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

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.

.gitea/scripts/main-red-watchdog.py

@@ -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}: "

.gitea/scripts/status-reaper.py

@@ -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]

.gitea/workflows/e2e-peer-visibility.yml

@@ -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).

.gitea/workflows/publish-workspace-server-image.yml

@@ -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.

CONTRIBUTING.md

@@ -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

docs/guides/external-agent-registration.md

@@ -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

docs/guides/external-workspace-quickstart.md

@@ -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>`

docs/guides/remote-workspaces.md

@@ -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 |

tests/e2e/_lib.sh

@@ -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

tests/e2e/lib/peer_visibility_assert.sh

@@ -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)

tests/e2e/test_peer_visibility_mcp_local.sh

@@ -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

tests/e2e/test_peer_visibility_mcp_staging.sh

@@ -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
@@ -218,25 +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)"
+  TOKEN_DIAG=""
   if [ -z "$WTOK" ]; then
-    TTOK_RESP=$(tenant_call GET "/admin/workspaces/$WID/test-token" 2>/dev/null || true)
-    WTOK=$(echo "$TTOK_RESP" | python3 -c "
-import sys, json
-try: d = json.load(sys.stdin)
-except Exception: print(''); sys.exit(0)
-print(d.get('auth_token') or '')
-" 2>/dev/null)
+    TTOK_FILE=$(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 "$TTOK_FILE"
   fi
-  [ -n "$WTOK" ] || fail "$rt workspace did not return or mint an auth_token — cannot drive its MCP call (resp: $(echo "$R" | head -c 300))"
+  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"

tests/test_main_red_watchdog.py

@@ -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"
+    )

tests/test_status_reaper.py

@@ -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(

workspace-server/internal/handlers/a2a_proxy.go

@@ -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 {

workspace-server/internal/handlers/a2a_proxy_test.go

@@ -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) {

workspace-server/internal/handlers/admin_workspace_tokens.go

@@ -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.",
+	})
+}

workspace-server/internal/handlers/admin_workspace_tokens_test.go

@@ -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())
+	}
+}

workspace-server/internal/handlers/mcp.go

@@ -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)
 // ─────────────────────────────────────────────────────────────────────────────

workspace-server/internal/handlers/mcp_test.go

@@ -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
 // ─────────────────────────────────────────────────────────────────────────────

workspace-server/internal/handlers/mcp_tools.go

@@ -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)
 }
-

workspace-server/internal/handlers/registry.go

@@ -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.

workspace-server/internal/router/router.go

@@ -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()