test(e2e): support empty auth headers on mac bash

Merge pull request 'chore(ci): mirror tenant image to staging ecr' (#1647 ) from chore/mirror-tenant-image-staging-ecr into main
chore(ci): mirror tenant image to staging ecr\n\nAdds optional staging ECR tags to the tenant image publish build. The primary publish path remains unchanged when staging publisher secrets are absent.
2026-05-21 12:41:04 -07:00 · 2026-05-21 19:39:14 +00:00 · 2026-05-21 12:31:54 -07:00 · 2026-05-21 18:38:11 +00:00 · 2026-05-21 11:22:16 -07:00 · 2026-05-21 18:21:40 +00:00
63 changed files with 3633 additions and 394 deletions
@@ -0,0 +1,174 @@
+#!/usr/bin/env python3
+"""Shared path-filter helper for Gitea Actions workflows.
+
+Computes changed files against the PR base SHA or push-before SHA and writes
+boolean outputs to GITHUB_OUTPUT. If the diff base is missing or untrusted, the
+helper fails open by setting every output in the selected profile to true.
+"""
+
+from __future__ import annotations
+
+import argparse
+import os
+import re
+import subprocess
+import sys
+from pathlib import Path
+
+
+PROFILES: dict[str, dict[str, str]] = {
+    "ci": {
+        "platform": r"^workspace-server/",
+        "canvas": r"^canvas/",
+        "python": r"^workspace/",
+        "scripts": r"^tests/e2e/|^scripts/|^infra/scripts/",
+    },
+    "handlers-postgres": {
+        "handlers": (
+            r"^workspace-server/internal/handlers/"
+            r"|^workspace-server/internal/wsauth/"
+            r"|^workspace-server/migrations/"
+            r"|^\.gitea/workflows/handlers-postgres-integration\.yml$"
+        ),
+    },
+    "e2e-api": {
+        "api": r"^workspace-server/|^tests/e2e/|^\.gitea/workflows/e2e-api\.yml$",
+    },
+}
+
+
+def classify(profile: str, paths: list[str]) -> dict[str, bool]:
+    patterns = PROFILES[profile]
+    return {
+        name: any(re.search(pattern, path) for path in paths)
+        for name, pattern in patterns.items()
+    }
+
+
+def all_true(profile: str) -> dict[str, bool]:
+    return {name: True for name in PROFILES[profile]}
+
+
+def resolve_base(event_name: str, pr_base_sha: str, push_before: str) -> str:
+    if event_name == "pull_request" and pr_base_sha:
+        return pr_base_sha
+    return push_before
+
+
+def is_zero_sha(value: str) -> bool:
+    return not value or bool(re.fullmatch(r"0+", value))
+
+
+def run_git(args: list[str], *, timeout: int = 30) -> subprocess.CompletedProcess[str]:
+    return subprocess.run(
+        ["git", *args],
+        check=False,
+        text=True,
+        stdout=subprocess.PIPE,
+        stderr=subprocess.PIPE,
+        timeout=timeout,
+    )
+
+
+def base_exists(base: str) -> bool:
+    return run_git(["cat-file", "-e", base]).returncode == 0
+
+
+def fetch_base(base: str, base_ref: str) -> None:
+    # Gitea may reject fetching an arbitrary unadvertised SHA from a shallow
+    # PR checkout. Fetch the advertised base branch first, then fall back to
+    # the SHA for hosts that allow it.
+    if base_ref:
+        run_git(["fetch", "--depth=1", "origin", base_ref])
+    if not base_exists(base):
+        run_git(["fetch", "--depth=1", "origin", base])
+
+
+def deepen_base_ref(base_ref: str) -> None:
+    if base_ref:
+        run_git(["fetch", "--deepen=200", "origin", base_ref], timeout=60)
+
+
+def merge_base(base: str) -> str | None:
+    proc = run_git(["merge-base", base, "HEAD"])
+    if proc.returncode != 0:
+        return None
+    value = proc.stdout.strip()
+    return value or None
+
+
+def changed_paths(base: str, *, use_merge_base: bool) -> list[str] | None:
+    compare_base = base
+    if use_merge_base:
+        compare_base = merge_base(base) or ""
+        if not compare_base:
+            return None
+
+    proc = run_git(["diff", "--name-only", compare_base, "HEAD"])
+    if proc.returncode != 0:
+        return None
+    return [line for line in proc.stdout.splitlines() if line]
+
+
+def write_outputs(values: dict[str, bool], output_path: str | None) -> None:
+    lines = [f"{name}={'true' if value else 'false'}" for name, value in values.items()]
+    if output_path:
+        with Path(output_path).open("a", encoding="utf-8") as fh:
+            for line in lines:
+                fh.write(line + "\n")
+    else:
+        for line in lines:
+            print(line)
+
+
+def detect(
+    profile: str,
+    event_name: str,
+    pr_base_sha: str,
+    push_before: str,
+    base_ref: str = "",
+) -> dict[str, bool]:
+    base = resolve_base(event_name, pr_base_sha, push_before)
+    if is_zero_sha(base):
+        return all_true(profile)
+
+    if not base_exists(base):
+        fetch_base(base, base_ref)
+    if not base_exists(base):
+        return all_true(profile)
+
+    use_merge_base = event_name == "pull_request"
+    if use_merge_base and base_ref and merge_base(base) is None:
+        deepen_base_ref(base_ref)
+
+    paths = changed_paths(base, use_merge_base=use_merge_base)
+    if paths is None:
+        return all_true(profile)
+    return classify(profile, paths)
+
+
+def parse_args(argv: list[str]) -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--profile", required=True, choices=sorted(PROFILES))
+    parser.add_argument("--event-name", default=os.environ.get("GITHUB_EVENT_NAME", ""))
+    parser.add_argument("--pr-base-sha", default="")
+    parser.add_argument("--base-ref", default="")
+    parser.add_argument("--push-before", default=os.environ.get("GITHUB_EVENT_BEFORE", ""))
+    return parser.parse_args(argv)
+
+
+def main(argv: list[str]) -> int:
+    args = parse_args(argv)
+    values = detect(
+        args.profile,
+        args.event_name,
+        args.pr_base_sha,
+        args.push_before,
+        args.base_ref,
+    )
+    write_outputs(values, os.environ.get("GITHUB_OUTPUT"))
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main(sys.argv[1:]))
@@ -61,6 +61,7 @@ import os
 import shutil
 import subprocess
 import sys
+import time
 import urllib.error
 import urllib.parse
 import urllib.request
@@ -89,6 +90,19 @@ API = f"https://{GITEA_HOST}/api/v1" if GITEA_HOST else ""
 # match by exact title without parsing.
 TITLE_PREFIX = "[main-red]"

+# Settling window (seconds) between initial red detection and the
+# pre-file recheck. The recheck filters out the two largest false-
+# positive classes seen in mc#1597..1630 (task #394, 2026-05-21):
+#   1. HEAD moved on (a new commit landed mid-tick) — the prior red SHA
+#      is no longer authoritative; let the next cron tick re-evaluate.
+#   2. Combined status recovered on the SAME SHA (transient
+#      cancel-cascade rolled forward to success on retry).
+# 90s is well below the hourly cron cadence; a real failure that
+# persists past it is the one we want surfaced.
+# Override with WATCHDOG_RECHECK_DELAY_SECS for tests / local probes
+# (the test suite stubs time.sleep to a no-op).
+RECHECK_DELAY_SECS = int(_env("WATCHDOG_RECHECK_DELAY_SECS", default="90"))
+

 def _require_runtime_env() -> None:
    """Enforce env contract — called from `main()` only.
@@ -172,6 +186,49 @@ def api(
        return status, {"_raw": raw.decode("utf-8", errors="replace")}


+# --------------------------------------------------------------------------
+# action_run.status resolver — extensibility hook for task #394.
+# --------------------------------------------------------------------------
+def _resolve_action_run_status(target_url: str) -> int | None:
+    """Resolve the underlying Gitea `action_run.status` integer for the
+    run referenced by `target_url`, returning None if the resolver
+    cannot reach an authoritative source from the runner.
+
+    Canonical Gitea 1.22.6 enum (per `models/actions/status.go` +
+    `reference_gitea_action_status_enum_corrected_2026_05_19`):
+        1=Success, 2=Failure, 3=Cancelled, 4=Skipped,
+        5=Waiting,  6=Running, 7=Blocked
+    Only `status == 2` is a real defect; status=3 is cancel-cascade and
+    status=1 is an emission artifact (Gitea wrote a 'failure' commit_status
+    row for a run that actually succeeded — observed empirically on
+    `publish-canvas-image` jobs at SHAs in mc#1597..1630).
+
+    CURRENT STATE (2026-05-20, verified): Gitea 1.22.6 exposes NO REST
+    endpoint for `action_run.status`. Probed:
+        /api/v1/repos/{o}/{r}/actions/runs/{id}   → HTTP 404
+        /api/v1/repos/{o}/{r}/actions/jobs/{id}   → HTTP 404
+        /api/v1/repos/{o}/{r}/actions/tasks/{id}  → HTTP 404
+        /swagger.v1.json paths containing 'actions' → secrets+variables+runners only
+    The SPA backend (`/{repo}/actions/runs/{id}/jobs/{idx}` POST) requires
+    a session CSRF token, unreachable from a runner. The only authoritative
+    source today is direct DB access (`mol_action_status` on op-host,
+    `docker exec molecule-postgres-1 psql ...`), which the runner cannot
+    reach.
+
+    Therefore: this hook returns None on every call. Callers MUST fall
+    back to the description-string filter (existing) plus the HEAD
+    recheck (this PR). When a future Gitea release (>=1.23 expected) or
+    an op-host proxy exposes the endpoint, replace the body of this
+    function with an `api(...)` call — the caller contract is stable.
+
+    See also:
+        - `reference_chronic_red_sweep_cancelled_vs_failed_filter`
+        - `feedback_gitea_status_enum_use_helper_not_raw_int`
+    """
+    _ = target_url  # noqa: F841 — intentional placeholder
+    return None
+
+
 # --------------------------------------------------------------------------
 # Gitea reads
 # --------------------------------------------------------------------------
@@ -614,6 +671,56 @@ def run_once(*, dry_run: bool = False) -> int:
    }

    if red:
+        # HEAD recheck (task #394 — guards mc#1597..1630 false-positive
+        # cluster). After the initial detection, wait RECHECK_DELAY_SECS
+        # (default 90s; tests stub time.sleep) and re-evaluate:
+        #
+        #   1. Re-fetch HEAD SHA. If HEAD moved, a new commit landed
+        #      mid-tick — the prior red SHA is no longer authoritative
+        #      and the next cron run will re-evaluate against the new
+        #      HEAD. Skip-file.
+        #
+        #   2. If HEAD unchanged, re-fetch the combined status. If it
+        #      recovered (combined state no longer in {failure,error}
+        #      after the cancel-cascade filter), a transient retry
+        #      rolled the run forward. Skip-file.
+        #
+        # Both paths emit a Loki event distinguishable from the real
+        # `main_red_detected` so obs queries can track filter activity.
+        # The settling window is well below the hourly cron cadence —
+        # genuine failures persist past it and are surfaced normally.
+        time.sleep(RECHECK_DELAY_SECS)
+
+        recheck_sha = get_head_sha(WATCH_BRANCH)
+        if recheck_sha != sha:
+            emit_loki_event("main_red_skipped_head_drift", sha, [])
+            print(
+                f"::notice::skip-file (HEAD moved): initial red at "
+                f"{sha[:10]} but HEAD is now {recheck_sha[:10]} on "
+                f"{WATCH_BRANCH}; next cron tick will re-evaluate."
+            )
+            return 0
+
+        recheck_status = get_combined_status(sha)
+        recheck_red, recheck_failed = is_red(recheck_status)
+        if not recheck_red:
+            emit_loki_event("main_red_skipped_recovered", sha, [])
+            print(
+                f"::notice::skip-file (recovered after settling): "
+                f"combined state at {sha[:10]} flipped to "
+                f"{recheck_status.get('state')!r} on recheck; "
+                f"initial red was a transient cancel-cascade."
+            )
+            return 0
+
+        # Still red after settling — file/update. Use the recheck data
+        # as authoritative so the issue body reflects the latest state.
+        failed = recheck_failed
+        debug["recheck_combined_state"] = recheck_status.get("state")
+        debug["recheck_failed_contexts"] = [
+            s.get("context") for s in failed
+        ]
+
        failed_ctxs = [s.get("context") for s in failed if s.get("context")]
        emit_loki_event("main_red_detected", sha, failed_ctxs)
        print(f"::warning::main is RED at {sha[:10]} on {WATCH_BRANCH}: "
@@ -47,7 +47,9 @@ What this script does, per `.gitea/workflows/status-reaper.yml` invocation:
         Parse context as `<workflow_name> / <job_name> (push)`.
         Look up workflow_name in the trigger map:
           - missing → log ::notice:: and skip (conservative).
-           - has_push_trigger=True → preserve (real defect signal).
+           - has_push_trigger=True and description == "Has been cancelled"
+             → compensate cancelled/superseded push noise.
+           - has_push_trigger=True otherwise → preserve (real defect signal).
           - has_push_trigger=False → POST a compensating
             `state=success` status to /statuses/{sha} with the same
             context (Gitea de-dups by context) and a description
@@ -141,6 +143,11 @@ PR_SHADOW_COMPENSATION_DESCRIPTION = (
    "shadowed by successful push status on same SHA; see "
    ".gitea/scripts/status-reaper.py)"
 )
+CANCELLED_PUSH_COMPENSATION_DESCRIPTION = (
+    "Compensated by status-reaper (push run was cancelled/superseded; "
+    "Gitea 1.22.6 reports cancelled runs as failure statuses)"
+)
+CANCELLED_DESCRIPTION = "Has been cancelled"

 # Context suffix the reaper acts on. Gitea hardcodes this for ALL
 # default-branch workflow runs.
@@ -476,7 +483,7 @@ def reap(
      {compensated, preserved_real_push, preserved_unknown,
       preserved_non_failure, preserved_non_push_suffix,
       preserved_unparseable, compensated_pr_shadowed_by_push_success,
-       preserved_pr_without_push_success,
+       preserved_pr_without_push_success, compensated_cancelled_push,
       compensated_contexts: [<context>, ...]}

    `compensated_contexts` is rev2-added so `reap_branch` can build
@@ -490,6 +497,7 @@ def reap(
        "preserved_non_push_suffix": 0,
        "preserved_unparseable": 0,
        "compensated_pr_shadowed_by_push_success": 0,
+        "compensated_cancelled_push": 0,
        "preserved_pr_without_push_success": 0,
        "compensated_contexts": [],
    }
@@ -567,8 +575,27 @@ def reap(
            counters["preserved_unknown"] += 1
            continue

+        if (s.get("description") or "").strip() == CANCELLED_DESCRIPTION:
+            # Gitea 1.22.6 maps cancelled action runs to failure commit
+            # statuses. During merge bursts, older push runs can be
+            # superseded and cancelled even though a newer run for the
+            # same branch is the real signal. Compensate only the exact
+            # Gitea cancellation description; real push failures remain red.
+            post_compensating_status(
+                sha,
+                context,
+                s.get("target_url"),
+                description=CANCELLED_PUSH_COMPENSATION_DESCRIPTION,
+                dry_run=dry_run,
+            )
+            counters["compensated"] += 1
+            counters["compensated_cancelled_push"] += 1
+            counters["compensated_contexts"].append(context)
+            continue
+
        if workflow_trigger_map[workflow_name]:
-            # Real push trigger → real defect signal. Preserve.
+            # Real push trigger with a non-cancelled failure description
+            # remains a defect signal. Preserve.
            counters["preserved_real_push"] += 1
            continue

@@ -674,6 +701,7 @@ def reap_branch(
            "preserved_non_push_suffix": 0,
            "preserved_unparseable": 0,
            "compensated_pr_shadowed_by_push_success": 0,
+            "compensated_cancelled_push": 0,
            "preserved_pr_without_push_success": 0,
            "compensated_per_sha": {},
            "skipped": True,
@@ -689,6 +717,7 @@ def reap_branch(
        "preserved_non_push_suffix": 0,
        "preserved_unparseable": 0,
        "compensated_pr_shadowed_by_push_success": 0,
+        "compensated_cancelled_push": 0,
        "preserved_pr_without_push_success": 0,
        "compensated_per_sha": {},
    }
@@ -728,6 +757,7 @@ def reap_branch(
            "preserved_non_push_suffix",
            "preserved_unparseable",
            "compensated_pr_shadowed_by_push_success",
+            "compensated_cancelled_push",
            "preserved_pr_without_push_success",
        ):
            aggregate[key] += per_sha[key]
@@ -86,46 +86,17 @@ jobs:
        with:
          fetch-depth: 0
      - id: check
+        env:
+          PR_BASE_SHA: ${{ github.event.pull_request.base.sha }}
+          PR_BASE_REF: ${{ github.event.pull_request.base.ref }}
+          PUSH_BEFORE: ${{ github.event.before }}
        run: |
-          # For PR events: diff against the base branch (not HEAD~1 of the branch,
-          # which may be unrelated after force-pushes). When a push updates a PR,
-          # both pull_request and push events fire — prefer the PR base so that
-          # the diff is always computed against the actual merge base, not the
-          # previous SHA on the branch which may be on a different history line.
-          BASE="${GITHUB_BASE_REF:-${{ github.event.before }}}"
-          # GITHUB_BASE_REF is set for PR events (the base branch name).
-          # For pull_request events we use the stored base.sha; for push events
-          # (or when base.sha is unavailable) fall back to github.event.before.
-          if [ "${{ github.event_name }}" = "pull_request" ] && [ -n "${{ github.event.pull_request.base.sha }}" ]; then
-            BASE="${{ github.event.pull_request.base.sha }}"
-          fi
-          # Fallback: if BASE is empty or all zeros (new branch), run everything
-          if [ -z "$BASE" ] || echo "$BASE" | grep -qE '^0+$'; then
-            echo "platform=true" >> "$GITHUB_OUTPUT"
-            echo "canvas=true" >> "$GITHUB_OUTPUT"
-            echo "python=true" >> "$GITHUB_OUTPUT"
-            echo "scripts=true" >> "$GITHUB_OUTPUT"
-            exit 0
-          fi
-          # Workflow-only edits are covered by the workflow lint family
-          # and by this workflow's always-present required jobs. Do not fan
-          # those edits out into Go/Canvas/Python/shellcheck work; the
-          # downstream jobs still emit their required contexts via no-op
-          # steps when their surface flag is false.
-          #
-          # If the diff itself cannot be trusted, fail open by running every
-          # surface instead of silently under-testing the PR.
-          if ! DIFF=$(git diff --name-only "$BASE" HEAD 2>/dev/null); then
-            echo "platform=true" >> "$GITHUB_OUTPUT"
-            echo "canvas=true" >> "$GITHUB_OUTPUT"
-            echo "python=true" >> "$GITHUB_OUTPUT"
-            echo "scripts=true" >> "$GITHUB_OUTPUT"
-            exit 0
-          fi
-          echo "platform=$(echo "$DIFF" | grep -qE '^workspace-server/' && echo true || echo false)" >> "$GITHUB_OUTPUT"
-          echo "canvas=$(echo "$DIFF" | grep -qE '^canvas/' && echo true || echo false)" >> "$GITHUB_OUTPUT"
-          echo "python=$(echo "$DIFF" | grep -qE '^workspace/' && echo true || echo false)" >> "$GITHUB_OUTPUT"
-          echo "scripts=$(echo "$DIFF" | grep -qE '^tests/e2e/|^scripts/|^infra/scripts/' && echo true || echo false)" >> "$GITHUB_OUTPUT"
+          python3 .gitea/scripts/detect-changes.py \
+            --profile ci \
+            --event-name "${{ github.event_name }}" \
+            --pr-base-sha "$PR_BASE_SHA" \
+            --base-ref "$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
@@ -133,6 +104,7 @@ jobs:
  # the name match works on PRs that don't touch workspace-server/).
  platform-build:
    name: Platform (Go)
+    needs: changes
    runs-on: ubuntu-latest
    # mc#774 (closed 2026-05-14): Phase 4 flip of the platform-build job.
    # Phase 4 (#656) originally flipped this to continue-on-error: false based on
@@ -153,29 +125,29 @@ jobs:
      run:
        working-directory: workspace-server
    steps:
-      - if: false
+      - if: ${{ github.event_name == 'pull_request' && needs.changes.outputs.platform != 'true' }}
        working-directory: .
-        run: echo "No platform/** changes — skipping real build steps; this job always runs to satisfy the required-check name on branch protection."
-      - if: always()
+        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' }}
        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
-      - if: always()
+      - if: ${{ github.event_name != 'pull_request' || needs.changes.outputs.platform == 'true' }}
        uses: actions/setup-go@40f1582b2485089dde7abd97c1529aa768e1baff # v5
        with:
          go-version: 'stable'
-      - if: always()
+      - if: ${{ github.event_name != 'pull_request' || needs.changes.outputs.platform == 'true' }}
        run: go mod download
-      - if: always()
+      - if: ${{ github.event_name != 'pull_request' || needs.changes.outputs.platform == 'true' }}
        run: go build ./cmd/server
      # CLI (molecli) moved to standalone repo: git.moleculesai.app/molecule-ai/molecule-cli
-      - if: always()
+      - if: ${{ github.event_name != 'pull_request' || needs.changes.outputs.platform == 'true' }}
        run: go vet ./...
-      - if: always()
+      - if: ${{ github.event_name != 'pull_request' || 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: always()
+      - if: ${{ github.event_name != 'pull_request' || needs.changes.outputs.platform == 'true' }}
        name: Run golangci-lint
        run: $(go env GOPATH)/bin/golangci-lint run --timeout 3m ./...
-      - if: always()
+      - if: ${{ github.event_name != 'pull_request' || needs.changes.outputs.platform == 'true' }}
        name: Diagnostic — per-package verbose 60s
        run: |
          set +e
@@ -191,7 +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: always()
+      - if: ${{ github.event_name != 'pull_request' || 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
@@ -199,7 +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: always()
+      - if: ${{ github.event_name != 'pull_request' || 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
@@ -213,7 +185,7 @@ jobs:
                   END {for (f in s) printf "%6.1f%%  %s\n", s[f]/c[f], f}' \
            | sort -n

-      - if: always()
+      - if: ${{ github.event_name != 'pull_request' || 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).
@@ -301,6 +273,7 @@ jobs:
  # siblings — verified empirically on PR #2314).
  canvas-build:
    name: Canvas (Next.js)
+    needs: changes
    runs-on: ubuntu-latest
    timeout-minutes: 20
    # Phase 4 (RFC #219 §1): confirmed green on main 2026-05-12.
@@ -309,20 +282,20 @@ jobs:
      run:
        working-directory: canvas
    steps:
-      - if: false
+      - if: ${{ github.event_name == 'pull_request' && needs.changes.outputs.canvas != 'true' }}
        working-directory: .
-        run: echo "No canvas/** changes — skipping real build steps; this job always runs to satisfy the required-check name on branch protection."
-      - if: always()
+        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' }}
        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
-      - if: always()
+      - if: ${{ github.event_name != 'pull_request' || needs.changes.outputs.canvas == 'true' }}
        uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e # v6.4.0
        with:
          node-version: '22'
-      - if: always()
-        run: rm -f package-lock.json && npm install
-      - if: always()
+      - if: ${{ github.event_name != 'pull_request' || needs.changes.outputs.canvas == 'true' }}
+        run: npm ci --include=optional --prefer-offline
+      - if: ${{ github.event_name != 'pull_request' || needs.changes.outputs.canvas == 'true' }}
        run: npm run build
-      - if: always()
+      - if: ${{ github.event_name != 'pull_request' || 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
@@ -331,7 +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: always()
+        if: ${{ github.event_name != 'pull_request' || 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
@@ -348,15 +321,16 @@ jobs:
  # Shellcheck (E2E scripts) — required check, always runs.
  shellcheck:
    name: Shellcheck (E2E scripts)
+    needs: changes
    runs-on: ubuntu-latest
    # Phase 4 (RFC #219 §1): confirmed green on main 2026-05-12.
    continue-on-error: false
    steps:
-      - if: false
-        run: echo "No tests/e2e/ or infra/scripts/ changes — skipping real shellcheck; this job always runs to satisfy the required-check name on branch protection."
-      - if: always()
+      - 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' }}
        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
-      - if: always()
+      - if: ${{ github.event_name != 'pull_request' || 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
@@ -367,16 +341,16 @@ jobs:
          find tests/e2e infra/scripts -type f -name '*.sh' -print0 \
            | xargs -0 shellcheck --severity=warning

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

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

-      - if: always()
+      - if: ${{ github.event_name != 'pull_request' || 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
@@ -386,7 +360,7 @@ jobs:
        run: |
          bash scripts/test-promote-tenant-image.sh

-      - if: always()
+      - if: ${{ github.event_name != 'pull_request' || 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
@@ -43,6 +43,18 @@ name: Continuous synthetic E2E (staging)

 on:
  schedule:
+    # Every 30 minutes, on :02 and :32. This keeps a recurring SaaS
+    # behavior probe while cutting runner occupancy from this workflow by
+    # roughly two thirds; fast liveness belongs in the lighter smoke/heartbeat
+    # probes, not in a full tenant/workspace synth every 10 minutes.
+    #
+    # Previous cadence was every 10 minutes (:02 :12 :22 :32 :42 :52).
+    # The current operator-host runner pool is the bottleneck, so full
+    # synth E2E is deliberately lower-cadence until it moves to a dedicated
+    # runner host or warm-runtime pool.
+    #
+    # Historical notes from the 10-minute shape:
+    #
    # Every 10 minutes, on :02 :12 :22 :32 :42 :52. Three constraints:
    #   1. Stay off the top-of-hour. GitHub Actions scheduler drops
    #      :00 firings under high load (own docs:
@@ -66,7 +78,7 @@ on:
    # fires = ~30 min cadence; closer to the 20-min target than the
    # current shape and provides a real degradation alarm if drops
    # get worse.
-    - cron: '2,12,22,32,42,52 * * * *'
+    - cron: '2,32 * * * *'
 permissions:
  contents: read
  # No issue-write here — failures surface as red runs in the workflow
@@ -132,31 +132,13 @@ jobs:
        with:
          fetch-depth: 0
      - id: decide
-        # Inline replacement for dorny/paths-filter — same pattern PR#372's
-        # ci.yml port used. Diffs against the PR base or push BEFORE SHA,
-        # then matches against the api-relevant path set.
        run: |
-          BASE="${GITHUB_BASE_REF:-${{ github.event.before }}}"
-          if [ "${{ github.event_name }}" = "pull_request" ] && [ -n "${{ github.event.pull_request.base.sha }}" ]; then
-            BASE="${{ github.event.pull_request.base.sha }}"
-          fi
-          if [ -z "$BASE" ] || echo "$BASE" | grep -qE '^0+$'; then
-            echo "api=true" >> "$GITHUB_OUTPUT"
-            exit 0
-          fi
-          if ! git cat-file -e "$BASE" 2>/dev/null; then
-            git fetch --depth=1 origin "$BASE" 2>/dev/null || true
-          fi
-          if ! git cat-file -e "$BASE" 2>/dev/null; then
-            echo "api=true" >> "$GITHUB_OUTPUT"
-            exit 0
-          fi
-          CHANGED=$(git diff --name-only "$BASE" HEAD)
-          if echo "$CHANGED" | grep -qE '^(workspace-server/|tests/e2e/|\.gitea/workflows/e2e-api\.yml$)'; then
-            echo "api=true" >> "$GITHUB_OUTPUT"
-          else
-            echo "api=false" >> "$GITHUB_OUTPUT"
-          fi
+          python3 .gitea/scripts/detect-changes.py \
+            --profile e2e-api \
+            --event-name "${{ github.event_name }}" \
+            --pr-base-sha "${{ github.event.pull_request.base.sha }}" \
+            --base-ref "${{ github.event.pull_request.base.ref }}" \
+            --push-before "${GITHUB_EVENT_BEFORE:-${{ github.event.before }}}"

  # ONE job (no job-level `if:`) that always runs and reports under the
  # required-check name `E2E API Smoke Test`. Real work is gated per-step
@@ -1,8 +1,10 @@
 name: E2E Chat

 # Comprehensive Playwright E2E for the unified chat stack (desktop
-# ChatTab + mobile MobileChat). Runs on every PR that touches canvas,
-# workspace-server, or this workflow file.
+# ChatTab + mobile MobileChat). Heavy browser execution is intentionally
+# outside the normal required PR path: PRs run it only after entering the
+# `merge-queue`, while push/main, nightly, and manual dispatch preserve
+# coverage without making every PR pay the full runtime/browser cost.
 #
 # Architecture:
 #   1. Ephemeral Postgres + Redis (docker, unique container names)
@@ -22,6 +24,11 @@ on:
    branches: [main, staging]
  pull_request:
    branches: [main, staging]
+  schedule:
+    # Nightly at 09:00 UTC. Keeps coverage for the currently non-required
+    # heavy browser lane without spending runner time on every PR.
+    - cron: '0 9 * * *'
+  workflow_dispatch:

 concurrency:
  group: e2e-chat-${{ github.event.pull_request.head.sha || github.sha }}
@@ -50,7 +57,14 @@ jobs:
        with:
          fetch-depth: 0
      - id: decide
+        env:
+          GITEA_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          QUEUE_LABEL: merge-queue
        run: |
+          if [ "${{ github.event_name }}" = "schedule" ] || [ "${{ github.event_name }}" = "workflow_dispatch" ]; then
+            echo "chat=true" >> "$GITHUB_OUTPUT"
+            exit 0
+          fi
          BASE="${GITHUB_BASE_REF:-${{ github.event.before }}}"
          if [ "${{ github.event_name }}" = "pull_request" ] && [ -n "${{ github.event.pull_request.base.sha }}" ]; then
            BASE="${{ github.event.pull_request.base.sha }}"
@@ -67,9 +81,26 @@ jobs:
            exit 0
          fi
          CHANGED=$(git diff --name-only "$BASE" HEAD)
-          if echo "$CHANGED" | grep -qE '^(canvas/|workspace-server/|\.gitea/workflows/e2e-chat\.yml$)'; then
+          if ! echo "$CHANGED" | grep -qE '^(canvas/|workspace-server/|\.gitea/workflows/e2e-chat\.yml$)'; then
+            echo "chat=false" >> "$GITHUB_OUTPUT"
+            exit 0
+          fi
+          if [ "${{ github.event_name }}" != "pull_request" ]; then
+            echo "chat=true" >> "$GITHUB_OUTPUT"
+            exit 0
+          fi
+
+          authfile=$(mktemp)
+          chmod 600 "$authfile"
+          printf 'header = "Authorization: token %s"\n' "$GITEA_TOKEN" > "$authfile"
+          labels=$(curl -fsS -K "$authfile" \
+            "${{ github.server_url }}/api/v1/repos/${{ github.repository }}/issues/${{ github.event.pull_request.number }}/labels" \
+            | python3 -c 'import json,sys; print("\n".join(label.get("name","") for label in json.load(sys.stdin)))')
+          rm -f "$authfile"
+          if printf '%s\n' "$labels" | grep -qx "$QUEUE_LABEL"; then
            echo "chat=true" >> "$GITHUB_OUTPUT"
          else
+            echo "PR is not in merge-queue; skipping heavy E2E Chat for normal PR path."
            echo "chat=false" >> "$GITHUB_OUTPUT"
          fi

@@ -230,7 +261,14 @@ jobs:
      - name: Install Playwright browsers
        if: needs.detect-changes.outputs.chat == 'true'
        working-directory: canvas
-        run: npx playwright install --with-deps chromium
+        run: |
+          PREBAKED_PLAYWRIGHT=/ms-playwright
+          if [ -d "${PREBAKED_PLAYWRIGHT}" ] && find "${PREBAKED_PLAYWRIGHT}" -maxdepth 3 -type f -name 'chrome' | grep -q .; then
+            echo "Using prebaked Playwright Chromium from ${PREBAKED_PLAYWRIGHT}"
+            echo "PLAYWRIGHT_BROWSERS_PATH=${PREBAKED_PLAYWRIGHT}" >> "$GITHUB_ENV"
+            exit 0
+          fi
+          npx playwright install --with-deps chromium

      - name: Start canvas dev server (background)
        if: needs.detect-changes.outputs.chat == 'true'
@@ -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).
@@ -16,9 +16,9 @@ name: E2E Staging Canvas (Playwright)
 # e2e-staging-saas.yml (which tests the API shape) by exercising the
 # actual browser + canvas bundle against live staging.
 #
-# Triggers: push to main/staging or PR touching canvas sources + this workflow,
-# manual dispatch, and weekly cron to catch browser/runtime drift even
-# when canvas is quiet.
+# Triggers: push to main, PR touching canvas sources + this workflow only
+# after the PR enters `merge-queue`, manual dispatch, and scheduled cron to
+# catch browser/runtime drift even when canvas is quiet.
 # Added staging to push/pull_request branches so the auto-promote gate
 # check (--event push --branch staging) can see a completed run for this
 # workflow — mirrors what PR #1891 does for e2e-api.yml.
@@ -37,9 +37,10 @@ on:
  pull_request:
    branches: [main]
  schedule:
-    # Weekly on Sunday 08:00 UTC — catches Chrome / Playwright / Next.js
+    # Nightly at 08:00 UTC — catches Chrome / Playwright / Next.js
    # release-note-shaped regressions that don't ride in with a PR.
-    - cron: '0 8 * * 0'
+    - cron: '0 8 * * *'
+  workflow_dispatch:

 concurrency:
  # Per-SHA grouping (changed 2026-04-28 from a single global group). The
@@ -79,10 +80,13 @@ jobs:
        with:
          fetch-depth: 0
      - id: decide
+        env:
+          GITEA_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          QUEUE_LABEL: merge-queue
        # Inline replacement for dorny/paths-filter — see e2e-api.yml.
-        # Cron triggers always run real work (no diff context).
+        # Cron and manual triggers always run real work (no diff context).
        run: |
-          if [ "${{ github.event_name }}" = "schedule" ]; then
+          if [ "${{ github.event_name }}" = "schedule" ] || [ "${{ github.event_name }}" = "workflow_dispatch" ]; then
            echo "canvas=true" >> "$GITHUB_OUTPUT"
            exit 0
          fi
@@ -102,9 +106,26 @@ jobs:
            exit 0
          fi
          CHANGED=$(git diff --name-only "$BASE" HEAD)
-          if echo "$CHANGED" | grep -qE '^(canvas/|\.gitea/workflows/e2e-staging-canvas\.yml$)'; then
+          if ! echo "$CHANGED" | grep -qE '^(canvas/|\.gitea/workflows/e2e-staging-canvas\.yml$)'; then
+            echo "canvas=false" >> "$GITHUB_OUTPUT"
+            exit 0
+          fi
+          if [ "${{ github.event_name }}" != "pull_request" ]; then
+            echo "canvas=true" >> "$GITHUB_OUTPUT"
+            exit 0
+          fi
+
+          authfile=$(mktemp)
+          chmod 600 "$authfile"
+          printf 'header = "Authorization: token %s"\n' "$GITEA_TOKEN" > "$authfile"
+          labels=$(curl -fsS -K "$authfile" \
+            "${{ github.server_url }}/api/v1/repos/${{ github.repository }}/issues/${{ github.event.pull_request.number }}/labels" \
+            | python3 -c 'import json,sys; print("\n".join(label.get("name","") for label in json.load(sys.stdin)))')
+          rm -f "$authfile"
+          if printf '%s\n' "$labels" | grep -qx "$QUEUE_LABEL"; then
            echo "canvas=true" >> "$GITHUB_OUTPUT"
          else
+            echo "PR is not in merge-queue; skipping heavy E2E Staging Canvas for normal PR path."
            echo "canvas=false" >> "$GITHUB_OUTPUT"
          fi

@@ -169,7 +190,14 @@ jobs:
      - name: Install Playwright browsers
        if: needs.detect-changes.outputs.canvas == 'true'
        timeout-minutes: 10
-        run: npx playwright install --with-deps chromium
+        run: |
+          PREBAKED_PLAYWRIGHT=/ms-playwright
+          if [ -d "${PREBAKED_PLAYWRIGHT}" ] && find "${PREBAKED_PLAYWRIGHT}" -maxdepth 3 -type f -name 'chrome' | grep -q .; then
+            echo "Using prebaked Playwright Chromium from ${PREBAKED_PLAYWRIGHT}"
+            echo "PLAYWRIGHT_BROWSERS_PATH=${PREBAKED_PLAYWRIGHT}" >> "$GITHUB_ENV"
+            exit 0
+          fi
+          npx playwright install --with-deps chromium

      - name: Run staging canvas E2E
        if: needs.detect-changes.outputs.canvas == 'true'
@@ -13,8 +13,12 @@ name: gitea-merge-queue
 #   - add `merge-queue-hold` to pause a queued PR without removing it

 on:
-  schedule:
-    - cron: '*/5 * * * *'
+  # Schedule moved to operator-config:
+  #   /etc/cron.d/molecule-core-merge-queue ->
+  #   /usr/local/bin/molecule-core-cron-bot.sh merge-queue
+  #
+  # The queue bot still processes one PR per tick, but no longer occupies
+  # one of the shared Actions runners just to poll.
  workflow_dispatch:

 permissions:
@@ -101,36 +101,13 @@ jobs:
          # not present in the shallow checkout.
          fetch-depth: 2
      - id: filter
-        # Inline replacement for dorny/paths-filter — see e2e-api.yml.
        run: |
-          # Gitea Actions evaluates github.event.before to empty string in shell
-          # scripts. Use GITHUB_EVENT_BEFORE shell env var instead (Gitea
-          # correctly populates it for push events). PR case uses template var.
-          BASE=""
-          if [ "${{ github.event_name }}" = "pull_request" ] && [ -n "${{ github.event.pull_request.base.sha }}" ]; then
-            BASE="${{ github.event.pull_request.base.sha }}"
-          elif [ -n "$GITHUB_EVENT_BEFORE" ]; then
-            BASE="$GITHUB_EVENT_BEFORE"
-          fi
-          if [ -z "$BASE" ] || echo "$BASE" | grep -qE '^0+$'; then
-            echo "handlers=true" >> "$GITHUB_OUTPUT"
-            exit 0
-          fi
-          # timeout 30 guards against the case where BASE points to a ref that
-          # git can resolve but cat-file hangs (rare on corrupted objects).
-          if ! timeout 30 git cat-file -e "$BASE" 2>/dev/null; then
-            git fetch --depth=1 origin "$BASE" 2>/dev/null || true
-          fi
-          if ! timeout 30 git cat-file -e "$BASE" 2>/dev/null; then
-            echo "handlers=true" >> "$GITHUB_OUTPUT"
-            exit 0
-          fi
-          CHANGED=$(git diff --name-only "$BASE" HEAD)
-          if echo "$CHANGED" | grep -qE '^(workspace-server/internal/handlers/|workspace-server/internal/wsauth/|workspace-server/migrations/|\.gitea/workflows/handlers-postgres-integration\.yml$)'; then
-            echo "handlers=true" >> "$GITHUB_OUTPUT"
-          else
-            echo "handlers=false" >> "$GITHUB_OUTPUT"
-          fi
+          python3 .gitea/scripts/detect-changes.py \
+            --profile handlers-postgres \
+            --event-name "${{ github.event_name }}" \
+            --pr-base-sha "${{ github.event.pull_request.base.sha }}" \
+            --base-ref "${{ github.event.pull_request.base.ref }}" \
+            --push-before "${GITHUB_EVENT_BEFORE:-}"

  # Single-job-with-per-step-if pattern: always runs to satisfy the
  # required-check name on branch protection; real work gates on the
@@ -25,8 +25,11 @@ 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
+# Optional secrets: AWS_STAGING_ECR_ACCESS_KEY_ID, AWS_STAGING_ECR_SECRET_ACCESS_KEY
 #
 # 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 +68,7 @@ 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:
@@ -135,6 +139,18 @@ jobs:
        run: |
          echo "sha=${GITHUB_SHA::7}" >> "$GITHUB_OUTPUT"

+      # Keep Buildx state inside the job temp dir. The publish runner's
+      # inherited DOCKER_CONFIG can point at a host-owned ECR config path
+      # (/home/hongming/.docker-ecr), which caused setup-buildx-action to
+      # fail before image build with EACCES creating buildx/certs.
+      - name: Prepare writable Docker config
+        run: |
+          set -euo pipefail
+          export DOCKER_CONFIG="$RUNNER_TEMP/docker-config"
+          mkdir -p "$DOCKER_CONFIG/buildx/certs"
+          echo "DOCKER_CONFIG=$DOCKER_CONFIG" >> "$GITHUB_ENV"
+          docker buildx version
+
      # Build + push platform image (inline ECR auth — mirrors the operator-host
      # approach; credentials come from GITHUB_SECRET_AWS_ACCESS_KEY_ID /
      # GITHUB_SECRET_AWS_SECRET_ACCESS_KEY in Gitea Actions).
@@ -170,21 +186,46 @@ jobs:
            --push .

      # Build + push tenant image (Go platform + Next.js canvas in one image).
+      # When staging ECR publisher credentials are configured, push the same
+      # build to the staging account too so fresh staging/E2E tenants can pull
+      # without cross-account ECR permissions.
      - 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 }}
          REPO: ${{ github.repository }}
          AWS_ACCESS_KEY_ID: ${{ secrets.AWS_ACCESS_KEY_ID }}
          AWS_SECRET_ACCESS_KEY: ${{ secrets.AWS_SECRET_ACCESS_KEY }}
+          AWS_STAGING_ECR_ACCESS_KEY_ID: ${{ secrets.AWS_STAGING_ECR_ACCESS_KEY_ID }}
+          AWS_STAGING_ECR_SECRET_ACCESS_KEY: ${{ secrets.AWS_STAGING_ECR_SECRET_ACCESS_KEY }}
          AWS_DEFAULT_REGION: us-east-2
        run: |
          set -euo pipefail
          ECR_REGISTRY="${TENANT_IMAGE_NAME%%/*}"
          aws ecr get-login-password --region us-east-2 | \
            docker login --username AWS --password-stdin "${ECR_REGISTRY}"
+
+          build_tags=(
+            --tag "${TENANT_IMAGE_NAME}:${TAG_SHA}"
+            --tag "${TENANT_IMAGE_NAME}:${TAG_LATEST}"
+          )
+          if [ -n "${AWS_STAGING_ECR_ACCESS_KEY_ID:-}" ] && [ -n "${AWS_STAGING_ECR_SECRET_ACCESS_KEY:-}" ]; then
+            STAGING_ECR_REGISTRY="${STAGING_TENANT_IMAGE_NAME%%/*}"
+            AWS_ACCESS_KEY_ID="${AWS_STAGING_ECR_ACCESS_KEY_ID}" \
+            AWS_SECRET_ACCESS_KEY="${AWS_STAGING_ECR_SECRET_ACCESS_KEY}" \
+              aws ecr get-login-password --region us-east-2 | \
+              docker login --username AWS --password-stdin "${STAGING_ECR_REGISTRY}"
+            build_tags+=(
+              --tag "${STAGING_TENANT_IMAGE_NAME}:${TAG_SHA}"
+              --tag "${STAGING_TENANT_IMAGE_NAME}:${TAG_LATEST}"
+            )
+          else
+            echo "::notice::Skipping staging ECR tenant push; AWS_STAGING_ECR_ACCESS_KEY_ID/AWS_STAGING_ECR_SECRET_ACCESS_KEY are not configured."
+          fi
+
          docker buildx build \
            --file ./workspace-server/Dockerfile.tenant \
            --build-arg NEXT_PUBLIC_PLATFORM_URL= \
@@ -193,8 +234,7 @@ jobs:
            --label "org.opencontainers.image.revision=${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.
@@ -53,19 +53,12 @@ name: status-reaper
 # `inputs:` block here. Gitea 1.22.6 rejects the whole workflow as
 # "unknown on type" when `workflow_dispatch.inputs.X` is present.
 on:
-  # SCHEDULE RE-ENABLED 2026-05-12 rev3 — interim disable (mc#645) reverted now that
-  # rev3 widens DEFAULT_SWEEP_LIMIT 10 → 30 (covers retroactive-failure timing window).
-  # Sibling watchdog re-enabled in the same PR with timeout-minutes raised 5 → 15.
-  schedule:
-    # Every 5 minutes. Off-zero alignment with sibling cron workflows:
-    # ci-required-drift (`:17`), main-red-watchdog (`:05`),
-    # railway-pin-audit (`:23`). 5-min cadence gives a tight enough
-    # close on schedule-triggered false-reds that main-red-watchdog
-    # (hourly :05) almost never files an issue on the false case.
-    # rev3 keeps `*/5` unchanged per hongming-pc2 03:25Z review:
-    # "trades window-width-cheap for cadence-loady" — N=30 widens
-    # the lookback cheaply without doubling runner load via `*/2`.
-    - cron: '*/5 * * * *'
+  # Schedule moved to operator-config:
+  #   /etc/cron.d/molecule-core-status-reaper ->
+  #   /usr/local/bin/molecule-core-cron-bot.sh status-reaper
+  #
+  # This keeps the 5-minute compensation cadence but stops a maintenance
+  # bot from consuming Gitea Actions runner slots during PR merge waves.
  workflow_dispatch:

 # Compensating-status POST needs write on repo statuses; no other
@@ -127,7 +127,11 @@ cd workspace-server && go test -race ./...
 cd canvas && npm test

 # Workspace runtime (Python)
-cd workspace && python -m pytest -v
+# Runtime code is SSOT in molecule-ai-workspace-runtime, not molecule-core/workspace.
+cd ../molecule-ai-workspace-runtime
+python -m venv .venv && source .venv/bin/activate
+pip install --index-url https://git.moleculesai.app/api/packages/molecule-ai/pypi/simple/ -e . pytest pytest-asyncio
+pytest -q

 # E2E API tests (requires running platform)
 bash tests/e2e/test_api.sh
@@ -159,6 +163,19 @@ and run CI manually.
 | review-check-tests | `review-check.sh` evaluator regression suite (13 scenarios) |
 | ops-scripts | Python unittest suite for `scripts/*.py` |

+### Workspace runtime SSOT
+
+Runtime code lives in
+[`molecule-ai-workspace-runtime`](https://git.moleculesai.app/molecule-ai/molecule-ai-workspace-runtime).
+Do not reintroduce `molecule-core/workspace/` or vendored `molecule_runtime/`
+copies in consumers. Core and templates consume the published runtime package
+from the Gitea package registry.
+
+For local external MCP agents, multi-workspace config is
+`MOLECULE_WORKSPACES=[{"id":"...","token":"...","platform_url":"..."}]`.
+`platform_url` selects the tenant; `org_id` is not part of this config.
+Workspace IDs can differ across orgs.
+
 ## Local Testing

 ### review-check.sh
@@ -68,14 +68,103 @@ export function Toolbar() {
    return c;
  }, [nodes]);

+  /**
+   * Stop All - task #377 fix.
+   *
+   * BEFORE this PR: directly POSTed `/workspaces/:id/restart`, which tears
+   * the container down and back up. That kills in-flight tool subprocesses
+   * (e.g. `bash -c 'sleep 600'`) but is heavy and discards any in-progress
+   * agent state. It also bypasses the runtime-side fast cancel path (task
+   * #377 PR#40 in template-claude-code) - meaning flipping
+   * `MOLECULE_STOP_PROPAGATE=true` would produce zero canary signal because
+   * nothing ever invokes `executor.cancel()` in production.
+   *
+   * AFTER this PR (two-phase polite cancel):
+   *
+   * 1. POST `tasks/cancel` (A2A JSON-RPC) to each active workspace's
+   *    `/workspaces/:id/a2a` proxy. The platform proxies the envelope to
+   *    the workspace runtime; the a2a-sdk framework dispatches `tasks/cancel`
+   *    to `AgentExecutor.cancel()` (a2a-sdk 1.0.3
+   *    `a2a/compat/v0_3/types.py` line 1125 pins the wire literal as
+   *    `Literal["tasks/cancel"]`; A2A protocol spec section 9.4.5 maps the
+   *    abstract `CancelTask` operation to that wire string). The runtime's
+   *    executor cancel path signals the CLI subprocess group with
+   *    SIGTERM/grace/SIGKILL (template-claude-code PR#40 `stop_propagate.py`).
+   *
+   * 2. Poll the canvas store (the platform pushes `TASK_UPDATED` over WS
+   *    on `active_tasks` changes - `canvas-events.ts` line 400) for up to
+   *    `STOP_ALL_DRAIN_TIMEOUT_MS`. A workspace whose `activeTasks` drops
+   *    to 0 is considered drained and is NOT restarted.
+   *
+   * 3. For any workspace that DID NOT drain inside the timeout - runtime
+   *    is on an old image without the cancel path, or the cancel
+   *    propagation is stuck - fall back to the original heavy
+   *    `/workspaces/:id/restart`. The original behavior is preserved as a
+   *    floor so a stuck workspace still gets stopped; the polite path is
+   *    a fast top-up that lets well-behaved workspaces cancel without
+   *    losing context.
+   *
+   * The polite-cancel envelope mirrors `ScheduleTab.handleRunNow` (line 168)
+   * which is the only other place in canvas that POSTs `/workspaces/:id/a2a`
+   * directly. Method string `tasks/cancel` and empty `params` match the
+   * a2a-sdk shape verified above. The proxy adds `jsonrpc:"2.0"` and `id`
+   * via `normalizeA2APayload` server-side, so the canvas envelope omits them.
+   */
  const stopAll = useCallback(async () => {
    setStopping(true);
    const active = nodes.filter((n) => (n.data.activeTasks as number) > 0);
+    const activeIds = active.map((n) => n.id);
+
+    // Phase 1 - polite cancel on every active workspace in parallel.
+    // Errors are swallowed (same shape as the pre-fix /restart
+    // Promise.all): a 4xx/5xx on tasks/cancel just means we fall through
+    // to /restart for that workspace below.
    await Promise.all(
-      active.map((n) =>
-        api.post(`/workspaces/${n.id}/restart`).catch(() => {})
+      activeIds.map((id) =>
+        api
+          .post(`/workspaces/${id}/a2a`, {
+            method: "tasks/cancel",
+            params: {},
+          })
+          .catch(() => {})
      )
    );
+
+    // Phase 2 - poll the store for activeTasks reaching 0, with a hard
+    // timeout. STOP_ALL_DRAIN_TIMEOUT_MS is sized to cover the runtime's
+    // own SIGTERM-grace (5s in template-claude-code stop_propagate.py
+    // `_SIGTERM_GRACE_S`) plus a small WS round-trip buffer for the
+    // TASK_UPDATED push. STOP_ALL_POLL_INTERVAL_MS keeps the poll cheap
+    // (no animation jitter, no busy-wait).
+    const STOP_ALL_DRAIN_TIMEOUT_MS = 8000;
+    const STOP_ALL_POLL_INTERVAL_MS = 250;
+    const deadline = Date.now() + STOP_ALL_DRAIN_TIMEOUT_MS;
+    let undrained = new Set(activeIds);
+    while (undrained.size > 0 && Date.now() < deadline) {
+      await new Promise((r) => setTimeout(r, STOP_ALL_POLL_INTERVAL_MS));
+      const fresh = useCanvasStore.getState().nodes;
+      const stillActive = new Set<string>();
+      for (const id of undrained) {
+        const n = fresh.find((x) => x.id === id);
+        // Missing node (workspace deleted mid-cancel) is treated as
+        // drained - there's nothing left to restart and reporting it as
+        // "still running" would be a lie.
+        if (n && (n.data.activeTasks as number) > 0) stillActive.add(id);
+      }
+      undrained = stillActive;
+    }
+
+    // Phase 3 - hard-restart anything that did not drain. This is the
+    // same call shape as the pre-fix Stop All, so behavior is strictly a
+    // superset: undrained workspaces still get the heavy stop, drained
+    // ones are spared.
+    if (undrained.size > 0) {
+      await Promise.all(
+        Array.from(undrained).map((id) =>
+          api.post(`/workspaces/${id}/restart`).catch(() => {})
+        )
+      );
+    }
    setStopping(false);
  }, [nodes]);

@@ -131,14 +131,30 @@ const defaultStore = {
  batchDelete: vi.fn(() => Promise.resolve()),
 };

-vi.mock("@/store/canvas", () => ({
-  useCanvasStore: vi.fn((selector: (s: typeof defaultStore) => unknown) =>
+vi.mock("@/store/canvas", () => {
+  // useCanvasStore is used in two shapes:
+  //   1. As a hook: `useCanvasStore((s) => s.x)` — selector path.
+  //   2. As a static accessor: `useCanvasStore.getState().nodes` —
+  //      used by stopAll's drain-poll loop (task #377 Toolbar fix) and
+  //      restartAll's success-clear loop. Both read the LIVE
+  //      defaultStore object so tests that mutate `defaultStore.nodes`
+  //      mid-flight (e.g. simulating a TASK_UPDATED that drops
+  //      activeTasks to 0) see the update on the next poll tick.
+  const hook = vi.fn((selector: (s: typeof defaultStore) => unknown) =>
    selector(defaultStore)
-  ),
-}));
+  ) as unknown as ((selector: (s: typeof defaultStore) => unknown) => unknown) & {
+    getState: () => typeof defaultStore;
+  };
+  hook.getState = () => defaultStore;
+  return { useCanvasStore: hook };
+});

 // ── Component under test ───────────────────────────────────────────────────────
 import { Toolbar } from "../Toolbar";
+// Imported AFTER vi.mock("@/lib/api", ...) above (hoisted) so this
+// resolves to the mock module; gives the new task #377 tests a typed
+// handle on api.post without a CJS require() (Vitest runs ESM).
+import { api as mockedApi } from "@/lib/api";

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

@@ -315,3 +331,157 @@ describe("Toolbar — ? shortcut opens shortcuts dialog", () => {
    expect(screen.queryByTestId("shortcuts-dialog")).toBeNull();
  });
 });
+
+// ── Toolbar — Stop All polite-cancel flow (task #377) ───────────────────────
+
+describe("Toolbar — Stop All polite cancel before restart (#377)", () => {
+  // `api` resolves to the top-level vi.mock factory's mocked `post`.
+  // We type-cast so TS allows mockReset/mockResolvedValue/mockImplementation
+  // calls without leaking the mock surface into the production type.
+  const api = mockedApi as unknown as { post: ReturnType<typeof vi.fn> };
+
+  /**
+   * Build a working set of two active workspaces so the assertions can
+   * distinguish per-id behavior (drained vs undrained) within one test.
+   */
+  const seedTwoActive = () => {
+    defaultStore.nodes = toStoreNodes(makeNodes(["online", "online"], [2, 2]));
+  };
+
+  /**
+   * Drive an async useCallback handler to completion. Vitest's fake
+   * timers don't see microtasks unless we yield between advances; the
+   * helper interleaves `vi.advanceTimersByTimeAsync` with macrotask
+   * yields so pending fetch resolutions and setTimeout callbacks both
+   * settle before the assertion runs.
+   */
+  const advanceUntilSettled = async (ms: number) => {
+    await vi.advanceTimersByTimeAsync(ms);
+    // One extra tick lets any chained .then() after a setTimeout
+    // resolution fire before the test moves on.
+    await Promise.resolve();
+  };
+
+  beforeEach(() => {
+    vi.useFakeTimers();
+    api.post.mockReset();
+  });
+
+  afterEach(() => {
+    vi.useRealTimers();
+  });
+
+  it("phase 1: issues tasks/cancel via /workspaces/:id/a2a BEFORE any /restart", async () => {
+    seedTwoActive();
+    // Hold both tasks/cancel responses open so the click handler is
+    // observably paused at phase 1. We don't actually need to resolve
+    // them for the order assertion — just inspect the call log.
+    let resolveCancels!: () => void;
+    const cancelGate = new Promise<void>((r) => { resolveCancels = r; });
+    api.post.mockImplementation(async (path: string) => {
+      if (path.endsWith("/a2a")) {
+        await cancelGate;
+      }
+      return undefined;
+    });
+
+    render(<Toolbar />);
+    const btn = screen.getByRole("button", { name: /stop all running tasks/i });
+    fireEvent.click(btn);
+
+    // Yield once so the click handler enters phase 1 and dispatches the
+    // two /a2a POSTs.
+    await Promise.resolve();
+    await Promise.resolve();
+
+    const a2aCalls = api.post.mock.calls.filter((c) => String(c[0]).endsWith("/a2a"));
+    const restartCalls = api.post.mock.calls.filter((c) => String(c[0]).endsWith("/restart"));
+    expect(a2aCalls.length).toBe(2);
+    expect(restartCalls.length).toBe(0);
+
+    // Each /a2a POST carries the canonical tasks/cancel envelope.
+    for (const call of a2aCalls) {
+      expect(call[1]).toEqual({ method: "tasks/cancel", params: {} });
+    }
+
+    // Release the gate so the test cleanup doesn't dangle.
+    resolveCancels();
+    await advanceUntilSettled(10_000);
+  });
+
+  it("phase 2: when activeTasks drains to 0 during the poll window, /restart is NOT called", async () => {
+    seedTwoActive();
+    api.post.mockResolvedValue(undefined);
+
+    render(<Toolbar />);
+    fireEvent.click(screen.getByRole("button", { name: /stop all running tasks/i }));
+
+    // Let phase 1 fire (the two tasks/cancel calls).
+    await Promise.resolve();
+    await Promise.resolve();
+
+    // Simulate the platform pushing TASK_UPDATED with active_tasks=0
+    // on both workspaces — emulate by mutating the store directly,
+    // which is what canvas-events.ts does in production.
+    defaultStore.nodes = toStoreNodes(makeNodes(["online", "online"], [0, 0]));
+
+    // Advance past the first poll interval (250ms) so the loop sees
+    // the drained store and exits early.
+    await advanceUntilSettled(400);
+    // Drain any remaining timers so the handler returns cleanly.
+    await advanceUntilSettled(10_000);
+
+    const restartCalls = api.post.mock.calls.filter((c) => String(c[0]).endsWith("/restart"));
+    expect(restartCalls.length).toBe(0);
+  });
+
+  it("phase 3: when activeTasks does NOT drain inside the timeout, falls through to /restart for each stuck workspace", async () => {
+    seedTwoActive();
+    api.post.mockResolvedValue(undefined);
+
+    render(<Toolbar />);
+    fireEvent.click(screen.getByRole("button", { name: /stop all running tasks/i }));
+
+    // Phase 1 dispatch.
+    await Promise.resolve();
+    await Promise.resolve();
+
+    // Do NOT drain — activeTasks stays at 2 for both. Advance past the
+    // 8000ms drain timeout plus a buffer so phase 3's /restart POSTs fire.
+    await advanceUntilSettled(9_000);
+    await advanceUntilSettled(1_000);
+
+    const a2aCalls = api.post.mock.calls.filter((c) => String(c[0]).endsWith("/a2a"));
+    const restartCalls = api.post.mock.calls.filter((c) => String(c[0]).endsWith("/restart"));
+    expect(a2aCalls.length).toBe(2);
+    expect(restartCalls.length).toBe(2);
+
+    // Order check: every /a2a call comes before every /restart call.
+    const lastA2AIdx = Math.max(
+      ...api.post.mock.calls.map((c, i) => (String(c[0]).endsWith("/a2a") ? i : -1))
+    );
+    const firstRestartIdx = Math.min(
+      ...api.post.mock.calls.map((c, i) => (String(c[0]).endsWith("/restart") ? i : Infinity))
+    );
+    expect(lastA2AIdx).toBeLessThan(firstRestartIdx);
+  });
+
+  it("phase 3 selective: drains only one of two workspaces — /restart is called only for the stuck one", async () => {
+    seedTwoActive();
+    api.post.mockResolvedValue(undefined);
+
+    render(<Toolbar />);
+    fireEvent.click(screen.getByRole("button", { name: /stop all running tasks/i }));
+
+    await Promise.resolve();
+    await Promise.resolve();
+
+    // ws-0 drains immediately, ws-1 stays stuck for the full timeout.
+    defaultStore.nodes = toStoreNodes(makeNodes(["online", "online"], [0, 2]));
+    await advanceUntilSettled(9_500);
+
+    const restartCalls = api.post.mock.calls.filter((c) => String(c[0]).endsWith("/restart"));
+    expect(restartCalls.length).toBe(1);
+    expect(restartCalls[0][0]).toBe("/workspaces/ws-1/restart");
+  });
+});
@@ -0,0 +1,181 @@
+'use client';
+
+import { useCallback, useEffect, useState } from 'react';
+import { api } from '@/lib/api';
+import { fetchSession, type Session } from '@/lib/auth';
+import { getTenantSlug } from '@/lib/tenant';
+import { Spinner } from '@/components/Spinner';
+
+/**
+ * Organization-identity surface inside SettingsPanel.
+ *
+ * Closes a chronic UX gap where users (and our own AI agents) had to
+ * call /cp/auth/me or /cp/orgs from browser devtools to read their
+ * org_id UUID. Now: a copy-buttoned view of name + slug + UUID for the
+ * currently-active org, plus a switcher list when the user belongs to
+ * multiple orgs.
+ *
+ * Data path:
+ *   1. fetchSession() → /cp/auth/me → current org_id
+ *   2. api.get('/cp/orgs') → list of all orgs the user belongs to
+ *   3. Match by id === session.org_id; fall back to host-slug match
+ *      if the session probe loses the race.
+ *
+ * Read-only — this tab never mutates. Org creation/switching lives at
+ * /orgs (the post-signup landing page).
+ */
+
+interface Org {
+  id: string;
+  slug: string;
+  name: string;
+  status?: string;
+}
+
+// /cp/orgs may return a bare array or {orgs: []} — see orgs/page.tsx
+// for the same defensive unwrap.
+type OrgsResponse = Org[] | { orgs?: Org[] };
+
+export function OrgInfoTab() {
+  const [orgs, setOrgs] = useState<Org[] | null>(null);
+  const [session, setSession] = useState<Session | null>(null);
+  const [error, setError] = useState<string | null>(null);
+  const [loading, setLoading] = useState(true);
+
+  useEffect(() => {
+    let cancelled = false;
+    (async () => {
+      try {
+        const [sess, body] = await Promise.all([
+          fetchSession().catch(() => null),
+          api.get<OrgsResponse>('/cp/orgs'),
+        ]);
+        if (cancelled) return;
+        setSession(sess);
+        setOrgs(Array.isArray(body) ? body : body.orgs ?? []);
+      } catch (e) {
+        if (!cancelled) setError(e instanceof Error ? e.message : 'Failed to load org info');
+      } finally {
+        if (!cancelled) setLoading(false);
+      }
+    })();
+    return () => {
+      cancelled = true;
+    };
+  }, []);
+
+  const tenantSlug = getTenantSlug();
+  const currentOrg =
+    orgs?.find((o) => session && o.id === session.org_id) ??
+    orgs?.find((o) => tenantSlug && o.slug === tenantSlug) ??
+    null;
+  const otherOrgs = orgs?.filter((o) => o.id !== currentOrg?.id) ?? [];
+
+  if (loading) {
+    return (
+      <div
+        role="status"
+        aria-live="polite"
+        className="flex items-center justify-center gap-2 py-6 text-ink-mid text-xs"
+      >
+        <Spinner /> Loading organization…
+      </div>
+    );
+  }
+  if (error) {
+    return (
+      <div className="p-4">
+        <div className="px-3 py-2 bg-red-950/40 border border-red-800/50 rounded-lg text-[10px] text-bad">
+          {error}
+        </div>
+      </div>
+    );
+  }
+  if (!currentOrg) {
+    return (
+      <div className="p-4">
+        <p className="text-xs text-ink-mid">
+          No organization found for this session. If this is unexpected, sign out and back in, or visit{' '}
+          <a href="/orgs" className="underline">/orgs</a>.
+        </p>
+      </div>
+    );
+  }
+
+  return (
+    <div className="p-4 space-y-4">
+      <div>
+        <h3 className="text-sm font-semibold text-ink mb-1">Current Organization</h3>
+        <p className="text-[10px] text-ink-mid leading-relaxed">
+          IDs you can paste into API calls, support tickets, or CLI arguments. The UUID never changes;
+          the slug is the URL subdomain.
+        </p>
+      </div>
+      <OrgIdentityCard org={currentOrg} highlighted />
+      {otherOrgs.length > 0 && (
+        <div className="space-y-2 pt-2">
+          <h4 className="text-[11px] font-semibold text-ink-mid uppercase tracking-wider">
+            Your other organizations ({otherOrgs.length})
+          </h4>
+          {otherOrgs.map((o) => (
+            <OrgIdentityCard key={o.id} org={o} />
+          ))}
+        </div>
+      )}
+    </div>
+  );
+}
+
+function OrgIdentityCard({ org, highlighted }: { org: Org; highlighted?: boolean }) {
+  return (
+    <div
+      className={`rounded-lg border p-3 space-y-2 ${
+        highlighted ? 'border-accent/40 bg-accent-strong/5' : 'border-line/40 bg-surface-card/40'
+      }`}
+      data-testid={`org-card-${org.slug}`}
+    >
+      <div className="flex items-baseline justify-between gap-2">
+        <span className="text-[12px] font-medium text-ink truncate">{org.name}</span>
+        {org.status && (
+          <span className="text-[9px] text-ink-mid uppercase tracking-wider shrink-0">{org.status}</span>
+        )}
+      </div>
+      <IdentityRow label="Slug" value={org.slug} />
+      <IdentityRow label="UUID" value={org.id} mono />
+    </div>
+  );
+}
+
+function IdentityRow({ label, value, mono }: { label: string; value: string; mono?: boolean }) {
+  const [copied, setCopied] = useState(false);
+  const onCopy = useCallback(() => {
+    // Best-effort: jsdom + old Safari throw synchronously on writeText.
+    try {
+      navigator.clipboard.writeText(value);
+    } catch {
+      /* user can still triple-click select */
+    }
+    setCopied(true);
+    setTimeout(() => setCopied(false), 2000);
+  }, [value]);
+  return (
+    <div className="flex items-center gap-2">
+      <span className="text-[10px] text-ink-mid w-10 shrink-0">{label}</span>
+      <code
+        className={`flex-1 text-[11px] text-ink bg-surface-sunken/60 px-2 py-1 rounded select-all break-all ${
+          mono ? 'font-mono' : ''
+        }`}
+      >
+        {value}
+      </code>
+      <button
+        type="button"
+        onClick={onCopy}
+        aria-label={`Copy ${label}`}
+        className="shrink-0 px-2 py-1 bg-surface-card/60 hover:bg-surface-card border border-line/40 rounded text-[10px] text-ink-mid hover:text-ink transition-colors focus:outline-none focus-visible:ring-2 focus-visible:ring-accent focus-visible:ring-offset-1"
+      >
+        {copied ? 'Copied' : 'Copy'}
+      </button>
+    </div>
+  );
+}
@@ -8,6 +8,7 @@ import { useKeyboardShortcut } from '@/hooks/use-keyboard-shortcut';
 import { SecretsTab } from './SecretsTab';
 import { TokensTab } from './TokensTab';
 import { OrgTokensTab } from './OrgTokensTab';
+import { OrgInfoTab } from './OrgInfoTab';
 import { UnsavedChangesGuard } from './UnsavedChangesGuard';

 /** Module-level ref so TopBar's SettingsButton can receive focus back on close. */
@@ -116,6 +117,9 @@ export function SettingsPanel({ workspaceId }: SettingsPanelProps) {
                <Tabs.Trigger value="org-tokens" className="settings-panel__tab">
                  Org API Keys
                </Tabs.Trigger>
+                <Tabs.Trigger value="org-info" className="settings-panel__tab">
+                  Organization
+                </Tabs.Trigger>
              </Tabs.List>

              <Tabs.Content value="api-keys" className="settings-panel__content">
@@ -129,6 +133,10 @@ export function SettingsPanel({ workspaceId }: SettingsPanelProps) {
              <Tabs.Content value="org-tokens" className="settings-panel__content">
                <OrgTokensTab />
              </Tabs.Content>
+
+              <Tabs.Content value="org-info" className="settings-panel__content">
+                <OrgInfoTab />
+              </Tabs.Content>
            </Tabs.Root>

            <div className="settings-panel__footer">
@@ -0,0 +1,207 @@
+// @vitest-environment jsdom
+/**
+ * Tests for OrgInfoTab — surfaces current org name/slug/UUID with copy
+ * buttons, plus a list of the user's other orgs when applicable.
+ *
+ * Covers (≥3 cases per the closing-the-UX-gap brief):
+ *   - Loading state (spinner + aria-live)
+ *   - Renders current org matched by session.org_id, with UUID + slug + name
+ *   - Copy button writes the UUID to navigator.clipboard
+ *   - Falls back to host-slug match when session lookup fails
+ *   - Lists other orgs when user belongs to multiple
+ *   - Error banner when /cp/orgs throws
+ *   - Empty/no-match state renders the recovery hint, not a crash
+ */
+import React from "react";
+import { render, screen, fireEvent, cleanup, act, waitFor } from "@testing-library/react";
+import { afterEach, beforeEach, describe, expect, it, vi } from "vitest";
+import { OrgInfoTab } from "../OrgInfoTab";
+
+const mockGet = vi.fn();
+const mockFetchSession = vi.fn();
+const mockGetTenantSlug = vi.fn();
+
+vi.mock("@/lib/api", () => ({
+  api: { get: (...args: unknown[]) => mockGet(...args) },
+}));
+vi.mock("@/lib/auth", () => ({
+  fetchSession: (...args: unknown[]) => mockFetchSession(...args),
+}));
+vi.mock("@/lib/tenant", () => ({
+  getTenantSlug: (...args: unknown[]) => mockGetTenantSlug(...args),
+}));
+
+// Stub clipboard
+vi.stubGlobal("navigator", {
+  clipboard: { writeText: vi.fn().mockResolvedValue(undefined) },
+});
+
+beforeEach(() => {
+  vi.useRealTimers();
+  mockGet.mockReset();
+  mockFetchSession.mockReset();
+  mockGetTenantSlug.mockReset();
+  mockGetTenantSlug.mockReturnValue("");
+  vi.mocked(navigator.clipboard.writeText).mockReset();
+});
+
+afterEach(() => {
+  cleanup();
+});
+
+async function flush() {
+  await act(async () => {
+    await Promise.resolve();
+    await Promise.resolve();
+  });
+}
+
+const AGENTS_TEAM = {
+  id: "2355b568-0799-4cc7-9e7f-806747f9958c",
+  slug: "agents-team",
+  name: "Agents Team",
+  status: "running",
+};
+const OTHER_ORG = {
+  id: "11111111-1111-4111-8111-111111111111",
+  slug: "skunkworks",
+  name: "Skunkworks",
+  status: "running",
+};
+
+// ─── Loading ─────────────────────────────────────────────────────────────────
+
+describe("OrgInfoTab — loading", () => {
+  it("shows spinner while fetching", () => {
+    mockGet.mockImplementation(() => new Promise(() => {}));
+    mockFetchSession.mockImplementation(() => new Promise(() => {}));
+    render(<OrgInfoTab />);
+    const status = screen.getByRole("status");
+    expect(status).toBeTruthy();
+    expect(status.getAttribute("aria-live")).toBe("polite");
+    expect(status.textContent).toContain("Loading organization");
+  });
+});
+
+// ─── Current org renders + copy ──────────────────────────────────────────────
+
+describe("OrgInfoTab — current org", () => {
+  it("renders the org matched by session.org_id with name, slug, UUID", async () => {
+    mockFetchSession.mockResolvedValue({
+      user_id: "u-1",
+      org_id: AGENTS_TEAM.id,
+      email: "hongming@moleculesai.app",
+    });
+    mockGet.mockResolvedValue([AGENTS_TEAM, OTHER_ORG]);
+
+    render(<OrgInfoTab />);
+    await flush();
+    await waitFor(() => screen.getByText("Current Organization"));
+
+    // Name shown
+    expect(screen.getByText("Agents Team")).toBeTruthy();
+    // Slug shown
+    expect(screen.getByText("agents-team")).toBeTruthy();
+    // UUID shown
+    expect(screen.getByText(AGENTS_TEAM.id)).toBeTruthy();
+  });
+
+  it("copy-UUID button writes the UUID to navigator.clipboard", async () => {
+    mockFetchSession.mockResolvedValue({
+      user_id: "u-1",
+      org_id: AGENTS_TEAM.id,
+      email: "hongming@moleculesai.app",
+    });
+    mockGet.mockResolvedValue([AGENTS_TEAM]);
+
+    render(<OrgInfoTab />);
+    await flush();
+    await waitFor(() => screen.getByText(AGENTS_TEAM.id));
+
+    const copyUuid = screen.getByRole("button", { name: /Copy UUID/i });
+    fireEvent.click(copyUuid);
+
+    expect(navigator.clipboard.writeText).toHaveBeenCalledWith(AGENTS_TEAM.id);
+    // Optimistic "Copied" label flip
+    await waitFor(() =>
+      expect(
+        screen.getByRole("button", { name: /Copy UUID/i }).textContent,
+      ).toContain("Copied"),
+    );
+  });
+
+  it("copy-Slug button writes the slug to navigator.clipboard", async () => {
+    mockFetchSession.mockResolvedValue({
+      user_id: "u-1",
+      org_id: AGENTS_TEAM.id,
+      email: "hongming@moleculesai.app",
+    });
+    mockGet.mockResolvedValue([AGENTS_TEAM]);
+
+    render(<OrgInfoTab />);
+    await flush();
+    await waitFor(() => screen.getByText(AGENTS_TEAM.slug));
+
+    fireEvent.click(screen.getByRole("button", { name: /Copy Slug/i }));
+    expect(navigator.clipboard.writeText).toHaveBeenCalledWith(AGENTS_TEAM.slug);
+  });
+});
+
+// ─── Fallback: host-slug match when session fails ────────────────────────────
+
+describe("OrgInfoTab — fallbacks", () => {
+  it("falls back to host-slug match when fetchSession rejects", async () => {
+    mockFetchSession.mockRejectedValue(new Error("session probe failed"));
+    mockGetTenantSlug.mockReturnValue("agents-team");
+    mockGet.mockResolvedValue({ orgs: [AGENTS_TEAM, OTHER_ORG] }); // wrapped shape
+
+    render(<OrgInfoTab />);
+    await flush();
+    await waitFor(() => screen.getByText("Current Organization"));
+
+    expect(screen.getByText("Agents Team")).toBeTruthy();
+    expect(screen.getByText(AGENTS_TEAM.id)).toBeTruthy();
+  });
+
+  it("lists other orgs the user belongs to under a separate header", async () => {
+    mockFetchSession.mockResolvedValue({
+      user_id: "u-1",
+      org_id: AGENTS_TEAM.id,
+      email: "hongming@moleculesai.app",
+    });
+    mockGet.mockResolvedValue([AGENTS_TEAM, OTHER_ORG]);
+
+    render(<OrgInfoTab />);
+    await flush();
+    await waitFor(() => screen.getByText(/Your other organizations/));
+
+    expect(screen.getByText("Skunkworks")).toBeTruthy();
+    expect(screen.getByText(OTHER_ORG.id)).toBeTruthy();
+  });
+});
+
+// ─── Error + empty handling ──────────────────────────────────────────────────
+
+describe("OrgInfoTab — error + empty", () => {
+  it("renders an error banner when /cp/orgs throws", async () => {
+    mockFetchSession.mockResolvedValue(null);
+    mockGet.mockRejectedValue(new Error("API GET /cp/orgs: 500 boom"));
+
+    render(<OrgInfoTab />);
+    await flush();
+    await waitFor(() => screen.getByText(/500 boom/));
+    expect(screen.queryByText("Current Organization")).toBeNull();
+  });
+
+  it("renders the recovery hint when no org matches (no crash)", async () => {
+    mockFetchSession.mockResolvedValue(null);
+    mockGetTenantSlug.mockReturnValue("");
+    mockGet.mockResolvedValue([]);
+
+    render(<OrgInfoTab />);
+    await flush();
+    await waitFor(() =>
+      screen.getByText(/No organization found for this session/),
+    );
+  });
+});
@@ -8,3 +8,4 @@ export { SearchBar } from './SearchBar';
 export { EmptyState } from './EmptyState';
 export { DeleteConfirmDialog } from './DeleteConfirmDialog';
 export { UnsavedChangesGuard } from './UnsavedChangesGuard';
+export { OrgInfoTab } from './OrgInfoTab';
@@ -649,7 +649,17 @@ function WaitingBubbles({ visible }: { visible: CommMessage[] }) {
    if (!prev || m.timestamp > prev.timestamp) tailByPeer.set(m.peerId, m);
  }
  const waitingPeers = Array.from(tailByPeer.values()).filter(
-    (m) => m.flow === "out" && (m.status === "pending" || m.status === "queued"),
+    // Task #227 — also light the indicator for status="dispatched": that's
+    // the platform's marker for a poll-mode delegation that's been
+    // recorded into the peer's inbox but not yet picked up. Without this
+    // arm, external/MCP peer threads showed an outbound bubble and then
+    // dead silence until the eventual reply landed — no parity with the
+    // native push-path "pending" indicator.
+    (m) =>
+      m.flow === "out" &&
+      (m.status === "pending" ||
+        m.status === "queued" ||
+        m.status === "dispatched"),
  );
  if (waitingPeers.length === 0) return null;
  return (
@@ -688,7 +698,9 @@ function WaitingBubbles({ visible }: { visible: CommMessage[] }) {
              <span className="text-[10px]">
                {m.status === "queued"
                  ? `${m.peerName} is busy — reply will arrive when they're free`
-                  : `Waiting for ${m.peerName}…`}
+                  : m.status === "dispatched"
+                    ? `Queued — ${m.peerName} will pick up on next poll`
+                    : `Waiting for ${m.peerName}…`}
              </span>
            </span>
          </div>
@@ -41,6 +41,19 @@ describe("inferA2AErrorHint", () => {
    expect(inferA2AErrorHint("RuntimeException in tool call")).toMatch(/runtime threw an exception/);
  });

+  it("points at the Activity tab (the real in-product logs surface), not 'workspace/container logs' (internal#212)", () => {
+    // Pre-#212 these hints sent users to "workspace logs" / "container
+    // logs" — neither has a UI affordance in the canvas. Activity tab
+    // is the in-product surface where the full row lives. Lock the
+    // copy so a future refactor cannot re-introduce the dangling
+    // pointer.
+    expect(inferA2AErrorHint("Agent error: boom")).toMatch(/Activity tab/);
+    expect(inferA2AErrorHint("some completely novel error nobody has matched yet")).toMatch(/Activity tab/);
+    // And the two strings together must not regress to the old text.
+    expect(inferA2AErrorHint("Agent error: boom")).not.toMatch(/container logs/);
+    expect(inferA2AErrorHint("some novel error")).not.toMatch(/workspace logs/);
+  });
+
  it("recognises peer-unreachable cases (Activity-tab originals)", () => {
    expect(inferA2AErrorHint("workspace not found")).toMatch(/can't be reached/);
    expect(inferA2AErrorHint("not accessible")).toMatch(/can't be reached/);
@@ -53,7 +66,8 @@ describe("inferA2AErrorHint", () => {

  it("returns a generic fallback for unrecognised text", () => {
    const hint = inferA2AErrorHint("some completely novel error nobody has matched yet");
-    expect(hint).toMatch(/Check the workspace logs|delivery failure/);
+    // Fallback now sends the user to the Activity tab (post-#212).
+    expect(hint).toMatch(/Activity tab|delivery failure/);
  });

  it("Claude SDK wedge wins over the more general timeout pattern", () => {
@@ -38,7 +38,11 @@ export function inferA2AErrorHint(detail: string): string {
    return "The connection to the remote agent dropped before a reply arrived. Usually a transient network blip — retry once. If it repeats, the remote container may have crashed mid-request; check its logs.";
  }
  if (t.includes("agent error") || t.includes("exception")) {
-    return "The remote agent's runtime threw an exception. Check the workspace's container logs for the traceback. Restart usually clears transient runtime crashes.";
+    // internal#212 closeout: end users have no "container logs" surface
+    // in the canvas; the Activity tab IS the user-visible logs surface
+    // (full row carries request/response body + error_detail). Point
+    // there so the hint is actionable from inside the product.
+    return "The remote agent's runtime threw an exception. Open the Activity tab for the full row (request body, response, error_detail) — Restart usually clears transient runtime crashes.";
  }
  if (
    t.includes("not found") ||
@@ -50,5 +54,9 @@ export function inferA2AErrorHint(detail: string): string {
  if (detail === "") {
    return "The remote agent returned no error detail (the underlying httpx exception had an empty message — typically a connection-reset or silent timeout). A workspace restart is the safe first move.";
  }
-  return "The remote agent reported a delivery failure. Check the workspace logs or try restarting.";
+  // internal#212 closeout: "workspace logs" pointed at a tab that does
+  // not exist — Activity tab is the in-product logs surface. Keep the
+  // hint generic enough for the unrecognised-detail fallback but point
+  // the user at a real affordance.
+  return "The remote agent reported a delivery failure. Open the Activity tab for the full row, or try restarting the workspace.";
 }
@@ -0,0 +1,178 @@
+// @vitest-environment jsdom
+//
+// Task #227 — external/MCP workspace progress UX parity.
+//
+// ws-server's `proxyA2ARequest` poll-mode short-circuit
+// (workspace-server/internal/handlers/a2a_proxy.go:402-432) returns a
+// synthetic `{status:"queued", delivery_mode:"poll", method:"message/send"}`
+// HTTP 200 within ~50ms when the target workspace is registered with
+// `delivery_mode=poll` — i.e. an operator's laptop running
+// `molecule-mcp-claude-channel`, a hermes/codex MCP bridge, or a Cursor
+// MCP client. The real agent reply arrives separately via the
+// AGENT_MESSAGE WebSocket event after the agent's next
+// `wait_for_message` poll (could be 1s, could be 60s).
+//
+// Pre-#227 behaviour: useChatSend treated the queued-200 as a successful
+// round-trip — extractReplyText returned "", no agent bubble was
+// created, `releaseSendGuards` flipped `sending` off, and the user saw
+// dead silence between their user bubble and the eventual reply with
+// NO progress indicator. That's the user-reported gap this task fixes.
+//
+// These tests pin the new behaviour: on a queued-200, the hook MUST NOT
+// call onAgentMessage (no empty bubble) AND MUST NOT call
+// releaseSendGuards (spinner persists). The eventual AGENT_MESSAGE WS
+// event is what clears the spinner — that path is covered by
+// useChatSocket.test.tsx already.
+
+import { describe, it, expect, vi, beforeEach } from "vitest";
+import { renderHook, act } from "@testing-library/react";
+
+// Capture the api.post invocations + control responses per-test.
+const apiPostMock = vi.fn<
+  (url: string, body?: unknown, opts?: unknown) => Promise<unknown>
+>();
+vi.mock("@/lib/api", () => ({
+  api: {
+    post: (url: string, body?: unknown, opts?: unknown) =>
+      apiPostMock(url, body, opts),
+    get: vi.fn(),
+  },
+}));
+
+// uploads — tests don't go through the upload path; stub the helpers
+// useChatSend imports so the module loads.
+vi.mock("../../uploads", () => ({
+  uploadChatFiles: vi.fn(),
+  FileTooLargeError: class FileTooLargeError extends Error {},
+}));
+
+// types — re-export the createMessage helper unchanged; only the
+// uploads stub matters above.
+import { useChatSend } from "../useChatSend";
+
+beforeEach(() => {
+  apiPostMock.mockReset();
+});
+
+describe("useChatSend — poll-mode (external/MCP) queued-200 handling — task #227", () => {
+  it("does NOT call onAgentMessage when the synthetic {status:'queued'} response lands (no empty bubble)", async () => {
+    // Mock the platform's poll-mode short-circuit response shape exactly
+    // as ws-server's `proxyA2ARequest` returns it (a2a_proxy.go:420-431).
+    apiPostMock.mockResolvedValueOnce({
+      status: "queued",
+      delivery_mode: "poll",
+      method: "message/send",
+    });
+
+    const onUserMessage = vi.fn();
+    const onAgentMessage = vi.fn();
+
+    const { result } = renderHook(() =>
+      useChatSend("ws-poll-target", {
+        getHistoryMessages: () => [],
+        onUserMessage,
+        onAgentMessage,
+      }),
+    );
+
+    await act(async () => {
+      await result.current.sendMessage("hello external workspace");
+      // Yield one microtask so the .then runs.
+      await Promise.resolve();
+    });
+
+    // User bubble fires — the user typed, that part is unconditional.
+    expect(onUserMessage).toHaveBeenCalledTimes(1);
+    // CRITICAL: no agent bubble. extractReplyText on a queued envelope
+    // returns "" — the pre-#227 code would still have hit the
+    // "releaseSendGuards + no bubble" path, BUT it would have ended
+    // `sending`. The new code returns early BEFORE that release, so the
+    // contract under test is "no synthesised empty bubble".
+    expect(onAgentMessage).not.toHaveBeenCalled();
+  });
+
+  it("keeps `sending` true after a queued-200 — the spinner must persist until the real AGENT_MESSAGE arrives", async () => {
+    apiPostMock.mockResolvedValueOnce({
+      status: "queued",
+      delivery_mode: "poll",
+      method: "message/send",
+    });
+
+    const { result } = renderHook(() =>
+      useChatSend("ws-poll-target", {
+        getHistoryMessages: () => [],
+      }),
+    );
+
+    await act(async () => {
+      await result.current.sendMessage("waiting for the operator laptop");
+      await Promise.resolve();
+    });
+
+    // The spinner-driving state is `sending`. On a queued-200, it must
+    // remain true — clearing it here is the exact bug task #227
+    // resurfaces (collapsing the spinner before the agent has even seen
+    // the message).
+    expect(result.current.sending).toBe(true);
+  });
+
+  it("ALSO keeps `sending` true even after a follow-up microtask flush — guards against an accidental late release", async () => {
+    // Defense: ensure no chained .then / .finally accidentally calls
+    // releaseSendGuards on the queued path. Run several microtask
+    // ticks and re-assert.
+    apiPostMock.mockResolvedValueOnce({
+      status: "queued",
+      delivery_mode: "poll",
+    });
+
+    const { result } = renderHook(() =>
+      useChatSend("ws-poll-target", {
+        getHistoryMessages: () => [],
+      }),
+    );
+
+    await act(async () => {
+      await result.current.sendMessage("late-release-guard");
+      // Flush multiple microtask ticks.
+      await Promise.resolve();
+      await Promise.resolve();
+      await Promise.resolve();
+    });
+
+    expect(result.current.sending).toBe(true);
+  });
+
+  it("push-mode (real reply parts) still flips sending=false + creates an agent bubble — non-regression for the default path", async () => {
+    // Sanity-check the push path still works: a real reply must call
+    // onAgentMessage and flip sending=false. Without this assertion an
+    // overzealous "return early on any non-result body" would silently
+    // break the dominant push-mode path.
+    apiPostMock.mockResolvedValueOnce({
+      result: {
+        parts: [{ kind: "text", text: "hi from native workspace" }],
+      },
+    });
+
+    const onAgentMessage = vi.fn();
+    const { result } = renderHook(() =>
+      useChatSend("ws-native-push", {
+        getHistoryMessages: () => [],
+        onAgentMessage,
+      }),
+    );
+
+    await act(async () => {
+      await result.current.sendMessage("native push test");
+      await Promise.resolve();
+    });
+
+    expect(onAgentMessage).toHaveBeenCalledTimes(1);
+    const msg = onAgentMessage.mock.calls[0][0] as {
+      role: string;
+      content: string;
+    };
+    expect(msg.role).toBe("agent");
+    expect(msg.content).toBe("hi from native workspace");
+    expect(result.current.sending).toBe(false);
+  });
+});
@@ -116,6 +116,77 @@ describe("useChatSocket — surface error_detail to onSendError (internal#212)",
    expect(reason.length).toBeGreaterThan(0);
  });

+  // Task #227 — external/MCP (poll-mode) workspace progress UX.
+  //
+  // ws-server's `proxyA2ARequest` poll-mode short-circuit fires the
+  // ACTIVITY_LOGGED a2a_receive with status="ok" and NO duration_ms (no
+  // reply yet — the request is queued for the agent's next poll). Before
+  // task #227 the (status==="ok" && durationMs) guard silently dropped
+  // this row, so the chat UI had ZERO progress signal between "user
+  // typed" and "agent eventually polled and replied". Lock the queued
+  // line in so future refactors don't regress to the silent-drop state.
+  it("emits a 'queued — will pick up on next poll' activity line when a2a_receive status=ok has no duration_ms (poll-mode)", () => {
+    const onActivityLog = vi.fn();
+    renderHook(() =>
+      useChatSocket("ws-self", {
+        onActivityLog,
+      }),
+    );
+
+    expect(capturedHandler).not.toBeNull();
+    act(() => {
+      capturedHandler!({
+        event: "ACTIVITY_LOGGED",
+        workspace_id: "ws-self",
+        payload: {
+          activity_type: "a2a_receive",
+          method: "message/send",
+          status: "ok",
+          target_id: "ws-self",
+          // No duration_ms — this is the queued-for-poll signal.
+        },
+        timestamp: "2026-05-20T00:00:00Z",
+      });
+    });
+
+    expect(onActivityLog).toHaveBeenCalledTimes(1);
+    const line = onActivityLog.mock.calls[0][0] as string;
+    // The line MUST be present (not the empty-string silent-drop pattern)
+    // and MUST mention the queued state so the user has actionable signal.
+    expect(line.length).toBeGreaterThan(0);
+    expect(line.toLowerCase()).toMatch(/queued|poll/);
+  });
+
+  // Pair with the above: poll-mode acknowledgement must NOT prematurely
+  // call onSendComplete — the spinner has to stay up until the actual
+  // AGENT_MESSAGE reply lands. (The reply-success path with duration_ms
+  // still calls onSendComplete; that's the push-mode case.)
+  it("does NOT call onSendComplete on a poll-mode queued a2a_receive (spinner must persist)", () => {
+    const onSendComplete = vi.fn();
+    renderHook(() =>
+      useChatSocket("ws-self", {
+        onSendComplete,
+      }),
+    );
+
+    act(() => {
+      capturedHandler!({
+        event: "ACTIVITY_LOGGED",
+        workspace_id: "ws-self",
+        payload: {
+          activity_type: "a2a_receive",
+          method: "message/send",
+          status: "ok",
+          target_id: "ws-self",
+          // No duration_ms.
+        },
+        timestamp: "2026-05-20T00:00:00Z",
+      });
+    });
+
+    expect(onSendComplete).not.toHaveBeenCalled();
+  });
+
  it("ignores errors targeted at a different workspace's peer", () => {
    // Defense against a race where the WS hub fans out to all clients —
    // each chat panel must only react when target_id matches its own
@@ -22,6 +22,28 @@ interface A2AResponse {
    parts?: A2APart[];
    artifacts?: Array<{ parts: A2APart[] }>;
  };
+  /** Set by ws-server's poll-mode short-circuit in `proxyA2ARequest`
+   *  (a2a_proxy.go:416-431) when the target workspace is registered as
+   *  `delivery_mode=poll` — e.g. an operator's laptop running
+   *  `molecule-mcp-claude-channel`, a hermes/codex MCP bridge, or a
+   *  Cursor MCP client. The HTTP 200 carries the synthetic envelope
+   *  `{status:"queued", delivery_mode:"poll", method:"message/send"}`
+   *  immediately (~50ms), BEFORE the agent has produced a reply.
+   *
+   *  Task #227 routing: when this field is "queued" the caller must NOT
+   *  treat the 200 as "agent done" — there are no `result.parts` yet
+   *  (the reply will arrive separately via the AGENT_MESSAGE WS event
+   *  after the agent's next poll). Keep the spinner up; the eventual
+   *  AGENT_MESSAGE flips `sending` off via the existing useChatSocket
+   *  `onSendComplete` path. Without this distinction the spinner
+   *  disappeared immediately and external/MCP workspaces had no progress
+   *  UX between send and reply. */
+  status?: string;
+  /** Companion to `status` — "poll" when the queued short-circuit fired.
+   *  Defensive: we key the poll-mode-skip decision on status==="queued"
+   *  (the canonical signal) rather than on this field, but it's surfaced
+   *  here so future debugging / tests can assert on the full envelope. */
+  delivery_mode?: string;
 }

 export function extractReplyText(resp: A2AResponse): string {
@@ -195,6 +217,30 @@ export function useChatSend(workspaceId: string, options: UseChatSendOptions) {
            sendInFlightRef.current = false;
            return;
          }
+          // Task #227 — poll-mode (external/MCP workspace) queued-200
+          // short-circuit. ws-server's `proxyA2ARequest` returns
+          // `{status:"queued", delivery_mode:"poll", ...}` immediately
+          // when the target has no URL (delivery_mode=poll), BEFORE the
+          // agent has produced any reply. There is no `result.parts`
+          // payload here — the actual reply will arrive separately via
+          // the AGENT_MESSAGE WebSocket event after the agent's next
+          // `wait_for_message` poll.
+          //
+          // Keep the spinner up by deliberately NOT calling
+          // releaseSendGuards: the user-facing "thinking" state must
+          // persist until the AGENT_MESSAGE lands (handled by the
+          // useChatSocket `onAgentMessage`/`onSendComplete` path) or an
+          // explicit error fires (`onSendError` from an ACTIVITY_LOGGED
+          // status="error"). Don't synthesise an empty agent bubble.
+          //
+          // sendInFlightRef stays true intentionally — it's the dedup
+          // guard for the user typing two messages back-to-back; for
+          // poll mode the second message would race the first agent's
+          // reply, so blocking is correct (matches push-mode behaviour
+          // where `sending` blocks the textarea).
+          if (resp?.status === "queued") {
+            return;
+          }
          const replyText = extractReplyText(resp);
          const replyFiles = extractFilesFromTask(
            (resp?.result ?? {}) as Record<string, unknown>,
@@ -62,6 +62,25 @@ export function useChatSocket(
            line = `← ${targetName} responded (${sec}s)`;
            const own = (targetId || msg.workspace_id) === workspaceId;
            if (own) callbacksRef.current.onSendComplete?.();
+          } else if (status === "ok" && !durationMs) {
+            // Task #227 — poll-mode (external/MCP workspace) queued receipt.
+            // ws-server `logA2AReceiveQueued` writes a "received but no
+            // reply yet" row with status="ok" and NO duration_ms, then
+            // immediately returns the synthetic {status:"queued"} 200 to
+            // the caller. Before this branch the row was silently dropped
+            // by the (status==="ok" && durationMs) guard above — leaving
+            // the chat UI with zero progress signal for the entire window
+            // between "user typed" and "agent eventually polled and
+            // replied". Surface the queued state explicitly so the user
+            // sees acknowledgement (matches the queued-delegation
+            // indicator in AgentCommsPanel.WaitingBubbles).
+            //
+            // We intentionally do NOT call onSendComplete here: the
+            // outbound is not done — only acknowledged. The MyChatPanel
+            // spinner stays up until the actual AGENT_MESSAGE reply lands
+            // (poll path) or an explicit error fires (which still hits
+            // the status==="error" branch below).
+            line = `⧗ ${targetName} queued — agent will pick up on next poll`;
          } else if (status === "error") {
            line = `⚠ ${targetName} error`;
            const own = (targetId || msg.workspace_id) === workspaceId;
@@ -523,6 +523,9 @@ export function buildNodesAndEdges(
        // that don't yet include these columns in the GET response.
        broadcastEnabled: ws.broadcast_enabled ?? false,
        talkToUserEnabled: ws.talk_to_user_enabled ?? true,
+        // A2A delivery mode (task #227). Absent on older ws-server builds
+        // — leave undefined so the chat UI's "?? 'push'" fallback applies.
+        deliveryMode: ws.delivery_mode,
      },
    };
    if (hasParent) {
@@ -106,6 +106,28 @@ export interface WorkspaceNodeData extends Record<string, unknown> {
   *  send_message_to_user / POST /notify return 403 and the canvas
   *  shows a "not enabled" state with a button to re-enable. Default true. */
  talkToUserEnabled?: boolean;
+  /** A2A inbound delivery mode for this workspace — "push" (default —
+   *  synchronous HTTP dispatch by ws-server `proxyA2ARequest`) or "poll"
+   *  (workspace has no URL; ws-server logs the request and the agent
+   *  consumes it via `wait_for_message` / GET /activity?since_id=).
+   *
+   *  Why surfaced to the UI: poll-mode targets (external/MCP workspaces:
+   *  `molecule-mcp-claude-channel` on an operator laptop, hermes/codex
+   *  bridge clients, Cursor MCP) acknowledge a canvas `message/send` with
+   *  a synthetic `{status:"queued"}` 200 within ~50ms. Without this flag
+   *  the chat UI cannot tell that gap from a real round-trip — the
+   *  spinner disappears immediately and the user sees dead silence until
+   *  the agent eventually polls and replies via the AGENT_MESSAGE WS
+   *  event (could be seconds, could be minutes). Task #227 — render a
+   *  "queued — agent will pick up on next poll" state for poll-mode
+   *  sends so external/MCP workspaces have progress UX parity with
+   *  native runtimes (claude-code / codex / hermes / openclaw).
+   *
+   *  Sourced from the GET /workspaces response (`delivery_mode` snake_case
+   *  field, mapped here in canvas-topology.ts). Absent on older platform
+   *  builds — that fallthrough is treated as "push" to match
+   *  ws-server's `lookupDeliveryMode` default. */
+  deliveryMode?: string;
 }

 export type PanelTab = "details" | "skills" | "chat" | "terminal" | "config" | "schedule" | "channels" | "files" | "memory" | "traces" | "events" | "activity" | "audit";
@@ -342,6 +342,16 @@ export interface WorkspaceData {
  /** Workspace ability flags (migration 20260514). */
  broadcast_enabled?: boolean;
  talk_to_user_enabled?: boolean;
+  /** A2A delivery mode for inbound messages — "push" (default, synchronous
+   *  HTTP dispatch to `url`) or "poll" (queued to activity_logs, agent
+   *  picks up via `wait_for_message` / GET /activity?since_id=). Surfaced
+   *  in the GET /workspaces response since #2339 PR 1; older platform
+   *  versions return it absent so the canvas treats absent as "push" (the
+   *  documented default in `lookupDeliveryMode`). Used by the chat UI to
+   *  render an "agent will pick up on next poll" indicator instead of
+   *  collapsing the spinner the moment the synchronous queued-200 returns
+   *  (task #227 — external/MCP workspaces had no progress UX). */
+  delivery_mode?: string;
 }

 let socket: ReconnectingSocket | null = null;
@@ -285,6 +285,39 @@ Canvas requests (no `X-Workspace-ID` header) and system callers

 ---

+## Multiple Workspaces From One Local MCP Bridge
+
+The standalone runtime package includes `molecule-mcp`, a local MCP bridge for
+external agents such as Claude Code, Codex, Hermes, and other tools that run
+outside the platform container fleet. One local bridge can serve multiple
+external workspaces by setting `MOLECULE_WORKSPACES`:
+
+```json
+[
+  {
+    "id": "workspace-id-local-to-hongming-org",
+    "token": "...",
+    "platform_url": "https://hongming.moleculesai.app"
+  },
+  {
+    "id": "different-workspace-id-local-to-agents-team-org",
+    "token": "...",
+    "platform_url": "https://agents-team.moleculesai.app"
+  }
+]
+```
+
+`platform_url` is the tenant routing key. The bridge registers, heartbeats,
+polls inboxes, and sends outbound A2A calls against the URL attached to the
+workspace that is doing the work.
+
+Do not add `org_id` to this config. The tenant already comes from
+`platform_url`, and the bearer token is issued by that tenant. Workspace IDs
+also do not need to be shared across orgs; each tenant can return its own
+workspace ID and token for the same local agent process.
+
+---
+
 ## Canvas Appearance

 External workspaces appear on the canvas with a purple **REMOTE** badge
@@ -135,6 +135,33 @@ The `id` field is your workspace ID — remember it.

 ---

+## Optional — one local MCP bridge, multiple tenants
+
+If your local agent runtime uses `molecule-mcp`, one process can serve more
+than one external workspace:
+
+```bash
+export MOLECULE_WORKSPACES='[
+  {
+    "id": "workspace-id-local-to-you-org",
+    "token": "...",
+    "platform_url": "https://you.moleculesai.app"
+  },
+  {
+    "id": "different-workspace-id-local-to-team-org",
+    "token": "...",
+    "platform_url": "https://team.moleculesai.app"
+  }
+]'
+molecule-mcp
+```
+
+Use the workspace ID and token returned by each tenant. The IDs may differ
+across orgs. `org_id` is not required here because `platform_url` selects the
+tenant and the token is tenant-scoped.
+
+---
+
 ## Step 4 — Chat with it

 1. Open your Molecule canvas at `https://<TENANT>`
@@ -125,6 +125,33 @@ The agent appears on the canvas with a **purple REMOTE badge** within seconds. F

 ---

+## Multi-Tenant Local MCP Bridge
+
+For local MCP-driven agents, use the standalone runtime's `molecule-mcp`
+entrypoint. A single local bridge can serve multiple external workspaces by
+setting `MOLECULE_WORKSPACES`:
+
+```json
+[
+  {
+    "id": "workspace-id-local-to-acme",
+    "token": "...",
+    "platform_url": "https://acme.moleculesai.app"
+  },
+  {
+    "id": "different-workspace-id-local-to-ops",
+    "token": "...",
+    "platform_url": "https://ops.moleculesai.app"
+  }
+]
+```
+
+`platform_url` selects the tenant for registration, heartbeat, inbox polling,
+and outbound A2A routing. `org_id` is not required in this config, and the
+workspace IDs do not need to match across tenants.
+
+---
+
 ## What Phase 30 Covers

 | Phase | What shipped | Endpoint |
@@ -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
@@ -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)
@@ -24,7 +24,8 @@
 #
 # Only PROVISIONING differs from staging:
 #   - staging: POST /cp/admin/orgs (cold EC2 tenant) + per-tenant admin
-#     token + each workspace's auth_token from the POST /workspaces resp.
+#     token + each workspace's MCP bearer from create response or an admin
+#     token-mint fallback.
 #   - local:   POST /workspaces directly against the local stack
 #     (BASE, default http://localhost:8080), MCP bearer minted via
 #     GET /admin/workspaces/:id/test-token (e2e_mint_test_token —
@@ -103,7 +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 $rc
 }
@@ -112,7 +113,7 @@ trap teardown EXIT INT TERM
 # Pre-sweep workspaces a prior crashed run of THIS script left behind
 # (name prefix match only — never a blanket delete). The trap fires on
 # normal exit, but a kill -9 / SIGPIPE can bypass it.
-PRIOR=$(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-")))
@@ -121,7 +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 ─────────────────────────────────────────────
@@ -132,10 +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 "$PROBE" | grep -qi 'production\|disabled\|not found.*endpoint'; then
    echo "::error::GET /admin/workspaces/:id/test-token disabled (MOLECULE_ENV=production?). Cannot mint a local MCP bearer." >&2
    exit 1
@@ -240,7 +241,7 @@ else
 fi
 log "1/5 provisioning parent ($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
@@ -290,7 +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 "$WID" ]; then
@@ -40,8 +40,10 @@
 #   drives: POST /cp/admin/orgs (provision), GET
 #   /cp/admin/orgs/:slug/admin-token (per-tenant token), DELETE
 #   /cp/admin/tenants/:slug (teardown). The per-tenant admin token drives
-#   tenant workspace creation; each workspace's OWN auth_token (returned by
-#   POST /workspaces) drives its MCP call.
+#   tenant workspace creation; each workspace's OWN auth_token drives its
+#   MCP call. External-like runtimes may return the token in POST
+#   /workspaces; managed container runtimes usually require the admin token
+#   mint fallback below.
 #
 # Required env:
 #   MOLECULE_ADMIN_TOKEN   CP admin bearer — Railway staging CP_ADMIN_API_TOKEN
@@ -104,6 +106,46 @@ tenant_call() {
    -H "Content-Type: application/json" "$@"
 }

+tenant_call_capture() {
+  local method="$1" path="$2" out="$3"; shift 3
+  curl -sS -o "$out" -w "%{http_code}" -X "$method" "$TENANT_URL$path" \
+    -H "Authorization: Bearer $TENANT_TOKEN" \
+    -H "X-Molecule-Org-Id: $ORG_ID" \
+    -H "Content-Type: application/json" "$@"
+}
+
+redact_token_body() {
+  python3 -c '
+import json, re, sys
+raw = sys.stdin.read()
+try:
+    data = json.loads(raw)
+except Exception:
+    print(re.sub(r"(?i)([a-z0-9_]*token)=([^&\\s]+)", r"\1=<redacted>", raw)[:500])
+    raise SystemExit(0)
+
+def scrub(v):
+    if isinstance(v, dict):
+        return {k: ("<redacted>" if "token" in k.lower() else scrub(val)) for k, val in v.items()}
+    if isinstance(v, list):
+        return [scrub(x) for x in v]
+    return v
+
+print(json.dumps(scrub(data), separators=(",", ":"))[:500])
+'
+}
+
+extract_auth_token() {
+  python3 -c "
+import sys, json
+try:
+    d = json.load(sys.stdin)
+except Exception:
+    print(''); sys.exit(0)
+print(d.get('auth_token') or d.get('connection', {}).get('auth_token') or '')
+" 2>/dev/null
+}
+
 # ─── Scoped teardown ───────────────────────────────────────────────────
 # Deletes ONLY the org this run created (DELETE /cp/admin/tenants/$SLUG
 # with the {"confirm":$SLUG} fat-finger guard). Never a cluster-wide
@@ -218,16 +260,31 @@ 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}")
  WID=$(echo "$R" | python3 -c "import sys,json; print(json.load(sys.stdin).get('id',''))" 2>/dev/null)
-  # auth_token is top-level for container runtimes; external-like nest it
-  # under connection.auth_token (verified vs staging response shape).
-  WTOK=$(echo "$R" | python3 -c "
-import sys, json
-try: d = json.load(sys.stdin)
-except Exception: print(''); sys.exit(0)
-print(d.get('auth_token') or d.get('connection', {}).get('auth_token') or '')
-" 2>/dev/null)
+  # 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.
+  WTOK=$(echo "$R" | extract_auth_token)
  [ -n "$WID" ] || fail "$rt workspace create failed: $(echo "$R" | head -c 300)"
-  [ -n "$WTOK" ] || fail "$rt workspace did not return an auth_token — cannot drive its MCP call (resp: $(echo "$R" | head -c 300))"
+  TOKEN_DIAG=""
+  if [ -z "$WTOK" ]; then
+    TTOK_FILE=$(mktemp)
+    TTOK_CODE=$(tenant_call_capture POST "/admin/workspaces/$WID/tokens" "$TTOK_FILE" 2>/dev/null || echo "curl_error")
+    TTOK_RESP=$(cat "$TTOK_FILE" 2>/dev/null || true)
+    WTOK=$(echo "$TTOK_RESP" | extract_auth_token)
+    TOKEN_DIAG="POST /admin/workspaces/$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]="$WID"
  WS_TOKENS[$rt]="$WTOK"
  ALL_WS_IDS="$ALL_WS_IDS $WID"
@@ -0,0 +1,193 @@
+"""Tests for `.gitea/scripts/detect-changes.py`."""
+
+from __future__ import annotations
+
+import importlib.util
+from pathlib import Path
+
+
+REPO_ROOT = Path(__file__).resolve().parents[1]
+SCRIPT = REPO_ROOT / ".gitea" / "scripts" / "detect-changes.py"
+
+
+def load_module():
+    spec = importlib.util.spec_from_file_location("detect_changes", SCRIPT)
+    assert spec is not None
+    module = importlib.util.module_from_spec(spec)
+    assert spec.loader is not None
+    spec.loader.exec_module(module)
+    return module
+
+
+def test_ci_profile_classifies_surfaces():
+    mod = load_module()
+
+    assert mod.classify("ci", ["workspace-server/internal/handlers/a2a_proxy.go"]) == {
+        "platform": True,
+        "canvas": False,
+        "python": False,
+        "scripts": False,
+    }
+    assert mod.classify("ci", ["canvas/src/app/page.tsx"]) == {
+        "platform": False,
+        "canvas": True,
+        "python": False,
+        "scripts": False,
+    }
+    assert mod.classify("ci", ["tests/e2e/test_model_slug.sh"]) == {
+        "platform": False,
+        "canvas": False,
+        "python": False,
+        "scripts": True,
+    }
+    assert mod.classify("ci", [".gitea/workflows/ci.yml", "README.md"]) == {
+        "platform": False,
+        "canvas": False,
+        "python": False,
+        "scripts": False,
+    }
+
+
+def test_handlers_postgres_profile_is_narrower_than_workspace_server():
+    mod = load_module()
+
+    assert mod.classify("handlers-postgres", ["workspace-server/internal/handlers/a2a_proxy.go"]) == {
+        "handlers": True,
+    }
+    assert mod.classify("handlers-postgres", ["workspace-server/internal/provisioner/provisioner.go"]) == {
+        "handlers": False,
+    }
+
+
+def test_e2e_api_profile_covers_api_inputs():
+    mod = load_module()
+
+    assert mod.classify("e2e-api", ["workspace-server/internal/handlers/workspace.go"]) == {
+        "api": True,
+    }
+    assert mod.classify("e2e-api", ["tests/e2e/test_api.sh"]) == {"api": True}
+    assert mod.classify("e2e-api", ["canvas/src/app/page.tsx"]) == {"api": False}
+
+
+def test_fail_open_all_true_for_missing_base():
+    mod = load_module()
+
+    assert mod.all_true("ci") == {
+        "platform": True,
+        "canvas": True,
+        "python": True,
+        "scripts": True,
+    }
+
+
+def test_fetch_base_prefers_advertised_base_ref(monkeypatch):
+    mod = load_module()
+    calls: list[list[str]] = []
+    exists_checks = 0
+
+    def fake_base_exists(base: str) -> bool:
+        nonlocal exists_checks
+        exists_checks += 1
+        return exists_checks >= 1
+
+    def fake_run_git(args: list[str], *, timeout: int = 30):
+        calls.append(args)
+
+        class Result:
+            returncode = 0
+            stdout = ""
+            stderr = ""
+
+        return Result()
+
+    monkeypatch.setattr(mod, "base_exists", fake_base_exists)
+    monkeypatch.setattr(mod, "run_git", fake_run_git)
+
+    mod.fetch_base("abc123", "main")
+
+    assert calls == [["fetch", "--depth=1", "origin", "main"]]
+
+
+def test_fetch_base_falls_back_to_sha_when_ref_fetch_does_not_materialize(monkeypatch):
+    mod = load_module()
+    calls: list[list[str]] = []
+
+    monkeypatch.setattr(mod, "base_exists", lambda _base: False)
+
+    def fake_run_git(args: list[str], *, timeout: int = 30):
+        calls.append(args)
+
+        class Result:
+            returncode = 0
+            stdout = ""
+            stderr = ""
+
+        return Result()
+
+    monkeypatch.setattr(mod, "run_git", fake_run_git)
+
+    mod.fetch_base("abc123", "main")
+
+    assert calls == [
+        ["fetch", "--depth=1", "origin", "main"],
+        ["fetch", "--depth=1", "origin", "abc123"],
+    ]
+
+
+def test_changed_paths_uses_merge_base_for_pull_request(monkeypatch):
+    mod = load_module()
+    calls: list[list[str]] = []
+
+    def fake_run_git(args: list[str], *, timeout: int = 30):
+        calls.append(args)
+
+        class Result:
+            returncode = 0
+            stdout = "workspace/agent.py\n"
+            stderr = ""
+
+        if args[0] == "merge-base":
+            Result.stdout = "merge123\n"
+        return Result()
+
+    monkeypatch.setattr(mod, "run_git", fake_run_git)
+
+    assert mod.changed_paths("base123", use_merge_base=True) == ["workspace/agent.py"]
+    assert calls == [
+        ["merge-base", "base123", "HEAD"],
+        ["diff", "--name-only", "merge123", "HEAD"],
+    ]
+
+
+def test_detect_deepens_base_ref_when_pr_merge_base_missing(monkeypatch):
+    mod = load_module()
+    calls: list[tuple[str, str | None]] = []
+    merge_base_calls = 0
+
+    monkeypatch.setattr(mod, "base_exists", lambda _base: True)
+
+    def fake_merge_base(base: str):
+        nonlocal merge_base_calls
+        merge_base_calls += 1
+        if merge_base_calls == 1:
+            return None
+        return "merge123"
+
+    def fake_deepen_base_ref(base_ref: str):
+        calls.append(("deepen", base_ref))
+
+    def fake_changed_paths(base: str, *, use_merge_base: bool):
+        calls.append(("changed", str(use_merge_base)))
+        return [".gitea/workflows/ci.yml"]
+
+    monkeypatch.setattr(mod, "merge_base", fake_merge_base)
+    monkeypatch.setattr(mod, "deepen_base_ref", fake_deepen_base_ref)
+    monkeypatch.setattr(mod, "changed_paths", fake_changed_paths)
+
+    assert mod.detect("ci", "pull_request", "base123", "", "main") == {
+        "platform": False,
+        "canvas": False,
+        "python": False,
+        "scripts": False,
+    }
+    assert calls == [("deepen", "main"), ("changed", "True")]
@@ -0,0 +1,28 @@
+from pathlib import Path
+
+import yaml
+
+
+ROOT = Path(__file__).resolve().parents[1]
+
+
+def workflow_on(path: Path):
+    doc = yaml.safe_load(path.read_text())
+    return doc.get("on") or doc.get(True)
+
+
+def test_browser_e2e_workflows_are_not_unconditional_pr_heavy_lanes():
+    workflows = [
+        ROOT / ".gitea/workflows/e2e-chat.yml",
+        ROOT / ".gitea/workflows/e2e-staging-canvas.yml",
+    ]
+
+    for path in workflows:
+        text = path.read_text()
+        events = workflow_on(path)
+
+        assert "workflow_dispatch" in events
+        assert "schedule" in events
+        assert "merge-queue" in text
+        assert "/issues/${{ github.event.pull_request.number }}/labels" in text
+        assert "PR is not in merge-queue" in text
@@ -26,9 +26,11 @@ import re
 import subprocess
 import sys
 import textwrap
+import importlib.util
 from pathlib import Path

 import pytest  # noqa: F401  (declares the dep)
+import yaml

 REPO_ROOT = Path(__file__).resolve().parents[1]
 SCRIPT = REPO_ROOT / ".gitea" / "scripts" / "lint-workflow-yaml.py"
@@ -616,16 +618,24 @@ def test_rule10_docker_info_head_in_separate_step_without_pipefail_passes(tmp_pa

 CI_WORKFLOW = REPO_ROOT / ".gitea" / "workflows" / "ci.yml"
 CI_SURFACES = ("platform", "canvas", "python", "scripts")
+DETECT_CHANGES_SCRIPT = REPO_ROOT / ".gitea" / "scripts" / "detect-changes.py"
+
+
+def _load_detect_changes():
+    spec = importlib.util.spec_from_file_location("detect_changes", DETECT_CHANGES_SCRIPT)
+    assert spec is not None
+    module = importlib.util.module_from_spec(spec)
+    assert spec.loader is not None
+    spec.loader.exec_module(module)
+    return module


 def _ci_change_patterns() -> dict[str, re.Pattern[str]]:
-    text = CI_WORKFLOW.read_text(encoding="utf-8")
-    patterns: dict[str, re.Pattern[str]] = {}
-    for surface, pattern in re.findall(
-        r'echo "(platform|canvas|python|scripts)=.*?grep -qE \'([^\']+)\'',
-        text,
-    ):
-        patterns[surface] = re.compile(pattern)
+    detect_changes = _load_detect_changes()
+    patterns = {
+        surface: re.compile(pattern)
+        for surface, pattern in detect_changes.PROFILES["ci"].items()
+    }
    assert set(patterns) == set(CI_SURFACES)
    return patterns

@@ -693,3 +703,58 @@ def test_ci_change_detector_docs_and_meta_scripts_do_not_trigger_surfaces():
        "python": False,
        "scripts": False,
    }
+
+
+def test_ci_platform_go_pr_steps_are_path_scoped():
+    doc = yaml.safe_load(CI_WORKFLOW.read_text(encoding="utf-8"))
+    platform = doc["jobs"]["platform-build"]
+    assert platform.get("needs") == "changes"
+
+    expensive_steps = [
+        step
+        for step in platform["steps"]
+        if step.get("uses")
+        or step.get("run", "").startswith("go ")
+        or "golangci-lint" in step.get("run", "")
+    ]
+    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
+
+
+def test_ci_canvas_nextjs_pr_steps_are_path_scoped():
+    doc = yaml.safe_load(CI_WORKFLOW.read_text(encoding="utf-8"))
+    canvas = doc["jobs"]["canvas-build"]
+    assert canvas.get("needs") == "changes"
+
+    expensive_steps = [
+        step
+        for step in canvas["steps"]
+        if step.get("uses")
+        or step.get("run", "").startswith("npm ")
+        or step.get("run", "").startswith("npx ")
+    ]
+    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
+
+
+def test_ci_shellcheck_pr_steps_are_path_scoped():
+    doc = yaml.safe_load(CI_WORKFLOW.read_text(encoding="utf-8"))
+    shellcheck = doc["jobs"]["shellcheck"]
+    assert shellcheck.get("needs") == "changes"
+
+    expensive_steps = [
+        step
+        for step in shellcheck["steps"]
+        if step.get("uses") or step.get("run", "").startswith(("bash ", "find ", "shellcheck "))
+    ]
+    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
@@ -56,6 +56,21 @@ SCRIPT_PATH = (
 )


+@pytest.fixture(autouse=True)
+def _stub_time_sleep(monkeypatch):
+    """Autouse: stub time.sleep across every test.
+
+    The watchdog's RECHECK_DELAY_SECS (default 90s) is wired into
+    run_once() via time.sleep(). Without this stub, integration-style
+    tests that exercise run_once() would each block for 90s — a
+    pre-fix `pytest -q` ran in ~0.1s; the unstubbed equivalent took
+    >4 minutes (task #394 review evidence). Stubbing here keeps the
+    suite fast and deterministic without requiring every red-path test
+    to remember the patch.
+    """
+    monkeypatch.setattr("time.sleep", lambda s: None)
+
+
@pytest.fixture(scope="module")
 def wd_module():
    """Import the script as a module under a known env."""
@@ -809,3 +824,214 @@ def test_require_runtime_env_exits_when_missing(wd_module, monkeypatch):
    with pytest.raises(SystemExit) as excinfo:
        wd_module._require_runtime_env()
    assert excinfo.value.code == 2
+
+
+# --------------------------------------------------------------------------
+# Action-run status filter + HEAD-recheck (task #394, mc#1597..1630)
+#
+# The existing cancel-cascade filter matched description=='Has been
+# cancelled' EXACTLY, but a 7-day DB sweep on 2026-05-20 showed that
+# only 76/702 (~11%) of action_run.status=3 (Cancelled) entries carry
+# that string — 89% are written as 'Failing after Ns', indistinguishable
+# from real action_run.status=2 (Failure) at the commit_status layer.
+#
+# Gitea 1.22.6 has NO REST endpoint exposing action_run.status, so the
+# canonical filter (status=2 only) cannot run from a Gitea Actions
+# runner. The next-best signal is the HEAD-recheck: re-fetch HEAD SHA
+# (or its combined status) right before filing. If HEAD moved on or
+# combined state recovered, the prior "red" was a transient
+# cancel-cascade and we skip-file.
+#
+# References:
+#   - reference_chronic_red_sweep_cancelled_vs_failed_filter
+#   - feedback_gitea_status_enum_use_helper_not_raw_int
+#   - reference_gitea_action_status_enum_corrected_2026_05_19
+#   - triage evidence 2026-05-21 04:55 (6 cancellation + 1 emission
+#     artifact across mc#1597,1605,1609,1613,1626,1627,1630)
+# --------------------------------------------------------------------------
+def test_head_recheck_skips_file_when_head_moved(wd_module, monkeypatch, capsys):
+    """When initial tick sees red at SHA_A but HEAD has since moved to
+    SHA_B (next commit landed mid-tick), the watchdog must NOT file.
+    Re-evaluation happens on the next cron tick against the new SHA.
+
+    REGRESSION CLASS: this guards mc#1597..#1630 — 7 false-positives
+    filed in 24h because cancel-cascade fired commit_status=failure
+    rows on SHAs that were already superseded by new merges."""
+    SHA_A = SHA_RED
+    SHA_B = SHA_GREEN
+    failed_ctx = [
+        {"context": "ci/test", "status": "failure",
+         "target_url": "/r/runs/100/jobs/0",
+         "description": "Failing after 12s"},
+    ]
+    # First branches read returns SHA_A; the second (recheck) returns SHA_B
+    # → watchdog detects HEAD drift and skip-files.
+    branches_responses = iter([
+        (200, _branches_response(SHA_A)),
+        (200, _branches_response(SHA_B)),
+    ])
+
+    def fake_api(method, path, *, body=None, query=None, expect_json=True):
+        if method == "GET" and path == "/repos/owner/repo/branches/main":
+            return next(branches_responses)
+        if method == "GET" and path == f"/repos/owner/repo/commits/{SHA_A}/status":
+            return (200, _combined_status("failure", failed_ctx))
+        if method == "POST" and path == "/repos/owner/repo/issues":
+            raise AssertionError(
+                "watchdog filed a phantom issue despite HEAD moving away "
+                "from the red SHA (regression: mc#1597..1630)"
+            )
+        if method == "GET" and path == "/repos/owner/repo/issues":
+            return (200, [])
+        raise AssertionError(f"unexpected api call: {method} {path}")
+
+    # Settling delay is no-op'd by the _stub_time_sleep autouse fixture.
+    monkeypatch.setattr(wd_module, "api", fake_api)
+    wd_module.run_once(dry_run=False)
+    captured = capsys.readouterr()
+    assert "head drift" in captured.out.lower() or "head moved" in captured.out.lower(), (
+        f"expected a notice about HEAD drift, got: {captured.out!r}"
+    )
+
+
+def test_head_recheck_skips_file_when_recheck_status_recovered(
+    wd_module, monkeypatch, capsys,
+):
+    """When initial tick sees red at SHA, but the post-settling recheck
+    on the SAME SHA shows combined status recovered (e.g. transient
+    cancel-cascade rolled forward to success on retry), skip-file.
+
+    This catches the mid-flight cancel-cascade window — the second
+    largest false-positive cluster in mc#1597..1630."""
+    failed_ctx_initial = [
+        {"context": "ci/test", "status": "failure",
+         "target_url": "/r/runs/100/jobs/0",
+         "description": "Failing after 12s"},
+    ]
+    recovered_ctx = [
+        {"context": "ci/test", "status": "success",
+         "target_url": "/r/runs/100/jobs/0",
+         "description": "Successful in 30s"},
+    ]
+    # Same SHA across both branch reads; status flips from failure→success
+    # between the two combined-status reads.
+    status_responses = iter([
+        (200, _combined_status("failure", failed_ctx_initial)),
+        (200, _combined_status("success", recovered_ctx)),
+    ])
+
+    def fake_api(method, path, *, body=None, query=None, expect_json=True):
+        if method == "GET" and path == "/repos/owner/repo/branches/main":
+            return (200, _branches_response(SHA_RED))
+        if method == "GET" and path == f"/repos/owner/repo/commits/{SHA_RED}/status":
+            return next(status_responses)
+        if method == "POST" and path == "/repos/owner/repo/issues":
+            raise AssertionError(
+                "watchdog filed a phantom issue despite combined status "
+                "recovering on recheck (mid-flight cancel-cascade window)"
+            )
+        if method == "GET" and path == "/repos/owner/repo/issues":
+            return (200, [])
+        raise AssertionError(f"unexpected api call: {method} {path}")
+
+    monkeypatch.setattr(wd_module, "api", fake_api)
+    wd_module.run_once(dry_run=False)
+    captured = capsys.readouterr()
+    assert "recovered" in captured.out.lower() or "settled" in captured.out.lower(), (
+        f"expected a notice about post-settling recovery, got: {captured.out!r}"
+    )
+
+
+def test_head_recheck_files_when_still_red_after_settling(
+    wd_module, monkeypatch,
+):
+    """When BOTH the initial detection AND the post-settling recheck
+    show the same SHA still red, file the issue. This is the genuine-
+    failure path the watchdog is designed to surface.
+
+    Locks the over-filter: a future change that always-skips after
+    recheck would dismiss real failures."""
+    failed_ctx = [
+        {"context": "ci/test", "status": "failure",
+         "target_url": "/r/runs/100/jobs/0",
+         "description": "Failing after 12s"},
+    ]
+    post_filed = {"value": False}
+
+    def fake_api(method, path, *, body=None, query=None, expect_json=True):
+        if method == "GET" and path == "/repos/owner/repo/branches/main":
+            return (200, _branches_response(SHA_RED))
+        if method == "GET" and path == f"/repos/owner/repo/commits/{SHA_RED}/status":
+            return (200, _combined_status("failure", failed_ctx))
+        if method == "GET" and path == "/repos/owner/repo/issues":
+            return (200, [])
+        if method == "GET" and path == "/repos/owner/repo/labels":
+            return (200, [{"id": 9, "name": "tier:high"}])
+        if method == "POST" and path == "/repos/owner/repo/issues":
+            post_filed["value"] = True
+            return (201, {"number": 999})
+        if method == "POST" and path == "/repos/owner/repo/issues/999/labels":
+            return (200, [])
+        raise AssertionError(f"unexpected api call: {method} {path}")
+
+    monkeypatch.setattr(wd_module, "api", fake_api)
+    wd_module.run_once(dry_run=False)
+    assert post_filed["value"], (
+        "genuine-failure path was skip-filed — head-recheck over-filter "
+        "regression (would suppress all real main-red alarms)"
+    )
+
+
+def test_head_recheck_skips_when_initial_was_only_cancel_cascade(
+    wd_module, monkeypatch,
+):
+    """Belt-and-braces: combined-status failure caused exclusively by
+    description='Has been cancelled' entries should still be filtered
+    by the EXISTING cancel-cascade filter — head-recheck must not
+    accidentally bypass it. Regression guard for the existing mc#1564
+    fix."""
+    failed_ctx = [
+        {"context": "ci/test", "status": "failure",
+         "description": "Has been cancelled"},
+    ]
+
+    def fake_api(method, path, *, body=None, query=None, expect_json=True):
+        if method == "GET" and path == "/repos/owner/repo/branches/main":
+            return (200, _branches_response(SHA_RED))
+        if method == "GET" and path == f"/repos/owner/repo/commits/{SHA_RED}/status":
+            return (200, _combined_status("failure", failed_ctx))
+        if method == "POST" and path == "/repos/owner/repo/issues":
+            raise AssertionError(
+                "cancel-cascade-only entry must be filtered before any "
+                "head-recheck logic runs"
+            )
+        if method == "GET" and path == "/repos/owner/repo/issues":
+            return (200, [])
+        # No commit-status recheck should happen because is_red() returned False
+        raise AssertionError(f"unexpected api call: {method} {path}")
+
+    monkeypatch.setattr(wd_module, "api", fake_api)
+    wd_module.run_once(dry_run=False)
+    # success: no AssertionError raised, no POST
+
+
+def test_resolve_action_run_status_returns_none_on_no_endpoint(wd_module):
+    """The action_run.status REST endpoint does NOT exist in Gitea
+    1.22.6 (verified empirically 2026-05-20 — /api/v1/.../actions/runs/N
+    returns HTTP 404 across all probe variants). The resolver must
+    return None gracefully so callers fall back to the description-
+    string + head-recheck heuristics.
+
+    This pins the extensibility hook: when a future Gitea release (or
+    an op-host proxy) exposes the endpoint, the resolver implementation
+    can be swapped in without touching the caller contract."""
+    # The function exists and is callable
+    assert hasattr(wd_module, "_resolve_action_run_status")
+    # A typical target_url shape from real Gitea commit_status rows:
+    target_url = "/molecule-ai/molecule-core/actions/runs/75020/jobs/0"
+    # Return None when no endpoint available
+    out = wd_module._resolve_action_run_status(target_url)
+    assert out is None, (
+        "resolver must return None when the action_run.status endpoint "
+        "isn't reachable — callers depend on the None-fallback path"
+    )
@@ -442,6 +442,46 @@ def test_reap_preserves_real_push(sr_module, monkeypatch):
    assert calls == []  # NO POST


+def test_reap_compensates_cancelled_real_push_status(sr_module, monkeypatch):
+    """Gitea 1.22.6 maps cancelled push runs to failure statuses.
+
+    A real push workflow with description exactly "Has been cancelled"
+    is cancel-cascade noise, not a defect signal. Status-reaper should
+    compensate it even though the workflow has a push trigger.
+    """
+    calls = []
+
+    def fake_api(method, path, *, body=None, query=None, expect_json=True):
+        calls.append((method, path, body))
+        return (201, {})
+
+    monkeypatch.setattr(sr_module, "api", fake_api)
+
+    workflow_map = {"ci": True}
+    combined = {
+        "state": "failure",
+        "statuses": [
+            {
+                "context": "ci / test (push)",
+                "status": "failure",
+                "description": "Has been cancelled",
+                "target_url": "https://example.test/actions/runs/1",
+            }
+        ],
+    }
+
+    counters = sr_module.reap(workflow_map, combined, SHA, dry_run=False)
+
+    assert counters["compensated"] == 1
+    assert counters["compensated_cancelled_push"] == 1
+    assert counters["preserved_real_push"] == 0
+    assert len(calls) == 1
+    assert calls[0][0] == "POST"
+    assert calls[0][1] == f"/repos/owner/repo/statuses/{SHA}"
+    assert calls[0][2]["context"] == "ci / test (push)"
+    assert calls[0][2]["state"] == "success"
+
+
 def test_reap_preserves_unknown_workflow(sr_module, monkeypatch, capsys):
    """Workflow not in map → ::notice:: + skip (conservative)."""
    monkeypatch.setattr(
@@ -686,11 +686,22 @@ func (h *WorkspaceHandler) resolveAgentURL(ctx context.Context, workspaceID stri
 		_ = db.CacheURL(ctx, workspaceID, agentURL)
 	}

-	// When the platform runs inside Docker, 127.0.0.1:{host_port} is
-	// unreachable (it's the platform container's own localhost, not the
-	// Docker host). Rewrite to the container's Docker-bridge hostname.
+	// When the platform runs inside Docker, a managed workspace's
+	// 127.0.0.1:{host_port} URL points at the Docker host and must be
+	// rewritten to the workspace container's Docker-bridge hostname.
+	// External runtimes are not managed containers; their local test/runtime
+	// URL is the target and must not be synthesized into ws-<id>:8000.
 	if strings.HasPrefix(agentURL, "http://127.0.0.1:") && h.provisioner != nil && platformInDocker {
-		agentURL = provisioner.InternalURL(workspaceID)
+		var wsRuntime string
+		if err := db.DB.QueryRowContext(ctx,
+			`SELECT COALESCE(runtime, 'langgraph') FROM workspaces WHERE id = $1`,
+			workspaceID,
+		).Scan(&wsRuntime); err != nil {
+			log.Printf("ProxyA2A: runtime lookup before Docker URL rewrite failed for %s: %v", workspaceID, err)
+		}
+		if !isExternalLikeRuntime(wsRuntime) {
+			agentURL = provisioner.InternalURL(workspaceID)
+		}
 	}
 	// SSRF defence: reject private/metadata URLs before making outbound call.
 	if err := isSafeURL(agentURL); err != nil {
@@ -1511,6 +1511,35 @@ func TestResolveAgentURL_DockerRewrite(t *testing.T) {
 	}
 }

+func TestResolveAgentURL_ExternalRuntimeLoopbackNotRewrittenInDocker(t *testing.T) {
+	mock := setupTestDB(t)
+	mr := setupTestRedis(t)
+	allowLoopbackForTest(t)
+	handler := NewWorkspaceHandler(newTestBroadcaster(), nil, "http://localhost:8080", t.TempDir())
+	waitForHandlerAsyncBeforeDBCleanup(t, handler)
+	handler.provisioner = &stubLocalProv{}
+
+	restore := setPlatformInDockerForTest(true)
+	defer restore()
+
+	agentURL := "http://127.0.0.1:55555"
+	mr.Set("ws:ws-external:url", agentURL)
+	mock.ExpectQuery("SELECT COALESCE\\(runtime").
+		WithArgs("ws-external").
+		WillReturnRows(sqlmock.NewRows([]string{"runtime"}).AddRow("external"))
+
+	url, perr := handler.resolveAgentURL(context.Background(), "ws-external")
+	if perr != nil {
+		t.Fatalf("unexpected error: %+v", perr)
+	}
+	if url != agentURL {
+		t.Errorf("external runtime loopback URL must not be rewritten; got %q want %q", url, agentURL)
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Errorf("unmet sqlmock expectations: %v", err)
+	}
+}
+
 // --- dispatchA2A direct unit tests ---

 func TestDispatchA2A_BuildRequestError(t *testing.T) {
@@ -0,0 +1,72 @@
+package handlers
+
+import (
+	"database/sql"
+	"fmt"
+	"log"
+	"net/http"
+
+	"github.com/Molecule-AI/molecule-monorepo/platform/internal/db"
+	"github.com/Molecule-AI/molecule-monorepo/platform/internal/wsauth"
+	"github.com/gin-gonic/gin"
+)
+
+// AdminWorkspaceTokenHandler lets tenant admins mint the first workspace
+// bearer for managed SaaS workspaces whose runtime receives its token later
+// through registry registration.
+type AdminWorkspaceTokenHandler struct{}
+
+func NewAdminWorkspaceTokenHandler() *AdminWorkspaceTokenHandler {
+	return &AdminWorkspaceTokenHandler{}
+}
+
+// Create handles POST /admin/workspaces/:id/tokens. The route must be mounted
+// behind AdminAuth; the plaintext token is returned exactly once.
+func (h *AdminWorkspaceTokenHandler) Create(c *gin.Context) {
+	workspaceID := c.Param("id")
+	if !validWorkspaceID(workspaceID) {
+		c.JSON(http.StatusBadRequest, gin.H{"error": "invalid workspace id"})
+		return
+	}
+
+	var existing string
+	err := db.DB.QueryRowContext(c.Request.Context(),
+		`SELECT id FROM workspaces WHERE id = $1 AND status <> 'removed'`,
+		workspaceID).Scan(&existing)
+	if err != nil {
+		if err == sql.ErrNoRows {
+			c.JSON(http.StatusNotFound, gin.H{"error": "workspace not found"})
+			return
+		}
+		log.Printf("admin workspace tokens: workspace lookup failed for %s: %v", workspaceID, err)
+		c.JSON(http.StatusInternalServerError, gin.H{"error": "workspace lookup failed"})
+		return
+	}
+
+	var count int
+	if err := db.DB.QueryRowContext(c.Request.Context(),
+		`SELECT COUNT(*) FROM workspace_auth_tokens WHERE workspace_id = $1 AND revoked_at IS NULL`,
+		workspaceID).Scan(&count); err != nil {
+		log.Printf("admin workspace tokens: count failed for %s: %v", workspaceID, err)
+		c.JSON(http.StatusInternalServerError, gin.H{"error": "failed to count tokens"})
+		return
+	}
+	if count >= maxTokensPerWorkspace {
+		c.JSON(http.StatusTooManyRequests, gin.H{"error": fmt.Sprintf("maximum %d active tokens per workspace", maxTokensPerWorkspace)})
+		return
+	}
+
+	token, err := wsauth.IssueToken(c.Request.Context(), db.DB, workspaceID)
+	if err != nil {
+		log.Printf("admin workspace tokens: issue failed for %s: %v", workspaceID, err)
+		c.JSON(http.StatusInternalServerError, gin.H{"error": "failed to create token"})
+		return
+	}
+
+	log.Printf("admin workspace tokens: issued token for workspace %s", workspaceID)
+	c.JSON(http.StatusCreated, gin.H{
+		"auth_token":   token,
+		"workspace_id": workspaceID,
+		"message":      "Save this token now — it cannot be retrieved again.",
+	})
+}
@@ -0,0 +1,102 @@
+package handlers
+
+import (
+	"encoding/json"
+	"errors"
+	"net/http"
+	"testing"
+
+	"github.com/DATA-DOG/go-sqlmock"
+	"github.com/gin-gonic/gin"
+)
+
+func TestAdminWorkspaceTokenHandler_Create_HappyPath(t *testing.T) {
+	mock, cleanup := withMockDB(t)
+	defer cleanup()
+
+	mock.ExpectQuery(`SELECT id FROM workspaces WHERE id = \$1 AND status <> 'removed'`).
+		WithArgs(wsUUID1).
+		WillReturnRows(sqlmock.NewRows([]string{"id"}).AddRow(wsUUID1))
+	mock.ExpectQuery(`SELECT COUNT\(\*\) FROM workspace_auth_tokens`).
+		WithArgs(wsUUID1).
+		WillReturnRows(sqlmock.NewRows([]string{"count"}).AddRow(0))
+	mock.ExpectExec(`INSERT INTO workspace_auth_tokens`).
+		WithArgs(wsUUID1, sqlmock.AnyArg(), sqlmock.AnyArg()).
+		WillReturnResult(sqlmock.NewResult(1, 1))
+
+	w := makeReq(t, NewAdminWorkspaceTokenHandler().Create, "POST",
+		"/admin/workspaces/"+wsUUID1+"/tokens", gin.Params{{Key: "id", Value: wsUUID1}})
+
+	if w.Code != http.StatusCreated {
+		t.Fatalf("expected 201, got %d: %s", w.Code, w.Body.String())
+	}
+	var body struct {
+		AuthToken   string `json:"auth_token"`
+		WorkspaceID string `json:"workspace_id"`
+	}
+	if err := json.Unmarshal(w.Body.Bytes(), &body); err != nil {
+		t.Fatalf("decode: %v", err)
+	}
+	if body.AuthToken == "" || body.WorkspaceID != wsUUID1 {
+		t.Fatalf("unexpected body: %+v", body)
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Errorf("unmet: %v", err)
+	}
+}
+
+func TestAdminWorkspaceTokenHandler_Create_MissingWorkspace(t *testing.T) {
+	mock, cleanup := withMockDB(t)
+	defer cleanup()
+
+	mock.ExpectQuery(`SELECT id FROM workspaces WHERE id = \$1 AND status <> 'removed'`).
+		WithArgs(wsUUID1).
+		WillReturnRows(sqlmock.NewRows([]string{"id"}))
+
+	w := makeReq(t, NewAdminWorkspaceTokenHandler().Create, "POST",
+		"/admin/workspaces/"+wsUUID1+"/tokens", gin.Params{{Key: "id", Value: wsUUID1}})
+
+	if w.Code != http.StatusNotFound {
+		t.Fatalf("expected 404, got %d: %s", w.Code, w.Body.String())
+	}
+}
+
+func TestAdminWorkspaceTokenHandler_Create_RateLimited(t *testing.T) {
+	mock, cleanup := withMockDB(t)
+	defer cleanup()
+
+	mock.ExpectQuery(`SELECT id FROM workspaces WHERE id = \$1 AND status <> 'removed'`).
+		WithArgs(wsUUID1).
+		WillReturnRows(sqlmock.NewRows([]string{"id"}).AddRow(wsUUID1))
+	mock.ExpectQuery(`SELECT COUNT\(\*\) FROM workspace_auth_tokens`).
+		WithArgs(wsUUID1).
+		WillReturnRows(sqlmock.NewRows([]string{"count"}).AddRow(maxTokensPerWorkspace))
+
+	w := makeReq(t, NewAdminWorkspaceTokenHandler().Create, "POST",
+		"/admin/workspaces/"+wsUUID1+"/tokens", gin.Params{{Key: "id", Value: wsUUID1}})
+
+	if w.Code != http.StatusTooManyRequests {
+		t.Fatalf("expected 429, got %d: %s", w.Code, w.Body.String())
+	}
+}
+
+func TestAdminWorkspaceTokenHandler_Create_IssueFails(t *testing.T) {
+	mock, cleanup := withMockDB(t)
+	defer cleanup()
+
+	mock.ExpectQuery(`SELECT id FROM workspaces WHERE id = \$1 AND status <> 'removed'`).
+		WithArgs(wsUUID1).
+		WillReturnRows(sqlmock.NewRows([]string{"id"}).AddRow(wsUUID1))
+	mock.ExpectQuery(`SELECT COUNT\(\*\) FROM workspace_auth_tokens`).
+		WithArgs(wsUUID1).
+		WillReturnRows(sqlmock.NewRows([]string{"count"}).AddRow(0))
+	mock.ExpectExec(`INSERT INTO workspace_auth_tokens`).
+		WillReturnError(errors.New("disk full"))
+
+	w := makeReq(t, NewAdminWorkspaceTokenHandler().Create, "POST",
+		"/admin/workspaces/"+wsUUID1+"/tokens", gin.Params{{Key: "id", Value: wsUUID1}})
+
+	if w.Code != http.StatusInternalServerError {
+		t.Fatalf("expected 500, got %d: %s", w.Code, w.Body.String())
+	}
+}
@@ -122,8 +122,22 @@ func (h *DelegationHandler) Delegate(c *gin.Context) {

 	// #548 — prevent self-delegation: a workspace delegating to itself
 	// acquires _run_lock twice on the same mutex, deadlocking permanently.
+	//
+	// #383 — the error message is the agent-visible string when this 400
+	// fires on the SDK's _delegate_sync_via_polling path. The previous
+	// terse "self-delegation not permitted" was correct but indistinct
+	// from a transient rate-limit or auth failure, so the LLM would
+	// re-attempt every 2-3s in a tight loop (chloe-dong tenant external
+	// workspace, 2026-05-20). The expanded message is explicit about
+	// (a) what just happened, (b) why it cannot succeed, (c) what to do
+	// instead — so the agent's retry heuristic recognizes the path as
+	// terminal and stops.
 	if sourceID == body.TargetID {
-		c.JSON(http.StatusBadRequest, gin.H{"error": "self-delegation not permitted"})
+		c.JSON(http.StatusBadRequest, gin.H{
+			"error":  "self-delegation not permitted",
+			"reason": "the source workspace and target workspace are the same; you cannot delegate a task to yourself",
+			"hint":   "do the work yourself, or pick a different peer via list_peers — retrying with the same target_id will fail every time",
+		})
 		return
 	}

@@ -256,24 +256,43 @@ func (h *DiscoveryHandler) Peers(c *gin.Context) {
 		peers = append(peers, siblings...)
 	}

-	// Children
+	// Children — exclude self defensively. A child row whose parent_id
+	// equals the requesting workspaceID can never legitimately be the
+	// caller (a workspace can't be its own child), but a future data-
+	// integrity defect (e.g. self-loop introduced by a buggy register
+	// path) would otherwise smuggle the caller back into its own peer
+	// list. The agent then attempts `delegate_task(<own_id>)`, which
+	// either deadlocks _run_lock (sync path) or hits the platform's
+	// self-delegation 400 in a tight loop (#383). The `w.id != $2`
+	// clause makes self-delegation-via-peer-list impossible regardless
+	// of DB state.
 	children, _ := queryPeerMaps(`
 		SELECT w.id, w.name, COALESCE(w.role, ''), w.tier, w.status,
 			   COALESCE(w.agent_card, 'null'::jsonb), COALESCE(w.url, ''),
 			   w.parent_id, w.active_tasks
-		FROM workspaces w WHERE w.parent_id = $1 AND w.status != 'removed'`, workspaceID)
+		FROM workspaces w WHERE w.parent_id = $1 AND w.id != $2 AND w.status != 'removed'`,
+		workspaceID, workspaceID)
 	peers = append(peers, children...)

-	// Parent
+	// Parent — same defense-in-depth. A workspace whose parent_id points
+	// to itself is data corruption, but the peer-list endpoint must not
+	// propagate that corruption back to the agent as a "peer who is also
+	// you" entry.
 	if parentID.Valid {
 		parent, _ := queryPeerMaps(`
 			SELECT w.id, w.name, COALESCE(w.role, ''), w.tier, w.status,
 				   COALESCE(w.agent_card, 'null'::jsonb), COALESCE(w.url, ''),
 				   w.parent_id, w.active_tasks
-			FROM workspaces w WHERE w.id = $1 AND w.status != 'removed'`, parentID.String)
+			FROM workspaces w WHERE w.id = $1 AND w.id != $2 AND w.status != 'removed'`,
+			parentID.String, workspaceID)
 		peers = append(peers, parent...)
 	}

+	// #383 final-line defense: even if a future code path adds a query
+	// that doesn't filter self, strip the caller's own row before
+	// returning. Cheap O(n) over a peer set bounded at <50 rows.
+	peers = excludeSelfFromPeers(peers, workspaceID)
+
 	peers = filterPeersByQuery(peers, c.Query("q"))

 	if peers == nil {
@@ -282,6 +301,32 @@ func (h *DiscoveryHandler) Peers(c *gin.Context) {
 	c.JSON(http.StatusOK, peers)
 }

+// excludeSelfFromPeers strips any peer entry whose ``id`` equals
+// ``workspaceID`` (the caller's own row). Final-line defense for #383
+// (self-delegation 400-loop on external workspaces): a peer-list that
+// includes the requester's own row is the root mechanism by which an
+// agent ends up delegating to itself. The pre-DB filters in Peers
+// already enforce `w.id != $caller` on each branch; this function
+// guarantees the contract holds regardless of which query path
+// returned the row, including future ones added without a self-filter.
+//
+// O(n) over a peer set bounded at <50 rows per `Peers` comment — well
+// below the hot-path overhead of the existing filterPeersByQuery.
+func excludeSelfFromPeers(peers []map[string]interface{}, workspaceID string) []map[string]interface{} {
+	if len(peers) == 0 {
+		return peers
+	}
+	out := make([]map[string]interface{}, 0, len(peers))
+	for _, p := range peers {
+		id, _ := p["id"].(string)
+		if id == workspaceID {
+			continue
+		}
+		out = append(out, p)
+	}
+	return out
+}
+
 // filterPeersByQuery returns peers whose name or role case-insensitively
 // contains q. Whitespace-trimmed empty q is a no-op (returns input unchanged).
 func filterPeersByQuery(peers []map[string]interface{}, q string) []map[string]interface{} {
@@ -125,14 +125,14 @@ func TestPeers_WithParent(t *testing.T) {
 		WillReturnRows(sqlmock.NewRows(peerCols).
 			AddRow("ws-sibling-2", "Sibling Two", "worker", 1, "online", []byte("null"), "http://localhost:8002", "ws-parent", 0))

-	// Expect children query
-	mock.ExpectQuery("SELECT w.id, w.name.*WHERE w.parent_id = \\$1 AND w.status").
-		WithArgs("ws-sibling-1").
+	// Expect children query — #383 added explicit `w.id != $2` self-filter
+	mock.ExpectQuery("SELECT w.id, w.name.*WHERE w.parent_id = \\$1 AND w.id != \\$2 AND w.status").
+		WithArgs("ws-sibling-1", "ws-sibling-1").
 		WillReturnRows(sqlmock.NewRows(peerCols))

-	// Expect parent query
-	mock.ExpectQuery("SELECT w.id, w.name.*WHERE w.id = \\$1 AND w.status").
-		WithArgs("ws-parent").
+	// Expect parent query — #383 added explicit `w.id != $2` self-filter
+	mock.ExpectQuery("SELECT w.id, w.name.*WHERE w.id = \\$1 AND w.id != \\$2 AND w.status").
+		WithArgs("ws-parent", "ws-sibling-1").
 		WillReturnRows(sqlmock.NewRows(peerCols).
 			AddRow("ws-parent", "Parent PM", "manager", 2, "online", []byte("null"), "http://localhost:8001", nil, 1))

@@ -228,9 +228,9 @@ func TestPeers_RootWorkspace_NoPeers(t *testing.T) {
 		WithArgs("ws-root-alone").
 		WillReturnRows(sqlmock.NewRows(peerCols))

-	// Children — none
-	mock.ExpectQuery("SELECT w.id, w.name.*WHERE w.parent_id = \\$1").
-		WithArgs("ws-root-alone").
+	// Children — none. #383 added explicit `w.id != $2` self-filter.
+	mock.ExpectQuery("SELECT w.id, w.name.*WHERE w.parent_id = \\$1 AND w.id != \\$2").
+		WithArgs("ws-root-alone", "ws-root-alone").
 		WillReturnRows(sqlmock.NewRows(peerCols))

 	// No parent query since parent_id is NULL
@@ -282,12 +282,14 @@ func peersFilterFixture(t *testing.T) (*DiscoveryHandler, sqlmock.Sqlmock) {
 			AddRow("ws-alpha", "Alpha Researcher", "researcher", 1, "online", []byte("null"), "http://a", "ws-pm", 0).
 			AddRow("ws-beta", "Beta Designer", "designer", 1, "online", []byte("null"), "http://b", "ws-pm", 0))

-	mock.ExpectQuery("SELECT w.id, w.name.*WHERE w.parent_id = \\$1 AND w.status").
-		WithArgs("ws-self").
+	// #383 — children query gained explicit `w.id != $2` self-filter.
+	mock.ExpectQuery("SELECT w.id, w.name.*WHERE w.parent_id = \\$1 AND w.id != \\$2 AND w.status").
+		WithArgs("ws-self", "ws-self").
 		WillReturnRows(sqlmock.NewRows(cols))

-	mock.ExpectQuery("SELECT w.id, w.name.*WHERE w.id = \\$1 AND w.status").
-		WithArgs("ws-pm").
+	// #383 — parent query gained explicit `w.id != $2` self-filter.
+	mock.ExpectQuery("SELECT w.id, w.name.*WHERE w.id = \\$1 AND w.id != \\$2 AND w.status").
+		WithArgs("ws-pm", "ws-self").
 		WillReturnRows(sqlmock.NewRows(cols).
 			AddRow("ws-pm", "PM Workspace", "manager", 2, "online", []byte("null"), "http://pm", nil, 1))

@@ -966,8 +968,9 @@ func TestPeers_DevModeFailOpen_AllowsBearerlessRequest(t *testing.T) {
 	mock.ExpectQuery("SELECT w.id.+WHERE w.parent_id IS NULL AND w.id").
 		WithArgs("ws-dev").
 		WillReturnRows(sqlmock.NewRows(peerCols))
-	mock.ExpectQuery("SELECT w.id.+WHERE w.parent_id = \\$1 AND w.status").
-		WithArgs("ws-dev").
+	// #383 — children query gained explicit `w.id != $2` self-filter.
+	mock.ExpectQuery("SELECT w.id.+WHERE w.parent_id = \\$1 AND w.id != \\$2 AND w.status").
+		WithArgs("ws-dev", "ws-dev").
 		WillReturnRows(sqlmock.NewRows(peerCols))

 	w := httptest.NewRecorder()
@@ -1030,3 +1033,183 @@ func TestPeers_DevModeFailOpen_ClosedInProduction(t *testing.T) {
 		t.Fatalf("expected 401 in production, got %d: %s", w.Code, w.Body.String())
 	}
 }
+
+// ==================== Peers — #383 self never appears in result ====================
+
+// TestPeers_ExcludeSelf_DefenseInDepth verifies the final-line filter in
+// Peers strips any row whose id matches the caller. The pre-DB SQL filters
+// already do this, but a future code path that omits the `w.id != $caller`
+// clause must not be able to smuggle a self-row through. This test
+// simulates that future-defect case by mocking the children query to
+// (incorrectly) return a row whose id matches the caller, and asserts the
+// final filter still drops it.
+//
+// Root cause class for #383: an agent that sees its own row in /peers
+// proceeds to delegate_task to itself, hitting the platform's
+// self-delegation 400 in a tight loop. The fix in discovery.go is
+// defense-in-depth: even if the SQL filter regresses, this handler-level
+// filter prevents the 400-loop from materializing.
+func TestPeers_ExcludeSelf_DefenseInDepth(t *testing.T) {
+	mock := setupTestDB(t)
+	setupTestRedis(t)
+	handler := NewDiscoveryHandler()
+
+	const selfID = "ws-xiaodong"
+
+	// parent_id lookup — workspace has a parent.
+	mock.ExpectQuery("SELECT parent_id FROM workspaces WHERE id =").
+		WithArgs(selfID).
+		WillReturnRows(sqlmock.NewRows([]string{"parent_id"}).AddRow("ws-parent"))
+
+	peerCols := []string{"id", "name", "role", "tier", "status", "agent_card", "url", "parent_id", "active_tasks"}
+
+	// Siblings — returns one legitimate sibling. The SQL filter excludes
+	// self at the source.
+	mock.ExpectQuery("SELECT w.id, w.name.*WHERE w.parent_id = \\$1 AND w.id != \\$2").
+		WithArgs("ws-parent", selfID).
+		WillReturnRows(sqlmock.NewRows(peerCols).
+			AddRow("ws-sibling", "Sibling", "worker", 1, "online", []byte("null"), "http://localhost:8002", "ws-parent", 0))
+
+	// Children — simulates the data-integrity defect class: the DB
+	// (incorrectly) returns the caller's own row in the children set.
+	// In real production this would require a workspace whose
+	// parent_id points to itself — corruption only, but the handler
+	// must not propagate it.
+	mock.ExpectQuery("SELECT w.id, w.name.*WHERE w.parent_id = \\$1 AND w.id != \\$2 AND w.status").
+		WithArgs(selfID, selfID).
+		WillReturnRows(sqlmock.NewRows(peerCols).
+			AddRow(selfID, "Self As Child", "worker", 1, "online", []byte("null"), "http://localhost:8001", selfID, 0).
+			AddRow("ws-child", "Real Child", "worker", 1, "online", []byte("null"), "http://localhost:8003", selfID, 0))
+
+	// Parent — explicit `w.id != $2` clause so the parent path is also
+	// self-filtered. parentID.String = "ws-parent" != selfID, so the
+	// row is included.
+	mock.ExpectQuery("SELECT w.id, w.name.*WHERE w.id = \\$1 AND w.id != \\$2 AND w.status").
+		WithArgs("ws-parent", selfID).
+		WillReturnRows(sqlmock.NewRows(peerCols).
+			AddRow("ws-parent", "Parent", "manager", 2, "online", []byte("null"), "http://localhost:8004", nil, 1))
+
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Params = gin.Params{{Key: "id", Value: selfID}}
+	c.Request = httptest.NewRequest("GET", "/registry/"+selfID+"/peers", nil)
+
+	handler.Peers(c)
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("expected 200, got %d: %s", w.Code, w.Body.String())
+	}
+
+	var peers []map[string]interface{}
+	if err := json.Unmarshal(w.Body.Bytes(), &peers); err != nil {
+		t.Fatalf("failed to parse response: %v", err)
+	}
+
+	// The defense-in-depth filter must drop the self row even though
+	// the (mocked-defective) children query returned it.
+	for _, p := range peers {
+		if id, _ := p["id"].(string); id == selfID {
+			t.Fatalf("peer list contains caller's own id %q — self-delegation defense regressed; full list: %+v", selfID, peers)
+		}
+	}
+
+	// Sanity: the three legitimate peers (sibling, real child, parent)
+	// must all be present. Catches an over-aggressive filter that
+	// strips legitimate rows.
+	expectedIDs := map[string]bool{"ws-sibling": false, "ws-child": false, "ws-parent": false}
+	for _, p := range peers {
+		if id, _ := p["id"].(string); id != "" {
+			if _, ok := expectedIDs[id]; ok {
+				expectedIDs[id] = true
+			}
+		}
+	}
+	for id, found := range expectedIDs {
+		if !found {
+			t.Errorf("legitimate peer %q missing from response; got %+v", id, peers)
+		}
+	}
+
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Errorf("unmet sqlmock expectations: %v", err)
+	}
+}
+
+// TestExcludeSelfFromPeers_Unit exercises the helper directly so the
+// defense-in-depth contract is asserted independently of SQL mocking.
+// Pure-function tests run in microseconds and pin the filter shape
+// (empty input, no-match passthrough, single-row drop, multi-row drop,
+// preserves order) so future edits to the helper can't silently
+// regress to "returns input unchanged".
+func TestExcludeSelfFromPeers_Unit(t *testing.T) {
+	t.Run("empty input returns empty slice", func(t *testing.T) {
+		out := excludeSelfFromPeers(nil, "ws-self")
+		if len(out) != 0 {
+			t.Errorf("expected empty, got %+v", out)
+		}
+	})
+
+	t.Run("no self in list passes through unchanged", func(t *testing.T) {
+		in := []map[string]interface{}{
+			{"id": "ws-a", "name": "A"},
+			{"id": "ws-b", "name": "B"},
+		}
+		out := excludeSelfFromPeers(in, "ws-self")
+		if len(out) != 2 {
+			t.Fatalf("expected 2, got %d (%+v)", len(out), out)
+		}
+		if out[0]["id"] != "ws-a" || out[1]["id"] != "ws-b" {
+			t.Errorf("order not preserved: %+v", out)
+		}
+	})
+
+	t.Run("self row dropped, others preserved", func(t *testing.T) {
+		in := []map[string]interface{}{
+			{"id": "ws-a", "name": "A"},
+			{"id": "ws-self", "name": "Me"},
+			{"id": "ws-b", "name": "B"},
+		}
+		out := excludeSelfFromPeers(in, "ws-self")
+		if len(out) != 2 {
+			t.Fatalf("expected 2, got %d (%+v)", len(out), out)
+		}
+		if out[0]["id"] != "ws-a" || out[1]["id"] != "ws-b" {
+			t.Errorf("expected [ws-a, ws-b], got %+v", out)
+		}
+	})
+
+	t.Run("multiple self rows all dropped", func(t *testing.T) {
+		// Pathological — should never happen, but the contract is
+		// "no row with id==workspaceID survives", not "at most one
+		// such row is dropped". Pin it.
+		in := []map[string]interface{}{
+			{"id": "ws-self", "name": "Me1"},
+			{"id": "ws-a", "name": "A"},
+			{"id": "ws-self", "name": "Me2"},
+		}
+		out := excludeSelfFromPeers(in, "ws-self")
+		if len(out) != 1 {
+			t.Fatalf("expected 1, got %d (%+v)", len(out), out)
+		}
+		if out[0]["id"] != "ws-a" {
+			t.Errorf("expected [ws-a], got %+v", out)
+		}
+	})
+
+	t.Run("row with missing id key is preserved (not a self-collision)", func(t *testing.T) {
+		// A peer row with no "id" key shouldn't be silently dropped
+		// by the self-filter — it's a malformed row class that
+		// belongs to a different defect.
+		in := []map[string]interface{}{
+			{"name": "no-id-row"},
+			{"id": "ws-self", "name": "Me"},
+		}
+		out := excludeSelfFromPeers(in, "ws-self")
+		if len(out) != 1 {
+			t.Fatalf("expected 1, got %d (%+v)", len(out), out)
+		}
+		if out[0]["name"] != "no-id-row" {
+			t.Errorf("expected no-id-row preserved, got %+v", out)
+		}
+	})
+}
@@ -432,9 +432,10 @@ func TestExtended_Peers(t *testing.T) {
 		WillReturnRows(sqlmock.NewRows([]string{"id", "name", "role", "tier", "status", "agent_card", "url", "parent_id", "active_tasks"}).
 			AddRow("ws-sibling", "Sibling Agent", "worker", 1, "online", []byte("null"), "http://localhost:9001", nil, 0))

-	// Expect children query (workspaces with parent_id = ws-peer)
+	// Expect children query (workspaces with parent_id = ws-peer, excluding self)
+	// Query now binds (parent_id, self_id) for the self-filter guard added in #383.
 	mock.ExpectQuery("SELECT w.id, w.name").
-		WithArgs("ws-peer").
+		WithArgs("ws-peer", "ws-peer").
 		WillReturnRows(sqlmock.NewRows([]string{"id", "name", "role", "tier", "status", "agent_card", "url", "parent_id", "active_tasks"}))

 	// No parent query since workspace is root-level
@@ -84,6 +84,7 @@ type mcpTool struct {
 type MCPHandler struct {
 	database    *sql.DB
 	broadcaster *events.Broadcaster
+	a2aProxy    func(ctx context.Context, workspaceID string, body []byte, callerID string, logActivity bool) (int, []byte, error)

 	// memv2 is the v2 memory plugin wiring (RFC #2728). nil-safe:
 	// every v2 tool calls memoryV2Available() first and returns a
@@ -98,6 +99,14 @@ func NewMCPHandler(database *sql.DB, broadcaster *events.Broadcaster) *MCPHandle
 	return &MCPHandler{database: database, broadcaster: broadcaster}
 }

+func (h *MCPHandler) proxyA2ARequest(ctx context.Context, workspaceID string, body []byte, callerID string, logActivity bool) (int, []byte, error) {
+	if h.a2aProxy != nil {
+		return h.a2aProxy(ctx, workspaceID, body, callerID, logActivity)
+	}
+	wh := NewWorkspaceHandler(h.broadcaster, nil, "", "")
+	return wh.ProxyA2ARequest(ctx, workspaceID, body, callerID, logActivity)
+}
+
 // ─────────────────────────────────────────────────────────────────────────────
 // Tool definitions (mirrors workspace/a2a_mcp_server.py TOOLS list)
 // ─────────────────────────────────────────────────────────────────────────────
@@ -53,6 +53,15 @@ func mcpPost(t *testing.T, h *MCPHandler, workspaceID string, body interface{})
 	return w
 }

+func expectCanCommunicateSiblings(mock sqlmock.Sqlmock, callerID, targetID, parentID string) {
+	mock.ExpectQuery(`SELECT id, parent_id FROM workspaces WHERE id = \$1`).
+		WithArgs(callerID).
+		WillReturnRows(sqlmock.NewRows([]string{"id", "parent_id"}).AddRow(callerID, parentID))
+	mock.ExpectQuery(`SELECT id, parent_id FROM workspaces WHERE id = \$1`).
+		WithArgs(targetID).
+		WillReturnRows(sqlmock.NewRows([]string{"id", "parent_id"}).AddRow(targetID, parentID))
+}
+
 // ─────────────────────────────────────────────────────────────────────────────
 // initialize
 // ─────────────────────────────────────────────────────────────────────────────
@@ -178,6 +187,98 @@ func TestMCPHandler_ToolsList_ContainsExpectedTools(t *testing.T) {
 	}
 }

+func TestMCPHandler_DelegateTask_RoutesThroughPlatformA2AProxy(t *testing.T) {
+	h, mock := newMCPHandler(t)
+	callerID := "11111111-1111-1111-1111-111111111111"
+	targetID := "22222222-2222-2222-2222-222222222222"
+	parentID := "33333333-3333-3333-3333-333333333333"
+
+	expectCanCommunicateSiblings(mock, callerID, targetID, parentID)
+	mock.ExpectExec(`(?s)INSERT INTO activity_logs.*'delegation'.*'delegate'`).
+		WithArgs(callerID, callerID, targetID, "Delegating to "+targetID, sqlmock.AnyArg()).
+		WillReturnResult(sqlmock.NewResult(1, 1))
+	mock.ExpectExec(`UPDATE activity_logs`).
+		WithArgs("dispatched", "", callerID, sqlmock.AnyArg()).
+		WillReturnResult(sqlmock.NewResult(0, 1))
+
+	var gotTarget, gotCaller string
+	h.a2aProxy = func(ctx context.Context, workspaceID string, body []byte, callerID string, logActivity bool) (int, []byte, error) {
+		gotTarget = workspaceID
+		gotCaller = callerID
+		if !logActivity {
+			t.Fatal("delegate_task should log through platform A2A proxy")
+		}
+		if !strings.Contains(string(body), "do work") {
+			t.Fatalf("A2A body missing task text: %s", string(body))
+		}
+		return 200, []byte(`{"result":{"message":{"parts":[{"text":"done"}]}}}`), nil
+	}
+
+	out, err := h.toolDelegateTask(context.Background(), callerID, map[string]interface{}{
+		"workspace_id": targetID,
+		"task":         "do work",
+	}, mcpCallTimeout)
+	if err != nil {
+		t.Fatalf("delegate_task returned error: %v", err)
+	}
+	if out != "done" {
+		t.Fatalf("delegate_task response = %q, want done", out)
+	}
+	if gotTarget != targetID || gotCaller != callerID {
+		t.Fatalf("proxy called with target=%q caller=%q, want target=%q caller=%q", gotTarget, gotCaller, targetID, callerID)
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Fatalf("unmet expectations: %v", err)
+	}
+}
+
+func TestMCPHandler_DelegateTaskAsync_RoutesThroughPlatformA2AProxy(t *testing.T) {
+	h, mock := newMCPHandler(t)
+	callerID := "11111111-1111-1111-1111-111111111111"
+	targetID := "22222222-2222-2222-2222-222222222222"
+	parentID := "33333333-3333-3333-3333-333333333333"
+
+	expectCanCommunicateSiblings(mock, callerID, targetID, parentID)
+	mock.ExpectExec(`(?s)INSERT INTO activity_logs.*'delegation'.*'delegate'`).
+		WithArgs(callerID, callerID, targetID, "Delegating to "+targetID, sqlmock.AnyArg()).
+		WillReturnResult(sqlmock.NewResult(1, 1))
+	mock.ExpectExec(`UPDATE activity_logs`).
+		WithArgs("dispatched", "", callerID, sqlmock.AnyArg()).
+		WillReturnResult(sqlmock.NewResult(0, 1))
+
+	called := make(chan struct{}, 1)
+	h.a2aProxy = func(ctx context.Context, workspaceID string, body []byte, proxyCallerID string, logActivity bool) (int, []byte, error) {
+		if workspaceID != targetID || proxyCallerID != callerID {
+			t.Fatalf("unexpected proxy route target=%q caller=%q", workspaceID, proxyCallerID)
+		}
+		if !strings.Contains(string(body), "async work") {
+			t.Fatalf("A2A body missing task text: %s", string(body))
+		}
+		called <- struct{}{}
+		return 200, []byte(`{"result":{"message":{"parts":[{"text":"accepted"}]}}}`), nil
+	}
+
+	out, err := h.toolDelegateTaskAsync(context.Background(), callerID, map[string]interface{}{
+		"workspace_id": targetID,
+		"task":         "async work",
+	})
+	if err != nil {
+		t.Fatalf("delegate_task_async returned error: %v", err)
+	}
+	if !strings.Contains(out, `"status":"dispatched"`) {
+		t.Fatalf("delegate_task_async response = %s", out)
+	}
+	waitGlobalAsyncForTest()
+	select {
+	case <-called:
+	default:
+		t.Fatal("async delegate did not call platform A2A proxy")
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Fatalf("unmet expectations: %v", err)
+	}
+}
+
 // ─────────────────────────────────────────────────────────────────────────────
 // notifications/initialized
 // ─────────────────────────────────────────────────────────────────────────────
@@ -7,24 +7,19 @@ package handlers
 // and A2A response parsing helpers.

 import (
-	"bytes"
 	"context"
 	"database/sql"
 	"encoding/json"
 	"errors"
 	"fmt"
-	"io"
 	"log"
-	"net/http"
 	"os"
-	"strings"
 	"time"

-	"github.com/Molecule-AI/molecule-monorepo/platform/internal/db"
-	"github.com/Molecule-AI/molecule-monorepo/platform/internal/provisioner"
 	"github.com/Molecule-AI/molecule-monorepo/platform/internal/registry"
 	"github.com/google/uuid"
 )
+
 // insertMCPDelegationRow writes a delegation activity row so the canvas
 // Agent Comms tab can show the task text for MCP-initiated delegations.
 // Mirrors insertDelegationRow (delegation.go) for the MCP tool path.
@@ -190,15 +185,6 @@ func (h *MCPHandler) toolDelegateTask(ctx context.Context, callerID string, args
 		// Non-fatal: still make the A2A call even if activity log write fails.
 	}

-	agentURL, err := mcpResolveURL(ctx, h.database, targetID)
-	if err != nil {
-		return "", err
-	}
-	// SSRF defence: reject private/metadata URLs before making outbound call.
-	if err := isSafeURL(agentURL); err != nil {
-		return "", fmt.Errorf("invalid workspace URL: %w", err)
-	}
-
 	a2aBody, err := json.Marshal(map[string]interface{}{
 		"jsonrpc": "2.0",
 		"id":      uuid.New().String(),
@@ -218,36 +204,17 @@ func (h *MCPHandler) toolDelegateTask(ctx context.Context, callerID string, args
 	reqCtx, cancel := context.WithTimeout(ctx, timeout)
 	defer cancel()

-	httpReq, err := http.NewRequestWithContext(reqCtx, "POST", agentURL+"/a2a", bytes.NewReader(a2aBody))
-	if err != nil {
-		return "", fmt.Errorf("failed to create request: %w", err)
-	}
-	httpReq.Header.Set("Content-Type", "application/json")
-	// X-Workspace-ID identifies this caller to the A2A proxy. The /workspaces/:id/a2a
-	// endpoint is intentionally outside WorkspaceAuth (agents do not hold bearer tokens
-	// to peer workspaces). Access control is enforced by CanCommunicate above, which
-	// already validated callerID → targetID before this request is constructed.
-	// callerID was authenticated by WorkspaceAuth on the MCP bridge entry point,
-	// so this header reflects a verified caller identity, not a spoofable value.
-	httpReq.Header.Set("X-Workspace-ID", callerID)
-
-	resp, err := http.DefaultClient.Do(httpReq)
+	status, body, err := h.proxyA2ARequest(reqCtx, targetID, a2aBody, callerID, true)
 	if err != nil {
 		updateMCPDelegationStatus(ctx, h.database, callerID, delegationID, "failed", err.Error())
-		return "", fmt.Errorf("A2A call failed: %w", err)
+		return "", fmt.Errorf("A2A proxy failed: %w", err)
+	}
+	if status < 200 || status >= 300 {
+		updateMCPDelegationStatus(ctx, h.database, callerID, delegationID, "failed", fmt.Sprintf("A2A proxy returned status %d", status))
+		return "", fmt.Errorf("A2A proxy returned status %d", status)
 	}
-	defer func() { _ = resp.Body.Close() }()
-
-	// A 200/500 from the peer still means the call was dispatched — only
-	// network errors are truly "failed". Status 'dispatched' is correct for
-	// any HTTP response (peer's A2A layer handles the actual processing).
 	updateMCPDelegationStatus(ctx, h.database, callerID, delegationID, "dispatched", "")

-	body, err := io.ReadAll(io.LimitReader(resp.Body, 1<<20))
-	if err != nil {
-		return "", fmt.Errorf("failed to read response: %w", err)
-	}
-
 	return extractA2AText(body), nil
 }

@@ -278,24 +245,13 @@ func (h *MCPHandler) toolDelegateTaskAsync(ctx context.Context, callerID string,

 	// Fire and forget in a detached goroutine. Use a background context so
 	// the call is not cancelled when the HTTP request completes.
-	// RFC internal#524 Layer 1: globalGoAsync — the detached call reads
-	// db.DB (mcpResolveURL + updateMCPDelegationStatus) and must be
-	// drained by drainTestAsync before any t.Cleanup-driven db.DB swap.
+	// RFC internal#524 Layer 1: globalGoAsync — the detached call reads db.DB
+	// through the platform A2A proxy and must be drained by drainTestAsync
+	// before any t.Cleanup-driven db.DB swap.
 	globalGoAsync(func() {
 		bgCtx, cancel := context.WithTimeout(context.Background(), mcpAsyncCallTimeout)
 		defer cancel()

-		agentURL, err := mcpResolveURL(bgCtx, h.database, targetID)
-		if err != nil {
-			log.Printf("MCPHandler.delegate_task_async: resolve URL for %s: %v", targetID, err)
-			return
-		}
-		// SSRF defence: reject private/metadata URLs before making outbound call.
-		if err := isSafeURL(agentURL); err != nil {
-			log.Printf("MCPHandler.delegate_task_async: unsafe URL for %s: %v", targetID, err)
-			return
-		}
-
 		a2aBody, _ := json.Marshal(map[string]interface{}{
 			"jsonrpc": "2.0",
 			"id":      delegationID,
@@ -309,22 +265,15 @@ func (h *MCPHandler) toolDelegateTaskAsync(ctx context.Context, callerID string,
 			},
 		})

-		httpReq, err := http.NewRequestWithContext(bgCtx, "POST", agentURL+"/a2a", bytes.NewReader(a2aBody))
-		if err != nil {
-			log.Printf("MCPHandler.delegate_task_async: create request: %v", err)
+		status, _, err := h.proxyA2ARequest(bgCtx, targetID, a2aBody, callerID, true)
+		if err != nil || status < 200 || status >= 300 {
+			if err != nil {
+				log.Printf("MCPHandler.delegate_task_async: A2A proxy to %s: %v", targetID, err)
+			} else {
+				log.Printf("MCPHandler.delegate_task_async: A2A proxy to %s returned status %d", targetID, status)
+			}
 			return
 		}
-		httpReq.Header.Set("Content-Type", "application/json")
-		httpReq.Header.Set("X-Workspace-ID", callerID)
-
-		resp, err := http.DefaultClient.Do(httpReq)
-		if err != nil {
-			log.Printf("MCPHandler.delegate_task_async: A2A call to %s: %v", targetID, err)
-			return
-		}
-		defer func() { _ = resp.Body.Close() }()
-		// Drain response so the connection can be reused.
-		_, _ = io.Copy(io.Discard, resp.Body)
 	})

 	return fmt.Sprintf(`{"task_id":%q,"status":"dispatched","target_id":%q}`, delegationID, targetID), nil
@@ -405,7 +354,6 @@ func (h *MCPHandler) toolSendMessageToUser(ctx context.Context, workspaceID stri
 	return "Message sent.", nil
 }

-
 func (h *MCPHandler) toolCommitMemory(ctx context.Context, workspaceID string, args map[string]interface{}) (string, error) {
 	// PR-6 (RFC #2728) compat shim: when the v2 plugin is wired
 	// (MEMORY_PLUGIN_URL set), translate legacy scope→namespace and
@@ -534,56 +482,6 @@ func (h *MCPHandler) toolRecallMemory(ctx context.Context, workspaceID string, a
 // Helpers
 // ─────────────────────────────────────────────────────────────────────────────

-// mcpResolveURL returns a routable URL for a workspace's A2A server.
-//
-// Resolution order:
-//  1. Docker-internal URL cache (set by provisioner; correct when platform is in Docker)
-//  2. Redis URL cache
-//  3. DB `url` column fallback, with 127.0.0.1→Docker bridge rewrite when in Docker
-//
-// SECURITY (F1083 / #1130): all three paths run the returned URL through
-// validateAgentURL to block SSRF targets (private IPs, loopback, cloud metadata).
-func mcpResolveURL(ctx context.Context, database *sql.DB, workspaceID string) (string, error) {
-	if platformInDocker {
-		if url, err := db.GetCachedInternalURL(ctx, workspaceID); err == nil && url != "" {
-			if err := validateAgentURL(url); err != nil {
-				return "", fmt.Errorf("workspace %s: forbidden URL from internal cache: %w", workspaceID, err)
-			}
-			return url, nil
-		}
-	}
-	if url, err := db.GetCachedURL(ctx, workspaceID); err == nil && url != "" {
-		if platformInDocker && strings.HasPrefix(url, "http://127.0.0.1:") {
-			return provisioner.InternalURL(workspaceID), nil
-		}
-		if err := validateAgentURL(url); err != nil {
-			return "", fmt.Errorf("workspace %s: forbidden URL from Redis cache: %w", workspaceID, err)
-		}
-		return url, nil
-	}
-
-	var urlStr sql.NullString
-	var status string
-	if err := database.QueryRowContext(ctx,
-		`SELECT url, status FROM workspaces WHERE id = $1`, workspaceID,
-	).Scan(&urlStr, &status); err != nil {
-		if err == sql.ErrNoRows {
-			return "", fmt.Errorf("workspace %s not found", workspaceID)
-		}
-		return "", fmt.Errorf("workspace lookup failed: %w", err)
-	}
-	if !urlStr.Valid || urlStr.String == "" {
-		return "", fmt.Errorf("workspace %s has no URL (status: %s)", workspaceID, status)
-	}
-	if platformInDocker && strings.HasPrefix(urlStr.String, "http://127.0.0.1:") {
-		return provisioner.InternalURL(workspaceID), nil
-	}
-	if err := validateAgentURL(urlStr.String); err != nil {
-		return "", fmt.Errorf("workspace %s: forbidden URL from DB: %w", workspaceID, err)
-	}
-	return urlStr.String, nil
-}
-
 // extractA2AText extracts human-readable text from an A2A JSON-RPC response body.
 // Falls back to the raw JSON when no text part can be found.
 func extractA2AText(body []byte) string {
@@ -632,4 +530,3 @@ func extractA2AText(body []byte) string {
 	b, _ := json.Marshal(result)
 	return string(b)
 }
-
@@ -112,7 +112,7 @@ func (h *RegistryHandler) SetQueueDrainFunc(f QueueDrainFunc) {
 // Go's net.ParseIP.To4() before Contains() runs, so the IPv4 rules above
 // catch those without a separate entry.
 //
-// F1083/#1130 (SSRF on mcpResolveURL / a2a_proxy resolveAgentURL): in
+// F1083/#1130 (SSRF on direct A2A URL resolution): in
 // addition to blocking IP literals, DNS names are now resolved and each
 // returned IP is checked against the blocklist. This closes the gap where
 // an attacker could register agent.example.com pointing to 169.254.169.254.
@@ -234,9 +234,13 @@ func (h *TemplatesHandler) ReplaceFiles(c *gin.Context) {
 			"source":    "ec2-ssh",
 		})
 		if h.wh != nil {
-			// RFC internal#524 Layer 1: per-handler goAsync (drains via h.wh.waitAsyncForTest)
-			wsID := workspaceID
-			h.wh.goAsync(func() { h.wh.RestartByID(wsID) })
+			// internal#624: 15s per-workspace debounce around the file-write
+			// → RestartByID trigger. Canvas Save / ReplaceFiles fires N PUTs
+			// in a burst; without this each PUT chains into the
+			// coalesceRestart drain loop. The helper still uses goAsync
+			// internally (drains via h.wh.waitAsyncForTest), preserving
+			// RFC internal#524 Layer 1.
+			h.wh.maybeRestartAfterFileWrite(workspaceID)
 		}
 		return
 	}
@@ -270,9 +274,13 @@ func (h *TemplatesHandler) ReplaceFiles(c *gin.Context) {
 			"source":    "container",
 		})
 		if h.wh != nil {
-			// RFC internal#524 Layer 1: per-handler goAsync (drains via h.wh.waitAsyncForTest)
-			wsID := workspaceID
-			h.wh.goAsync(func() { h.wh.RestartByID(wsID) })
+			// internal#624: 15s per-workspace debounce around the file-write
+			// → RestartByID trigger. Canvas Save / ReplaceFiles fires N PUTs
+			// in a burst; without this each PUT chains into the
+			// coalesceRestart drain loop. The helper still uses goAsync
+			// internally (drains via h.wh.waitAsyncForTest), preserving
+			// RFC internal#524 Layer 1.
+			h.wh.maybeRestartAfterFileWrite(workspaceID)
 		}
 		return
 	}
@@ -292,8 +300,12 @@ func (h *TemplatesHandler) ReplaceFiles(c *gin.Context) {

 	c.JSON(http.StatusOK, gin.H{"status": "replaced", "workspace": workspaceID, "files": len(body.Files), "source": "volume"})
 	if h.wh != nil {
-		// RFC internal#524 Layer 1: per-handler goAsync (drains via h.wh.waitAsyncForTest)
-		wsID := workspaceID
-		h.wh.goAsync(func() { h.wh.RestartByID(wsID) })
+		// internal#624: 15s per-workspace debounce around the file-write
+		// → RestartByID trigger. Canvas Save / ReplaceFiles fires N PUTs
+		// in a burst; without this each PUT chains into the
+		// coalesceRestart drain loop. The helper still uses goAsync
+		// internally (drains via h.wh.waitAsyncForTest), preserving
+		// RFC internal#524 Layer 1.
+		h.wh.maybeRestartAfterFileWrite(workspaceID)
 	}
 }
@@ -570,9 +570,13 @@ func (h *TemplatesHandler) WriteFile(c *gin.Context) {
 		}
 		c.JSON(http.StatusOK, gin.H{"status": "saved", "path": filePath})
 		if h.wh != nil {
-			// RFC internal#524 Layer 1: per-handler goAsync (drains via h.wh.waitAsyncForTest)
-			wsID := workspaceID
-			h.wh.goAsync(func() { h.wh.RestartByID(wsID) })
+			// internal#624: 15s per-workspace debounce around the file-write
+			// → RestartByID trigger. Canvas Save fires N PUTs in a burst;
+			// without this each PUT chains into the coalesceRestart drain
+			// loop and produces back-to-back EC2 recreate cycles. The
+			// helper still uses goAsync internally (drains via
+			// h.wh.waitAsyncForTest), preserving RFC internal#524 Layer 1.
+			h.wh.maybeRestartAfterFileWrite(workspaceID)
 		}
 		return
 	}
@@ -586,9 +590,13 @@ func (h *TemplatesHandler) WriteFile(c *gin.Context) {
 		}
 		c.JSON(http.StatusOK, gin.H{"status": "saved", "path": filePath})
 		if h.wh != nil {
-			// RFC internal#524 Layer 1: per-handler goAsync (drains via h.wh.waitAsyncForTest)
-			wsID := workspaceID
-			h.wh.goAsync(func() { h.wh.RestartByID(wsID) })
+			// internal#624: 15s per-workspace debounce around the file-write
+			// → RestartByID trigger. Canvas Save fires N PUTs in a burst;
+			// without this each PUT chains into the coalesceRestart drain
+			// loop and produces back-to-back EC2 recreate cycles. The
+			// helper still uses goAsync internally (drains via
+			// h.wh.waitAsyncForTest), preserving RFC internal#524 Layer 1.
+			h.wh.maybeRestartAfterFileWrite(workspaceID)
 		}
 		return
 	}
@@ -602,9 +610,13 @@ func (h *TemplatesHandler) WriteFile(c *gin.Context) {
 	}
 	c.JSON(http.StatusOK, gin.H{"status": "saved", "path": filePath})
 	if h.wh != nil {
-		// RFC internal#524 Layer 1: per-handler goAsync (drains via h.wh.waitAsyncForTest)
-		wsID := workspaceID
-		h.wh.goAsync(func() { h.wh.RestartByID(wsID) })
+		// internal#624: 15s per-workspace debounce around the file-write
+		// → RestartByID trigger. Canvas Save fires N PUTs in a burst;
+		// without this each PUT chains into the coalesceRestart drain
+		// loop and produces back-to-back EC2 recreate cycles. The
+		// helper still uses goAsync internally (drains via
+		// h.wh.waitAsyncForTest), preserving RFC internal#524 Layer 1.
+		h.wh.maybeRestartAfterFileWrite(workspaceID)
 	}
 }

@@ -657,9 +669,13 @@ func (h *TemplatesHandler) DeleteFile(c *gin.Context) {
 		}
 		c.JSON(http.StatusOK, gin.H{"status": "deleted", "path": filePath})
 		if h.wh != nil {
-			// RFC internal#524 Layer 1: per-handler goAsync (drains via h.wh.waitAsyncForTest)
-			wsID := workspaceID
-			h.wh.goAsync(func() { h.wh.RestartByID(wsID) })
+			// internal#624: 15s per-workspace debounce around the file-write
+			// → RestartByID trigger. Canvas Save fires N PUTs in a burst;
+			// without this each PUT chains into the coalesceRestart drain
+			// loop and produces back-to-back EC2 recreate cycles. The
+			// helper still uses goAsync internally (drains via
+			// h.wh.waitAsyncForTest), preserving RFC internal#524 Layer 1.
+			h.wh.maybeRestartAfterFileWrite(workspaceID)
 		}
 		return
 	}
@@ -677,9 +693,13 @@ func (h *TemplatesHandler) DeleteFile(c *gin.Context) {
 		}
 		c.JSON(http.StatusOK, gin.H{"status": "deleted", "path": filePath})
 		if h.wh != nil {
-			// RFC internal#524 Layer 1: per-handler goAsync (drains via h.wh.waitAsyncForTest)
-			wsID := workspaceID
-			h.wh.goAsync(func() { h.wh.RestartByID(wsID) })
+			// internal#624: 15s per-workspace debounce around the file-write
+			// → RestartByID trigger. Canvas Save fires N PUTs in a burst;
+			// without this each PUT chains into the coalesceRestart drain
+			// loop and produces back-to-back EC2 recreate cycles. The
+			// helper still uses goAsync internally (drains via
+			// h.wh.waitAsyncForTest), preserving RFC internal#524 Layer 1.
+			h.wh.maybeRestartAfterFileWrite(workspaceID)
 		}
 		return
 	}
@@ -692,8 +712,12 @@ func (h *TemplatesHandler) DeleteFile(c *gin.Context) {
 	}
 	c.JSON(http.StatusOK, gin.H{"status": "deleted", "path": filePath})
 	if h.wh != nil {
-		// RFC internal#524 Layer 1: per-handler goAsync (drains via h.wh.waitAsyncForTest)
-		wsID := workspaceID
-		h.wh.goAsync(func() { h.wh.RestartByID(wsID) })
+		// internal#624: 15s per-workspace debounce around the file-write
+		// → RestartByID trigger. Canvas Save fires N PUTs in a burst;
+		// without this each PUT chains into the coalesceRestart drain
+		// loop and produces back-to-back EC2 recreate cycles. The
+		// helper still uses goAsync internally (drains via
+		// h.wh.waitAsyncForTest), preserving RFC internal#524 Layer 1.
+		h.wh.maybeRestartAfterFileWrite(workspaceID)
 	}
 }
@@ -869,14 +869,31 @@ func applyRuntimeModelEnv(envVars map[string]string, runtime, model string) {
 // Returns nil map + error string on decrypt failure. Shared by both Docker
 // and control plane provisioning paths to avoid duplication.
 //
+// The second return value (globalKeys) records which keys originated from
+// the operator-controlled `global_secrets` table — used by RFC#523 Layer 1
+// to constrain its forbidden-key check to the operator-bleed channel,
+// instead of blanket-blocking by name across BOTH provenance channels (the
+// over-fire that breaks the legitimate user flow of pasting their own
+// GitHub PAT into the canvas Secrets tab → workspace_secrets row). See
+// `feedback_upstream_docs_first_before_hypothesizing`: RFC#523's threat
+// model (issue molecule-ai/internal#523 §"Threat model") names operator-
+// scope tokens being injected via provision-time env / operator-side
+// stores — NOT the user's own scoped PAT they explicitly authorized via
+// the per-workspace Secrets tab.
+//
+// The merged map preserves the existing precedence semantic (workspace
+// rows overwrite global rows on key collision); only the provenance side-
+// channel is new. Existing single-return callers can ignore globalKeys.
+//
 // F1086 / #1206: the returned error string is the SAFE-CANNED message that
 // gets persisted to workspaces.last_sample_error AND broadcast as the
 // WORKSPACE_PROVISION_FAILED payload. Internal detail (the secret key name,
 // the encryption version, the decrypt-error text) is logged here, never
 // returned to the caller, so it can't leak via the canvas event stream
 // (cf. TestProvisionWorkspace_NoInternalErrorsInBroadcast).
-func loadWorkspaceSecrets(ctx context.Context, workspaceID string) (map[string]string, string) {
+func loadWorkspaceSecrets(ctx context.Context, workspaceID string) (map[string]string, map[string]struct{}, string) {
 	envVars := map[string]string{}
+	globalKeys := map[string]struct{}{}
 	globalRows, globalErr := db.DB.QueryContext(ctx,
 		`SELECT key, encrypted_value, encryption_version FROM global_secrets`)
 	if globalErr == nil {
@@ -889,9 +906,10 @@ func loadWorkspaceSecrets(ctx context.Context, workspaceID string) (map[string]s
 				decrypted, decErr := crypto.DecryptVersioned(v, ver)
 				if decErr != nil {
 					log.Printf("Provisioner: FATAL — failed to decrypt global secret %s (version=%d): %v — aborting provision of workspace %s", k, ver, decErr, workspaceID)
-					return nil, "failed to decrypt global secret"
+					return nil, nil, "failed to decrypt global secret"
 				}
 				envVars[k] = string(decrypted)
+				globalKeys[k] = struct{}{}
 			}
 		}
 		if err := globalRows.Err(); err != nil {
@@ -910,16 +928,22 @@ func loadWorkspaceSecrets(ctx context.Context, workspaceID string) (map[string]s
 				decrypted, decErr := crypto.DecryptVersioned(v, ver)
 				if decErr != nil {
 					log.Printf("Provisioner: FATAL — failed to decrypt workspace secret %s (version=%d) for %s: %v — aborting provision", k, ver, workspaceID, decErr)
-					return nil, "failed to decrypt workspace secret"
+					return nil, nil, "failed to decrypt workspace secret"
 				}
 				envVars[k] = string(decrypted)
+				// User-authored workspace_secrets value supersedes any
+				// global_secrets row of the same key — including dropping
+				// the operator-bleed provenance flag. The user explicitly
+				// re-set the value via the canvas Secrets tab, so it is
+				// no longer "the operator-store version."
+				delete(globalKeys, k)
 			}
 		}
 		if err := wsRows.Err(); err != nil {
 			log.Printf("Provisioner: workspace_secrets rows.Err workspace=%s: %v", workspaceID, err)
 		}
 	}
-	return envVars, ""
+	return envVars, globalKeys, ""
 }

 // provisionWorkspaceCP provisions a workspace via the control plane API.
@@ -135,6 +135,15 @@ func isForbiddenTenantEnvKey(key string) bool {
 // message and the structured-extra payload that goes to the
 // canvas Events tab. Sorting makes the message stable across
 // Go's randomized map iteration.
+//
+// PROVENANCE NOTE: this helper checks by env-var name ONLY and is
+// unaware of where each value came from. Production provision code
+// uses findForbiddenTenantEnvKeysFromGlobals instead, restricting
+// the check to keys originating from the operator-controlled
+// global_secrets table — see the doc-comment on that function and
+// the RFC#523 Layer 1 block in prepareProvisionContext. This name-
+// only helper is kept for the workspace_secrets-write CI lint
+// (Layer 3) and for tests that pin the deny-set definition.
 func findForbiddenTenantEnvKeys(envVars map[string]string) []string {
 	if len(envVars) == 0 {
 		return []string{}
@@ -149,6 +158,48 @@ func findForbiddenTenantEnvKeys(envVars map[string]string) []string {
 	return found
 }

+// findForbiddenTenantEnvKeysFromGlobals is the provenance-aware
+// variant used by RFC#523 Layer 1 in prepareProvisionContext. It
+// restricts the forbidden-key scan to keys whose value originated
+// from the operator-controlled `global_secrets` table.
+//
+// Fixes the over-fire reported by CTO empirical 2026-05-20: a user
+// who explicitly pastes their own scoped GitHub PAT under
+// GITHUB_TOKEN into the canvas Secrets tab (a `workspace_secrets`
+// row) was being blocked alongside the genuine operator-bleed case.
+// RFC#523's threat model (issue molecule-ai/internal#523 §"Threat
+// model") names operator-scope tokens injected via operator-side
+// stores; user-authored workspace_secrets is out of scope.
+//
+// globalSecretKeys is the set returned as the second value from
+// loadWorkspaceSecrets. A key that exists in BOTH stores is treated
+// as workspace_secrets (user override wins) — loadWorkspaceSecrets
+// drops the global flag when the workspace row is read.
+//
+// Empty/nil globalSecretKeys means no operator-side source was
+// loaded (e.g. tests, or table empty); the scan returns no hits.
+// Deterministic sort order, same as findForbiddenTenantEnvKeys.
+func findForbiddenTenantEnvKeysFromGlobals(envVars map[string]string, globalSecretKeys map[string]struct{}) []string {
+	if len(envVars) == 0 || len(globalSecretKeys) == 0 {
+		return []string{}
+	}
+	found := make([]string, 0)
+	for k := range globalSecretKeys {
+		if _, present := envVars[k]; !present {
+			// Defensive: a key flagged as global-origin must also
+			// be in the resolved env-set. If not, skip — the
+			// loadWorkspaceSecrets contract guarantees this never
+			// happens, but the helper stays total.
+			continue
+		}
+		if isForbiddenTenantEnvKey(k) {
+			found = append(found, k)
+		}
+	}
+	sort.Strings(found)
+	return found
+}
+
 // formatForbiddenTenantEnvError builds the safe-canned user-facing
 // message for a provision aborted because forbidden env keys are
 // present in the resolved env-set. The message names the
@@ -150,6 +150,106 @@ func TestFindForbiddenTenantEnvKeys_SingleAndMultipleSorted(t *testing.T) {
 	}
 }

+// TestFindForbiddenTenantEnvKeysFromGlobals pins the provenance-aware
+// behaviour added 2026-05-20 to fix the RFC#523 Layer 1 over-fire: a
+// user-set workspace_secrets row with key=GITHUB_TOKEN must NOT be
+// flagged, while a global_secrets row of the same key MUST be.
+//
+// Cross-references the empirical bug: CTO 2026-05-20 hit
+// `provision aborted: env var "GITHUB_TOKEN" is operator-scope...`
+// after pasting their own scoped PAT into the canvas Secrets tab
+// (workspace_secrets) — the original blanket check fired on the
+// merged env-set regardless of provenance.
+func TestFindForbiddenTenantEnvKeysFromGlobals_UserSetAllowed(t *testing.T) {
+	// User pasted their own PAT via canvas Secrets tab —
+	// workspace_secrets row only. globalSecretKeys is empty for
+	// this key, so the check MUST not fire.
+	envVars := map[string]string{
+		"GITHUB_TOKEN":      "ghp_FAKEUSERPAT_user_set_via_canvas",
+		"ANTHROPIC_API_KEY": "sk-ant-keep",
+	}
+	globalKeys := map[string]struct{}{} // nothing from global_secrets
+	got := findForbiddenTenantEnvKeysFromGlobals(envVars, globalKeys)
+	if len(got) != 0 {
+		t.Errorf("user-set workspace_secrets with GITHUB_TOKEN: got %v; want empty (provenance-allowed)", got)
+	}
+}
+
+func TestFindForbiddenTenantEnvKeysFromGlobals_OperatorLeakBlocked(t *testing.T) {
+	// Operator-store bleed — GITHUB_TOKEN sourced from global_secrets.
+	// This is the literal RFC#523 §"Threat model" attack vector.
+	// Check MUST fire and name GITHUB_TOKEN.
+	envVars := map[string]string{
+		"GITHUB_TOKEN":      "ghp_OPERATOR_LEAK_from_global_secrets",
+		"ANTHROPIC_API_KEY": "sk-ant-keep",
+	}
+	globalKeys := map[string]struct{}{
+		"GITHUB_TOKEN":      {},
+		"ANTHROPIC_API_KEY": {},
+	}
+	got := findForbiddenTenantEnvKeysFromGlobals(envVars, globalKeys)
+	if len(got) != 1 || got[0] != "GITHUB_TOKEN" {
+		t.Errorf("operator-leak GITHUB_TOKEN in global_secrets: got %v; want [GITHUB_TOKEN]", got)
+	}
+}
+
+func TestFindForbiddenTenantEnvKeysFromGlobals_UserOverrideOfGlobalAllowed(t *testing.T) {
+	// Both stores have the key; loadWorkspaceSecrets drops the global
+	// flag when the workspace row supersedes (caller contract).
+	// Simulate that here: globalKeys does NOT contain GITHUB_TOKEN
+	// because workspace_secrets re-set it. Allowed.
+	envVars := map[string]string{
+		"GITHUB_TOKEN": "ghp_USER_RESET_after_global_was_present",
+	}
+	globalKeys := map[string]struct{}{} // workspace overrode → flag dropped
+	got := findForbiddenTenantEnvKeysFromGlobals(envVars, globalKeys)
+	if len(got) != 0 {
+		t.Errorf("user-override of global GITHUB_TOKEN: got %v; want empty", got)
+	}
+}
+
+func TestFindForbiddenTenantEnvKeysFromGlobals_MultipleOperatorLeaks(t *testing.T) {
+	// Multiple operator-leaked tokens — must return sorted slice.
+	envVars := map[string]string{
+		"GITHUB_TOKEN":           "leak1",
+		"CP_ADMIN_API_TOKEN":     "leak2",
+		"MOLECULE_OPERATOR_HOST": "leak3",
+		"RAILWAY_TOKEN":          "leak4",
+		"ANTHROPIC_API_KEY":      "user-allowed",
+	}
+	globalKeys := map[string]struct{}{
+		"GITHUB_TOKEN":           {},
+		"CP_ADMIN_API_TOKEN":     {},
+		"MOLECULE_OPERATOR_HOST": {},
+		"RAILWAY_TOKEN":          {},
+	}
+	got := findForbiddenTenantEnvKeysFromGlobals(envVars, globalKeys)
+	want := []string{"CP_ADMIN_API_TOKEN", "GITHUB_TOKEN", "MOLECULE_OPERATOR_HOST", "RAILWAY_TOKEN"}
+	if len(got) != len(want) {
+		t.Fatalf("operator-leak multi: got %v; want %v", got, want)
+	}
+	for i := range want {
+		if got[i] != want[i] {
+			t.Errorf("operator-leak multi[%d] = %q; want %q (full got=%v)", i, got[i], want[i], got)
+		}
+	}
+}
+
+func TestFindForbiddenTenantEnvKeysFromGlobals_EmptyInputs(t *testing.T) {
+	if got := findForbiddenTenantEnvKeysFromGlobals(nil, nil); len(got) != 0 {
+		t.Errorf("nil/nil: got %v; want empty", got)
+	}
+	if got := findForbiddenTenantEnvKeysFromGlobals(map[string]string{}, map[string]struct{}{}); len(got) != 0 {
+		t.Errorf("empty/empty: got %v; want empty", got)
+	}
+	// Non-empty envVars but no global provenance — nothing came from
+	// global_secrets, so nothing to block (even if a workspace_secrets
+	// row exists for GITHUB_TOKEN).
+	if got := findForbiddenTenantEnvKeysFromGlobals(map[string]string{"GITHUB_TOKEN": "ghp_user"}, map[string]struct{}{}); len(got) != 0 {
+		t.Errorf("workspace-only GITHUB_TOKEN: got %v; want empty", got)
+	}
+}
+
 func TestFormatForbiddenTenantEnvError_Phrasing(t *testing.T) {
 	// Empty input — defensive total function.
 	if msg := formatForbiddenTenantEnvError(nil); !strings.Contains(msg, "RFC#523") {
@@ -120,38 +120,52 @@ func (h *WorkspaceHandler) prepareProvisionContext(
 	payload models.CreateWorkspacePayload,
 	resetClaudeSession bool,
 ) (*preparedProvisionContext, *provisionAbort) {
-	envVars, decryptErr := loadWorkspaceSecrets(ctx, workspaceID)
+	envVars, globalSecretKeys, decryptErr := loadWorkspaceSecrets(ctx, workspaceID)
 	if decryptErr != "" {
 		return nil, &provisionAbort{Msg: decryptErr}
 	}

-	// RFC#523 Layer 1 (task #146): refuse to start a tenant workspace
-	// when any forbidden operator-scope env var is present in the
-	// resolved secret-load env-set. Runs IMMEDIATELY after
-	// loadWorkspaceSecrets and BEFORE applyAgentGitHTTPCreds — the
-	// per-agent persona injection sets a fallback GITEA_USER/GITEA_TOKEN
-	// pair that the buildContainerEnv forensic #145 guard will strip
-	// later. We want THIS layer to catch leaks from the operator-
-	// controlled stores (global_secrets, workspace_secrets) only, not
-	// the deliberate per-agent platform injection that lives downstream.
+	// RFC#523 Layer 1 (issue molecule-ai/internal#523): refuse to start a
+	// tenant workspace when any forbidden operator-scope env var is
+	// present in the operator-controlled store (global_secrets).
 	//
-	// Threat model is "an upstream secret-writer accidentally widened
-	// the propagation set" — e.g. an operator pastes GITEA_TOKEN into
-	// a workspace_secrets row. Caught here, surfaced loudly to the
-	// canvas Events tab, fail-closed. The existing forensic #145 guard
-	// in provisioner.buildContainerEnv / CPProvisioner.Start stays as
-	// defense-in-depth: it silently strips at container-env-build time.
+	// PROVENANCE-AWARE — fix for the over-fire reported by CTO empirical
+	// 2026-05-20: the original implementation ran this check on the
+	// merged env-set, which conflated two very different sources:
+	//
+	//   1. global_secrets — operator-side store. ANY operator-scope token
+	//      here is an upstream bleed (e.g. tenant_secrets_seed.go pre-
+	//      4f45d37 propagating CP-env GITHUB_TOKEN into every fresh
+	//      tenant's row). RFC#523's literal threat model.
+	//
+	//   2. workspace_secrets — user-set via the canvas Secrets tab,
+	//      authenticated as the workspace owner. If the user pastes
+	//      their own scoped GitHub PAT under GITHUB_TOKEN so the agent
+	//      can push to their personal repos, that is the system working
+	//      as designed — not the leak RFC#523 was written to catch.
+	//
+	// The provenance side-channel from loadWorkspaceSecrets tells us
+	// which keys came from global_secrets (workspace_secrets writes
+	// override and clear the flag, since the user explicitly re-set
+	// the value). We restrict the abort to that set.
+	//
+	// Defense-in-depth NOT removed: provisioner.buildContainerEnv still
+	// runs the forensic #145 silent-strip (lower-confidence late layer),
+	// and workspace/entrypoint.sh has Layer 2 inside the container. If a
+	// real operator-scope token slips into workspace_secrets some other
+	// way, the later layers (and the per-workspace SG, and the per-tenant
+	// VPC isolation) are still in force.
 	//
 	// Key names (not values) are echoed in the user-facing error so
 	// the operator can locate and remove the offending row. Per memory
 	// `feedback_passwords_in_chat_are_burned`, key names are not
 	// secret; values would be.
-	if forbidden := findForbiddenTenantEnvKeys(envVars); len(forbidden) > 0 {
+	if forbidden := findForbiddenTenantEnvKeysFromGlobals(envVars, globalSecretKeys); len(forbidden) > 0 {
 		msg := formatForbiddenTenantEnvError(forbidden)
-		log.Printf("Provisioner: ABORT workspace=%s — forbidden operator-scope env keys present: %v (RFC#523)", workspaceID, forbidden)
+		log.Printf("Provisioner: ABORT workspace=%s — forbidden operator-scope env keys present in global_secrets: %v (RFC#523)", workspaceID, forbidden)
 		return nil, &provisionAbort{
 			Msg:   msg,
-			Extra: map[string]interface{}{"error": msg, "forbidden_env_keys": forbidden, "rfc": "523"},
+			Extra: map[string]interface{}{"error": msg, "forbidden_env_keys": forbidden, "rfc": "523", "source": "global_secrets"},
 		}
 	}

@@ -70,6 +70,97 @@ var restartDebounceWindow = 60 * time.Second
 // workspace-server yet — that's a separate RFC.
 var restartByIDDropCounter atomic.Uint64

+// fileWriteRestartDebounceWindow is the per-workspace coalescing window for
+// the file-write → RestartByID trigger fired by templates.go's WriteFile,
+// DeleteFile, and ReplaceFiles handlers (and template_import.go's variants).
+//
+// Background (internal#624 2026-05-20): canvas Save fires N PUT /files
+// requests in a 30-60s burst (claude-code SEO agent observed 10-17 files in
+// 60s). Each successful write previously fired `goAsync(RestartByID)`. The
+// 60s self-fire debounce in RestartByID itself catches calls 1-60s, but
+// writes at T+65s+ pass the debounce, set pending=true on a still-running
+// coalesceRestart cycle, and drain immediately into cycle 2 — which DELETEs
+// + recreates EC2 mid-burst, returning 500 EC2InstanceStateInvalidException
+// on the in-flight user PUTs.
+//
+// 15s is sized to absorb a canvas Save burst (writes typically land within
+// a 5-10s window) while still letting a deliberate "edit, wait, edit again"
+// pattern restart twice. Bigger than that would silently swallow legitimate
+// rapid-iteration edits; smaller would let burst tails leak through.
+var fileWriteRestartDebounceWindow = 15 * time.Second
+
+// fileWriteRestartLastFireAt records the last time `maybeRestartAfterFileWrite`
+// actually fired a restart for each workspace. sync.Map (not RWMutex+map)
+// because writes happen on every successful file-write handler, reads on
+// every subsequent file-write handler call — both per-workspace — and the
+// keys are sparse + long-lived. Stored as int64 unix-nano so the load/store
+// path can stay lock-free (atomic.Int64 inside sync.Map.Value is fine, but
+// time.Time itself isn't atomically loadable).
+var fileWriteRestartLastFireAt sync.Map // map[workspaceID]*atomic.Int64
+
+// fileWriteRestartDropCounter counts how many file-write restart triggers
+// were silently coalesced. Same observability rationale as
+// restartByIDDropCounter — package-level atomic so tests can assert the
+// drop fired and ops can correlate with "user clicked Save 10 times,
+// only saw 1 restart cycle".
+var fileWriteRestartDropCounter atomic.Uint64
+
+// maybeRestartAfterFileWrite is the call-site debounce wrapper for the 9
+// file-write trigger sites in templates.go + template_import.go. Replaces
+// the direct `goAsync(func() { wh.RestartByID(wsID) })` pattern with a
+// 15s per-workspace coalescing window:
+//
+//   - First call (no prior fire OR last fire >15s ago): records the
+//     current timestamp and fires goAsync(RestartByID).
+//   - Subsequent calls within 15s of the last fire: silently dropped,
+//     drop counter incremented.
+//
+// This is the call-site-layer protection (internal#624 Path A). The drain-
+// loop layer in coalesceRestart (Path B, re-stamping restartStartedAt per
+// iteration) is the platform-layer defense in depth — together they close
+// the file-write tight-loop class regardless of which entry point fires.
+//
+// Stateless on the handler so any handler with access to a WorkspaceHandler
+// can use it; the per-workspace state lives in the package-level sync.Map.
+func (h *WorkspaceHandler) maybeRestartAfterFileWrite(workspaceID string) {
+	now := time.Now().UnixNano()
+
+	// LoadOrStore the per-workspace last-fire stamp. First write for a
+	// brand-new workspace falls through the CompareAndSwap below because
+	// the zero-init value (0) is far enough in the past to satisfy the
+	// "last fire >15s ago" predicate.
+	sv, _ := fileWriteRestartLastFireAt.LoadOrStore(workspaceID, new(atomic.Int64))
+	stamp := sv.(*atomic.Int64)
+
+	// CAS loop: read last, decide, swap. We use CAS instead of Lock/Unlock
+	// because the typical case is "thousands of writes, one restart per
+	// 15s" — uncontended atomic is ~5ns vs ~30ns mutex. Bounded retry
+	// because in the rare contended case (two writes finishing nanoseconds
+	// apart) one will win the swap and the other will see the new stamp,
+	// drop, and bail.
+	for retry := 0; retry < 4; retry++ {
+		last := stamp.Load()
+		elapsed := time.Duration(now - last)
+		if last != 0 && elapsed < fileWriteRestartDebounceWindow {
+			// Within debounce window — drop silently.
+			fileWriteRestartDropCounter.Add(1)
+			log.Printf("maybeRestartAfterFileWrite: %s — coalesced "+
+				"(last fire %s ago < %s window; total dropped=%d)",
+				workspaceID, elapsed.Round(time.Millisecond),
+				fileWriteRestartDebounceWindow,
+				fileWriteRestartDropCounter.Load())
+			return
+		}
+		if stamp.CompareAndSwap(last, now) {
+			break
+		}
+		// Another writer beat us to the stamp update. Re-read and retry;
+		// the retry will almost certainly see the new value and drop.
+	}
+
+	h.goAsync(func() { h.RestartByID(workspaceID) })
+}
+
 // 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
@@ -513,6 +604,27 @@ func coalesceRestart(workspaceID string, cycle func()) {
 	// inside provisionWorkspace, so any writes that committed since the
 	// last cycle are picked up. Continues until no pending request was
 	// observed at the top of an iteration.
+	//
+	// internal#624 Path B (defense in depth for the file-write tight-loop
+	// class): re-stamp restartStartedAt at the top of every drain iteration
+	// past the first. The original design (stamp only on false→true edge)
+	// treated all drained pending as "one event from the debounce's POV",
+	// which is correct for the secrets-batch use case but lets a file-write
+	// burst at T+65s of a 60s drain pipe straight into another full cycle.
+	// Re-stamping closes that hole — each drained cycle gets its own fresh
+	// debounce window, so any RestartByID arriving during cycle N is
+	// dropped by shouldDebounceRestart instead of accumulating into
+	// pending=true for cycle N+1.
+	//
+	// The original "one cycle picks up everyone who arrived during it"
+	// semantic still holds for the secrets-write path: callers that hit
+	// coalesceRestart during cycle 1 still set pending=true and still get
+	// their effects landed in cycle 2. What changes is that callers
+	// arriving during cycle 2 (via RestartByID) now hit the re-stamped
+	// debounce and are dropped instead of being chained into cycle 3,
+	// which is exactly the chain that produced the 22:08-22:10 thrash on
+	// 3fe84b89.
+	iteration := 0
 	for {
 		state.mu.Lock()
 		if !state.pending {
@@ -520,7 +632,13 @@ func coalesceRestart(workspaceID string, cycle func()) {
 			return // defer clears running
 		}
 		state.pending = false
+		if iteration > 0 {
+			// Re-stamp for drained iterations only; the false→true edge
+			// already stamped at the top of coalesceRestart.
+			state.restartStartedAt = time.Now()
+		}
 		state.mu.Unlock()
+		iteration++

 		cycle()
 	}
@@ -0,0 +1,316 @@
+package handlers
+
+// Tests for internal#624 — file-write → RestartByID tight-loop fix.
+//
+// Empirical chain (Loki 2026-05-20 22:00-22:11Z on workspace
+// 3fe84b89-eb65-42fc-ad1f-5c93582ca3e7, claude-code SEO Agent):
+//
+//   1. Canvas Save writes 10-17 files in a 30-60s window.
+//   2. Each successful PUT /files at templates.go:575 / 591 / 607 / 662 /
+//      682 / 697 (and template_import.go:239 / 275 / 297) fires
+//      `goAsync(func() { wh.RestartByID(wsID) })`.
+//   3. RestartByID's existing 60s self-fire debounce catches calls 1-60s
+//      after the cycle starts. But writes at T+65s+ pass the debounce,
+//      set pending=true on the still-running coalesceRestart cycle, and
+//      drain IMMEDIATELY into cycle 2 — no re-debounce because the
+//      original drain loop re-uses the same restartStartedAt.
+//   4. Cycle 2 DELETEs+recreates EC2 mid-burst → user sees
+//      EC2InstanceStateInvalidException 500 on the in-flight PUTs.
+//
+// Fix: two layers (both shipped in the same PR).
+//
+//   Path A (call-site debounce): every file-write trigger goes through
+//   maybeRestartAfterFileWrite, which silently drops re-fires within 15s
+//   of the last fire for the same workspace.
+//
+//   Path B (drain-loop re-stamp): coalesceRestart now re-stamps
+//   restartStartedAt at the top of each drained iteration, so any
+//   RestartByID arriving during a drained cycle hits a fresh 60s window
+//   and is dropped by shouldDebounceRestart instead of chaining further.
+
+import (
+	"sync"
+	"sync/atomic"
+	"testing"
+	"time"
+)
+
+// resetFileWriteDebounceState wipes the package-level sync.Map + drop
+// counter for the given workspace ID. Tests must call this between
+// scenarios because fileWriteRestartLastFireAt is shared.
+func resetFileWriteDebounceState(workspaceID string) {
+	fileWriteRestartLastFireAt.Delete(workspaceID)
+	fileWriteRestartDropCounter.Store(0)
+}
+
+// newFileWriteDebounceHandler constructs a minimal *WorkspaceHandler with
+// no provisioner so RestartByID short-circuits at HasProvisioner()=false
+// — we only care that maybeRestartAfterFileWrite reaches goAsync at all.
+// The asyncWG inside goAsync lets us wait for the goroutine to finish so
+// we can deterministically observe whether RestartByID was scheduled.
+func newFileWriteDebounceHandler(t *testing.T) *WorkspaceHandler {
+	t.Helper()
+	return NewWorkspaceHandler(newTestBroadcaster(), nil, "http://localhost:8080", t.TempDir())
+}
+
+// TestMaybeRestartAfterFileWrite_FirstWriteRestarts — the baseline case:
+// the very first call for a workspace must actually fire goAsync (i.e.
+// no debounce-drop on the first PUT). Without this the helper would
+// silently swallow every legitimate single-file save.
+func TestMaybeRestartAfterFileWrite_FirstWriteRestarts(t *testing.T) {
+	const wsID = "fw-debounce-first"
+	resetFileWriteDebounceState(wsID)
+
+	h := newFileWriteDebounceHandler(t)
+	h.maybeRestartAfterFileWrite(wsID)
+
+	// Drop counter must NOT have incremented — the call fired.
+	if got := fileWriteRestartDropCounter.Load(); got != 0 {
+		t.Errorf("first call to maybeRestartAfterFileWrite must fire (drop counter must stay 0), got %d", got)
+	}
+
+	// Last-fire timestamp must be populated (non-zero) so the next call
+	// will compare against it.
+	sv, ok := fileWriteRestartLastFireAt.Load(wsID)
+	if !ok {
+		t.Fatal("first call must register the workspace in fileWriteRestartLastFireAt")
+	}
+	stamp := sv.(*atomic.Int64).Load()
+	if stamp == 0 {
+		t.Error("first call must record a non-zero last-fire timestamp")
+	}
+
+	// Wait for the spawned goroutine to finish so it doesn't leak into
+	// the next test (RestartByID will short-circuit on no-provisioner).
+	h.waitAsyncForTest()
+}
+
+// TestMaybeRestartAfterFileWrite_SecondWriteWithin15sSkipped — the core
+// fix: a second call within fileWriteRestartDebounceWindow of the first
+// MUST NOT fire RestartByID. The drop counter must increment by exactly
+// one and the last-fire timestamp must remain the FIRST call's stamp
+// (proof that the second call did not overwrite it).
+func TestMaybeRestartAfterFileWrite_SecondWriteWithin15sSkipped(t *testing.T) {
+	const wsID = "fw-debounce-second-within"
+	resetFileWriteDebounceState(wsID)
+
+	h := newFileWriteDebounceHandler(t)
+
+	// First call — fires.
+	h.maybeRestartAfterFileWrite(wsID)
+	h.waitAsyncForTest()
+
+	sv, _ := fileWriteRestartLastFireAt.Load(wsID)
+	firstStamp := sv.(*atomic.Int64).Load()
+
+	// Second call immediately — must be dropped.
+	h.maybeRestartAfterFileWrite(wsID)
+
+	if got := fileWriteRestartDropCounter.Load(); got != 1 {
+		t.Errorf("second call within 15s must increment drop counter by exactly 1, got %d", got)
+	}
+
+	// The CAS-loop must NOT have overwritten the first-call stamp — the
+	// debounce branch short-circuits before the CompareAndSwap.
+	stampAfter := sv.(*atomic.Int64).Load()
+	if stampAfter != firstStamp {
+		t.Errorf("dropped call must NOT update last-fire stamp (preserves debounce window); "+
+			"first=%d after=%d", firstStamp, stampAfter)
+	}
+}
+
+// TestMaybeRestartAfterFileWrite_ManyWritesInBurstCoalesceToOne — the
+// "bonus" regression test called out in the issue: 10 simulated PUTs
+// over 60s (compressed to a tight loop, all within 15s) must produce
+// exactly 1 RestartByID schedule and 9 drops. Models the canvas Save
+// burst shape that triggered the prod incident.
+func TestMaybeRestartAfterFileWrite_ManyWritesInBurstCoalesceToOne(t *testing.T) {
+	const wsID = "fw-debounce-burst"
+	resetFileWriteDebounceState(wsID)
+
+	h := newFileWriteDebounceHandler(t)
+
+	// 10 rapid-fire calls — simulates 10 PUTs landing inside the canvas
+	// Save burst window.
+	const burstSize = 10
+	for i := 0; i < burstSize; i++ {
+		h.maybeRestartAfterFileWrite(wsID)
+	}
+	h.waitAsyncForTest()
+
+	// One fired (call #1) + 9 dropped.
+	if got := fileWriteRestartDropCounter.Load(); got != burstSize-1 {
+		t.Errorf("expected %d drops for a %d-call burst (only call #1 fires), got %d",
+			burstSize-1, burstSize, got)
+	}
+}
+
+// TestMaybeRestartAfterFileWrite_AfterWindowExpiresFiresAgain — outside
+// the debounce window, the helper must release and fire again. Shrinks
+// fileWriteRestartDebounceWindow to 5ms so we don't sleep 15s in CI.
+// Important: without this, a legitimate "user edited, walked away for
+// a minute, edited again" would never restart and config changes would
+// never reach the agent.
+func TestMaybeRestartAfterFileWrite_AfterWindowExpiresFiresAgain(t *testing.T) {
+	const wsID = "fw-debounce-window-expires"
+	resetFileWriteDebounceState(wsID)
+
+	orig := fileWriteRestartDebounceWindow
+	fileWriteRestartDebounceWindow = 5 * time.Millisecond
+	defer func() { fileWriteRestartDebounceWindow = orig }()
+
+	h := newFileWriteDebounceHandler(t)
+
+	h.maybeRestartAfterFileWrite(wsID) // fires
+	h.waitAsyncForTest()
+
+	// Wait past the window.
+	time.Sleep(20 * time.Millisecond)
+
+	h.maybeRestartAfterFileWrite(wsID) // must fire again
+	h.waitAsyncForTest()
+
+	// Drop counter must still be 0 — both calls fired.
+	if got := fileWriteRestartDropCounter.Load(); got != 0 {
+		t.Errorf("second call after window expiry must fire (not drop), got %d drops", got)
+	}
+}
+
+// TestMaybeRestartAfterFileWrite_DifferentWorkspacesIndependent — the
+// per-workspace state map must isolate: a burst on workspace A must not
+// affect workspace B's debounce. Pinning so a future "use a single
+// global atomic" refactor breaks loudly.
+func TestMaybeRestartAfterFileWrite_DifferentWorkspacesIndependent(t *testing.T) {
+	const wsA = "fw-debounce-ws-a"
+	const wsB = "fw-debounce-ws-b"
+	resetFileWriteDebounceState(wsA)
+	resetFileWriteDebounceState(wsB)
+
+	h := newFileWriteDebounceHandler(t)
+
+	// 5 calls on A, all but one drop.
+	for i := 0; i < 5; i++ {
+		h.maybeRestartAfterFileWrite(wsA)
+	}
+	h.waitAsyncForTest()
+
+	dropsAfterA := fileWriteRestartDropCounter.Load()
+
+	// First call on B — must fire (its own independent window).
+	h.maybeRestartAfterFileWrite(wsB)
+	h.waitAsyncForTest()
+
+	// B's call must not have incremented the drop counter — it fired.
+	if got := fileWriteRestartDropCounter.Load(); got != dropsAfterA {
+		t.Errorf("workspace B's first call must fire (not share workspace A's debounce); "+
+			"drops after A=%d, drops after B=%d", dropsAfterA, got)
+	}
+
+	// Both workspaces must have their own last-fire entries.
+	if _, ok := fileWriteRestartLastFireAt.Load(wsA); !ok {
+		t.Error("workspace A missing from fileWriteRestartLastFireAt")
+	}
+	if _, ok := fileWriteRestartLastFireAt.Load(wsB); !ok {
+		t.Error("workspace B missing from fileWriteRestartLastFireAt")
+	}
+}
+
+// TestMaybeRestartAfterFileWrite_ConcurrentCallsSafelyDebounced — the
+// CAS-loop contract: many goroutines hitting the helper concurrently
+// must still produce at most one fired call (drops = N-1). Pinning the
+// "thousands of writes, one restart" performance shape called out in
+// the helper's comment. Uses sync.WaitGroup to release all goroutines
+// in a tight burst so the CAS is genuinely contended.
+func TestMaybeRestartAfterFileWrite_ConcurrentCallsSafelyDebounced(t *testing.T) {
+	const wsID = "fw-debounce-concurrent"
+	resetFileWriteDebounceState(wsID)
+
+	h := newFileWriteDebounceHandler(t)
+
+	const goroutines = 50
+	start := make(chan struct{})
+	var wg sync.WaitGroup
+	for i := 0; i < goroutines; i++ {
+		wg.Add(1)
+		go func() {
+			defer wg.Done()
+			<-start // hold every goroutine at the gate
+			h.maybeRestartAfterFileWrite(wsID)
+		}()
+	}
+	close(start) // release the herd
+	wg.Wait()
+	h.waitAsyncForTest()
+
+	// Exactly N-1 drops: one goroutine wins the CAS and fires, all
+	// other N-1 see a fresh stamp and drop into the debounce branch.
+	if got := fileWriteRestartDropCounter.Load(); got != goroutines-1 {
+		t.Errorf("expected %d drops for %d concurrent callers (exactly one fires), got %d",
+			goroutines-1, goroutines, got)
+	}
+}
+
+// TestCoalesceRestart_DrainRespectsRestartedAtBetweenIterations —
+// Path B regression: when coalesceRestart drains a pending request into
+// a follow-up cycle, the restartStartedAt timestamp must be re-stamped
+// for that follow-up iteration. Without this, a RestartByID arriving
+// during cycle 2 would hit a stale 60s window (computed from cycle 1's
+// start) and could pass the debounce just because cycle 1 + cycle 2's
+// runtime exceeded 60s combined.
+//
+// The test fires cycle 1 → completes → sets pending=true to trigger
+// cycle 2 → asserts that restartStartedAt was advanced for the drained
+// iteration. The cycle function itself just records the wall-clock at
+// which it observed restartStartedAt, so the test can compare cycle 1's
+// stamp vs cycle 2's stamp.
+func TestCoalesceRestart_DrainRespectsRestartedAtBetweenIterations(t *testing.T) {
+	const wsID = "fw-debounce-drain-restamp"
+	resetRestartStatesFor(wsID)
+
+	// Capture the restartStartedAt observed at the top of each cycle
+	// iteration. The cycle reads it directly from the state map so we
+	// see what coalesceRestart wrote.
+	var stamps []time.Time
+	var stampsMu sync.Mutex
+	cycleCount := 0
+	cycle := func() {
+		sv, _ := restartStates.Load(wsID)
+		state := sv.(*restartState)
+		state.mu.Lock()
+		stampsMu.Lock()
+		stamps = append(stamps, state.restartStartedAt)
+		stampsMu.Unlock()
+		state.mu.Unlock()
+
+		cycleCount++
+		if cycleCount == 1 {
+			// While inside cycle 1, set pending=true so the drain loop
+			// runs cycle 2 next iteration. Mirrors the prod shape: a
+			// PUT lands during cycle 1, sets pending=true via
+			// RestartByID → coalesceRestart's pending branch.
+			state.mu.Lock()
+			state.pending = true
+			state.mu.Unlock()
+
+			// Sleep briefly so cycle 2's stamp is observably later
+			// than cycle 1's. Without a real wall-clock gap the
+			// assertion can't tell re-stamp from no-op.
+			time.Sleep(20 * time.Millisecond)
+		}
+	}
+
+	coalesceRestart(wsID, cycle)
+
+	stampsMu.Lock()
+	defer stampsMu.Unlock()
+	if len(stamps) != 2 {
+		t.Fatalf("expected 2 cycle iterations (original + drained pending), got %d", len(stamps))
+	}
+	if !stamps[1].After(stamps[0]) {
+		t.Errorf("Path B regression: cycle 2's restartStartedAt (%v) must be AFTER "+
+			"cycle 1's (%v) — drained iterations must re-stamp so the self-fire "+
+			"debounce window resets per cycle. Without this, a RestartByID arriving "+
+			"during cycle 2 sees a stale window and can chain into cycle 3.",
+			stamps[1], stamps[0])
+	}
+}
@@ -397,6 +397,8 @@ func Setup(hub *ws.Hub, broadcaster *events.Broadcaster, prov *provisioner.Provi
 		wsAuth.GET("/tokens", tokh.List)
 		wsAuth.POST("/tokens", tokh.Create)
 		wsAuth.DELETE("/tokens/:tokenId", tokh.Revoke)
+		adminTokH := handlers.NewAdminWorkspaceTokenHandler()
+		r.POST("/admin/workspaces/:id/tokens", middleware.AdminAuth(db.DB), adminTokH.Create)

 		// Memory
 		memh := handlers.NewMemoryHandler()