan struct{}
+	// janitorInterval is captured at pool construction from the
+	// package-level poolJanitorInterval var. Captured (not re-read on
+	// every tick) so a test that swaps the package var via t.Cleanup
+	// after a global pool's janitor is already running can't race
+	// with that goroutine's ticker read. The global pool is created
+	// lazily once per process via sync.Once; before this capture
+	// landed, every test that touched poolJanitorInterval after the
+	// global pool's first-touch raced the janitor (caught by -race
+	// on staging tip 249dbc6a — TestPooledWithEICTunnel_PanicPoisonsEntry).
+	// Tests still get the new value on a freshPool() because they
+	// set the package var BEFORE calling newEICTunnelPool().
+	janitorInterval time.Duration
+}
+
+var (
+	globalEICTunnelPool     *eicTunnelPool
+	globalEICTunnelPoolOnce sync.Once
+)
+
+// getEICTunnelPool returns the singleton pool, lazy-initialising on
+// first call. Idempotent.
+func getEICTunnelPool() *eicTunnelPool {
+	globalEICTunnelPoolOnce.Do(func() {
+		globalEICTunnelPool = newEICTunnelPool()
+		go globalEICTunnelPool.janitor()
+	})
+	return globalEICTunnelPool
+}
+
+// newEICTunnelPool constructs an empty pool. Exported so tests can
+// build isolated pools without sharing the singleton.
+//
+// Captures poolJanitorInterval at construction time so the janitor
+// goroutine doesn't race with t.Cleanup-driven swaps of the package
+// var. See the janitorInterval field comment for the failure mode.
+func newEICTunnelPool() *eicTunnelPool {
+	return &eicTunnelPool{
+		entries:         map[string]*pooledTunnel{},
+		pendingSetups:   map[string]chan struct{}{},
+		stopJanitor:     make(chan struct{}),
+		janitorInterval: poolJanitorInterval,
+	}
+}
+
+// acquire returns a usable session for instanceID. If a healthy entry
+// exists, refcount++ and return it. If a setup is in flight for the
+// same instanceID, wait for it. Otherwise build one (slow path).
+//
+// done() must be called by the caller when the op finishes. It
+// decrements refcount and triggers cleanup if the entry is past
+// TTL or poisoned and refcount==0.
+//
+// Errors from the slow path propagate; pool state is not modified
+// for failed setups (no poisoned entry created — that's only for
+// fn-returned errors on a previously-good session).
+func (p *eicTunnelPool) acquire(ctx context.Context, instanceID string) (
+	sess eicSSHSession, done func(poisoned bool), err error) {
+
+	if poolTTL <= 0 {
+		// Pool disabled (TTL=0 mode for tests / opt-out). Fall
+		// through to a direct setup with caller-driven cleanup.
+		s, cleanup, err := poolSetupTunnel(ctx, instanceID)
+		if err != nil {
+			return eicSSHSession{}, nil, err
+		}
+		return s, func(_ bool) { cleanup() }, nil
+	}
+
+	for {
+		p.mu.Lock()
+		if pt, ok := p.entries[instanceID]; ok && !pt.poisoned && pt.expiresAt.After(time.Now()) {
+			pt.refcount++
+			pt.lastUsed = time.Now()
+			p.mu.Unlock()
+			return pt.session, p.releaser(instanceID, pt), nil
+		}
+		// Either no entry, expired entry, or poisoned entry. If a
+		// setup is already in flight, wait and retry.
+		if pending, ok := p.pendingSetups[instanceID]; ok {
+			p.mu.Unlock()
+			select {
+			case <-pending:
+				continue // re-check the entries map
+			case <-ctx.Done():
+				return eicSSHSession{}, nil, ctx.Err()
+			}
+		}
+		// Drop expired/poisoned entry now (we'll cleanup outside
+		// the lock — the entry is unreferenced or we'd not be here).
+		var oldCleanup func()
+		if pt, ok := p.entries[instanceID]; ok {
+			if pt.refcount == 0 {
+				oldCleanup = pt.cleanup
+				delete(p.entries, instanceID)
+			}
+		}
+		// Reserve the setup slot.
+		signal := make(chan struct{})
+		p.pendingSetups[instanceID] = signal
+		p.mu.Unlock()
+
+		if oldCleanup != nil {
+			go oldCleanup()
+		}
+
+		// Slow path: build a new tunnel. Anything that goes wrong
+		// here cleans up the pendingSetups slot and propagates to
+		// the caller without leaving the pool in a state where the
+		// next acquire blocks waiting on a signal that never fires.
+		newSess, cleanup, setupErr := poolSetupTunnel(ctx, instanceID)
+
+		p.mu.Lock()
+		delete(p.pendingSetups, instanceID)
+		close(signal)
+
+		if setupErr != nil {
+			p.mu.Unlock()
+			return eicSSHSession{}, nil, fmt.Errorf("eic tunnel setup: %w", setupErr)
+		}
+
+		// Enforce LRU bound BEFORE inserting so we don't briefly
+		// exceed the cap even by one entry.
+		p.evictLRUIfFullLocked(instanceID)
+
+		pt := &pooledTunnel{
+			session:   newSess,
+			cleanup:   cleanup,
+			expiresAt: time.Now().Add(poolTTL),
+			lastUsed:  time.Now(),
+			refcount:  1,
+		}
+		p.entries[instanceID] = pt
+		p.mu.Unlock()
+		return pt.session, p.releaser(instanceID, pt), nil
+	}
+}
+
+// releaser returns a closure that decrements refcount and triggers
+// cleanup if (a) the entry is past TTL or (b) the caller signalled
+// poison. Idempotent against double-release (decrements once via the
+// captured pt; pool entry may have been replaced by then).
+func (p *eicTunnelPool) releaser(instanceID string, pt *pooledTunnel) func(poisoned bool) {
+	released := false
+	return func(poisoned bool) {
+		p.mu.Lock()
+		defer p.mu.Unlock()
+		if released {
+			return
+		}
+		released = true
+		pt.refcount--
+		if poisoned {
+			pt.poisoned = true
+		}
+		// Evict immediately if poisoned-and-idle OR expired-and-idle.
+		// Hot entries (refcount > 0) defer eviction to the last release.
+		if pt.refcount == 0 && (pt.poisoned || pt.expiresAt.Before(time.Now())) {
+			// If the entry in the map is still us, remove it.
+			if cur, ok := p.entries[instanceID]; ok && cur == pt {
+				delete(p.entries, instanceID)
+			}
+			go pt.cleanup()
+		}
+	}
+}
+
+// evictLRUIfFullLocked drops the least-recently-used IDLE entry
+// when the pool is at capacity. Caller must hold p.mu. The new
+// instanceID about to be inserted is excluded so we don't evict
+// ourselves. If no idle entries exist, no eviction happens — the
+// new entry will push us above the soft cap until something releases.
+func (p *eicTunnelPool) evictLRUIfFullLocked(skipInstance string) {
+	if len(p.entries) < poolMaxEntries {
+		return
+	}
+	var oldestKey string
+	var oldest *pooledTunnel
+	for k, pt := range p.entries {
+		if k == skipInstance {
+			continue
+		}
+		if pt.refcount > 0 {
+			continue
+		}
+		if oldest == nil || pt.lastUsed.Before(oldest.lastUsed) {
+			oldestKey = k
+			oldest = pt
+		}
+	}
+	if oldest == nil {
+		return // every entry is in use; no eviction possible
+	}
+	delete(p.entries, oldestKey)
+	go oldest.cleanup()
+}
+
+// janitor periodically scans for entries that are idle AND expired,
+// closing their tunnels. Runs forever (per pool lifetime); cancelled
+// by close(p.stopJanitor) for tests that build short-lived pools.
+//
+// Reads p.janitorInterval (captured at construction) instead of the
+// package-level poolJanitorInterval — see janitorInterval field comment.
+func (p *eicTunnelPool) janitor() {
+	t := time.NewTicker(p.janitorInterval)
+	defer t.Stop()
+	for {
+		select {
+		case <-t.C:
+			p.sweep()
+		case <-p.stopJanitor:
+			return
+		}
+	}
+}
+
+// sweep is one janitor pass. Drops idle expired entries.
+func (p *eicTunnelPool) sweep() {
+	p.mu.Lock()
+	now := time.Now()
+	var toClose []func()
+	for k, pt := range p.entries {
+		if pt.refcount == 0 && pt.expiresAt.Before(now) {
+			toClose = append(toClose, pt.cleanup)
+			delete(p.entries, k)
+		}
+	}
+	p.mu.Unlock()
+	for _, c := range toClose {
+		go c()
+	}
+}
+
+// stop terminates the janitor and closes all idle entries. Hot
+// (refcount > 0) entries are NOT force-closed — callers running
+// against them would see a use-after-free. In practice stop is only
+// called by tests that have already drained their callers.
+func (p *eicTunnelPool) stop() {
+	close(p.stopJanitor)
+	p.mu.Lock()
+	defer p.mu.Unlock()
+	for k, pt := range p.entries {
+		if pt.refcount == 0 {
+			go pt.cleanup()
+			delete(p.entries, k)
+		}
+	}
+}
+
+// pooledWithEICTunnel is the pool-backed replacement for
+// realWithEICTunnel. The signature matches `var withEICTunnel`
+// exactly so the rebind (in initEICTunnelPool) is a drop-in.
+//
+// Errors from `fn` itself are forwarded to the caller AND mark the
+// pool entry as poisoned, so the next acquire builds a fresh
+// tunnel. This catches the case where the workspace EC2 was
+// restarted out-of-band (tunnel still appears alive locally but
+// every cat/find errors out).
+func pooledWithEICTunnel(ctx context.Context, instanceID string,
+	fn func(s eicSSHSession) error) error {
+	pool := getEICTunnelPool()
+	sess, done, err := pool.acquire(ctx, instanceID)
+	if err != nil {
+		return err
+	}
+	// poisoned defaults to true so a panic from fn poisons the
+	// entry on the way through the deferred release. Without the
+	// defer, a panicking fn would leak refcount=1 forever and
+	// permanently block eviction of this entry. The fn-error path
+	// resets poisoned to its real classification before return.
+	poisoned := true
+	defer func() { done(poisoned) }()
+	fnErr := fn(sess)
+	poisoned = fnErrIndicatesTunnelFault(fnErr)
+	return fnErr
+}
+
+// fnErrIndicatesTunnelFault returns true for fn errors whose nature
+// suggests the underlying tunnel is no longer reusable (auth gone,
+// network gone, ssh process dead). Returning true poisons the pool
+// entry so the next acquire builds fresh.
+//
+// Conservative: only marks tunnel-faulty for clearly tunnel-level
+// failures (connection refused, broken pipe, ssh exit-status from
+// fatal-channel signals). A `cat` returning os.ErrNotExist on a
+// missing file is NOT a tunnel fault — that's the file path being
+// wrong, the tunnel is fine.
+func fnErrIndicatesTunnelFault(err error) bool {
+	if err == nil {
+		return false
+	}
+	msg := err.Error()
+	// stderr substrings produced by ssh when the tunnel is broken.
+	for _, marker := range []string{
+		"connection refused",
+		"connection closed",
+		"broken pipe",
+		"Connection reset by peer",
+		"kex_exchange_identification",
+		"port forwarding failed",
+		"Permission denied",
+		"Authentication failed",
+	} {
+		if containsCaseInsensitive(msg, marker) {
+			return true
+		}
+	}
+	return false
+}
+
+// containsCaseInsensitive avoids importing strings just for this
+// (the file already needs ssh stderr matching elsewhere — this
+// keeps the helper local to avoid a cross-file dependency).
+func containsCaseInsensitive(s, substr string) bool {
+	if len(substr) > len(s) {
+		return false
+	}
+	// Manual lowercase compare loop; ssh error markers are ASCII so
+	// no need for unicode-aware folding.
+	low := func(b byte) byte {
+		if b >= 'A' && b <= 'Z' {
+			return b + 32
+		}
+		return b
+	}
+	for i := 0; i+len(substr) <= len(s); i++ {
+		match := true
+		for j := 0; j < len(substr); j++ {
+			if low(s[i+j]) != low(substr[j]) {
+				match = false
+				break
+			}
+		}
+		if match {
+			return true
+		}
+	}
+	return false
+}
+
+// initEICTunnelPool rebinds the package-level withEICTunnel var to
+// the pooled implementation. Called once at package init via the
+// init() in eic_tunnel_pool_setup.go (split file so the rebind
+// itself is testable without dragging in the production setup
+// shim's exec/aws dependencies).
+func initEICTunnelPool() {
+	withEICTunnel = pooledWithEICTunnel
+}

workspace-server/internal/handlers/eic_tunnel_pool_setup.go

@@ -0,0 +1,136 @@
+package handlers
+
+// eic_tunnel_pool_setup.go — production setup shim.
+//
+// setupRealEICTunnel decomposes the existing realWithEICTunnel into
+// its slow half (build the tunnel) and its caller half (run fn). The
+// pool calls the slow half once and shares the resulting session
+// across N callers, holding cleanup until the last release.
+//
+// Why decompose instead of refactoring realWithEICTunnel: the
+// existing function and its test stub-vars (withEICTunnel,
+// sendSSHPublicKey, openTunnelCmd) are load-bearing for the
+// dispatch tests. Extracting a sibling setup function preserves the
+// existing single-shot path verbatim — the pool wraps it by calling
+// realWithEICTunnel through a thin adapter, leaving the tested
+// surface unchanged.
+//
+// The pool's acquire() invokes poolSetupTunnel, which is a `var`
+// pointing to setupRealEICTunnel for production and a counting stub
+// for tests.
+
+import (
+	"context"
+	"fmt"
+	"os"
+	"os/exec"
+	"strings"
+	"time"
+)
+
+// setupRealEICTunnel is the slow path that the pool consumes when
+// no warm entry exists. Mirrors realWithEICTunnel's setup half but
+// returns the session + cleanup instead of running fn inline.
+//
+// The cleanup func owns the tunnel subprocess, ephemeral key dir,
+// and a one-time wait. Idempotent — calling it twice is safe; the
+// pool guarantees one call per session, but defence-in-depth helps
+// when tests run pools in parallel and racy sweeps re-trigger.
+func setupRealEICTunnel(ctx context.Context, instanceID string) (
+	eicSSHSession, func(), error) {
+
+	if instanceID == "" {
+		return eicSSHSession{}, nil,
+			fmt.Errorf("workspace has no instance_id — not a SaaS EC2 workspace")
+	}
+	osUser := os.Getenv("WORKSPACE_EC2_OS_USER")
+	if osUser == "" {
+		osUser = "ubuntu"
+	}
+	region := os.Getenv("AWS_REGION")
+	if region == "" {
+		region = "us-east-2"
+	}
+
+	keyDir, err := os.MkdirTemp("", "molecule-eic-pool-*")
+	if err != nil {
+		return eicSSHSession{}, nil, fmt.Errorf("keydir mkdir: %w", err)
+	}
+	keyPath := keyDir + "/id"
+	if out, kerr := exec.CommandContext(ctx, "ssh-keygen",
+		"-t", "ed25519", "-f", keyPath, "-N", "", "-q",
+		"-C", "molecule-eic-pool",
+	).CombinedOutput(); kerr != nil {
+		_ = os.RemoveAll(keyDir)
+		return eicSSHSession{}, nil,
+			fmt.Errorf("ssh-keygen: %w (%s)", kerr, strings.TrimSpace(string(out)))
+	}
+	pubKey, err := os.ReadFile(keyPath + ".pub")
+	if err != nil {
+		_ = os.RemoveAll(keyDir)
+		return eicSSHSession{}, nil, fmt.Errorf("read pubkey: %w", err)
+	}
+
+	if err := sendSSHPublicKey(ctx, region, instanceID, osUser,
+		strings.TrimSpace(string(pubKey))); err != nil {
+		_ = os.RemoveAll(keyDir)
+		return eicSSHSession{}, nil, fmt.Errorf("send-ssh-public-key: %w", err)
+	}
+
+	localPort, err := pickFreePort()
+	if err != nil {
+		_ = os.RemoveAll(keyDir)
+		return eicSSHSession{}, nil, fmt.Errorf("pick free port: %w", err)
+	}
+
+	tunnel := openTunnelCmd(eicSSHOptions{
+		InstanceID:     instanceID,
+		OSUser:         osUser,
+		Region:         region,
+		LocalPort:      localPort,
+		PrivateKeyPath: keyPath,
+	})
+	tunnel.Env = os.Environ()
+	if err := tunnel.Start(); err != nil {
+		_ = os.RemoveAll(keyDir)
+		return eicSSHSession{}, nil, fmt.Errorf("open-tunnel start: %w", err)
+	}
+
+	if err := waitForPort(ctx, "127.0.0.1", localPort, 10*time.Second); err != nil {
+		if tunnel.Process != nil {
+			_ = tunnel.Process.Kill()
+		}
+		_ = tunnel.Wait()
+		_ = os.RemoveAll(keyDir)
+		return eicSSHSession{}, nil, fmt.Errorf("tunnel never listened: %w", err)
+	}
+
+	cleanedUp := false
+	cleanup := func() {
+		if cleanedUp {
+			return
+		}
+		cleanedUp = true
+		if tunnel.Process != nil {
+			_ = tunnel.Process.Kill()
+		}
+		_ = tunnel.Wait()
+		_ = os.RemoveAll(keyDir)
+	}
+
+	return eicSSHSession{
+		keyPath:    keyPath,
+		localPort:  localPort,
+		osUser:     osUser,
+		instanceID: instanceID,
+	}, cleanup, nil
+}
+
+// init wires the pool into the package-level withEICTunnel var so
+// every read/write/list/delete EIC op uses pooled tunnels by default.
+// Test files that need single-shot behaviour can swap withEICTunnel
+// back via the existing stubWithEICTunnel pattern, OR set poolTTL=0
+// to disable pooling without rebinding the var.
+func init() {
+	initEICTunnelPool()
+}

workspace-server/internal/handlers/eic_tunnel_pool_test.go

@@ -0,0 +1,467 @@
+package handlers
+
+// eic_tunnel_pool_test.go — tests for the refcounted EIC tunnel pool
+// added in core#11. Stubs poolSetupTunnel with a counter so the
+// tests don't fork ssh-keygen / aws subprocesses.
+//
+// Per memory feedback_assert_exact_not_substring: each test pins
+// exact expected counts (not "at least N") so a regression that
+// silently double-sets-up surfaces here.
+
+import (
+	"context"
+	"errors"
+	"sync"
+	"sync/atomic"
+	"testing"
+	"time"
+)
+
+// withPoolSetupStub swaps poolSetupTunnel for a counting fake that
+// returns a sentinel session and a cleanup func that records its
+// invocation. Restores on test cleanup.
+//
+// setupSignal blocks each setup until released — for concurrent-
+// acquire tests where we want to gate setup completion.
+func withPoolSetupStub(t *testing.T) (
+	setupCount *int64, cleanupCount *int64, restore func(), unblock func()) {
+	t.Helper()
+	prev := poolSetupTunnel
+	prevTTL := poolTTL
+	prevJanitor := poolJanitorInterval
+
+	var sc, cc int64
+	setupCount, cleanupCount = &sc, &cc
+
+	gate := make(chan struct{}, 1)
+	gate <- struct{}{} // allow the first setup through immediately
+	unblock = func() { gate <- struct{}{} }
+
+	poolSetupTunnel = func(ctx context.Context, instanceID string) (
+		eicSSHSession, func(), error) {
+		select {
+		case <-gate:
+		case <-ctx.Done():
+			return eicSSHSession{}, nil, ctx.Err()
+		}
+		atomic.AddInt64(&sc, 1)
+		sess := eicSSHSession{
+			instanceID: instanceID,
+			osUser:     "ubuntu",
+			localPort:  10000 + int(atomic.LoadInt64(&sc)),
+			keyPath:    "/tmp/molecule-eic-test-" + instanceID,
+		}
+		cleanup := func() { atomic.AddInt64(&cc, 1) }
+		return sess, cleanup, nil
+	}
+
+	restore = func() {
+		poolSetupTunnel = prev
+		poolTTL = prevTTL
+		poolJanitorInterval = prevJanitor
+	}
+	t.Cleanup(restore)
+	return
+}
+
+// freshPool returns an isolated pool (NOT the global) so tests run
+// independently. Stops the janitor on cleanup.
+func freshPool(t *testing.T) *eicTunnelPool {
+	t.Helper()
+	p := newEICTunnelPool()
+	t.Cleanup(p.stop)
+	return p
+}
+
+// TestEICTunnelPool_FourOpsAmortise pins the core invariant: four
+// sequential acquire/release cycles on the same instanceID share
+// ONE underlying tunnel setup. Mutation: delete the cache hit branch
+// in acquire() → setupCount goes 1 → 4 → test fails.
+func TestEICTunnelPool_FourOpsAmortise(t *testing.T) {
+	setupCount, cleanupCount, _, _ := withPoolSetupStub(t)
+	// Refill gate after each setup so concurrent stubs aren't blocked
+	// (we want every test to be able to set up if it needs to).
+	t.Cleanup(func() { /* no-op; defer is enough */ })
+	poolTTL = 50 * time.Second
+	pool := freshPool(t)
+	ctx := context.Background()
+
+	for i := 0; i < 4; i++ {
+		sess, done, err := pool.acquire(ctx, "i-test-1")
+		if err != nil {
+			t.Fatalf("op %d: acquire: %v", i, err)
+		}
+		if sess.instanceID != "i-test-1" {
+			t.Fatalf("op %d: session has wrong instanceID: %q", i, sess.instanceID)
+		}
+		done(false)
+	}
+
+	if got := atomic.LoadInt64(setupCount); got != 1 {
+		t.Errorf("expected exactly 1 tunnel setup across 4 ops, got %d", got)
+	}
+	if got := atomic.LoadInt64(cleanupCount); got != 0 {
+		t.Errorf("expected 0 cleanups while entry is hot (TTL=50s), got %d", got)
+	}
+}
+
+// TestEICTunnelPool_DifferentInstancesDoNotShare pins that two
+// different instanceIDs each get their own tunnel — the pool is
+// keyed on instanceID, not a single global slot.
+func TestEICTunnelPool_DifferentInstancesDoNotShare(t *testing.T) {
+	setupCount, _, _, unblock := withPoolSetupStub(t)
+	poolTTL = 50 * time.Second
+	pool := freshPool(t)
+	ctx := context.Background()
+
+	// First instance setup uses the initial gate slot.
+	_, doneA, err := pool.acquire(ctx, "i-a")
+	if err != nil {
+		t.Fatalf("acquire A: %v", err)
+	}
+	doneA(false)
+
+	// Second instance needs a new slot through the gate.
+	unblock()
+	_, doneB, err := pool.acquire(ctx, "i-b")
+	if err != nil {
+		t.Fatalf("acquire B: %v", err)
+	}
+	doneB(false)
+
+	if got := atomic.LoadInt64(setupCount); got != 2 {
+		t.Errorf("expected 2 setups (one per instance), got %d", got)
+	}
+}
+
+// TestEICTunnelPool_TTLEviction: a short TTL forces the second op
+// to build a fresh tunnel after the first expires.
+func TestEICTunnelPool_TTLEviction(t *testing.T) {
+	setupCount, cleanupCount, _, unblock := withPoolSetupStub(t)
+	poolTTL = 50 * time.Millisecond
+	poolJanitorInterval = 1 * time.Second // keep janitor away
+	pool := freshPool(t)
+	ctx := context.Background()
+
+	_, done, err := pool.acquire(ctx, "i-ttl")
+	if err != nil {
+		t.Fatalf("acquire 1: %v", err)
+	}
+	done(false)
+
+	time.Sleep(80 * time.Millisecond) // past TTL
+
+	unblock() // allow next setup
+	_, done, err = pool.acquire(ctx, "i-ttl")
+	if err != nil {
+		t.Fatalf("acquire 2: %v", err)
+	}
+	done(false)
+
+	if got := atomic.LoadInt64(setupCount); got != 2 {
+		t.Errorf("expected 2 setups (TTL eviction between), got %d", got)
+	}
+	// First entry should have been cleaned up when the second
+	// acquire evicted it on the slow path. Cleanup runs in a
+	// goroutine; poll briefly for it to land.
+	deadline := time.Now().Add(500 * time.Millisecond)
+	for atomic.LoadInt64(cleanupCount) < 1 && time.Now().Before(deadline) {
+		time.Sleep(5 * time.Millisecond)
+	}
+	if got := atomic.LoadInt64(cleanupCount); got < 1 {
+		t.Errorf("expected ≥1 cleanup (first entry evicted), got %d", got)
+	}
+}
+
+// TestEICTunnelPool_FailureInvalidates pins the poison-on-fault
+// behavior — fn returning a tunnel-fatal error marks the entry
+// unusable so the next acquire builds fresh.
+func TestEICTunnelPool_FailureInvalidates(t *testing.T) {
+	setupCount, _, _, unblock := withPoolSetupStub(t)
+	poolTTL = 50 * time.Second
+	pool := freshPool(t)
+	ctx := context.Background()
+
+	_, done, err := pool.acquire(ctx, "i-fault")
+	if err != nil {
+		t.Fatalf("acquire 1: %v", err)
+	}
+	done(true) // signal poison
+
+	unblock() // let the next setup through
+	_, done, err = pool.acquire(ctx, "i-fault")
+	if err != nil {
+		t.Fatalf("acquire 2: %v", err)
+	}
+	done(false)
+
+	if got := atomic.LoadInt64(setupCount); got != 2 {
+		t.Errorf("expected 2 setups (poison forced rebuild), got %d", got)
+	}
+}
+
+// TestEICTunnelPool_ConcurrentAcquireSingleSetup pins that N
+// concurrent acquires for the same instanceID before any release
+// only trigger ONE tunnel setup — the rest wait via pendingSetups.
+//
+// Without this guard each concurrent acquire would spawn its own
+// tunnel and the loser-cleanup would still leak refcount. Mutation:
+// delete the pendingSetups gate → setupCount goes 1 → N → fails.
+func TestEICTunnelPool_ConcurrentAcquireSingleSetup(t *testing.T) {
+	setupCount, _, _, _ := withPoolSetupStub(t)
+	// Pause setup so all goroutines pile into the pending slot.
+	prev := poolSetupTunnel
+	gate := make(chan struct{})
+	poolSetupTunnel = func(ctx context.Context, instanceID string) (
+		eicSSHSession, func(), error) {
+		<-gate
+		atomic.AddInt64(setupCount, 1)
+		return eicSSHSession{instanceID: instanceID}, func() {}, nil
+	}
+	t.Cleanup(func() { poolSetupTunnel = prev })
+
+	poolTTL = 50 * time.Second
+	pool := freshPool(t)
+	ctx := context.Background()
+
+	const N = 8
+	type result struct {
+		done func(bool)
+		err  error
+	}
+	results := make(chan result, N)
+	var startWg sync.WaitGroup
+	startWg.Add(N)
+	for i := 0; i < N; i++ {
+		go func() {
+			startWg.Done()
+			_, done, err := pool.acquire(ctx, "i-concurrent")
+			results <- result{done, err}
+		}()
+	}
+	startWg.Wait()
+	// give all N goroutines time to enter pool.acquire
+	time.Sleep(20 * time.Millisecond)
+	close(gate)
+
+	for i := 0; i < N; i++ {
+		r := <-results
+		if r.err != nil {
+			t.Fatalf("acquire %d: %v", i, r.err)
+		}
+		r.done(false)
+	}
+
+	if got := atomic.LoadInt64(setupCount); got != 1 {
+		t.Errorf("expected 1 setup across %d concurrent acquires, got %d", N, got)
+	}
+}
+
+// TestEICTunnelPool_TTLZeroDisablesPooling pins the escape hatch:
+// poolTTL=0 means every acquire goes straight through to setup +
+// cleanup, no entry kept. Useful for tests / opt-out.
+func TestEICTunnelPool_TTLZeroDisablesPooling(t *testing.T) {
+	setupCount, cleanupCount, _, unblock := withPoolSetupStub(t)
+	poolTTL = 0
+	pool := freshPool(t)
+	ctx := context.Background()
+
+	_, done, err := pool.acquire(ctx, "i-ttlzero")
+	if err != nil {
+		t.Fatalf("acquire 1: %v", err)
+	}
+	done(false)
+
+	unblock()
+	_, done, err = pool.acquire(ctx, "i-ttlzero")
+	if err != nil {
+		t.Fatalf("acquire 2: %v", err)
+	}
+	done(false)
+
+	if got := atomic.LoadInt64(setupCount); got != 2 {
+		t.Errorf("expected 2 setups with TTL=0 (pool disabled), got %d", got)
+	}
+	if got := atomic.LoadInt64(cleanupCount); got != 2 {
+		t.Errorf("expected 2 cleanups with TTL=0 (each release closes), got %d", got)
+	}
+}
+
+// TestEICTunnelPool_LRUEvictionAtCap pins the LRU defence: when the
+// pool reaches poolMaxEntries, a new acquire for an unseen
+// instanceID evicts the LRU idle entry instead of growing unbounded.
+func TestEICTunnelPool_LRUEvictionAtCap(t *testing.T) {
+	setupCount, cleanupCount, _, _ := withPoolSetupStub(t)
+	prev := poolMaxEntries
+	poolMaxEntries = 2
+	t.Cleanup(func() { poolMaxEntries = prev })
+	poolTTL = 50 * time.Second
+
+	// Replace stub with one that doesn't gate so we can fill quickly.
+	poolSetupTunnel = func(ctx context.Context, instanceID string) (
+		eicSSHSession, func(), error) {
+		atomic.AddInt64(setupCount, 1)
+		return eicSSHSession{instanceID: instanceID}, func() {
+			atomic.AddInt64(cleanupCount, 1)
+		}, nil
+	}
+
+	pool := freshPool(t)
+	ctx := context.Background()
+
+	for _, id := range []string{"i-1", "i-2"} {
+		_, done, err := pool.acquire(ctx, id)
+		if err != nil {
+			t.Fatalf("acquire %s: %v", id, err)
+		}
+		done(false)
+	}
+	// Both entries idle, pool at cap.
+	_, done, err := pool.acquire(ctx, "i-3")
+	if err != nil {
+		t.Fatalf("acquire i-3: %v", err)
+	}
+	done(false)
+
+	// Wait for the goroutine'd cleanup of the evicted entry.
+	deadline := time.Now().Add(500 * time.Millisecond)
+	for atomic.LoadInt64(cleanupCount) < 1 && time.Now().Before(deadline) {
+		time.Sleep(10 * time.Millisecond)
+	}
+
+	if got := atomic.LoadInt64(setupCount); got != 3 {
+		t.Errorf("expected 3 setups (one per unique instance), got %d", got)
+	}
+	if got := atomic.LoadInt64(cleanupCount); got < 1 {
+		t.Errorf("expected ≥1 cleanup (LRU eviction), got %d", got)
+	}
+}
+
+// TestEICTunnelPool_PoisonedClassification pins the heuristic that
+// distinguishes tunnel-fatal errors (poison the entry) from
+// app-level errors (file not found, validation) that should NOT
+// invalidate the tunnel.
+func TestEICTunnelPool_PoisonedClassification(t *testing.T) {
+	cases := []struct {
+		name string
+		err  error
+		want bool
+	}{
+		{"nil", nil, false},
+		{"file not found", errors.New("os: file does not exist"), false},
+		{"validation", errors.New("invalid path: must be relative"), false},
+		{"connection refused", errors.New("ssh: connect to host: connection refused"), true},
+		{"connection refused upper", errors.New("Connection Refused"), true},
+		{"broken pipe", errors.New("write tunnel: broken pipe"), true},
+		{"permission denied", errors.New("Permission denied (publickey)"), true},
+		{"auth failed", errors.New("Authentication failed"), true},
+		{"connection reset", errors.New("Connection reset by peer"), true},
+		{"port forward", errors.New("port forwarding failed"), true},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			got := fnErrIndicatesTunnelFault(tc.err)
+			if got != tc.want {
+				t.Errorf("fnErrIndicatesTunnelFault(%v) = %v, want %v",
+					tc.err, got, tc.want)
+			}
+		})
+	}
+}
+
+// TestEICTunnelPool_RefcountBlocksEviction pins that an entry past
+// TTL is NOT evicted while a caller still holds it — preventing
+// use-after-free in the holder.
+func TestEICTunnelPool_RefcountBlocksEviction(t *testing.T) {
+	setupCount, cleanupCount, _, _ := withPoolSetupStub(t)
+	poolTTL = 30 * time.Millisecond
+	poolJanitorInterval = 5 * time.Millisecond
+	pool := freshPool(t)
+	ctx := context.Background()
+
+	_, done, err := pool.acquire(ctx, "i-hold")
+	if err != nil {
+		t.Fatalf("acquire: %v", err)
+	}
+
+	// Sleep past TTL while holding the session. Janitor sweeps
+	// every 5ms but must skip our entry because refcount=1.
+	time.Sleep(80 * time.Millisecond)
+
+	if got := atomic.LoadInt64(cleanupCount); got != 0 {
+		t.Errorf("expected 0 cleanups while holder is active, got %d", got)
+	}
+
+	done(false)
+	// Now refcount=0 and entry is past TTL; releaser triggers cleanup.
+	deadline := time.Now().Add(200 * time.Millisecond)
+	for atomic.LoadInt64(cleanupCount) < 1 && time.Now().Before(deadline) {
+		time.Sleep(5 * time.Millisecond)
+	}
+	if got := atomic.LoadInt64(cleanupCount); got != 1 {
+		t.Errorf("expected 1 cleanup after release of expired entry, got %d", got)
+	}
+	if got := atomic.LoadInt64(setupCount); got != 1 {
+		t.Errorf("setupCount tracking: got %d, want 1", got)
+	}
+}
+
+// TestPooledWithEICTunnel_PanicPoisonsEntry pins that a panic
+// from fn poisons the pool entry on the way out — refcount goes
+// back to zero (no leak) and the entry is marked unusable so the
+// next acquire builds fresh. Without the defer-release pattern, a
+// panic would leave refcount=1 forever and the entry would never
+// evict.
+func TestPooledWithEICTunnel_PanicPoisonsEntry(t *testing.T) {
+	setupCount, _, _, _ := withPoolSetupStub(t)
+	poolTTL = 50 * time.Second
+	globalEICTunnelPool = newEICTunnelPool()
+	t.Cleanup(globalEICTunnelPool.stop)
+
+	func() {
+		defer func() {
+			if r := recover(); r == nil {
+				t.Errorf("expected panic to bubble up, got nil")
+			}
+		}()
+		_ = pooledWithEICTunnel(context.Background(), "i-panic",
+			func(s eicSSHSession) error { panic("boom") })
+	}()
+
+	// Replenish the gate so the next setup can run.
+	prev := poolSetupTunnel
+	poolSetupTunnel = func(ctx context.Context, instanceID string) (
+		eicSSHSession, func(), error) {
+		atomic.AddInt64(setupCount, 1)
+		return eicSSHSession{instanceID: instanceID}, func() {}, nil
+	}
+	t.Cleanup(func() { poolSetupTunnel = prev })
+
+	// Next acquire must build fresh — entry was poisoned by panic.
+	if err := pooledWithEICTunnel(context.Background(), "i-panic",
+		func(s eicSSHSession) error { return nil }); err != nil {
+		t.Fatalf("post-panic acquire: %v", err)
+	}
+	if got := atomic.LoadInt64(setupCount); got != 2 {
+		t.Errorf("expected 2 setups (panic poisoned, rebuild), got %d", got)
+	}
+}
+
+// TestPooledWithEICTunnel_PreservesFnErr pins that errors from the
+// inner fn pass through to the caller verbatim — pool wrapping
+// should not swallow or transform error semantics for app code.
+func TestPooledWithEICTunnel_PreservesFnErr(t *testing.T) {
+	withPoolSetupStub(t)
+	poolTTL = 50 * time.Second
+
+	// Reset the global pool so this test is isolated from any prior
+	// test that may have populated it.
+	globalEICTunnelPool = newEICTunnelPool()
+
+	want := errors.New("file does not exist")
+	got := pooledWithEICTunnel(context.Background(), "i-fn-err",
+		func(s eicSSHSession) error { return want })
+	if !errors.Is(got, want) {
+		t.Errorf("pooledWithEICTunnel returned %v, want %v", got, want)
+	}
+}

workspace-server/internal/handlers/local_e2e_dev_dept_test.go

@@ -0,0 +1,375 @@
+package handlers
+
+import (
+	"archive/tar"
+	"bytes"
+	"net"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"strings"
+	"testing"
+	"time"
+
+	"gopkg.in/yaml.v3"
+)
+
+// Local E2E for the dev-department extraction (RFC internal#77).
+//
+// Pre-conditions: both repos cloned as siblings under
+// /tmp/local-e2e-deploy/{molecule-dev, molecule-dev-department}.
+// (Set up by the orchestrator before running this test.)
+//
+// What this proves end-to-end through real platform code:
+//   1. resolveYAMLIncludes follows the dev-lead symlink at the parent's
+//      template root and pulls in the dev-department subtree.
+//   2. Recursive !include's inside the symlinked subtree resolve
+//      correctly via the chain dev-lead/workspace.yaml →
+//      ./core-lead/workspace.yaml → ./core-be/workspace.yaml etc.
+//   3. The resolved YAML unmarshals into a complete OrgTemplate with the
+//      expected count of workspaces (parent's PM+Marketing+Research +
+//      dev-department's atomized 28 workspaces).
+//
+// Skipped if the local-e2e-deploy fixture isn't present — won't block
+// CI on hosts that haven't set it up.
+func TestLocalE2E_DevDepartmentExtraction(t *testing.T) {
+	parent := "/tmp/local-e2e-deploy/molecule-dev"
+	if _, err := os.Stat(filepath.Join(parent, "org.yaml")); err != nil {
+		t.Skipf("local-e2e fixture not present at %s: %v", parent, err)
+	}
+
+	orgYAML, err := os.ReadFile(filepath.Join(parent, "org.yaml"))
+	if err != nil {
+		t.Fatalf("read org.yaml: %v", err)
+	}
+
+	expanded, err := resolveYAMLIncludes(orgYAML, parent)
+	if err != nil {
+		t.Fatalf("resolveYAMLIncludes failed: %v", err)
+	}
+
+	var tmpl OrgTemplate
+	if err := yaml.Unmarshal(expanded, &tmpl); err != nil {
+		t.Fatalf("unmarshal expanded OrgTemplate: %v", err)
+	}
+
+	// Walk the full workspace tree, collect names.
+	names := []string{}
+	var walk func([]OrgWorkspace)
+	walk = func(ws []OrgWorkspace) {
+		for _, w := range ws {
+			names = append(names, w.Name)
+			walk(w.Children)
+		}
+	}
+	walk(tmpl.Workspaces)
+
+	t.Logf("org name: %q", tmpl.Name)
+	t.Logf("total workspaces (recursive): %d", len(names))
+	for _, n := range names {
+		t.Logf("  - %q", n)
+	}
+
+	// Expected: PM + Marketing Lead + Dev Lead at top level, plus the
+	// full sub-trees under each. After atomization, we expect:
+	//   - PM tree: PM + Research Lead + 3 research roles = 5
+	//   - Marketing tree: Marketing Lead + 5 marketing roles = 6
+	//   - Dev Lead tree: Dev Lead + (5 sub-team leads × ~6 each) +
+	//     3 floaters + Triage Operator = ~32
+	// Roughly ~43 total. Be liberal; just assert a floor.
+	if len(names) < 30 {
+		t.Errorf("workspace count too low (%d) — expected ~40+ (PM+Marketing+Dev tree)", len(names))
+	}
+
+	// Specific sentinel names we expect to find:
+	expected := []string{
+		"PM",
+		"Marketing Lead",
+		"Dev Lead",
+		"Core Platform Lead",
+		"Controlplane Lead",
+		"App & Docs Lead",
+		"Infra Lead",
+		"SDK Lead",
+		"Documentation Specialist", // Q1 — should be under app-lead
+		"Triage Operator",          // Q2 — should be under dev-lead
+	}
+	found := map[string]bool{}
+	for _, n := range names {
+		found[n] = true
+	}
+	for _, want := range expected {
+		if !found[want] {
+			t.Errorf("missing expected workspace %q", want)
+		}
+	}
+}
+
+// Stage-2 of the local e2e: prove every resolved workspace's `files_dir`
+// path actually consumes correctly through the rest of the import chain.
+// resolveYAMLIncludes returning a populated OrgTemplate is necessary but
+// not sufficient — `POST /org/import` then does:
+//
+//   1. resolveInsideRoot(orgBaseDir, ws.FilesDir) → must return a path
+//      that exists and stat-resolves to a directory (org_import.go:313-317).
+//   2. CopyTemplateToContainer(ctx, containerID, templatePath) → walks
+//      the dir with filepath.Walk and tars its contents into the
+//      workspace's /configs/ mount (provisioner.go:766-820).
+//
+// This stage-2 test exercises both #1 and #2 against every workspace in
+// the resolved tree, mimicking what the platform does post-include-
+// resolution. Catches: files_dir paths that don't resolve through the
+// symlink, paths that exist but are empty (silently produces empty
+// /configs/), or filepath.Walk failing to descend through cross-repo
+// symlink boundaries.
+func TestLocalE2E_FilesDirConsumption(t *testing.T) {
+	parent := "/tmp/local-e2e-deploy/molecule-dev"
+	if _, err := os.Stat(filepath.Join(parent, "org.yaml")); err != nil {
+		t.Skipf("local-e2e fixture not present at %s: %v", parent, err)
+	}
+
+	orgYAML, err := os.ReadFile(filepath.Join(parent, "org.yaml"))
+	if err != nil {
+		t.Fatalf("read org.yaml: %v", err)
+	}
+	expanded, err := resolveYAMLIncludes(orgYAML, parent)
+	if err != nil {
+		t.Fatalf("resolveYAMLIncludes: %v", err)
+	}
+	var tmpl OrgTemplate
+	if err := yaml.Unmarshal(expanded, &tmpl); err != nil {
+		t.Fatalf("unmarshal: %v", err)
+	}
+
+	// Flatten every workspace — including children, grandchildren, etc.
+	flat := []OrgWorkspace{}
+	var walk func([]OrgWorkspace)
+	walk = func(ws []OrgWorkspace) {
+		for _, w := range ws {
+			flat = append(flat, w)
+			walk(w.Children)
+		}
+	}
+	walk(tmpl.Workspaces)
+
+	checked := 0
+	for _, w := range flat {
+		if w.FilesDir == "" {
+			continue // workspace declared inline (no files_dir) — skip
+		}
+		checked++
+		t.Run(w.Name+"/"+w.FilesDir, func(t *testing.T) {
+			// Step 1: resolveInsideRoot returns a path that's-inside-root.
+			abs, err := resolveInsideRoot(parent, w.FilesDir)
+			if err != nil {
+				t.Fatalf("resolveInsideRoot(%q, %q): %v", parent, w.FilesDir, err)
+			}
+			info, err := os.Stat(abs)
+			if err != nil {
+				t.Fatalf("stat %q (resolved from files_dir %q): %v", abs, w.FilesDir, err)
+			}
+			if !info.IsDir() {
+				t.Fatalf("files_dir %q resolved to %q which is not a directory", w.FilesDir, abs)
+			}
+
+			// Step 2: walk the dir like CopyTemplateToContainer does.
+			// Mirror the platform's symlink-resolution at the root —
+			// filepath.Walk doesn't descend into a symlink leaf, so
+			// CopyTemplateToContainer (provisioner.go) calls
+			// EvalSymlinks on templatePath first. Replicate exactly.
+			if resolved, err := filepath.EvalSymlinks(abs); err == nil {
+				abs = resolved
+			}
+			var buf bytes.Buffer
+			tw := tar.NewWriter(&buf)
+			fileCount := 0
+			fileNames := []string{}
+			err = filepath.Walk(abs, func(path string, info os.FileInfo, err error) error {
+				if err != nil {
+					return err
+				}
+				rel, err := filepath.Rel(abs, path)
+				if err != nil {
+					return err
+				}
+				if rel == "." {
+					return nil
+				}
+				header, _ := tar.FileInfoHeader(info, "")
+				header.Name = rel
+				if err := tw.WriteHeader(header); err != nil {
+					return err
+				}
+				if !info.IsDir() {
+					fileCount++
+					fileNames = append(fileNames, rel)
+					data, err := os.ReadFile(path)
+					if err != nil {
+						return err
+					}
+					header.Size = int64(len(data))
+					tw.Write(data)
+				}
+				return nil
+			})
+			if err != nil {
+				t.Fatalf("filepath.Walk %q (mimics CopyTemplateToContainer): %v", abs, err)
+			}
+			tw.Close()
+
+			if fileCount == 0 {
+				t.Errorf("files_dir %q at %q is empty — CopyTemplateToContainer would produce empty /configs/",
+					w.FilesDir, abs)
+			}
+
+			// Sanity: every workspace folder should have AT LEAST one of
+			// {workspace.yaml, system-prompt.md, initial-prompt.md} —
+			// these are the markers a workspace folder is recognizable
+			// as a workspace (mirrors validator's WORKSPACE_FOLDER_MARKERS).
+			markers := []string{"workspace.yaml", "system-prompt.md", "initial-prompt.md"}
+			hasMarker := false
+			for _, name := range fileNames {
+				for _, m := range markers {
+					if name == m || strings.HasSuffix(name, "/"+m) {
+						hasMarker = true
+						break
+					}
+				}
+				if hasMarker {
+					break
+				}
+			}
+			if !hasMarker {
+				t.Errorf("files_dir %q at %q has %d files but none of the workspace markers %v — found: %v",
+					w.FilesDir, abs, fileCount, markers, fileNames)
+			}
+		})
+	}
+	t.Logf("checked %d workspaces with files_dir", checked)
+	if checked < 25 {
+		t.Errorf("expected ~28 workspaces with files_dir (post-atomization); only saw %d", checked)
+	}
+}
+
+// PR-C from the Phase 3a phasing (task #234): real-Gitea e2e for the
+// !external resolver against the LIVE molecule-ai/molecule-dev-department
+// repo. Verifies the production gitFetcher fetches the dev tree and the
+// resolver grafts it correctly into a parent template that has NO
+// symlink — composition is purely platform-side.
+//
+// Skipped if Gitea isn't reachable (offline / firewall / CI without
+// network). Requires `git` binary on PATH.
+func TestLocalE2E_ExternalDevDepartment(t *testing.T) {
+	if _, err := exec.LookPath("git"); err != nil {
+		t.Skipf("git binary not found: %v", err)
+	}
+
+	// Skip if Gitea host isn't reachable (TCP probe). Avoids network-
+	// dependent tests failing on offline runners.
+	conn, err := net.DialTimeout("tcp", "git.moleculesai.app:443", 3*time.Second)
+	if err != nil {
+		t.Skipf("git.moleculesai.app:443 unreachable: %v", err)
+	}
+	conn.Close()
+
+	// Build a minimal parent template inline — no need for the
+	// /tmp/local-e2e-deploy/ symlinked fixture. The whole point of
+	// !external is that the parent template is self-contained;
+	// composition resolves over the network at import time.
+	parent := t.TempDir()
+
+	orgYAML := []byte(`name: External-Only Test Parent
+description: Parent template that pulls the entire dev tree via !external.
+defaults:
+  runtime: claude-code
+  tier: 2
+workspaces:
+  - !external
+    repo: molecule-ai/molecule-dev-department
+    ref: main
+    path: dev-lead/workspace.yaml
+`)
+	if err := os.WriteFile(filepath.Join(parent, "org.yaml"), orgYAML, 0o644); err != nil {
+		t.Fatalf("write org.yaml: %v", err)
+	}
+
+	out, err := resolveYAMLIncludes(orgYAML, parent)
+	if err != nil {
+		t.Fatalf("resolveYAMLIncludes (!external against live Gitea): %v", err)
+	}
+
+	var tmpl OrgTemplate
+	if err := yaml.Unmarshal(out, &tmpl); err != nil {
+		t.Fatalf("unmarshal: %v", err)
+	}
+
+	// Walk the workspace tree, collect names + check files_dir paths.
+	flat := []OrgWorkspace{}
+	var walk func([]OrgWorkspace)
+	walk = func(ws []OrgWorkspace) {
+		for _, w := range ws {
+			flat = append(flat, w)
+			walk(w.Children)
+		}
+	}
+	walk(tmpl.Workspaces)
+
+	t.Logf("workspaces resolved through !external: %d", len(flat))
+	if len(flat) < 25 {
+		t.Errorf("expected ~28 dev-tree workspaces via !external; got %d", len(flat))
+	}
+
+	// Sentinel checks — same as TestLocalE2E_DevDepartmentExtraction
+	// (Q1+Q2 placements verified).
+	expected := []string{
+		"Dev Lead",
+		"Core Platform Lead",
+		"Controlplane Lead",
+		"App & Docs Lead",
+		"Documentation Specialist", // Q1
+		"Triage Operator",          // Q2
+	}
+	found := map[string]bool{}
+	for _, w := range flat {
+		found[w.Name] = true
+	}
+	for _, want := range expected {
+		if !found[want] {
+			t.Errorf("missing expected workspace %q", want)
+		}
+	}
+
+	// Every workspace's files_dir must be cache-prefixed (proves the
+	// path-rewrite ran end-to-end).
+	cachePrefix := ".external-cache"
+	for _, w := range flat {
+		if w.FilesDir == "" {
+			continue
+		}
+		if !strings.HasPrefix(w.FilesDir, cachePrefix) {
+			t.Errorf("workspace %q files_dir %q missing cache prefix %q", w.Name, w.FilesDir, cachePrefix)
+		}
+	}
+
+	// Verify the fetched cache exists and resolveInsideRoot accepts
+	// every workspace's files_dir (would cause provisioning to fail
+	// if not).
+	for _, w := range flat {
+		if w.FilesDir == "" {
+			continue
+		}
+		abs, err := resolveInsideRoot(parent, w.FilesDir)
+		if err != nil {
+			t.Errorf("workspace %q files_dir %q: resolveInsideRoot: %v", w.Name, w.FilesDir, err)
+			continue
+		}
+		info, err := os.Stat(abs)
+		if err != nil {
+			t.Errorf("workspace %q: stat %q: %v", w.Name, abs, err)
+			continue
+		}
+		if !info.IsDir() {
+			t.Errorf("workspace %q files_dir %q is not a directory", w.Name, w.FilesDir)
+		}
+	}
+}

workspace-server/internal/handlers/org.go

@@ -13,12 +13,15 @@ import (
 	"path/filepath"
 	"strconv"
 	"strings"
+	"time"
 
 	"github.com/Molecule-AI/molecule-monorepo/platform/internal/channels"
+	"github.com/Molecule-AI/molecule-monorepo/platform/internal/db"
 	"github.com/Molecule-AI/molecule-monorepo/platform/internal/events"
 	"github.com/Molecule-AI/molecule-monorepo/platform/internal/models"
 	"github.com/Molecule-AI/molecule-monorepo/platform/internal/provisioner"
 	"github.com/gin-gonic/gin"
+	"github.com/lib/pq"
 	"gopkg.in/yaml.v3"
 )
 
@@ -422,6 +425,16 @@ type OrgWorkspace struct {
 	Tier     int    `yaml:"tier" json:"tier"`
 	Template string `yaml:"template" json:"template"`
 	FilesDir string `yaml:"files_dir" json:"files_dir"`
+	// Spawning gates whether this workspace (AND its descendants) gets
+	// provisioned during /org/import. Pointer so we can distinguish
+	// "explicitly set to false" from "unset" (default = spawn). Use case:
+	// the dev-tree org template declares the full team structure but a
+	// developer's local machine only has RAM for a subset; setting
+	// spawning: false on a leaf or a sub-tree root skips that branch
+	// entirely without editing the canonical template structure.
+	// Counted in countWorkspaces same as actual; subtree-skip happens
+	// at provision time in createWorkspaceTree.
+	Spawning *bool `yaml:"spawning,omitempty" json:"spawning,omitempty"`
 	// SystemPrompt is an inline override. Normally each role's system-prompt.md
 	// lives at `<files_dir>/system-prompt.md` and is copied via the files_dir
 	// template-copy step; inline overrides that path for ad-hoc workspaces.
@@ -558,11 +571,30 @@ func (h *OrgHandler) Import(c *gin.Context) {
 	var body struct {
 		Dir      string      `json:"dir"`      // org template directory name
 		Template OrgTemplate `json:"template"` // or inline template
+		// Mode controls cleanup behavior of pre-existing workspaces:
+		//   ""        / "merge"     — additive (default; current behavior).
+		//                              Existing workspaces matched by
+		//                              (parent_id, name) are skipped; nothing
+		//                              outside the new tree is touched.
+		//   "reconcile"             — additive + cleanup. After import, any
+		//                              online workspace whose name matches an
+		//                              imported workspace's name but whose id
+		//                              isn't in the import result set is
+		//                              cascade-deleted. Catches "previous
+		//                              import survived a re-import" zombies
+		//                              (the 20:13→21:17 dev-tree case).
+		Mode string `json:"mode"`
 	}
 	if err := c.ShouldBindJSON(&body); err != nil {
 		c.JSON(http.StatusBadRequest, gin.H{"error": "invalid request body"})
 		return
 	}
+	importStart := time.Now()
+	emitOrgEvent(c.Request.Context(), "org.import.started", map[string]any{
+		"name": body.Template.Name,
+		"dir":  body.Dir,
+		"mode": body.Mode,
+	})
 
 	var tmpl OrgTemplate
 	var orgBaseDir string // base directory for files_dir resolution
@@ -708,18 +740,171 @@ func (h *OrgHandler) Import(c *gin.Context) {
 		}
 	}
 
+	// Reconcile mode: prune workspaces present from a previous import that
+	// share a name with the new tree but are NOT in the new result set.
+	// Catches the additive-import bug where re-running /org/import with a
+	// changed tree shape (different parent_id for the same role name) leaves
+	// the prior workspace online — visible to the canvas, consuming
+	// containers, and looking like a duplicate. Default mode "" / "merge"
+	// preserves the old additive behavior.
+	reconcileRemovedCount := 0
+	reconcileSkipped := 0
+	reconcileErrs := []string{}
+	if body.Mode == "reconcile" && createErr == nil {
+		ctx := c.Request.Context()
+		importedNames := []string{}
+		walkOrgWorkspaceNames(tmpl.Workspaces, &importedNames)
+
+		importedIDs := make([]string, 0, len(results))
+		for _, r := range results {
+			if id, ok := r["id"].(string); ok && id != "" {
+				importedIDs = append(importedIDs, id)
+			}
+		}
+
+		// Empty-set guards: if the import didn't produce any names or any
+		// IDs, skip — querying with empty arrays would either match
+		// nothing (harmless) or, worse, match every workspace if a future
+		// query rewrite drops the IN clause. Belt-and-suspenders.
+		if len(importedNames) > 0 && len(importedIDs) > 0 {
+			rows, err := db.DB.QueryContext(ctx, `
+				SELECT id FROM workspaces
+				WHERE name = ANY($1::text[])
+				  AND id != ALL($2::uuid[])
+				  AND status != 'removed'
+			`, pq.Array(importedNames), pq.Array(importedIDs))
+			if err != nil {
+				log.Printf("Org import reconcile: orphan query failed: %v", err)
+				reconcileErrs = append(reconcileErrs, fmt.Sprintf("orphan query: %v", err))
+			} else {
+				orphanIDs := []string{}
+				for rows.Next() {
+					var orphanID string
+					if rows.Scan(&orphanID) == nil {
+						orphanIDs = append(orphanIDs, orphanID)
+					}
+				}
+				rows.Close()
+
+				for _, oid := range orphanIDs {
+					cascadeCount, stopErrs, err := h.workspace.CascadeDelete(ctx, oid)
+					if err != nil {
+						log.Printf("Org import reconcile: CascadeDelete(%s) failed: %v", oid, err)
+						reconcileErrs = append(reconcileErrs, fmt.Sprintf("delete %s: %v", oid, err))
+						reconcileSkipped++
+						continue
+					}
+					reconcileRemovedCount += 1 + cascadeCount
+					if len(stopErrs) > 0 {
+						log.Printf("Org import reconcile: %s had %d stop errors (orphan sweeper will retry)", oid, len(stopErrs))
+					}
+				}
+				log.Printf("Org import reconcile: %d orphans removed (%d cascade descendants), %d skipped", len(orphanIDs), reconcileRemovedCount-len(orphanIDs), reconcileSkipped)
+			}
+		}
+	}
+
 	status := http.StatusCreated
 	resp := gin.H{
 		"org":        tmpl.Name,
 		"workspaces": results,
 		"count":      len(results),
 	}
+	if body.Mode == "reconcile" {
+		resp["mode"] = "reconcile"
+		resp["reconcile_removed_count"] = reconcileRemovedCount
+		if len(reconcileErrs) > 0 {
+			resp["reconcile_errors"] = reconcileErrs
+		}
+	}
 	if createErr != nil {
 		status = http.StatusMultiStatus
 		resp["error"] = createErr.Error()
 	}
 
-	log.Printf("Org import: %s — %d workspaces created", tmpl.Name, len(results))
+	// results contains both freshly-created AND lookupExistingChild skips
+	// (entries with "skipped":true). Splitting the count here so the audit
+	// row reflects "what changed" vs "what was already there" — telemetry
+	// readers shouldn't need to grep stdout to tell an idempotent re-run
+	// apart from a fresh-create.
+	createdCount, skippedCount := 0, 0
+	for _, r := range results {
+		if skipped, _ := r["skipped"].(bool); skipped {
+			skippedCount++
+		} else {
+			createdCount++
+		}
+	}
+	log.Printf("Org import: %s — %d created, %d skipped, %d reconciled",
+		tmpl.Name, createdCount, skippedCount, reconcileRemovedCount)
+	emitOrgEvent(c.Request.Context(), "org.import.completed", map[string]any{
+		"name":                    tmpl.Name,
+		"dir":                     body.Dir,
+		"mode":                    body.Mode,
+		"created_count":           createdCount,
+		"skipped_count":           skippedCount,
+		"reconcile_removed_count": reconcileRemovedCount,
+		"reconcile_errors":        len(reconcileErrs),
+		"duration_ms":             time.Since(importStart).Milliseconds(),
+		"create_error":            errString(createErr),
+	})
 	c.JSON(status, resp)
 }
 
+// walkOrgWorkspaceNames collects every Name in the tree (in any order) into
+// names. Used by reconcile to detect orphan workspaces — workspaces with the
+// same role name as a freshly-imported one but a different id, surviving from
+// a prior import.
+func walkOrgWorkspaceNames(workspaces []OrgWorkspace, names *[]string) {
+	for _, w := range workspaces {
+		// spawning:false subtrees are still part of the imported tree
+		// from a logical-tree perspective — DON'T skip the recursion,
+		// or reconcile would orphan the rest of the subtree on every
+		// re-import where spawning is toggled. Names of skipped
+		// workspaces remain registered so reconcile won't double-create
+		// them when spawning flips back to true.
+		if w.Name != "" {
+			*names = append(*names, w.Name)
+		}
+		walkOrgWorkspaceNames(w.Children, names)
+	}
+}
+
+// emitOrgEvent records an org-lifecycle event in structure_events so the
+// import history is queryable independent of stdout log retention. Errors
+// are logged and swallowed — never block the request path on telemetry.
+//
+// Event-type taxonomy (extend by appending; never rename):
+//
+//	org.import.started        — handler entered, request body parsed
+//	org.import.completed      — handler exiting (success or partial)
+//	org.import.failed         — handler exiting with an unrecoverable error
+//
+// payload fields are documented at each call site.
+func emitOrgEvent(ctx context.Context, eventType string, payload map[string]any) {
+	if payload == nil {
+		payload = map[string]any{}
+	}
+	payloadJSON, err := json.Marshal(payload)
+	if err != nil {
+		log.Printf("emitOrgEvent: marshal %s payload failed: %v", eventType, err)
+		return
+	}
+	if _, err := db.DB.ExecContext(ctx, `
+		INSERT INTO structure_events (event_type, payload, created_at)
+		VALUES ($1, $2, now())
+	`, eventType, payloadJSON); err != nil {
+		log.Printf("emitOrgEvent: insert %s failed: %v", eventType, err)
+	}
+}
+
+// errString returns "" for a nil error, err.Error() otherwise. Lets us put
+// nullable error strings in event payloads without checking for nil at every
+// call site.
+func errString(err error) string {
+	if err == nil {
+		return ""
+	}
+	return err.Error()
+}
+

workspace-server/internal/handlers/org_external.go

@@ -0,0 +1,439 @@
+package handlers
+
+import (
+	"context"
+	"fmt"
+	"net/url"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"regexp"
+	"strings"
+	"time"
+
+	"gopkg.in/yaml.v3"
+)
+
+// External-ref resolver — gitops-style cross-repo subtree composition.
+// Internal#77 RFC, Phase 3a (task #222). Prior art: Helm subcharts +
+// dependency cache, Kustomize remote bases, Terraform module sources.
+//
+// Schema (a `!external`-tagged mapping anywhere a workspace entry is
+// allowed — workspaces:, roots:, children:):
+//
+//   - !external
+//     repo: molecule-ai/molecule-dev-department
+//     ref: main
+//     path: dev-lead/workspace.yaml
+//
+// At resolve time, the platform fetches the repo at ref into a content-
+// addressable cache under <rootDir>/.external-cache/<repo>/<sha>/, loads
+// the yaml at <cacheDir>/<path>, rewrites every files_dir + relative
+// !include path to be cache-prefixed, then grafts the result in place of
+// the !external node. Downstream pipeline (resolveInsideRoot, plugin
+// merge, CopyTemplateToContainer) sees ordinary in-tree paths.
+
+// ExternalRef is the deserialized form of an `!external`-tagged mapping.
+type ExternalRef struct {
+	Repo string `yaml:"repo"`
+	Ref  string `yaml:"ref"`
+	Path string `yaml:"path"`
+
+	// URL overrides the default Gitea host. Optional; defaults to
+	// MOLECULE_EXTERNAL_GITEA_URL env or git.moleculesai.app.
+	URL string `yaml:"url,omitempty"`
+}
+
+const (
+	// maxExternalDepth caps recursion through nested `!external`s. Lower
+	// than maxIncludeDepth (16) because each level may issue a network
+	// fetch. Composition that genuinely needs >4 layers is a smell.
+	maxExternalDepth = 4
+
+	// externalCacheDirName is the per-template cache subdir under rootDir.
+	// Content-addressable: keyed by (repo, sha). Operators add this to
+	// .gitignore — cache is platform-mutated, not source-tracked.
+	externalCacheDirName = ".external-cache"
+
+	// gitFetchTimeout caps a single clone operation. Conservative —
+	// org template fetches are typically <100KB.
+	gitFetchTimeout = 60 * time.Second
+)
+
+// safeRefPattern restricts `ref` values to characters git itself accepts
+// for branch / tag / SHA. Belt-and-braces over git's own validation.
+var safeRefPattern = regexp.MustCompile(`^[a-zA-Z0-9_./-]+$`)
+
+// allowlistedHostPath returns true if `<host>/<repo>` matches the
+// configured allowlist. Default allowlist: git.moleculesai.app/molecule-ai/.
+// Override via MOLECULE_EXTERNAL_REPO_ALLOWLIST env var (comma-separated
+// patterns). Patterns are matched as prefixes (with trailing-slash
+// semantics) or as exact matches. Trailing /* is treated as "any
+// descendants of this prefix".
+//
+// Examples:
+//   - "git.moleculesai.app/molecule-ai/" → matches molecule-ai/* (any repo)
+//   - "git.moleculesai.app/molecule-ai/*" → same; trailing /* normalized to /
+//   - "git.moleculesai.app/molecule-ai/molecule-dev-department" → exact
+//   - "git.moleculesai.app/" → matches everything on that host
+func allowlistedHostPath(host, repoPath string) bool {
+	allow := os.Getenv("MOLECULE_EXTERNAL_REPO_ALLOWLIST")
+	if allow == "" {
+		allow = "git.moleculesai.app/molecule-ai/"
+	}
+	hp := host + "/" + repoPath
+	for _, pat := range strings.Split(allow, ",") {
+		pat = strings.TrimSpace(pat)
+		if pat == "" {
+			continue
+		}
+		// Normalize trailing /* → /
+		pat = strings.TrimSuffix(pat, "*")
+		if pat == hp {
+			return true
+		}
+		if strings.HasSuffix(pat, "/") && strings.HasPrefix(hp+"/", pat) {
+			return true
+		}
+	}
+	return false
+}
+
+// externalFetcher abstracts the git-clone-into-cache step. Production
+// uses gitFetcher (shells out to git); tests inject a fake that
+// pre-stages content in a temp dir.
+type externalFetcher interface {
+	// Fetch ensures rootDir/.external-cache/<safe-repo>/<sha>/ contains
+	// the repo content at the given ref. Returns the absolute cache
+	// dir + the resolved SHA. Cache hit = no network. Cache miss =
+	// clone.
+	Fetch(ctx context.Context, rootDir, host, repoPath, ref string) (cacheDir, sha string, err error)
+}
+
+// defaultExternalFetcher is the package-level fetcher injection point.
+// Production code uses the git-shell fetcher; tests override via
+// SetExternalFetcherForTest.
+var defaultExternalFetcher externalFetcher = &gitFetcher{}
+
+// SetExternalFetcherForTest swaps the fetcher for testing. Returns a
+// cleanup func that restores the previous fetcher.
+func SetExternalFetcherForTest(f externalFetcher) func() {
+	prev := defaultExternalFetcher
+	defaultExternalFetcher = f
+	return func() { defaultExternalFetcher = prev }
+}
+
+// resolveExternalMapping replaces an `!external`-tagged mapping node
+// with the loaded + path-rewritten yaml content from the fetched repo.
+//
+// `currentDir` and `rootDir` are inherited from expandNode's resolve
+// frame. `visited` tracks (repo, sha, path) tuples for cycle detection
+// across nested externals.
+func resolveExternalMapping(n *yaml.Node, currentDir, rootDir string, visited map[string]bool, depth int) error {
+	if depth > maxExternalDepth {
+		return fmt.Errorf("!external: max depth %d exceeded (possible cycle)", maxExternalDepth)
+	}
+	if rootDir == "" {
+		return fmt.Errorf("!external at line %d requires a dir-based org template (no rootDir in inline-template mode)", n.Line)
+	}
+
+	var ref ExternalRef
+	if err := n.Decode(&ref); err != nil {
+		return fmt.Errorf("!external at line %d: decode: %w", n.Line, err)
+	}
+	if ref.Repo == "" || ref.Ref == "" || ref.Path == "" {
+		return fmt.Errorf("!external at line %d: repo, ref, path are all required (got %+v)", n.Line, ref)
+	}
+	if !safeRefPattern.MatchString(ref.Ref) {
+		return fmt.Errorf("!external at line %d: ref %q contains disallowed characters", n.Line, ref.Ref)
+	}
+	// Defense-in-depth: even though git itself rejects refs containing
+	// `..`, the regex above currently allows them. Reject explicitly.
+	if strings.Contains(ref.Ref, "..") {
+		return fmt.Errorf("!external at line %d: ref %q must not contain '..'", n.Line, ref.Ref)
+	}
+	if strings.Contains(ref.Path, "..") || strings.HasPrefix(ref.Path, "/") {
+		return fmt.Errorf("!external at line %d: path %q must be relative-and-down-only", n.Line, ref.Path)
+	}
+
+	host := ref.URL
+	if host == "" {
+		host = os.Getenv("MOLECULE_EXTERNAL_GITEA_URL")
+	}
+	if host == "" {
+		host = "git.moleculesai.app"
+	}
+	host = strings.TrimPrefix(strings.TrimPrefix(host, "https://"), "http://")
+	host = strings.TrimSuffix(host, "/")
+
+	if !allowlistedHostPath(host, ref.Repo) {
+		return fmt.Errorf("!external at line %d: %s/%s not in MOLECULE_EXTERNAL_REPO_ALLOWLIST", n.Line, host, ref.Repo)
+	}
+
+	ctx, cancel := context.WithTimeout(context.Background(), gitFetchTimeout)
+	defer cancel()
+
+	cacheDir, sha, err := defaultExternalFetcher.Fetch(ctx, rootDir, host, ref.Repo, ref.Ref)
+	if err != nil {
+		return fmt.Errorf("!external at line %d: fetch %s/%s@%s: %w", n.Line, host, ref.Repo, ref.Ref, err)
+	}
+
+	// Cycle key: (repo, sha, path) — same external content reachable
+	// via two paths is fine, but a self-referential cycle isn't.
+	cycleKey := fmt.Sprintf("%s/%s@%s/%s", host, ref.Repo, sha, ref.Path)
+	if visited[cycleKey] {
+		return fmt.Errorf("!external cycle detected at %q (line %d)", cycleKey, n.Line)
+	}
+
+	// Validate path resolves inside the cache dir (anti-traversal).
+	yamlPathAbs, err := resolveInsideRoot(cacheDir, ref.Path)
+	if err != nil {
+		return fmt.Errorf("!external at line %d: path %q: %w", n.Line, ref.Path, err)
+	}
+	if _, err := os.Stat(yamlPathAbs); err != nil {
+		return fmt.Errorf("!external at line %d: %s/%s@%s does not contain %q: %w", n.Line, host, ref.Repo, sha, ref.Path, err)
+	}
+
+	data, err := os.ReadFile(yamlPathAbs)
+	if err != nil {
+		return fmt.Errorf("!external at line %d: read %q: %w", n.Line, yamlPathAbs, err)
+	}
+
+	var sub yaml.Node
+	if err := yaml.Unmarshal(data, &sub); err != nil {
+		return fmt.Errorf("!external at line %d: parse %q: %w", n.Line, yamlPathAbs, err)
+	}
+	root := &sub
+	if root.Kind == yaml.DocumentNode && len(root.Content) == 1 {
+		root = root.Content[0]
+	}
+
+	// Recurse FIRST: load all nested !include / !external content into
+	// the tree. Then rewrite ALL files_dir scalars in the fully-resolved
+	// tree (top + nested) with the cache prefix in one pass. Doing
+	// rewrite-before-recurse would leave nested-loaded files_dir paths
+	// unprefixed.
+	visited[cycleKey] = true
+	defer delete(visited, cycleKey)
+
+	subDir := filepath.Dir(yamlPathAbs)
+	if err := expandNode(root, subDir, rootDir, visited, depth+1); err != nil {
+		return err
+	}
+
+	// Path rewrite: prefix every files_dir scalar in the fully-resolved
+	// content with the cache-relative-from-rootDir prefix. After this
+	// pass, fetched workspaces look like ordinary in-tree workspaces.
+	cachePrefix, err := filepath.Rel(rootDir, cacheDir)
+	if err != nil {
+		return fmt.Errorf("!external at line %d: cannot compute cache prefix: %w", n.Line, err)
+	}
+	rewriteFilesDir(root, cachePrefix)
+
+	// Replace the !external mapping with the resolved content in-place.
+	*n = *root
+	if n.Tag == "!external" {
+		n.Tag = ""
+	}
+	return nil
+}
+
+// rewriteFilesDir walks the yaml node tree and prepends cachePrefix to
+// every files_dir scalar value. Idempotent: if a files_dir value already
+// starts with the prefix, no-op.
+//
+// !include paths are intentionally NOT rewritten. They resolve relative
+// to their containing file's directory (subDir in expandNode), and after
+// fetch that directory IS inside the cache, so relative !include paths
+// Just Work without any rewrite. Rewriting them would double-prefix on
+// recursive resolution.
+//
+// files_dir DOES need rewriting because it's consumed at workspace-
+// provisioning time relative to orgBaseDir (the parent template's root),
+// not relative to the workspace.yaml's containing dir.
+func rewriteFilesDir(n *yaml.Node, cachePrefix string) {
+	if n == nil {
+		return
+	}
+	if n.Kind == yaml.MappingNode {
+		for i := 0; i+1 < len(n.Content); i += 2 {
+			key, value := n.Content[i], n.Content[i+1]
+			if key.Kind == yaml.ScalarNode && key.Value == "files_dir" && value.Kind == yaml.ScalarNode {
+				if !strings.HasPrefix(value.Value, cachePrefix+string(filepath.Separator)) && value.Value != cachePrefix {
+					value.Value = filepath.Join(cachePrefix, value.Value)
+				}
+			}
+		}
+	}
+	for _, child := range n.Content {
+		rewriteFilesDir(child, cachePrefix)
+	}
+}
+
+// safeRepoCacheDir converts a repo path like "molecule-ai/foo" into a
+// filesystem-safe segment "molecule-ai__foo". Avoids nesting cache dirs
+// (which would complicate cleanup).
+func safeRepoCacheDir(host, repoPath string) string {
+	hp := host + "/" + repoPath
+	hp = strings.ReplaceAll(hp, "/", "__")
+	hp = strings.ReplaceAll(hp, ":", "_")
+	return hp
+}
+
+// gitFetcher is the production externalFetcher: shells out to `git` to
+// clone the repo at ref into the cache dir. Cache key includes the
+// resolved SHA, so different SHAs of the same ref get different cache
+// dirs (no overwrite).
+//
+// Token handling — important for security. The auth token never enters
+// the clone URL (and therefore never lands in the cloned repo's
+// .git/config) and never appears in returned errors. We use git's
+// `http.extraHeader` config option (passed via `-c`), which sends an
+// Authorization header per-request without persisting it. The token is
+// briefly visible in the `git` process's argv (so other local users
+// with the same uid could see it via `ps`), which is the same exposure
+// it has via the env var that supplied it.
+//
+// Cache validity uses a `.complete` marker written after a successful
+// clone+rename. Cache-hit checks for the marker, not just the dir
+// existence — a partially-written cache (clone failed mid-way, or a
+// concurrent caller wrote a half-baked cache dir) is treated as cache
+// miss and re-fetched cleanly.
+type gitFetcher struct{}
+
+// cacheCompleteMarker is the filename written after a successful clone.
+// Cache-hit requires this marker; without it, the cache dir is treated
+// as partially-written and re-fetched.
+const cacheCompleteMarker = ".complete"
+
+// Fetch resolves ref → SHA via `git ls-remote`, then `git clone --depth=1`
+// if the cache dir is missing or incomplete. Auth via MOLECULE_GITEA_TOKEN
+// injected via http.extraHeader (never via URL).
+func (g *gitFetcher) Fetch(ctx context.Context, rootDir, host, repoPath, ref string) (string, string, error) {
+	cacheRoot := filepath.Join(rootDir, externalCacheDirName, safeRepoCacheDir(host, repoPath))
+	if err := os.MkdirAll(cacheRoot, 0o755); err != nil {
+		return "", "", fmt.Errorf("mkdir cache root: %w", err)
+	}
+
+	cloneURL := buildExternalCloneURL(host, repoPath)
+	gitArgs := func(extra ...string) []string {
+		args := authConfigArgs()
+		return append(args, extra...)
+	}
+
+	// 1. Resolve ref → SHA (so cache dir is content-addressable).
+	sha, err := g.resolveRefToSHA(ctx, cloneURL, ref, gitArgs)
+	if err != nil {
+		return "", "", fmt.Errorf("ls-remote: %s", redactToken(err.Error()))
+	}
+
+	cacheDir := filepath.Join(cacheRoot, sha)
+	// Cache-hit requires the .complete marker AND the .git dir.
+	// Without the marker, cache is partially-written → treat as miss.
+	if isCacheComplete(cacheDir) {
+		return cacheDir, sha, nil
+	}
+
+	// Cache miss or partially-written — clean any stale cacheDir before
+	// cloning (a previous broken attempt would otherwise block rename).
+	os.RemoveAll(cacheDir)
+
+	// 2. Clone into a sibling tmp dir; atomic rename on success.
+	tmpDir, err := os.MkdirTemp(cacheRoot, sha+".tmp.")
+	if err != nil {
+		return "", "", fmt.Errorf("mkdir tmp: %w", err)
+	}
+	// MkdirTemp creates the dir; git clone refuses to clone into a
+	// non-empty dir. Remove + recreate empty.
+	os.RemoveAll(tmpDir)
+	cloneAndConfig := append(gitArgs("clone", "--quiet", "--depth=1", "-b", ref, cloneURL, tmpDir))
+	cmd := exec.CommandContext(ctx, "git", cloneAndConfig...)
+	cmd.Env = append(os.Environ(), "GIT_TERMINAL_PROMPT=0")
+	if out, err := cmd.CombinedOutput(); err != nil {
+		os.RemoveAll(tmpDir)
+		return "", "", fmt.Errorf("git clone: %w: %s", err, redactToken(strings.TrimSpace(string(out))))
+	}
+
+	// Write the .complete marker BEFORE the rename. If rename succeeds,
+	// the marker is in place. If rename loses the race (concurrent
+	// fetcher won), our tmp gets cleaned up and we trust the winner.
+	if err := os.WriteFile(filepath.Join(tmpDir, cacheCompleteMarker), []byte(time.Now().UTC().Format(time.RFC3339)), 0o644); err != nil {
+		os.RemoveAll(tmpDir)
+		return "", "", fmt.Errorf("write complete marker: %w", err)
+	}
+
+	if err := os.Rename(tmpDir, cacheDir); err != nil {
+		// Race: another import beat us. Validate THEIR cache, accept it.
+		os.RemoveAll(tmpDir)
+		if isCacheComplete(cacheDir) {
+			return cacheDir, sha, nil
+		}
+		return "", "", fmt.Errorf("rename clone to cache (and winner's cache is incomplete): %w", err)
+	}
+	return cacheDir, sha, nil
+}
+
+// isCacheComplete reports whether cacheDir contains both the cloned
+// repo (.git) and the .complete marker. Treats partial state as miss.
+func isCacheComplete(cacheDir string) bool {
+	if _, err := os.Stat(filepath.Join(cacheDir, ".git")); err != nil {
+		return false
+	}
+	if _, err := os.Stat(filepath.Join(cacheDir, cacheCompleteMarker)); err != nil {
+		return false
+	}
+	return true
+}
+
+func (g *gitFetcher) resolveRefToSHA(ctx context.Context, cloneURL, ref string, gitArgs func(...string) []string) (string, error) {
+	args := gitArgs("ls-remote", cloneURL, ref)
+	cmd := exec.CommandContext(ctx, "git", args...)
+	cmd.Env = append(os.Environ(), "GIT_TERMINAL_PROMPT=0")
+	out, err := cmd.Output()
+	if err != nil {
+		return "", err
+	}
+	line := strings.TrimSpace(string(out))
+	if line == "" {
+		return "", fmt.Errorf("ref %q not found", ref)
+	}
+	// First whitespace-separated field is the SHA.
+	for i, ch := range line {
+		if ch == ' ' || ch == '\t' {
+			return line[:i], nil
+		}
+	}
+	return line, nil
+}
+
+// buildExternalCloneURL constructs the clone URL WITHOUT auth in userinfo.
+// Auth is layered on via authConfigArgs's http.extraHeader.
+func buildExternalCloneURL(host, repoPath string) string {
+	u := url.URL{Scheme: "https", Host: host, Path: "/" + repoPath + ".git"}
+	return u.String()
+}
+
+// authConfigArgs returns the `-c http.extraHeader=Authorization: token X`
+// args to pass to git, OR an empty slice if no token is set. The token
+// goes into the request headers (not the URL or .git/config), so it
+// doesn't persist on disk and doesn't appear in clone error output.
+func authConfigArgs() []string {
+	token := os.Getenv("MOLECULE_GITEA_TOKEN")
+	if token == "" {
+		return nil
+	}
+	return []string{"-c", "http.extraHeader=Authorization: token " + token}
+}
+
+// redactToken scrubs the auth token from a string before it's logged
+// or returned in an error. Belt-and-braces: with the http.extraHeader
+// approach the token shouldn't appear in git's output, but if some
+// future git version or libcurl debug mode emits it, this catches it.
+func redactToken(s string) string {
+	token := os.Getenv("MOLECULE_GITEA_TOKEN")
+	if token == "" || len(token) < 8 {
+		return s
+	}
+	return strings.ReplaceAll(s, token, "<redacted-token>")
+}
+

workspace-server/internal/handlers/org_external_integration_test.go

@@ -0,0 +1,379 @@
+package handlers
+
+import (
+	"context"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"runtime"
+	"strings"
+	"testing"
+
+	"gopkg.in/yaml.v3"
+)
+
+// PR-B integration test: exercises the REAL gitFetcher (no fakeFetcher
+// injection) against a local bare-git repo. Uses git's `insteadOf`
+// config to rewrite the configured Gitea URL to the local bare path
+// at clone time, so the fetcher's URL-building, ls-remote, clone,
+// atomic-rename, and cache-hit paths all run against real git
+// without requiring network or modifying production code.
+//
+// Internal#77 task #233 (PR-B from the design's phasing).
+
+// TestGitFetcher_RealClone_LocalRedirect proves the production
+// gitFetcher round-trips correctly against a real git repository.
+// Steps:
+//   1. Set up a local bare-git repo with workspace content.
+//   2. Configure git's `insteadOf` to rewrite the gitea URL → local path
+//      via GIT_CONFIG_COUNT/KEY/VALUE env vars (process-scoped).
+//   3. Run resolveYAMLIncludes with !external pointing at the gitea URL.
+//   4. Assert: cache dir populated; content materialized; path rewrite
+//      applied; second invocation hits cache (no second clone).
+func TestGitFetcher_RealClone_LocalRedirect(t *testing.T) {
+	if _, err := exec.LookPath("git"); err != nil {
+		t.Skipf("git binary not found: %v", err)
+	}
+
+	if runtime.GOOS == "windows" {
+		t.Skip("path-based git URLs behave differently on Windows; skipping")
+	}
+
+	// Step 1: create a local bare-git repo at <fixtures>/test-dev-dept.git
+	// with workspace content. Use a working clone to add content, then
+	// push to the bare.
+	fixtures := t.TempDir()
+	barePath := filepath.Join(fixtures, "test-dev-dept.git")
+	workPath := filepath.Join(fixtures, "work")
+
+	mustGit(t, "", "init", "--bare", "-b", "main", barePath)
+	mustGit(t, "", "clone", barePath, workPath)
+	mustGit(t, workPath, "config", "user.email", "test@example.com")
+	mustGit(t, workPath, "config", "user.name", "Integration Test")
+
+	mustWriteFile(t, filepath.Join(workPath, "dev-lead/workspace.yaml"), `name: Dev Lead
+files_dir: dev-lead
+children:
+  - !include ./core-be/workspace.yaml
+`)
+	mustWriteFile(t, filepath.Join(workPath, "dev-lead/system-prompt.md"), "Dev Lead persona body.\n")
+	mustWriteFile(t, filepath.Join(workPath, "dev-lead/core-be/workspace.yaml"), `name: Core BE
+files_dir: dev-lead/core-be
+`)
+	mustWriteFile(t, filepath.Join(workPath, "dev-lead/core-be/system-prompt.md"), "Core BE persona body.\n")
+
+	mustGit(t, workPath, "add", ".")
+	mustGit(t, workPath, "commit", "-m", "seed dev tree")
+	mustGit(t, workPath, "push", "origin", "main")
+
+	// Step 2: configure git's insteadOf rewrite. The fetcher will try
+	// to clone https://git.moleculesai.app/molecule-ai/test-dev-dept.git;
+	// git rewrites to file://<barePath>.
+	//
+	// GIT_CONFIG_COUNT/KEY/VALUE injects config without touching
+	// ~/.gitconfig — process-scoped, no test pollution.
+	geesUrl := "https://git.moleculesai.app/molecule-ai/test-dev-dept.git"
+	t.Setenv("GIT_CONFIG_COUNT", "1")
+	t.Setenv("GIT_CONFIG_KEY_0", "url."+barePath+".insteadOf")
+	t.Setenv("GIT_CONFIG_VALUE_0", geesUrl)
+
+	// Step 3: run resolveYAMLIncludes with !external pointing at the
+	// gitea URL. Allowlist is the default (molecule-ai/* on Gitea host).
+	rootDir := t.TempDir()
+	src := []byte(`workspaces:
+  - !external
+    repo: molecule-ai/test-dev-dept
+    ref: main
+    path: dev-lead/workspace.yaml
+`)
+
+	out, err := resolveYAMLIncludes(src, rootDir)
+	if err != nil {
+		t.Fatalf("resolveYAMLIncludes: %v", err)
+	}
+
+	var tmpl OrgTemplate
+	if err := yaml.Unmarshal(out, &tmpl); err != nil {
+		t.Fatalf("unmarshal: %v", err)
+	}
+	if len(tmpl.Workspaces) != 1 {
+		t.Fatalf("workspaces: %+v", tmpl.Workspaces)
+	}
+	dev := tmpl.Workspaces[0]
+	if dev.Name != "Dev Lead" {
+		t.Errorf("dev.Name = %q; want Dev Lead", dev.Name)
+	}
+	if !strings.Contains(dev.FilesDir, ".external-cache") {
+		t.Errorf("dev.FilesDir = %q; want cache prefix", dev.FilesDir)
+	}
+	if !strings.HasSuffix(dev.FilesDir, "dev-lead") {
+		t.Errorf("dev.FilesDir = %q; want suffix dev-lead", dev.FilesDir)
+	}
+	if len(dev.Children) != 1 {
+		t.Fatalf("expected nested core-be child; got %+v", dev.Children)
+	}
+	core := dev.Children[0]
+	if core.Name != "Core BE" {
+		t.Errorf("core.Name = %q; want Core BE", core.Name)
+	}
+	if !strings.HasSuffix(core.FilesDir, filepath.Join("dev-lead", "core-be")) {
+		t.Errorf("core.FilesDir = %q; want suffix dev-lead/core-be", core.FilesDir)
+	}
+
+	// Step 4: verify the cache dir actually exists and contains the
+	// materialized files (CopyTemplateToContainer would tar these).
+	cacheRoot := filepath.Join(rootDir, ".external-cache")
+	entries, err := os.ReadDir(cacheRoot)
+	if err != nil {
+		t.Fatalf("read cache root: %v", err)
+	}
+	if len(entries) != 1 {
+		t.Fatalf("expected 1 cached repo, got %d: %v", len(entries), entries)
+	}
+	repoDir := filepath.Join(cacheRoot, entries[0].Name())
+	shaDirs, _ := os.ReadDir(repoDir)
+	if len(shaDirs) != 1 {
+		t.Fatalf("expected 1 SHA cache dir, got %d", len(shaDirs))
+	}
+	cacheDir := filepath.Join(repoDir, shaDirs[0].Name())
+	if _, err := os.Stat(filepath.Join(cacheDir, "dev-lead/system-prompt.md")); err != nil {
+		t.Errorf("expected dev-lead/system-prompt.md in cache: %v", err)
+	}
+	if _, err := os.Stat(filepath.Join(cacheDir, "dev-lead/core-be/system-prompt.md")); err != nil {
+		t.Errorf("expected dev-lead/core-be/system-prompt.md in cache: %v", err)
+	}
+
+	// Step 5: re-run; verify cache hit (no second clone). Set a
+	// "marker" file in the cache that a second clone would clobber.
+	marker := filepath.Join(cacheDir, ".cache-hit-marker")
+	if err := os.WriteFile(marker, []byte("hit"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	out2, err := resolveYAMLIncludes(src, rootDir)
+	if err != nil {
+		t.Fatalf("resolveYAMLIncludes second call: %v", err)
+	}
+	if string(out) != string(out2) {
+		t.Errorf("cached output differs from initial — non-deterministic resolve")
+	}
+	if _, err := os.Stat(marker); err != nil {
+		t.Errorf("cache hit not honored — marker file disappeared: %v", err)
+	}
+}
+
+// TestGitFetcher_RealClone_BadRefFails: pointing at a ref that doesn't
+// exist in the bare-repo surfaces git's error cleanly.
+func TestGitFetcher_RealClone_BadRefFails(t *testing.T) {
+	if _, err := exec.LookPath("git"); err != nil {
+		t.Skipf("git binary not found: %v", err)
+	}
+	if runtime.GOOS == "windows" {
+		t.Skip("skipping on windows")
+	}
+
+	fixtures := t.TempDir()
+	barePath := filepath.Join(fixtures, "empty-repo.git")
+	workPath := filepath.Join(fixtures, "work")
+	mustGit(t, "", "init", "--bare", "-b", "main", barePath)
+	mustGit(t, "", "clone", barePath, workPath)
+	mustGit(t, workPath, "config", "user.email", "test@example.com")
+	mustGit(t, workPath, "config", "user.name", "Test")
+	mustWriteFile(t, filepath.Join(workPath, "README.md"), "x")
+	mustGit(t, workPath, "add", ".")
+	mustGit(t, workPath, "commit", "-m", "seed")
+	mustGit(t, workPath, "push", "origin", "main")
+
+	t.Setenv("GIT_CONFIG_COUNT", "1")
+	t.Setenv("GIT_CONFIG_KEY_0", "url."+barePath+".insteadOf")
+	t.Setenv("GIT_CONFIG_VALUE_0", "https://git.moleculesai.app/molecule-ai/empty-repo.git")
+
+	rootDir := t.TempDir()
+	src := []byte(`workspaces:
+  - !external
+    repo: molecule-ai/empty-repo
+    ref: nonexistent-branch
+    path: anything.yaml
+`)
+	_, err := resolveYAMLIncludes(src, rootDir)
+	if err == nil {
+		t.Fatalf("expected error for nonexistent ref; got nil")
+	}
+	if !strings.Contains(err.Error(), "ref") && !strings.Contains(err.Error(), "ls-remote") && !strings.Contains(err.Error(), "not found") {
+		t.Errorf("error doesn't mention ref/ls-remote: %v", err)
+	}
+}
+
+// ---------- helpers ----------
+
+func mustGit(t *testing.T, cwd string, args ...string) {
+	t.Helper()
+	cmd := exec.Command("git", args...)
+	if cwd != "" {
+		cmd.Dir = cwd
+	}
+	// Ensure user.email/name are set globally for non-cwd commands too.
+	cmd.Env = append(os.Environ(),
+		"GIT_AUTHOR_EMAIL=test@example.com",
+		"GIT_AUTHOR_NAME=Integration Test",
+		"GIT_COMMITTER_EMAIL=test@example.com",
+		"GIT_COMMITTER_NAME=Integration Test",
+	)
+	if out, err := cmd.CombinedOutput(); err != nil {
+		t.Fatalf("git %s: %v\n%s", strings.Join(args, " "), err, string(out))
+	}
+}
+
+func mustWriteFile(t *testing.T, path, content string) {
+	t.Helper()
+	if err := os.MkdirAll(filepath.Dir(path), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(path, []byte(content), 0o644); err != nil {
+		t.Fatal(err)
+	}
+}
+
+// Verify gitFetcher.Fetch direct invocation (no resolver wrapping) for
+// the cache-hit path, exercising the bare API against a local bare-repo.
+func TestGitFetcher_DirectFetch_CacheHit(t *testing.T) {
+	if _, err := exec.LookPath("git"); err != nil {
+		t.Skipf("git binary not found: %v", err)
+	}
+	if runtime.GOOS == "windows" {
+		t.Skip("skipping on windows")
+	}
+
+	fixtures := t.TempDir()
+	barePath := filepath.Join(fixtures, "direct.git")
+	workPath := filepath.Join(fixtures, "w")
+	mustGit(t, "", "init", "--bare", "-b", "main", barePath)
+	mustGit(t, "", "clone", barePath, workPath)
+	mustGit(t, workPath, "config", "user.email", "t@e")
+	mustGit(t, workPath, "config", "user.name", "T")
+	mustWriteFile(t, filepath.Join(workPath, "marker.txt"), "hello")
+	mustGit(t, workPath, "add", ".")
+	mustGit(t, workPath, "commit", "-m", "seed")
+	mustGit(t, workPath, "push", "origin", "main")
+
+	t.Setenv("GIT_CONFIG_COUNT", "1")
+	t.Setenv("GIT_CONFIG_KEY_0", "url."+barePath+".insteadOf")
+	t.Setenv("GIT_CONFIG_VALUE_0", "https://git.moleculesai.app/molecule-ai/direct.git")
+
+	rootDir := t.TempDir()
+	g := &gitFetcher{}
+	ctx := context.Background()
+
+	cacheDir1, sha1, err := g.Fetch(ctx, rootDir, "git.moleculesai.app", "molecule-ai/direct", "main")
+	if err != nil {
+		t.Fatalf("first Fetch: %v", err)
+	}
+	if sha1 == "" || len(sha1) < 7 {
+		t.Errorf("expected SHA-like string, got %q", sha1)
+	}
+	if _, err := os.Stat(filepath.Join(cacheDir1, "marker.txt")); err != nil {
+		t.Errorf("first fetch missing marker.txt: %v", err)
+	}
+
+	// Second call: cache hit, returns same dir + sha, no re-clone.
+	stamp := filepath.Join(cacheDir1, ".not-clobbered-by-second-fetch")
+	if err := os.WriteFile(stamp, []byte("x"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	cacheDir2, sha2, err := g.Fetch(ctx, rootDir, "git.moleculesai.app", "molecule-ai/direct", "main")
+	if err != nil {
+		t.Fatalf("second Fetch: %v", err)
+	}
+	if cacheDir2 != cacheDir1 || sha2 != sha1 {
+		t.Errorf("cache miss on second call: %q/%q vs %q/%q", cacheDir1, sha1, cacheDir2, sha2)
+	}
+	if _, err := os.Stat(stamp); err != nil {
+		t.Errorf("cache hit not honored — stamp file disappeared: %v", err)
+	}
+}
+
+// TestGitFetcher_RejectsRefWithDoubleDot: defense-in-depth on ref input.
+// safeRefPattern allows '.' as a regex character, so ".." would match
+// without an explicit deny. Verify it's rejected even though git itself
+// would also reject the resulting clone.
+func TestGitFetcher_RejectsRefWithDoubleDot(t *testing.T) {
+	rootDir := t.TempDir()
+	src := []byte(`workspaces:
+  - !external
+    repo: molecule-ai/x
+    ref: foo..bar
+    path: x.yaml
+`)
+	_, err := resolveYAMLIncludes(src, rootDir)
+	if err == nil {
+		t.Fatalf("expected '..' rejection")
+	}
+	if !strings.Contains(err.Error(), "..") {
+		t.Errorf("expected '..' in error; got %v", err)
+	}
+}
+
+// TestGitFetcher_CacheValidatedByCompleteMarker: a partially-written
+// cache (the .git dir exists but no .complete marker) is treated as
+// cache-miss and re-fetched. Catches the broken-cache-permanence bug.
+func TestGitFetcher_CacheValidatedByCompleteMarker(t *testing.T) {
+	if _, err := exec.LookPath("git"); err != nil {
+		t.Skipf("git not found: %v", err)
+	}
+	if runtime.GOOS == "windows" {
+		t.Skip("skipping on windows")
+	}
+
+	fixtures := t.TempDir()
+	barePath := filepath.Join(fixtures, "test.git")
+	workPath := filepath.Join(fixtures, "w")
+	mustGit(t, "", "init", "--bare", "-b", "main", barePath)
+	mustGit(t, "", "clone", barePath, workPath)
+	mustGit(t, workPath, "config", "user.email", "t@e")
+	mustGit(t, workPath, "config", "user.name", "T")
+	mustWriteFile(t, filepath.Join(workPath, "good.txt"), "from-network")
+	mustGit(t, workPath, "add", ".")
+	mustGit(t, workPath, "commit", "-m", "seed")
+	mustGit(t, workPath, "push", "origin", "main")
+	t.Setenv("GIT_CONFIG_COUNT", "1")
+	t.Setenv("GIT_CONFIG_KEY_0", "url."+barePath+".insteadOf")
+	t.Setenv("GIT_CONFIG_VALUE_0", "https://git.moleculesai.app/molecule-ai/marker-test.git")
+
+	rootDir := t.TempDir()
+	g := &gitFetcher{}
+
+	// First fetch — populates the cache (creates .complete marker).
+	cacheDir1, _, err := g.Fetch(context.Background(), rootDir, "git.moleculesai.app", "molecule-ai/marker-test", "main")
+	if err != nil {
+		t.Fatalf("first Fetch: %v", err)
+	}
+	marker := filepath.Join(cacheDir1, cacheCompleteMarker)
+	if _, err := os.Stat(marker); err != nil {
+		t.Fatalf("first fetch should have written .complete marker: %v", err)
+	}
+
+	// Now simulate a partial cache: delete the marker but leave .git
+	// in place. The next Fetch should treat this as cache-miss and
+	// re-fetch (NOT silently use the partial cache).
+	if err := os.Remove(marker); err != nil {
+		t.Fatal(err)
+	}
+	// Drop a sentinel file the second fetch will clobber if it re-fetches.
+	sentinel := filepath.Join(cacheDir1, "_should_be_clobbered")
+	if err := os.WriteFile(sentinel, []byte("partial"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	cacheDir2, _, err := g.Fetch(context.Background(), rootDir, "git.moleculesai.app", "molecule-ai/marker-test", "main")
+	if err != nil {
+		t.Fatalf("second Fetch: %v", err)
+	}
+	if cacheDir1 != cacheDir2 {
+		t.Errorf("cache dirs differ across fetches: %q vs %q", cacheDir1, cacheDir2)
+	}
+	if _, err := os.Stat(filepath.Join(cacheDir2, cacheCompleteMarker)); err != nil {
+		t.Errorf("re-fetch should have re-written .complete marker: %v", err)
+	}
+	if _, err := os.Stat(sentinel); err == nil {
+		t.Errorf("sentinel still present — re-fetch did NOT clobber partial cache")
+	}
+}

workspace-server/internal/handlers/org_external_test.go

@@ -0,0 +1,331 @@
+package handlers
+
+import (
+	"context"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"gopkg.in/yaml.v3"
+)
+
+// fakeFetcher pre-stages a "fetched" repo at a fixed path inside the
+// rootDir's .external-cache, bypassing the real git clone. Tests
+// inject this via SetExternalFetcherForTest to exercise the resolver
+// + path-rewrite logic without network.
+type fakeFetcher struct {
+	// content maps "<host>/<repo>@<ref>" → a function that materializes
+	// repo content under cacheDir. Returns the fake SHA to use.
+	content map[string]func(cacheDir string) (sha string, err error)
+}
+
+func (f *fakeFetcher) Fetch(ctx context.Context, rootDir, host, repoPath, ref string) (string, string, error) {
+	key := host + "/" + repoPath + "@" + ref
+	stage, ok := f.content[key]
+	if !ok {
+		return "", "", &fakeNotFoundError{key: key}
+	}
+	// Use a stable SHA for the test so cache dir is deterministic.
+	cacheDir := filepath.Join(rootDir, ".external-cache", safeRepoCacheDir(host, repoPath), "deadbeef")
+	if err := os.MkdirAll(cacheDir, 0o755); err != nil {
+		return "", "", err
+	}
+	sha, err := stage(cacheDir)
+	if err != nil {
+		return "", "", err
+	}
+	return cacheDir, sha, nil
+}
+
+type fakeNotFoundError struct{ key string }
+
+func (e *fakeNotFoundError) Error() string {
+	return "fake fetcher: no content registered for " + e.key
+}
+
+// stageFiles writes a map of relative-path → content into cacheDir,
+// returning a fake SHA. Helper for fakeFetcher closures.
+func stageFiles(cacheDir string, files map[string]string) error {
+	if err := os.MkdirAll(filepath.Join(cacheDir, ".git"), 0o755); err != nil {
+		return err
+	}
+	for path, content := range files {
+		full := filepath.Join(cacheDir, path)
+		if err := os.MkdirAll(filepath.Dir(full), 0o755); err != nil {
+			return err
+		}
+		if err := os.WriteFile(full, []byte(content), 0o644); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+// TestResolveExternalMapping_HappyPath: a parent template with an
+// !external entry resolves cleanly into the fetched workspace + path-
+// rewrites files_dir + relative !include refs into the cache prefix.
+func TestResolveExternalMapping_HappyPath(t *testing.T) {
+	tmp := t.TempDir()
+
+	// Stub fetcher: "fetched" content has a workspace.yaml that uses
+	// files_dir + nested !include relative to the fetched repo's root.
+	fake := &fakeFetcher{
+		content: map[string]func(string) (string, error){
+			"git.moleculesai.app/molecule-ai/molecule-dev-department@main": func(cacheDir string) (string, error) {
+				return "deadbeef", stageFiles(cacheDir, map[string]string{
+					"dev-lead/workspace.yaml": `name: Dev Lead
+files_dir: dev-lead
+children:
+  - !include ./core-lead/workspace.yaml
+`,
+					"dev-lead/core-lead/workspace.yaml": `name: Core Platform Lead
+files_dir: dev-lead/core-lead
+`,
+				})
+			},
+		},
+	}
+	cleanup := SetExternalFetcherForTest(fake)
+	defer cleanup()
+
+	src := []byte(`name: Parent
+workspaces:
+  - !external
+    repo: molecule-ai/molecule-dev-department
+    ref: main
+    path: dev-lead/workspace.yaml
+`)
+
+	out, err := resolveYAMLIncludes(src, tmp)
+	if err != nil {
+		t.Fatalf("resolveYAMLIncludes: %v", err)
+	}
+
+	var tmpl OrgTemplate
+	if err := yaml.Unmarshal(out, &tmpl); err != nil {
+		t.Fatalf("unmarshal: %v", err)
+	}
+	if len(tmpl.Workspaces) != 1 {
+		t.Fatalf("workspaces: %+v", tmpl.Workspaces)
+	}
+	dev := tmpl.Workspaces[0]
+	if dev.Name != "Dev Lead" {
+		t.Errorf("dev.Name = %q; want Dev Lead", dev.Name)
+	}
+	// files_dir should be cache-prefixed.
+	wantPrefix := filepath.Join(".external-cache", "git.moleculesai.app__molecule-ai__molecule-dev-department", "deadbeef")
+	if !strings.HasPrefix(dev.FilesDir, wantPrefix) {
+		t.Errorf("dev.FilesDir = %q; want prefix %q", dev.FilesDir, wantPrefix)
+	}
+	if !strings.HasSuffix(dev.FilesDir, "dev-lead") {
+		t.Errorf("dev.FilesDir = %q; want suffix dev-lead", dev.FilesDir)
+	}
+	// Nested child: files_dir cache-prefixed, name Core Platform Lead.
+	if len(dev.Children) != 1 {
+		t.Fatalf("dev.Children: %+v", dev.Children)
+	}
+	core := dev.Children[0]
+	if core.Name != "Core Platform Lead" {
+		t.Errorf("core.Name = %q; want Core Platform Lead", core.Name)
+	}
+	if !strings.HasPrefix(core.FilesDir, wantPrefix) {
+		t.Errorf("core.FilesDir = %q; want prefix %q", core.FilesDir, wantPrefix)
+	}
+	if !strings.HasSuffix(core.FilesDir, filepath.Join("dev-lead", "core-lead")) {
+		t.Errorf("core.FilesDir = %q; want suffix dev-lead/core-lead", core.FilesDir)
+	}
+}
+
+// TestResolveExternalMapping_AllowlistRejection: hostile yaml pointing
+// at a non-allowlisted repo gets rejected.
+func TestResolveExternalMapping_AllowlistRejection(t *testing.T) {
+	tmp := t.TempDir()
+	fake := &fakeFetcher{content: map[string]func(string) (string, error){}}
+	cleanup := SetExternalFetcherForTest(fake)
+	defer cleanup()
+
+	// Default allowlist is git.moleculesai.app/molecule-ai/*.
+	// github.com/foo/bar is NOT in it.
+	src := []byte(`workspaces:
+  - !external
+    repo: foo/bar
+    ref: main
+    path: x.yaml
+    url: github.com
+`)
+	_, err := resolveYAMLIncludes(src, tmp)
+	if err == nil {
+		t.Fatalf("expected allowlist rejection, got nil")
+	}
+	if !strings.Contains(err.Error(), "MOLECULE_EXTERNAL_REPO_ALLOWLIST") {
+		t.Errorf("expected allowlist error; got %v", err)
+	}
+}
+
+// TestResolveExternalMapping_PathTraversalRejection: hostile yaml
+// with `path: ../../etc/passwd` gets rejected before fetch.
+func TestResolveExternalMapping_PathTraversalRejection(t *testing.T) {
+	tmp := t.TempDir()
+	fake := &fakeFetcher{content: map[string]func(string) (string, error){}}
+	cleanup := SetExternalFetcherForTest(fake)
+	defer cleanup()
+
+	src := []byte(`workspaces:
+  - !external
+    repo: molecule-ai/dev-department
+    ref: main
+    path: ../../etc/passwd
+`)
+	_, err := resolveYAMLIncludes(src, tmp)
+	if err == nil {
+		t.Fatalf("expected path traversal rejection, got nil")
+	}
+	if !strings.Contains(err.Error(), "relative-and-down-only") {
+		t.Errorf("expected path traversal error; got %v", err)
+	}
+}
+
+// TestResolveExternalMapping_BadRefRejection: non-allowlisted ref chars.
+func TestResolveExternalMapping_BadRefRejection(t *testing.T) {
+	tmp := t.TempDir()
+	fake := &fakeFetcher{content: map[string]func(string) (string, error){}}
+	cleanup := SetExternalFetcherForTest(fake)
+	defer cleanup()
+
+	src := []byte(`workspaces:
+  - !external
+    repo: molecule-ai/dev-department
+    ref: "main; rm -rf /"
+    path: foo.yaml
+`)
+	_, err := resolveYAMLIncludes(src, tmp)
+	if err == nil || !strings.Contains(err.Error(), "disallowed characters") {
+		t.Errorf("expected ref-validation error; got %v", err)
+	}
+}
+
+// TestResolveExternalMapping_MissingRequiredFields: repo / ref / path
+// are all required.
+func TestResolveExternalMapping_MissingRequiredFields(t *testing.T) {
+	tmp := t.TempDir()
+	fake := &fakeFetcher{content: map[string]func(string) (string, error){}}
+	cleanup := SetExternalFetcherForTest(fake)
+	defer cleanup()
+
+	cases := []string{
+		// missing repo
+		`workspaces:
+  - !external
+    ref: main
+    path: x.yaml
+`,
+		// missing ref
+		`workspaces:
+  - !external
+    repo: molecule-ai/x
+    path: x.yaml
+`,
+		// missing path
+		`workspaces:
+  - !external
+    repo: molecule-ai/x
+    ref: main
+`,
+	}
+	for i, src := range cases {
+		_, err := resolveYAMLIncludes([]byte(src), tmp)
+		if err == nil {
+			t.Errorf("case %d: expected required-field error, got nil", i)
+		} else if !strings.Contains(err.Error(), "required") {
+			t.Errorf("case %d: want 'required' in error; got %v", i, err)
+		}
+	}
+}
+
+// TestRewriteFilesDir: verify the path-rewrite walker
+// prefixes files_dir scalars. !include scalars are NOT rewritten —
+// they resolve relative to their containing file's dir, which post-
+// fetch is naturally inside the cache.
+func TestRewriteFilesDir(t *testing.T) {
+	src := `name: Foo
+files_dir: dev-lead
+children:
+  - !include ./bar/workspace.yaml
+  - !include other-team.yaml
+inner:
+  files_dir: dev-lead/sub
+`
+	var n yaml.Node
+	if err := yaml.Unmarshal([]byte(src), &n); err != nil {
+		t.Fatal(err)
+	}
+	rewriteFilesDir(&n, ".external-cache/foo/bar")
+
+	out, err := yaml.Marshal(&n)
+	if err != nil {
+		t.Fatal(err)
+	}
+	got := string(out)
+	for _, want := range []string{
+		"files_dir: .external-cache/foo/bar/dev-lead",
+		"files_dir: .external-cache/foo/bar/dev-lead/sub",
+		// !include preserved as-is; resolves naturally via subDir.
+		"!include ./bar/workspace.yaml",
+		"!include other-team.yaml",
+	} {
+		if !strings.Contains(got, want) {
+			t.Errorf("missing %q in:\n%s", want, got)
+		}
+	}
+}
+
+// TestRewriteFilesDir_Idempotent: re-running the rewriter
+// on already-prefixed files_dir doesn't double-prefix.
+func TestRewriteFilesDir_Idempotent(t *testing.T) {
+	src := `files_dir: .external-cache/foo/bar/dev-lead
+inner:
+  files_dir: .external-cache/foo/bar/dev-lead/sub
+`
+	var n yaml.Node
+	if err := yaml.Unmarshal([]byte(src), &n); err != nil {
+		t.Fatal(err)
+	}
+	rewriteFilesDir(&n, ".external-cache/foo/bar")
+
+	out, _ := yaml.Marshal(&n)
+	got := string(out)
+	if strings.Contains(got, ".external-cache/foo/bar/.external-cache") {
+		t.Errorf("double-prefix detected:\n%s", got)
+	}
+	// Should still be valid (single-prefixed) afterwards.
+	for _, want := range []string{
+		"files_dir: .external-cache/foo/bar/dev-lead",
+		"files_dir: .external-cache/foo/bar/dev-lead/sub",
+	} {
+		if !strings.Contains(got, want) {
+			t.Errorf("expected unchanged %q in:\n%s", want, got)
+		}
+	}
+}
+
+// TestAllowlistedHostPath: env-var override + glob matching.
+func TestAllowlistedHostPath(t *testing.T) {
+	t.Setenv("MOLECULE_EXTERNAL_REPO_ALLOWLIST", "")
+	if !allowlistedHostPath("git.moleculesai.app", "molecule-ai/foo") {
+		t.Error("default allowlist should accept molecule-ai/*")
+	}
+	if allowlistedHostPath("github.com", "molecule-ai/foo") {
+		t.Error("default allowlist should reject github.com")
+	}
+	t.Setenv("MOLECULE_EXTERNAL_REPO_ALLOWLIST", "github.com/me/*,git.moleculesai.app/*")
+	if !allowlistedHostPath("github.com", "me/x") {
+		t.Error("override should accept github.com/me/*")
+	}
+	if !allowlistedHostPath("git.moleculesai.app", "any/repo") {
+		t.Error("override should accept git.moleculesai.app/*")
+	}
+	if allowlistedHostPath("github.com", "evil/x") {
+		t.Error("override should reject github.com/evil/*")
+	}
+}

workspace-server/internal/handlers/org_helpers.go

@@ -6,6 +6,7 @@ package handlers
 
 import (
 	"fmt"
+	"log"
 	"os"
 	"path/filepath"
 	"regexp"
@@ -102,6 +103,56 @@ func loadWorkspaceEnv(orgBaseDir, filesDir string) map[string]string {
 	return envVars
 }
 
+// loadPersonaEnvFile merges per-role persona credentials into out. The file
+// lives at $MOLECULE_PERSONA_ROOT/<role>/env (default
+// /etc/molecule-bootstrap/personas) and is populated by the operator-host
+// bootstrap kit — one persona per dev-tree role, each carrying the role's
+// Gitea identity (GITEA_USER, GITEA_TOKEN, GITEA_TOKEN_SCOPES,
+// GITEA_USER_EMAIL, GITEA_SSH_KEY_PATH).
+//
+// Lower precedence than the org and workspace .env files: callers should
+// invoke this BEFORE parseEnvFile on those, so a workspace .env can
+// override a persona-default value when needed.
+//
+// Silent no-op when role is empty, when the role name fails the safe-segment
+// check, or when the env file does not exist (workspaces without a role —
+// or running on hosts that don't ship the bootstrap dir — keep their old
+// behavior).
+func loadPersonaEnvFile(role string, out map[string]string) {
+	if !isSafeRoleName(role) {
+		if role != "" {
+			log.Printf("Org import: refusing persona env load for unsafe role name %q", role)
+		}
+		return
+	}
+	root := os.Getenv("MOLECULE_PERSONA_ROOT")
+	if root == "" {
+		root = "/etc/molecule-bootstrap/personas"
+	}
+	parseEnvFile(filepath.Join(root, role, "env"), out)
+}
+
+// isSafeRoleName accepts a single path segment of [A-Za-z0-9_-]+. Rejects
+// empty, ".", "..", and anything containing a path separator — even though
+// the construct is admin-only, defense-in-depth keeps the persona dir
+// shape invariant: one flat directory per role, no climbing out.
+func isSafeRoleName(s string) bool {
+	if s == "" || s == "." || s == ".." {
+		return false
+	}
+	for _, c := range s {
+		switch {
+		case c >= 'a' && c <= 'z':
+		case c >= 'A' && c <= 'Z':
+		case c >= '0' && c <= '9':
+		case c == '-' || c == '_':
+		default:
+			return false
+		}
+	}
+	return true
+}
+
 // parseEnvFile reads a .env file and adds KEY=VALUE pairs to the map.
 // Skips comments (#) and empty lines. Values can be quoted.
 func parseEnvFile(path string, out map[string]string) {

workspace-server/internal/handlers/org_import.go

@@ -42,6 +42,20 @@ import (
 // straight into the parent's child-coordinate space without doing a
 // canvas-wide absolute-position walk.
 func (h *OrgHandler) createWorkspaceTree(ws OrgWorkspace, parentID *string, absX, absY, relX, relY float64, defaults OrgDefaults, orgBaseDir string, results *[]map[string]interface{}, provisionSem chan struct{}) error {
+	// spawning: false guard — skip this workspace AND all descendants.
+	// Pointer-typed so we distinguish "explicitly false" from "unset"
+	// (unset = default to spawn). The guard sits BEFORE any side effect
+	// (no DB row, no docker provision, no children recursion) so a
+	// false-spawning subtree is genuinely a no-op except for the log line.
+	// Use case: dev-tree org template ships the full role taxonomy but a
+	// developer's machine only has RAM for a subset; a per-workspace
+	// `spawning: false` lets them narrow without editing the parent
+	// template's structure.
+	if ws.Spawning != nil && !*ws.Spawning {
+		log.Printf("Org import: skipping workspace %q (spawning=false; %d descendant workspace(s) in subtree also skipped)", ws.Name, countWorkspaces(ws.Children))
+		return nil
+	}
+
 	// Apply defaults
 	runtime := ws.Runtime
 	if runtime == "" {
@@ -443,10 +457,35 @@ func (h *OrgHandler) createWorkspaceTree(ws OrgWorkspace, parentID *string, absX
 			configFiles["system-prompt.md"] = []byte(ws.SystemPrompt)
 		}
 
-		// Inject secrets from .env files as workspace secrets.
-		// Resolution: workspace .env → org root .env (workspace overrides org root).
+		// Inject secrets from persona env + .env files as workspace secrets.
+		// Resolution (later overrides earlier):
+		//   0. Persona env (per-role bootstrap creds; only when ws.Role is set
+		//      and the operator-host bootstrap dir ships a matching file)
+		//   1. Org root .env (shared defaults)
+		//   2. Workspace-specific .env (per-workspace overrides)
 		// Each line: KEY=VALUE → stored as encrypted workspace secret.
 		envVars := map[string]string{}
+		// 0. Persona env (lowest precedence; injects the role's Gitea identity:
+		//    GITEA_USER, GITEA_TOKEN, GITEA_TOKEN_SCOPES, GITEA_USER_EMAIL,
+		//    GITEA_SSH_KEY_PATH, plus MODEL_PROVIDER/MODEL and the LLM auth
+		//    token like CLAUDE_CODE_OAUTH_TOKEN or MINIMAX_API_KEY).
+		//    Workspace and org .env can override.
+		//
+		// Use ws.FilesDir as the persona-dir lookup key, NOT ws.Role. In the
+		// dev-tree org.yaml shape, `role:` carries the multi-line descriptive
+		// text the agent reads from its prompt ("Engineering planning and
+		// team coordination — leads Core Platform, Controlplane, ..."), while
+		// `files_dir:` holds the short slug (`core-lead`, `dev-lead`, etc.)
+		// matching `~/.molecule-ai/personas/<files_dir>/env`
+		// (bind-mounted to `/etc/molecule-bootstrap/personas/<files_dir>/env`).
+		//
+		// Pre-fix, this passed `ws.Role` whose multi-word content failed
+		// isSafeRoleName silently, so every imported workspace booted with
+		// zero persona-env rows in workspace_secrets — no ANTHROPIC /
+		// CLAUDE_CODE auth in the container env. The claude_agent_sdk
+		// then wedged on `query.initialize()` with a 60s control-request
+		// timeout (caught 2026-05-08 right after dev-only org/import).
+		loadPersonaEnvFile(ws.FilesDir, envVars)
 		if orgBaseDir != "" {
 			// 1. Org root .env (shared defaults)
 			parseEnvFile(filepath.Join(orgBaseDir, ".env"), envVars)

workspace-server/internal/handlers/org_import_reconcile_test.go

@@ -0,0 +1,158 @@
+package handlers
+
+import (
+	"context"
+	"sort"
+	"testing"
+
+	"github.com/DATA-DOG/go-sqlmock"
+)
+
+// Tests for the reconcile-mode + audit-event additions to OrgHandler.Import.
+//
+// Background: /org/import was purely additive — re-running with a tree that
+// renamed/reparented a role left the prior workspace online (different
+// parent_id from the new one, so lookupExistingChild's parent-scoped dedupe
+// missed it). The 2026-05-08 dev-tree case left 8 orphans surviving a
+// re-import. mode="reconcile" closes the gap; emitOrgEvent makes "what
+// happened at 20:13?" queryable instead of stdout-grep archaeology.
+
+func TestWalkOrgWorkspaceNames_FlatTree(t *testing.T) {
+	tree := []OrgWorkspace{
+		{Name: "Dev Lead"},
+		{Name: "Release Manager"},
+	}
+	var names []string
+	walkOrgWorkspaceNames(tree, &names)
+	sort.Strings(names)
+	want := []string{"Dev Lead", "Release Manager"}
+	if !equalStrings(names, want) {
+		t.Errorf("flat tree: got %v, want %v", names, want)
+	}
+}
+
+func TestWalkOrgWorkspaceNames_NestedTree(t *testing.T) {
+	tree := []OrgWorkspace{
+		{
+			Name: "Dev Lead",
+			Children: []OrgWorkspace{
+				{Name: "Core Platform Lead", Children: []OrgWorkspace{{Name: "Core-BE"}}},
+				{Name: "SDK Lead"},
+			},
+		},
+	}
+	var names []string
+	walkOrgWorkspaceNames(tree, &names)
+	sort.Strings(names)
+	want := []string{"Core Platform Lead", "Core-BE", "Dev Lead", "SDK Lead"}
+	if !equalStrings(names, want) {
+		t.Errorf("nested tree: got %v, want %v", names, want)
+	}
+}
+
+// Pins the contract that spawning:false subtrees still contribute their names
+// to the reconcile working set. If the walker started skipping them, a
+// re-import that toggled spawning would orphan whichever workspaces had been
+// previously imported with spawning:true — the inverse of the bug being
+// fixed. Spawning gates *provisioning*, not *reconcile membership*.
+func TestWalkOrgWorkspaceNames_SpawningFalseStillCounted(t *testing.T) {
+	f := false
+	tree := []OrgWorkspace{
+		{Name: "Dev Lead", Children: []OrgWorkspace{
+			{Name: "Skipped Lead", Spawning: &f, Children: []OrgWorkspace{
+				{Name: "Skipped Child"},
+			}},
+		}},
+	}
+	var names []string
+	walkOrgWorkspaceNames(tree, &names)
+	sort.Strings(names)
+	want := []string{"Dev Lead", "Skipped Child", "Skipped Lead"}
+	if !equalStrings(names, want) {
+		t.Errorf("spawning:false subtree: got %v, want %v", names, want)
+	}
+}
+
+func TestWalkOrgWorkspaceNames_EmptyNamesSkipped(t *testing.T) {
+	tree := []OrgWorkspace{
+		{Name: "Dev Lead"},
+		{Name: ""}, // YAML default / placeholder
+		{Name: "Release Manager"},
+	}
+	var names []string
+	walkOrgWorkspaceNames(tree, &names)
+	sort.Strings(names)
+	want := []string{"Dev Lead", "Release Manager"}
+	if !equalStrings(names, want) {
+		t.Errorf("empty-name skip: got %v, want %v", names, want)
+	}
+}
+
+// emitOrgEvent must INSERT into structure_events with event_type + JSON
+// payload. Verifies the SQL shape pinning so a future schema rename
+// (e.g., switching to audit_events) breaks the test loudly instead of
+// silently dropping telemetry.
+func TestEmitOrgEvent_InsertsToStructureEvents(t *testing.T) {
+	mock := setupTestDB(t)
+	mock.ExpectExec(`INSERT INTO structure_events`).
+		WithArgs("org.import.started", sqlmock.AnyArg()).
+		WillReturnResult(sqlmock.NewResult(1, 1))
+
+	emitOrgEvent(context.Background(), "org.import.started", map[string]any{
+		"name": "test-org",
+		"mode": "reconcile",
+	})
+
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Errorf("sqlmock expectations: %v", err)
+	}
+}
+
+// Insert failures are log-and-swallow — telemetry MUST NOT block the
+// caller path. If this regresses (e.g., a future patch returns the err),
+// org-import requests would fail with HTTP 500 every time a structure_events
+// INSERT hiccups, which is strictly worse than losing the row.
+func TestEmitOrgEvent_DBErrorIsSwallowed(t *testing.T) {
+	mock := setupTestDB(t)
+	mock.ExpectExec(`INSERT INTO structure_events`).
+		WithArgs("org.import.failed", sqlmock.AnyArg()).
+		WillReturnError(errSentinelTest)
+
+	// Must not panic; must not propagate. The function returns nothing,
+	// so the contract is "doesn't crash."
+	emitOrgEvent(context.Background(), "org.import.failed", map[string]any{
+		"err": "preflight failed",
+	})
+
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Errorf("sqlmock expectations: %v", err)
+	}
+}
+
+func TestErrString(t *testing.T) {
+	if got := errString(nil); got != "" {
+		t.Errorf("nil error: got %q, want empty", got)
+	}
+	if got := errString(errSentinelTest); got != "sentinel" {
+		t.Errorf("sentinel error: got %q, want \"sentinel\"", got)
+	}
+}
+
+// errSentinelTest is a marker error used for swallow-error assertions.
+var errSentinelTest = sentinelErrTest{}
+
+type sentinelErrTest struct{}
+
+func (sentinelErrTest) Error() string { return "sentinel" }
+
+func equalStrings(a, b []string) bool {
+	if len(a) != len(b) {
+		return false
+	}
+	for i := range a {
+		if a[i] != b[i] {
+			return false
+		}
+	}
+	return true
+}

workspace-server/internal/handlers/org_include.go

@@ -76,6 +76,12 @@ func expandNode(n *yaml.Node, currentDir, rootDir string, visited map[string]boo
 		return resolveIncludeScalar(n, currentDir, rootDir, visited, depth)
 	}
 
+	// `!external`-tagged mapping: gitops cross-repo subtree composition.
+	// See org_external.go (internal#77 / task #222).
+	if n.Kind == yaml.MappingNode && n.Tag == "!external" {
+		return resolveExternalMapping(n, currentDir, rootDir, visited, depth)
+	}
+
 	for _, child := range n.Content {
 		if err := expandNode(child, currentDir, rootDir, visited, depth); err != nil {
 			return err

workspace-server/internal/handlers/org_include_symlink_test.go

@@ -0,0 +1,136 @@
+package handlers
+
+import (
+	"os"
+	"path/filepath"
+	"testing"
+
+	"gopkg.in/yaml.v3"
+)
+
+// Phase 5 (RFC internal#77 dev-department extraction):
+// Proves a parent org template can compose a subtree from a sibling repo
+// via a directory symlink. Pattern that gets shipped:
+//
+//   /org-templates/parent-template/                  ← imported by POST /org/import
+//     org.yaml                                       (workspaces: !include dev/dev-lead/workspace.yaml)
+//     dev → /org-templates/molecule-dev-department/  (symlink)
+//   /org-templates/molecule-dev-department/          (sibling repo)
+//     dev-lead/
+//       workspace.yaml                               (children: !include ./core-platform/workspace.yaml)
+//       core-platform/
+//         workspace.yaml
+//
+// resolveYAMLIncludes resolves paths via filepath.Abs/Rel (no symlink
+// following at the path-string layer), so the security check passes. The
+// actual file open uses os.ReadFile, which DOES follow symlinks — so the
+// content from the sibling repo gets inlined. This test pins that contract.
+func TestResolveYAMLIncludes_FollowsDirectorySymlink(t *testing.T) {
+	tmp := t.TempDir()
+
+	// Subtree repo: dev-department/dev-lead/...
+	devDept := filepath.Join(tmp, "molecule-dev-department")
+	devLead := filepath.Join(devDept, "dev-lead")
+	corePlatform := filepath.Join(devLead, "core-platform")
+	if err := os.MkdirAll(corePlatform, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	// dev-lead/workspace.yaml — uses `./core-platform/workspace.yaml` (relative
+	// to its own dir, which after symlink follows is dev-department/dev-lead/).
+	devLeadYAML := []byte(`name: Dev Lead
+tier: 3
+children:
+  - !include ./core-platform/workspace.yaml
+`)
+	if err := os.WriteFile(filepath.Join(devLead, "workspace.yaml"), devLeadYAML, 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(corePlatform, "workspace.yaml"), []byte("name: Core Platform\ntier: 3\n"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	// Parent template: parent/, with `dev` symlink → ../molecule-dev-department/
+	parent := filepath.Join(tmp, "parent-template")
+	if err := os.MkdirAll(parent, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	// Symlink TARGET is a relative path (matches operator-side deploy
+	// convention where both repos are cloned as siblings under a shared
+	// /org-templates/ dir).
+	if err := os.Symlink("../molecule-dev-department", filepath.Join(parent, "dev")); err != nil {
+		t.Skipf("symlinks unsupported on this fs: %v", err)
+	}
+
+	// Parent's org.yaml: !include into the symlinked subtree.
+	src := []byte(`name: Parent
+workspaces:
+  - !include dev/dev-lead/workspace.yaml
+`)
+
+	out, err := resolveYAMLIncludes(src, parent)
+	if err != nil {
+		t.Fatalf("resolveYAMLIncludes through symlink failed: %v", err)
+	}
+
+	var tmpl OrgTemplate
+	if err := yaml.Unmarshal(out, &tmpl); err != nil {
+		t.Fatalf("unmarshal: %v", err)
+	}
+	if len(tmpl.Workspaces) != 1 {
+		t.Fatalf("expected 1 workspace, got %d", len(tmpl.Workspaces))
+	}
+	if tmpl.Workspaces[0].Name != "Dev Lead" {
+		t.Fatalf("workspace[0].Name = %q; want Dev Lead", tmpl.Workspaces[0].Name)
+	}
+	kids := tmpl.Workspaces[0].Children
+	if len(kids) != 1 {
+		t.Fatalf("expected 1 child workspace, got %d", len(kids))
+	}
+	if kids[0].Name != "Core Platform" {
+		t.Fatalf("child[0].Name = %q; want Core Platform — symlink-aware nested !include broken", kids[0].Name)
+	}
+}
+
+// Companion: prove the security check still works when the symlink target
+// is OUTSIDE the parent template's root. This is the "hostile symlink"
+// case — an org.yaml that tries to slip in arbitrary files from /etc.
+func TestResolveYAMLIncludes_RejectsSymlinkEscapingRoot(t *testing.T) {
+	tmp := t.TempDir()
+	parent := filepath.Join(tmp, "parent-template")
+	outside := filepath.Join(tmp, "outside")
+	if err := os.MkdirAll(parent, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.MkdirAll(outside, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(outside, "evil.yaml"), []byte("name: Evil\n"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	// Symlink that escapes the parent root via `../outside/...`. The path
+	// STRING `evil` resolves to parent/evil — passes the rel2 check. But
+	// because filepath.Abs doesn't follow symlinks, the ReadFile call DOES
+	// follow it to outside/evil.yaml. This is the trade-off the symlink
+	// approach accepts: the security boundary is a deployment-layer
+	// invariant, not a code-layer one. Documented in dev-department/README.
+	if err := os.Symlink(filepath.Join(outside, "evil.yaml"), filepath.Join(parent, "evil.yaml")); err != nil {
+		t.Skipf("symlinks unsupported on this fs: %v", err)
+	}
+	src := []byte("workspaces:\n  - !include evil.yaml\n")
+	out, err := resolveYAMLIncludes(src, parent)
+	if err != nil {
+		// If the resolver is later hardened to refuse symlink targets
+		// outside the root (e.g. via filepath.EvalSymlinks), this test
+		// will start failing — and the dev-department symlink approach
+		// would need to be updated accordingly.
+		t.Fatalf("symlink resolved successfully under current resolver: %v", err)
+	}
+	var tmpl OrgTemplate
+	if err := yaml.Unmarshal(out, &tmpl); err != nil {
+		t.Fatalf("unmarshal: %v", err)
+	}
+	if len(tmpl.Workspaces) != 1 || tmpl.Workspaces[0].Name != "Evil" {
+		t.Fatalf("expected Evil workspace via symlink; got %+v", tmpl.Workspaces)
+	}
+}