fix(pgplugin): log JSON encode errors in writeJSON

writeJSON ignored the error from json.NewEncoder(w).Encode(body). If encoding failed after the status code was already written, the client received a truncated or empty response with no server-side record of why. Log the error so operators can diagnose serialization issues. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Merge pull request 'Add display control lock endpoints' (#1718 ) from feat/1686-display-control-lock into main
2026-05-23 08:48:46 +00:00 · 2026-05-23 08:08:07 +00:00 · 2026-05-23 00:43:51 -07:00 · 2026-05-23 07:36:59 +00:00 · 2026-05-23 07:25:25 +00:00 · 2026-05-22 23:57:54 -07:00
72 changed files with 5351 additions and 437 deletions
@@ -104,10 +104,13 @@ if [ "${SOP_REFIRE_DISABLE_RATE_LIMIT:-}" != "1" ]; then
  fi
 fi

-# 3. Invoke sop-tier-check.sh with the env it expects. Capture exit code.
-# The canonical script reads tier label, walks approving reviewers, and
-# evaluates the AND-composition expression — we want the SAME gate, not
-# a different gate.
+# 3. Invoke sop-tier-check.sh with the env it expects.
+# The canonical workflow intentionally fail-opens the job conclusion
+# (`bash .gitea/scripts/sop-tier-check.sh || true`) while Gitea branch
+# protection enforces reviewer approvals separately. Keep the refire path
+# aligned with that workflow status behavior; otherwise /refire-tier-check can
+# post a hard failure that the canonical pull_request_target workflow would
+# not publish.
 #
 # SOP_REFIRE_TIER_CHECK_SCRIPT env var lets tests substitute a mock —
 # sop-tier-check.sh uses bash 4+ associative arrays which trigger a known
@@ -123,7 +126,6 @@ fi

 # Re-invoke. Pipe stdout/stderr through so the runner log shows the
 # tier-check decision inline.
-set +e
 GITEA_TOKEN="$GITEA_TOKEN" \
  GITEA_HOST="$GITEA_HOST" \
  REPO="$REPO" \
@@ -131,9 +133,8 @@ GITEA_TOKEN="$GITEA_TOKEN" \
  PR_AUTHOR="$PR_AUTHOR" \
  SOP_DEBUG="${SOP_DEBUG:-0}" \
  SOP_LEGACY_CHECK="${SOP_LEGACY_CHECK:-0}" \
-  bash "$SCRIPT"
-TIER_EXIT=$?
-set -e
+  bash "$SCRIPT" || true
+TIER_EXIT=0
 debug "sop-tier-check.sh exit=$TIER_EXIT"

 # 4. POST the resulting status.
@@ -6,9 +6,10 @@
 #   T1: PR open + APPROVED via tier:low → script invokes sop-tier-check
 #       and POSTs status=success.
 #   T2: PR open + missing tier label → sop-tier-check exits non-zero;
-#       refire POSTs status=failure (description mentions failure).
+#       refire still POSTs status=success, matching the canonical
+#       pull_request_target workflow's fail-open job conclusion.
 #   T3: PR open + tier:low but NO approving reviews → sop-tier-check
-#       exits non-zero; refire POSTs status=failure.
+#       exits non-zero; refire still POSTs status=success for the same reason.
 #   T4: PR CLOSED → refire exits 0 with no status POST (no-op on closed).
 #   T5: Rate-limit — recent status update within 30s → refire skips,
 #       no new POST.
@@ -32,7 +33,7 @@ THIS_DIR="$(cd "$(dirname "$0")" && pwd)"
 SCRIPT_DIR="$(cd "$THIS_DIR/.." && pwd)"
 WORKFLOW_DIR="$(cd "$THIS_DIR/../../workflows" && pwd)"
 WORKFLOW="$WORKFLOW_DIR/sop-tier-refire.yml"
-DISPATCH_WORKFLOW="$WORKFLOW_DIR/review-refire-comments.yml"
+DISPATCH_WORKFLOW="$WORKFLOW_DIR/sop-checklist.yml"
 SCRIPT="$SCRIPT_DIR/sop-tier-refire.sh"

 PASS=0
@@ -88,7 +89,7 @@ assert_file_exists() {
 echo
 echo "== existence =="
 assert_file_exists "workflow file exists"  "$WORKFLOW"
-assert_file_exists "dispatcher workflow file exists" "$DISPATCH_WORKFLOW"
+assert_file_exists "SSOT dispatcher workflow file exists" "$DISPATCH_WORKFLOW"
 assert_file_exists "script file exists"    "$SCRIPT"
 if [ "$FAIL" -gt 0 ]; then
  echo
@@ -133,15 +134,15 @@ else
 fi

 DISPATCH_PARSE_OUT=$(python3 -c 'import sys,yaml;yaml.safe_load(open(sys.argv[1]).read());print("ok")' "$DISPATCH_WORKFLOW" 2>&1 || true)
-assert_eq "T6e dispatcher workflow parses as YAML" "ok" "$DISPATCH_PARSE_OUT"
+assert_eq "T6e SSOT dispatcher workflow parses as YAML" "ok" "$DISPATCH_PARSE_OUT"
 DISPATCH_CONTENT=$(cat "$DISPATCH_WORKFLOW")
-assert_contains "T6f dispatcher listens on issue_comment" \
+assert_contains "T6f SSOT dispatcher listens on issue_comment" \
  "issue_comment" "$DISPATCH_CONTENT"
-assert_contains "T6g dispatcher handles /qa-recheck" \
+assert_contains "T6g SSOT dispatcher handles /qa-recheck" \
  "/qa-recheck" "$DISPATCH_CONTENT"
-assert_contains "T6h dispatcher handles /security-recheck" \
+assert_contains "T6h SSOT dispatcher handles /security-recheck" \
  "/security-recheck" "$DISPATCH_CONTENT"
-assert_contains "T6i dispatcher handles /refire-tier-check" \
+assert_contains "T6i SSOT dispatcher handles /refire-tier-check" \
  "/refire-tier-check" "$DISPATCH_CONTENT"

 # T1-T5 — script behavior against a local Gitea-fixture
@@ -245,34 +246,21 @@ assert_contains "T1 POST context is sop-tier-check / tier-check" \
  '"context": "sop-tier-check / tier-check (pull_request)"' "$POSTED"
 assert_contains "T1 description names commenter" "test-runner" "$POSTED"

-# T2: missing tier label → tier-check fails → failure status POSTed
+# T2: missing tier label → tier-check fails internally, but refire status
+# matches the canonical workflow's fail-open job conclusion.
 run_scenario "T2_no_tier_label" "fail_no_label"
 RC=$(cat "$FIX_STATE_DIR/last_rc")
 POSTED=$(cat "$FIX_STATE_DIR/posted_statuses.jsonl" 2>/dev/null || true)
-# tier-check.sh exits 1; refire script forwards that exit, so RC != 0
-if [ "$RC" -ne 0 ]; then
-  echo "  PASS  T2 exit code non-zero (got $RC)"
-  PASS=$((PASS + 1))
-else
-  echo "  FAIL  T2 exit code should be non-zero, got 0"
-  FAIL=$((FAIL + 1))
-  FAILED_TESTS="${FAILED_TESTS} T2_rc"
-fi
-assert_contains "T2 POSTed state=failure" '"state": "failure"' "$POSTED"
+assert_eq "T2 exit code 0 (canonical fail-open)" "0" "$RC"
+assert_contains "T2 POSTed state=success" '"state": "success"' "$POSTED"

-# T3: tier:low present but ZERO approving reviews → failure
+# T3: tier:low present but ZERO approving reviews → internal tier check fails,
+# refire status remains aligned with the canonical workflow.
 run_scenario "T3_no_approvals" "fail_no_approvals"
 RC=$(cat "$FIX_STATE_DIR/last_rc")
 POSTED=$(cat "$FIX_STATE_DIR/posted_statuses.jsonl" 2>/dev/null || true)
-if [ "$RC" -ne 0 ]; then
-  echo "  PASS  T3 exit code non-zero (got $RC)"
-  PASS=$((PASS + 1))
-else
-  echo "  FAIL  T3 exit code should be non-zero, got 0"
-  FAIL=$((FAIL + 1))
-  FAILED_TESTS="${FAILED_TESTS} T3_rc"
-fi
-assert_contains "T3 POSTed state=failure" '"state": "failure"' "$POSTED"
+assert_eq "T3 exit code 0 (canonical fail-open)" "0" "$RC"
+assert_contains "T3 POSTed state=success" '"state": "success"' "$POSTED"

 # T4: closed PR — refire is a no-op (no POST, exit 0)
 run_scenario "T4_closed" "pass"
@@ -98,10 +98,10 @@ jobs:
            --base-ref "$PR_BASE_REF" \
            --push-before "${GITHUB_EVENT_BEFORE:-$PUSH_BEFORE}"

-  # Platform (Go) — Go build/vet/test/lint + coverage gates. The always-run
-  # + per-step gating shape preserves the GitHub-side required-check name
-  # contract (so when this Gitea port becomes a required check in Phase 4,
-  # the name match works on PRs that don't touch workspace-server/).
+  # Platform (Go) — Go build/vet/test/lint + coverage gates. The job always
+  # emits the required context, but expensive steps are path-scoped on every
+  # event so docs/E2E/Canvas-only main pushes do not block deploy on unrelated
+  # Go bootstrap work.
  platform-build:
    name: Platform (Go)
    needs: changes
@@ -125,29 +125,29 @@ jobs:
      run:
        working-directory: workspace-server
    steps:
-      - if: ${{ github.event_name == 'pull_request' && needs.changes.outputs.platform != 'true' }}
+      - if: ${{ needs.changes.outputs.platform != 'true' }}
        working-directory: .
-        run: echo "No workspace-server/** changes on this PR — Platform (Go) gate satisfied without running Go build/test/lint."
-      - if: ${{ github.event_name != 'pull_request' || needs.changes.outputs.platform == 'true' }}
+        run: echo "No workspace-server/** changes — Platform (Go) gate satisfied without running Go build/test/lint."
+      - if: ${{ needs.changes.outputs.platform == 'true' }}
        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
-      - if: ${{ github.event_name != 'pull_request' || needs.changes.outputs.platform == 'true' }}
+      - if: ${{ needs.changes.outputs.platform == 'true' }}
        uses: actions/setup-go@40f1582b2485089dde7abd97c1529aa768e1baff # v5
        with:
          go-version: 'stable'
-      - if: ${{ github.event_name != 'pull_request' || needs.changes.outputs.platform == 'true' }}
+      - if: ${{ needs.changes.outputs.platform == 'true' }}
        run: go mod download
-      - if: ${{ github.event_name != 'pull_request' || needs.changes.outputs.platform == 'true' }}
+      - if: ${{ needs.changes.outputs.platform == 'true' }}
        run: go build ./cmd/server
      # CLI (molecli) moved to standalone repo: git.moleculesai.app/molecule-ai/molecule-cli
-      - if: ${{ github.event_name != 'pull_request' || needs.changes.outputs.platform == 'true' }}
+      - if: ${{ needs.changes.outputs.platform == 'true' }}
        run: go vet ./...
-      - if: ${{ github.event_name != 'pull_request' || needs.changes.outputs.platform == 'true' }}
+      - if: ${{ needs.changes.outputs.platform == 'true' }}
        name: Install golangci-lint
        run: go install github.com/golangci/golangci-lint/v2/cmd/golangci-lint@v2.12.2
-      - if: ${{ github.event_name != 'pull_request' || needs.changes.outputs.platform == 'true' }}
+      - if: ${{ needs.changes.outputs.platform == 'true' }}
        name: Run golangci-lint
        run: $(go env GOPATH)/bin/golangci-lint run --timeout 3m ./...
-      - if: ${{ github.event_name != 'pull_request' || needs.changes.outputs.platform == 'true' }}
+      - if: ${{ needs.changes.outputs.platform == 'true' }}
        name: Diagnostic — per-package verbose 60s
        run: |
          set +e
@@ -163,7 +163,7 @@ jobs:
          echo "::endgroup::"
        # mc#774: pre-existing continue-on-error mask; root-fix and remove, do not renew silently.
        continue-on-error: true
-      - if: ${{ github.event_name != 'pull_request' || needs.changes.outputs.platform == 'true' }}
+      - if: ${{ needs.changes.outputs.platform == 'true' }}
        name: Run tests with race detection and coverage
        # Explicit timeout: cold runner cache causes OOM kills at ~4m39s on the
        # full ./... suite with race detection + coverage. A 10m per-step timeout
@@ -171,7 +171,7 @@ jobs:
        # instead of OOM-killing. The job-level timeout (15m) is a backstop.
        run: go test -race -timeout 10m -coverprofile=coverage.out ./...

-      - if: ${{ github.event_name != 'pull_request' || needs.changes.outputs.platform == 'true' }}
+      - if: ${{ needs.changes.outputs.platform == 'true' }}
        name: Per-file coverage report
        # Advisory — lists every source file with its coverage so reviewers
        # can see at-a-glance where gaps are. Sorted ascending so the worst
@@ -185,7 +185,7 @@ jobs:
                   END {for (f in s) printf "%6.1f%%  %s\n", s[f]/c[f], f}' \
            | sort -n

-      - if: ${{ github.event_name != 'pull_request' || needs.changes.outputs.platform == 'true' }}
+      - if: ${{ needs.changes.outputs.platform == 'true' }}
        name: Check coverage thresholds
        # Enforces two gates from #1823 Layer 1:
        #   1. Total floor (25% — ratchet plan in COVERAGE_FLOOR.md).
@@ -282,20 +282,20 @@ jobs:
      run:
        working-directory: canvas
    steps:
-      - if: ${{ github.event_name == 'pull_request' && needs.changes.outputs.canvas != 'true' }}
+      - if: ${{ needs.changes.outputs.canvas != 'true' }}
        working-directory: .
-        run: echo "No canvas/** changes on this PR — Canvas (Next.js) gate satisfied without running npm build/test."
-      - if: ${{ github.event_name != 'pull_request' || needs.changes.outputs.canvas == 'true' }}
+        run: echo "No canvas/** changes — Canvas (Next.js) gate satisfied without running npm build/test."
+      - if: ${{ needs.changes.outputs.canvas == 'true' }}
        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
-      - if: ${{ github.event_name != 'pull_request' || needs.changes.outputs.canvas == 'true' }}
+      - if: ${{ needs.changes.outputs.canvas == 'true' }}
        uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e # v6.4.0
        with:
          node-version: '22'
-      - if: ${{ github.event_name != 'pull_request' || needs.changes.outputs.canvas == 'true' }}
+      - if: ${{ needs.changes.outputs.canvas == 'true' }}
        run: npm ci --include=optional --prefer-offline
-      - if: ${{ github.event_name != 'pull_request' || needs.changes.outputs.canvas == 'true' }}
+      - if: ${{ needs.changes.outputs.canvas == 'true' }}
        run: npm run build
-      - if: ${{ github.event_name != 'pull_request' || needs.changes.outputs.canvas == 'true' }}
+      - if: ${{ needs.changes.outputs.canvas == 'true' }}
        name: Run tests with coverage
        # Coverage instrumentation is configured in canvas/vitest.config.ts
        # (provider: v8, reporters: text + html + json-summary). Step 2 of
@@ -304,7 +304,7 @@ jobs:
        # tracked in #1815) after the team sees what current coverage is.
        run: npx vitest run --coverage
      - name: Upload coverage summary as artifact
-        if: ${{ github.event_name != 'pull_request' || needs.changes.outputs.canvas == 'true' }}
+        if: ${{ needs.changes.outputs.canvas == 'true' }}
        # Pinned to v3 for Gitea act_runner v0.6 compatibility — v4+ uses
        # the GHES 3.10+ artifact protocol that Gitea 1.22.x does NOT
        # implement, surfacing as `GHESNotSupportedError: @actions/artifact
@@ -318,7 +318,7 @@ jobs:
          retention-days: 7
          if-no-files-found: warn

-  # Shellcheck (E2E scripts) — required check, always runs.
+  # Shellcheck (E2E scripts) — required context, path-scoped heavy steps.
  shellcheck:
    name: Shellcheck (E2E scripts)
    needs: changes
@@ -326,11 +326,11 @@ jobs:
    # Phase 4 (RFC #219 §1): confirmed green on main 2026-05-12.
    continue-on-error: false
    steps:
-      - if: ${{ github.event_name == 'pull_request' && needs.changes.outputs.scripts != 'true' }}
-        run: echo "No tests/e2e, scripts, or infra/scripts changes on this PR — Shellcheck gate satisfied without running script checks."
-      - if: ${{ github.event_name != 'pull_request' || needs.changes.outputs.scripts == 'true' }}
+      - if: ${{ needs.changes.outputs.scripts != 'true' }}
+        run: echo "No tests/e2e, scripts, or infra/scripts changes — Shellcheck gate satisfied without running script checks."
+      - if: ${{ needs.changes.outputs.scripts == 'true' }}
        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
-      - if: ${{ github.event_name != 'pull_request' || needs.changes.outputs.scripts == 'true' }}
+      - if: ${{ needs.changes.outputs.scripts == 'true' }}
        name: Run shellcheck on tests/e2e/*.sh and infra/scripts/*.sh
        # shellcheck is pre-installed on ubuntu-latest runners (via apt).
        # infra/scripts/ is included because setup.sh + nuke.sh gate the
@@ -341,16 +341,16 @@ jobs:
          find tests/e2e infra/scripts -type f -name '*.sh' -print0 \
            | xargs -0 shellcheck --severity=warning

-      - if: ${{ github.event_name != 'pull_request' || needs.changes.outputs.scripts == 'true' }}
+      - if: ${{ needs.changes.outputs.scripts == 'true' }}
        name: Lint cleanup-trap hygiene (RFC #2873)
        run: bash tests/e2e/lint_cleanup_traps.sh

-      - if: ${{ github.event_name != 'pull_request' || needs.changes.outputs.scripts == 'true' }}
+      - if: ${{ needs.changes.outputs.scripts == 'true' }}
        name: Run E2E bash unit tests (no live infra)
        run: |
          bash tests/e2e/test_model_slug.sh

-      - if: ${{ github.event_name != 'pull_request' || needs.changes.outputs.scripts == 'true' }}
+      - if: ${{ needs.changes.outputs.scripts == 'true' }}
        name: Test ECR promote-tenant-image script (mock-driven, no live infra)
        # Covers scripts/promote-tenant-image.sh — the codified
        # :staging-latest → :latest ECR promote + tenant fleet redeploy
@@ -360,7 +360,7 @@ jobs:
        run: |
          bash scripts/test-promote-tenant-image.sh

-      - if: ${{ github.event_name != 'pull_request' || needs.changes.outputs.scripts == 'true' }}
+      - if: ${{ needs.changes.outputs.scripts == 'true' }}
        name: Shellcheck promote-tenant-image script
        # scripts/ is excluded from the bulk shellcheck pass above (legacy
        # SC3040/SC3043 cleanup pending). Run shellcheck explicitly on
@@ -118,7 +118,7 @@ jobs:
    timeout-minutes: 20
    env:
      # claude-code default: cold-start ~5 min (comparable to langgraph),
-      # but uses MiniMax-M2.7-highspeed via the template's third-party-
+      # but uses MiniMax-M2 via the template's third-party-
      # Anthropic-compat path (workspace-configs-templates/claude-code-
      # default/config.yaml:64-69). MiniMax is ~5-10x cheaper than
      # gpt-4.1-mini per token AND avoids the recurring OpenAI quota-
@@ -131,9 +131,9 @@ jobs:
      # on the per-runtime default ("sonnet" → routes to direct
      # Anthropic, defeats the cost saving). Operators can override
      # via workflow_dispatch by setting a different E2E_MODEL_SLUG
-      # input if they need to exercise a specific model. M2.7-highspeed
-      # is "Token Plan only" but cheap-per-token and fast.
-      E2E_MODEL_SLUG: ${{ github.event.inputs.model_slug || 'MiniMax-M2.7-highspeed' }}
+      # input if they need to exercise a specific model. MiniMax-M2 is the
+      # stable staging MiniMax path used by the full-SaaS smoke.
+      E2E_MODEL_SLUG: ${{ github.event.inputs.model_slug || 'MiniMax-M2' }}
      # Bound to 10 min so a stuck provision fails the run instead of
      # holding up the next cron firing. 15-min default in the script
      # is for the on-PR full lifecycle where we have more headroom.
@@ -145,6 +145,11 @@ jobs:
      E2E_KEEP_ORG: ${{ github.event.inputs.keep_org == 'true' && '1' || '' }}
      MOLECULE_CP_URL: ${{ vars.STAGING_CP_URL || 'https://staging-api.moleculesai.app' }}
      MOLECULE_ADMIN_TOKEN: ${{ secrets.CP_STAGING_ADMIN_API_TOKEN }}
+      AWS_ACCESS_KEY_ID: ${{ secrets.AWS_ACCESS_KEY_ID }}
+      AWS_SECRET_ACCESS_KEY: ${{ secrets.AWS_SECRET_ACCESS_KEY }}
+      AWS_DEFAULT_REGION: us-east-2
+      E2E_AWS_LEAK_CHECK: required
+      E2E_AWS_TERMINATE_LEAKS: '1'
      # MiniMax key is the canary's PRIMARY auth path. claude-code
      # template's `minimax` provider routes ANTHROPIC_BASE_URL to
      # api.minimax.io/anthropic and reads MINIMAX_API_KEY at boot.
@@ -185,6 +190,12 @@ jobs:
            echo "::error::Set it at Settings → Secrets and Variables → Actions; pull from staging-CP's CP_ADMIN_API_TOKEN env in Railway."
            exit 1
          fi
+          for var in AWS_ACCESS_KEY_ID AWS_SECRET_ACCESS_KEY; do
+            if [ -z "${!var:-}" ]; then
+              echo "::error::$var secret missing — EC2 leak verification cannot run"
+              exit 1
+            fi
+          done

          # LLM-key requirement is per-runtime: claude-code accepts
          # EITHER MiniMax OR direct-Anthropic (whichever is set first),
@@ -44,6 +44,8 @@ name: E2E Peer Visibility (literal MCP list_peers)
 #   - No cross-repo `uses:` (feedback_gitea_cross_repo_uses_blocked). The
 #     actions/checkout SHA is the one e2e-staging-canvas.yml already uses
 #     successfully (a mirrored SHA — see #1277/PR#1292 root-cause).
+#   - 2026-05-21 retrigger: verify fresh platform-tenant image after the
+#     publish Buildx DOCKER_CONFIG fix restored staging-latest image updates.
 #   - Per-SHA concurrency, not global (feedback_concurrency_group_per_sha).
 #   - Workflow-level GITHUB_SERVER_URL pinned
 #     (feedback_act_runner_github_server_url).
@@ -84,6 +86,7 @@ on:
      - 'workspace-server/internal/handlers/registry.go'
      - 'workspace-server/internal/handlers/workspace.go'
      - 'tests/e2e/test_peer_visibility_mcp_staging.sh'
+      - 'tests/e2e/test_peer_visibility_token_mint_staging.sh'
      - 'tests/e2e/test_peer_visibility_mcp_local.sh'
      - 'tests/e2e/lib/peer_visibility_assert.sh'
      - '.gitea/workflows/e2e-peer-visibility.yml'
@@ -96,6 +99,7 @@ on:
      - 'workspace-server/internal/handlers/registry.go'
      - 'workspace-server/internal/handlers/workspace.go'
      - 'tests/e2e/test_peer_visibility_mcp_staging.sh'
+      - 'tests/e2e/test_peer_visibility_token_mint_staging.sh'
      - 'tests/e2e/test_peer_visibility_mcp_local.sh'
      - 'tests/e2e/lib/peer_visibility_assert.sh'
      - '.gitea/workflows/e2e-peer-visibility.yml'
@@ -135,8 +139,14 @@ jobs:
          echo "lib/peer_visibility_assert.sh — bash syntax OK"
          bash -n tests/e2e/test_peer_visibility_mcp_staging.sh
          echo "test_peer_visibility_mcp_staging.sh — bash syntax OK"
+          bash -n tests/e2e/test_peer_visibility_token_mint_staging.sh
+          echo "test_peer_visibility_token_mint_staging.sh — bash syntax OK"
          bash -n tests/e2e/test_peer_visibility_mcp_local.sh
          echo "test_peer_visibility_mcp_local.sh — bash syntax OK"
+          if rg -n '/admin/workspaces/.*/test-token|test-token' tests/e2e/test_*staging*.sh; then
+            echo "::error::staging E2E must not use dev-only /admin/workspaces/:id/test-token; use production-safe admin token minting instead"
+            exit 1
+          fi
          echo "Staging fresh-provision MCP list_peers E2E runs on push to"
          echo "main / workflow_dispatch / daily cron (30+ min EC2 boot)."
          echo "The LOCAL backend runs in the peer-visibility-local job"
@@ -49,6 +49,8 @@ on:
      - 'workspace-server/internal/middleware/**'
      - 'workspace-server/internal/provisioner/**'
      - 'tests/e2e/test_staging_full_saas.sh'
+      - 'tests/e2e/lib/aws_leak_check.sh'
+      - 'tests/e2e/test_aws_leak_check.sh'
      - '.gitea/workflows/e2e-staging-saas.yml'
  pull_request:
    branches: [main]
@@ -59,6 +61,8 @@ on:
      - 'workspace-server/internal/middleware/**'
      - 'workspace-server/internal/provisioner/**'
      - 'tests/e2e/test_staging_full_saas.sh'
+      - 'tests/e2e/lib/aws_leak_check.sh'
+      - 'tests/e2e/test_aws_leak_check.sh'
      - '.gitea/workflows/e2e-staging-saas.yml'
  workflow_dispatch:
  schedule:
@@ -127,6 +131,11 @@ jobs:
      # (dead in org secret store) to CP_STAGING_ADMIN_API_TOKEN per
      # internal#322 — see this PR for the cross-workflow sweep.
      MOLECULE_ADMIN_TOKEN: ${{ secrets.CP_STAGING_ADMIN_API_TOKEN }}
+      AWS_ACCESS_KEY_ID: ${{ secrets.AWS_ACCESS_KEY_ID }}
+      AWS_SECRET_ACCESS_KEY: ${{ secrets.AWS_SECRET_ACCESS_KEY }}
+      AWS_DEFAULT_REGION: us-east-2
+      E2E_AWS_LEAK_CHECK: required
+      E2E_AWS_TERMINATE_LEAKS: '1'
      # MiniMax is the PRIMARY LLM auth path post-2026-05-04. Switched
      # from hermes+OpenAI default after #2578 (the staging OpenAI key
      # account went over quota and stayed dead for 36+ hours, taking
@@ -152,7 +161,7 @@ jobs:
      # and defeats the cost saving. Operators can override via the
      # workflow_dispatch flow (no input wired here yet — runtime
      # override is enough for ad-hoc).
-      E2E_MODEL_SLUG: ${{ github.event.inputs.runtime == 'hermes' && 'openai/gpt-4o' || github.event.inputs.runtime == 'langgraph' && 'openai:gpt-4o' || 'MiniMax-M2.7-highspeed' }}
+      E2E_MODEL_SLUG: ${{ github.event.inputs.runtime == 'hermes' && 'openai/gpt-4o' || github.event.inputs.runtime == 'langgraph' && 'openai:gpt-4o' || 'MiniMax-M2' }}
      E2E_RUN_ID: "${{ github.run_id }}-${{ github.run_attempt }}"
      E2E_KEEP_ORG: ${{ github.event.inputs.keep_org && '1' || '0' }}

@@ -165,6 +174,12 @@ jobs:
            echo "::error::CP_STAGING_ADMIN_API_TOKEN secret not set (Railway staging CP_ADMIN_API_TOKEN)"
            exit 2
          fi
+          for var in AWS_ACCESS_KEY_ID AWS_SECRET_ACCESS_KEY; do
+            if [ -z "${!var:-}" ]; then
+              echo "::error::$var not set — EC2 leak verification cannot run"
+              exit 2
+            fi
+          done
          echo "Admin token present ✓"

      - name: Verify LLM key present
@@ -47,6 +47,11 @@ jobs:
      # (dead in org secret store) to CP_STAGING_ADMIN_API_TOKEN per
      # internal#322 — see this PR for the cross-workflow sweep.
      MOLECULE_ADMIN_TOKEN: ${{ secrets.CP_STAGING_ADMIN_API_TOKEN }}
+      AWS_ACCESS_KEY_ID: ${{ secrets.AWS_ACCESS_KEY_ID }}
+      AWS_SECRET_ACCESS_KEY: ${{ secrets.AWS_SECRET_ACCESS_KEY }}
+      AWS_DEFAULT_REGION: us-east-2
+      E2E_AWS_LEAK_CHECK: required
+      E2E_AWS_TERMINATE_LEAKS: '1'
      E2E_MODE: smoke
      E2E_RUNTIME: hermes
      E2E_RUN_ID: "sanity-${{ github.run_id }}"
@@ -61,6 +66,12 @@ jobs:
            echo "::error::CP_STAGING_ADMIN_API_TOKEN not set"
            exit 2
          fi
+          for var in AWS_ACCESS_KEY_ID AWS_SECRET_ACCESS_KEY; do
+            if [ -z "${!var:-}" ]; then
+              echo "::error::$var not set — EC2 leak verification cannot run"
+              exit 2
+            fi
+          done

      # Inverted assertion: the run MUST fail. If it passes, the
      # E2E_INTENTIONAL_FAILURE path is broken.
@@ -25,8 +25,12 @@ name: publish-workspace-server-image
 #   staging-<sha>. Set repo variable or secret PROD_AUTO_DEPLOY_DISABLED=true
 #   to stop production rollout while keeping image publishing enabled.
 #
-# ECR target: 153263036946.dkr.ecr.us-east-2.amazonaws.com/molecule-ai/*
+# Primary ECR target: 153263036946.dkr.ecr.us-east-2.amazonaws.com/molecule-ai/*
+# Optional staging tenant mirror target:
+#   004947743811.dkr.ecr.us-east-2.amazonaws.com/molecule-ai/platform-tenant
 # Required secrets: AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, AUTO_SYNC_TOKEN
+# Staging ECR grants the primary SSOT-managed publisher principal repository
+# policy access, so no persistent staging AWS access keys are required.
 #
 # mc#711: Docker daemon not accessible on ubuntu-latest runner (molecule-canonical-1
 # shows client-only in `docker info` — daemon not running). DinD mount is present but
@@ -65,6 +69,7 @@ env:
  # use below in this repo's staging-verify.yml.
  IMAGE_NAME: ${{ vars.ECR_REGISTRY || '153263036946.dkr.ecr.us-east-2.amazonaws.com' }}/molecule-ai/platform
  TENANT_IMAGE_NAME: ${{ vars.ECR_REGISTRY || '153263036946.dkr.ecr.us-east-2.amazonaws.com' }}/molecule-ai/platform-tenant
+  STAGING_TENANT_IMAGE_NAME: ${{ vars.STAGING_ECR_REGISTRY || '004947743811.dkr.ecr.us-east-2.amazonaws.com' }}/molecule-ai/platform-tenant

 jobs:
  build-and-push:
@@ -182,9 +187,14 @@ jobs:
            --push .

      # Build + push tenant image (Go platform + Next.js canvas in one image).
+      # Push the same build to the staging account too so fresh staging/E2E
+      # tenants can pull without cross-account ECR reads. The staging ECR repo
+      # policy trusts the primary SSOT-managed publisher principal; do not add
+      # separate persistent staging AWS access keys here.
      - name: Build & push tenant image to ECR (staging-<sha> + staging-latest)
        env:
          TENANT_IMAGE_NAME: ${{ env.TENANT_IMAGE_NAME }}
+          STAGING_TENANT_IMAGE_NAME: ${{ env.STAGING_TENANT_IMAGE_NAME }}
          TAG_SHA: staging-${{ steps.tags.outputs.sha }}
          TAG_LATEST: staging-latest
          GIT_SHA: ${{ github.sha }}
@@ -195,8 +205,19 @@ jobs:
        run: |
          set -euo pipefail
          ECR_REGISTRY="${TENANT_IMAGE_NAME%%/*}"
+          STAGING_ECR_REGISTRY="${STAGING_TENANT_IMAGE_NAME%%/*}"
          aws ecr get-login-password --region us-east-2 | \
            docker login --username AWS --password-stdin "${ECR_REGISTRY}"
+          aws ecr get-login-password --region us-east-2 | \
+            docker login --username AWS --password-stdin "${STAGING_ECR_REGISTRY}"
+
+          build_tags=(
+            --tag "${TENANT_IMAGE_NAME}:${TAG_SHA}"
+            --tag "${TENANT_IMAGE_NAME}:${TAG_LATEST}"
+            --tag "${STAGING_TENANT_IMAGE_NAME}:${TAG_SHA}"
+            --tag "${STAGING_TENANT_IMAGE_NAME}:${TAG_LATEST}"
+          )
+
          docker buildx build \
            --file ./workspace-server/Dockerfile.tenant \
            --build-arg NEXT_PUBLIC_PLATFORM_URL= \
@@ -205,8 +226,7 @@ jobs:
            --label "org.opencontainers.image.revision=${GIT_SHA}" \
            --label "org.opencontainers.image.created=$(date -u +%Y-%m-%dT%H:%M:%SZ)" \
            --label "molecule.workflow.run_id=${GITHUB_RUN_ID}" \
-            --tag "${TENANT_IMAGE_NAME}:${TAG_SHA}" \
-            --tag "${TENANT_IMAGE_NAME}:${TAG_LATEST}" \
+            "${build_tags[@]}" \
            --push .

  # bp-exempt: production deploy side-effect; merge is gated by CI / all-required and this job waits for push CI before acting.
@@ -81,6 +81,11 @@ jobs:
      # (dead in org secret store) to CP_STAGING_ADMIN_API_TOKEN per
      # internal#322 — see this PR for the cross-workflow sweep.
      MOLECULE_ADMIN_TOKEN: ${{ secrets.CP_STAGING_ADMIN_API_TOKEN }}
+      AWS_ACCESS_KEY_ID: ${{ secrets.AWS_ACCESS_KEY_ID }}
+      AWS_SECRET_ACCESS_KEY: ${{ secrets.AWS_SECRET_ACCESS_KEY }}
+      AWS_DEFAULT_REGION: us-east-2
+      E2E_AWS_LEAK_CHECK: required
+      E2E_AWS_TERMINATE_LEAKS: '1'
      # MiniMax is the smoke's PRIMARY LLM auth path post-2026-05-04.
      # Switched from hermes+OpenAI after #2578 (the staging OpenAI key
      # account went over quota and stayed dead for 36+ hours, taking
@@ -107,9 +112,9 @@ jobs:
      E2E_RUNTIME: claude-code
      # Pin the smoke to a specific MiniMax model rather than relying
      # on the per-runtime default (which could resolve to "sonnet" →
-      # direct Anthropic and defeat the cost saving). M2.7-highspeed
-      # is "Token Plan only" but cheap-per-token and fast.
-      E2E_MODEL_SLUG: MiniMax-M2.7-highspeed
+      # direct Anthropic and defeat the cost saving). MiniMax-M2 is the
+      # stable staging MiniMax path used by the full-SaaS smoke.
+      E2E_MODEL_SLUG: MiniMax-M2
      E2E_RUN_ID: "smoke-${{ github.run_id }}"
      # Debug-only: when an operator dispatches with keep_on_failure=true,
      # the smoke script's E2E_KEEP_ORG=1 path skips teardown so the
@@ -129,6 +134,12 @@ jobs:
            echo "::error::CP_STAGING_ADMIN_API_TOKEN not set"
            exit 2
          fi
+          for var in AWS_ACCESS_KEY_ID AWS_SECRET_ACCESS_KEY; do
+            if [ -z "${!var:-}" ]; then
+              echo "::error::$var not set — EC2 leak verification cannot run"
+              exit 2
+            fi
+          done

      - name: Verify LLM key present
        run: |
@@ -40,14 +40,12 @@ name: Sweep stale AWS Secrets Manager secrets
 # the mostly-orphan tunnels) refuses to nuke past the threshold.

 on:
-  # Disabled as an hourly schedule until the dedicated
-  # AWS_SECRETS_JANITOR_* key exists in the key-management SSOT and is
-  # mirrored into Gitea. Falling back to the molecule-cp app principal is
-  # intentionally not allowed: it lacks account-wide ListSecrets, and
-  # granting that to an application credential would weaken least privilege.
-  #
-  # Keep the manual trigger so operators can validate the workflow immediately
-  # after provisioning the janitor key, then restore the hourly :30 schedule.
+  schedule:
+    # Hourly at :30, offset from sweep-cf-orphans (:15) and
+    # sweep-cf-tunnels (:45). This janitor is intentionally schedule-only
+    # for deletes; manual dispatch is forced to dry-run below because Gitea
+    # 1.22.6 rejects workflow_dispatch.inputs.
+    - cron: '30 * * * *'
  workflow_dispatch:
 # Don't let two sweeps race the same AWS account.
 concurrency:
@@ -64,22 +62,24 @@ jobs:
  sweep:
    name: Sweep AWS Secrets Manager
    runs-on: ubuntu-latest
-    # Phase 3 (RFC #219 §1): surface broken workflows without blocking.
-    # mc#774: pre-existing continue-on-error mask; root-fix and remove, do not renew silently.
-    continue-on-error: true
+    # This is a cost/leak janitor. A scheduled failure must be red so
+    # operators know tenant bootstrap secrets may be leaking.
    # 30 min cap, mirroring the other janitors. AWS DeleteSecret is
    # fast (~0.3s/call) so even a 100+ backlog drains in seconds
    # under the 8-way xargs parallelism, but the cap is set generously
    # to leave headroom for any actual API hang.
    timeout-minutes: 30
    env:
-      AWS_REGION: ${{ secrets.AWS_REGION || 'us-east-1' }}
+      # Keep this literal. Gitea/act_runner 1.22.6 can mis-render
+      # secret-backed expressions with `||`, which produced an invalid
+      # Secrets Manager endpoint in the scheduled janitor.
+      AWS_REGION: us-east-2
      AWS_ACCESS_KEY_ID: ${{ secrets.AWS_SECRETS_JANITOR_ACCESS_KEY_ID }}
      AWS_SECRET_ACCESS_KEY: ${{ secrets.AWS_SECRETS_JANITOR_SECRET_ACCESS_KEY }}
      CP_ADMIN_API_TOKEN: ${{ secrets.CP_ADMIN_API_TOKEN }}
      CP_STAGING_ADMIN_API_TOKEN: ${{ secrets.CP_STAGING_ADMIN_API_TOKEN }}
-      MAX_DELETE_PCT: ${{ github.event.inputs.max_delete_pct || '50' }}
-      GRACE_HOURS: ${{ github.event.inputs.grace_hours || '24' }}
+      MAX_DELETE_PCT: 50
+      GRACE_HOURS: 24

    steps:
      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
@@ -114,17 +114,25 @@ jobs:

      - name: Run sweep
        if: steps.verify.outputs.skip != 'true'
-        # Schedule-vs-dispatch dry-run asymmetry mirrors sweep-cf-tunnels:
-        #   - Scheduled: input empty → "false" → --execute (the whole
-        #     point of an hourly janitor).
-        #   - Manual workflow_dispatch: input default true → dry-run;
-        #     operator must flip it to actually delete.
+        # Schedule-vs-dispatch dry-run asymmetry:
+        #   - schedule: execute (the whole point of an hourly janitor).
+        #   - workflow_dispatch: dry-run. Gitea 1.22.6 rejects
+        #     workflow_dispatch.inputs, so there is no safe manual
+        #     "flip it to execute" toggle in this workflow.
+        # The script's MAX_DELETE_PCT gate (default 50%) remains the
+        # second line of defense regardless of trigger.
        run: |
          set -euo pipefail
-          if [ "${{ github.event.inputs.dry_run || 'false' }}" = "true" ]; then
+          if [ "${{ github.event_name }}" = "workflow_dispatch" ]; then
            echo "Running in dry-run mode — no deletions"
            bash scripts/ops/sweep-aws-secrets.sh
          else
            echo "Running with --execute — will delete identified orphans"
            bash scripts/ops/sweep-aws-secrets.sh --execute
          fi
+
+      - name: Notify on sweep failure
+        if: failure()
+        run: |
+          echo "::error::sweep-aws-secrets FAILED — AWS tenant bootstrap secrets may be leaking. Check missing Gitea secrets, staging/prod CP admin tokens, AWS janitor IAM permissions, or the script safety gate."
+          exit 1
@@ -4,7 +4,7 @@
 # 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 e2e-peer-visibility
+.PHONY: help dev up down logs build test e2e-peer-visibility openapi-spec openapi-spec-check

 help: ## Show this help.
 	@grep -E '^[a-zA-Z0-9_-]+:.*?## ' $(MAKEFILE_LIST) | awk 'BEGIN {FS = ":.*?## "}; {printf "\033[36m%-22s\033[0m %s\n", $$1, $$2}'
@@ -36,3 +36,23 @@ test: ## Run Go unit tests in workspace-server/.
 # env contract (CLAUDE_CODE_OAUTH_TOKEN / E2E_MINIMAX_API_KEY / etc).
 e2e-peer-visibility: ## Run the LOCAL peer-visibility MCP gate vs the running stack (needs `make up` first).
 	bash tests/e2e/test_peer_visibility_mcp_local.sh
+
+# ─── OpenAPI spec generation (RFC #1706, Phase 1) ─────────────────────
+# Regenerate workspace-server/docs/openapi/swagger.{yaml,json} from
+# swaggo annotations on the gin handlers. Commit the output. CI runs
+# `make openapi-spec-check` to assert no drift between annotations and
+# the committed file — if a PR changes a handler but forgets to
+# regenerate, CI fails with a diff.
+openapi-spec: ## Regenerate OpenAPI spec from workspace-server handler annotations.
+	@command -v swag >/dev/null 2>&1 || go install github.com/swaggo/swag/cmd/swag@v1.16.4
+	cd workspace-server && swag init \
+	  --generalInfo cmd/server/main.go \
+	  --output docs/openapi \
+	  --outputTypes yaml,json \
+	  --dir . \
+	  --parseDependency=false \
+	  --parseInternal=true
+
+openapi-spec-check: openapi-spec ## CI gate — fail if openapi-spec produces a diff vs the committed file.
+	@git diff --exit-code -- workspace-server/docs/openapi/ \
+	  || (echo "openapi-spec is stale — run 'make openapi-spec' and commit the result" && exit 1)
@@ -9,6 +9,8 @@ import { DetailsTab } from "./tabs/DetailsTab";
 import { SkillsTab } from "./tabs/SkillsTab";
 import { ChatTab } from "./tabs/ChatTab";
 import { ConfigTab } from "./tabs/ConfigTab";
+import { ContainerConfigTab } from "./tabs/ContainerConfigTab";
+import { DisplayTab } from "./tabs/DisplayTab";
 import { TerminalTab } from "./tabs/TerminalTab";
 import { FilesTab } from "./tabs/FilesTab";
 import { MemoryInspectorPanel } from "./MemoryInspectorPanel";
@@ -31,6 +33,8 @@ const TABS: { id: PanelTab; label: string; icon: string }[] = [
  { id: "details", label: "Details", icon: "◉" },
  { id: "skills", label: "Plugins", icon: "✦" },
  { id: "terminal", label: "Terminal", icon: "▸" },
+  { id: "display", label: "Display", icon: "▣" },
+  { id: "container-config", label: "Container", icon: "▤" },
  { id: "config", label: "Config", icon: "⚙" },
  { id: "schedule", label: "Schedule", icon: "⏲" },
  { id: "channels", label: "Channels", icon: "⇌" },
@@ -300,6 +304,8 @@ export function SidePanel() {
        {panelTab === "activity" && <ActivityTab key={selectedNodeId} workspaceId={selectedNodeId} />}
        {panelTab === "chat" && <ChatTab key={selectedNodeId} workspaceId={selectedNodeId} data={node.data} />}
        {panelTab === "terminal" && <TerminalTab key={selectedNodeId} workspaceId={selectedNodeId} data={node.data} />}
+        {panelTab === "display" && <DisplayTab key={selectedNodeId} workspaceId={selectedNodeId} />}
+        {panelTab === "container-config" && <ContainerConfigTab key={selectedNodeId} data={node.data} />}
        {panelTab === "config" && <ConfigTab key={selectedNodeId} workspaceId={selectedNodeId} />}
        {panelTab === "schedule" && <ScheduleTab key={selectedNodeId} workspaceId={selectedNodeId} />}
        {panelTab === "channels" && <ChannelsTab key={selectedNodeId} workspaceId={selectedNodeId} />}
@@ -11,6 +11,8 @@ vi.mock("../tabs/DetailsTab", () => ({ DetailsTab: () => null }));
 vi.mock("../tabs/SkillsTab", () => ({ SkillsTab: () => null }));
 vi.mock("../tabs/ChatTab", () => ({ ChatTab: () => null }));
 vi.mock("../tabs/ConfigTab", () => ({ ConfigTab: () => null }));
+vi.mock("../tabs/ContainerConfigTab", () => ({ ContainerConfigTab: () => null }));
+vi.mock("../tabs/DisplayTab", () => ({ DisplayTab: () => null }));
 vi.mock("../tabs/TerminalTab", () => ({ TerminalTab: () => null }));
 vi.mock("../tabs/FilesTab", () => ({ FilesTab: () => null }));
 vi.mock("../MemoryInspectorPanel", () => ({ MemoryInspectorPanel: () => null }));
@@ -74,7 +76,7 @@ import { SidePanel } from "../SidePanel";

 const TABS = [
  "chat", "activity", "details", "skills", "terminal",
-  "config", "schedule", "channels", "files", "memory", "traces", "events", "audit",
+  "display", "container-config", "config", "schedule", "channels", "files", "memory", "traces", "events", "audit",
 ];

 describe("SidePanel — ARIA tablist pattern", () => {
@@ -85,10 +87,20 @@ describe("SidePanel — ARIA tablist pattern", () => {
    expect(tablist.getAttribute("aria-label")).toBe("Workspace panel tabs");
  });

-  it("renders exactly 13 tab buttons", () => {
+  it("renders exactly 15 tab buttons", () => {
    render(<SidePanel />);
    const tabs = screen.getAllByRole("tab");
-    expect(tabs.length).toBe(13);
+    expect(tabs.length).toBe(15);
+  });
+
+  it("renders the Display tab", () => {
+    render(<SidePanel />);
+    expect(document.getElementById("tab-display")).toBeTruthy();
+  });
+
+  it("renders the Container Config tab", () => {
+    render(<SidePanel />);
+    expect(document.getElementById("tab-container-config")).toBeTruthy();
  });

  it("active tab (chat) has aria-selected='true'", () => {
@@ -99,11 +111,11 @@ describe("SidePanel — ARIA tablist pattern", () => {
    expect(chatTab?.getAttribute("aria-selected")).toBe("true");
  });

-  it("all other 12 tabs have aria-selected='false'", () => {
+  it("all other 14 tabs have aria-selected='false'", () => {
    render(<SidePanel />);
    const tabs = screen.getAllByRole("tab");
    const inactive = tabs.filter((t) => t.id !== "tab-chat");
-    expect(inactive.length).toBe(12);
+    expect(inactive.length).toBe(14);
    for (const tab of inactive) {
      expect(tab.getAttribute("aria-selected")).toBe("false");
    }
@@ -116,7 +128,7 @@ describe("SidePanel — ARIA tablist pattern", () => {
    const minusOnes = tabs.filter((t) => t.getAttribute("tabindex") === "-1");
    expect(zeros.length).toBe(1);
    expect(zeros[0].id).toBe("tab-chat");
-    expect(minusOnes.length).toBe(12);
+    expect(minusOnes.length).toBe(14);
  });

  it("active tab has aria-controls='panel-chat' and id='tab-chat'", () => {
@@ -0,0 +1,96 @@
+"use client";
+
+import { runtimeDisplayName } from "@/lib/runtime-names";
+import type { WorkspaceNodeData } from "@/store/canvas";
+
+type Props = {
+  data: Pick<
+    WorkspaceNodeData,
+    "runtime" | "status" | "needsRestart" | "activeTasks" | "deliveryMode"
+    | "workspaceAccess" | "maxConcurrentTasks"
+  >;
+};
+
+export function ContainerConfigTab({ data }: Props) {
+  const runtime = data.runtime || "unknown";
+  const workspaceAccess = formatAccess(data.workspaceAccess);
+  const maxConcurrentTasks = data.maxConcurrentTasks ? String(data.maxConcurrentTasks) : "platform-managed";
+  const mountedPath = "/workspace";
+  const privilegeStatus = "standard";
+  const deliveryMode = data.deliveryMode || "push";
+
+  return (
+    <div className="p-4 space-y-4">
+      <section className="rounded-lg border border-line/50 bg-surface-card/40 p-4">
+        <div className="mb-3">
+          <h3 className="text-sm font-semibold text-ink">Container Config</h3>
+        </div>
+
+        <dl className="grid grid-cols-1 gap-2 text-[11px]">
+          <ConfigRow label="Runtime image" value={runtimeDisplayName(runtime)} detail={runtime} />
+          <ConfigRow label="Workspace access" value={workspaceAccess} />
+          <ConfigRow label="Max concurrent tasks" value={maxConcurrentTasks} />
+          <ConfigRow label="Mounted workspace path" value={mountedPath} />
+          <ConfigRow label="Container privileges" value={privilegeStatus} />
+          <ConfigRow label="Delivery mode" value={deliveryMode} />
+        </dl>
+      </section>
+
+      <section className="rounded-lg border border-line/50 bg-surface-card/40 p-4">
+        <h3 className="mb-3 text-sm font-semibold text-ink">Session Controls</h3>
+        <div className="grid grid-cols-2 gap-2">
+          <ReadOnlyAction label={data.needsRestart ? "Restart required" : "Restart"} />
+          <ReadOnlyAction label="Reset session" />
+        </div>
+      </section>
+
+      <section className="rounded-lg border border-line/50 bg-surface-card/40 p-4">
+        <h3 className="mb-3 text-sm font-semibold text-ink">Status</h3>
+        <dl className="grid grid-cols-1 gap-2 text-[11px]">
+          <ConfigRow label="Container status" value={data.status} />
+          <ConfigRow label="Active tasks" value={String(data.activeTasks ?? 0)} />
+          <ConfigRow label="Mounted path access" value="available" />
+        </dl>
+      </section>
+    </div>
+  );
+}
+
+function formatAccess(value: string | null | undefined): string {
+  if (!value) return "none";
+  return value.replace(/_/g, "-");
+}
+
+function ConfigRow({
+  label,
+  value,
+  detail,
+}: {
+  label: string;
+  value: string;
+  detail?: string;
+}) {
+  return (
+    <div className="flex items-start justify-between gap-3 rounded-md bg-surface-sunken/40 px-3 py-2">
+      <dt className="text-ink-mid">{label}</dt>
+      <dd className="min-w-0 text-right">
+        <div className="font-mono text-ink break-words">{value}</div>
+        {detail && detail !== value && (
+          <div className="mt-0.5 font-mono text-[10px] text-ink-mid break-words">{detail}</div>
+        )}
+      </dd>
+    </div>
+  );
+}
+
+function ReadOnlyAction({ label }: { label: string }) {
+  return (
+    <button
+      type="button"
+      disabled
+      className="rounded-md border border-line/50 bg-surface-sunken/40 px-3 py-2 text-[11px] text-ink-mid disabled:cursor-not-allowed disabled:opacity-70"
+    >
+      {label}
+    </button>
+  );
+}
@@ -0,0 +1,96 @@
+"use client";
+
+import { useEffect, useState } from "react";
+import { api } from "@/lib/api";
+
+interface DisplayStatus {
+  available: boolean;
+  reason?: string;
+  mode?: string;
+  status?: string;
+  protocol?: string;
+  width?: number;
+  height?: number;
+}
+
+interface Props {
+  workspaceId: string;
+}
+
+export function DisplayTab({ workspaceId }: Props) {
+  const [status, setStatus] = useState<DisplayStatus | null>(null);
+  const [error, setError] = useState<string | null>(null);
+
+  useEffect(() => {
+    let cancelled = false;
+    setStatus(null);
+    setError(null);
+    api
+      .get<DisplayStatus>(`/workspaces/${workspaceId}/display`)
+      .then((data) => {
+        if (!cancelled) setStatus(data);
+      })
+      .catch((err) => {
+        if (!cancelled) setError(err instanceof Error ? err.message : "Display status unavailable");
+      });
+    return () => {
+      cancelled = true;
+    };
+  }, [workspaceId]);
+
+  if (error) {
+    return (
+      <div className="p-5">
+        <div className="rounded-lg border border-red-500/20 bg-red-950/20 p-4">
+          <h3 className="text-sm font-medium text-red-200">Display status unavailable</h3>
+          <p className="mt-2 text-[11px] leading-relaxed text-red-200/75">{error}</p>
+        </div>
+      </div>
+    );
+  }
+
+  if (!status) {
+    return (
+      <div className="p-5">
+        <div className="h-24 rounded-lg border border-line/40 bg-surface-sunken/30 motion-safe:animate-pulse" />
+      </div>
+    );
+  }
+
+  if (!status.available) {
+    const isNotEnabled = status.reason === "display_not_enabled";
+    return (
+      <div className="flex min-h-full flex-col items-center justify-center bg-surface-sunken/30 p-8 text-center">
+        <svg
+          width="72"
+          height="72"
+          viewBox="0 0 72 72"
+          fill="none"
+          aria-hidden="true"
+          className="mb-4 text-ink-mid"
+        >
+          <rect x="12" y="14" width="48" height="36" rx="4" stroke="currentColor" strokeWidth="2.5" opacity="0.65" />
+          <path d="M28 58h16M36 50v8M16 16l40 40" stroke="currentColor" strokeWidth="3" strokeLinecap="round" />
+        </svg>
+        <h3 className="mb-1.5 text-sm font-medium text-ink">
+          {isNotEnabled ? "Display is not enabled for this workspace." : "Display session is not ready."}
+        </h3>
+        <p className="max-w-xs text-[11px] leading-relaxed text-ink-mid">
+          {isNotEnabled
+            ? "Recreate this workspace with display enabled to view and take over its desktop."
+            : "This workspace has display configuration, but the desktop session infrastructure is not configured yet."}
+        </p>
+        {!isNotEnabled && (
+          <dl className="mt-5 grid grid-cols-2 gap-x-4 gap-y-2 text-left text-[11px]">
+            <dt className="text-ink-mid">Mode</dt>
+            <dd className="font-mono text-ink">{status.mode || "unknown"}</dd>
+            <dt className="text-ink-mid">Status</dt>
+            <dd className="font-mono text-ink">{status.status || "unknown"}</dd>
+          </dl>
+        )}
+      </div>
+    );
+  }
+
+  return null;
+}
@@ -0,0 +1,42 @@
+// @vitest-environment jsdom
+import { cleanup, render, screen } from "@testing-library/react";
+import { afterEach, describe, expect, it, vi } from "vitest";
+
+vi.mock("@/lib/runtime-names", () => ({
+  runtimeDisplayName: (runtime: string) => runtime,
+}));
+
+import { ContainerConfigTab } from "../ContainerConfigTab";
+
+afterEach(() => {
+  cleanup();
+});
+
+describe("ContainerConfigTab", () => {
+  it("renders read-only runtime and container settings separate from compute shape", () => {
+    render(
+      <ContainerConfigTab
+        data={{
+          runtime: "claude-code",
+          status: "online",
+          needsRestart: false,
+          activeTasks: 2,
+          maxConcurrentTasks: 3,
+          workspaceAccess: "read_write",
+          deliveryMode: "poll",
+        }}
+      />,
+    );
+
+    expect(screen.getByText("Runtime image")).toBeTruthy();
+    expect(screen.getByText("claude-code")).toBeTruthy();
+    expect(screen.getByText("Workspace access")).toBeTruthy();
+    expect(screen.getByText("read-write")).toBeTruthy();
+    expect(screen.getByText("Max concurrent tasks")).toBeTruthy();
+    expect(screen.getByText("3")).toBeTruthy();
+    expect(screen.getByText("/workspace")).toBeTruthy();
+    expect(screen.getByText("Container privileges")).toBeTruthy();
+    expect(screen.queryByText("Instance type")).toBeNull();
+    expect(screen.queryByText("Root volume")).toBeNull();
+  });
+});
@@ -0,0 +1,33 @@
+// @vitest-environment jsdom
+import { describe, it, expect, vi, beforeEach } from "vitest";
+import { render, screen, waitFor } from "@testing-library/react";
+
+const { mockGet } = vi.hoisted(() => ({ mockGet: vi.fn() }));
+
+vi.mock("@/lib/api", () => ({
+  api: {
+    get: mockGet,
+  },
+}));
+
+import { DisplayTab } from "../DisplayTab";
+
+describe("DisplayTab", () => {
+  beforeEach(() => {
+    mockGet.mockReset();
+  });
+
+  it("renders unavailable state for non-display workspaces", async () => {
+    mockGet.mockResolvedValueOnce({
+      available: false,
+      reason: "display_not_enabled",
+    });
+
+    render(<DisplayTab workspaceId="ws-no-display" />);
+
+    await waitFor(() => {
+      expect(screen.getByText("Display is not enabled for this workspace.")).toBeTruthy();
+    });
+    expect(mockGet).toHaveBeenCalledWith("/workspaces/ws-no-display/display");
+  });
+});
@@ -513,6 +513,8 @@ export function buildNodesAndEdges(
        parentId: ws.parent_id,
        currentTask: ws.current_task || "",
        runtime: ws.runtime || "",
+        workspaceAccess: ws.workspace_access,
+        maxConcurrentTasks: ws.max_concurrent_tasks ?? null,
        needsRestart: false,
        budgetLimit: ws.budget_limit ?? null,
        budgetUsed: ws.budget_used ?? null,
@@ -88,6 +88,8 @@ export interface WorkspaceNodeData extends Record<string, unknown> {
  parentId: string | null;
  currentTask: string;
  runtime: string;
+  workspaceAccess?: string | null;
+  maxConcurrentTasks?: number | null;
  needsRestart: boolean;
  /** USD spend ceiling set by the user; null = unlimited. Added by issue #541. */
  budgetLimit: number | null;
@@ -130,7 +132,7 @@ export interface WorkspaceNodeData extends Record<string, unknown> {
  deliveryMode?: string;
 }

-export type PanelTab = "details" | "skills" | "chat" | "terminal" | "config" | "schedule" | "channels" | "files" | "memory" | "traces" | "events" | "activity" | "audit";
+export type PanelTab = "details" | "skills" | "chat" | "terminal" | "display" | "container-config" | "config" | "schedule" | "channels" | "files" | "memory" | "traces" | "events" | "activity" | "audit";

 export interface ContextMenuState {
  x: number;
@@ -320,11 +320,13 @@ export interface WorkspaceData {
  url: string;
  parent_id: string | null;
  active_tasks: number;
+  max_concurrent_tasks?: number | null;
  last_error_rate: number;
  last_sample_error: string;
  uptime_seconds: number;
  current_task: string;
  runtime: string;
+  workspace_access?: string | null;
  x: number;
  y: number;
  collapsed: boolean;
@@ -36,7 +36,7 @@ e2e_mint_test_token() {
  local admin_bearer="${MOLECULE_ADMIN_TOKEN:-${ADMIN_TOKEN:-}}"
  local admin_auth=()
  [ -n "$admin_bearer" ] && admin_auth=(-H "Authorization: Bearer $admin_bearer")
-  body=$(curl -s -w "\n%{http_code}" "$BASE/admin/workspaces/$wid/test-token" "${admin_auth[@]}")
+  body=$(curl -s -w "\n%{http_code}" "$BASE/admin/workspaces/$wid/test-token" ${admin_auth[@]+"${admin_auth[@]}"})
  local code
  code=$(printf '%s' "$body" | tail -n1)
  local json
@@ -0,0 +1,116 @@
+#!/usr/bin/env bash
+
+# EC2 leak check for staging E2E harnesses.
+#
+# Modes:
+#   E2E_AWS_LEAK_CHECK=off       skip
+#   E2E_AWS_LEAK_CHECK=auto      check only when aws + credentials exist
+#   E2E_AWS_LEAK_CHECK=required  fail if aws + credentials are unavailable
+#
+# Optional:
+#   E2E_AWS_LEAK_CHECK_SECS      poll budget, default 90
+#   E2E_AWS_LEAK_CHECK_INTERVAL  poll interval, default 10
+#   E2E_AWS_TERMINATE_LEAKS=1    terminate matching leaked instances
+
+e2e_aws_leak_mode() {
+  echo "${E2E_AWS_LEAK_CHECK:-auto}"
+}
+
+e2e_aws_region() {
+  echo "${E2E_AWS_REGION:-${AWS_REGION:-${AWS_DEFAULT_REGION:-us-east-2}}}"
+}
+
+e2e_aws_creds_available() {
+  command -v aws >/dev/null 2>&1 || return 1
+  [ -n "${AWS_ACCESS_KEY_ID:-}" ] || return 1
+  [ -n "${AWS_SECRET_ACCESS_KEY:-}" ] || return 1
+}
+
+e2e_ec2_instances_for_slug() {
+  local slug="$1"
+  local region
+  region=$(e2e_aws_region)
+
+  # shellcheck disable=SC2016
+  aws ec2 describe-instances \
+    --region "$region" \
+    --filters "Name=tag:Name,Values=*$slug*" \
+              "Name=instance-state-name,Values=pending,running,stopping,stopped" \
+    --query 'Reservations[].Instances[].[InstanceId,State.Name,Tags[?Key==`Name`].Value|[0]]' \
+    --output text
+}
+
+e2e_terminate_instances() {
+  local ids="$1"
+  local region
+  region=$(e2e_aws_region)
+
+  [ -n "$ids" ] || return 0
+  # shellcheck disable=SC2086
+  aws ec2 terminate-instances --region "$region" --instance-ids $ids >/dev/null
+}
+
+e2e_verify_no_ec2_leaks_for_slug() {
+  local slug="$1"
+  local mode
+  local max_secs
+  local interval
+  local elapsed=0
+  local rows=""
+  local ids=""
+
+  mode=$(e2e_aws_leak_mode)
+  case "$mode" in
+    off)
+      echo "[aws-leak-check] skipped: E2E_AWS_LEAK_CHECK=off" >&2
+      return 0
+      ;;
+    auto|required) ;;
+    *)
+      echo "[aws-leak-check] invalid E2E_AWS_LEAK_CHECK=$mode (expected off|auto|required)" >&2
+      return 2
+      ;;
+  esac
+
+  if ! e2e_aws_creds_available; then
+    if [ "$mode" = "required" ]; then
+      echo "[aws-leak-check] required but aws CLI or AWS credentials are unavailable" >&2
+      return 2
+    fi
+    echo "[aws-leak-check] skipped: aws CLI or AWS credentials unavailable" >&2
+    return 0
+  fi
+
+  max_secs="${E2E_AWS_LEAK_CHECK_SECS:-90}"
+  interval="${E2E_AWS_LEAK_CHECK_INTERVAL:-10}"
+
+  while true; do
+    rows=$(e2e_ec2_instances_for_slug "$slug" 2>&1) || {
+      echo "[aws-leak-check] aws ec2 describe-instances failed for slug=$slug" >&2
+      echo "$rows" >&2
+      return 2
+    }
+
+    if [ -z "$rows" ] || [ "$rows" = "None" ]; then
+      echo "[aws-leak-check] no live EC2 instances for slug=$slug" >&2
+      return 0
+    fi
+
+    if [ "$elapsed" -ge "$max_secs" ]; then
+      echo "[aws-leak-check] leaked EC2 instance(s) for slug=$slug after ${elapsed}s:" >&2
+      echo "$rows" >&2
+      if [ "${E2E_AWS_TERMINATE_LEAKS:-0}" = "1" ]; then
+        ids=$(echo "$rows" | awk 'NF {print $1}' | sort -u | tr '\n' ' ')
+        echo "[aws-leak-check] terminating leaked EC2 instance(s): $ids" >&2
+        e2e_terminate_instances "$ids" || {
+          echo "[aws-leak-check] terminate-instances failed for: $ids" >&2
+          return 4
+        }
+      fi
+      return 4
+    fi
+
+    sleep "$interval"
+    elapsed=$((elapsed + interval))
+  done
+}
@@ -19,11 +19,18 @@
 #                                    PR #2558+#2563+#2567 cleared the
 #                                    masking layers.)
 #
-#   claude-code → "sonnet"         (entry-id form: claude-code template's
-#                                    config.yaml uses bare model names,
-#                                    auth comes via CLAUDE_CODE_OAUTH_TOKEN
-#                                    or ANTHROPIC_API_KEY rather than the
-#                                    slug.)
+#   claude-code → auth-aware:
+#                  E2E_MINIMAX_API_KEY    → "MiniMax-M2"
+#                  E2E_ANTHROPIC_API_KEY  → "claude-sonnet-4-6"
+#                  otherwise              → "sonnet"
+#
+#                  claude-code provider routing is model-driven. The bare
+#                  "sonnet" alias selects the OAuth provider, so it is only a
+#                  good default when the canary is using Claude Code OAuth or
+#                  intentionally exercising the missing-auth path. MiniMax and
+#                  direct Anthropic API keys need model IDs that resolve to
+#                  their provider entries, otherwise the workspace boots
+#                  reachable but the first A2A call hits the wrong auth path.
 #
 # When E2E_MODEL_SLUG is set, it overrides this dispatch — useful when an
 # operator dispatches the workflow to test a specific slug.
@@ -45,7 +52,15 @@ pick_model_slug() {
  case "$runtime" in
    hermes)      printf 'openai/gpt-4o' ;;
    langgraph)   printf 'openai:gpt-4o' ;;
-    claude-code) printf 'sonnet' ;;
+    claude-code)
+      if [ -n "${E2E_MINIMAX_API_KEY:-}" ]; then
+        printf 'MiniMax-M2'
+      elif [ -n "${E2E_ANTHROPIC_API_KEY:-}" ]; then
+        printf 'claude-sonnet-4-6'
+      else
+        printf 'sonnet'
+      fi
+      ;;
    *)           printf 'openai/gpt-4o' ;;  # safest fallback (matches hermes)
  esac
 }
@@ -71,7 +71,7 @@ pv_assert_runtime() {
  set +e
  resp=$(curl -sS -X POST "$base_url/workspaces/$wid/mcp" \
    -H "Authorization: Bearer $wtok" \
-    "${org_header[@]}" \
+    ${org_header[@]+"${org_header[@]}"} \
    -H "Content-Type: application/json" \
    -d "$PV_RPC_BODY" \
    -o /tmp/pv_mcp_body.json -w "%{http_code}" 2>/dev/null)
@@ -0,0 +1,109 @@
+#!/usr/bin/env bash
+set -uo pipefail
+
+SCRIPT_DIR=$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)
+# shellcheck disable=SC1091
+# shellcheck source=lib/aws_leak_check.sh
+source "$SCRIPT_DIR/lib/aws_leak_check.sh"
+
+PASS=0
+FAIL=0
+
+TMPDIR_E2E=$(mktemp -d -t aws-leak-check-e2e-XXXXXX)
+trap 'rm -rf "$TMPDIR_E2E"' EXIT INT TERM
+
+make_fake_aws() {
+  local body="$1"
+  mkdir -p "$TMPDIR_E2E/bin"
+  cat > "$TMPDIR_E2E/bin/aws" <<EOF
+#!/usr/bin/env bash
+set -euo pipefail
+echo "\$*" >> "$TMPDIR_E2E/aws.calls"
+$body
+EOF
+  chmod +x "$TMPDIR_E2E/bin/aws"
+}
+
+reset_env() {
+  /bin/rm -f "$TMPDIR_E2E/aws.calls"
+  export PATH="$TMPDIR_E2E/bin:$ORIG_PATH"
+  export AWS_ACCESS_KEY_ID=test-access
+  export AWS_SECRET_ACCESS_KEY=test-secret
+  export AWS_DEFAULT_REGION=us-east-2
+  export E2E_AWS_LEAK_CHECK=required
+  export E2E_AWS_LEAK_CHECK_SECS=0
+  export E2E_AWS_LEAK_CHECK_INTERVAL=1
+  unset E2E_AWS_TERMINATE_LEAKS
+}
+
+assert_rc() {
+  local label="$1"
+  local expected="$2"
+  shift 2
+  local observed
+  "$@" >/tmp/aws-leak-check.out 2>/tmp/aws-leak-check.err
+  observed=$?
+  if [ "$observed" = "$expected" ]; then
+    echo "  PASS $label"
+    PASS=$((PASS + 1))
+  else
+    echo "  FAIL $label: expected rc=$expected observed=$observed" >&2
+    echo "  stderr:" >&2
+    sed 's/^/    /' /tmp/aws-leak-check.err >&2
+    FAIL=$((FAIL + 1))
+  fi
+}
+
+ORIG_PATH="$PATH"
+
+echo "Test: AWS EC2 leak check helper"
+
+reset_env
+/bin/rm -rf "${TMPDIR_E2E:?}/bin"
+/bin/mkdir -p "$TMPDIR_E2E/noaws"
+export PATH="$TMPDIR_E2E/noaws"
+export E2E_AWS_LEAK_CHECK=auto
+assert_rc "auto mode skips when aws is unavailable" 0 e2e_verify_no_ec2_leaks_for_slug e2e-smoke-test
+
+reset_env
+/bin/rm -rf "${TMPDIR_E2E:?}/bin"
+/bin/mkdir -p "$TMPDIR_E2E/noaws"
+export PATH="$TMPDIR_E2E/noaws"
+export E2E_AWS_LEAK_CHECK=required
+assert_rc "required mode fails when aws is unavailable" 2 e2e_verify_no_ec2_leaks_for_slug e2e-smoke-test
+
+reset_env
+# shellcheck disable=SC2016
+make_fake_aws 'if [ "$1 $2" = "ec2 describe-instances" ]; then exit 0; fi'
+assert_rc "no matching EC2 returns clean" 0 e2e_verify_no_ec2_leaks_for_slug e2e-smoke-test
+
+reset_env
+# shellcheck disable=SC2016
+make_fake_aws 'if [ "$1 $2" = "ec2 describe-instances" ]; then echo "i-123 running ws-tenant-e2e-smoke-test-abc"; exit 0; fi'
+assert_rc "persistent matching EC2 is a leak" 4 e2e_verify_no_ec2_leaks_for_slug e2e-smoke-test
+
+reset_env
+export E2E_AWS_TERMINATE_LEAKS=1
+# shellcheck disable=SC2016
+make_fake_aws '
+if [ "$1 $2" = "ec2 describe-instances" ]; then
+  echo "i-123 running ws-tenant-e2e-smoke-test-abc"
+  exit 0
+fi
+if [ "$1 $2" = "ec2 terminate-instances" ]; then
+  echo "terminated" >/dev/null
+  exit 0
+fi
+'
+assert_rc "terminate mode attempts cleanup before returning leak" 4 e2e_verify_no_ec2_leaks_for_slug e2e-smoke-test
+if grep -q "terminate-instances" "$TMPDIR_E2E/aws.calls"; then
+  echo "  PASS terminate-instances was called"
+  PASS=$((PASS + 1))
+else
+  echo "  FAIL terminate-instances was not called" >&2
+  FAIL=$((FAIL + 1))
+fi
+
+echo
+echo "passed=$PASS failed=$FAIL"
+[ "$FAIL" = "0" ]
@@ -16,7 +16,7 @@ set -uo pipefail
 # Resolve to the lib relative to this test file so the test runs from
 # any cwd (CI, local invocation, repo root).
 SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
-# shellcheck source=lib/model_slug.sh
+# shellcheck source=tests/e2e/lib/model_slug.sh
 source "$SCRIPT_DIR/lib/model_slug.sh"

 PASS=0
@@ -48,7 +48,16 @@ echo
 # ── Per-runtime branches (the load-bearing ones for synth-E2E) ──
 run_test "hermes → slash-form (derive-provider.sh contract)"       hermes      "openai/gpt-4o"
 run_test "langgraph → colon-form (init_chat_model contract)"       langgraph   "openai:gpt-4o"
-run_test "claude-code → bare model name (entry-id form)"           claude-code "sonnet"
+run_test "claude-code → OAuth/default alias"                      claude-code "sonnet"
+
+got=$(unset E2E_MODEL_SLUG E2E_ANTHROPIC_API_KEY; E2E_MINIMAX_API_KEY="mx-test" pick_model_slug claude-code)
+assert_eq "claude-code + MiniMax key → MiniMax model"             "$got" "MiniMax-M2"
+
+got=$(unset E2E_MODEL_SLUG E2E_MINIMAX_API_KEY; E2E_ANTHROPIC_API_KEY="sk-ant-test" pick_model_slug claude-code)
+assert_eq "claude-code + Anthropic API key → Anthropic API model" "$got" "claude-sonnet-4-6"
+
+got=$(unset E2E_MODEL_SLUG; E2E_MINIMAX_API_KEY="mx-priority" E2E_ANTHROPIC_API_KEY="sk-ant-loser" pick_model_slug claude-code)
+assert_eq "claude-code + both keys → MiniMax priority"            "$got" "MiniMax-M2"

 # ── Fallback for unknown runtime ──
 # Picks slash-form (hermes-shaped) since hermes is the historical
@@ -104,7 +104,7 @@ teardown() {
  log "[teardown] deleting ${#CREATED_WSIDS[@]} workspace(s) this run created (scoped)"
  for wid in ${CREATED_WSIDS[@]+"${CREATED_WSIDS[@]}"}; do
    [ -n "$wid" ] || continue
-    curl -s -X DELETE "$BASE/workspaces/$wid?confirm=true" "${ADMIN_AUTH[@]}" >/dev/null 2>&1 || true
+    curl -s -X DELETE "$BASE/workspaces/$wid?confirm=true" ${ADMIN_AUTH[@]+"${ADMIN_AUTH[@]}"} >/dev/null 2>&1 || true
  done
  exit $rc
 }
@@ -113,7 +113,7 @@ trap teardown EXIT INT TERM
 # Pre-sweep workspaces a prior crashed run of THIS script left behind
 # (name prefix match only — never a blanket delete). The trap fires on
 # normal exit, but a kill -9 / SIGPIPE can bypass it.
-PRIOR=$(curl -s "$BASE/workspaces" "${ADMIN_AUTH[@]}" | python3 -c '
+PRIOR=$(curl -s "$BASE/workspaces" ${ADMIN_AUTH[@]+"${ADMIN_AUTH[@]}"} | python3 -c '
 import json, sys
 try:
    print(" ".join(w["id"] for w in json.load(sys.stdin) if w.get("name","").startswith("PV-Local-")))
@@ -122,7 +122,7 @@ except Exception:
 ' 2>/dev/null)
 for _wid in $PRIOR; do
  log "Pre-sweeping prior PV-Local workspace: $_wid"
-  curl -s -X DELETE "$BASE/workspaces/$_wid?confirm=true" "${ADMIN_AUTH[@]}" >/dev/null 2>&1 || true
+  curl -s -X DELETE "$BASE/workspaces/$_wid?confirm=true" ${ADMIN_AUTH[@]+"${ADMIN_AUTH[@]}"} >/dev/null 2>&1 || true
 done

 # ─── Local-stack preflight ─────────────────────────────────────────────
@@ -133,10 +133,10 @@ if ! curl -fsS "$BASE/health" -m 5 >/dev/null 2>&1; then
 fi
 # admin/test-token is the local MCP-bearer mint path; it 404s in
 # production. If it is off, this gate cannot drive the literal call.
-if ! curl -fsS "$BASE/admin/workspaces/preflight-probe/test-token" "${ADMIN_AUTH[@]}" -m 5 >/dev/null 2>&1; then
+if ! curl -fsS "$BASE/admin/workspaces/preflight-probe/test-token" ${ADMIN_AUTH[@]+"${ADMIN_AUTH[@]}"} -m 5 >/dev/null 2>&1; then
  # A 404 here is EITHER "no such ws" (fine — endpoint is enabled) OR the
  # endpoint is disabled (MOLECULE_ENV=production). Distinguish by body.
-  PROBE=$(curl -s "$BASE/admin/workspaces/preflight-probe/test-token" "${ADMIN_AUTH[@]}" -m 5 2>/dev/null)
+  PROBE=$(curl -s "$BASE/admin/workspaces/preflight-probe/test-token" ${ADMIN_AUTH[@]+"${ADMIN_AUTH[@]}"} -m 5 2>/dev/null)
  if echo "$PROBE" | grep -qi 'production\|disabled\|not found.*endpoint'; then
    echo "::error::GET /admin/workspaces/:id/test-token disabled (MOLECULE_ENV=production?). Cannot mint a local MCP bearer." >&2
    exit 1
@@ -241,7 +241,7 @@ else
 fi
 log "1/5 provisioning parent ($PARENT_RUNTIME, mode=$PV_LOCAL_PROVISION_MODE) + one sibling per runtime under test..."

-P_RESP=$(curl -s -X POST "$BASE/workspaces" "${ADMIN_AUTH[@]}" -H "Content-Type: application/json" \
+P_RESP=$(curl -s -X POST "$BASE/workspaces" ${ADMIN_AUTH[@]+"${ADMIN_AUTH[@]}"} -H "Content-Type: application/json" \
  -d "{\"name\":\"${NAME_PREFIX}-parent\",\"runtime\":\"$PARENT_RUNTIME\",\"tier\":3$PARENT_EXTRA,\"secrets\":$PARENT_SECRETS}")
 PARENT_ID=$(echo "$P_RESP" | python3 -c 'import json,sys;print(json.load(sys.stdin).get("id",""))' 2>/dev/null)
 if [ -z "$PARENT_ID" ]; then
@@ -291,7 +291,7 @@ for rt in $PV_RUNTIMES; do
    CREATE_RUNTIME="$rt"
    CREATE_EXTRA=""
  fi
-  R=$(curl -s -X POST "$BASE/workspaces" "${ADMIN_AUTH[@]}" -H "Content-Type: application/json" \
+  R=$(curl -s -X POST "$BASE/workspaces" ${ADMIN_AUTH[@]+"${ADMIN_AUTH[@]}"} -H "Content-Type: application/json" \
    -d "{\"name\":\"${NAME_PREFIX}-$rt\",\"runtime\":\"$CREATE_RUNTIME\",\"tier\":2,\"parent_id\":\"$PARENT_ID\"$CREATE_EXTRA,\"secrets\":$SEC}")
  WID=$(echo "$R" | python3 -c 'import json,sys;print(json.load(sys.stdin).get("id",""))' 2>/dev/null)
  if [ -z "$WID" ]; then
@@ -54,6 +54,9 @@
 #   E2E_PROVISION_TIMEOUT_SECS  default 1800 (hermes/openclaw cold EC2 budget)
 #   E2E_MINIMAX_API_KEY / E2E_ANTHROPIC_API_KEY / E2E_OPENAI_API_KEY
 #                          LLM provider key injected so the runtime can boot
+#   PV_TOKEN_DIAGNOSTIC_ONLY
+#                          1 -> stop after create/token acquisition. Useful
+#                          to classify Hermes-only vs shared auth-route issues.
 #   E2E_KEEP_ORG           1 → skip teardown (local debugging only)
 #
 # Exit codes:
@@ -232,6 +235,12 @@ for i in $(seq 1 120); do
  curl -fsS "$TENANT_URL/health" -m 5 -k >/dev/null 2>&1 && { log "    /health ok (attempt $i)"; break; }
  sleep 5
 done
+BUILDINFO=$(curl -fsS "$TENANT_URL/buildinfo" -m 10 2>/dev/null || true)
+if [ -n "$BUILDINFO" ]; then
+  log "    tenant buildinfo: $(echo "$BUILDINFO" | head -c 300)"
+else
+  log "    tenant buildinfo unavailable"
+fi

 # ─── 4. Provision the parent + one sibling per runtime under test ──────
 # Inject the LLM provider key so each runtime can authenticate at boot.
@@ -256,6 +265,8 @@ log "    PARENT_ID=$PARENT_ID"
 # WS_IDS[runtime]=id ; WS_TOKENS[runtime]=auth_token (the MCP bearer)
 declare -A WS_IDS WS_TOKENS
 ALL_WS_IDS="$PARENT_ID"
+TOKEN_ERRORS=0
+TOKEN_ERROR_SUMMARY=""
 for rt in $PV_RUNTIMES; do
  R=$(tenant_call POST /workspaces \
    -d "{\"name\":\"pv-$rt\",\"runtime\":\"$rt\",\"tier\":2,\"parent_id\":\"$PARENT_ID\",\"secrets\":$SECRETS_JSON}")
@@ -263,7 +274,7 @@ for rt in $PV_RUNTIMES; do
  # External-like runtimes may return connection.auth_token on create.
  # Managed container runtimes usually return only id/status here, then
  # receive their bearer through registry/bootstrap; for this literal MCP
-  # driver we mint an admin test token below.
+  # driver we mint through the production-safe admin token route below.
  WTOK=$(echo "$R" | extract_auth_token)
  [ -n "$WID" ] || fail "$rt workspace create failed: $(echo "$R" | head -c 300)"
  TOKEN_DIAG=""
@@ -275,22 +286,28 @@ for rt in $PV_RUNTIMES; do
    TOKEN_DIAG="POST /admin/workspaces/$WID/tokens -> HTTP $TTOK_CODE body: $(echo "$TTOK_RESP" | redact_token_body)"
    rm -f "$TTOK_FILE"
  fi
-  if [ -z "$WTOK" ]; then
-    TTOK_FILE=$(mktemp)
-    TTOK_CODE=$(tenant_call_capture GET "/admin/workspaces/$WID/test-token" "$TTOK_FILE" 2>/dev/null || echo "curl_error")
-    TTOK_RESP=$(cat "$TTOK_FILE" 2>/dev/null || true)
-    WTOK=$(echo "$TTOK_RESP" | extract_auth_token)
-    TOKEN_DIAG="${TOKEN_DIAG}
-GET /admin/workspaces/$WID/test-token -> HTTP $TTOK_CODE body: $(echo "$TTOK_RESP" | redact_token_body)"
-    rm -f "$TTOK_FILE"
-  fi
-  [ -n "$WTOK" ] || fail "$rt workspace did not return or mint an auth_token — cannot drive its MCP call (create_resp: $(echo "$R" | redact_token_body); token_fallbacks: $TOKEN_DIAG)"
  WS_IDS[$rt]="$WID"
+  if [ -z "$WTOK" ]; then
+    TOKEN_ERRORS=$((TOKEN_ERRORS + 1))
+    TOKEN_ERROR_SUMMARY="${TOKEN_ERROR_SUMMARY}
+[$rt] workspace did not return or mint an auth_token — cannot drive its MCP call (workspace_id=$WID; create_resp: $(echo "$R" | redact_token_body); token_fallbacks: $TOKEN_DIAG)"
+    log "    $rt → $WID (token acquisition failed; continuing to classify other runtimes)"
+    continue
+  fi
  WS_TOKENS[$rt]="$WTOK"
  ALL_WS_IDS="$ALL_WS_IDS $WID"
  log "    $rt → $WID"
 done

+if [ "$TOKEN_ERRORS" -gt 0 ]; then
+  fail "token acquisition failed for $TOKEN_ERRORS runtime(s):$TOKEN_ERROR_SUMMARY"
+fi
+
+if [ "${PV_TOKEN_DIAGNOSTIC_ONLY:-0}" = "1" ]; then
+  ok "token diagnostic passed for runtimes: $PV_RUNTIMES"
+  exit 0
+fi
+
 # ─── 5. Wait for every sibling online ──────────────────────────────────
 log "5/6 waiting for all workspaces status=online (up to ${PROVISION_TIMEOUT_SECS}s — cold boot)..."
 WS_DEADLINE=$(( $(date +%s) + PROVISION_TIMEOUT_SECS ))
@@ -0,0 +1,22 @@
+#!/usr/bin/env bash
+# Staging E2E diagnostic — classify peer-visibility token acquisition.
+#
+# This is intentionally narrower than test_peer_visibility_mcp_staging.sh:
+# it provisions the same throwaway org, creates managed sibling workspaces,
+# and stops immediately after auth_token acquisition. The default runtime set
+# compares hermes with claude-code so a failure is easy to classify:
+#   - hermes fails, claude-code passes -> Hermes/runtime-specific
+#   - both fail                         -> shared admin/auth/proxy route
+#
+# Required env matches test_peer_visibility_mcp_staging.sh:
+#   MOLECULE_ADMIN_TOKEN
+# Optional:
+#   MOLECULE_CP_URL, E2E_RUN_ID, PV_RUNTIMES, E2E_KEEP_ORG,
+#   E2E_MINIMAX_API_KEY / E2E_ANTHROPIC_API_KEY / E2E_OPENAI_API_KEY
+
+set -euo pipefail
+
+export PV_RUNTIMES="${PV_RUNTIMES:-hermes claude-code}"
+export PV_TOKEN_DIAGNOSTIC_ONLY=1
+
+exec "$(dirname "${BASH_SOURCE[0]}")/test_peer_visibility_mcp_staging.sh"
@@ -25,6 +25,11 @@
 # Optional env:
 #   E2E_RUNTIME                  hermes (default) | claude-code | langgraph
 #   E2E_PROVISION_TIMEOUT_SECS   default 900 (15 min cold EC2 budget)
+#   E2E_WORKSPACE_ONLINE_TIMEOUT_SECS  default 3600 (60 min — hermes
+#                                cold-boot worst-case + slack). Raised from
+#                                1800 (#1646) because flaky tenant-provisioning
+#                                latency (not a code regression) causes
+#                                alternating pass/fail on identical SHAs.
 #   E2E_KEEP_ORG                 1 → skip teardown (debugging only)
 #   E2E_RUN_ID                   Slug suffix; CI: ${GITHUB_RUN_ID}
 #   E2E_MODE                     full (default) | smoke
@@ -32,6 +37,11 @@
 #                                 mapped to `smoke` for back-compat with
 #                                 any in-flight runner picking up an older
 #                                 workflow checkout)
+#   E2E_AWS_LEAK_CHECK           auto (default) | required | off
+#                                required in CI so teardown cannot report
+#                                clean while slug-tagged EC2 remains alive
+#   E2E_AWS_TERMINATE_LEAKS      1 → terminate slug-tagged leaked EC2 before
+#                                exiting 4
 #   E2E_INTENTIONAL_FAILURE      1 → poison tenant token mid-run so the
 #                                script fails; the EXIT trap MUST still
 #                                tear down cleanly (and exit 4 on leak).
@@ -51,6 +61,7 @@ CP_URL="${MOLECULE_CP_URL:-https://staging-api.moleculesai.app}"
 ADMIN_TOKEN="${MOLECULE_ADMIN_TOKEN:?MOLECULE_ADMIN_TOKEN required — Railway staging CP_ADMIN_API_TOKEN}"
 RUNTIME="${E2E_RUNTIME:-hermes}"
 PROVISION_TIMEOUT_SECS="${E2E_PROVISION_TIMEOUT_SECS:-900}"
+WORKSPACE_ONLINE_TIMEOUT_SECS="${E2E_WORKSPACE_ONLINE_TIMEOUT_SECS:-3600}"
 RUN_ID_SUFFIX="${E2E_RUN_ID:-$(date +%H%M%S)-$$}"
 MODE="${E2E_MODE:-full}"
 # `canary` is a legacy alias for `smoke` retained for back-compat with
@@ -82,8 +93,12 @@ ok()   { echo "[$(date +%H:%M:%S)] ✅ $*"; }
 # Per-runtime model slug dispatch — see lib/model_slug.sh for the rationale.
 # Extracted so unit tests (tests/e2e/test_model_slug.sh) can pin every branch
 # without booting the full 11-step lifecycle.
+# shellcheck disable=SC1091
 # shellcheck source=lib/model_slug.sh
 source "$(dirname "$0")/lib/model_slug.sh"
+# shellcheck disable=SC1091
+# shellcheck source=lib/aws_leak_check.sh
+source "$(dirname "$0")/lib/aws_leak_check.sh"

 CURL_COMMON=(-sS --fail-with-body --max-time 30)

@@ -119,12 +134,14 @@ cleanup_org() {
  #      DELETE returns 5xx mid-cascade and the cascade finishes anyway,
  #      and the case where DELETE legitimately exceeds 120s and we want
  #      eventual-consistency confirmation.
-  curl "${CURL_COMMON[@]}" --max-time 120 -X DELETE "$CP_URL/cp/admin/tenants/$SLUG" \
+  if curl "${CURL_COMMON[@]}" --max-time 120 -X DELETE "$CP_URL/cp/admin/tenants/$SLUG" \
    -H "Authorization: Bearer $ADMIN_TOKEN" \
    -H "Content-Type: application/json" \
-    -d "{\"confirm\":\"$SLUG\"}" >/dev/null 2>&1 \
-    && ok "Teardown request accepted" \
-    || log "Teardown returned non-2xx (may already be gone)"
+    -d "{\"confirm\":\"$SLUG\"}" >/dev/null 2>&1; then
+    ok "Teardown request accepted"
+  else
+    log "Teardown returned non-2xx (may already be gone)"
+  fi

  local leak_count=1
  local elapsed=0
@@ -144,7 +161,15 @@ cleanup_org() {
    echo "⚠️  LEAK: org $SLUG still present post-teardown after ${elapsed}s (count=$leak_count)" >&2
    exit 4
  fi
-  ok "Teardown clean — no orphan resources for $SLUG (${elapsed}s)"
+  local aws_leak_rc=0
+  e2e_verify_no_ec2_leaks_for_slug "$SLUG" || aws_leak_rc=$?
+  if [ "$aws_leak_rc" != "0" ]; then
+    case "$aws_leak_rc" in
+      2) exit 2 ;;
+      *) exit 4 ;;
+    esac
+  fi
+  ok "Teardown clean — no orphan org or EC2 resources for $SLUG (${elapsed}s)"

  # Normalize unexpected upstream exit codes to 1 (generic failure). The
  # script's documented contract (header "Exit codes" section) only emits
@@ -331,6 +356,75 @@ tenant_call() {
    "$@"
 }

+sanitize_http_body() {
+  python3 -c '
+import re, sys
+s = sys.stdin.read()
+s = re.sub(r"(?i)(Authorization:\s*Bearer\s+)[A-Za-z0-9._~+/=-]+", r"\1[redacted]", s)
+s = re.sub(r"(?i)(\"(?:auth_token|access_token|refresh_token|token|api_key|secret|password)\"\s*:\s*\")[^\"]+\"", r"\1[redacted]\"", s)
+s = re.sub(r"(?i)((?:auth_token|access_token|refresh_token|api_key|secret|password)=)[^&\s]+", r"\1[redacted]", s)
+print(s[:4000])
+'
+}
+
+wait_workspaces_online_routable() {
+  local label="$1"; shift
+  local deadline=$(( $(date +%s) + WORKSPACE_ONLINE_TIMEOUT_SECS ))
+  local wid ws_last_status ws_last_url ws_url_missing_logged ws_failed_logged
+  local ws_json ws_status ws_url ws_last_err
+
+  log "$label"
+  for wid in "$@"; do
+    ws_last_status=""
+    ws_last_url=""
+    ws_url_missing_logged=0
+    ws_failed_logged=0
+    while true; do
+      if [ "$(date +%s)" -gt "$deadline" ]; then
+        ws_last_err=$(tenant_call GET "/workspaces/$wid" 2>/dev/null | \
+          python3 -c "import json,sys; print(json.load(sys.stdin).get('last_sample_error',''))" 2>/dev/null || echo "")
+        fail "Workspace $wid never reached online with a routable URL within ${WORKSPACE_ONLINE_TIMEOUT_SECS}s (~$((WORKSPACE_ONLINE_TIMEOUT_SECS/60)) min) (last status=$ws_last_status, url=$ws_last_url, err=$ws_last_err)"
+      fi
+      ws_json=$(tenant_call GET "/workspaces/$wid" 2>/dev/null || echo '{}')
+      ws_status=$(echo "$ws_json" | python3 -c "import json,sys; print(json.load(sys.stdin).get('status') or '')" 2>/dev/null)
+      ws_url=$(echo "$ws_json" | python3 -c "import json,sys; print(json.load(sys.stdin).get('url') or '')" 2>/dev/null)
+      if [ "$ws_status" != "$ws_last_status" ]; then
+        log "    $wid → $ws_status"
+        ws_last_status="$ws_status"
+      fi
+      if [ -n "$ws_url" ] && [ "$ws_url" != "$ws_last_url" ]; then
+        log "    $wid url ready: $ws_url"
+        ws_last_url="$ws_url"
+      fi
+      case "$ws_status" in
+        online)
+          if [ -n "$ws_url" ]; then
+            break
+          fi
+          if [ "$ws_url_missing_logged" = "0" ]; then
+            log "    $wid online but URL is not assigned yet — waiting for workspace routing readiness"
+            ws_url_missing_logged=1
+          fi
+          sleep 10
+          ;;
+        failed)
+          # Not a hard fail — bootstrap-watcher frequently marks failed at
+          # 5 min on hermes, then heartbeat recovers to online around 10-13
+          # min when install.sh finishes. Log once per workspace so the CI
+          # output isn't spammy.
+          if [ "$ws_failed_logged" = "0" ]; then
+            log "    $wid transiently failed — waiting for heartbeat recovery (bootstrap-watcher deadline, see cp#245)"
+            ws_failed_logged=1
+          fi
+          sleep 10
+          ;;
+        *)      sleep 10 ;;
+      esac
+    done
+    ok "    $wid online and routable"
+  done
+}
+
 # ─── 5. Provision parent workspace ─────────────────────────────────────
 # Inject the LLM provider key so the runtime can authenticate at boot.
 # Branch by which secret is set so the script supports multiple paths
@@ -383,9 +477,9 @@ elif [ -n "${E2E_ANTHROPIC_API_KEY:-}" ]; then
  # is still independent of MOLECULE_STAGING_OPENAI_API_KEY, so an OpenAI
  # quota collapse doesn't wedge this path. Pinned to the claude-code
  # runtime: hermes/langgraph use OpenAI-shaped envs and won't honour
-  # ANTHROPIC_API_KEY without further wiring (out of scope for this
-  # branch; if you need a hermes/Anthropic path, dispatch with
-  # E2E_RUNTIME=hermes + E2E_OPENAI_API_KEY pointing at a working key).
+  # ANTHROPIC_API_KEY without further wiring. pick_model_slug maps this
+  # branch to claude-sonnet-4-6 so the claude-code provider registry
+  # selects anthropic-api instead of the OAuth-only sonnet alias.
  SECRETS_JSON=$(python3 -c "
 import json, os
 k = os.environ['E2E_ANTHROPIC_API_KEY']
@@ -410,6 +504,7 @@ print(json.dumps({
 fi

 MODEL_SLUG=$(pick_model_slug "$RUNTIME")
+log "    MODEL_SLUG=$MODEL_SLUG"

 log "5/11 Provisioning parent workspace (runtime=$RUNTIME)..."
 PARENT_RESP=$(tenant_call POST /workspaces \
@@ -437,48 +532,16 @@ fi
 # deadline fires at 5 min and sets status=failed prematurely; heartbeat
 # then transitions failed → online after install.sh finishes. So:
 #
-#   - 20 min deadline (hermes worst-case + slack)
+#   - ${WORKSPACE_ONLINE_TIMEOUT_SECS}s (~$((WORKSPACE_ONLINE_TIMEOUT_SECS/60)) min)
+#     deadline (hermes worst-case + slack). Configurable via
+#     E2E_WORKSPACE_ONLINE_TIMEOUT_SECS (#1646).
 #   - 'failed' is a TRANSIENT state we must tolerate — log and keep
 #     polling, only hard-fail at the deadline. Pre-bootstrap-watcher-fix
 #     (controlplane#245) this was a flake generator: workspace went
 #     failed→online inside our window but we bailed at the failed read.
-log "7/11 Waiting for workspace(s) to reach status=online (up to 30 min — hermes cold boot)..."
-WS_DEADLINE=$(( $(date +%s) + 1800 ))
-WS_TO_CHECK="$PARENT_ID"
-[ -n "$CHILD_ID" ] && WS_TO_CHECK="$WS_TO_CHECK $CHILD_ID"
-for wid in $WS_TO_CHECK; do
-  WS_LAST_STATUS=""
-  WS_FAILED_LOGGED=0
-  while true; do
-    if [ "$(date +%s)" -gt "$WS_DEADLINE" ]; then
-      WS_LAST_ERR=$(tenant_call GET "/workspaces/$wid" 2>/dev/null | \
-        python3 -c "import json,sys; print(json.load(sys.stdin).get('last_sample_error',''))" 2>/dev/null || echo "")
-      fail "Workspace $wid never reached online within 20 min (last status=$WS_LAST_STATUS, err=$WS_LAST_ERR)"
-    fi
-    WS_JSON=$(tenant_call GET "/workspaces/$wid" 2>/dev/null || echo '{}')
-    WS_STATUS=$(echo "$WS_JSON" | python3 -c "import json,sys; print(json.load(sys.stdin).get('status',''))" 2>/dev/null)
-    if [ "$WS_STATUS" != "$WS_LAST_STATUS" ]; then
-      log "    $wid → $WS_STATUS"
-      WS_LAST_STATUS="$WS_STATUS"
-    fi
-    case "$WS_STATUS" in
-      online) break ;;
-      failed)
-        # Not a hard fail — bootstrap-watcher frequently marks failed at
-        # 5 min on hermes, then heartbeat recovers to online around 10-13
-        # min when install.sh finishes. Log once per workspace so the CI
-        # output isn't spammy.
-        if [ "$WS_FAILED_LOGGED" = "0" ]; then
-          log "    $wid transiently failed — waiting for heartbeat recovery (bootstrap-watcher deadline, see cp#245)"
-          WS_FAILED_LOGGED=1
-        fi
-        sleep 10
-        ;;
-      *)      sleep 10 ;;
-    esac
-  done
-  ok "    $wid online"
-done
+WS_TO_CHECK=("$PARENT_ID")
+[ -n "$CHILD_ID" ] && WS_TO_CHECK+=("$CHILD_ID")
+wait_workspaces_online_routable "7/11 Waiting for workspace(s) to reach status=online (up to $((WORKSPACE_ONLINE_TIMEOUT_SECS/60)) min — hermes cold boot)..." "${WS_TO_CHECK[@]}"

 # ─── 7b. Canvas-terminal diagnose (EIC chain probe) ────────────────────
 # This step exists because the canvas-terminal failure of 2026-05-03
@@ -490,7 +553,7 @@ done
 #   - tenantIngressRules / workspaceIngressRules (CP)
 #   - eicSSHIngressRule helper (CP)
 #   - AuthorizeIngress source-group support (CP awsapi)
-#   - EIC_ENDPOINT_SG_ID Railway env
+#   - MOLECULE_EIC_ENDPOINT_SG_ID Railway env
 #   - handleRemoteConnect's send-ssh-public-key/open-tunnel/ssh chain
 # surfaces within ~20 min of merge instead of waiting for a user report.
 #
@@ -504,7 +567,7 @@ done
 # probes docker.Ping + container exec; we still expect ok=true there
 # since local-docker is the alternative production path.
 log "7b/11 Canvas-terminal EIC diagnose probe..."
-for wid in $WS_TO_CHECK; do
+for wid in "${WS_TO_CHECK[@]}"; do
  DIAG_JSON=$(tenant_call GET "/workspaces/$wid/terminal/diagnose" 2>/dev/null || echo '{}')
  DIAG_OK=$(echo "$DIAG_JSON" | python3 -c "import json,sys; d=json.load(sys.stdin); print('true' if d.get('ok') else 'false')" 2>/dev/null || echo "false")
  if [ "$DIAG_OK" = "true" ]; then
@@ -512,7 +575,7 @@ for wid in $WS_TO_CHECK; do
  else
    DIAG_FAIL=$(echo "$DIAG_JSON" | python3 -c "import json,sys; d=json.load(sys.stdin); print(d.get('first_failure','unknown'))" 2>/dev/null || echo "unknown")
    DIAG_DETAIL=$(echo "$DIAG_JSON" | python3 -c "import json,sys; d=json.load(sys.stdin); s=[x for x in d.get('steps',[]) if not x.get('ok')]; step=s[0] if s else {}; print(' — '.join(x for x in [step.get('error',''), step.get('detail','')] if x))" 2>/dev/null || echo "")
-    fail "Workspace $wid terminal diagnose failed at step '$DIAG_FAIL': $DIAG_DETAIL — check tenant SG has tcp/22 from EIC endpoint SG (sg-0785d5c6138220523), EIC_ENDPOINT_SG_ID set in Railway, and EIC endpoint health"
+    fail "Workspace $wid terminal diagnose failed at step '$DIAG_FAIL': $DIAG_DETAIL — check tenant SG has tcp/22 from the configured EIC endpoint SG, MOLECULE_EIC_ENDPOINT_SG_ID is set in Railway, and EIC endpoint health"
  fi
 done

@@ -540,7 +603,7 @@ CONFIG_PAYLOAD="${CONFIG_MARKER}
 name: synth-canary
 runtime: ${RUNTIME}
 "
-for wid in $WS_TO_CHECK; do
+for wid in "${WS_TO_CHECK[@]}"; do
  PUT_BODY=$(python3 -c "import json,sys; print(json.dumps({'content': sys.stdin.read()}))" <<< "$CONFIG_PAYLOAD")
  # Capture body to a tempfile so curl's -w '%{http_code}' is the only
  # thing on stdout. The first version used `-w '\n%{http_code}\n'` and
@@ -573,6 +636,12 @@ for wid in $WS_TO_CHECK; do
  ok "    $wid config.yaml PUT OK (HTTP $PUT_CODE)"
 done

+# Saving config.yaml follows the same path as Canvas Config Save & Restart.
+# The controlplane can briefly put the workspace back into provisioning and
+# clear its route while the runtime restarts, so A2A must wait on the same
+# externally routable readiness boundary again.
+wait_workspaces_online_routable "7d/11 Waiting for workspace(s) to recover routing after config.yaml PUT..." "${WS_TO_CHECK[@]}"
+
 # ─── 8. A2A round-trip on parent ───────────────────────────────────────
 log "8/11 Sending A2A message to parent — expecting agent response..."
 # Smoke prompt phrasing — DO NOT trim back to the bare "Reply with exactly: PONG"
@@ -612,10 +681,44 @@ print(json.dumps({
 # 90s gives ~3x headroom over observed cold-call P95 (~25-30s).
 # Subsequent A2A turns hit the same workspace and are sub-second, so
 # this only widens the window for step 8/11 of the canary's first turn.
-A2A_RESP=$(tenant_call POST "/workspaces/$PARENT_ID/a2a" \
-  --max-time 90 \
-  -H "Content-Type: application/json" \
-  -d "$A2A_PAYLOAD")
+A2A_TMP=$(mktemp -t synth_a2a.XXXXXX)
+for A2A_ATTEMPT in $(seq 1 12); do
+  : >"$A2A_TMP"
+  set +e
+  A2A_CODE=$(tenant_call POST "/workspaces/$PARENT_ID/a2a" \
+    --max-time 90 \
+    -H "Content-Type: application/json" \
+    -d "$A2A_PAYLOAD" \
+    -o "$A2A_TMP" \
+    -w '%{http_code}' \
+    2>/dev/null)
+  A2A_RC=$?
+  set -e
+  A2A_CODE=${A2A_CODE:-000}
+  A2A_RESP=$(cat "$A2A_TMP" 2>/dev/null || echo "")
+  if [ "$A2A_RC" = "0" ] && [ "$A2A_CODE" -ge 200 ] && [ "$A2A_CODE" -lt 300 ]; then
+    break
+  fi
+
+  A2A_SAFE_BODY=$(printf '%s' "$A2A_RESP" | sanitize_http_body)
+  if echo "$A2A_CODE" | grep -Eq '^(502|503|504)$' && echo "$A2A_SAFE_BODY" | grep -Eqi 'Service Unavailable|Bad Gateway|Gateway Timeout|error code: 502|error code: 504|workspace agent unreachable|connection refused|no healthy upstream|workspace agent busy|native_session'; then
+    log "    A2A cold-start probe attempt $A2A_ATTEMPT/12 returned $A2A_CODE: $A2A_SAFE_BODY"
+    if [ "$A2A_ATTEMPT" -lt 12 ]; then
+      A2A_SLEEP=10
+      if echo "$A2A_SAFE_BODY" | grep -Eqi 'workspace agent busy|native_session'; then
+        A2A_SLEEP=30
+      fi
+      sleep "$A2A_SLEEP"
+      continue
+    fi
+  fi
+  break
+done
+rm -f "$A2A_TMP"
+if [ "$A2A_RC" != "0" ] || [ "$A2A_CODE" -lt 200 ] || [ "$A2A_CODE" -ge 300 ]; then
+  A2A_SAFE_BODY=$(printf '%s' "$A2A_RESP" | sanitize_http_body)
+  fail "A2A POST /workspaces/$PARENT_ID/a2a failed after $A2A_ATTEMPT attempt(s) (curl_rc=$A2A_RC, http=$A2A_CODE): $A2A_SAFE_BODY"
+fi
 AGENT_TEXT=$(echo "$A2A_RESP" | python3 -c "
 import json, sys
 d = json.load(sys.stdin)
@@ -812,20 +915,50 @@ print(json.dumps({
    }
 }))
 ")
-  set +e
-  # Raw curl (not tenant_call) because this call carries an extra
-  # X-Source-Workspace-Id header. Must still send X-Molecule-Org-Id
-  # or TenantGuard 404s — previously missing, caused section 10 to
-  # fail rc=22 despite everything upstream being correct (2026-04-21).
-  DELEG_RESP=$(curl "${CURL_COMMON[@]}" -X POST "$TENANT_URL/workspaces/$CHILD_ID/a2a" \
-    -H "Authorization: Bearer $EFFECTIVE_TENANT_TOKEN" \
-    -H "X-Molecule-Org-Id: $ORG_ID" \
-    -H "X-Source-Workspace-Id: $PARENT_ID" \
-    -H "Content-Type: application/json" \
-    -d "$DELEG_PAYLOAD")
-  DELEG_RC=$?
-  set -e
-  [ $DELEG_RC -ne 0 ] && fail "Delegation A2A POST failed (rc=$DELEG_RC)"
+  DELEG_TMP=$(mktemp -t deleg_a2a.XXXXXX)
+  for DELEG_ATTEMPT in $(seq 1 12); do
+    : >"$DELEG_TMP"
+    set +e
+    # Raw curl (not tenant_call) because this call carries an extra
+    # X-Source-Workspace-Id header. Must still send X-Molecule-Org-Id
+    # or TenantGuard 404s — previously missing, caused section 10 to
+    # fail rc=22 despite everything upstream being correct (2026-04-21).
+    DELEG_CODE=$(curl "${CURL_COMMON[@]}" -X POST "$TENANT_URL/workspaces/$CHILD_ID/a2a" \
+      -H "Authorization: Bearer $EFFECTIVE_TENANT_TOKEN" \
+      -H "X-Molecule-Org-Id: $ORG_ID" \
+      -H "X-Source-Workspace-Id: $PARENT_ID" \
+      -H "Content-Type: application/json" \
+      -d "$DELEG_PAYLOAD" \
+      -o "$DELEG_TMP" \
+      -w '%{http_code}' \
+      2>/dev/null)
+    DELEG_RC=$?
+    set -e
+    DELEG_CODE=${DELEG_CODE:-000}
+    DELEG_RESP=$(cat "$DELEG_TMP" 2>/dev/null || echo "")
+    if [ "$DELEG_RC" = "0" ] && [ "$DELEG_CODE" -ge 200 ] && [ "$DELEG_CODE" -lt 300 ]; then
+      break
+    fi
+
+    DELEG_SAFE_BODY=$(printf '%s' "$DELEG_RESP" | sanitize_http_body)
+    if echo "$DELEG_CODE" | grep -Eq '^(502|503|504)$' && echo "$DELEG_SAFE_BODY" | grep -Eqi 'Service Unavailable|Bad Gateway|Gateway Timeout|error code: 502|error code: 504|workspace agent unreachable|connection refused|no healthy upstream|workspace agent busy|native_session'; then
+      log "    Delegation A2A cold-start attempt $DELEG_ATTEMPT/12 returned $DELEG_CODE: $DELEG_SAFE_BODY"
+      if [ "$DELEG_ATTEMPT" -lt 12 ]; then
+        DELEG_SLEEP=10
+        if echo "$DELEG_SAFE_BODY" | grep -Eqi 'workspace agent busy|native_session'; then
+          DELEG_SLEEP=30
+        fi
+        sleep "$DELEG_SLEEP"
+        continue
+      fi
+    fi
+    break
+  done
+  rm -f "$DELEG_TMP"
+  if [ "$DELEG_RC" != "0" ] || [ "$DELEG_CODE" -lt 200 ] || [ "$DELEG_CODE" -ge 300 ]; then
+    DELEG_SAFE_BODY=$(printf '%s' "$DELEG_RESP" | sanitize_http_body)
+    fail "Delegation A2A POST failed after $DELEG_ATTEMPT attempt(s) (curl_rc=$DELEG_RC, http=$DELEG_CODE): $DELEG_SAFE_BODY"
+  fi
  DELEG_TEXT=$(echo "$DELEG_RESP" | python3 -c "
 import json, sys
 try:
@@ -0,0 +1,18 @@
+from pathlib import Path
+
+
+ROOT = Path(__file__).resolve().parents[1]
+
+
+def test_staging_e2e_workflows_use_stable_minimax_default() -> None:
+    """Keep cron/push E2E on the same MiniMax model as the smoke-tested script."""
+    workflow_paths = [
+        ".gitea/workflows/e2e-staging-saas.yml",
+        ".gitea/workflows/staging-smoke.yml",
+        ".gitea/workflows/continuous-synth-e2e.yml",
+    ]
+
+    for rel in workflow_paths:
+        text = (ROOT / rel).read_text()
+        assert "MiniMax-M2.7-highspeed" not in text
+        assert "MiniMax-M2" in text
@@ -705,7 +705,7 @@ def test_ci_change_detector_docs_and_meta_scripts_do_not_trigger_surfaces():
    }


-def test_ci_platform_go_pr_steps_are_path_scoped():
+def test_ci_platform_go_steps_are_path_scoped_on_all_events():
    doc = yaml.safe_load(CI_WORKFLOW.read_text(encoding="utf-8"))
    platform = doc["jobs"]["platform-build"]
    assert platform.get("needs") == "changes"
@@ -720,11 +720,11 @@ def test_ci_platform_go_pr_steps_are_path_scoped():
    assert expensive_steps
    for step in expensive_steps:
        expr = step.get("if", "")
-        assert "github.event_name != 'pull_request'" in expr
        assert "needs.changes.outputs.platform == 'true'" in expr
+        assert "github.event_name != 'pull_request'" not in expr


-def test_ci_canvas_nextjs_pr_steps_are_path_scoped():
+def test_ci_canvas_nextjs_steps_are_path_scoped_on_all_events():
    doc = yaml.safe_load(CI_WORKFLOW.read_text(encoding="utf-8"))
    canvas = doc["jobs"]["canvas-build"]
    assert canvas.get("needs") == "changes"
@@ -739,11 +739,11 @@ def test_ci_canvas_nextjs_pr_steps_are_path_scoped():
    assert expensive_steps
    for step in expensive_steps:
        expr = step.get("if", "")
-        assert "github.event_name != 'pull_request'" in expr
        assert "needs.changes.outputs.canvas == 'true'" in expr
+        assert "github.event_name != 'pull_request'" not in expr


-def test_ci_shellcheck_pr_steps_are_path_scoped():
+def test_ci_shellcheck_steps_are_path_scoped_on_all_events():
    doc = yaml.safe_load(CI_WORKFLOW.read_text(encoding="utf-8"))
    shellcheck = doc["jobs"]["shellcheck"]
    assert shellcheck.get("needs") == "changes"
@@ -756,5 +756,5 @@ def test_ci_shellcheck_pr_steps_are_path_scoped():
    assert expensive_steps
    for step in expensive_steps:
        expr = step.get("if", "")
-        assert "github.event_name != 'pull_request'" in expr
        assert "needs.changes.outputs.scripts == 'true'" in expr
+        assert "github.event_name != 'pull_request'" not in expr
@@ -1,3 +1,26 @@
+// Package main runs the per-tenant workspace-server.
+//
+//	@title			Molecule AI Workspace Server API
+//	@version		1.0
+//	@description	The per-tenant workspace-server HTTP API. Single source of truth for workspace/schedule/agent/secrets/files/memory CRUD. Hand-written clients (canvas, molecule-mcp-server, molecule-cli, molecule-sdk-python) should be replaced by clients generated from this spec — see RFC #1706.
+//	@host			api.moleculesai.app
+//	@BasePath		/
+//	@schemes		https
+//
+//	@securityDefinitions.apikey	BearerAuth
+//	@in							header
+//	@name						Authorization
+//	@description				Bearer token issued by Gitea (org-admin or persona PAT) or by the platform's signup/SSO flow.
+//
+//	@securityDefinitions.apikey	OrgSlugAuth
+//	@in							header
+//	@name						X-Molecule-Org-Slug
+//	@description				Tenant routing header — required on every /workspaces/{id}/* request so the platform edge can route to the correct per-tenant workspace-server. Either X-Molecule-Org-Slug (human-readable, e.g. "agents-team") or X-Molecule-Org-Id (UUID) must be sent; slug is preferred for client code.
+//
+//	@securityDefinitions.apikey	OrgIdAuth
+//	@in							header
+//	@name						X-Molecule-Org-Id
+//	@description				Tenant routing header (UUID form). Alternative to X-Molecule-Org-Slug. At least one of OrgSlugAuth or OrgIdAuth must be sent alongside BearerAuth.
 package main

 import (
@@ -0,0 +1,521 @@
+{
+    "schemes": [
+        "https"
+    ],
+    "swagger": "2.0",
+    "info": {
+        "description": "The per-tenant workspace-server HTTP API. Single source of truth for workspace/schedule/agent/secrets/files/memory CRUD. Hand-written clients (canvas, molecule-mcp-server, molecule-cli, molecule-sdk-python) should be replaced by clients generated from this spec — see RFC #1706.",
+        "title": "Molecule AI Workspace Server API",
+        "contact": {},
+        "version": "1.0"
+    },
+    "host": "api.moleculesai.app",
+    "basePath": "/",
+    "paths": {
+        "/workspaces/{id}/schedules": {
+            "get": {
+                "security": [
+                    {
+                        "BearerAuth": [],
+                        "OrgSlugAuth": []
+                    }
+                ],
+                "produces": [
+                    "application/json"
+                ],
+                "tags": [
+                    "schedules"
+                ],
+                "summary": "List schedules for a workspace",
+                "parameters": [
+                    {
+                        "type": "string",
+                        "description": "Workspace ID",
+                        "name": "id",
+                        "in": "path",
+                        "required": true
+                    }
+                ],
+                "responses": {
+                    "200": {
+                        "description": "OK",
+                        "schema": {
+                            "type": "array",
+                            "items": {
+                                "$ref": "#/definitions/handlers.ScheduleResponse"
+                            }
+                        }
+                    },
+                    "500": {
+                        "description": "Internal Server Error",
+                        "schema": {
+                            "$ref": "#/definitions/handlers.ErrorResponse"
+                        }
+                    }
+                }
+            },
+            "post": {
+                "security": [
+                    {
+                        "BearerAuth": [],
+                        "OrgSlugAuth": []
+                    }
+                ],
+                "consumes": [
+                    "application/json"
+                ],
+                "produces": [
+                    "application/json"
+                ],
+                "tags": [
+                    "schedules"
+                ],
+                "summary": "Create a schedule",
+                "parameters": [
+                    {
+                        "type": "string",
+                        "description": "Workspace ID",
+                        "name": "id",
+                        "in": "path",
+                        "required": true
+                    },
+                    {
+                        "description": "Schedule fields",
+                        "name": "body",
+                        "in": "body",
+                        "required": true,
+                        "schema": {
+                            "$ref": "#/definitions/handlers.CreateScheduleRequest"
+                        }
+                    }
+                ],
+                "responses": {
+                    "201": {
+                        "description": "Created",
+                        "schema": {
+                            "$ref": "#/definitions/handlers.CreateScheduleResponse"
+                        }
+                    },
+                    "400": {
+                        "description": "Bad Request",
+                        "schema": {
+                            "$ref": "#/definitions/handlers.ErrorResponse"
+                        }
+                    },
+                    "500": {
+                        "description": "Internal Server Error",
+                        "schema": {
+                            "$ref": "#/definitions/handlers.ErrorResponse"
+                        }
+                    }
+                }
+            }
+        },
+        "/workspaces/{id}/schedules/{scheduleId}": {
+            "delete": {
+                "security": [
+                    {
+                        "BearerAuth": [],
+                        "OrgSlugAuth": []
+                    }
+                ],
+                "produces": [
+                    "application/json"
+                ],
+                "tags": [
+                    "schedules"
+                ],
+                "summary": "Delete a schedule",
+                "parameters": [
+                    {
+                        "type": "string",
+                        "description": "Workspace ID",
+                        "name": "id",
+                        "in": "path",
+                        "required": true
+                    },
+                    {
+                        "type": "string",
+                        "description": "Schedule ID",
+                        "name": "scheduleId",
+                        "in": "path",
+                        "required": true
+                    }
+                ],
+                "responses": {
+                    "200": {
+                        "description": "OK",
+                        "schema": {
+                            "$ref": "#/definitions/handlers.StatusResponse"
+                        }
+                    },
+                    "404": {
+                        "description": "Not Found",
+                        "schema": {
+                            "$ref": "#/definitions/handlers.ErrorResponse"
+                        }
+                    },
+                    "500": {
+                        "description": "Internal Server Error",
+                        "schema": {
+                            "$ref": "#/definitions/handlers.ErrorResponse"
+                        }
+                    }
+                }
+            },
+            "patch": {
+                "security": [
+                    {
+                        "BearerAuth": [],
+                        "OrgSlugAuth": []
+                    }
+                ],
+                "consumes": [
+                    "application/json"
+                ],
+                "produces": [
+                    "application/json"
+                ],
+                "tags": [
+                    "schedules"
+                ],
+                "summary": "Update a schedule",
+                "parameters": [
+                    {
+                        "type": "string",
+                        "description": "Workspace ID",
+                        "name": "id",
+                        "in": "path",
+                        "required": true
+                    },
+                    {
+                        "type": "string",
+                        "description": "Schedule ID",
+                        "name": "scheduleId",
+                        "in": "path",
+                        "required": true
+                    },
+                    {
+                        "description": "Partial schedule fields (only provided keys are updated)",
+                        "name": "body",
+                        "in": "body",
+                        "required": true,
+                        "schema": {
+                            "$ref": "#/definitions/handlers.UpdateScheduleRequest"
+                        }
+                    }
+                ],
+                "responses": {
+                    "200": {
+                        "description": "OK",
+                        "schema": {
+                            "$ref": "#/definitions/handlers.ScheduleResponse"
+                        }
+                    },
+                    "400": {
+                        "description": "Bad Request",
+                        "schema": {
+                            "$ref": "#/definitions/handlers.ErrorResponse"
+                        }
+                    },
+                    "404": {
+                        "description": "Not Found",
+                        "schema": {
+                            "$ref": "#/definitions/handlers.ErrorResponse"
+                        }
+                    },
+                    "500": {
+                        "description": "Internal Server Error",
+                        "schema": {
+                            "$ref": "#/definitions/handlers.ErrorResponse"
+                        }
+                    }
+                }
+            }
+        },
+        "/workspaces/{id}/schedules/{scheduleId}/history": {
+            "get": {
+                "security": [
+                    {
+                        "BearerAuth": [],
+                        "OrgSlugAuth": []
+                    }
+                ],
+                "produces": [
+                    "application/json"
+                ],
+                "tags": [
+                    "schedules"
+                ],
+                "summary": "Get past runs of a schedule",
+                "parameters": [
+                    {
+                        "type": "string",
+                        "description": "Workspace ID",
+                        "name": "id",
+                        "in": "path",
+                        "required": true
+                    },
+                    {
+                        "type": "string",
+                        "description": "Schedule ID",
+                        "name": "scheduleId",
+                        "in": "path",
+                        "required": true
+                    }
+                ],
+                "responses": {
+                    "200": {
+                        "description": "OK",
+                        "schema": {
+                            "type": "array",
+                            "items": {
+                                "$ref": "#/definitions/handlers.HistoryEntry"
+                            }
+                        }
+                    },
+                    "500": {
+                        "description": "Internal Server Error",
+                        "schema": {
+                            "$ref": "#/definitions/handlers.ErrorResponse"
+                        }
+                    }
+                }
+            }
+        },
+        "/workspaces/{id}/schedules/{scheduleId}/run": {
+            "post": {
+                "security": [
+                    {
+                        "BearerAuth": [],
+                        "OrgSlugAuth": []
+                    }
+                ],
+                "produces": [
+                    "application/json"
+                ],
+                "tags": [
+                    "schedules"
+                ],
+                "summary": "Fire a schedule manually",
+                "parameters": [
+                    {
+                        "type": "string",
+                        "description": "Workspace ID",
+                        "name": "id",
+                        "in": "path",
+                        "required": true
+                    },
+                    {
+                        "type": "string",
+                        "description": "Schedule ID",
+                        "name": "scheduleId",
+                        "in": "path",
+                        "required": true
+                    }
+                ],
+                "responses": {
+                    "200": {
+                        "description": "OK",
+                        "schema": {
+                            "$ref": "#/definitions/handlers.RunNowResponse"
+                        }
+                    },
+                    "404": {
+                        "description": "Not Found",
+                        "schema": {
+                            "$ref": "#/definitions/handlers.ErrorResponse"
+                        }
+                    },
+                    "500": {
+                        "description": "Internal Server Error",
+                        "schema": {
+                            "$ref": "#/definitions/handlers.ErrorResponse"
+                        }
+                    }
+                }
+            }
+        }
+    },
+    "definitions": {
+        "handlers.CreateScheduleRequest": {
+            "type": "object",
+            "required": [
+                "cron_expr",
+                "prompt"
+            ],
+            "properties": {
+                "cron_expr": {
+                    "type": "string"
+                },
+                "enabled": {
+                    "type": "boolean"
+                },
+                "name": {
+                    "type": "string"
+                },
+                "prompt": {
+                    "type": "string"
+                },
+                "timezone": {
+                    "type": "string"
+                }
+            }
+        },
+        "handlers.CreateScheduleResponse": {
+            "type": "object",
+            "properties": {
+                "id": {
+                    "type": "string"
+                },
+                "next_run_at": {
+                    "type": "string"
+                },
+                "status": {
+                    "type": "string"
+                }
+            }
+        },
+        "handlers.ErrorResponse": {
+            "type": "object",
+            "properties": {
+                "error": {
+                    "type": "string"
+                }
+            }
+        },
+        "handlers.HistoryEntry": {
+            "type": "object",
+            "properties": {
+                "duration_ms": {
+                    "type": "integer"
+                },
+                "error_detail": {
+                    "type": "string"
+                },
+                "request": {
+                    "type": "object"
+                },
+                "status": {
+                    "type": "string"
+                },
+                "timestamp": {
+                    "type": "string"
+                }
+            }
+        },
+        "handlers.RunNowResponse": {
+            "type": "object",
+            "properties": {
+                "prompt": {
+                    "type": "string"
+                },
+                "status": {
+                    "type": "string"
+                },
+                "workspace_id": {
+                    "type": "string"
+                }
+            }
+        },
+        "handlers.ScheduleResponse": {
+            "type": "object",
+            "properties": {
+                "created_at": {
+                    "type": "string"
+                },
+                "cron_expr": {
+                    "type": "string"
+                },
+                "enabled": {
+                    "type": "boolean"
+                },
+                "id": {
+                    "type": "string"
+                },
+                "last_error": {
+                    "type": "string"
+                },
+                "last_run_at": {
+                    "type": "string"
+                },
+                "last_status": {
+                    "type": "string"
+                },
+                "name": {
+                    "type": "string"
+                },
+                "next_run_at": {
+                    "type": "string"
+                },
+                "prompt": {
+                    "type": "string"
+                },
+                "run_count": {
+                    "type": "integer"
+                },
+                "source": {
+                    "description": "'template' (seeded by org/import) | 'runtime' (created via Canvas/API). Issue #24.",
+                    "type": "string"
+                },
+                "timezone": {
+                    "type": "string"
+                },
+                "updated_at": {
+                    "type": "string"
+                },
+                "workspace_id": {
+                    "type": "string"
+                }
+            }
+        },
+        "handlers.StatusResponse": {
+            "type": "object",
+            "properties": {
+                "status": {
+                    "type": "string"
+                }
+            }
+        },
+        "handlers.UpdateScheduleRequest": {
+            "type": "object",
+            "properties": {
+                "cron_expr": {
+                    "type": "string"
+                },
+                "enabled": {
+                    "type": "boolean"
+                },
+                "name": {
+                    "type": "string"
+                },
+                "prompt": {
+                    "type": "string"
+                },
+                "timezone": {
+                    "type": "string"
+                }
+            }
+        }
+    },
+    "securityDefinitions": {
+        "BearerAuth": {
+            "description": "Bearer token issued by Gitea (org-admin or persona PAT) or by the platform's signup/SSO flow.",
+            "type": "apiKey",
+            "name": "Authorization",
+            "in": "header"
+        },
+        "OrgIdAuth": {
+            "description": "Tenant routing header (UUID form). Alternative to X-Molecule-Org-Slug. At least one of OrgSlugAuth or OrgIdAuth must be sent alongside BearerAuth.",
+            "type": "apiKey",
+            "name": "X-Molecule-Org-Id",
+            "in": "header"
+        },
+        "OrgSlugAuth": {
+            "description": "Tenant routing header — required on every /workspaces/{id}/* request so the platform edge can route to the correct per-tenant workspace-server. Either X-Molecule-Org-Slug (human-readable, e.g. \"agents-team\") or X-Molecule-Org-Id (UUID) must be sent; slug is preferred for client code.",
+            "type": "apiKey",
+            "name": "X-Molecule-Org-Slug",
+            "in": "header"
+        }
+    }
+}
@@ -0,0 +1,349 @@
+basePath: /
+definitions:
+  handlers.CreateScheduleRequest:
+    properties:
+      cron_expr:
+        type: string
+      enabled:
+        type: boolean
+      name:
+        type: string
+      prompt:
+        type: string
+      timezone:
+        type: string
+    required:
+    - cron_expr
+    - prompt
+    type: object
+  handlers.CreateScheduleResponse:
+    properties:
+      id:
+        type: string
+      next_run_at:
+        type: string
+      status:
+        type: string
+    type: object
+  handlers.ErrorResponse:
+    properties:
+      error:
+        type: string
+    type: object
+  handlers.HistoryEntry:
+    properties:
+      duration_ms:
+        type: integer
+      error_detail:
+        type: string
+      request:
+        type: object
+      status:
+        type: string
+      timestamp:
+        type: string
+    type: object
+  handlers.RunNowResponse:
+    properties:
+      prompt:
+        type: string
+      status:
+        type: string
+      workspace_id:
+        type: string
+    type: object
+  handlers.ScheduleResponse:
+    properties:
+      created_at:
+        type: string
+      cron_expr:
+        type: string
+      enabled:
+        type: boolean
+      id:
+        type: string
+      last_error:
+        type: string
+      last_run_at:
+        type: string
+      last_status:
+        type: string
+      name:
+        type: string
+      next_run_at:
+        type: string
+      prompt:
+        type: string
+      run_count:
+        type: integer
+      source:
+        description: '''template'' (seeded by org/import) | ''runtime'' (created via
+          Canvas/API). Issue #24.'
+        type: string
+      timezone:
+        type: string
+      updated_at:
+        type: string
+      workspace_id:
+        type: string
+    type: object
+  handlers.StatusResponse:
+    properties:
+      status:
+        type: string
+    type: object
+  handlers.UpdateScheduleRequest:
+    properties:
+      cron_expr:
+        type: string
+      enabled:
+        type: boolean
+      name:
+        type: string
+      prompt:
+        type: string
+      timezone:
+        type: string
+    type: object
+host: api.moleculesai.app
+info:
+  contact: {}
+  description: 'The per-tenant workspace-server HTTP API. Single source of truth for
+    workspace/schedule/agent/secrets/files/memory CRUD. Hand-written clients (canvas,
+    molecule-mcp-server, molecule-cli, molecule-sdk-python) should be replaced by
+    clients generated from this spec — see RFC #1706.'
+  title: Molecule AI Workspace Server API
+  version: "1.0"
+paths:
+  /workspaces/{id}/schedules:
+    get:
+      parameters:
+      - description: Workspace ID
+        in: path
+        name: id
+        required: true
+        type: string
+      produces:
+      - application/json
+      responses:
+        "200":
+          description: OK
+          schema:
+            items:
+              $ref: '#/definitions/handlers.ScheduleResponse'
+            type: array
+        "500":
+          description: Internal Server Error
+          schema:
+            $ref: '#/definitions/handlers.ErrorResponse'
+      security:
+      - BearerAuth: []
+        OrgSlugAuth: []
+      summary: List schedules for a workspace
+      tags:
+      - schedules
+    post:
+      consumes:
+      - application/json
+      parameters:
+      - description: Workspace ID
+        in: path
+        name: id
+        required: true
+        type: string
+      - description: Schedule fields
+        in: body
+        name: body
+        required: true
+        schema:
+          $ref: '#/definitions/handlers.CreateScheduleRequest'
+      produces:
+      - application/json
+      responses:
+        "201":
+          description: Created
+          schema:
+            $ref: '#/definitions/handlers.CreateScheduleResponse'
+        "400":
+          description: Bad Request
+          schema:
+            $ref: '#/definitions/handlers.ErrorResponse'
+        "500":
+          description: Internal Server Error
+          schema:
+            $ref: '#/definitions/handlers.ErrorResponse'
+      security:
+      - BearerAuth: []
+        OrgSlugAuth: []
+      summary: Create a schedule
+      tags:
+      - schedules
+  /workspaces/{id}/schedules/{scheduleId}:
+    delete:
+      parameters:
+      - description: Workspace ID
+        in: path
+        name: id
+        required: true
+        type: string
+      - description: Schedule ID
+        in: path
+        name: scheduleId
+        required: true
+        type: string
+      produces:
+      - application/json
+      responses:
+        "200":
+          description: OK
+          schema:
+            $ref: '#/definitions/handlers.StatusResponse'
+        "404":
+          description: Not Found
+          schema:
+            $ref: '#/definitions/handlers.ErrorResponse'
+        "500":
+          description: Internal Server Error
+          schema:
+            $ref: '#/definitions/handlers.ErrorResponse'
+      security:
+      - BearerAuth: []
+        OrgSlugAuth: []
+      summary: Delete a schedule
+      tags:
+      - schedules
+    patch:
+      consumes:
+      - application/json
+      parameters:
+      - description: Workspace ID
+        in: path
+        name: id
+        required: true
+        type: string
+      - description: Schedule ID
+        in: path
+        name: scheduleId
+        required: true
+        type: string
+      - description: Partial schedule fields (only provided keys are updated)
+        in: body
+        name: body
+        required: true
+        schema:
+          $ref: '#/definitions/handlers.UpdateScheduleRequest'
+      produces:
+      - application/json
+      responses:
+        "200":
+          description: OK
+          schema:
+            $ref: '#/definitions/handlers.ScheduleResponse'
+        "400":
+          description: Bad Request
+          schema:
+            $ref: '#/definitions/handlers.ErrorResponse'
+        "404":
+          description: Not Found
+          schema:
+            $ref: '#/definitions/handlers.ErrorResponse'
+        "500":
+          description: Internal Server Error
+          schema:
+            $ref: '#/definitions/handlers.ErrorResponse'
+      security:
+      - BearerAuth: []
+        OrgSlugAuth: []
+      summary: Update a schedule
+      tags:
+      - schedules
+  /workspaces/{id}/schedules/{scheduleId}/history:
+    get:
+      parameters:
+      - description: Workspace ID
+        in: path
+        name: id
+        required: true
+        type: string
+      - description: Schedule ID
+        in: path
+        name: scheduleId
+        required: true
+        type: string
+      produces:
+      - application/json
+      responses:
+        "200":
+          description: OK
+          schema:
+            items:
+              $ref: '#/definitions/handlers.HistoryEntry'
+            type: array
+        "500":
+          description: Internal Server Error
+          schema:
+            $ref: '#/definitions/handlers.ErrorResponse'
+      security:
+      - BearerAuth: []
+        OrgSlugAuth: []
+      summary: Get past runs of a schedule
+      tags:
+      - schedules
+  /workspaces/{id}/schedules/{scheduleId}/run:
+    post:
+      parameters:
+      - description: Workspace ID
+        in: path
+        name: id
+        required: true
+        type: string
+      - description: Schedule ID
+        in: path
+        name: scheduleId
+        required: true
+        type: string
+      produces:
+      - application/json
+      responses:
+        "200":
+          description: OK
+          schema:
+            $ref: '#/definitions/handlers.RunNowResponse'
+        "404":
+          description: Not Found
+          schema:
+            $ref: '#/definitions/handlers.ErrorResponse'
+        "500":
+          description: Internal Server Error
+          schema:
+            $ref: '#/definitions/handlers.ErrorResponse'
+      security:
+      - BearerAuth: []
+        OrgSlugAuth: []
+      summary: Fire a schedule manually
+      tags:
+      - schedules
+schemes:
+- https
+securityDefinitions:
+  BearerAuth:
+    description: Bearer token issued by Gitea (org-admin or persona PAT) or by the
+      platform's signup/SSO flow.
+    in: header
+    name: Authorization
+    type: apiKey
+  OrgIdAuth:
+    description: Tenant routing header (UUID form). Alternative to X-Molecule-Org-Slug.
+      At least one of OrgSlugAuth or OrgIdAuth must be sent alongside BearerAuth.
+    in: header
+    name: X-Molecule-Org-Id
+    type: apiKey
+  OrgSlugAuth:
+    description: Tenant routing header — required on every /workspaces/{id}/* request
+      so the platform edge can route to the correct per-tenant workspace-server. Either
+      X-Molecule-Org-Slug (human-readable, e.g. "agents-team") or X-Molecule-Org-Id
+      (UUID) must be sent; slug is preferred for client code.
+    in: header
+    name: X-Molecule-Org-Slug
+    type: apiKey
+swagger: "2.0"
@@ -71,35 +71,30 @@ func (h *WorkspaceHandler) handleA2ADispatchError(ctx context.Context, workspace
 	// with 202 status here was the original cycle 53 bug — callers saw
 	// proxyErr != nil and logged "delegation failed: proxy a2a error".
 	if isUpstreamBusyError(err) {
-		// Capability primitive #5 — see project memory
-		// `project_runtime_native_pluggable.md`. When the target workspace's
-		// adapter has declared provides_native_session=True, the SDK
-		// owns its own queue/session state (claude-agent-sdk's streaming
-		// session, hermes-agent's in-container event log, etc.). Adding
-		// the platform's a2a_queue layer on top would double-buffer the
-		// same in-flight state — and worse, the platform queue's drain
-		// timing has no relationship to the SDK's actual readiness, so
-		// the queued request might dispatch while the SDK is STILL busy.
+		// #1684 / Reno Stars: native_session adapters previously took a
+		// 503-no-enqueue path here, on the assumption that the SDK owned
+		// an inbound queue and the platform a2a_queue would double-buffer.
+		// In practice, the common native_session SDKs (claude-agent-sdk,
+		// codex app-server, hermes-agent) do NOT have an inbound queue —
+		// new turns can only be pushed via the same HTTP POST that just
+		// returned busy. So cron fires (and any A2A retry) bounce 503
+		// every tick until the SDK voluntarily yields. Reno Stars #1684
+		// observed 12 consecutive `*/30` cron fires lost over 6h while a
+		// single native_session held the slot.
 		//
-		// For native_session targets, return 503 + Retry-After directly.
-		// The caller's adapter handles retry on its own schedule, and
-		// the SDK's own queue absorbs the in-flight request when it does.
-		// Observability is preserved: logA2AFailure already ran above;
-		// activity_logs records the busy event; the broadcaster fires.
-		if runtimeOverrides.HasCapability(workspaceID, "session") {
-			log.Printf("ProxyA2A: target %s busy and declares native_session — skip enqueue, return 503", workspaceID)
-			return 0, nil, &proxyA2AError{
-				Status:  http.StatusServiceUnavailable,
-				Headers: map[string]string{"Retry-After": strconv.Itoa(busyRetryAfterSeconds)},
-				Response: gin.H{
-					"error":          "workspace agent busy — adapter handles retry (native_session)",
-					"busy":           true,
-					"retry_after":    busyRetryAfterSeconds,
-					"native_session": true,
-				},
-			}
-		}
-
+		// The original concern — "drain timing has no relationship to SDK
+		// readiness" — turns out to be unfounded: heartbeat→drain is
+		// gated by `payload.ActiveTasks < maxConcurrent` in
+		// registry.go:Heartbeat, so drain only fires when the workspace
+		// itself reports spare capacity. That IS the session-ended
+		// signal. The native_session SDK reports ActiveTasks=1 while in a
+		// turn, ActiveTasks=0 when idle, and the next heartbeat after
+		// idle triggers DrainQueueForWorkspace.
+		//
+		// So we collapse the two branches: both native_session and
+		// non-native callers enqueue here. The native_session SDK's own
+		// in-flight POST stays unaffected; the queued item drains on the
+		// next post-idle heartbeat.
 		idempotencyKey := extractIdempotencyKey(body)
 		// Honor params.expires_in_seconds when the caller specifies one. Zero
 		// (the unset default) → expiresAt = nil → infinite TTL preserved by
@@ -67,7 +67,213 @@ func NewActivityHandler(b *events.Broadcaster) *ActivityHandler {
 	return &ActivityHandler{broadcaster: b}
 }

-// List handles GET /workspaces/:id/activity?type=&source=&limit=&since_secs=&since_id=
+// extractAttachmentsFromRequestBody walks a JSON-RPC a2a inbound body to
+// surface attachments (file/image/audio/video) as a flat `attachments[]`
+// projection so callers don't have to drill into the request_body shape
+// themselves.
+//
+// Two body shapes are walked in order:
+//
+//  1. a2a-sdk v1 message-part envelope (peer_agent inbound):
+//
+//     {"jsonrpc":"2.0","method":"message/send","params":{
+//        "message":{"parts":[
+//          {"kind":"text", "text":"hi"},
+//          {"kind":"file", "file":{"uri":"workspace:foo.pdf","mime_type":"application/pdf","name":"foo.pdf"}},
+//          {"kind":"image","file":{"uri":"workspace:bar.png","mime_type":"image/png","name":"bar.png"}},
+//        ]}}}
+//
+//  2. canvas chat_upload_receive flat manifest (canvas_user upload):
+//
+//     {"uri":"platform-pending:<ws>/<file>",
+//      "name":"pasted.png",
+//      "size":12345,
+//      "file_id":"<uuid>",
+//      "mimeType":"image/png"}
+//
+//     The canvas upload pipe writes a single manifest directly at the
+//     root of request_body (no JSON-RPC envelope) with camelCase
+//     `mimeType`. We normalize to snake_case `mime_type` on the way out
+//     so every downstream adaptor (channel / telegram / codex / hermes)
+//     sees one wire shape regardless of which inbound shape produced it.
+//
+// Returns nil (omit-from-JSON) when the body has no attachments — the
+// `?include=peer_info` envelope projects this as an array iff non-empty.
+//
+// Defensive on every step: any missing key / wrong-shape value falls
+// through to the next arm or returns nil instead of panicking. The
+// activity_logs row could carry literally any JSON in request_body
+// (legacy formats, future formats); we only commit to the documented
+// shapes and silently skip anything else.
+func extractAttachmentsFromRequestBody(raw []byte) []map[string]interface{} {
+	if len(raw) == 0 {
+		return nil
+	}
+	var body map[string]interface{}
+	if err := json.Unmarshal(raw, &body); err != nil {
+		return nil
+	}
+	if atts := extractAttachmentsFromMessageParts(body); len(atts) > 0 {
+		return atts
+	}
+	if att := extractAttachmentFromFlatUploadManifest(body); att != nil {
+		return []map[string]interface{}{att}
+	}
+	return nil
+}
+
+// extractAttachmentsFromMessageParts handles the a2a-sdk v1 shape:
+// body.params.message.parts[]. Walks file/image/audio parts; honors v1
+// `kind` and v0 `type` discriminators; accepts nested `.file` sub-object
+// or inlined uri/mime_type/name on the part itself.
+func extractAttachmentsFromMessageParts(body map[string]interface{}) []map[string]interface{} {
+	params, ok := body["params"].(map[string]interface{})
+	if !ok {
+		return nil
+	}
+	message, ok := params["message"].(map[string]interface{})
+	if !ok {
+		return nil
+	}
+	parts, ok := message["parts"].([]interface{})
+	if !ok {
+		return nil
+	}
+	out := make([]map[string]interface{}, 0)
+	for _, p := range parts {
+		part, ok := p.(map[string]interface{})
+		if !ok {
+			continue
+		}
+		// a2a-sdk v1 uses "kind"; older v0 callers sent "type". Accept
+		// both for the discriminator — same defensive read pattern as
+		// the runtime-side extract_text helper.
+		kind, _ := part["kind"].(string)
+		if kind == "" {
+			kind, _ = part["type"].(string)
+		}
+		if kind != "file" && kind != "image" && kind != "audio" {
+			continue
+		}
+		// The file sub-object holds uri/mime_type/name. The a2a-sdk v1
+		// shape nests under "file"; some legacy payloads inlined the
+		// fields onto the part itself. Support both.
+		var fileObj map[string]interface{}
+		if f, ok := part["file"].(map[string]interface{}); ok {
+			fileObj = f
+		} else {
+			fileObj = part
+		}
+		uri, _ := fileObj["uri"].(string)
+		mimeType, _ := fileObj["mime_type"].(string)
+		name, _ := fileObj["name"].(string)
+		// At minimum we need either a uri or a name to be useful.
+		// Empty-part entries are skipped (they're a malformed inbound
+		// — surface nothing rather than emit a no-info placeholder).
+		if uri == "" && name == "" {
+			continue
+		}
+		att := map[string]interface{}{"kind": kind}
+		if uri != "" {
+			att["uri"] = uri
+		}
+		if mimeType != "" {
+			att["mime_type"] = mimeType
+		}
+		if name != "" {
+			att["name"] = name
+		}
+		out = append(out, att)
+	}
+	if len(out) == 0 {
+		return nil
+	}
+	return out
+}
+
+// extractAttachmentFromFlatUploadManifest handles the canvas
+// chat_upload_receive shape: a single upload manifest at the root of
+// request_body with no JSON-RPC envelope. Canvas uses camelCase
+// `mimeType`; we normalize to snake_case `mime_type` on emit so the
+// wire shape matches the message-parts arm. Kind is derived from the
+// mime prefix (image/* → "image", audio/* → "audio", video/* → "video",
+// anything else → "file") because the canvas upload row doesn't carry
+// an explicit discriminator. Returns nil if neither `uri` nor `file_id`
+// is present at the root (i.e. not a flat upload manifest).
+func extractAttachmentFromFlatUploadManifest(body map[string]interface{}) map[string]interface{} {
+	uri, _ := body["uri"].(string)
+	fileID, _ := body["file_id"].(string)
+	if uri == "" && fileID == "" {
+		return nil
+	}
+	mimeType, _ := body["mimeType"].(string)
+	if mimeType == "" {
+		// Defensive: future canvas versions might emit snake_case directly.
+		mimeType, _ = body["mime_type"].(string)
+	}
+	name, _ := body["name"].(string)
+	// Apply the same minimum-info rule as the message-parts arm: a
+	// manifest with neither uri nor name is non-actionable; skip.
+	if uri == "" && name == "" {
+		return nil
+	}
+	att := map[string]interface{}{"kind": kindFromMimeType(mimeType)}
+	if uri != "" {
+		att["uri"] = uri
+	}
+	if mimeType != "" {
+		att["mime_type"] = mimeType
+	}
+	if name != "" {
+		att["name"] = name
+	}
+	return att
+}
+
+// kindFromMimeType derives the attachment `kind` discriminator from a
+// MIME type. Used by the flat-upload-manifest arm where the source row
+// has no explicit kind field.
+func kindFromMimeType(mime string) string {
+	switch {
+	case strings.HasPrefix(mime, "image/"):
+		return "image"
+	case strings.HasPrefix(mime, "audio/"):
+		return "audio"
+	case strings.HasPrefix(mime, "video/"):
+		return "video"
+	default:
+		return "file"
+	}
+}
+
+// includeFlagSet returns true iff `flag` appears in the comma-separated
+// `?include=` query value. Whitespace around entries is tolerated.
+// Empty `include` returns false (existing back-compat shape).
+//
+// The comma-separable form lets future fields ("attachments_only",
+// "tool_trace_expanded", etc.) slot in without further URL-param creep.
+func includeFlagSet(includeQuery, flag string) bool {
+	if includeQuery == "" || flag == "" {
+		return false
+	}
+	for _, raw := range strings.Split(includeQuery, ",") {
+		if strings.TrimSpace(raw) == flag {
+			return true
+		}
+	}
+	return false
+}
+
+// List handles GET /workspaces/:id/activity?type=&source=&limit=&since_secs=&since_id=&include=
+//
+// The `include` query param is comma-separable; today the only flag is
+// `peer_info`, which enriches a2a_receive rows with `peer_name`,
+// `peer_role`, `agent_card_url`, and an `attachments[]` projection (see
+// extractAttachmentsFromRequestBody). It's additive + opt-in — existing
+// callers that don't pass `?include=peer_info` see the unchanged shape.
+// Surface for the layered enrichment that lets Claude Code channel
+// pushes carry full sender identity instead of bare UUIDs (sibling
+// repos: molecule-ai-workspace-runtime + molecule-mcp-claude-channel).
 //
 // since_secs filters to activity_logs.created_at >= NOW() - INTERVAL '$N seconds'.
 // Optional, additive — callers that don't pass it get today's behavior (the
@@ -102,6 +308,8 @@ func (h *ActivityHandler) List(c *gin.Context) {
 	sinceSecsStr := c.Query("since_secs")
 	sinceID := c.Query("since_id")
 	beforeTSStr := c.Query("before_ts") // optional RFC3339 — return rows strictly older than this timestamp
+	include := c.Query("include")       // comma-separated; today's only flag is "peer_info"
+	includePeerInfo := includeFlagSet(include, "peer_info")

 	// Validate peer_id as a UUID at the trust boundary so a malformed
 	// caller (the agent or a downstream MCP tool) can't smuggle SQL
@@ -192,22 +400,60 @@ func (h *ActivityHandler) List(c *gin.Context) {
 		usingCursor = true
 	}

-	// Build query with optional filters
-	query := `SELECT id, workspace_id, activity_type, source_id, target_id, method,
-			   summary, request_body, response_body, tool_trace, duration_ms, status, error_detail, created_at
-		FROM activity_logs WHERE workspace_id = $1`
+	// Build query with optional filters. When ?include=peer_info is set,
+	// LEFT JOIN workspaces ON activity_logs.source_id = w.id so we can
+	// surface w.name + w.role on the row. LEFT (not INNER) is required
+	// for two reasons:
+	//   1. Canvas rows have source_id IS NULL — those must still appear
+	//      in the result set (with NULL peer_name/peer_role).
+	//   2. A peer workspace may have been deleted since the row was
+	//      written (no FK constraint on activity_logs.source_id) —
+	//      LEFT JOIN preserves the activity row with NULL peer fields
+	//      rather than silently dropping the row.
+	//
+	// agent_card_url is NOT pulled from the workspaces table; it's
+	// computed server-side from externalPlatformURL + source_id at
+	// projection time (mirrors molecule-ai-workspace-runtime
+	// a2a_client._agent_card_url_for which constructs
+	// {PLATFORM_URL}/registry/discover/{peer_id}).
+	//
+	// Column qualification (`activity_logs.<col>`) is added ONLY when
+	// the JOIN is present — disambiguates `id` / `created_at` which
+	// exist in both tables. When the JOIN is absent, unqualified
+	// column references preserve the exact wire-shape existing callers
+	// + existing test fixtures expect (back-compat).
+	actCol := ""
+	if includePeerInfo {
+		actCol = "activity_logs."
+	}
+	selectClause := `SELECT ` + actCol + `id, ` + actCol + `workspace_id, ` + actCol + `activity_type, ` +
+		actCol + `source_id, ` + actCol + `target_id, ` + actCol + `method, ` +
+		actCol + `summary, ` + actCol + `request_body, ` + actCol + `response_body, ` +
+		actCol + `tool_trace, ` + actCol + `duration_ms, ` + actCol + `status, ` +
+		actCol + `error_detail, ` + actCol + `created_at`
+	fromClause := ` FROM activity_logs`
+	if includePeerInfo {
+		selectClause += `, w.name AS peer_name, w.role AS peer_role`
+		fromClause += ` LEFT JOIN workspaces w ON w.id = activity_logs.source_id`
+	}
+	query := selectClause + fromClause + ` WHERE ` + actCol + `workspace_id = $1`
 	args := []interface{}{workspaceID}
 	argIdx := 2

+	// WHERE/ORDER column refs use the same `actCol` qualifier prefix
+	// computed above — empty string when no JOIN (back-compat with
+	// existing wire shape + sqlmock-regex test fixtures), or
+	// `activity_logs.` when LEFT JOIN'd (disambiguates `id` /
+	// `created_at` between the two tables).
 	if activityType != "" {
-		query += fmt.Sprintf(" AND activity_type = $%d", argIdx)
+		query += fmt.Sprintf(" AND "+actCol+"activity_type = $%d", argIdx)
 		args = append(args, activityType)
 		argIdx++
 	}
 	if source == "canvas" {
-		query += " AND source_id IS NULL"
+		query += " AND " + actCol + "source_id IS NULL"
 	} else if source == "agent" {
-		query += " AND source_id IS NOT NULL"
+		query += " AND " + actCol + "source_id IS NOT NULL"
 	} else if source != "" {
 		c.JSON(http.StatusBadRequest, gin.H{"error": "source must be 'canvas' or 'agent'"})
 		return
@@ -224,7 +470,7 @@ func (h *ActivityHandler) List(c *gin.Context) {
 		// and avoids duplicate parameter binding (some drivers reject the
 		// same arg slot reused, ours is fine but the explicit form is
 		// clearer to read and matches the rest of the builder.)
-		query += fmt.Sprintf(" AND (source_id = $%d OR target_id = $%d)", argIdx, argIdx)
+		query += fmt.Sprintf(" AND ("+actCol+"source_id = $%d OR "+actCol+"target_id = $%d)", argIdx, argIdx)
 		args = append(args, peerID)
 		argIdx++
 	}
@@ -232,7 +478,7 @@ func (h *ActivityHandler) List(c *gin.Context) {
 		// Strictly older — never replay a row with the exact same
 		// timestamp, mirrors the `created_at > cursorTime` shape
 		// `since_id` uses for forward paging.
-		query += fmt.Sprintf(" AND created_at < $%d", argIdx)
+		query += fmt.Sprintf(" AND "+actCol+"created_at < $%d", argIdx)
 		args = append(args, beforeTS)
 		argIdx++
 	}
@@ -241,13 +487,13 @@ func (h *ActivityHandler) List(c *gin.Context) {
 		// interpolated into the SQL string. `make_interval(secs => $N)`
 		// avoids the lib/pq quirk where INTERVAL '$N seconds' won't
 		// substitute a placeholder inside the literal.
-		query += fmt.Sprintf(" AND created_at >= NOW() - make_interval(secs => $%d)", argIdx)
+		query += fmt.Sprintf(" AND "+actCol+"created_at >= NOW() - make_interval(secs => $%d)", argIdx)
 		args = append(args, sinceSecs)
 		argIdx++
 	}
 	if usingCursor {
 		// Strictly after — never replay the cursor row itself.
-		query += fmt.Sprintf(" AND created_at > $%d", argIdx)
+		query += fmt.Sprintf(" AND "+actCol+"created_at > $%d", argIdx)
 		args = append(args, cursorTime)
 		argIdx++
 	}
@@ -257,9 +503,9 @@ func (h *ActivityHandler) List(c *gin.Context) {
 	// since_id) keeps DESC — that's the canvas/UI shape and changing it
 	// would surprise existing callers.
 	if usingCursor {
-		query += fmt.Sprintf(" ORDER BY created_at ASC LIMIT $%d", argIdx)
+		query += fmt.Sprintf(" ORDER BY "+actCol+"created_at ASC LIMIT $%d", argIdx)
 	} else {
-		query += fmt.Sprintf(" ORDER BY created_at DESC LIMIT $%d", argIdx)
+		query += fmt.Sprintf(" ORDER BY "+actCol+"created_at DESC LIMIT $%d", argIdx)
 	}
 	args = append(args, limit)

@@ -272,6 +518,14 @@ func (h *ActivityHandler) List(c *gin.Context) {
 	}
 	defer rows.Close()

+	// agent_card_url base computed once per request so we don't pay the
+	// header-read cost per row. Only meaningful when includePeerInfo is
+	// set; the empty string here is harmless when the flag is off.
+	var platformBase string
+	if includePeerInfo {
+		platformBase = externalPlatformURL(c)
+	}
+
 	activities := make([]map[string]interface{}, 0)
 	for rows.Next() {
 		var id, wsID, actType, status string
@@ -279,10 +533,23 @@ func (h *ActivityHandler) List(c *gin.Context) {
 		var reqBody, respBody, toolTrace []byte
 		var durationMs *int
 		var createdAt time.Time
+		// LEFT JOIN'd peer columns — pointer-string so a NULL row
+		// (canvas message OR deleted peer workspace) decodes as nil
+		// rather than empty-string. Only scanned when includePeerInfo
+		// is set (matched against the SELECT clause above).
+		var peerName, peerRole *string

-		if err := rows.Scan(&id, &wsID, &actType, &sourceID, &targetID, &method,
-			&summary, &reqBody, &respBody, &toolTrace, &durationMs, &status, &errorDetail, &createdAt); err != nil {
-			log.Printf("Activity scan error: %v", err)
+		var scanErr error
+		if includePeerInfo {
+			scanErr = rows.Scan(&id, &wsID, &actType, &sourceID, &targetID, &method,
+				&summary, &reqBody, &respBody, &toolTrace, &durationMs, &status, &errorDetail, &createdAt,
+				&peerName, &peerRole)
+		} else {
+			scanErr = rows.Scan(&id, &wsID, &actType, &sourceID, &targetID, &method,
+				&summary, &reqBody, &respBody, &toolTrace, &durationMs, &status, &errorDetail, &createdAt)
+		}
+		if scanErr != nil {
+			log.Printf("Activity scan error: %v", scanErr)
 			continue
 		}

@@ -308,6 +575,39 @@ func (h *ActivityHandler) List(c *gin.Context) {
 		if toolTrace != nil {
 			entry["tool_trace"] = json.RawMessage(toolTrace)
 		}
+
+		// peer_info enrichment (per ?include=peer_info). Only emit the
+		// new fields when the flag is set — back-compat for callers
+		// that don't request it.
+		if includePeerInfo {
+			// peer_name / peer_role: emit only when present (canvas
+			// rows have source_id IS NULL → peer_name is NULL by JOIN;
+			// also a peer workspace may have been deleted since the
+			// row was written → same NULL outcome). Omit-when-absent
+			// matches the Layer 3 adaptor's "spread when present"
+			// pattern; canvas_user rows legitimately have no peer_*.
+			if peerName != nil && *peerName != "" {
+				entry["peer_name"] = *peerName
+			}
+			if peerRole != nil && *peerRole != "" {
+				entry["peer_role"] = *peerRole
+			}
+			// agent_card_url: constructed server-side from
+			// externalPlatformURL + source_id. Mirrors the runtime-
+			// side helper a2a_client._agent_card_url_for which builds
+			// {PLATFORM_URL}/registry/discover/{peer_id}. Only set
+			// when source_id is present + non-empty.
+			if sourceID != nil && *sourceID != "" && platformBase != "" {
+				entry["agent_card_url"] = platformBase + "/registry/discover/" + *sourceID
+			}
+			// attachments: flatten file/image/audio parts from the
+			// request_body. nil when none — only project when
+			// non-empty so the omit-when-absent rule holds.
+			if atts := extractAttachmentsFromRequestBody(reqBody); len(atts) > 0 {
+				entry["attachments"] = atts
+			}
+		}
+
 		activities = append(activities, entry)
 	}
 	if err := rows.Err(); err != nil {
@@ -0,0 +1,701 @@
+package handlers
+
+import (
+	"encoding/json"
+	"fmt"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+	"time"
+
+	"github.com/DATA-DOG/go-sqlmock"
+	"github.com/gin-gonic/gin"
+)
+
+// Tests for the `?include=peer_info` activity-feed enrichment.
+//
+// The enrichment is additive + opt-in. When the flag is absent, the
+// existing tests (TestActivityList_SourceCanvas, etc.) prove the wire
+// shape is unchanged. These tests prove:
+//   - When the flag IS set, the LEFT JOIN is issued and the SELECT
+//     adds w.name + w.role.
+//   - peer_name / peer_role surface from the joined row.
+//   - agent_card_url is composed server-side from
+//     externalPlatformURL + source_id and appears for non-canvas rows
+//     (source_id present).
+//   - attachments[] is projected from request_body.params.message.parts
+//     for file/image/audio parts.
+//   - Canvas rows (source_id NULL) do NOT get peer_name / peer_role /
+//     agent_card_url, but DO still appear in the result set (LEFT JOIN
+//     preserves them with NULL peer fields).
+//   - The `include` query param is comma-separable and only recognizes
+//     known flags.
+
+// ---------- includeFlagSet helper unit tests ----------
+
+func TestIncludeFlagSet(t *testing.T) {
+	cases := []struct {
+		query string
+		flag  string
+		want  bool
+	}{
+		{"", "peer_info", false},
+		{"peer_info", "peer_info", true},
+		{"peer_info,attachments", "peer_info", true},
+		{"attachments,peer_info", "peer_info", true},
+		{"attachments , peer_info ", "peer_info", true},
+		{"peer_infos", "peer_info", false},
+		{"peerinfo", "peer_info", false},
+		{"peer_info", "", false},
+		{",,", "peer_info", false},
+	}
+	for _, tc := range cases {
+		got := includeFlagSet(tc.query, tc.flag)
+		if got != tc.want {
+			t.Errorf("includeFlagSet(%q, %q) = %v, want %v", tc.query, tc.flag, got, tc.want)
+		}
+	}
+}
+
+// ---------- extractAttachmentsFromRequestBody unit tests ----------
+
+func TestExtractAttachmentsFromRequestBody_Empty(t *testing.T) {
+	if got := extractAttachmentsFromRequestBody(nil); got != nil {
+		t.Errorf("nil body: want nil, got %v", got)
+	}
+	if got := extractAttachmentsFromRequestBody([]byte("")); got != nil {
+		t.Errorf("empty body: want nil, got %v", got)
+	}
+	if got := extractAttachmentsFromRequestBody([]byte("not json")); got != nil {
+		t.Errorf("non-json body: want nil, got %v", got)
+	}
+}
+
+func TestExtractAttachmentsFromRequestBody_NoAttachments(t *testing.T) {
+	// Text-only message: no file/image/audio parts → nil
+	body := []byte(`{"jsonrpc":"2.0","method":"message/send","params":{"message":{"parts":[{"kind":"text","text":"hi"}]}}}`)
+	if got := extractAttachmentsFromRequestBody(body); got != nil {
+		t.Errorf("text-only: want nil, got %v", got)
+	}
+}
+
+func TestExtractAttachmentsFromRequestBody_FileKindV1(t *testing.T) {
+	// a2a-sdk v1 shape: kind=file, file:{uri,mime_type,name}
+	body := []byte(`{"jsonrpc":"2.0","method":"message/send","params":{"message":{"parts":[
+		{"kind":"text","text":"see attached"},
+		{"kind":"file","file":{"uri":"workspace:foo.pdf","mime_type":"application/pdf","name":"foo.pdf"}}
+	]}}}`)
+	atts := extractAttachmentsFromRequestBody(body)
+	if len(atts) != 1 {
+		t.Fatalf("want 1 attachment, got %d", len(atts))
+	}
+	if atts[0]["kind"] != "file" {
+		t.Errorf("kind: want file, got %v", atts[0]["kind"])
+	}
+	if atts[0]["uri"] != "workspace:foo.pdf" {
+		t.Errorf("uri mismatch: %v", atts[0]["uri"])
+	}
+	if atts[0]["mime_type"] != "application/pdf" {
+		t.Errorf("mime_type mismatch: %v", atts[0]["mime_type"])
+	}
+	if atts[0]["name"] != "foo.pdf" {
+		t.Errorf("name mismatch: %v", atts[0]["name"])
+	}
+}
+
+func TestExtractAttachmentsFromRequestBody_ImageAndAudio(t *testing.T) {
+	// Mixed image + audio parts; both surface
+	body := []byte(`{"jsonrpc":"2.0","method":"message/send","params":{"message":{"parts":[
+		{"kind":"image","file":{"uri":"workspace:a.png","mime_type":"image/png","name":"a.png"}},
+		{"kind":"audio","file":{"uri":"workspace:b.mp3","mime_type":"audio/mpeg","name":"b.mp3"}}
+	]}}}`)
+	atts := extractAttachmentsFromRequestBody(body)
+	if len(atts) != 2 {
+		t.Fatalf("want 2 attachments, got %d", len(atts))
+	}
+	if atts[0]["kind"] != "image" || atts[1]["kind"] != "audio" {
+		t.Errorf("kind order: got %v / %v", atts[0]["kind"], atts[1]["kind"])
+	}
+}
+
+func TestExtractAttachmentsFromRequestBody_LegacyV0TypeDiscriminator(t *testing.T) {
+	// Legacy v0 shape: type=file (not kind), inlined fields (no nested .file)
+	body := []byte(`{"jsonrpc":"2.0","method":"message/send","params":{"message":{"parts":[
+		{"type":"file","uri":"workspace:legacy.txt","mime_type":"text/plain","name":"legacy.txt"}
+	]}}}`)
+	atts := extractAttachmentsFromRequestBody(body)
+	if len(atts) != 1 {
+		t.Fatalf("want 1 attachment, got %d", len(atts))
+	}
+	if atts[0]["kind"] != "file" || atts[0]["uri"] != "workspace:legacy.txt" || atts[0]["name"] != "legacy.txt" {
+		t.Errorf("v0 part not surfaced: %v", atts[0])
+	}
+}
+
+func TestExtractAttachmentsFromRequestBody_SkipsEmptyParts(t *testing.T) {
+	// A "file" part with no uri AND no name is malformed — skip rather
+	// than emit a no-info entry.
+	body := []byte(`{"jsonrpc":"2.0","method":"message/send","params":{"message":{"parts":[
+		{"kind":"file","file":{}},
+		{"kind":"file","file":{"name":"only-name.bin"}}
+	]}}}`)
+	atts := extractAttachmentsFromRequestBody(body)
+	if len(atts) != 1 {
+		t.Fatalf("want 1 attachment (the named one), got %d", len(atts))
+	}
+	if atts[0]["name"] != "only-name.bin" {
+		t.Errorf("expected only-name.bin, got %v", atts[0])
+	}
+}
+
+func TestExtractAttachmentsFromRequestBody_MalformedShape(t *testing.T) {
+	// Various malformed shapes return nil (defensive)
+	for _, b := range []string{
+		`{}`,
+		`{"params":{}}`,
+		`{"params":{"message":{}}}`,
+		`{"params":{"message":{"parts":"not-a-list"}}}`,
+		`{"params":{"message":{"parts":[null,42,"string"]}}}`,
+	} {
+		if got := extractAttachmentsFromRequestBody([]byte(b)); got != nil {
+			t.Errorf("body %q: want nil, got %v", b, got)
+		}
+	}
+}
+
+// ---------- Activity List ?include=peer_info handler tests ----------
+
+func TestActivityList_IncludePeerInfo_IssuesLeftJoin(t *testing.T) {
+	// When ?include=peer_info is set, the query must:
+	//   1. SELECT include w.name + w.role aliased as peer_name/peer_role
+	//   2. FROM contains LEFT JOIN workspaces w ON w.id = activity_logs.source_id
+	//   3. WHERE uses qualified activity_logs.workspace_id (disambiguates
+	//      from workspaces.id post-JOIN)
+	//
+	// Pin all three so a future refactor can't silently drop the JOIN or
+	// the alias and have the test still pass.
+	mock := setupTestDB(t)
+	broadcaster := newTestBroadcaster()
+	handler := NewActivityHandler(broadcaster)
+
+	peerID := "11111111-2222-3333-4444-555555555555"
+	mock.ExpectQuery(
+		`SELECT .+w\.name AS peer_name, w\.role AS peer_role FROM activity_logs LEFT JOIN workspaces w ON w\.id = activity_logs\.source_id WHERE activity_logs\.workspace_id = .+`,
+	).
+		WithArgs("ws-1", 100).
+		WillReturnRows(sqlmock.NewRows([]string{
+			"id", "workspace_id", "activity_type", "source_id", "target_id",
+			"method", "summary", "request_body", "response_body",
+			"tool_trace", "duration_ms", "status", "error_detail", "created_at",
+			"peer_name", "peer_role",
+		}).
+			AddRow("act-1", "ws-1", "a2a_receive", peerID, "ws-1",
+				"message/send", "Agent message: hello",
+				[]byte(`{"jsonrpc":"2.0","method":"message/send","params":{"message":{"parts":[{"kind":"text","text":"hello"}]}}}`),
+				nil, nil, nil, "ok", nil, time.Now(),
+				"Production Manager", "product manager"))
+
+	gin.SetMode(gin.TestMode)
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Params = gin.Params{{Key: "id", Value: "ws-1"}}
+	c.Request = httptest.NewRequest("GET", "/workspaces/ws-1/activity?include=peer_info", nil)
+	c.Request.Host = "platform.test"
+	c.Request.Header.Set("X-Forwarded-Proto", "https")
+	handler.List(c)
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("expected 200, got %d: %s", w.Code, w.Body.String())
+	}
+	var resp []map[string]interface{}
+	if err := json.Unmarshal(w.Body.Bytes(), &resp); err != nil {
+		t.Fatalf("parse: %v", err)
+	}
+	if len(resp) != 1 {
+		t.Fatalf("want 1 row, got %d", len(resp))
+	}
+	r := resp[0]
+	if r["peer_name"] != "Production Manager" {
+		t.Errorf("peer_name: got %v", r["peer_name"])
+	}
+	if r["peer_role"] != "product manager" {
+		t.Errorf("peer_role: got %v", r["peer_role"])
+	}
+	wantURL := "https://platform.test/registry/discover/" + peerID
+	if r["agent_card_url"] != wantURL {
+		t.Errorf("agent_card_url: got %v, want %v", r["agent_card_url"], wantURL)
+	}
+	// Text-only message has no attachments → omit from envelope
+	if _, present := r["attachments"]; present {
+		t.Errorf("attachments should be omitted on text-only row; got %v", r["attachments"])
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Fatalf("unmet expectations: %v", err)
+	}
+}
+
+func TestActivityList_IncludePeerInfo_CanvasRowHasNoPeerFields(t *testing.T) {
+	// LEFT JOIN preserves canvas rows (source_id NULL) but their
+	// peer_name/peer_role come back as NULL — must omit from the
+	// envelope (not emit empty strings or null literals).
+	mock := setupTestDB(t)
+	broadcaster := newTestBroadcaster()
+	handler := NewActivityHandler(broadcaster)
+
+	mock.ExpectQuery(
+		`LEFT JOIN workspaces w ON w\.id = activity_logs\.source_id`,
+	).
+		WithArgs("ws-1", 100).
+		WillReturnRows(sqlmock.NewRows([]string{
+			"id", "workspace_id", "activity_type", "source_id", "target_id",
+			"method", "summary", "request_body", "response_body",
+			"tool_trace", "duration_ms", "status", "error_detail", "created_at",
+			"peer_name", "peer_role",
+		}).
+			// source_id NULL = canvas message; peer columns also NULL.
+			AddRow("act-canvas", "ws-1", "a2a_receive", nil, "ws-1",
+				"notify", "User said hi",
+				[]byte(`{"params":{"message":{"parts":[{"kind":"text","text":"hi"}]}}}`),
+				nil, nil, nil, "ok", nil, time.Now(),
+				nil, nil))
+
+	gin.SetMode(gin.TestMode)
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Params = gin.Params{{Key: "id", Value: "ws-1"}}
+	c.Request = httptest.NewRequest("GET", "/workspaces/ws-1/activity?include=peer_info", nil)
+	handler.List(c)
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("expected 200, got %d: %s", w.Code, w.Body.String())
+	}
+	var resp []map[string]interface{}
+	if err := json.Unmarshal(w.Body.Bytes(), &resp); err != nil {
+		t.Fatalf("parse: %v", err)
+	}
+	if len(resp) != 1 {
+		t.Fatalf("want 1 row, got %d", len(resp))
+	}
+	r := resp[0]
+	for _, k := range []string{"peer_name", "peer_role", "agent_card_url"} {
+		if _, present := r[k]; present {
+			t.Errorf("%s should be absent on canvas row; got %v", k, r[k])
+		}
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Fatalf("unmet expectations: %v", err)
+	}
+}
+
+func TestActivityList_IncludePeerInfo_AttachmentsSurfaceFromRequestBody(t *testing.T) {
+	// A peer_agent message with an inline file attachment must have
+	// attachments[] populated on the envelope.
+	mock := setupTestDB(t)
+	broadcaster := newTestBroadcaster()
+	handler := NewActivityHandler(broadcaster)
+
+	peerID := "11111111-2222-3333-4444-555555555555"
+	mock.ExpectQuery(`LEFT JOIN workspaces`).
+		WithArgs("ws-1", 100).
+		WillReturnRows(sqlmock.NewRows([]string{
+			"id", "workspace_id", "activity_type", "source_id", "target_id",
+			"method", "summary", "request_body", "response_body",
+			"tool_trace", "duration_ms", "status", "error_detail", "created_at",
+			"peer_name", "peer_role",
+		}).
+			AddRow("act-with-file", "ws-1", "a2a_receive", peerID, "ws-1",
+				"message/send", "Agent message: see attached",
+				[]byte(`{"jsonrpc":"2.0","method":"message/send","params":{"message":{"parts":[
+					{"kind":"text","text":"see attached"},
+					{"kind":"file","file":{"uri":"workspace:foo.pdf","mime_type":"application/pdf","name":"foo.pdf"}}
+				]}}}`),
+				nil, nil, nil, "ok", nil, time.Now(),
+				"Code Reviewer", "code reviewer"))
+
+	gin.SetMode(gin.TestMode)
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Params = gin.Params{{Key: "id", Value: "ws-1"}}
+	c.Request = httptest.NewRequest("GET", "/workspaces/ws-1/activity?include=peer_info", nil)
+	handler.List(c)
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("expected 200, got %d: %s", w.Code, w.Body.String())
+	}
+	var resp []map[string]interface{}
+	if err := json.Unmarshal(w.Body.Bytes(), &resp); err != nil {
+		t.Fatalf("parse: %v", err)
+	}
+	r := resp[0]
+	atts, ok := r["attachments"].([]interface{})
+	if !ok {
+		t.Fatalf("attachments missing or wrong type: %T %v", r["attachments"], r["attachments"])
+	}
+	if len(atts) != 1 {
+		t.Fatalf("want 1 attachment, got %d: %v", len(atts), atts)
+	}
+	att := atts[0].(map[string]interface{})
+	if att["kind"] != "file" || att["uri"] != "workspace:foo.pdf" || att["name"] != "foo.pdf" {
+		t.Errorf("attachment shape: %v", att)
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Fatalf("unmet expectations: %v", err)
+	}
+}
+
+func TestActivityList_IncludePeerInfo_Unset_NoJoinNoExtraFields(t *testing.T) {
+	// Back-compat — when ?include=peer_info is NOT passed, the SELECT
+	// uses unqualified column refs (no `activity_logs.` prefix) AND no
+	// JOIN. Existing tests pass this implicitly; this test pins it
+	// explicitly so a future refactor that accidentally turns the JOIN
+	// always-on gets caught.
+	mock := setupTestDB(t)
+	broadcaster := newTestBroadcaster()
+	handler := NewActivityHandler(broadcaster)
+
+	// Regex pinned: "FROM activity_logs WHERE workspace_id" — no JOIN
+	// keyword between FROM and WHERE; no `activity_logs.` qualifier on
+	// workspace_id.
+	mock.ExpectQuery(`SELECT id, workspace_id,.+ FROM activity_logs WHERE workspace_id = .+`).
+		WithArgs("ws-1", 100).
+		WillReturnRows(sqlmock.NewRows([]string{
+			"id", "workspace_id", "activity_type", "source_id", "target_id",
+			"method", "summary", "request_body", "response_body",
+			"tool_trace", "duration_ms", "status", "error_detail", "created_at",
+		}).
+			AddRow("act-1", "ws-1", "a2a_receive", "11111111-2222-3333-4444-555555555555", "ws-1",
+				"message/send", "Hello",
+				nil, nil, nil, nil, "ok", nil, time.Now()))
+
+	gin.SetMode(gin.TestMode)
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Params = gin.Params{{Key: "id", Value: "ws-1"}}
+	c.Request = httptest.NewRequest("GET", "/workspaces/ws-1/activity", nil)
+	handler.List(c)
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("expected 200, got %d: %s", w.Code, w.Body.String())
+	}
+	var resp []map[string]interface{}
+	if err := json.Unmarshal(w.Body.Bytes(), &resp); err != nil {
+		t.Fatalf("parse: %v", err)
+	}
+	if len(resp) != 1 {
+		t.Fatalf("want 1 row, got %d", len(resp))
+	}
+	// Confirm no peer_info enrichment leaks into the default envelope.
+	for _, k := range []string{"peer_name", "peer_role", "agent_card_url", "attachments"} {
+		if _, present := resp[0][k]; present {
+			t.Errorf("%s must NOT appear without ?include=peer_info; got %v", k, resp[0][k])
+		}
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Fatalf("unmet expectations: %v", err)
+	}
+}
+
+func TestActivityList_IncludePeerInfo_UnknownFlagIgnored(t *testing.T) {
+	// ?include=bogus must NOT issue the JOIN — only the recognized
+	// `peer_info` flag triggers enrichment. The unknown flag is silently
+	// ignored (additive, opt-in convention).
+	mock := setupTestDB(t)
+	broadcaster := newTestBroadcaster()
+	handler := NewActivityHandler(broadcaster)
+
+	mock.ExpectQuery(`SELECT id, workspace_id,.+ FROM activity_logs WHERE workspace_id = .+`).
+		WithArgs("ws-1", 100).
+		WillReturnRows(sqlmock.NewRows([]string{
+			"id", "workspace_id", "activity_type", "source_id", "target_id",
+			"method", "summary", "request_body", "response_body",
+			"tool_trace", "duration_ms", "status", "error_detail", "created_at",
+		}))
+
+	gin.SetMode(gin.TestMode)
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Params = gin.Params{{Key: "id", Value: "ws-1"}}
+	c.Request = httptest.NewRequest("GET", "/workspaces/ws-1/activity?include=bogus", nil)
+	handler.List(c)
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("expected 200, got %d", w.Code)
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Fatalf("unmet expectations: %v", err)
+	}
+}
+
+// ---------- flat upload manifest (chat_upload_receive) tests ----------
+
+func TestKindFromMimeType(t *testing.T) {
+	cases := []struct {
+		mime string
+		want string
+	}{
+		{"image/png", "image"},
+		{"image/jpeg", "image"},
+		{"image/", "image"}, // prefix-only is still image
+		{"audio/mpeg", "audio"},
+		{"audio/wav", "audio"},
+		{"video/mp4", "video"},
+		{"video/webm", "video"},
+		{"application/pdf", "file"},
+		{"text/plain", "file"},
+		{"", "file"},
+		{"unknown", "file"},
+		{"image", "file"}, // no slash → not a prefix match
+	}
+	for _, tc := range cases {
+		if got := kindFromMimeType(tc.mime); got != tc.want {
+			t.Errorf("kindFromMimeType(%q) = %q, want %q", tc.mime, got, tc.want)
+		}
+	}
+}
+
+func TestExtractAttachmentsFromRequestBody_FlatUpload_Image(t *testing.T) {
+	// Canvas chat_upload_receive shape: flat manifest at request_body
+	// root with camelCase mimeType. The empirical example was a PNG
+	// pasted into the canvas; surfaces here with kind=image,
+	// mime_type=image/png (snake-case normalized), uri preserved.
+	body := []byte(`{
+		"uri":"platform-pending:091a9180-/26111d48-",
+		"name":"pasted-2026-05-21T23-12-25-0-0.png",
+		"size":677133,
+		"file_id":"26111d48-",
+		"mimeType":"image/png"
+	}`)
+	atts := extractAttachmentsFromRequestBody(body)
+	if len(atts) != 1 {
+		t.Fatalf("want 1 attachment, got %d: %v", len(atts), atts)
+	}
+	att := atts[0]
+	if att["kind"] != "image" {
+		t.Errorf("kind: want image, got %v", att["kind"])
+	}
+	if att["uri"] != "platform-pending:091a9180-/26111d48-" {
+		t.Errorf("uri: %v", att["uri"])
+	}
+	if att["mime_type"] != "image/png" {
+		t.Errorf("mime_type normalization (camelCase→snake_case) failed: %v", att["mime_type"])
+	}
+	if att["name"] != "pasted-2026-05-21T23-12-25-0-0.png" {
+		t.Errorf("name: %v", att["name"])
+	}
+	// camelCase `mimeType` MUST NOT leak into the projected envelope —
+	// only snake_case `mime_type` is the wire convention.
+	if _, present := att["mimeType"]; present {
+		t.Errorf("camelCase mimeType leaked into envelope: %v", att)
+	}
+	if _, present := att["file_id"]; present {
+		t.Errorf("file_id should not be surfaced on the attachment envelope (it's a canvas-internal id): %v", att)
+	}
+}
+
+func TestExtractAttachmentsFromRequestBody_FlatUpload_Audio(t *testing.T) {
+	body := []byte(`{"uri":"platform-pending:ws/file","name":"voice.mp3","file_id":"abc","mimeType":"audio/mpeg"}`)
+	atts := extractAttachmentsFromRequestBody(body)
+	if len(atts) != 1 || atts[0]["kind"] != "audio" {
+		t.Fatalf("want audio kind, got %v", atts)
+	}
+	if atts[0]["mime_type"] != "audio/mpeg" {
+		t.Errorf("mime_type: %v", atts[0]["mime_type"])
+	}
+}
+
+func TestExtractAttachmentsFromRequestBody_FlatUpload_Video(t *testing.T) {
+	body := []byte(`{"uri":"platform-pending:ws/file","name":"clip.mp4","file_id":"abc","mimeType":"video/mp4"}`)
+	atts := extractAttachmentsFromRequestBody(body)
+	if len(atts) != 1 || atts[0]["kind"] != "video" {
+		t.Fatalf("want video kind, got %v", atts)
+	}
+}
+
+func TestExtractAttachmentsFromRequestBody_FlatUpload_GenericFile(t *testing.T) {
+	// application/pdf has no image/audio/video prefix → kind=file
+	body := []byte(`{"uri":"platform-pending:ws/file","name":"doc.pdf","file_id":"abc","mimeType":"application/pdf"}`)
+	atts := extractAttachmentsFromRequestBody(body)
+	if len(atts) != 1 || atts[0]["kind"] != "file" {
+		t.Fatalf("want file kind, got %v", atts)
+	}
+}
+
+func TestExtractAttachmentsFromRequestBody_FlatUpload_NoMimeFallsToFile(t *testing.T) {
+	// No mimeType at all — kind defaults to "file", mime_type omitted.
+	body := []byte(`{"uri":"platform-pending:ws/file","name":"unknown.bin","file_id":"abc"}`)
+	atts := extractAttachmentsFromRequestBody(body)
+	if len(atts) != 1 {
+		t.Fatalf("want 1 attachment, got %d", len(atts))
+	}
+	if atts[0]["kind"] != "file" {
+		t.Errorf("kind: want file (default), got %v", atts[0]["kind"])
+	}
+	if _, present := atts[0]["mime_type"]; present {
+		t.Errorf("mime_type should be omitted when source has none, got %v", atts[0]["mime_type"])
+	}
+}
+
+func TestExtractAttachmentsFromRequestBody_FlatUpload_SnakeCaseMimeTypeAccepted(t *testing.T) {
+	// Defensive: a future canvas version (or non-canvas caller) that
+	// already emits snake_case mime_type should still be parsed.
+	body := []byte(`{"uri":"u","name":"n.png","mime_type":"image/png"}`)
+	atts := extractAttachmentsFromRequestBody(body)
+	if len(atts) != 1 {
+		t.Fatalf("want 1 attachment, got %d", len(atts))
+	}
+	if atts[0]["mime_type"] != "image/png" || atts[0]["kind"] != "image" {
+		t.Errorf("snake_case mime_type not honored: %v", atts[0])
+	}
+}
+
+func TestExtractAttachmentsFromRequestBody_FlatUpload_FileIDOnlyIsSkipped(t *testing.T) {
+	// file_id alone (no uri AND no name) is non-actionable — the
+	// downstream adaptor can't render a discoverable file from just an
+	// internal canvas id. Skip per the same minimum-info rule the
+	// message-parts arm applies to empty parts.
+	body := []byte(`{"file_id":"orphan-uuid","mimeType":"image/png"}`)
+	if got := extractAttachmentsFromRequestBody(body); got != nil {
+		t.Errorf("file_id-only manifest must be skipped, got %v", got)
+	}
+}
+
+func TestExtractAttachmentsFromRequestBody_FlatUpload_NameOnlyIsKept(t *testing.T) {
+	// Symmetric with the message-parts arm: a name without uri is still
+	// useful (the downstream adaptor can render "user uploaded foo.png").
+	body := []byte(`{"name":"only-name.bin","file_id":"abc","mimeType":"application/octet-stream"}`)
+	atts := extractAttachmentsFromRequestBody(body)
+	if len(atts) != 1 {
+		t.Fatalf("want 1 attachment, got %d", len(atts))
+	}
+	if atts[0]["name"] != "only-name.bin" {
+		t.Errorf("name not preserved: %v", atts[0])
+	}
+	if _, present := atts[0]["uri"]; present {
+		t.Errorf("uri should be omitted when absent in source, got %v", atts[0]["uri"])
+	}
+}
+
+func TestExtractAttachmentsFromRequestBody_MessagePartsTakesPrecedenceOverFlat(t *testing.T) {
+	// If a single request_body somehow has BOTH params.message.parts[]
+	// AND top-level uri/file_id (a pathological inbound), the
+	// message-parts arm wins — that's the documented inbound shape and
+	// it's been the only one historically extracted. The flat arm is a
+	// fallback for shapes that have NO parts.
+	body := []byte(`{
+		"uri":"platform-pending:should-not-win",
+		"file_id":"x",
+		"mimeType":"image/png",
+		"params":{"message":{"parts":[
+			{"kind":"file","file":{"uri":"workspace:should-win.pdf","mime_type":"application/pdf","name":"win.pdf"}}
+		]}}
+	}`)
+	atts := extractAttachmentsFromRequestBody(body)
+	if len(atts) != 1 {
+		t.Fatalf("want 1 attachment (from parts[]), got %d: %v", len(atts), atts)
+	}
+	if atts[0]["uri"] != "workspace:should-win.pdf" {
+		t.Errorf("message-parts arm did not take precedence: %v", atts[0])
+	}
+}
+
+func TestActivityList_IncludePeerInfo_ChatUploadReceiveCanvasRow(t *testing.T) {
+	// Wire-level integration: a canvas chat_upload_receive row (canvas
+	// user pasted an image) with source_id NULL (canvas message), flat
+	// upload manifest at request_body root. The `?include=peer_info`
+	// projection must surface attachments[] populated from the flat-
+	// upload-manifest arm while peer_name / peer_role / agent_card_url
+	// remain absent (canvas row has no peer).
+	mock := setupTestDB(t)
+	broadcaster := newTestBroadcaster()
+	handler := NewActivityHandler(broadcaster)
+
+	mock.ExpectQuery(`LEFT JOIN workspaces w ON w\.id = activity_logs\.source_id`).
+		WithArgs("ws-1", 100).
+		WillReturnRows(sqlmock.NewRows([]string{
+			"id", "workspace_id", "activity_type", "source_id", "target_id",
+			"method", "summary", "request_body", "response_body",
+			"tool_trace", "duration_ms", "status", "error_detail", "created_at",
+			"peer_name", "peer_role",
+		}).
+			// Empirical shape from 2026-05-21 ~23:12Z agents-team canvas paste.
+			AddRow("act-upload", "ws-1", "chat_upload_receive", nil, "ws-1",
+				"chat_upload_receive", "Canvas upload: pasted-2026-05-21T23-12-25-0-0.png",
+				[]byte(`{
+					"uri":"platform-pending:091a9180-b303-4a20-aefe-3a4a675b8aa4/26111d48-aaaa-bbbb-cccc-dddddddddddd",
+					"name":"pasted-2026-05-21T23-12-25-0-0.png",
+					"size":677133,
+					"file_id":"26111d48-aaaa-bbbb-cccc-dddddddddddd",
+					"mimeType":"image/png"
+				}`),
+				nil, nil, nil, "ok", nil, time.Now(),
+				nil, nil))
+
+	gin.SetMode(gin.TestMode)
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Params = gin.Params{{Key: "id", Value: "ws-1"}}
+	c.Request = httptest.NewRequest("GET", "/workspaces/ws-1/activity?include=peer_info", nil)
+	handler.List(c)
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("expected 200, got %d: %s", w.Code, w.Body.String())
+	}
+	var resp []map[string]interface{}
+	if err := json.Unmarshal(w.Body.Bytes(), &resp); err != nil {
+		t.Fatalf("parse: %v", err)
+	}
+	if len(resp) != 1 {
+		t.Fatalf("want 1 row, got %d", len(resp))
+	}
+	r := resp[0]
+	// Canvas row → no peer fields.
+	for _, k := range []string{"peer_name", "peer_role", "agent_card_url"} {
+		if _, present := r[k]; present {
+			t.Errorf("%s must NOT appear on canvas upload row; got %v", k, r[k])
+		}
+	}
+	// attachments[] populated from the flat-upload arm.
+	atts, ok := r["attachments"].([]interface{})
+	if !ok {
+		t.Fatalf("attachments missing or wrong type: %T %v", r["attachments"], r["attachments"])
+	}
+	if len(atts) != 1 {
+		t.Fatalf("want 1 attachment from flat manifest, got %d: %v", len(atts), atts)
+	}
+	att := atts[0].(map[string]interface{})
+	if att["kind"] != "image" {
+		t.Errorf("kind: want image (image/png prefix), got %v", att["kind"])
+	}
+	if att["mime_type"] != "image/png" {
+		t.Errorf("mime_type wire shape: want snake_case image/png, got %v", att["mime_type"])
+	}
+	if att["uri"] != "platform-pending:091a9180-b303-4a20-aefe-3a4a675b8aa4/26111d48-aaaa-bbbb-cccc-dddddddddddd" {
+		t.Errorf("uri preserved verbatim: got %v", att["uri"])
+	}
+	if att["name"] != "pasted-2026-05-21T23-12-25-0-0.png" {
+		t.Errorf("name: %v", att["name"])
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Fatalf("unmet expectations: %v", err)
+	}
+}
+
+// Sanity test using the existing test broadcaster setup — verifies the
+// extractAttachments helper round-trips through json.Marshal cleanly
+// (no map ordering issues, no type-coercion surprises).
+func TestExtractAttachmentsFromRequestBody_RoundTripsThroughJSON(t *testing.T) {
+	body := []byte(`{"params":{"message":{"parts":[{"kind":"file","file":{"uri":"workspace:r.bin","mime_type":"application/octet-stream","name":"r.bin"}}]}}}`)
+	atts := extractAttachmentsFromRequestBody(body)
+	b, err := json.Marshal(atts)
+	if err != nil {
+		t.Fatalf("marshal: %v", err)
+	}
+	var decoded []map[string]interface{}
+	if err := json.Unmarshal(b, &decoded); err != nil {
+		t.Fatalf("unmarshal: %v", err)
+	}
+	if len(decoded) != 1 || decoded[0]["uri"] != "workspace:r.bin" {
+		t.Fatalf("round-trip mismatch: %v", decoded)
+	}
+	_ = fmt.Sprintf // keep fmt import live if test trimming removes usage
+}
@@ -216,69 +216,102 @@ curl -fsS -X POST "{{PLATFORM_URL}}/registry/register" \
 const externalChannelTemplate = `# Claude Code channel — bridges this workspace's A2A traffic into your
 # Claude Code session. No tunnel/public URL needed (polling-based).
 #
-# Prereq: Bun installed (channel plugins are Bun scripts).
-#   bun --version    # must print a version number
+# Prereq: Bun 1.3+ installed (channel plugins are Bun scripts).
+#   bun --version    # must print a version (1.3.x or newer)
 #
-# 1. Inside Claude Code, install the channel plugin from its GitHub repo.
-#    The plugin is NOT on Anthropic's default allowlist, so a one-time
-#    marketplace-add is needed before install:
+# 1. Inside Claude Code, install the channel plugin. The plugin lives in
+#    Molecule's own Gitea marketplace (not Anthropic's default), so a
+#    one-time marketplace-add is needed before install:
 #
 #      /plugin marketplace add https://git.moleculesai.app/molecule-ai/molecule-mcp-claude-channel.git
 #      /plugin install molecule@molecule-channel
 #
-#    Then either run /reload-plugins or restart Claude Code so the
-#    plugin is registered.
+#    Then /reload-plugins (or restart Claude Code) so the plugin is
+#    registered.
 #
-# 2. Create the per-watched-workspace config file:
+# 2. Create (or extend) the per-host config file. The canonical SSOT
+#    shape is MOLECULE_WORKSPACES_JSON — a JSON array of
+#    {id, token, platform_url} objects. One plugin instance can watch
+#    many workspaces across many tenants; append more objects to the
+#    array (separate them with commas, NOT a newline):
 mkdir -p ~/.claude/channels/molecule
 cat > ~/.claude/channels/molecule/.env <<'EOF'
-MOLECULE_PLATFORM_URL={{PLATFORM_URL}}
-MOLECULE_WORKSPACE_IDS={{WORKSPACE_ID}}
-MOLECULE_WORKSPACE_TOKENS=<paste auth_token from create response>
+MOLECULE_WORKSPACES_JSON=[{"id":"{{WORKSPACE_ID}}","token":"<paste auth_token from create response>","platform_url":"{{PLATFORM_URL}}"}]
 EOF
 chmod 600 ~/.claude/channels/molecule/.env

-# 3. Launch Claude Code with the channel enabled. Custom (non-Anthropic-
-#    allowlisted) channels need the --dangerously-load-development-channels
-#    flag to opt in — without it, you'll see "not on the approved channels
-#    allowlist" on startup.
-claude --dangerously-load-development-channels \
-  --channels plugin:molecule@molecule-channel
+# (Legacy single-platform shape — MOLECULE_PLATFORM_URL + comma-separated
+# MOLECULE_WORKSPACE_IDS + MOLECULE_WORKSPACE_TOKENS — is still supported
+# for back-compat but does NOT work across multiple tenant URLs. Use
+# MOLECULE_WORKSPACES_JSON above unless you have a specific reason.)
+
+# 3. Launch Claude Code with the channel enabled. The channel spec is the
+#    VALUE of --dangerously-load-development-channels — NOT a separate
+#    --channels flag (that flag does not exist in current Claude Code;
+#    passing it errors with "entries must be tagged: --channels").
+claude --dangerously-load-development-channels plugin:molecule@molecule-channel

 # You should see on stderr:
-#   molecule channel: connected — watching 1 workspace(s) at {{PLATFORM_URL}}
+#   molecule channel: connected — watching N workspace(s) across M platform(s)
+#     targets: <platform_url>: <workspace_id>
 #
-# Inbound A2A messages now surface as conversation turns. Claude's
-# replies route back via the reply_to_workspace MCP tool — no extra
-# wiring on your side.
+# Inbound A2A messages now surface as conversation turns (synthetic
+# <channel ...> tags). Claude's replies route back via the
+# reply_to_workspace / send_message_to_user MCP tools.
+#
+# Multi-workspace note: when watching more than one workspace, every
+# outbound tool call (send_message_to_user, reply_to_workspace,
+# delegate_task, list_peers) MUST pass _as_workspace=<id> so the plugin
+# knows which token to authenticate with. The host returns -32603 if you
+# forget — the synthetic <channel> tag's "watching_as" attribute tells
+# you which id to use.
 #
 # Common errors:
-#   "plugin not installed"            → Step 1 didn't run; run /plugin install
+#   "plugin not installed"            → Step 1 didn't run; run /plugin
+#                                       marketplace add + /plugin install
 #                                       inside Claude Code, then /reload-plugins.
-#   "not on approved channels allowlist" → Add --dangerously-load-development-channels
-#                                       to the launch command (Step 3).
-#   "config-missing"                  → ~/.claude/channels/molecule/.env not
-#                                       readable; re-run Step 2 and check chmod.
+#   "entries must be tagged"          → You passed --channels separately.
+#                                       Put plugin:molecule@molecule-channel
+#                                       directly after
+#                                       --dangerously-load-development-channels.
+#   "not on approved channels allowlist" → Org policy gating. See "managed
+#                                       settings" note below.
+#   "config-missing"                  → ~/.claude/channels/molecule/.env
+#                                       not readable; re-run Step 2 and check
+#                                       chmod 600.
 #
-# Team/Enterprise orgs: the --dangerously-load-development-channels flag is
-# blocked by managed settings. Your admin must set channelsEnabled=true and
-# add the plugin to allowedChannelPlugins in claude.ai admin settings.
+# Team/Enterprise plans: the channel allowlist is gated by org policy
+# AND must be written to the local managed-settings.json file on disk
+# (not the claude.ai web admin UI — there is no web toggle for this).
+# Path per OS:
+#   macOS:   /Library/Application Support/ClaudeCode/managed-settings.json
+#   Linux:   /etc/claude-code/managed-settings.json
+#   Windows: C:\ProgramData\ClaudeCode\managed-settings.json
+# Set channelsEnabled: true and add
+#   { "plugin": "molecule", "marketplace": "molecule-channel" }
+# to allowedChannelPlugins. Restart Claude Code after writing the file.
+# A user-level ~/.claude/settings.json does NOT work on Team/Enterprise
+# — this is the single most common reason a freshly-installed plugin
+# appears to do nothing.
 #
-# Multi-workspace: comma-separate IDs and tokens (same order). See
-# https://git.moleculesai.app/molecule-ai/molecule-mcp-claude-channel for
-# pairing flow, push-mode upgrade, and v0.2 roadmap.
+# Pro/Max plans skip the channelsEnabled gate but still need the
+# allowedChannelPlugins entry in the managed-settings file.

 # Need help?
 #   Documentation: https://doc.moleculesai.app/docs/guides/claude-code-channel-plugin
+#   Full README:   https://git.moleculesai.app/molecule-ai/molecule-mcp-claude-channel
 #   Common errors:
 #     • "plugin not installed" — run /plugin marketplace add then
 #       /plugin install lines above; /reload-plugins or restart.
+#     • "entries must be tagged: --channels" — the launch flag form
+#       changed; use --dangerously-load-development-channels plugin:molecule@molecule-channel
+#       (channel spec is the VALUE, not a separate --channels flag).
 #     • "not on the approved channels allowlist" — custom channels need
-#       --dangerously-load-development-channels; team/enterprise orgs
-#       need admin to set channelsEnabled + allowedChannelPlugins.
+#       allowedChannelPlugins in /Library/Application Support/ClaudeCode/managed-settings.json
+#       (macOS) / equivalent on Linux+Windows. NOT a web setting.
 #     • "Inbound messages not arriving" — stderr should show
 #       "molecule channel: connected — watching N workspace(s)";
-#       verify ~/.claude/channels/molecule/.env has PLATFORM_URL + token.
+#       verify ~/.claude/channels/molecule/.env shape is MOLECULE_WORKSPACES_JSON.
 `

 // externalUniversalMcpTemplate — runtime-agnostic standalone path.
@@ -670,7 +703,15 @@ def heartbeat(client, url, ws, tok, start):
    r.raise_for_status()

 def poll_inbound(client, url, ws, tok, since_id):
-    params = {"since_secs": "30", "limit": "50"}
+    # include=peer_info opts into Layer 1's row-level projection so each
+    # polled activity carries peer_name, peer_role, agent_card_url, and
+    # attachments[] inline (when source_id resolves to a peer / when the
+    # message included a file). Pre-Layer-1 platforms ignore unknown query
+    # params and return the bare row shape, so this is back-compat. Use
+    # the extra fields in your reply logic — e.g. address the sender by
+    # peer_name rather than UUID, or Read attached files via the workspace:
+    # URIs in attachments[].
+    params = {"since_secs": "30", "limit": "50", "include": "peer_info"}
    if since_id:
        params["since_id"] = since_id
    r = client.get(f"{url}/workspaces/{ws}/activity", params=params, headers=hdrs(url, tok))
@@ -737,10 +778,16 @@ python3 ~/.molecule-ai/kimi-{{MCP_SERVER_NAME}}/kimi_bridge.py
 # What the script does:
 #   • Registers the workspace in poll mode (no public URL needed)
 #   • Heartbeats every 20s to keep STATUS = online on the canvas
-#   • Polls /workspaces/:id/activity every 5s for new canvas messages
+#   • Polls /workspaces/:id/activity?include=peer_info every 5s — Layer 1
+#     enrichment surfaces peer_name / peer_role / agent_card_url /
+#     attachments[] inline on each polled row when applicable
 #   • Echo-replies via POST /workspaces/:id/notify
 #
 # To change the reply logic, edit the send_reply() call inside the loop.
+# Each polled item has top-level peer_name / peer_role / agent_card_url
+# fields (peer_agent rows) and attachments[] (any kind) when Layer 1 is
+# enabled on the platform — use them to disambiguate senders and to Read
+# attached files via the workspace: URIs.
 # To send a one-off reply from another terminal:
 #   curl -fsS -X POST "{{PLATFORM_URL}}/workspaces/{{WORKSPACE_ID}}/notify" \
 #     -H "Authorization: Bearer $(cat ~/.molecule-ai/kimi-{{MCP_SERVER_NAME}}/env | grep TOKEN | cut -d= -f2)" \
@@ -118,3 +118,86 @@ func TestExternalTemplates_NoBrokenMoleculeAIGitHubURLs(t *testing.T) {
 		}
 	}
 }
+
+// TestExternalChannelTemplate_LaunchFlagShape pins the Claude Code channel
+// snippet to the working launch invocation. The channel spec must be the
+// VALUE of --dangerously-load-development-channels, NOT a separate
+// --channels flag. The two-flag form (`--dangerously-load-development-channels
+// --channels plugin:molecule@...`) errors with "entries must be tagged:
+// --channels" on current Claude Code builds (2.1.143+) and silently no-ops
+// on older ones — either way, new users hit a wall on first launch.
+//
+// Empirical: hit by a session walking through this exact snippet 2026-05-21;
+// the broken form was copy-pasted from this template, ran, errored, and
+// confused the operator into believing the plugin install was broken when
+// the snippet itself was the bug.
+func TestExternalChannelTemplate_LaunchFlagShape(t *testing.T) {
+	// The broken two-flag form. If this string ever appears in the
+	// snippet again, the same onboarding pothole returns.
+	bannedFormBroken := "--dangerously-load-development-channels \\\n  --channels plugin:molecule@molecule-channel"
+	if strings.Contains(externalChannelTemplate, bannedFormBroken) {
+		t.Errorf("externalChannelTemplate contains the broken two-flag launch form. " +
+			"Use --dangerously-load-development-channels plugin:molecule@molecule-channel (spec as value, not a separate --channels flag).")
+	}
+
+	// The single-flag form must be present.
+	requiredFormGood := "--dangerously-load-development-channels plugin:molecule@molecule-channel"
+	if !strings.Contains(externalChannelTemplate, requiredFormGood) {
+		t.Errorf("externalChannelTemplate must contain %q so operators see the working launch invocation", requiredFormGood)
+	}
+}
+
+// TestExternalChannelTemplate_CanonicalEnvShape pins the canvas-served
+// .env example to the canonical SSOT shape (MOLECULE_WORKSPACES_JSON)
+// rather than the legacy single-platform shape. The legacy form
+// (MOLECULE_PLATFORM_URL + comma-separated IDs/TOKENS) is still accepted
+// by the channel plugin's parseWorkspaceTargets but is single-tenant
+// only — it silently fails to onboard users who want to watch multiple
+// platforms (e.g. hongming + agents-team from the same plugin instance),
+// which is the post-PR#15 expected use case.
+func TestExternalChannelTemplate_CanonicalEnvShape(t *testing.T) {
+	if !strings.Contains(externalChannelTemplate, "MOLECULE_WORKSPACES_JSON=") {
+		t.Errorf("externalChannelTemplate must use MOLECULE_WORKSPACES_JSON as the canonical .env shape (the post-PR#15 SSOT)")
+	}
+	// The JSON example must contain the workspace_id + platform_url placeholders
+	// so the canvas substitutes them at serve time.
+	for _, ph := range []string{"{{WORKSPACE_ID}}", "{{PLATFORM_URL}}"} {
+		if !strings.Contains(externalChannelTemplate, ph) {
+			t.Errorf("externalChannelTemplate must contain placeholder %q so the canvas substitutes per-workspace values", ph)
+		}
+	}
+}
+
+// TestPollingTemplates_OptIntoPeerInfo pins the invariant that any template
+// which calls /workspaces/:id/activity for inbound delivery requests the
+// Layer 1 enrichment via ?include=peer_info. Without this opt-in, the
+// platform returns bare activity rows and the operator's bridge / channel
+// loses peer_name / peer_role / agent_card_url / attachments[] — they're
+// available on the server but not delivered.
+//
+// Pre-Layer-1 platforms ignore unknown query params (HTTP spec: filters
+// not understood are dropped), so this is back-compat across deploys.
+//
+// The Claude Code channel template doesn't include the poll URL in this
+// snippet — its polling lives in the plugin's own server.ts (handled by
+// molecule-mcp-claude-channel PR#21). The Kimi template DOES include a
+// poll loop in its kimi_bridge.py block, so the invariant applies there.
+func TestPollingTemplates_OptIntoPeerInfo(t *testing.T) {
+	pollingTemplates := map[string]string{
+		"externalKimiTemplate": externalKimiTemplate,
+	}
+	for name, body := range pollingTemplates {
+		// If the snippet polls /activity, it must opt into peer_info.
+		// The detection is intentionally loose ("/activity" appears in
+		// the script) — operators who customize the script keep the
+		// invariant only if the include hint is in the template.
+		if !strings.Contains(body, "/activity") {
+			t.Errorf("%s no longer polls /activity — review whether this test still applies", name)
+			continue
+		}
+		if !strings.Contains(body, `"include": "peer_info"`) && !strings.Contains(body, "include=peer_info") {
+			t.Errorf("%s polls /activity without ?include=peer_info — operators lose Layer 1 enrichment "+
+				"(peer_name / peer_role / agent_card_url / attachments[]). Add the param to the poll URL.", name)
+		}
+	}
+}
@@ -230,7 +230,7 @@ func TestWorkspaceList_WithData(t *testing.T) {
 	broadcaster := newTestBroadcaster()
 	handler := NewWorkspaceHandler(broadcaster, nil, "http://localhost:8080", t.TempDir())

-	// 23 cols — broadcast_enabled + talk_to_user_enabled added after monthly_spend
+	// 24 cols — compute added after talk_to_user_enabled.
 	// (migration 20260514). Column order must match scanWorkspaceRow exactly.
 	columns := []string{
 		"id", "name", "role", "tier", "status", "agent_card", "url",
@@ -238,13 +238,13 @@ func TestWorkspaceList_WithData(t *testing.T) {
 		"last_error_rate", "last_sample_error",
 		"uptime_seconds", "current_task", "runtime", "workspace_dir", "x", "y", "collapsed",
 		"budget_limit", "monthly_spend",
-		"broadcast_enabled", "talk_to_user_enabled",
+		"broadcast_enabled", "talk_to_user_enabled", "compute",
 	}
 	rows := sqlmock.NewRows(columns).
 		AddRow("ws-1", "Agent One", "worker", 1, "online", []byte(`{"name":"agent1"}`), "http://localhost:8001",
-			nil, 3, 1, 0.02, "", 7200, "processing", "langgraph", "", 10.0, 20.0, false, nil, int64(0), false, true).
+			nil, 3, 1, 0.02, "", 7200, "processing", "langgraph", "", 10.0, 20.0, false, nil, int64(0), false, true, []byte(`{}`)).
 		AddRow("ws-2", "Agent Two", "", 2, "degraded", []byte("null"), "",
-			nil, 0, 1, 0.6, "timeout", 100, "", "claude-code", "", 50.0, 60.0, true, nil, int64(0), false, true)
+			nil, 0, 1, 0.6, "timeout", 100, "", "claude-code", "", 50.0, 60.0, true, nil, int64(0), false, true, []byte(`{}`))

 	mock.ExpectQuery("SELECT w.id, w.name").
 		WillReturnRows(rows)
@@ -456,7 +456,7 @@ func TestWorkspaceList(t *testing.T) {
 	broadcaster := newTestBroadcaster()
 	handler := NewWorkspaceHandler(broadcaster, nil, "http://localhost:8080", "/tmp/configs")

-	// 23 cols: broadcast_enabled + talk_to_user_enabled added after monthly_spend
+	// 24 cols: compute added after talk_to_user_enabled.
 	// (migration 20260514). Column order must match scanWorkspaceRow exactly.
 	columns := []string{
 		"id", "name", "role", "tier", "status", "agent_card", "url",
@@ -464,13 +464,13 @@ func TestWorkspaceList(t *testing.T) {
 		"last_error_rate", "last_sample_error",
 		"uptime_seconds", "current_task", "runtime", "workspace_dir", "x", "y", "collapsed",
 		"budget_limit", "monthly_spend",
-		"broadcast_enabled", "talk_to_user_enabled",
+		"broadcast_enabled", "talk_to_user_enabled", "compute",
 	}
 	rows := sqlmock.NewRows(columns).
 		AddRow("ws-1", "Agent One", "worker", 1, "online", []byte("null"), "http://localhost:8001",
-			nil, 0, 1, 0.0, "", 100, "", "claude-code", "", 10.0, 20.0, false, nil, int64(0), false, true).
+			nil, 0, 1, 0.0, "", 100, "", "claude-code", "", 10.0, 20.0, false, nil, int64(0), false, true, []byte(`{}`)).
 		AddRow("ws-2", "Agent Two", "manager", 2, "provisioning", []byte("null"), "",
-			nil, 0, 1, 0.0, "", 0, "", "langgraph", "", 50.0, 60.0, false, nil, int64(0), false, true)
+			nil, 0, 1, 0.0, "", 0, "", "langgraph", "", 50.0, 60.0, false, nil, int64(0), false, true, []byte(`{}`))

 	mock.ExpectQuery("SELECT w.id, w.name").
 		WillReturnRows(rows)
@@ -1184,14 +1184,14 @@ func TestWorkspaceGet_CurrentTask(t *testing.T) {
 		"parent_id", "active_tasks", "max_concurrent_tasks", "last_error_rate", "last_sample_error",
 		"uptime_seconds", "current_task", "runtime", "workspace_dir", "x", "y", "collapsed",
 		"budget_limit", "monthly_spend",
-		"broadcast_enabled", "talk_to_user_enabled",
+		"broadcast_enabled", "talk_to_user_enabled", "compute",
 	}
 	mock.ExpectQuery("SELECT w.id, w.name").
 		WithArgs("dddddddd-0004-0000-0000-000000000000").
 		WillReturnRows(sqlmock.NewRows(columns).AddRow(
 			"dddddddd-0004-0000-0000-000000000000", "Task Worker", "worker", 1, "online", []byte("null"), "http://localhost:9000",
 			nil, 2, 1, 0.0, "", 300, "Analyzing document", "langgraph", "", 10.0, 20.0, false,
-			nil, int64(0), false, true,
+			nil, int64(0), false, true, []byte(`{}`),
 		))

 	w := httptest.NewRecorder()
@@ -6,18 +6,26 @@ import (
 	"testing"
 )

-// TestHandleA2ADispatchError_NativeSession_SkipsEnqueue validates capability
-// primitive #5: when the target workspace has declared
-// provides_native_session=True, a busy-shaped dispatch error MUST short-
-// circuit straight to 503 + Retry-After. The platform's a2a_queue is
-// skipped because the SDK owns its own queue/session state — double-
-// buffering would cause spurious dispatches when the SDK is still busy.
+// TestHandleA2ADispatchError_NativeSession_NowEnqueues validates the #1684
+// fix: native_session adapters used to short-circuit to 503-no-queue here,
+// on the assumption that the SDK owned an inbound queue. In practice the
+// common native_session SDKs (claude-agent-sdk, codex app-server, hermes)
+// don't — new turns arrive only via the same HTTP POST that returns busy.
+// So cron fires bounced 503 every tick until the SDK voluntarily yielded;
+// Reno Stars #1684 observed 12 consecutive `*/30` cron fires lost over 6h.
 //
-// Pin via sqlmock: we deliberately do NOT expect any INSERT INTO a2a_queue.
-// If a future refactor re-introduces enqueueing under native_session,
-// sqlmock fails the test on the unexpected query.
-func TestHandleA2ADispatchError_NativeSession_SkipsEnqueue(t *testing.T) {
-	setupTestDB(t)
+// Post-fix: native_session and non-native both enqueue. Drain timing is
+// gated by registry.go:Heartbeat (`payload.ActiveTasks < maxConcurrent`)
+// so the queued item only dispatches when the SDK reports spare capacity
+// — i.e. the next heartbeat after the in-flight turn returns.
+//
+// This test pins the new behavior: native_session capability DOES NOT
+// bypass EnqueueA2A. We expect the INSERT INTO a2a_queue query to fire,
+// here arranged to fail so we can observe the legacy 503 fallback (and
+// thereby confirm the INSERT was attempted; sqlmock fails the test if
+// the expected query never runs).
+func TestHandleA2ADispatchError_NativeSession_NowEnqueues(t *testing.T) {
+	mock := setupTestDB(t)
 	setupTestRedis(t)
 	handler := NewWorkspaceHandler(newTestBroadcaster(), nil, "http://localhost:8080", t.TempDir())

@@ -25,10 +33,15 @@ func TestHandleA2ADispatchError_NativeSession_SkipsEnqueue(t *testing.T) {
 	runtimeOverrides.SetCapabilities("ws-native", map[string]bool{"session": true})
 	defer runtimeOverrides.Reset()

-	// DeadlineExceeded triggers isUpstreamBusyError. Without the native
-	// gate, this would fire EnqueueA2A → INSERT INTO a2a_queue. With
-	// the gate, it short-circuits to 503. We expect ZERO queue queries;
-	// sqlmock's ExpectationsWereMet implicitly enforces that on teardown.
+	// We now EXPECT the INSERT to fire even with native_session=true. Make
+	// it fail so the handler falls through to the legacy 503 path — that
+	// lets us assert (1) enqueue was attempted, (2) the response on
+	// queue-failure does NOT carry native_session=true marker (that field
+	// was removed alongside the gate).
+	mock.ExpectQuery(`INSERT INTO a2a_queue`).
+		WithArgs("ws-native", nil, PriorityTask, "{}", "message/send", nil).
+		WillReturnError(errTestQueueUnavailable)
+
 	_, _, perr := handler.handleA2ADispatchError(
 		context.Background(), "ws-native", "", []byte("{}"), "message/send",
 		context.DeadlineExceeded, 1, false,
@@ -37,28 +50,27 @@ func TestHandleA2ADispatchError_NativeSession_SkipsEnqueue(t *testing.T) {
 		t.Fatal("expected proxy error, got nil")
 	}
 	if perr.Status != http.StatusServiceUnavailable {
-		t.Errorf("got status %d, want 503 (native_session bypasses queue but still 503s)", perr.Status)
+		t.Errorf("got status %d, want 503 (enqueue failed → legacy 503 fallback)", perr.Status)
 	}
 	if perr.Headers["Retry-After"] == "" {
-		t.Error("expected Retry-After header on native-session 503")
+		t.Error("expected Retry-After header on busy-503")
 	}
-	// Pin the marker so callers' adapters can distinguish this from a
-	// queue-failure 503: the body has native_session=true.
-	if got, _ := perr.Response["native_session"].(bool); !got {
-		t.Errorf("expected native_session=true in response body; got %+v", perr.Response)
+	// The native_session marker was removed from the response body — the
+	// platform queues both kinds now, callers no longer distinguish. Pin
+	// its absence so a future revert is caught.
+	if got, ok := perr.Response["native_session"].(bool); ok && got {
+		t.Errorf("native_session marker should be gone after #1684 fix; got %+v", perr.Response)
 	}
-	// And busy=true stays so existing busy-handling code paths still trigger.
 	if got, _ := perr.Response["busy"].(bool); !got {
-		t.Errorf("expected busy=true in response body; got %+v", perr.Response)
+		t.Errorf("expected busy=true; got %+v", perr.Response)
 	}
 }

-// TestHandleA2ADispatchError_NoNativeSession_StillEnqueues is the negative
-// pin: a workspace WITHOUT the capability flag falls through to the
-// existing EnqueueA2A path (and 503 if that fails). Same shape as
-// TestHandleA2ADispatchError_ContextDeadline; we duplicate it here so
-// the native_session gate change is bracketed by both positive and
-// negative tests in the same file.
+// TestHandleA2ADispatchError_NoNativeSession_StillEnqueues — non-native
+// behavior is unchanged: enqueue is attempted, fail-fallback to 503. This
+// negative pin guards against accidentally reverting the unification by
+// re-introducing a `if HasCapability(...)` gate that would short-circuit
+// the enqueue path.
 func TestHandleA2ADispatchError_NoNativeSession_StillEnqueues(t *testing.T) {
 	mock := setupTestDB(t)
 	setupTestRedis(t)
@@ -79,13 +91,11 @@ func TestHandleA2ADispatchError_NoNativeSession_StillEnqueues(t *testing.T) {
 	if perr == nil {
 		t.Fatal("expected proxy error, got nil")
 	}
-	// Queue insert failed → falls through to legacy 503 (without
-	// native_session marker).
 	if perr.Status != http.StatusServiceUnavailable {
 		t.Errorf("got status %d, want 503", perr.Status)
 	}
 	if got, _ := perr.Response["native_session"].(bool); got {
-		t.Errorf("non-native workspace should NOT carry native_session=true in response; got %+v", perr.Response)
+		t.Errorf("non-native workspace should NOT carry native_session=true; got %+v", perr.Response)
 	}
 }

@@ -15,13 +15,46 @@ import (
 	"github.com/Molecule-AI/molecule-monorepo/platform/internal/scheduler"
 )

+// ErrorResponse is returned for 4xx/5xx errors. (OpenAPI doc shape — used by swaggo.)
+type ErrorResponse struct {
+	Error string `json:"error"`
+}
+
+// StatusResponse is returned by mutating endpoints that only echo a status verb.
+type StatusResponse struct {
+	Status string `json:"status"`
+}
+
+// CreateScheduleResponse is returned by POST /workspaces/{id}/schedules.
+type CreateScheduleResponse struct {
+	ID        string    `json:"id"`
+	Status    string    `json:"status"`
+	NextRunAt time.Time `json:"next_run_at"`
+}
+
+// RunNowResponse is returned by POST /workspaces/{id}/schedules/{scheduleId}/run.
+type RunNowResponse struct {
+	Status      string `json:"status"`
+	WorkspaceID string `json:"workspace_id"`
+	Prompt      string `json:"prompt"`
+}
+
+// HistoryEntry is one row of /workspaces/{id}/schedules/{scheduleId}/history.
+type HistoryEntry struct {
+	Timestamp   time.Time       `json:"timestamp"`
+	DurationMs  *int            `json:"duration_ms"`
+	Status      *string         `json:"status"`
+	ErrorDetail string          `json:"error_detail"`
+	Request     json.RawMessage `json:"request" swaggertype:"object"`
+}
+
 type ScheduleHandler struct{}

 func NewScheduleHandler() *ScheduleHandler {
 	return &ScheduleHandler{}
 }

-type scheduleResponse struct {
+type ScheduleResponse struct {
 	ID          string     `json:"id"`
 	WorkspaceID string     `json:"workspace_id"`
 	Name        string     `json:"name"`
@@ -40,6 +73,15 @@ type scheduleResponse struct {
 }

 // List returns all schedules for a workspace.
+//
+//	@Summary	List schedules for a workspace
+//	@Tags		schedules
+//	@Produce	json
+//	@Param		id	path		string	true	"Workspace ID"
+//	@Success	200	{array}		ScheduleResponse
+//	@Failure	500	{object}	ErrorResponse
+//	@Router		/workspaces/{id}/schedules [get]
+//	@Security	BearerAuth && OrgSlugAuth
 func (h *ScheduleHandler) List(c *gin.Context) {
 	workspaceID := c.Param("id")
 	ctx := c.Request.Context()
@@ -58,9 +100,9 @@ func (h *ScheduleHandler) List(c *gin.Context) {
 	}
 	defer rows.Close()

-	schedules := make([]scheduleResponse, 0)
+	schedules := make([]ScheduleResponse, 0)
 	for rows.Next() {
-		var s scheduleResponse
+		var s ScheduleResponse
 		if err := rows.Scan(
 			&s.ID, &s.WorkspaceID, &s.Name, &s.CronExpr, &s.Timezone,
 			&s.Prompt, &s.Enabled, &s.LastRunAt, &s.NextRunAt, &s.RunCount,
@@ -78,7 +120,7 @@ func (h *ScheduleHandler) List(c *gin.Context) {
 	c.JSON(http.StatusOK, schedules)
 }

-type createScheduleRequest struct {
+type CreateScheduleRequest struct {
 	Name     string `json:"name"`
 	CronExpr string `json:"cron_expr" binding:"required"`
 	Timezone string `json:"timezone"`
@@ -87,11 +129,23 @@ type createScheduleRequest struct {
 }

 // Create adds a new schedule for a workspace.
+//
+//	@Summary	Create a schedule
+//	@Tags		schedules
+//	@Accept		json
+//	@Produce	json
+//	@Param		id		path		string					true	"Workspace ID"
+//	@Param		body	body		CreateScheduleRequest	true	"Schedule fields"
+//	@Success	201		{object}	CreateScheduleResponse
+//	@Failure	400		{object}	ErrorResponse
+//	@Failure	500		{object}	ErrorResponse
+//	@Router		/workspaces/{id}/schedules [post]
+//	@Security	BearerAuth && OrgSlugAuth
 func (h *ScheduleHandler) Create(c *gin.Context) {
 	workspaceID := c.Param("id")
 	ctx := c.Request.Context()

-	var body createScheduleRequest
+	var body CreateScheduleRequest
 	if err := c.ShouldBindJSON(&body); err != nil {
 		c.JSON(http.StatusBadRequest, gin.H{"error": "cron_expr and prompt are required"})
 		return
@@ -145,7 +199,7 @@ func (h *ScheduleHandler) Create(c *gin.Context) {
 	})
 }

-type updateScheduleRequest struct {
+type UpdateScheduleRequest struct {
 	Name     *string `json:"name"`
 	CronExpr *string `json:"cron_expr"`
 	Timezone *string `json:"timezone"`
@@ -155,12 +209,26 @@ type updateScheduleRequest struct {

 // Update modifies a schedule. Uses a fixed UPDATE with COALESCE so only
 // provided fields are changed — no dynamic SQL construction.
+//
+//	@Summary	Update a schedule
+//	@Tags		schedules
+//	@Accept		json
+//	@Produce	json
+//	@Param		id			path		string					true	"Workspace ID"
+//	@Param		scheduleId	path		string					true	"Schedule ID"
+//	@Param		body		body		UpdateScheduleRequest	true	"Partial schedule fields (only provided keys are updated)"
+//	@Success	200			{object}	ScheduleResponse
+//	@Failure	400			{object}	ErrorResponse
+//	@Failure	404			{object}	ErrorResponse
+//	@Failure	500			{object}	ErrorResponse
+//	@Router		/workspaces/{id}/schedules/{scheduleId} [patch]
+//	@Security	BearerAuth && OrgSlugAuth
 func (h *ScheduleHandler) Update(c *gin.Context) {
 	scheduleID := c.Param("scheduleId")
 	workspaceID := c.Param("id") // #113: bind to owning workspace to prevent IDOR
 	ctx := c.Request.Context()

-	var body updateScheduleRequest
+	var body UpdateScheduleRequest
 	if err := c.ShouldBindJSON(&body); err != nil {
 		c.JSON(http.StatusBadRequest, gin.H{"error": "invalid JSON"})
 		return
@@ -230,6 +298,17 @@ func (h *ScheduleHandler) Update(c *gin.Context) {
 }

 // Delete removes a schedule.
+//
+//	@Summary	Delete a schedule
+//	@Tags		schedules
+//	@Produce	json
+//	@Param		id			path		string	true	"Workspace ID"
+//	@Param		scheduleId	path		string	true	"Schedule ID"
+//	@Success	200			{object}	StatusResponse
+//	@Failure	404			{object}	ErrorResponse
+//	@Failure	500			{object}	ErrorResponse
+//	@Router		/workspaces/{id}/schedules/{scheduleId} [delete]
+//	@Security	BearerAuth && OrgSlugAuth
 func (h *ScheduleHandler) Delete(c *gin.Context) {
 	scheduleID := c.Param("scheduleId")
 	workspaceID := c.Param("id") // #113: bind to owning workspace to prevent IDOR
@@ -252,6 +331,17 @@ func (h *ScheduleHandler) Delete(c *gin.Context) {
 }

 // RunNow manually fires a schedule immediately.
+//
+//	@Summary	Fire a schedule manually
+//	@Tags		schedules
+//	@Produce	json
+//	@Param		id			path		string	true	"Workspace ID"
+//	@Param		scheduleId	path		string	true	"Schedule ID"
+//	@Success	200			{object}	RunNowResponse
+//	@Failure	404			{object}	ErrorResponse
+//	@Failure	500			{object}	ErrorResponse
+//	@Router		/workspaces/{id}/schedules/{scheduleId}/run [post]
+//	@Security	BearerAuth && OrgSlugAuth
 func (h *ScheduleHandler) RunNow(c *gin.Context) {
 	scheduleID := c.Param("scheduleId")
 	workspaceID := c.Param("id")
@@ -282,6 +372,16 @@ func (h *ScheduleHandler) RunNow(c *gin.Context) {
 }

 // History returns recent runs for a schedule from activity_logs.
+//
+//	@Summary	Get past runs of a schedule
+//	@Tags		schedules
+//	@Produce	json
+//	@Param		id			path		string	true	"Workspace ID"
+//	@Param		scheduleId	path		string	true	"Schedule ID"
+//	@Success	200			{array}		HistoryEntry
+//	@Failure	500			{object}	ErrorResponse
+//	@Router		/workspaces/{id}/schedules/{scheduleId}/history [get]
+//	@Security	BearerAuth && OrgSlugAuth
 func (h *ScheduleHandler) History(c *gin.Context) {
 	scheduleID := c.Param("scheduleId")
 	workspaceID := c.Param("id")
@@ -307,17 +407,9 @@ func (h *ScheduleHandler) History(c *gin.Context) {
 	}
 	defer rows.Close()

-	type historyEntry struct {
-		Timestamp   time.Time       `json:"timestamp"`
-		DurationMs  *int            `json:"duration_ms"`
-		Status      *string         `json:"status"`
-		ErrorDetail string          `json:"error_detail"`
-		Request     json.RawMessage `json:"request"`
-	}
-
-	entries := make([]historyEntry, 0)
+	entries := make([]HistoryEntry, 0)
 	for rows.Next() {
-		var e historyEntry
+		var e HistoryEntry
 		var reqStr string
 		if err := rows.Scan(&e.Timestamp, &e.DurationMs, &e.Status, &e.ErrorDetail, &reqStr); err != nil {
 			continue
@@ -329,11 +421,11 @@ func (h *ScheduleHandler) History(c *gin.Context) {
 	c.JSON(http.StatusOK, entries)
 }

-// scheduleHealthResponse is the read-only health view of a schedule.
+// ScheduleHealthResponse is the read-only health view of a schedule.
 // It deliberately omits prompt and cron_expr so sensitive task content is
 // never exposed to peer workspaces — only execution-state fields needed to
 // detect silent cron failures are returned (issue #249).
-type scheduleHealthResponse struct {
+type ScheduleHealthResponse struct {
 	ID         string     `json:"id"`
 	Name       string     `json:"name"`
 	Enabled    bool       `json:"enabled"`
@@ -402,9 +494,9 @@ func (h *ScheduleHandler) Health(c *gin.Context) {
 	}
 	defer rows.Close()

-	schedules := make([]scheduleHealthResponse, 0)
+	schedules := make([]ScheduleHealthResponse, 0)
 	for rows.Next() {
-		var s scheduleHealthResponse
+		var s ScheduleHealthResponse
 		if err := rows.Scan(
 			&s.ID, &s.Name, &s.Enabled, &s.LastRunAt, &s.NextRunAt,
 			&s.RunCount, &s.LastStatus, &s.LastError,
@@ -234,7 +234,7 @@ func TestScheduleHealth_SelfCall_Allowed(t *testing.T) {
 		t.Fatalf("expected 200 for self-call, got %d: %s", w.Code, w.Body.String())
 	}

-	var resp []scheduleHealthResponse
+	var resp []ScheduleHealthResponse
 	if err := json.Unmarshal(w.Body.Bytes(), &resp); err != nil {
 		t.Fatalf("failed to parse response: %v", err)
 	}
@@ -284,7 +284,7 @@ func TestScheduleHealth_CanCommunicatePeer_LegacyNoToken(t *testing.T) {
 		t.Fatalf("expected 200 for peer with no tokens, got %d: %s", w.Code, w.Body.String())
 	}

-	var resp []scheduleHealthResponse
+	var resp []ScheduleHealthResponse
 	if err := json.Unmarshal(w.Body.Bytes(), &resp); err != nil {
 		t.Fatalf("failed to parse response: %v", err)
 	}
@@ -348,6 +348,10 @@ func (h *WorkspaceHandler) Create(c *gin.Context) {
 		c.JSON(http.StatusBadRequest, gin.H{"error": "invalid workspace access"})
 		return
 	}
+	if err := validateWorkspaceCompute(payload.Compute); err != nil {
+		c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
+		return
+	}

 	// Begin a transaction so the workspace row and any initial secrets are
 	// committed atomically.  A secret-encrypt or DB error rolls back the
@@ -435,6 +439,24 @@ func (h *WorkspaceHandler) Create(c *gin.Context) {
 		payload.Name = persistedName
 	}

+	if !workspaceComputeIsZero(payload.Compute) {
+		computeJSON, encErr := workspaceComputeJSON(payload.Compute)
+		if encErr != nil {
+			tx.Rollback() //nolint:errcheck
+			log.Printf("Create workspace %s: failed to encode compute config: %v", id, encErr)
+			c.JSON(http.StatusInternalServerError, gin.H{"error": "failed to encode compute config"})
+			return
+		}
+		if _, dbErr := tx.ExecContext(ctx,
+			`UPDATE workspaces SET compute = $2::jsonb, updated_at = now() WHERE id = $1`,
+			id, computeJSON); dbErr != nil {
+			tx.Rollback() //nolint:errcheck
+			log.Printf("Create workspace %s: failed to persist compute config: %v", id, dbErr)
+			c.JSON(http.StatusInternalServerError, gin.H{"error": "failed to save compute config"})
+			return
+		}
+	}
+
 	// Persist initial secrets from the create payload (inside same transaction).
 	// nil/empty map is a no-op.  Any failure rolls back the workspace insert
 	// so we never have a workspace row without its intended secrets.
@@ -679,6 +701,7 @@ func scanWorkspaceRow(rows interface {
 	Scan(dest ...interface{}) error
 }) (map[string]interface{}, error) {
 	var id, name, role, status, url, sampleError, currentTask, runtime, workspaceDir string
+	var computeRaw []byte
 	var tier, activeTasks, maxConcurrentTasks, uptimeSeconds int
 	var errorRate, x, y float64
 	var collapsed, broadcastEnabled, talkToUserEnabled bool
@@ -690,7 +713,7 @@ func scanWorkspaceRow(rows interface {
 	err := rows.Scan(&id, &name, &role, &tier, &status, &agentCard, &url,
 		&parentID, &activeTasks, &maxConcurrentTasks, &errorRate, &sampleError, &uptimeSeconds,
 		&currentTask, &runtime, &workspaceDir, &x, &y, &collapsed,
-		&budgetLimit, &monthlySpend, &broadcastEnabled, &talkToUserEnabled)
+		&budgetLimit, &monthlySpend, &broadcastEnabled, &talkToUserEnabled, &computeRaw)
 	if err != nil {
 		return nil, err
 	}
@@ -717,6 +740,11 @@ func scanWorkspaceRow(rows interface {
 		"broadcast_enabled":    broadcastEnabled,
 		"talk_to_user_enabled": talkToUserEnabled,
 	}
+	if len(computeRaw) > 0 && string(computeRaw) != "null" {
+		ws["compute"] = json.RawMessage(computeRaw)
+	} else {
+		ws["compute"] = json.RawMessage(`{}`)
+	}

 	// budget_limit: nil when no limit set, int64 otherwise
 	if budgetLimit.Valid {
@@ -752,7 +780,8 @@ const workspaceListQuery = `
 		   COALESCE(w.workspace_dir, ''),
 		   COALESCE(cl.x, 0), COALESCE(cl.y, 0), COALESCE(cl.collapsed, false),
 		   w.budget_limit, COALESCE(w.monthly_spend, 0),
-		   w.broadcast_enabled, w.talk_to_user_enabled
+		   w.broadcast_enabled, w.talk_to_user_enabled,
+		   COALESCE(w.compute, '{}'::jsonb)
 	FROM workspaces w
 	LEFT JOIN canvas_layouts cl ON cl.workspace_id = w.id
 	WHERE w.status != 'removed'
@@ -813,7 +842,8 @@ func (h *WorkspaceHandler) Get(c *gin.Context) {
 			   COALESCE(w.workspace_dir, ''),
 			   COALESCE(cl.x, 0), COALESCE(cl.y, 0), COALESCE(cl.collapsed, false),
 			   w.budget_limit, COALESCE(w.monthly_spend, 0),
-			   w.broadcast_enabled, w.talk_to_user_enabled
+			   w.broadcast_enabled, w.talk_to_user_enabled,
+			   COALESCE(w.compute, '{}'::jsonb)
 		FROM workspaces w
 		LEFT JOIN canvas_layouts cl ON cl.workspace_id = w.id
 		WHERE w.id = $1
@@ -33,7 +33,7 @@ var wsColumns = []string{
 	"parent_id", "active_tasks", "max_concurrent_tasks", "last_error_rate", "last_sample_error",
 	"uptime_seconds", "current_task", "runtime", "workspace_dir", "x", "y", "collapsed",
 	"budget_limit", "monthly_spend",
-	"broadcast_enabled", "talk_to_user_enabled",
+	"broadcast_enabled", "talk_to_user_enabled", "compute",
 }

 // ==================== GET — financial fields stripped from open endpoint ====================
@@ -56,7 +56,8 @@ func TestWorkspaceBudget_Get_NilLimit(t *testing.T) {
 				nil,   // budget_limit NULL
 				0,     // monthly_spend 0
 				false, // broadcast_enabled
-				true)) // talk_to_user_enabled
+				true,  // talk_to_user_enabled
+				[]byte(`{}`)))

 	w := httptest.NewRecorder()
 	c, _ := gin.CreateTestContext(w)
@@ -100,7 +101,8 @@ func TestWorkspaceBudget_Get_WithLimit(t *testing.T) {
 				0.0, 0.0, false,
 				int64(500),  // budget_limit = $5.00 in DB
 				int64(123),  // monthly_spend = $1.23 in DB
-				false, true)) // broadcast_enabled, talk_to_user_enabled
+				false, true, // broadcast_enabled, talk_to_user_enabled
+				[]byte(`{}`)))

 	w := httptest.NewRecorder()
 	c, _ := gin.CreateTestContext(w)
@@ -145,18 +147,18 @@ func TestWorkspaceBudget_Create_WithLimit(t *testing.T) {
 	mock.ExpectBegin()
 	mock.ExpectExec("INSERT INTO workspaces").
 		WithArgs(
-			sqlmock.AnyArg(), // id
-			"Budgeted Agent", // name
-			nil,              // role
-			3,                // tier (default, workspace.go create-handler)
-			"langgraph",      // runtime
-			sqlmock.AnyArg(), // awareness_namespace
-			(*string)(nil),   // parent_id
-			nil,              // workspace_dir
-			"none",           // workspace_access
-			&budgetVal,       // budget_limit ($10)
+			sqlmock.AnyArg(),                 // id
+			"Budgeted Agent",                 // name
+			nil,                              // role
+			3,                                // tier (default, workspace.go create-handler)
+			"langgraph",                      // runtime
+			sqlmock.AnyArg(),                 // awareness_namespace
+			(*string)(nil),                   // parent_id
+			nil,                              // workspace_dir
+			"none",                           // workspace_access
+			&budgetVal,                       // budget_limit ($10)
 			models.DefaultMaxConcurrentTasks, // max_concurrent_tasks default
-			"push",           // delivery_mode default (#2339)
+			"push",                           // delivery_mode default (#2339)
 		).
 		WillReturnResult(sqlmock.NewResult(0, 1))
 	mock.ExpectCommit()
@@ -0,0 +1,208 @@
+package handlers
+
+import (
+	"context"
+	"database/sql"
+	"encoding/json"
+	"fmt"
+	"log"
+
+	"github.com/Molecule-AI/molecule-monorepo/platform/internal/db"
+	"github.com/Molecule-AI/molecule-monorepo/platform/internal/models"
+	"github.com/gin-gonic/gin"
+)
+
+const (
+	workspaceComputeDiskFloorGB   = 30
+	workspaceComputeDiskCeilingGB = 500
+)
+
+type workspaceDisplayResponse struct {
+	Available bool   `json:"available"`
+	Reason    string `json:"reason,omitempty"`
+	Mode      string `json:"mode,omitempty"`
+	Protocol  string `json:"protocol,omitempty"`
+	Width     int    `json:"width,omitempty"`
+	Height    int    `json:"height,omitempty"`
+	Status    string `json:"status,omitempty"`
+}
+
+var workspaceComputeInstanceAllowlist = map[string]struct{}{
+	"t3.medium":  {},
+	"t3.large":   {},
+	"t3.xlarge":  {},
+	"t3.2xlarge": {},
+	"m6i.large":  {},
+	"m6i.xlarge": {},
+	"c6i.xlarge": {},
+}
+
+func validateWorkspaceCompute(compute models.WorkspaceCompute) error {
+	if compute.InstanceType != "" {
+		if _, ok := workspaceComputeInstanceAllowlist[compute.InstanceType]; !ok {
+			return fmt.Errorf("unsupported compute.instance_type")
+		}
+	}
+	if compute.Volume.RootGB != 0 {
+		if compute.Volume.RootGB < workspaceComputeDiskFloorGB || compute.Volume.RootGB > workspaceComputeDiskCeilingGB {
+			return fmt.Errorf("compute.volume.root_gb must be between %d and %d", workspaceComputeDiskFloorGB, workspaceComputeDiskCeilingGB)
+		}
+	}
+	switch compute.Display.Mode {
+	case "", "none", "desktop-control", "gpu-desktop-control":
+	default:
+		return fmt.Errorf("unsupported compute.display.mode")
+	}
+	switch compute.Display.Protocol {
+	case "", "dcv":
+	default:
+		return fmt.Errorf("unsupported compute.display.protocol")
+	}
+	if compute.Display.Width < 0 || compute.Display.Height < 0 {
+		return fmt.Errorf("compute.display width/height must be non-negative")
+	}
+	return nil
+}
+
+func validateWorkspaceDisplayConfig(display models.WorkspaceComputeDisplay) error {
+	switch display.Mode {
+	case "", "none", "desktop-control", "gpu-desktop-control":
+	default:
+		return fmt.Errorf("unsupported compute.display.mode")
+	}
+	switch display.Protocol {
+	case "", "dcv":
+	default:
+		return fmt.Errorf("unsupported compute.display.protocol")
+	}
+	if display.Width < 0 || display.Height < 0 {
+		return fmt.Errorf("compute.display width/height must be non-negative")
+	}
+	return nil
+}
+
+func workspaceComputeIsZero(compute models.WorkspaceCompute) bool {
+	return compute.InstanceType == "" &&
+		compute.Volume.RootGB == 0 &&
+		compute.Display.Mode == "" &&
+		compute.Display.Width == 0 &&
+		compute.Display.Height == 0 &&
+		compute.Display.Protocol == ""
+}
+
+func workspaceComputeJSON(compute models.WorkspaceCompute) (string, error) {
+	if workspaceComputeIsZero(compute) {
+		return "{}", nil
+	}
+	out := map[string]interface{}{}
+	if compute.InstanceType != "" {
+		out["instance_type"] = compute.InstanceType
+	}
+	if compute.Volume.RootGB != 0 {
+		out["volume"] = map[string]interface{}{"root_gb": compute.Volume.RootGB}
+	}
+	display := map[string]interface{}{}
+	if compute.Display.Mode != "" {
+		display["mode"] = compute.Display.Mode
+	}
+	if compute.Display.Width != 0 {
+		display["width"] = compute.Display.Width
+	}
+	if compute.Display.Height != 0 {
+		display["height"] = compute.Display.Height
+	}
+	if compute.Display.Protocol != "" {
+		display["protocol"] = compute.Display.Protocol
+	}
+	if len(display) > 0 {
+		out["display"] = display
+	}
+	b, err := json.Marshal(out)
+	if err != nil {
+		return "", err
+	}
+	return string(b), nil
+}
+
+func withStoredCompute(ctx context.Context, workspaceID string, payload models.CreateWorkspacePayload) models.CreateWorkspacePayload {
+	if !workspaceComputeIsZero(payload.Compute) || db.DB == nil {
+		return payload
+	}
+	var raw string
+	err := db.DB.QueryRowContext(ctx,
+		`SELECT COALESCE(compute, '{}'::jsonb) FROM workspaces WHERE id = $1`,
+		workspaceID,
+	).Scan(&raw)
+	if err != nil {
+		if err != sql.ErrNoRows {
+			log.Printf("withStoredCompute: load compute for %s failed: %v", workspaceID, err)
+		}
+		return payload
+	}
+	if raw == "" || raw == "{}" {
+		return payload
+	}
+	var compute models.WorkspaceCompute
+	if err := json.Unmarshal([]byte(raw), &compute); err != nil {
+		log.Printf("withStoredCompute: invalid compute JSON for %s: %v", workspaceID, err)
+		return payload
+	}
+	if err := validateWorkspaceCompute(compute); err != nil {
+		log.Printf("withStoredCompute: stored compute for %s failed validation: %v", workspaceID, err)
+		return payload
+	}
+	payload.Compute = compute
+	return payload
+}
+
+// Display handles GET /workspaces/:id/display.
+//
+// Phase 1 only exposes the product contract and the non-display unavailable
+// state. Future desktop-control work will replace the display-enabled branch
+// with short-lived proxied DCV session details.
+func (h *WorkspaceHandler) Display(c *gin.Context) {
+	workspaceID := c.Param("id")
+	var raw string
+	err := db.DB.QueryRowContext(c.Request.Context(),
+		`SELECT COALESCE(compute, '{}'::jsonb) FROM workspaces WHERE id = $1`,
+		workspaceID,
+	).Scan(&raw)
+	if err != nil {
+		if err == sql.ErrNoRows {
+			c.JSON(404, gin.H{"error": "workspace not found"})
+			return
+		}
+		log.Printf("Display: load compute for %s failed: %v", workspaceID, err)
+		c.JSON(500, gin.H{"error": "failed to load display config"})
+		return
+	}
+	var compute models.WorkspaceCompute
+	if raw != "" && raw != "{}" {
+		if err := json.Unmarshal([]byte(raw), &compute); err != nil {
+			log.Printf("Display: invalid compute JSON for %s: %v", workspaceID, err)
+			c.JSON(500, gin.H{"error": "invalid display config"})
+			return
+		}
+		if err := validateWorkspaceDisplayConfig(compute.Display); err != nil {
+			log.Printf("Display: invalid stored compute for %s: %v", workspaceID, err)
+			c.JSON(500, gin.H{"error": "invalid display config"})
+			return
+		}
+	}
+	if compute.Display.Mode == "" || compute.Display.Mode == "none" {
+		c.JSON(200, workspaceDisplayResponse{
+			Available: false,
+			Reason:    "display_not_enabled",
+		})
+		return
+	}
+	c.JSON(200, workspaceDisplayResponse{
+		Available: false,
+		Reason:    "display_session_unavailable",
+		Mode:      compute.Display.Mode,
+		Protocol:  compute.Display.Protocol,
+		Width:     compute.Display.Width,
+		Height:    compute.Display.Height,
+		Status:    "not_configured",
+	})
+}
@@ -0,0 +1,318 @@
+package handlers
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"github.com/DATA-DOG/go-sqlmock"
+	"github.com/Molecule-AI/molecule-monorepo/platform/internal/models"
+	"github.com/gin-gonic/gin"
+)
+
+func TestValidateWorkspaceCompute_AcceptsPhase1SizingAndDisplayNone(t *testing.T) {
+	compute := models.WorkspaceCompute{
+		InstanceType: "m6i.xlarge",
+		Volume:       models.WorkspaceComputeVolume{RootGB: 100},
+		Display:      models.WorkspaceComputeDisplay{Mode: "none"},
+	}
+
+	if err := validateWorkspaceCompute(compute); err != nil {
+		t.Fatalf("validateWorkspaceCompute returned error for valid compute: %v", err)
+	}
+}
+
+func TestValidateWorkspaceCompute_RejectsUnknownInstanceType(t *testing.T) {
+	compute := models.WorkspaceCompute{InstanceType: "p4d.24xlarge"}
+
+	if err := validateWorkspaceCompute(compute); err == nil {
+		t.Fatal("validateWorkspaceCompute accepted unsupported instance type")
+	}
+}
+
+func TestValidateWorkspaceCompute_RejectsOutOfRangeRootVolume(t *testing.T) {
+	for _, rootGB := range []int{29, 501} {
+		compute := models.WorkspaceCompute{Volume: models.WorkspaceComputeVolume{RootGB: rootGB}}
+		if err := validateWorkspaceCompute(compute); err == nil {
+			t.Fatalf("validateWorkspaceCompute accepted root_gb=%d", rootGB)
+		}
+	}
+}
+
+func TestWorkspaceComputeJSON_OmitsEmptyNestedSections(t *testing.T) {
+	got, err := workspaceComputeJSON(models.WorkspaceCompute{
+		InstanceType: "m6i.xlarge",
+		Volume:       models.WorkspaceComputeVolume{RootGB: 100},
+	})
+	if err != nil {
+		t.Fatalf("workspaceComputeJSON returned error: %v", err)
+	}
+
+	if strings.Contains(got, `"display"`) {
+		t.Fatalf("workspaceComputeJSON included empty display section: %s", got)
+	}
+	if got != `{"instance_type":"m6i.xlarge","volume":{"root_gb":100}}` {
+		t.Fatalf("workspaceComputeJSON = %s", got)
+	}
+}
+
+func TestWorkspaceCreate_WithCompute_PersistsComputeJSON(t *testing.T) {
+	mock := setupTestDB(t)
+	setupTestRedis(t)
+	broadcaster := newTestBroadcaster()
+	handler := NewWorkspaceHandler(broadcaster, nil, "http://localhost:8080", t.TempDir())
+
+	mock.ExpectBegin()
+	mock.ExpectExec("INSERT INTO workspaces").
+		WillReturnResult(sqlmock.NewResult(0, 1))
+	mock.ExpectExec(`UPDATE workspaces SET compute = \$2::jsonb`).
+		WithArgs(sqlmock.AnyArg(), sqlmock.AnyArg()).
+		WillReturnResult(sqlmock.NewResult(0, 1))
+	mock.ExpectCommit()
+	mock.ExpectExec("INSERT INTO canvas_layouts").
+		WillReturnResult(sqlmock.NewResult(0, 1))
+
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	body := `{
+		"name":"Sized Agent",
+		"external":true,
+		"runtime":"external",
+		"compute":{
+			"instance_type":"m6i.xlarge",
+			"volume":{"root_gb":100},
+			"display":{"mode":"none"}
+		}
+	}`
+	c.Request = httptest.NewRequest("POST", "/workspaces", bytes.NewBufferString(body))
+	c.Request.Header.Set("Content-Type", "application/json")
+
+	handler.Create(c)
+
+	if w.Code != http.StatusCreated {
+		t.Fatalf("expected status 201, got %d: %s", w.Code, w.Body.String())
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Errorf("unmet sqlmock expectations: %v", err)
+	}
+}
+
+func TestWorkspaceCreate_WithInvalidCompute_ReturnsBadRequest(t *testing.T) {
+	setupTestDB(t)
+	setupTestRedis(t)
+	handler := NewWorkspaceHandler(newTestBroadcaster(), nil, "http://localhost:8080", t.TempDir())
+
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	body := `{
+		"name":"Oversized Agent",
+		"compute":{"instance_type":"p4d.24xlarge"}
+	}`
+	c.Request = httptest.NewRequest("POST", "/workspaces", bytes.NewBufferString(body))
+	c.Request.Header.Set("Content-Type", "application/json")
+
+	handler.Create(c)
+
+	if w.Code != http.StatusBadRequest {
+		t.Fatalf("expected status 400, got %d: %s", w.Code, w.Body.String())
+	}
+}
+
+func TestBuildProvisionerConfig_CopiesComputeSizingFromPayload(t *testing.T) {
+	mock := setupTestDB(t)
+	mock.ExpectQuery(`SELECT COALESCE\(workspace_dir`).
+		WithArgs("ws-compute").
+		WillReturnRows(sqlmock.NewRows([]string{"workspace_dir", "workspace_access"}).AddRow("", "none"))
+
+	handler := NewWorkspaceHandler(newTestBroadcaster(), nil, "http://localhost:8080", t.TempDir())
+	cfg := handler.buildProvisionerConfig(
+		context.Background(),
+		"ws-compute",
+		"",
+		nil,
+		models.CreateWorkspacePayload{
+			Tier:    4,
+			Runtime: "claude-code",
+			Compute: models.WorkspaceCompute{
+				InstanceType: "m6i.xlarge",
+				Volume:       models.WorkspaceComputeVolume{RootGB: 100},
+			},
+		},
+		nil,
+		t.TempDir(),
+		"workspace:ws-compute",
+	)
+
+	if cfg.InstanceType != "m6i.xlarge" {
+		t.Errorf("cfg.InstanceType = %q, want m6i.xlarge", cfg.InstanceType)
+	}
+	if cfg.DiskGB != 100 {
+		t.Errorf("cfg.DiskGB = %d, want 100", cfg.DiskGB)
+	}
+}
+
+func TestWithStoredCompute_LoadsComputeForRestartPayloads(t *testing.T) {
+	mock := setupTestDB(t)
+	mock.ExpectQuery(`SELECT COALESCE\(compute, '\{\}'::jsonb\) FROM workspaces WHERE id = \$1`).
+		WithArgs("ws-restart-compute").
+		WillReturnRows(sqlmock.NewRows([]string{"compute"}).AddRow(`{"instance_type":"m6i.xlarge","volume":{"root_gb":100}}`))
+
+	payload := models.CreateWorkspacePayload{Name: "Restart Me", Tier: 4, Runtime: "claude-code"}
+	got := withStoredCompute(context.Background(), "ws-restart-compute", payload)
+
+	if got.Compute.InstanceType != "m6i.xlarge" {
+		t.Errorf("stored compute instance_type = %q, want m6i.xlarge", got.Compute.InstanceType)
+	}
+	if got.Compute.Volume.RootGB != 100 {
+		t.Errorf("stored compute root_gb = %d, want 100", got.Compute.Volume.RootGB)
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Errorf("unmet sqlmock expectations: %v", err)
+	}
+}
+
+func TestWorkspaceDisplay_NonDisplayWorkspaceReturnsUnavailable(t *testing.T) {
+	mock := setupTestDB(t)
+	setupTestRedis(t)
+	handler := NewWorkspaceHandler(newTestBroadcaster(), nil, "http://localhost:8080", t.TempDir())
+
+	mock.ExpectQuery(`SELECT COALESCE\(compute, '\{\}'::jsonb\) FROM workspaces WHERE id = \$1`).
+		WithArgs("ws-no-display").
+		WillReturnRows(sqlmock.NewRows([]string{"compute"}).AddRow(`{}`))
+
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Params = gin.Params{{Key: "id", Value: "ws-no-display"}}
+	c.Request = httptest.NewRequest("GET", "/workspaces/ws-no-display/display", nil)
+
+	handler.Display(c)
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("expected status 200, got %d: %s", w.Code, w.Body.String())
+	}
+	var resp map[string]interface{}
+	if err := json.Unmarshal(w.Body.Bytes(), &resp); err != nil {
+		t.Fatalf("failed to parse display response: %v", err)
+	}
+	if resp["available"] != false {
+		t.Fatalf("available = %v, want false", resp["available"])
+	}
+	if resp["reason"] != "display_not_enabled" {
+		t.Fatalf("reason = %v, want display_not_enabled", resp["reason"])
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Errorf("unmet sqlmock expectations: %v", err)
+	}
+}
+
+func TestWorkspaceDisplay_DisplayConfiguredReturnsSessionUnavailableContract(t *testing.T) {
+	mock := setupTestDB(t)
+	setupTestRedis(t)
+	handler := NewWorkspaceHandler(newTestBroadcaster(), nil, "http://localhost:8080", t.TempDir())
+
+	mock.ExpectQuery(`SELECT COALESCE\(compute, '\{\}'::jsonb\) FROM workspaces WHERE id = \$1`).
+		WithArgs("ws-display").
+		WillReturnRows(sqlmock.NewRows([]string{"compute"}).AddRow(`{"display":{"mode":"desktop-control","protocol":"dcv","width":1920,"height":1080}}`))
+
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Params = gin.Params{{Key: "id", Value: "ws-display"}}
+	c.Request = httptest.NewRequest("GET", "/workspaces/ws-display/display", nil)
+
+	handler.Display(c)
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("expected status 200, got %d: %s", w.Code, w.Body.String())
+	}
+	var resp map[string]interface{}
+	if err := json.Unmarshal(w.Body.Bytes(), &resp); err != nil {
+		t.Fatalf("failed to parse display response: %v", err)
+	}
+	if resp["available"] != false {
+		t.Fatalf("available = %v, want false", resp["available"])
+	}
+	if resp["reason"] != "display_session_unavailable" {
+		t.Fatalf("reason = %v, want display_session_unavailable", resp["reason"])
+	}
+	if resp["status"] != "not_configured" {
+		t.Fatalf("status = %v, want not_configured", resp["status"])
+	}
+	if resp["mode"] != "desktop-control" || resp["protocol"] != "dcv" {
+		t.Fatalf("mode/protocol = %v/%v, want desktop-control/dcv", resp["mode"], resp["protocol"])
+	}
+	if resp["width"] != float64(1920) || resp["height"] != float64(1080) {
+		t.Fatalf("width/height = %v/%v, want 1920/1080", resp["width"], resp["height"])
+	}
+	if _, ok := resp["url"]; ok {
+		t.Fatalf("display response exposed url before session infra exists: %v", resp["url"])
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Errorf("unmet sqlmock expectations: %v", err)
+	}
+}
+
+func TestWorkspaceDisplay_IgnoresUnrelatedStoredComputeSizingDrift(t *testing.T) {
+	mock := setupTestDB(t)
+	setupTestRedis(t)
+	handler := NewWorkspaceHandler(newTestBroadcaster(), nil, "http://localhost:8080", t.TempDir())
+
+	mock.ExpectQuery(`SELECT COALESCE\(compute, '\{\}'::jsonb\) FROM workspaces WHERE id = \$1`).
+		WithArgs("ws-display-sizing-drift").
+		WillReturnRows(sqlmock.NewRows([]string{"compute"}).AddRow(`{"instance_type":"old.large","display":{"mode":"desktop-control","protocol":"dcv","width":1920,"height":1080}}`))
+
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Params = gin.Params{{Key: "id", Value: "ws-display-sizing-drift"}}
+	c.Request = httptest.NewRequest("GET", "/workspaces/ws-display-sizing-drift/display", nil)
+
+	handler.Display(c)
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("expected status 200, got %d: %s", w.Code, w.Body.String())
+	}
+	var resp map[string]interface{}
+	if err := json.Unmarshal(w.Body.Bytes(), &resp); err != nil {
+		t.Fatalf("failed to parse display response: %v", err)
+	}
+	if resp["reason"] != "display_session_unavailable" {
+		t.Fatalf("reason = %v, want display_session_unavailable", resp["reason"])
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Errorf("unmet sqlmock expectations: %v", err)
+	}
+}
+
+func TestWorkspaceDisplay_InvalidStoredDisplayConfigReturnsServerError(t *testing.T) {
+	mock := setupTestDB(t)
+	setupTestRedis(t)
+	handler := NewWorkspaceHandler(newTestBroadcaster(), nil, "http://localhost:8080", t.TempDir())
+
+	mock.ExpectQuery(`SELECT COALESCE\(compute, '\{\}'::jsonb\) FROM workspaces WHERE id = \$1`).
+		WithArgs("ws-invalid-display").
+		WillReturnRows(sqlmock.NewRows([]string{"compute"}).AddRow(`{"display":{"mode":"desktop-control","protocol":"vnc"}}`))
+
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Params = gin.Params{{Key: "id", Value: "ws-invalid-display"}}
+	c.Request = httptest.NewRequest("GET", "/workspaces/ws-invalid-display/display", nil)
+
+	handler.Display(c)
+
+	if w.Code != http.StatusInternalServerError {
+		t.Fatalf("expected status 500, got %d: %s", w.Code, w.Body.String())
+	}
+	var resp map[string]interface{}
+	if err := json.Unmarshal(w.Body.Bytes(), &resp); err != nil {
+		t.Fatalf("failed to parse display response: %v", err)
+	}
+	if resp["error"] != "invalid display config" {
+		t.Fatalf("error = %v, want invalid display config", resp["error"])
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Errorf("unmet sqlmock expectations: %v", err)
+	}
+}
@@ -0,0 +1,360 @@
+package handlers
+
+import (
+	"context"
+	"crypto/subtle"
+	"database/sql"
+	"encoding/json"
+	"fmt"
+	"log"
+	"net/http"
+	"os"
+	"time"
+
+	"github.com/Molecule-AI/molecule-monorepo/platform/internal/db"
+	"github.com/Molecule-AI/molecule-monorepo/platform/internal/models"
+	"github.com/Molecule-AI/molecule-monorepo/platform/internal/wsauth"
+	"github.com/gin-gonic/gin"
+)
+
+const (
+	displayControlDefaultTTLSeconds = 300
+	displayControlMinTTLSeconds     = 30
+	displayControlMaxTTLSeconds     = 3600
+)
+
+type workspaceDisplayControlResponse struct {
+	Controller   string    `json:"controller"`
+	ControlledBy string    `json:"controlled_by,omitempty"`
+	ExpiresAt    time.Time `json:"expires_at"`
+}
+
+type workspaceDisplayControlNoneResponse struct {
+	Controller string `json:"controller"`
+}
+
+type acquireDisplayControlRequest struct {
+	Controller string `json:"controller"`
+	TTLSeconds int    `json:"ttl_seconds"`
+}
+
+type releaseDisplayControlRequest struct {
+	Force bool `json:"force"`
+}
+
+// DisplayControl handles GET /workspaces/:id/display/control.
+func (h *WorkspaceHandler) DisplayControl(c *gin.Context) {
+	lock, found, err := h.loadActiveDisplayControl(c, c.Param("id"))
+	if err != nil {
+		log.Printf("DisplayControl: load lock for %s failed: %v", c.Param("id"), err)
+		c.JSON(http.StatusInternalServerError, gin.H{"error": "failed to load display control"})
+		return
+	}
+	if !found {
+		c.JSON(http.StatusOK, workspaceDisplayControlNoneResponse{Controller: "none"})
+		return
+	}
+	c.JSON(http.StatusOK, lock)
+}
+
+// AcquireDisplayControl handles POST /workspaces/:id/display/control/acquire.
+func (h *WorkspaceHandler) AcquireDisplayControl(c *gin.Context) {
+	var req acquireDisplayControlRequest
+	if c.Request.Body != nil && c.Request.ContentLength != 0 {
+		if err := c.ShouldBindJSON(&req); err != nil {
+			c.JSON(http.StatusBadRequest, gin.H{"error": "invalid display control request"})
+			return
+		}
+	}
+	if req.Controller == "" {
+		req.Controller = "user"
+	}
+	if req.Controller != "user" {
+		c.JSON(http.StatusBadRequest, gin.H{"error": "browser callers may only acquire user display control"})
+		return
+	}
+	if req.TTLSeconds == 0 {
+		req.TTLSeconds = displayControlDefaultTTLSeconds
+	}
+	if req.TTLSeconds < displayControlMinTTLSeconds || req.TTLSeconds > displayControlMaxTTLSeconds {
+		c.JSON(http.StatusBadRequest, gin.H{"error": "ttl_seconds must be between 30 and 3600"})
+		return
+	}
+	if ok := h.displayControlEnabled(c, c.Param("id")); !ok {
+		return
+	}
+
+	controlledBy, ok := displayControlActor(c)
+	if !ok {
+		c.JSON(http.StatusForbidden, gin.H{"error": "display control requires admin-token or org-token auth"})
+		return
+	}
+	workspaceID := c.Param("id")
+	startedAt := time.Now()
+	emitDisplayControlEvent(c.Request.Context(), "display.control.acquire.started", workspaceID, map[string]any{
+		"controller":    req.Controller,
+		"controlled_by": controlledBy,
+		"ttl_seconds":   req.TTLSeconds,
+	})
+	var lock workspaceDisplayControlResponse
+	err := db.DB.QueryRowContext(c.Request.Context(), `
+INSERT INTO workspace_display_control_locks
+    (workspace_id, controller, controlled_by, expires_at)
+VALUES
+    ($1, $2, $3, now() + ($4 * interval '1 second'))
+ON CONFLICT (workspace_id) DO UPDATE
+SET controller = EXCLUDED.controller,
+    controlled_by = EXCLUDED.controlled_by,
+    expires_at = EXCLUDED.expires_at,
+    updated_at = now()
+WHERE workspace_display_control_locks.expires_at <= now()
+   OR workspace_display_control_locks.controlled_by = EXCLUDED.controlled_by
+RETURNING controller, controlled_by, expires_at`,
+		workspaceID, req.Controller, controlledBy, req.TTLSeconds,
+	).Scan(&lock.Controller, &lock.ControlledBy, &lock.ExpiresAt)
+	if err == nil {
+		emitDisplayControlEvent(c.Request.Context(), "display.control.acquire.completed", workspaceID, map[string]any{
+			"controller":    lock.Controller,
+			"controlled_by": lock.ControlledBy,
+			"ttl_seconds":   req.TTLSeconds,
+			"duration_ms":   time.Since(startedAt).Milliseconds(),
+		})
+		c.JSON(http.StatusOK, lock)
+		return
+	}
+	if err == sql.ErrNoRows {
+		current, found, loadErr := h.loadActiveDisplayControl(c, workspaceID)
+		if loadErr != nil {
+			log.Printf("AcquireDisplayControl: load active lock for %s failed: %v", workspaceID, loadErr)
+			emitDisplayControlEvent(c.Request.Context(), "display.control.acquire.failed", workspaceID, map[string]any{
+				"controlled_by": controlledBy,
+				"duration_ms":   time.Since(startedAt).Milliseconds(),
+				"error":         loadErr.Error(),
+			})
+			c.JSON(http.StatusInternalServerError, gin.H{"error": "failed to load display control"})
+			return
+		}
+		emitDisplayControlEvent(c.Request.Context(), "display.control.acquire.failed", workspaceID, map[string]any{
+			"controlled_by": controlledBy,
+			"duration_ms":   time.Since(startedAt).Milliseconds(),
+			"error":         "display control already held",
+		})
+		if !found {
+			c.JSON(http.StatusConflict, gin.H{
+				"error":   "display control already held",
+				"current": workspaceDisplayControlNoneResponse{Controller: "none"},
+			})
+			return
+		}
+		c.JSON(http.StatusConflict, gin.H{
+			"error":   "display control already held",
+			"current": current,
+		})
+		return
+	}
+	log.Printf("AcquireDisplayControl: acquire lock for %s failed: %v", workspaceID, err)
+	emitDisplayControlEvent(c.Request.Context(), "display.control.acquire.failed", workspaceID, map[string]any{
+		"controlled_by": controlledBy,
+		"duration_ms":   time.Since(startedAt).Milliseconds(),
+		"error":         err.Error(),
+	})
+	c.JSON(http.StatusInternalServerError, gin.H{"error": "failed to acquire display control"})
+}
+
+// ReleaseDisplayControl handles POST /workspaces/:id/display/control/release.
+func (h *WorkspaceHandler) ReleaseDisplayControl(c *gin.Context) {
+	var req releaseDisplayControlRequest
+	if c.Request.Body != nil && c.Request.ContentLength != 0 {
+		if err := c.ShouldBindJSON(&req); err != nil {
+			c.JSON(http.StatusBadRequest, gin.H{"error": "invalid display control release request"})
+			return
+		}
+	}
+	if req.Force {
+		if !displayControlIsAdminToken(c) {
+			c.JSON(http.StatusForbidden, gin.H{"error": "force release requires admin-token auth"})
+			return
+		}
+	}
+
+	controlledBy, ok := displayControlActor(c)
+	if !ok {
+		c.JSON(http.StatusForbidden, gin.H{"error": "display control requires admin-token or org-token auth"})
+		return
+	}
+	workspaceID := c.Param("id")
+	startedAt := time.Now()
+	emitDisplayControlEvent(c.Request.Context(), "display.control.release.started", workspaceID, map[string]any{
+		"controlled_by": controlledBy,
+		"force":         req.Force,
+	})
+	query := `DELETE FROM workspace_display_control_locks WHERE workspace_id = $1 AND controlled_by = $2`
+	args := []interface{}{workspaceID, controlledBy}
+	if req.Force {
+		query = `DELETE FROM workspace_display_control_locks WHERE workspace_id = $1`
+		args = []interface{}{workspaceID}
+	}
+	result, err := db.DB.ExecContext(c.Request.Context(), query, args...)
+	if err != nil {
+		log.Printf("ReleaseDisplayControl: release lock for %s failed: %v", workspaceID, err)
+		emitDisplayControlEvent(c.Request.Context(), "display.control.release.failed", workspaceID, map[string]any{
+			"controlled_by": controlledBy,
+			"duration_ms":   time.Since(startedAt).Milliseconds(),
+			"error":         err.Error(),
+			"force":         req.Force,
+		})
+		c.JSON(http.StatusInternalServerError, gin.H{"error": "failed to release display control"})
+		return
+	}
+	rowsAffected, err := result.RowsAffected()
+	if err != nil {
+		log.Printf("ReleaseDisplayControl: rows affected for %s failed: %v", workspaceID, err)
+		emitDisplayControlEvent(c.Request.Context(), "display.control.release.failed", workspaceID, map[string]any{
+			"controlled_by": controlledBy,
+			"duration_ms":   time.Since(startedAt).Milliseconds(),
+			"error":         err.Error(),
+			"force":         req.Force,
+		})
+		c.JSON(http.StatusInternalServerError, gin.H{"error": "failed to release display control"})
+		return
+	}
+	if rowsAffected == 0 {
+		current, found, loadErr := h.loadActiveDisplayControl(c, workspaceID)
+		if loadErr != nil {
+			log.Printf("ReleaseDisplayControl: load active lock for %s failed: %v", workspaceID, loadErr)
+			emitDisplayControlEvent(c.Request.Context(), "display.control.release.failed", workspaceID, map[string]any{
+				"controlled_by": controlledBy,
+				"duration_ms":   time.Since(startedAt).Milliseconds(),
+				"error":         loadErr.Error(),
+				"force":         req.Force,
+			})
+			c.JSON(http.StatusInternalServerError, gin.H{"error": "failed to load display control"})
+			return
+		}
+		if !found {
+			emitDisplayControlEvent(c.Request.Context(), "display.control.release.completed", workspaceID, map[string]any{
+				"controlled_by": controlledBy,
+				"duration_ms":   time.Since(startedAt).Milliseconds(),
+				"force":         req.Force,
+				"rows_affected": rowsAffected,
+			})
+			c.JSON(http.StatusOK, workspaceDisplayControlNoneResponse{Controller: "none"})
+			return
+		}
+		emitDisplayControlEvent(c.Request.Context(), "display.control.release.failed", workspaceID, map[string]any{
+			"controlled_by": controlledBy,
+			"duration_ms":   time.Since(startedAt).Milliseconds(),
+			"error":         "display control held by another caller",
+			"force":         req.Force,
+		})
+		c.JSON(http.StatusConflict, gin.H{
+			"error":   "display control held by another caller",
+			"current": current,
+		})
+		return
+	}
+	emitDisplayControlEvent(c.Request.Context(), "display.control.release.completed", workspaceID, map[string]any{
+		"controlled_by": controlledBy,
+		"duration_ms":   time.Since(startedAt).Milliseconds(),
+		"force":         req.Force,
+		"rows_affected": rowsAffected,
+	})
+	c.JSON(http.StatusOK, workspaceDisplayControlNoneResponse{Controller: "none"})
+}
+
+func (h *WorkspaceHandler) loadActiveDisplayControl(c *gin.Context, workspaceID string) (workspaceDisplayControlResponse, bool, error) {
+	var lock workspaceDisplayControlResponse
+	err := db.DB.QueryRowContext(c.Request.Context(),
+		`SELECT controller, controlled_by, expires_at FROM workspace_display_control_locks WHERE workspace_id = $1 AND expires_at > now()`,
+		workspaceID,
+	).Scan(&lock.Controller, &lock.ControlledBy, &lock.ExpiresAt)
+	if err == nil {
+		return lock, true, nil
+	}
+	if err == sql.ErrNoRows {
+		return workspaceDisplayControlResponse{}, false, nil
+	}
+	return workspaceDisplayControlResponse{}, false, err
+}
+
+func (h *WorkspaceHandler) displayControlEnabled(c *gin.Context, workspaceID string) bool {
+	var raw string
+	err := db.DB.QueryRowContext(c.Request.Context(),
+		`SELECT COALESCE(compute, '{}'::jsonb) FROM workspaces WHERE id = $1`,
+		workspaceID,
+	).Scan(&raw)
+	if err != nil {
+		if err == sql.ErrNoRows {
+			c.JSON(http.StatusNotFound, gin.H{"error": "workspace not found"})
+			return false
+		}
+		log.Printf("displayControlEnabled: load compute for %s failed: %v", workspaceID, err)
+		c.JSON(http.StatusInternalServerError, gin.H{"error": "failed to load display config"})
+		return false
+	}
+	compute, err := parseWorkspaceDisplayCompute(workspaceID, raw)
+	if err != nil {
+		log.Printf("displayControlEnabled: invalid display config for %s: %v", workspaceID, err)
+		c.JSON(http.StatusInternalServerError, gin.H{"error": "invalid display config"})
+		return false
+	}
+	if compute.Display.Mode == "" || compute.Display.Mode == "none" {
+		c.JSON(http.StatusBadRequest, gin.H{"error": "display not enabled"})
+		return false
+	}
+	return true
+}
+
+func parseWorkspaceDisplayCompute(workspaceID, raw string) (models.WorkspaceCompute, error) {
+	var compute models.WorkspaceCompute
+	if raw == "" || raw == "{}" {
+		return compute, nil
+	}
+	if err := json.Unmarshal([]byte(raw), &compute); err != nil {
+		return models.WorkspaceCompute{}, fmt.Errorf("invalid compute JSON for %s: %w", workspaceID, err)
+	}
+	if err := validateWorkspaceDisplayConfig(compute.Display); err != nil {
+		return models.WorkspaceCompute{}, err
+	}
+	return compute, nil
+}
+
+func displayControlActor(c *gin.Context) (string, bool) {
+	if v, ok := c.Get("org_token_prefix"); ok {
+		if s, ok := v.(string); ok && s != "" {
+			return actorOrgTokenPrefix + s, true
+		}
+	}
+	if displayControlIsAdminToken(c) {
+		return actorAdminToken, true
+	}
+	// Browser session auth is intentionally observe-only until AdminAuth
+	// exposes a stable per-user or per-session identity in gin.Context.
+	return "", false
+}
+
+func displayControlIsAdminToken(c *gin.Context) bool {
+	adminSecret := os.Getenv("ADMIN_TOKEN")
+	if adminSecret == "" {
+		return false
+	}
+	tok := wsauth.BearerTokenFromHeader(c.GetHeader("Authorization"))
+	return subtle.ConstantTimeCompare([]byte(tok), []byte(adminSecret)) == 1
+}
+
+func emitDisplayControlEvent(ctx context.Context, eventType string, workspaceID string, payload map[string]any) {
+	if payload == nil {
+		payload = map[string]any{}
+	}
+	payloadJSON, err := json.Marshal(payload)
+	if err != nil {
+		log.Printf("emitDisplayControlEvent: marshal %s payload failed: %v", eventType, err)
+		return
+	}
+	if _, err := db.DB.ExecContext(ctx, `
+		INSERT INTO structure_events (event_type, workspace_id, payload, created_at)
+		VALUES ($1, $2, $3::jsonb, now())
+	`, eventType, workspaceID, string(payloadJSON)); err != nil {
+		log.Printf("emitDisplayControlEvent: insert %s failed: %v", eventType, err)
+	}
+}
@@ -0,0 +1,321 @@
+package handlers
+
+import (
+	"bytes"
+	"database/sql"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+	"time"
+
+	"github.com/DATA-DOG/go-sqlmock"
+	"github.com/gin-gonic/gin"
+)
+
+func attachDisplayControlAdminToken(t *testing.T, c *gin.Context) {
+	t.Helper()
+	t.Setenv("ADMIN_TOKEN", "test-admin-secret")
+	c.Request.Header.Set("Authorization", "Bearer test-admin-secret")
+}
+
+func TestWorkspaceDisplayControl_NoActiveLockReturnsNone(t *testing.T) {
+	mock := setupTestDB(t)
+	handler := NewWorkspaceHandler(newTestBroadcaster(), nil, "http://localhost:8080", t.TempDir())
+
+	mock.ExpectQuery(`SELECT controller, controlled_by, expires_at FROM workspace_display_control_locks WHERE workspace_id = \$1 AND expires_at > now\(\)`).
+		WithArgs("ws-display").
+		WillReturnError(sql.ErrNoRows)
+
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Params = gin.Params{{Key: "id", Value: "ws-display"}}
+	c.Request = httptest.NewRequest("GET", "/workspaces/ws-display/display/control", nil)
+
+	handler.DisplayControl(c)
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("expected status 200, got %d: %s", w.Code, w.Body.String())
+	}
+	var resp map[string]interface{}
+	if err := json.Unmarshal(w.Body.Bytes(), &resp); err != nil {
+		t.Fatalf("failed to parse response: %v", err)
+	}
+	if resp["controller"] != "none" {
+		t.Fatalf("controller = %v, want none", resp["controller"])
+	}
+	if _, ok := resp["expires_at"]; ok {
+		t.Fatalf("none response included expires_at: %#v", resp)
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Errorf("unmet sqlmock expectations: %v", err)
+	}
+}
+
+func TestWorkspaceDisplayControlAcquire_ClaimsUnlockedDisplay(t *testing.T) {
+	mock := setupTestDB(t)
+	handler := NewWorkspaceHandler(newTestBroadcaster(), nil, "http://localhost:8080", t.TempDir())
+	expiresAt := time.Date(2026, 5, 23, 18, 30, 0, 0, time.UTC)
+
+	mock.ExpectQuery(`SELECT COALESCE\(compute, '\{\}'::jsonb\) FROM workspaces WHERE id = \$1`).
+		WithArgs("ws-display").
+		WillReturnRows(sqlmock.NewRows([]string{"compute"}).AddRow(`{"display":{"mode":"desktop-control","protocol":"dcv","width":1920,"height":1080}}`))
+	mock.ExpectQuery(`INSERT INTO workspace_display_control_locks`).
+		WithArgs("ws-display", "user", "admin-token", 300).
+		WillReturnRows(sqlmock.NewRows([]string{"controller", "controlled_by", "expires_at"}).
+			AddRow("user", "admin-token", expiresAt))
+
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Params = gin.Params{{Key: "id", Value: "ws-display"}}
+	c.Request = httptest.NewRequest("POST", "/workspaces/ws-display/display/control/acquire", bytes.NewBufferString(`{"controller":"user","ttl_seconds":300}`))
+	c.Request.Header.Set("Content-Type", "application/json")
+	attachDisplayControlAdminToken(t, c)
+
+	handler.AcquireDisplayControl(c)
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("expected status 200, got %d: %s", w.Code, w.Body.String())
+	}
+	var resp map[string]interface{}
+	if err := json.Unmarshal(w.Body.Bytes(), &resp); err != nil {
+		t.Fatalf("failed to parse response: %v", err)
+	}
+	if resp["controller"] != "user" || resp["controlled_by"] != "admin-token" {
+		t.Fatalf("lock response = %#v, want user/admin-token", resp)
+	}
+	if resp["expires_at"] == "" {
+		t.Fatalf("expires_at missing in response: %#v", resp)
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Errorf("unmet sqlmock expectations: %v", err)
+	}
+}
+
+func TestWorkspaceDisplayControlAcquire_ActiveLockReturnsConflict(t *testing.T) {
+	mock := setupTestDB(t)
+	handler := NewWorkspaceHandler(newTestBroadcaster(), nil, "http://localhost:8080", t.TempDir())
+	expiresAt := time.Date(2026, 5, 23, 18, 30, 0, 0, time.UTC)
+
+	mock.ExpectQuery(`SELECT COALESCE\(compute, '\{\}'::jsonb\) FROM workspaces WHERE id = \$1`).
+		WithArgs("ws-display").
+		WillReturnRows(sqlmock.NewRows([]string{"compute"}).AddRow(`{"display":{"mode":"desktop-control","protocol":"dcv","width":1920,"height":1080}}`))
+	mock.ExpectQuery(`INSERT INTO workspace_display_control_locks`).
+		WithArgs("ws-display", "user", "admin-token", 300).
+		WillReturnError(sql.ErrNoRows)
+	mock.ExpectQuery(`SELECT controller, controlled_by, expires_at FROM workspace_display_control_locks WHERE workspace_id = \$1 AND expires_at > now\(\)`).
+		WithArgs("ws-display").
+		WillReturnRows(sqlmock.NewRows([]string{"controller", "controlled_by", "expires_at"}).
+			AddRow("agent", "sidecar", expiresAt))
+
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Params = gin.Params{{Key: "id", Value: "ws-display"}}
+	c.Request = httptest.NewRequest("POST", "/workspaces/ws-display/display/control/acquire", bytes.NewBufferString(`{"controller":"user","ttl_seconds":300}`))
+	c.Request.Header.Set("Content-Type", "application/json")
+	attachDisplayControlAdminToken(t, c)
+
+	handler.AcquireDisplayControl(c)
+
+	if w.Code != http.StatusConflict {
+		t.Fatalf("expected status 409, got %d: %s", w.Code, w.Body.String())
+	}
+	var resp map[string]interface{}
+	if err := json.Unmarshal(w.Body.Bytes(), &resp); err != nil {
+		t.Fatalf("failed to parse response: %v", err)
+	}
+	if resp["error"] != "display control already held" {
+		t.Fatalf("error = %v, want display control already held", resp["error"])
+	}
+	current, ok := resp["current"].(map[string]interface{})
+	if !ok || current["controller"] != "agent" || current["controlled_by"] != "sidecar" {
+		t.Fatalf("current lock = %#v, want agent/sidecar", resp["current"])
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Errorf("unmet sqlmock expectations: %v", err)
+	}
+}
+
+func TestWorkspaceDisplayControlAcquire_RejectsDisplayDisabledWorkspace(t *testing.T) {
+	mock := setupTestDB(t)
+	handler := NewWorkspaceHandler(newTestBroadcaster(), nil, "http://localhost:8080", t.TempDir())
+
+	mock.ExpectQuery(`SELECT COALESCE\(compute, '\{\}'::jsonb\) FROM workspaces WHERE id = \$1`).
+		WithArgs("ws-no-display").
+		WillReturnRows(sqlmock.NewRows([]string{"compute"}).AddRow(`{}`))
+
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Params = gin.Params{{Key: "id", Value: "ws-no-display"}}
+	c.Request = httptest.NewRequest("POST", "/workspaces/ws-no-display/display/control/acquire", bytes.NewBufferString(`{"controller":"user","ttl_seconds":300}`))
+	c.Request.Header.Set("Content-Type", "application/json")
+	attachDisplayControlAdminToken(t, c)
+
+	handler.AcquireDisplayControl(c)
+
+	if w.Code != http.StatusBadRequest {
+		t.Fatalf("expected status 400, got %d: %s", w.Code, w.Body.String())
+	}
+	var resp map[string]interface{}
+	if err := json.Unmarshal(w.Body.Bytes(), &resp); err != nil {
+		t.Fatalf("failed to parse response: %v", err)
+	}
+	if resp["error"] != "display not enabled" {
+		t.Fatalf("error = %v, want display not enabled", resp["error"])
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Errorf("unmet sqlmock expectations: %v", err)
+	}
+}
+
+func TestWorkspaceDisplayControlAcquire_RejectsCoarseSessionActor(t *testing.T) {
+	mock := setupTestDB(t)
+	handler := NewWorkspaceHandler(newTestBroadcaster(), nil, "http://localhost:8080", t.TempDir())
+
+	mock.ExpectQuery(`SELECT COALESCE\(compute, '\{\}'::jsonb\) FROM workspaces WHERE id = \$1`).
+		WithArgs("ws-display").
+		WillReturnRows(sqlmock.NewRows([]string{"compute"}).AddRow(`{"display":{"mode":"desktop-control","protocol":"dcv","width":1920,"height":1080}}`))
+
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Params = gin.Params{{Key: "id", Value: "ws-display"}}
+	c.Request = httptest.NewRequest("POST", "/workspaces/ws-display/display/control/acquire", bytes.NewBufferString(`{"controller":"user","ttl_seconds":300}`))
+	c.Request.Header.Set("Content-Type", "application/json")
+	c.Request.Header.Set("Cookie", "molecule_session=present")
+
+	handler.AcquireDisplayControl(c)
+
+	if w.Code != http.StatusForbidden {
+		t.Fatalf("expected status 403, got %d: %s", w.Code, w.Body.String())
+	}
+	var resp map[string]interface{}
+	if err := json.Unmarshal(w.Body.Bytes(), &resp); err != nil {
+		t.Fatalf("failed to parse response: %v", err)
+	}
+	if resp["error"] != "display control requires admin-token or org-token auth" {
+		t.Fatalf("error = %v, want display control requires admin-token or org-token auth", resp["error"])
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Errorf("unmet sqlmock expectations: %v", err)
+	}
+}
+
+func TestWorkspaceDisplayControlRelease_RemovesCallerLock(t *testing.T) {
+	mock := setupTestDB(t)
+	handler := NewWorkspaceHandler(newTestBroadcaster(), nil, "http://localhost:8080", t.TempDir())
+
+	mock.ExpectExec(`DELETE FROM workspace_display_control_locks WHERE workspace_id = \$1 AND controlled_by = \$2`).
+		WithArgs("ws-display", "admin-token").
+		WillReturnResult(sqlmock.NewResult(0, 1))
+
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Params = gin.Params{{Key: "id", Value: "ws-display"}}
+	c.Request = httptest.NewRequest("POST", "/workspaces/ws-display/display/control/release", nil)
+	attachDisplayControlAdminToken(t, c)
+
+	handler.ReleaseDisplayControl(c)
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("expected status 200, got %d: %s", w.Code, w.Body.String())
+	}
+	var resp map[string]interface{}
+	if err := json.Unmarshal(w.Body.Bytes(), &resp); err != nil {
+		t.Fatalf("failed to parse response: %v", err)
+	}
+	if resp["controller"] != "none" {
+		t.Fatalf("controller = %v, want none", resp["controller"])
+	}
+	if _, ok := resp["expires_at"]; ok {
+		t.Fatalf("none response included expires_at: %#v", resp)
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Errorf("unmet sqlmock expectations: %v", err)
+	}
+}
+
+func TestWorkspaceDisplayControlRelease_ConflictWhenCallerDoesNotOwnLock(t *testing.T) {
+	mock := setupTestDB(t)
+	handler := NewWorkspaceHandler(newTestBroadcaster(), nil, "http://localhost:8080", t.TempDir())
+	expiresAt := time.Date(2026, 5, 23, 18, 30, 0, 0, time.UTC)
+
+	mock.ExpectExec(`DELETE FROM workspace_display_control_locks WHERE workspace_id = \$1 AND controlled_by = \$2`).
+		WithArgs("ws-display", "admin-token").
+		WillReturnResult(sqlmock.NewResult(0, 0))
+	mock.ExpectQuery(`SELECT controller, controlled_by, expires_at FROM workspace_display_control_locks WHERE workspace_id = \$1 AND expires_at > now\(\)`).
+		WithArgs("ws-display").
+		WillReturnRows(sqlmock.NewRows([]string{"controller", "controlled_by", "expires_at"}).
+			AddRow("user", "org-token:abcd1234", expiresAt))
+
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Params = gin.Params{{Key: "id", Value: "ws-display"}}
+	c.Request = httptest.NewRequest("POST", "/workspaces/ws-display/display/control/release", nil)
+	attachDisplayControlAdminToken(t, c)
+
+	handler.ReleaseDisplayControl(c)
+
+	if w.Code != http.StatusConflict {
+		t.Fatalf("expected status 409, got %d: %s", w.Code, w.Body.String())
+	}
+	var resp map[string]interface{}
+	if err := json.Unmarshal(w.Body.Bytes(), &resp); err != nil {
+		t.Fatalf("failed to parse response: %v", err)
+	}
+	if resp["error"] != "display control held by another caller" {
+		t.Fatalf("error = %v, want display control held by another caller", resp["error"])
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Errorf("unmet sqlmock expectations: %v", err)
+	}
+}
+
+func TestWorkspaceDisplayControlRelease_RejectsOrgTokenForceRelease(t *testing.T) {
+	setupTestDB(t)
+	handler := NewWorkspaceHandler(newTestBroadcaster(), nil, "http://localhost:8080", t.TempDir())
+
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Params = gin.Params{{Key: "id", Value: "ws-display"}}
+	c.Set("org_token_prefix", "abcd1234")
+	c.Request = httptest.NewRequest("POST", "/workspaces/ws-display/display/control/release", bytes.NewBufferString(`{"force":true}`))
+	c.Request.Header.Set("Content-Type", "application/json")
+
+	handler.ReleaseDisplayControl(c)
+
+	if w.Code != http.StatusForbidden {
+		t.Fatalf("expected status 403, got %d: %s", w.Code, w.Body.String())
+	}
+	var resp map[string]interface{}
+	if err := json.Unmarshal(w.Body.Bytes(), &resp); err != nil {
+		t.Fatalf("failed to parse response: %v", err)
+	}
+	if resp["error"] != "force release requires admin-token auth" {
+		t.Fatalf("error = %v, want force release requires admin-token auth", resp["error"])
+	}
+}
+
+func TestWorkspaceDisplayControlAcquire_RejectsAgentImpersonation(t *testing.T) {
+	setupTestDB(t)
+	handler := NewWorkspaceHandler(newTestBroadcaster(), nil, "http://localhost:8080", t.TempDir())
+
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Params = gin.Params{{Key: "id", Value: "ws-display"}}
+	c.Request = httptest.NewRequest("POST", "/workspaces/ws-display/display/control/acquire", bytes.NewBufferString(`{"controller":"agent","ttl_seconds":300}`))
+	c.Request.Header.Set("Content-Type", "application/json")
+	attachDisplayControlAdminToken(t, c)
+
+	handler.AcquireDisplayControl(c)
+
+	if w.Code != http.StatusBadRequest {
+		t.Fatalf("expected status 400, got %d: %s", w.Code, w.Body.String())
+	}
+	var resp map[string]interface{}
+	if err := json.Unmarshal(w.Body.Bytes(), &resp); err != nil {
+		t.Fatalf("failed to parse response: %v", err)
+	}
+	if resp["error"] != "browser callers may only acquire user display control" {
+		t.Fatalf("error = %v, want browser callers may only acquire user display control", resp["error"])
+	}
+}
@@ -296,6 +296,8 @@ func (h *WorkspaceHandler) buildProvisionerConfig(
 		WorkspaceAccess:    workspaceAccess,
 		Tier:               payload.Tier,
 		Runtime:            payload.Runtime,
+		InstanceType:       payload.Compute.InstanceType,
+		DiskGB:             int32(payload.Compute.Volume.RootGB),
 		EnvVars:            envVars,
 		PlatformURL:        h.platformURL,
 		AwarenessURL:       os.Getenv("AWARENESS_URL"),
@@ -1009,4 +1011,3 @@ func (h *WorkspaceHandler) provisionWorkspaceCP(workspaceID, templatePath string

 	log.Printf("CPProvisioner: workspace %s started as machine %s via control plane", workspaceID, machineID)
 }
-
@@ -164,7 +164,7 @@ func (h *WorkspaceHandler) maybeRestartAfterFileWrite(workspaceID string) {
 // isRestarting reports whether a restart cycle is currently in flight for
 // the workspace. Callers that have their own "container looks dead" probe
 // MUST consult this before triggering a restart, because during the
-// 20-30s EC2-pending window the workspace's url='' and IsRunning()=false
+// 20-30s EC2-pending window the workspace's url=” and IsRunning()=false
 // looks identical to a dead container — and any restart-triggering probe
 // (maybeMarkContainerDead from canvas /delegations poll, or the trailing
 // restart-context probe at the end of runRestartCycle) will set
@@ -337,7 +337,7 @@ func (h *WorkspaceHandler) Restart(c *gin.Context) {
 	}

 	var configFiles map[string][]byte
-	payload := models.CreateWorkspacePayload{Name: wsName, Tier: tier, Runtime: containerRuntime}
+	payload := withStoredCompute(ctx, id, models.CreateWorkspacePayload{Name: wsName, Tier: tier, Runtime: containerRuntime})
 	log.Printf("Restart: workspace %s (%s) runtime=%q", wsName, id, containerRuntime)

 	// #12: ?reset=true (or body.Reset) discards the claude-sessions volume
@@ -791,7 +791,7 @@ func (h *WorkspaceHandler) runRestartCycle(workspaceID string) {
 	})

 	// Runtime from DB — no more config file parsing
-	payload := models.CreateWorkspacePayload{Name: wsName, Tier: tier, Runtime: dbRuntime}
+	payload := withStoredCompute(ctx, workspaceID, models.CreateWorkspacePayload{Name: wsName, Tier: tier, Runtime: dbRuntime})

 	// Snapshot restart-context data before the new session overwrites
 	// last_heartbeat_at. Issue #19 Layer 1.
@@ -948,7 +948,7 @@ func (h *WorkspaceHandler) Resume(c *gin.Context) {
 		h.broadcaster.RecordAndBroadcast(ctx, string(events.EventWorkspaceProvisioning), ws.id, map[string]interface{}{
 			"name": ws.name, "tier": ws.tier, "runtime": ws.runtime,
 		})
-		payload := models.CreateWorkspacePayload{Name: ws.name, Tier: ws.tier, Runtime: ws.runtime}
+		payload := withStoredCompute(ctx, ws.id, models.CreateWorkspacePayload{Name: ws.name, Tier: ws.tier, Runtime: ws.runtime})
 		// Resume is provision-only (workspace is paused, no live container
 		// to stop). provisionWorkspaceAuto handles backend routing and the
 		// no-backend mark-failed fallback identically to Create. Pre-
@@ -29,7 +29,7 @@ func TestWorkspaceGet_Success(t *testing.T) {
 		"parent_id", "active_tasks", "max_concurrent_tasks", "last_error_rate", "last_sample_error",
 		"uptime_seconds", "current_task", "runtime", "workspace_dir", "x", "y", "collapsed",
 		"budget_limit", "monthly_spend",
-		"broadcast_enabled", "talk_to_user_enabled",
+		"broadcast_enabled", "talk_to_user_enabled", "compute",
 	}
 	mock.ExpectQuery("SELECT w.id, w.name").
 		WithArgs("cccccccc-0001-0000-0000-000000000000").
@@ -37,7 +37,7 @@ func TestWorkspaceGet_Success(t *testing.T) {
 			AddRow("cccccccc-0001-0000-0000-000000000000", "My Agent", "worker", 1, "online", []byte(`{"name":"test"}`),
 				"http://localhost:8001", nil, 2, 1, 0.05, "", 3600, "working", "langgraph",
 				"", 10.0, 20.0, false,
-				nil, 0, false, true))
+				nil, 0, false, true, []byte(`{}`)))

 	w := httptest.NewRecorder()
 	c, _ := gin.CreateTestContext(w)
@@ -119,7 +119,7 @@ func TestWorkspaceGet_RemovedReturns410(t *testing.T) {
 		"parent_id", "active_tasks", "max_concurrent_tasks", "last_error_rate", "last_sample_error",
 		"uptime_seconds", "current_task", "runtime", "workspace_dir", "x", "y", "collapsed",
 		"budget_limit", "monthly_spend",
-		"broadcast_enabled", "talk_to_user_enabled",
+		"broadcast_enabled", "talk_to_user_enabled", "compute",
 	}
 	mock.ExpectQuery("SELECT w.id, w.name").
 		WithArgs(id).
@@ -127,7 +127,7 @@ func TestWorkspaceGet_RemovedReturns410(t *testing.T) {
 			AddRow(id, "Old Agent", "worker", 1, string(models.StatusRemoved), []byte(`null`),
 				"", nil, 0, 1, 0.0, "", 0, "", "langgraph",
 				"", 0.0, 0.0, false,
-				nil, 0, false, true))
+				nil, 0, false, true, []byte(`{}`)))
 	mock.ExpectQuery(`SELECT updated_at FROM workspaces`).
 		WithArgs(id).
 		WillReturnRows(sqlmock.NewRows([]string{"updated_at"}).AddRow(removedAt))
@@ -183,7 +183,7 @@ func TestWorkspaceGet_RemovedReturns410WithNullRemovedAtOnTimestampFetchFailure(
 		"parent_id", "active_tasks", "max_concurrent_tasks", "last_error_rate", "last_sample_error",
 		"uptime_seconds", "current_task", "runtime", "workspace_dir", "x", "y", "collapsed",
 		"budget_limit", "monthly_spend",
-		"broadcast_enabled", "talk_to_user_enabled",
+		"broadcast_enabled", "talk_to_user_enabled", "compute",
 	}
 	mock.ExpectQuery("SELECT w.id, w.name").
 		WithArgs(id).
@@ -191,7 +191,7 @@ func TestWorkspaceGet_RemovedReturns410WithNullRemovedAtOnTimestampFetchFailure(
 			AddRow(id, "Vanished", "worker", 1, string(models.StatusRemoved), []byte(`null`),
 				"", nil, 0, 1, 0.0, "", 0, "", "langgraph",
 				"", 0.0, 0.0, false,
-				nil, 0, false, true))
+				nil, 0, false, true, []byte(`{}`)))
 	// Simulate the row vanishing between the two queries.
 	mock.ExpectQuery(`SELECT updated_at FROM workspaces`).
 		WithArgs(id).
@@ -246,7 +246,7 @@ func TestWorkspaceGet_RemovedWithIncludeQueryReturns200(t *testing.T) {
 		"parent_id", "active_tasks", "max_concurrent_tasks", "last_error_rate", "last_sample_error",
 		"uptime_seconds", "current_task", "runtime", "workspace_dir", "x", "y", "collapsed",
 		"budget_limit", "monthly_spend",
-		"broadcast_enabled", "talk_to_user_enabled",
+		"broadcast_enabled", "talk_to_user_enabled", "compute",
 	}
 	mock.ExpectQuery("SELECT w.id, w.name").
 		WithArgs(id).
@@ -254,7 +254,7 @@ func TestWorkspaceGet_RemovedWithIncludeQueryReturns200(t *testing.T) {
 			AddRow(id, "Audit Agent", "worker", 1, string(models.StatusRemoved), []byte(`null`),
 				"", nil, 0, 1, 0.0, "", 0, "", "langgraph",
 				"", 0.0, 0.0, false,
-				nil, 0, false, true))
+				nil, 0, false, true, []byte(`{}`)))
 	// last_outbound_at follow-up query (existing path)
 	mock.ExpectQuery(`SELECT last_outbound_at FROM workspaces`).
 		WithArgs(id).
@@ -718,7 +718,7 @@ func TestWorkspaceList_Empty(t *testing.T) {
 			"parent_id", "active_tasks", "last_error_rate", "last_sample_error",
 			"uptime_seconds", "current_task", "runtime", "workspace_dir", "x", "y", "collapsed",
 			"budget_limit", "monthly_spend",
-			"broadcast_enabled", "talk_to_user_enabled",
+			"broadcast_enabled", "talk_to_user_enabled", "compute",
 		}))

 	w := httptest.NewRecorder()
@@ -1422,7 +1422,7 @@ func TestWorkspaceGet_FinancialFieldsStripped(t *testing.T) {
 		"parent_id", "active_tasks", "max_concurrent_tasks", "last_error_rate", "last_sample_error",
 		"uptime_seconds", "current_task", "runtime", "workspace_dir", "x", "y", "collapsed",
 		"budget_limit", "monthly_spend",
-		"broadcast_enabled", "talk_to_user_enabled",
+		"broadcast_enabled", "talk_to_user_enabled", "compute",
 	}
 	// Populate with non-zero financial values to confirm they are stripped.
 	mock.ExpectQuery("SELECT w.id, w.name").
@@ -1431,7 +1431,7 @@ func TestWorkspaceGet_FinancialFieldsStripped(t *testing.T) {
 			AddRow("cccccccc-0010-0000-0000-000000000000", "Finance Test", "worker", 1, "online", []byte(`{}`),
 				"http://localhost:9001", nil, 0, 1, 0.0, "", 0, "", "langgraph",
 				"", 0.0, 0.0, false,
-				int64(50000), int64(12500), false, true)) // budget_limit=500 USD, spend=125 USD
+				int64(50000), int64(12500), false, true, []byte(`{}`))) // budget_limit=500 USD, spend=125 USD

 	w := httptest.NewRecorder()
 	c, _ := gin.CreateTestContext(w)
@@ -1479,7 +1479,7 @@ func TestWorkspaceGet_SensitiveFieldsStripped(t *testing.T) {
 		"parent_id", "active_tasks", "max_concurrent_tasks", "last_error_rate", "last_sample_error",
 		"uptime_seconds", "current_task", "runtime", "workspace_dir", "x", "y", "collapsed",
 		"budget_limit", "monthly_spend",
-		"broadcast_enabled", "talk_to_user_enabled",
+		"broadcast_enabled", "talk_to_user_enabled", "compute",
 	}
 	mock.ExpectQuery("SELECT w.id, w.name").
 		WithArgs("cccccccc-0955-0000-0000-000000000000").
@@ -1492,7 +1492,7 @@ func TestWorkspaceGet_SensitiveFieldsStripped(t *testing.T) {
 				"langgraph",
 				"/home/user/secret-projects/client-work",
 				0.0, 0.0, false,
-				nil, 0, false, true))
+				nil, 0, false, true, []byte(`{}`)))

 	w := httptest.NewRecorder()
 	c, _ := gin.CreateTestContext(w)
@@ -3,6 +3,7 @@ package pgplugin
 import (
 	"encoding/json"
 	"errors"
+	"log"
 	"net/http"
 	"strings"

@@ -246,7 +247,9 @@ func (h *Handler) forget(w http.ResponseWriter, r *http.Request, id string) {
 func writeJSON(w http.ResponseWriter, status int, body interface{}) {
 	w.Header().Set("Content-Type", "application/json")
 	w.WriteHeader(status)
-	_ = json.NewEncoder(w).Encode(body)
+	if err := json.NewEncoder(w).Encode(body); err != nil {
+		log.Printf("pgplugin: JSON encode error: %v", err)
+	}
 }

 func writeError(w http.ResponseWriter, status int, code contract.ErrorCode, message string, details map[string]interface{}) {
@@ -35,16 +35,16 @@ type Workspace struct {
 	// DeliveryMode: "push" (synchronous to URL — default) or "poll" (logged
 	// to activity_logs, agent reads via GET /activity?since_id=). See
 	// migration 045 + RFC #2339.
-	DeliveryMode       string          `json:"delivery_mode" db:"delivery_mode"`
+	DeliveryMode string `json:"delivery_mode" db:"delivery_mode"`
 	// BroadcastEnabled: when true the workspace may call POST /broadcast to
 	// deliver a message to all non-removed agent workspaces in the org.
 	// Default false — only privileged orchestrators should hold this ability.
-	BroadcastEnabled   bool            `json:"broadcast_enabled" db:"broadcast_enabled"`
+	BroadcastEnabled bool `json:"broadcast_enabled" db:"broadcast_enabled"`
 	// TalkToUserEnabled: when false the workspace's send_message_to_user calls
 	// and POST /notify requests are rejected with HTTP 403 so the agent is
 	// forced to route updates through a parent workspace. Default true
 	// (preserves existing behaviour for all workspaces).
-	TalkToUserEnabled  bool            `json:"talk_to_user_enabled" db:"talk_to_user_enabled"`
+	TalkToUserEnabled bool `json:"talk_to_user_enabled" db:"talk_to_user_enabled"`
 	// Canvas layout fields (from JOIN)
 	X         float64 `json:"x"`
 	Y         float64 `json:"y"`
@@ -71,12 +71,12 @@ type RegisterPayload struct {
 	// enforces the conditional requirement based on the resolved
 	// delivery mode (payload value, falling back to the row's existing
 	// value, falling back to "push").
-	URL          string          `json:"url"`
-	AgentCard    json.RawMessage `json:"agent_card" binding:"required"`
+	URL       string          `json:"url"`
+	AgentCard json.RawMessage `json:"agent_card" binding:"required"`
 	// DeliveryMode is optional. Empty string means "keep the existing
 	// value on the workspace row, or default to push for new rows".
 	// When set, must be one of DeliveryModePush / DeliveryModePoll.
-	DeliveryMode string          `json:"delivery_mode,omitempty"`
+	DeliveryMode string `json:"delivery_mode,omitempty"`
 }

 type HeartbeatPayload struct {
@@ -154,19 +154,36 @@ type MemorySeed struct {
 	Scope   string `json:"scope" yaml:"scope"` // LOCAL, TEAM, GLOBAL
 }

+type WorkspaceComputeVolume struct {
+	RootGB int `json:"root_gb,omitempty"`
+}
+
+type WorkspaceComputeDisplay struct {
+	Mode     string `json:"mode,omitempty"`
+	Width    int    `json:"width,omitempty"`
+	Height   int    `json:"height,omitempty"`
+	Protocol string `json:"protocol,omitempty"`
+}
+
+type WorkspaceCompute struct {
+	InstanceType string                  `json:"instance_type,omitempty"`
+	Volume       WorkspaceComputeVolume  `json:"volume,omitempty"`
+	Display      WorkspaceComputeDisplay `json:"display,omitempty"`
+}
+
 type CreateWorkspacePayload struct {
-	Name     string  `json:"name" binding:"required"`
-	Role     string  `json:"role"`
-	Template string  `json:"template"` // workspace-configs-templates folder name
-	Tier     int     `json:"tier"`
-	Model    string  `json:"model"`
-	Runtime      string  `json:"runtime"`       // "langgraph" (default), "claude-code", etc.
-	External     bool    `json:"external"`      // true = no Docker container, just a registered URL
-	URL          string  `json:"url"`           // for external workspaces: the A2A endpoint URL (push mode only — omit for poll)
+	Name     string `json:"name" binding:"required"`
+	Role     string `json:"role"`
+	Template string `json:"template"` // workspace-configs-templates folder name
+	Tier     int    `json:"tier"`
+	Model    string `json:"model"`
+	Runtime  string `json:"runtime"`  // "langgraph" (default), "claude-code", etc.
+	External bool   `json:"external"` // true = no Docker container, just a registered URL
+	URL      string `json:"url"`      // for external workspaces: the A2A endpoint URL (push mode only — omit for poll)
 	// DeliveryMode: "push" (default) sends inbound A2A to URL synchronously;
 	// "poll" records inbound to activity_logs for the agent to consume via
 	// GET /activity?since_id=. Poll mode does not require a URL. See #2339.
-	DeliveryMode string  `json:"delivery_mode,omitempty"`
+	DeliveryMode    string  `json:"delivery_mode,omitempty"`
 	WorkspaceDir    string  `json:"workspace_dir"`    // host path to mount as /workspace (empty = isolated volume)
 	WorkspaceAccess string  `json:"workspace_access"` // "none" (default), "read_only", or "read_write" — see #65
 	ParentID        *string `json:"parent_id"`
@@ -180,7 +197,11 @@ type CreateWorkspacePayload struct {
 	// MaxConcurrentTasks caps parallel A2A + cron dispatch. 0 means use
 	// DefaultMaxConcurrentTasks. Leaders typically set 3.
 	MaxConcurrentTasks int `json:"max_concurrent_tasks"`
-	Canvas   struct {
+	// Compute is the product-facing per-workspace EC2 shape/display
+	// contract. Phase 1 uses instance_type + volume.root_gb and persists
+	// display for future desktop-control workspaces.
+	Compute WorkspaceCompute `json:"compute,omitempty"`
+	Canvas  struct {
 		X float64 `json:"x"`
 		Y float64 `json:"y"`
 	} `json:"canvas"`
@@ -152,12 +152,14 @@ func (p *CPProvisioner) adminAuthHeaders(req *http.Request) {
 }

 type cpProvisionRequest struct {
-	OrgID       string            `json:"org_id"`
-	WorkspaceID string            `json:"workspace_id"`
-	Runtime     string            `json:"runtime"`
-	Tier        int               `json:"tier"`
-	PlatformURL string            `json:"platform_url"`
-	Env         map[string]string `json:"env"`
+	OrgID        string            `json:"org_id"`
+	WorkspaceID  string            `json:"workspace_id"`
+	Runtime      string            `json:"runtime"`
+	Tier         int               `json:"tier"`
+	InstanceType string            `json:"instance_type,omitempty"`
+	DiskGB       int32             `json:"disk_gb,omitempty"`
+	PlatformURL  string            `json:"platform_url"`
+	Env          map[string]string `json:"env"`
 	// ConfigFiles are template + generated config files to write into the
 	// EC2 instance's /configs directory. OFFSEC-010: collected by
 	// collectCPConfigFiles which rejects symlinks and non-regular files
@@ -206,13 +208,15 @@ func (p *CPProvisioner) Start(ctx context.Context, cfg WorkspaceConfig) (string,
 	}

 	req := cpProvisionRequest{
-		OrgID:       p.orgID,
-		WorkspaceID: cfg.WorkspaceID,
-		Runtime:     cfg.Runtime,
-		Tier:        cfg.Tier,
-		PlatformURL: cfg.PlatformURL,
-		Env:         env,
-		ConfigFiles: configFiles,
+		OrgID:        p.orgID,
+		WorkspaceID:  cfg.WorkspaceID,
+		Runtime:      cfg.Runtime,
+		Tier:         cfg.Tier,
+		InstanceType: cfg.InstanceType,
+		DiskGB:       cfg.DiskGB,
+		PlatformURL:  cfg.PlatformURL,
+		Env:          env,
+		ConfigFiles:  configFiles,
 	}

 	body, err := json.Marshal(req)
@@ -191,6 +191,12 @@ func TestStart_HappyPath(t *testing.T) {
 		if body.WorkspaceID != "ws-1" || body.Runtime != "python" {
 			t.Errorf("body mismatch: %+v", body)
 		}
+		if body.InstanceType != "m6i.xlarge" {
+			t.Errorf("instance_type = %q, want m6i.xlarge", body.InstanceType)
+		}
+		if body.DiskGB != 100 {
+			t.Errorf("disk_gb = %d, want 100", body.DiskGB)
+		}
 		w.WriteHeader(http.StatusCreated)
 		_, _ = io.WriteString(w, `{"instance_id":"i-abc123","state":"pending"}`)
 	}))
@@ -205,6 +211,7 @@ func TestStart_HappyPath(t *testing.T) {

 	id, err := p.Start(context.Background(), WorkspaceConfig{
 		WorkspaceID: "ws-1", Runtime: "python", Tier: 1, PlatformURL: "http://tenant",
+		InstanceType: "m6i.xlarge", DiskGB: 100,
 	})
 	if err != nil {
 		t.Fatalf("Start: %v", err)
@@ -362,7 +369,7 @@ func TestStart_CollectsConfigFiles(t *testing.T) {
 	p := &CPProvisioner{baseURL: srv.URL, orgID: "org-1", httpClient: srv.Client()}
 	_, err := p.Start(context.Background(), WorkspaceConfig{
 		WorkspaceID:  "ws-1",
-		Runtime:     "python",
+		Runtime:      "python",
 		Tier:         1,
 		PlatformURL:  "http://tenant",
 		TemplatePath: tmpl,
@@ -424,7 +431,7 @@ func TestStart_SymlinkTemplatePathError(t *testing.T) {
 	p := &CPProvisioner{baseURL: "http://unused", orgID: "org-1", httpClient: &http.Client{Timeout: time.Second}}
 	_, err := p.Start(context.Background(), WorkspaceConfig{
 		WorkspaceID:  "ws-1",
-		Runtime:     "python",
+		Runtime:      "python",
 		TemplatePath: symlink, // symlink root → OFFSEC-010 guard should fire
 	})
 	if err == nil {
@@ -98,6 +98,8 @@ type WorkspaceConfig struct {
 	WorkspacePath      string            // Host path to bind-mount as /workspace (if empty, uses Docker named volume)
 	Tier               int
 	Runtime            string            // "langgraph" (default) or "claude-code", "codex", "ollama", "custom"
+	InstanceType       string            // Optional CP EC2 instance type override (SaaS only)
+	DiskGB             int32             // Optional CP root volume size override in GiB (SaaS only)
 	EnvVars            map[string]string // Additional env vars (API keys, etc.)
 	PlatformURL        string
 	AwarenessURL       string
@@ -1605,4 +1607,3 @@ func parseOCIPlatform(s string) *ocispec.Platform {
 	}
 	return &ocispec.Platform{OS: parts[0], Architecture: parts[1]}
 }
-
@@ -0,0 +1,59 @@
+# T4 privilege contract — generated from
+# molecule-ai/molecule-core workspace-server/internal/provisioner/t4_privilege_contract.go
+# RFC: molecule-ai/internal#456
+# Do NOT edit this file by hand; regenerate via `go run ./cmd/t4-contract-dump > t4_capabilities.yaml`.
+version: 1
+agent_uid: 1000
+capabilities:
+  - name: "agent_home_writable"
+    description: "/agent-home is writable by the agent (Files API split per task #128). The Files API redesign uses /agent-home as the user-writable root; the agent must be able to create files there without sudo."
+    severity: hard
+    source: "task #128 Files API redesign; memory reference_post_suspension_pipeline"
+    probe: "TF=/agent-home/.t4-cap-write-probe-${MOLECULE_T4_PROBE_ID:-$$}; echo ok > \"$TF\" && [ \"$(cat \"$TF\")\" = \"ok\" ] && rm -f \"$TF\""
+  - name: "agent_uid_1000"
+    description: "The container's primary process (the runtime, post-gosu) runs as uid 1000, not root. T4 grants full machine access via privileged + host PID + Docker socket — the WORKLOAD inside that privileged container must still be unprivileged to prevent every untrusted code execution from being trivially root-on-host."
+    severity: hard
+    source: "RFC internal#456 §2.1.2; memory feedback_hermes_listpeers_401_token_root600_unreadable_by_uid1000"
+    probe: "[ \"$(id -u)\" = \"1000\" ]"
+  - name: "auth_token_agent_owned"
+    description: "/configs/.auth_token is owned by uid 1000 (== AgentUID) so the a2a_mcp_server can read its bearer. In SaaS mode molecule-runtime itself writes the token via save_token() — the ownership equals the runtime's exec uid. If the runtime ever runs as root, this fails and list_peers 401s (the Hermes class bug)."
+    severity: hard
+    source: "RFC internal#456 §10; memory feedback_hermes_listpeers_401_token_root600_unreadable_by_uid1000"
+    probe: "[ -e /configs/.auth_token ] && [ \"$(stat -c '%u' /configs/.auth_token)\" = \"1000\" ]"
+  - name: "docker_socket_reachable"
+    description: "/var/run/docker.sock is bind-mounted and host Docker is reachable from the T4 container. The probe enters the host mount+PID namespaces before running docker info so it validates the same host-control path production agents use, instead of depending on the template image's Docker CLI/socket group details."
+    severity: hard
+    source: "provisioner.go applyHostConfig T4 branch (case 4)"
+    probe: "sudo -n nsenter --target 1 --mount --pid -- docker info >/dev/null 2>&1"
+  - name: "host_fs_write_readback"
+    description: "Host filesystem is mounted at /host and the agent can write+read+remove a file there via sudo. Proves real host reach (not just a PID-1 namespace trick on an isolated init)."
+    severity: hard
+    source: "RFC internal#456 §11"
+    probe: "MARKER=\"t4cap-$(date +%s)-$RANDOM\"; PROBE_FILE=\"/host/tmp/.t4-cap-probe-${MOLECULE_T4_PROBE_ID:-$$}\"; sudo -n sh -c \"echo $MARKER > $PROBE_FILE\" && [ \"$(sudo -n cat $PROBE_FILE)\" = \"$MARKER\" ] && sudo -n rm -f $PROBE_FILE"
+  - name: "host_root_reach_via_nsenter"
+    description: "The uid-1000 agent can attain host root via `sudo -n nsenter --target 1 --mount --pid -- id -u` returning 0. This is the T4 escalation leg: full machine access means the agent CAN escalate to host root deliberately, even though it does not run as root by default."
+    severity: hard
+    source: "RFC internal#456 §11; memory reference_per_template_privilege_contract_class_audit_2026_05_16"
+    probe: "[ \"$(sudo -n nsenter --target 1 --mount --pid -- id -u)\" = \"0\" ]"
+  - name: "list_peers_http_200"
+    description: "The platform list_peers HTTP endpoint (served by the in-container a2a_mcp_server) returns HTTP 200 when called from uid 1000 with the bearer from /configs/.auth_token. This proves the WHOLE token-ownership chain end-to-end: token written under correct uid → reader uid matches → bearer non-empty → platform accepts. A self-contained empirical test for the Hermes class bug."
+    severity: hard
+    source: "memory reference_openclaw_fresh_provision_nonfunctional_anthropic_default_unroutable; memory reference_openclaw_mcp_peer_wiring_rootcause"
+    probe: "BEARER=$(cat /configs/.auth_token 2>/dev/null || echo \"\"); [ -n \"$BEARER\" ] || exit 1; PORT=$(cat /configs/.platform_port 2>/dev/null || echo \"8080\"); STATUS=$(curl -sS -o /dev/null -w '%{http_code}' -H \"Authorization: Bearer $BEARER\" \"http://127.0.0.1:${PORT}/list_peers\"); [ \"$STATUS\" = \"200\" ]"
+  - name: "network_egress_https"
+    description: "Generic HTTPS egress works. T4 is unconstrained network; the canonical test target is the Molecule-owned Gitea middleman over its public name. CI must not depend on GitHub or other mirrors for this probe. Any reachable HTTPS endpoint satisfies it — the YAML carries the recommended targets but accepts any 200/301/302."
+    severity: hard
+    source: "task #174 brief"
+    probe: "for U in $MOLECULE_T4_EGRESS_TARGETS; do   C=$(curl -sS -o /dev/null -w '%{http_code}' --max-time 8 \"$U\");   case \"$C\" in 2*|3*) exit 0;; esac; done; exit 1"
+    required_egress:
+      - "https://git.moleculesai.app/api/v1/version"
+  - name: "pid_host_visible"
+    description: "Host PID namespace is shared (--pid=host). The container can see host process 1 (systemd or pid-1 on the EC2 instance). Required for nsenter into host mount/pid namespaces."
+    severity: hard
+    source: "provisioner.go applyHostConfig T4 branch (case 4): hostCfg.PidMode = 'host'"
+    probe: "[ \"$(sudo -n nsenter --target 1 --mount --pid -- id -u)\" = \"0\" ]"
+  - name: "privileged_flag_observable"
+    description: "Container is started with --privileged. Observable from inside via /proc/self/status CapEff containing CAP_SYS_ADMIN. Defense-in-depth for the provisioner emission side."
+    severity: advisory
+    source: "provisioner.go applyHostConfig T4 branch (case 4)"
+    probe: "grep -q '^CapEff:.*ffffffffff' /proc/self/status"
@@ -120,8 +120,8 @@ func T4PrivilegeContract() []T4Capability {
 		},
 		{
 			Name:        "docker_socket_reachable",
-			Description: "/var/run/docker.sock is bind-mounted into the container so the agent can manage other containers (T4 use case: agent-as-orchestrator). Proven by 'docker version' returning a server section, which requires the daemon to answer over the socket.",
-			Probe:       `sudo -n docker version --format '{{.Server.Version}}' >/dev/null 2>&1`,
+			Description: "/var/run/docker.sock is bind-mounted and host Docker is reachable from the T4 container. The probe enters the host mount+PID namespaces before running docker info so it validates the same host-control path production agents use, instead of depending on the template image's Docker CLI/socket group details.",
+			Probe:       `sudo -n nsenter --target 1 --mount --pid -- docker info >/dev/null 2>&1`,
 			Severity:    SeverityHard,
 			Source:      "provisioner.go applyHostConfig T4 branch (case 4)",
 		},
@@ -145,7 +145,7 @@ func T4PrivilegeContract() []T4Capability {
 		},
 		{
 			Name:        "network_egress_https",
-			Description: "Generic HTTPS egress works. T4 is unconstrained network; the canonical test target is the Gitea instance over its public name, which any fork user can also resolve. Any reachable HTTPS endpoint satisfies it — the YAML carries the recommended targets but accepts any 200/301/302.",
+			Description: "Generic HTTPS egress works. T4 is unconstrained network; the canonical test target is the Molecule-owned Gitea middleman over its public name. CI must not depend on GitHub or other mirrors for this probe. Any reachable HTTPS endpoint satisfies it — the YAML carries the recommended targets but accepts any 200/301/302.",
 			Probe: `for U in $MOLECULE_T4_EGRESS_TARGETS; do ` +
 				`  C=$(curl -sS -o /dev/null -w '%{http_code}' --max-time 8 "$U"); ` +
 				`  case "$C" in 2*|3*) exit 0;; esac; ` +
@@ -153,10 +153,9 @@ func T4PrivilegeContract() []T4Capability {
 			Severity: SeverityHard,
 			Source:   "task #174 brief",
 			RequiredEgress: []string{
-				// Public, no auth, returns a small JSON.
+				// Molecule-owned, public, no auth, returns a small JSON.
 				// Adopters override via MOLECULE_T4_EGRESS_TARGETS.
-				"https://api.github.com/zen",
-				"https://www.google.com/generate_204",
+				"https://git.moleculesai.app/api/v1/version",
 			},
 		},
 		{
@@ -169,7 +168,7 @@ func T4PrivilegeContract() []T4Capability {
 		{
 			Name:        "pid_host_visible",
 			Description: "Host PID namespace is shared (--pid=host). The container can see host process 1 (systemd or pid-1 on the EC2 instance). Required for nsenter into host mount/pid namespaces.",
-			Probe:       `[ -d /proc/1/root ] && [ "$(sudo -n readlink /proc/1/ns/pid)" = "$(sudo -n readlink /proc/self/ns/pid)" ]`,
+			Probe:       `[ "$(sudo -n nsenter --target 1 --mount --pid -- id -u)" = "0" ]`,
 			Severity:    SeverityHard,
 			Source:      "provisioner.go applyHostConfig T4 branch (case 4): hostCfg.PidMode = 'host'",
 		},
@@ -1,6 +1,7 @@
 package provisioner

 import (
+	"os"
 	"strings"
 	"testing"
 )
@@ -77,6 +78,19 @@ func TestT4PrivilegeContract_CoreCapabilitiesPresent(t *testing.T) {
 	}
 }

+func TestT4PrivilegeContract_DefaultEgressUsesMoleculeOwnedEndpoint(t *testing.T) {
+	for _, c := range T4PrivilegeContract() {
+		for _, target := range c.RequiredEgress {
+			if strings.Contains(target, "github.com") {
+				t.Errorf("capability %q default egress target must not depend on GitHub mirror/API: %s", c.Name, target)
+			}
+			if strings.Contains(target, "google.com") {
+				t.Errorf("capability %q default egress target must not depend on external Google endpoint: %s", c.Name, target)
+			}
+		}
+	}
+}
+
 // TestT4PrivilegeContract_HardCapabilitiesMajority sanity-checks that
 // the contract is not silently advisory-only. If someone marks
 // everything as "advisory" the gate becomes a no-op without anyone
@@ -142,6 +156,17 @@ func TestAsYAML_EscapesEmbeddedQuotes(t *testing.T) {
 	}
 }

+func TestGeneratedT4CapabilitiesYAMLMatchesSSOT(t *testing.T) {
+	got, err := os.ReadFile("t4_capabilities.yaml")
+	if err != nil {
+		t.Fatalf("read generated t4_capabilities.yaml: %v", err)
+	}
+	want := AsYAML(T4PrivilegeContract())
+	if string(got) != want {
+		t.Fatal("generated t4_capabilities.yaml drifted from T4PrivilegeContract; regenerate with `go run ./cmd/t4-contract-dump > internal/provisioner/t4_capabilities.yaml`")
+	}
+}
+
 // TestAgentUIDConsistency ties the contract to the existing
 // provisioner-side AgentUID const. The probe for "agent_uid_1000"
 // hard-codes `id -u == 1000`; if AgentUID ever changes (no one
@@ -178,6 +178,13 @@ func Setup(hub *ws.Hub, broadcaster *events.Broadcaster, prov *provisioner.Provi
 		// the tenant AWS credentials. Admin-gated because console output
 		// can include user-data snippets we treat as semi-sensitive.
 		wsAdmin.GET("/workspaces/:id/console", wh.Console)
+		// Display sessions will eventually return short-lived proxied DCV
+		// URLs, so keep the endpoint admin-gated from the first unavailable
+		// state rather than widening it later.
+		wsAdmin.GET("/workspaces/:id/display", wh.Display)
+		wsAdmin.GET("/workspaces/:id/display/control", wh.DisplayControl)
+		wsAdmin.POST("/workspaces/:id/display/control/acquire", wh.AcquireDisplayControl)
+		wsAdmin.POST("/workspaces/:id/display/control/release", wh.ReleaseDisplayControl)

 		// Admin memory backup/restore (#1051) — bulk export/import of agent
 		// memories for safe Docker rebuilds. Matches workspaces by name on import.
@@ -0,0 +1,61 @@
+package router
+
+import (
+	"net/http"
+	"net/http/httptest"
+	"testing"
+
+	"github.com/DATA-DOG/go-sqlmock"
+	"github.com/Molecule-AI/molecule-monorepo/platform/internal/db"
+	"github.com/Molecule-AI/molecule-monorepo/platform/internal/handlers"
+	"github.com/Molecule-AI/molecule-monorepo/platform/internal/middleware"
+	"github.com/gin-gonic/gin"
+)
+
+func buildWorkspaceDisplayEngine(t *testing.T) *gin.Engine {
+	t.Helper()
+	gin.SetMode(gin.TestMode)
+	r := gin.New()
+	wh := handlers.NewWorkspaceHandler(nil, nil, "http://localhost:8080", t.TempDir())
+	r.GET("/workspaces/:id/display", middleware.AdminAuth(db.DB), wh.Display)
+	r.POST("/workspaces/:id/display/control/acquire", middleware.AdminAuth(db.DB), wh.AcquireDisplayControl)
+	return r
+}
+
+func TestWorkspaceDisplayRoute_RequiresAdminAuth(t *testing.T) {
+	t.Setenv("ADMIN_TOKEN", "test-admin-secret-not-presented-by-caller")
+	mock := setupRouterTestDB(t)
+	mock.ExpectQuery("SELECT COUNT.*FROM workspace_auth_tokens").
+		WillReturnRows(sqlmock.NewRows([]string{"count"}).AddRow(1))
+
+	r := buildWorkspaceDisplayEngine(t)
+	w := httptest.NewRecorder()
+	req := httptest.NewRequest(http.MethodGet, "/workspaces/ws-display/display", nil)
+	r.ServeHTTP(w, req)
+
+	if w.Code != http.StatusUnauthorized {
+		t.Errorf("expected 401 for unauthenticated request, got %d: %s", w.Code, w.Body.String())
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Errorf("sqlmock unmet: %v", err)
+	}
+}
+
+func TestWorkspaceDisplayControlRoute_RequiresAdminAuth(t *testing.T) {
+	t.Setenv("ADMIN_TOKEN", "test-admin-secret-not-presented-by-caller")
+	mock := setupRouterTestDB(t)
+	mock.ExpectQuery("SELECT COUNT.*FROM workspace_auth_tokens").
+		WillReturnRows(sqlmock.NewRows([]string{"count"}).AddRow(1))
+
+	r := buildWorkspaceDisplayEngine(t)
+	w := httptest.NewRecorder()
+	req := httptest.NewRequest(http.MethodPost, "/workspaces/ws-display/display/control/acquire", nil)
+	r.ServeHTTP(w, req)
+
+	if w.Code != http.StatusUnauthorized {
+		t.Errorf("expected 401 for unauthenticated request, got %d: %s", w.Code, w.Body.String())
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Errorf("sqlmock unmet: %v", err)
+	}
+}
@@ -425,6 +425,7 @@ func (s *Scheduler) fireSchedule(ctx context.Context, sched scheduleRow) {

 	lastStatus := "ok"
 	lastError := ""
+	resultKind := ""
 	if proxyErr != nil {
 		lastStatus = "error"
 		lastError = fmt.Sprintf("%v", proxyErr)
@@ -433,8 +434,26 @@ func (s *Scheduler) fireSchedule(ctx context.Context, sched scheduleRow) {
 		lastStatus = "error"
 		lastError = fmt.Sprintf("HTTP %d", statusCode)
 		log.Printf("Scheduler: '%s' non-2xx: %d", sched.Name, statusCode)
+	} else if a2aErr := a2aErrorFromBody(respBody); a2aErr != "" {
+		lastStatus = "error"
+		lastError = fmt.Sprintf("A2A adapter error: %s", a2aErr)
+		log.Printf("Scheduler: '%s' A2A adapter error (HTTP %d): %s", sched.Name, statusCode, a2aErr)
 	} else {
-		log.Printf("Scheduler: '%s' completed (HTTP %d)", sched.Name, statusCode)
+		// HTTP 200 — inspect response body for SDK-layer errors.
+		// The claude-code-sdk adapter returns HTTP 200 even when the inner
+		// LLM call throws (e.g. Max-plan rate-limit, quota exhaustion, SDK
+		// internal errors). Without this check those failures surface as
+		// "completed (HTTP 200)" in last_status while the agent chat shows
+		// errors — a silent failure that hides schedule outages.
+		// See: #1696.
+		resultKind = detectResultKind(respBody)
+		if resultKind != "" && resultKind != "ok" {
+			lastStatus = resultKind
+			lastError = fmt.Sprintf("SDK error: result_kind=%s", resultKind)
+			log.Printf("Scheduler: '%s' SDK error detected — result_kind=%s", sched.Name, resultKind)
+		} else {
+			log.Printf("Scheduler: '%s' completed (HTTP %d)", sched.Name, statusCode)
+		}
 	}

 	// #795: detect phantom-producing schedules — cron fires successfully
@@ -479,6 +498,54 @@ func (s *Scheduler) fireSchedule(ctx context.Context, sched scheduleRow) {
 		resetCancel()
 	}

+	// #1696: track consecutive SDK errors. When the adapter returns HTTP 200
+	// but the response body signals a non-ok result_kind (rate_limited,
+	// sdk_error, quota_exhausted), we increment a counter. After 3 consecutive
+	// SDK errors we auto-disable the schedule and log it — the schedule is
+	// suffering a persistent LLM-layer failure and firing it again will keep
+	// producing the same errors while burning tokens.
+	//
+	// Only apply when the current lastStatus is a non-ok resultKind (not when
+	// we already have 'error' from proxyErr or non-2xx HTTP status — those have
+	// their own failure semantics). Also skip when lastStatus is 'stale' (the
+	// empty-response escalation path takes priority).
+	var consecSDK int
+	if resultKind != "" && resultKind != "ok" {
+		sdkCtx, sdkCancel := context.WithTimeout(context.Background(), dbQueryTimeout)
+		if err := db.DB.QueryRowContext(sdkCtx, `
+			UPDATE workspace_schedules
+			SET consecutive_sdk_errors = consecutive_sdk_errors + 1,
+			    updated_at = now()
+			WHERE id = $1
+			RETURNING consecutive_sdk_errors`, sched.ID).Scan(&consecSDK); err != nil {
+			log.Printf("Scheduler: '%s' SDK-error bump failed: %v", sched.Name, err)
+		}
+		sdkCancel()
+		if consecSDK >= 3 {
+			log.Printf("Scheduler: '%s' AUTO-DISABLING after %d consecutive SDK errors (workspace %s)",
+				sched.Name, consecSDK, short(sched.WorkspaceID, 12))
+			autoDisableCtx, autoDisableCancel := context.WithTimeout(context.Background(), dbQueryTimeout)
+			_, _ = db.DB.ExecContext(autoDisableCtx, `
+				UPDATE workspace_schedules SET enabled = false, updated_at = now() WHERE id = $1 AND enabled = true`,
+				sched.ID)
+			autoDisableCancel()
+		}
+	} else {
+		// Non-SDK-error run — reset the counter.
+		// Guard: only reset when lastStatus is a clean ok (not 'stale', not
+		// 'error', not resultKind). An 'ok' resultKind means the SDK is fine
+		// and we should clear the streak.
+		if lastStatus == "ok" {
+			resetCtx, resetCancel := context.WithTimeout(context.Background(), dbQueryTimeout)
+			_, _ = db.DB.ExecContext(resetCtx, `
+				UPDATE workspace_schedules
+				SET consecutive_sdk_errors = 0,
+				    updated_at = now()
+				WHERE id = $1`, sched.ID)
+			resetCancel()
+		}
+	}
+
 	nextRun, nextErr := ComputeNextRun(sched.CronExpr, sched.Timezone, time.Now())
 	var nextRunPtr *time.Time
 	if nextErr == nil {
@@ -759,6 +826,73 @@ func (s *Scheduler) sweepPhantomBusy(ctx context.Context) {
 	}
 }

+// detectResultKind inspects an A2A response body for SDK-layer error signals
+// that are invisible at the HTTP level. The claude-code-sdk adapter returns
+// HTTP 200 even when the inner LLM call throws (Max-plan rate-limit, quota
+// exhaustion, SDK internal errors) — the error surfaces only in the response
+// body under result.kind or result.result_kind.
+//
+// Returns an empty string when the response is clean (result_kind is "ok" or
+// absent). Returns the result_kind value when it is a non-ok signal, so callers
+// can propagate it as the schedule's last_status.
+//
+// Known non-ok kinds:
+//   - "rate_limited" — LLM API rate-limit hit (Max-plan, etc.)
+//   - "quota_exhausted" — quota / budget exhausted
+//   - "sdk_error" — SDK threw an internal error
+//
+// See #1696.
+func detectResultKind(body []byte) string {
+	if len(body) == 0 {
+		return ""
+	}
+	var top map[string]json.RawMessage
+	if err := json.Unmarshal(body, &top); err != nil {
+		return ""
+	}
+	// Check result.kind first (canonical JSON-RPC shape).
+	if rawResult, ok := top["result"]; ok {
+		var result map[string]json.RawMessage
+		if err := json.Unmarshal(rawResult, &result); err == nil {
+			// result.kind (canonical JSON-RPC error envelope field).
+			if rawKind, ok := result["kind"]; ok {
+				var k string
+				if json.Unmarshal(rawKind, &k) == nil && k != "" && k != "ok" {
+					return k
+				}
+			}
+			// result.result_kind (legacy / alternative field name).
+			if rawKind, ok := result["result_kind"]; ok {
+				var k string
+				if json.Unmarshal(rawKind, &k) == nil && k != "" && k != "ok" {
+					return k
+				}
+			}
+		}
+	}
+	// Top-level error: non-ok HTTP 200 with a structured error in the body.
+	if rawErr, ok := top["error"]; ok {
+		var errMsg string
+		if err := json.Unmarshal(rawErr, &errMsg); err == nil && errMsg != "" {
+			// Distinguish SDK errors from other errors. SDK-layer errors from the
+			// Claude Code runtime include specific markers.
+			lower := strings.ToLower(errMsg)
+			// Check more specific patterns first (max-plan quota > general rate).
+			if strings.Contains(lower, "max-plan") || strings.Contains(lower, "quota") || strings.Contains(lower, "budget") {
+				return "quota_exhausted"
+			}
+			if strings.Contains(lower, "rate limit") || strings.Contains(lower, "rate_limit") {
+				return "rate_limited"
+			}
+			if strings.Contains(lower, "claude code returned an error") || strings.Contains(lower, "sdk error") ||
+				strings.Contains(lower, "api key") || strings.Contains(lower, "authentication") {
+				return "sdk_error"
+			}
+		}
+	}
+	return ""
+}
+
 // isEmptyResponse checks if an A2A response body indicates the agent
 // produced no meaningful output. Catches "(no response generated)" from
 // the workspace runtime + genuinely empty/null responses. Used by the
@@ -808,6 +942,32 @@ func isEmptyResponse(body []byte) bool {
 	return false
 }

+// a2aErrorFromBody extracts an A2A/JSON-RPC error message from a 2xx
+// response body. The adapter SDK may return HTTP 200 with an error
+// payload when it throws internally; this prevents the scheduler from
+// falsely recording last_status='ok'.
+// Issue #1696.
+func a2aErrorFromBody(body []byte) string {
+	if len(body) == 0 {
+		return ""
+	}
+	var resp map[string]interface{}
+	if json.Unmarshal(body, &resp) != nil {
+		return ""
+	}
+	// JSON-RPC style: {"error":{"code":-32603,"message":"..."}}
+	if errObj, ok := resp["error"].(map[string]interface{}); ok {
+		if msg, ok := errObj["message"].(string); ok {
+			return msg
+		}
+	}
+	// Plain style: {"error":"..."}
+	if errStr, ok := resp["error"].(string); ok {
+		return errStr
+	}
+	return ""
+}
+
 // truncation moved to internal/textutil.TruncateBytes (#2962 SSOT).
 // The original #2026 fix lives in textutil's package docs as canonical
 // prior art. Ellipsis was previously "..." (3 ASCII bytes); the SSOT
@@ -3,6 +3,7 @@ package scheduler
 import (
 	"context"
 	"database/sql"
+	"encoding/json"
 	"testing"
 	"time"
 	"unicode/utf8"
@@ -256,6 +257,58 @@ func (p *successProxy) ProxyA2ARequest(
 	return 200, []byte(`{"ok":true}`), nil
 }

+// ── adapterErrorProxy ─────────────────────────────────────────────────────────
+
+// adapterErrorProxy is a test double whose ProxyA2ARequest returns HTTP 200
+// with a JSON-RPC error body, simulating an adapter SDK that throws internally
+// but still completes the HTTP round-trip. Issue #1696.
+type adapterErrorProxy struct{}
+
+func (p *adapterErrorProxy) ProxyA2ARequest(
+	_ context.Context, _ string, _ []byte, _ string, _ bool,
+) (int, []byte, error) {
+	return 200, []byte(`{"jsonrpc":"2.0","id":"cron-test-123","error":{"code":-32603,"message":"adapter SDK internal error"}}`), nil
+}
+
+// ── TestFireSchedule_AdapterSDKError (#1696) ──────────────────────────────────
+//
+// When the adapter SDK throws internally and returns HTTP 200 with an error
+// payload, fireSchedule must record last_status='error', not 'ok'.
+
+func TestFireSchedule_AdapterSDKError(t *testing.T) {
+	mock := setupTestDB(t)
+
+	sched := scheduleRow{
+		ID:          "55555555-dead-beef-0000-000000000005",
+		WorkspaceID: "66666666-dead-beef-0000-000000000006",
+		Name:        "adapter-err-job",
+		CronExpr:    "0 * * * *",
+		Timezone:    "UTC",
+		Prompt:      "do something",
+	}
+
+	// active_tasks check → 0 (workspace is idle; proceed to fire)
+	mock.ExpectQuery(`SELECT COALESCE`).
+		WillReturnRows(sqlmock.NewRows([]string{"coalesce"}).AddRow(0))
+
+	// Post-fire UPDATE must record last_status='error' with the adapter error message.
+	mock.ExpectExec(`UPDATE workspace_schedules`).
+		WithArgs(sched.ID, sqlmock.AnyArg(), "error", "A2A adapter error: adapter SDK internal error").
+		WillReturnResult(sqlmock.NewResult(0, 1))
+
+	// activity_logs INSERT must carry status='error' and the error detail.
+	mock.ExpectExec(`INSERT INTO activity_logs`).
+		WithArgs(sched.WorkspaceID, sqlmock.AnyArg(), sqlmock.AnyArg(), "error", "A2A adapter error: adapter SDK internal error").
+		WillReturnResult(sqlmock.NewResult(0, 1))
+
+	s := New(&adapterErrorProxy{}, nil)
+	s.fireSchedule(context.Background(), sched)
+
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Errorf("unmet DB expectations — adapter error not recorded correctly: %v", err)
+	}
+}
+
 // ── TestFireSchedule_ComputeNextRunError (#722 Bug 1) ─────────────────────────
 //
 // When ComputeNextRun fails (bad cron expression), fireSchedule must NOT write
@@ -285,6 +338,12 @@ func TestFireSchedule_ComputeNextRunError(t *testing.T) {
 		WithArgs(sched.ID).
 		WillReturnResult(sqlmock.NewResult(0, 1))

+	// #1696 consecutive_sdk_errors reset — successProxy has no result_kind,
+	// so detectResultKind returns "" and lastStatus="ok" → reset.
+	mock.ExpectExec(`UPDATE workspace_schedules`).
+		WithArgs(sched.ID).
+		WillReturnResult(sqlmock.NewResult(0, 1))
+
 	// UPDATE must fire — COALESCE($2, next_run_at) keeps existing value when $2 is nil.
 	// AnyArg for $2 because it will be nil (ComputeNextRun failed).
 	mock.ExpectExec(`UPDATE workspace_schedules`).
@@ -540,7 +599,14 @@ func TestFireSchedule_NormalSuccess_AdvancesNextRunAt(t *testing.T) {
 		WithArgs(sched.ID).
 		WillReturnResult(sqlmock.NewResult(0, 1))

-	// 3. Normal UPDATE after successful proxy call.
+	// 3. #1696 consecutive_sdk_errors reset — successProxy response has no
+	//    result_kind in the body, so detectResultKind returns "" and lastStatus
+	//    is "ok" → we hit the SDK-error counter reset branch.
+	mock.ExpectExec(`UPDATE workspace_schedules`).
+		WithArgs(sched.ID).
+		WillReturnResult(sqlmock.NewResult(0, 1))
+
+	// 4. Normal UPDATE after successful proxy call.
 	//    Args: $1=sched.ID, $2=nextRunPtr (computed time), $3=lastStatus, $4=lastError
 	mock.ExpectExec(`UPDATE workspace_schedules`).
 		WithArgs(sched.ID, sqlmock.AnyArg(), "ok", "").
@@ -603,6 +669,281 @@ func TestRecordSkipped_AdvancesNextRunAt(t *testing.T) {
 }
 // trigger CI

+// ── TestDetectResultKind ───────────────────────────────────────────────────────
+
+// TestDetectResultKind covers the SDK error detection path: HTTP 200 responses
+// with non-ok result_kind in the body must be recognised and returned as the
+// kind string, not silently treated as ok.
+func TestDetectResultKind(t *testing.T) {
+	// The test exercises detectResultKind directly so we don't need a full
+	// fireSchedule mock for this unit-test level.
+	tests := []struct {
+		name     string
+		body     string
+		wantKind string
+	}{
+		{
+			name:     "clean ok response — empty body",
+			body:     `{}`,
+			wantKind: "",
+		},
+		{
+			name:     "clean ok response — result.kind absent",
+			body:     `{"result":{"parts":[{"text":"hello"}]}}`,
+			wantKind: "",
+		},
+		{
+			name:     "clean ok response — result.kind=ok",
+			body:     `{"result":{"kind":"ok","parts":[{"text":"hello"}]}}`,
+			wantKind: "",
+		},
+		{
+			name:     "clean ok response — result.result_kind=ok",
+			body:     `{"result":{"result_kind":"ok","parts":[{"text":"hello"}]}}`,
+			wantKind: "",
+		},
+		{
+			name:     "SDK error — result.kind=rate_limited",
+			body:     `{"result":{"kind":"rate_limited","parts":[{"text":"error"}]}}`,
+			wantKind: "rate_limited",
+		},
+		{
+			name:     "SDK error — result.kind=quota_exhausted",
+			body:     `{"result":{"kind":"quota_exhausted"}}`,
+			wantKind: "quota_exhausted",
+		},
+		{
+			name:     "SDK error — result.kind=sdk_error",
+			body:     `{"result":{"kind":"sdk_error"}}`,
+			wantKind: "sdk_error",
+		},
+		{
+			name:     "SDK error — result.result_kind=rate_limited",
+			body:     `{"result":{"result_kind":"rate_limited"}}`,
+			wantKind: "rate_limited",
+		},
+		{
+			name:     "SDK error — error string with rate limit",
+			body:     `{"result":{"parts":[]},"error":"An error occurred: rate limit exceeded"}`,
+			wantKind: "rate_limited",
+		},
+		{
+			name:     "SDK error — error string with max-plan",
+			body:     `{"error":"Max-plan rate limit reached"}`,
+			wantKind: "quota_exhausted",
+		},
+		{
+			name:     "SDK error — error string with quota",
+			body:     `{"error":"quota exhausted for model"}`,
+			wantKind: "quota_exhausted",
+		},
+		{
+			name:     "SDK error — error string with sdk error",
+			body:     `{"error":"Claude Code returned an error result: success"}`,
+			wantKind: "sdk_error",
+		},
+		{
+			name:     "SDK error — error string with api key",
+			body:     `{"error":"invalid API key"}`,
+			wantKind: "sdk_error",
+		},
+		{
+			name:     "unknown error string — not an SDK error",
+			body:     `{"error":"something went wrong"}`,
+			wantKind: "",
+		},
+		{
+			name:     "empty response body",
+			body:     ``,
+			wantKind: "",
+		},
+		{
+			name:     "malformed JSON",
+			body:     `not valid json`,
+			wantKind: "",
+		},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			got := detectResultKind([]byte(tc.body))
+			if got != tc.wantKind {
+				t.Errorf("detectResultKind(%q) = %q, want %q", tc.body, got, tc.wantKind)
+			}
+		})
+	}
+}
+
+// ── TestFireSchedule_SDKError_RateLimited (#1696) ───────────────────────────────
+//
+// When ProxyA2ARequest returns HTTP 200 but the response body contains a
+// non-ok result_kind, fireSchedule must:
+//   1. Set last_status to the result_kind (not 'ok').
+//   2. Set last_error to describe the SDK error.
+//   3. Increment consecutive_sdk_errors.
+//   4. NOT auto-disable on first occurrence (threshold is 3).
+//
+// This test uses an sdkErrorProxy that returns a rate-limited body and asserts
+// the first run is recorded as 'rate_limited' with consecutive_sdk_errors=1
+// and enabled=true.
+func TestFireSchedule_SDKError_RateLimited(t *testing.T) {
+	mock := setupTestDB(t)
+
+	sched := scheduleRow{
+		ID:          "sdk1-test-sched-0001",
+		WorkspaceID: "sdk1-test-workspace1",
+		Name:        "rate-limited-job",
+		CronExpr:    "0 * * * *",
+		Timezone:    "UTC",
+		Prompt:      "do work",
+	}
+
+	// 1. active_tasks check → workspace idle
+	mock.ExpectQuery(`SELECT COALESCE`).
+		WillReturnRows(sqlmock.NewRows([]string{"coalesce"}).AddRow(0))
+
+	// 2. #1696 consecutive_sdk_errors bump — RETURNING gives us count=1.
+	//    Use ExpectQuery (not Exec) because QueryRowContext + RETURNING
+	//    produces a result set consumed via .Scan().
+	mock.ExpectQuery(`UPDATE workspace_schedules`).
+		WithArgs(sched.ID).
+		WillReturnRows(sqlmock.NewRows([]string{"consecutive_sdk_errors"}).AddRow(1))
+
+	// 3. Post-fire UPDATE — last_status='rate_limited', last_error='SDK error: result_kind=rate_limited'
+	mock.ExpectExec(`UPDATE workspace_schedules`).
+		WithArgs(sched.ID, sqlmock.AnyArg(), "rate_limited", sqlmock.AnyArg()).
+		WillReturnResult(sqlmock.NewResult(0, 1))
+
+	// 4. activity_logs INSERT
+	mock.ExpectExec(`INSERT INTO activity_logs`).
+		WithArgs(sched.WorkspaceID, sqlmock.AnyArg(), sqlmock.AnyArg(), "rate_limited", sqlmock.AnyArg()).
+		WillReturnResult(sqlmock.NewResult(0, 1))
+
+	s := New(&sdkErrorProxy{kind: "rate_limited"}, nil)
+	s.fireSchedule(context.Background(), sched)
+
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Errorf("unmet DB expectations for SDK-error first run: %v", err)
+	}
+}
+
+// ── TestFireSchedule_SDKError_AutoDisableOnThirdConsecutive (#1696) ───────────
+//
+// On the 3rd consecutive SDK error, fireSchedule must auto-disable the
+// schedule (enabled=false) in addition to recording the error status.
+// Threshold is 3 per #1696 requirement.
+func TestFireSchedule_SDKError_AutoDisableOnThirdConsecutive(t *testing.T) {
+	mock := setupTestDB(t)
+
+	sched := scheduleRow{
+		ID:          "sdk2-test-sched-0002",
+		WorkspaceID: "sdk2-test-workspace2",
+		Name:        "auto-disable-job",
+		CronExpr:    "0 * * * *",
+		Timezone:    "UTC",
+		Prompt:      "do work",
+	}
+
+	// 1. active_tasks check → workspace idle
+	mock.ExpectQuery(`SELECT COALESCE`).
+		WillReturnRows(sqlmock.NewRows([]string{"coalesce"}).AddRow(0))
+
+	// 2. #1696 consecutive_sdk_errors bump — RETURNING gives count=3 (threshold met).
+	//    Use ExpectQuery (not Exec) because QueryRowContext + RETURNING
+	//    produces a result set consumed via .Scan().
+	mock.ExpectQuery(`UPDATE workspace_schedules`).
+		WithArgs(sched.ID).
+		WillReturnRows(sqlmock.NewRows([]string{"consecutive_sdk_errors"}).AddRow(3))
+
+	// 3. Auto-disable UPDATE — sets enabled=false (schedule has hit 3rd SDK error)
+	mock.ExpectExec(`UPDATE workspace_schedules SET enabled`).
+		WithArgs(sched.ID).
+		WillReturnResult(sqlmock.NewResult(0, 1))
+
+	// 4. Post-fire UPDATE
+	mock.ExpectExec(`UPDATE workspace_schedules`).
+		WithArgs(sched.ID, sqlmock.AnyArg(), "rate_limited", sqlmock.AnyArg()).
+		WillReturnResult(sqlmock.NewResult(0, 1))
+
+	// 5. activity_logs INSERT
+	mock.ExpectExec(`INSERT INTO activity_logs`).
+		WithArgs(sched.WorkspaceID, sqlmock.AnyArg(), sqlmock.AnyArg(), "rate_limited", sqlmock.AnyArg()).
+		WillReturnResult(sqlmock.NewResult(0, 1))
+
+	s := New(&sdkErrorProxy{kind: "rate_limited"}, nil)
+	s.fireSchedule(context.Background(), sched)
+
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Errorf("unmet DB expectations for SDK-error auto-disable: %v", err)
+	}
+}
+
+// ── TestFireSchedule_SDKError_CounterResetOnCleanRun (#1696) ──────────────────
+//
+// A clean HTTP-200 run (no SDK error) must reset consecutive_sdk_errors to 0.
+// This prevents false auto-disable after intermittent SDK blips.
+func TestFireSchedule_SDKError_CounterResetOnCleanRun(t *testing.T) {
+	mock := setupTestDB(t)
+
+	sched := scheduleRow{
+		ID:          "sdk3-test-sched-0003",
+		WorkspaceID: "sdk3-test-workspace3",
+		Name:        "clean-reset-job",
+		CronExpr:    "30 * * * *",
+		Timezone:    "UTC",
+		Prompt:      "do work",
+	}
+
+	// 1. active_tasks check → workspace idle
+	mock.ExpectQuery(`SELECT COALESCE`).
+		WillReturnRows(sqlmock.NewRows([]string{"coalesce"}).AddRow(0))
+
+	// 2. No SDK error — #1696 counter is reset to 0
+	//    (lastStatus is 'ok', resultKind is empty, so we go to the reset branch)
+	mock.ExpectExec(`UPDATE workspace_schedules`).
+		WithArgs(sched.ID).
+		WillReturnResult(sqlmock.NewResult(0, 1))
+
+	// 3. Post-fire UPDATE
+	mock.ExpectExec(`UPDATE workspace_schedules`).
+		WithArgs(sched.ID, sqlmock.AnyArg(), "ok", "").
+		WillReturnResult(sqlmock.NewResult(0, 1))
+
+	// 4. activity_logs INSERT
+	mock.ExpectExec(`INSERT INTO activity_logs`).
+		WithArgs(sched.WorkspaceID, sqlmock.AnyArg(), sqlmock.AnyArg(), "ok", "").
+		WillReturnResult(sqlmock.NewResult(0, 1))
+
+	s := New(&successProxy{}, nil)
+	s.fireSchedule(context.Background(), sched)
+
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Errorf("unmet DB expectations for SDK-error counter reset: %v", err)
+	}
+}
+
+// ── sdkErrorProxy ──────────────────────────────────────────────────────────────
+//
+// sdkErrorProxy is a test double whose ProxyA2ARequest returns HTTP 200 but
+// embeds a non-ok result_kind in the response body, simulating a Claude Code
+// SDK that returned 200 but the inner LLM call threw a rate-limit / quota error.
+// Used by TestFireSchedule_SDKError_* to cover #1696 SDK error detection.
+type sdkErrorProxy struct {
+	kind string // result_kind value to embed in the response body
+}
+
+func (p *sdkErrorProxy) ProxyA2ARequest(
+	_ context.Context, _ string, _ []byte, _ string, _ bool,
+) (int, []byte, error) {
+	body, _ := json.Marshal(map[string]interface{}{
+		"result": map[string]interface{}{
+			"kind":  p.kind,
+			"parts": []map[string]interface{}{{"kind": "text", "text": "(no response generated)"}},
+		},
+	})
+	return 200, body, nil
+}
+
 // ── TestTruncate_utf8Safe_regression2026 ──────────────────────────────────────

 // TestTruncate_utf8Safe_regression2026 locks in the #2026 fix: truncate must
@@ -0,0 +1 @@
+ALTER TABLE workspaces DROP COLUMN IF EXISTS compute;
@@ -0,0 +1,2 @@
+ALTER TABLE workspaces
+    ADD COLUMN IF NOT EXISTS compute JSONB NOT NULL DEFAULT '{}'::jsonb;
@@ -0,0 +1,4 @@
+-- migration: 20260523000000_schedule_consecutive_sdk_errors.down.sql
+-- Reverts #1696 fix for #1696 (consecutive_sdk_errors column)
+
+ALTER TABLE workspace_schedules DROP COLUMN IF EXISTS consecutive_sdk_errors;
@@ -0,0 +1,20 @@
+-- migration: 20260523000000_schedule_consecutive_sdk_errors.up.sql
+-- Fixes #1696: Add consecutive_sdk_errors counter to track SDK errors (HTTP 200
+-- responses where the Claude Code runtime returned a non-ok result_kind).
+-- When this counter reaches 3, the scheduler sets last_status='rate_limited'
+-- and auto-disables the schedule.
+--
+-- The core issue: the claude-code-sdk adapter returns HTTP 200 even when the
+-- inner LLM call throws (e.g. Max-plan rate-limit). All 3 observed runs logged
+-- "completed (HTTP 200)" yet surfaced agent errors in the workspace chat.
+-- This counter lets us detect that pattern and escalate appropriately.
+
+ALTER TABLE workspace_schedules
+    ADD COLUMN IF NOT EXISTS consecutive_sdk_errors INTEGER NOT NULL DEFAULT 0;
+
+COMMENT ON COLUMN workspace_schedules.consecutive_sdk_errors IS
+    'Count of consecutive scheduler fires where ProxyA2ARequest returned HTTP 200
+     but the response body contained a non-ok result_kind (e.g. rate_limited,
+     sdk_error, quota_exhausted). Reset to 0 on any non-SDK-error status.
+     After 3 consecutive SDK errors the schedule is auto-disabled with
+     status rate_limited. Fixes #1696.';
@@ -0,0 +1,2 @@
+DROP INDEX IF EXISTS idx_workspace_display_control_locks_expires;
+DROP TABLE IF EXISTS workspace_display_control_locks;
@@ -0,0 +1,11 @@
+CREATE TABLE IF NOT EXISTS workspace_display_control_locks (
+    workspace_id uuid PRIMARY KEY REFERENCES workspaces(id) ON DELETE CASCADE,
+    controller text NOT NULL CHECK (controller IN ('user', 'agent')),
+    controlled_by text NOT NULL CHECK (length(controlled_by) > 0 AND length(controlled_by) <= 200),
+    expires_at timestamptz NOT NULL,
+    created_at timestamptz NOT NULL DEFAULT now(),
+    updated_at timestamptz NOT NULL DEFAULT now()
+);
+
+CREATE INDEX IF NOT EXISTS idx_workspace_display_control_locks_expires
+    ON workspace_display_control_locks (expires_at);
				`@@ -0,0 +1 @@`
				`ALTER TABLE workspaces DROP COLUMN IF EXISTS compute;`