fix(2047): pass workspaceID to stripPluginMarkersFromMemory

Plugin uninstall had two sites where execAsRoot errors were discarded:
- Skill directory removal (plugins_install.go:125) — orphaned skill dirs
  if rm -rf failed silently
- CLAUDE.md marker stripping (plugins_install_pipeline.go:326) — stale
  plugin content left in CLAUDE.md if awk script failed

Both now log the error without failing the overall uninstall (best-effort
 cleanup), giving operators visibility into incomplete uninstalls.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

.gitea/scripts/ci-required-drift.py

@@ -8,7 +8,8 @@ pair diverges.
 Sources:
   A. `.gitea/workflows/ci.yml` jobs  (CI source — the actual job set)
   B. `status_check_contexts` in branch_protections (the merge gate)
-  C. `REQUIRED_CHECKS` env in audit-force-merge.yml (the audit env)
+  C. `REQUIRED_CHECKS_JSON` (preferred) or `REQUIRED_CHECKS` (legacy)
+     env in audit-force-merge.yml (the audit env)
 
 Three failure classes:
   F1  Job in (A) is not under the sentinel's `needs:` — sentinel
@@ -250,13 +251,21 @@ def sentinel_needs(ci_doc: dict) -> set[str]:
     return set(needs)
 
 
-def required_checks_env(audit_doc: dict) -> set[str]:
-    """Pull the REQUIRED_CHECKS env value from audit-force-merge.yml.
+def required_checks_env(audit_doc: dict, branch: str) -> set[str]:
+    """Pull the required-checks env value from audit-force-merge.yml.
+
     Walks the YAML AST per `feedback_behavior_based_ast_gates`: we do
-    NOT grep for `REQUIRED_CHECKS:` — that breaks under reformatting,
+    NOT grep for env keys — that breaks under reformatting,
     multi-job workflows, or a future move of the env to a different
-    step. Instead, look inside every job's every step's `env:` map."""
-    found: list[str] = []
+    step. Instead, look inside every job's every step's `env:` map.
+
+    Supports two variants:
+      - REQUIRED_CHECKS_JSON (preferred): JSON dict keyed by branch name.
+        We extract the array for the target branch.
+      - REQUIRED_CHECKS (legacy): newline-separated list of context names.
+    """
+    found_json: list[str] = []
+    found_legacy: list[str] = []
     jobs = audit_doc.get("jobs", {})
     if not isinstance(jobs, dict):
         sys.stderr.write(f"::warning::{AUDIT_WORKFLOW_PATH} has no jobs: mapping\n")
@@ -268,27 +277,67 @@ def required_checks_env(audit_doc: dict) -> set[str]:
             if not isinstance(step, dict):
                 continue
             step_env = step.get("env") or {}
-            if isinstance(step_env, dict) and "REQUIRED_CHECKS" in step_env:
-                v = step_env["REQUIRED_CHECKS"]
-                if isinstance(v, str):
-                    found.append(v)
-    if not found:
-        sys.stderr.write(
-            f"::error::REQUIRED_CHECKS env not found in any step of "
-            f"{AUDIT_WORKFLOW_PATH}\n"
-        )
-        sys.exit(3)
-    if len(found) > 1:
-        # Defensive: refuse to guess which one is canonical.
-        sys.stderr.write(
-            f"::error::REQUIRED_CHECKS env present in {len(found)} steps; ambiguous\n"
-        )
-        sys.exit(3)
-    raw = found[0]
-    # YAML block-scalars (`|`) leave a trailing newline + blanks; trim
-    # consistently with audit-force-merge.sh's parser so both sides
-    # produce identical sets.
-    return {line.strip() for line in raw.splitlines() if line.strip()}
+            if isinstance(step_env, dict):
+                if "REQUIRED_CHECKS_JSON" in step_env:
+                    v = step_env["REQUIRED_CHECKS_JSON"]
+                    if isinstance(v, str):
+                        found_json.append(v)
+                if "REQUIRED_CHECKS" in step_env:
+                    v = step_env["REQUIRED_CHECKS"]
+                    if isinstance(v, str):
+                        found_legacy.append(v)
+
+    # JSON variant takes precedence.
+    if found_json:
+        if len(found_json) > 1:
+            sys.stderr.write(
+                f"::error::REQUIRED_CHECKS_JSON env present in {len(found_json)} steps; ambiguous\n"
+            )
+            sys.exit(3)
+        try:
+            parsed = json.loads(found_json[0])
+        except json.JSONDecodeError as e:
+            sys.stderr.write(
+                f"::error::REQUIRED_CHECKS_JSON is not valid JSON: {e}\n"
+            )
+            sys.exit(3)
+        if not isinstance(parsed, dict):
+            sys.stderr.write(
+                f"::error::REQUIRED_CHECKS_JSON parsed to {type(parsed).__name__}, expected dict\n"
+            )
+            sys.exit(3)
+        branch_checks = parsed.get(branch)
+        if branch_checks is None:
+            sys.stderr.write(
+                f"::error::REQUIRED_CHECKS_JSON has no entry for branch '{branch}'\n"
+            )
+            sys.exit(3)
+        if not isinstance(branch_checks, list):
+            sys.stderr.write(
+                f"::error::REQUIRED_CHECKS_JSON['{branch}'] is {type(branch_checks).__name__}, expected list\n"
+            )
+            sys.exit(3)
+        return {str(item).strip() for item in branch_checks if str(item).strip()}
+
+    # Legacy variant fallback.
+    if found_legacy:
+        if len(found_legacy) > 1:
+            # Defensive: refuse to guess which one is canonical.
+            sys.stderr.write(
+                f"::error::REQUIRED_CHECKS env present in {len(found_legacy)} steps; ambiguous\n"
+            )
+            sys.exit(3)
+        raw = found_legacy[0]
+        # YAML block-scalars (`|`) leave a trailing newline + blanks; trim
+        # consistently with audit-force-merge.sh's parser so both sides
+        # produce identical sets.
+        return {line.strip() for line in raw.splitlines() if line.strip()}
+
+    sys.stderr.write(
+        f"::error::Neither REQUIRED_CHECKS_JSON nor REQUIRED_CHECKS env found in any step of "
+        f"{AUDIT_WORKFLOW_PATH}\n"
+    )
+    sys.exit(3)
 
 
 # --------------------------------------------------------------------------
@@ -330,7 +379,7 @@ def detect_drift(branch: str) -> tuple[list[str], dict]:
     jobs = ci_job_names(ci_doc)
     jobs_all = ci_jobs_all(ci_doc)
     needs = sentinel_needs(ci_doc)
-    env_set = required_checks_env(audit_doc)
+    env_set = required_checks_env(audit_doc, branch)
 
     # Protection
     # api() raises ApiError on non-2xx. Transient 5xx should fail loud.
@@ -524,7 +573,7 @@ def render_body(branch: str, findings: list[str], debug: dict) -> str:
             "- **F2**: rename the protection context to match an emitter, "
             "or remove it from `status_check_contexts` "
             "(PATCH `/api/v1/repos/{owner}/{repo}/branch_protections/{branch}`).",
-            "- **F3a / F3b**: bring `REQUIRED_CHECKS` env in "
+            "- **F3a / F3b**: bring `REQUIRED_CHECKS_JSON` (or `REQUIRED_CHECKS` legacy) env in "
             "`.gitea/workflows/audit-force-merge.yml` into set-equality with "
             "`status_check_contexts` (single PR, both files).",
             "",

.gitea/scripts/detect-changes.py

@@ -26,6 +26,10 @@ PROFILES: dict[str, dict[str, str]] = {
         "handlers": (
             r"^workspace-server/internal/handlers/"
             r"|^workspace-server/internal/wsauth/"
+            # #2149: the scheduler real-PG integration tests run in this same
+            # workflow (they reuse its migrated Postgres), so changes to the
+            # scheduler package must trigger the job too.
+            r"|^workspace-server/internal/scheduler/"
             r"|^workspace-server/migrations/"
             r"|^\.gitea/workflows/handlers-postgres-integration\.yml$"
         ),
@@ -174,3 +178,4 @@ def main(argv: list[str]) -> int:
 
 if __name__ == "__main__":
     sys.exit(main(sys.argv[1:]))
+

.gitea/scripts/sop-checklist.py

@@ -1228,10 +1228,13 @@ def main(argv: list[str] | None = None) -> int:
                 )
 
         na_desc = ", ".join(sorted(na_descs)) if na_descs else "(none)"
-        na_status_state = "success" if na_descs else "pending"
+        # internal#818: na-declarations is an informational context, not a merge
+        # gate. An empty declaration list is a terminal success state — pending
+        # here poisons the PR combined status.
+        na_status_state = "success"
         # review-check.sh reads the description to discover which gates are N/A.
         # Include the gate names so it can grep for them.
-        na_description = f"N/A: {na_desc}" if na_descs else "N/A: (none)"
+        na_description = f"N/A: {na_desc}"
 
         if not args.dry_run:
             client.post_status(

.gitea/scripts/sop-tier-check.sh

@@ -114,6 +114,19 @@ if [ -z "$WHOAMI" ]; then
 fi
 echo "::notice::token resolves to user: $WHOAMI"
 
+# 0.5 Read PR head SHA so we can reject stale approvals after head moves
+# (internal#816). Reviews carry the commit_id they were submitted against.
+HEAD_SHA=$(curl -sS -H "$AUTH" "${API}/repos/${OWNER}/${NAME}/pulls/${PR_NUMBER}" | jq -r '.head.sha // ""') || true
+if [ -z "$HEAD_SHA" ]; then
+  echo "::error::Failed to fetch PR head SHA — token may be invalid."
+  if [ "${SOP_FAIL_OPEN:-}" = "1" ]; then
+    echo "::warning::SOP_FAIL_OPEN=1 — exiting 0 so CI does not block."
+    exit 0
+  fi
+  exit 1
+fi
+debug "pr-head-sha=$HEAD_SHA"
+
 # 1. Read tier label. || true ensures set -euo pipefail does not abort the
 # script if curl or jq fails (e.g. 401 from empty token).
 LABELS=$(curl -sS -H "$AUTH" "${API}/repos/${OWNER}/${NAME}/issues/${PR_NUMBER}/labels" | jq -r '.[].name') || true
@@ -265,7 +278,7 @@ if [ $_REVIEWS_EXIT -ne 0 ] || [ -z "$REVIEWS" ]; then
   fi
   exit 1
 fi
-APPROVERS=$(echo "$REVIEWS" | jq -r '[.[] | select(.state=="APPROVED") | .user.login] | unique | .[]') || true
+APPROVERS=$(echo "$REVIEWS" | jq -r --arg head_sha "$HEAD_SHA" '[.[] | select(.state=="APPROVED" and .commit_id == $head_sha) | .user.login] | unique | .[]') || true
 if [ -z "$APPROVERS" ]; then
   echo "::error::No approving reviews on this PR. Set SOP_DEBUG=1 and re-run for diagnostics."
   exit 1

.gitea/scripts/tests/test_ci_required_drift.py

@@ -1,4 +1,5 @@
 import importlib.util
+import json
 import sys
 from pathlib import Path
 from unittest.mock import patch
@@ -36,6 +37,76 @@ def _make_audit_doc(required_checks: list[str]) -> dict:
     }
 
 
+def _make_audit_doc_json(required_checks_json: dict) -> dict:
+    return {
+        "jobs": {
+            "audit": {
+                "steps": [
+                    {"env": {"REQUIRED_CHECKS_JSON": json.dumps(required_checks_json)}}
+                ]
+            }
+        }
+    }
+
+
+# ---------------------------------------------------------------------------
+# required_checks_env — dual-variant parsing
+# ---------------------------------------------------------------------------
+
+def test_required_checks_env_prefers_json_over_legacy():
+    doc = {
+        "jobs": {
+            "audit": {
+                "steps": [
+                    {
+                        "env": {
+                            "REQUIRED_CHECKS_JSON": json.dumps(
+                                {"main": ["ctx-a"], "staging": ["ctx-b"]}
+                            ),
+                            "REQUIRED_CHECKS": "ctx-legacy\nctx-old",
+                        }
+                    }
+                ]
+            }
+        }
+    }
+    assert drift.required_checks_env(doc, "main") == {"ctx-a"}
+    assert drift.required_checks_env(doc, "staging") == {"ctx-b"}
+
+
+def test_required_checks_env_falls_back_to_legacy():
+    doc = _make_audit_doc(["legacy-ctx"])
+    assert drift.required_checks_env(doc, "main") == {"legacy-ctx"}
+
+
+def test_required_checks_env_json_missing_branch_fails():
+    doc = _make_audit_doc_json({"staging": ["ctx-b"]})
+    try:
+        drift.required_checks_env(doc, "main")
+    except SystemExit as exc:
+        assert exc.code == 3
+    else:
+        raise AssertionError("expected SystemExit(3)")
+
+
+def test_required_checks_env_json_malformed_fails():
+    doc = {
+        "jobs": {
+            "audit": {
+                "steps": [
+                    {"env": {"REQUIRED_CHECKS_JSON": "not-json"}}
+                ]
+            }
+        }
+    }
+    try:
+        drift.required_checks_env(doc, "main")
+    except SystemExit as exc:
+        assert exc.code == 3
+    else:
+        raise AssertionError("expected SystemExit(3)")
+
+
 # ---------------------------------------------------------------------------
 # sentinel_needs
 # ---------------------------------------------------------------------------

.gitea/scripts/tests/test_prod_auto_deploy.py

@@ -571,3 +571,44 @@ def test_current_branch_head_none_without_token():
 def test_current_branch_head_none_on_non_200(monkeypatch):
     monkeypatch.setattr(prod, "_api_json_optional", lambda _u, _t: (500, None))
     assert prod.current_branch_head({"GITEA_TOKEN": "secret"}) is None
+
+
+# --- #2213: superseded check must fire BEFORE production side effects ----------
+#
+# Real incident shape: two main pushes land ~2 min apart. The OLDER deploy job
+# (GITHUB_SHA=7a72516, target staging-7a72516) started LATE — main head was
+# already 7f25373. The #2194 guard only protected the *verify* step, so the
+# older job still:
+#   1. rolled the canary (hongming) BACKWARD to staging-7a72516 (the #2213 red,
+#      seen as the newer job's verify reading hongming on the old SHA), then
+#   2. promoted :latest backward to the older image,
+# before finally skipping verify. The workflow now calls this same superseded
+# check BEFORE the redeploy + promote steps and gates both off when it fires.
+# These tests pin the contract that check-superseded relies on for the exact
+# incident shape.
+
+
+def test_superseded_by_fires_for_older_job_when_newer_already_head(monkeypatch):
+    # Older job (7a72516) re-checks the head just before rollout and finds the
+    # newer merge (7f25373) already owns main -> superseded -> skip side effects.
+    monkeypatch.setattr(
+        prod, "current_branch_head", lambda _env: "7f25373309eca54a36f08c371ff783c3a47c3f8d"
+    )
+    newer = prod.superseded_by(
+        {"GITHUB_SHA": "7a72516f7e7ba1a710c4f393fef08be8d22e1866"}
+    )
+    assert newer == "7f25373309eca54a36f08c371ff783c3a47c3f8d"
+
+
+def test_superseded_by_none_for_latest_job_so_it_still_rolls(monkeypatch):
+    # The newer job (7f25373) IS the head -> NOT superseded -> it proceeds to
+    # roll the fleet and verify, so a genuinely-behind tenant still fails loud.
+    monkeypatch.setattr(
+        prod, "current_branch_head", lambda _env: "7f25373309eca54a36f08c371ff783c3a47c3f8d"
+    )
+    assert (
+        prod.superseded_by(
+            {"GITHUB_SHA": "7f25373309eca54a36f08c371ff783c3a47c3f8d"}
+        )
+        is None
+    )

.gitea/scripts/tests/test_sop_checklist.py

@@ -1299,3 +1299,108 @@ class TestGetCIStatus(unittest.TestCase):
         self.assertEqual(
             sop.get_ci_status(client, "o", "r", "sha1"), "unknown"
         )
+
+
+# ---------------------------------------------------------------------------
+# internal#818 — na-declarations status must be terminal success
+# ---------------------------------------------------------------------------
+
+
+class TestNaDeclarationsStatusTerminal(unittest.TestCase):
+    """Regression for internal#818: the na-declarations context is
+    informational, not a merge gate.  An empty N/A declaration list must
+    post `success` (not `pending`) so it does not poison the PR combined
+    status."""
+
+    def _run_with_fake_client(self, fake_client_class):
+        """Swap GiteaClient temporarily and invoke main() with a fake token."""
+        orig_client = sop.GiteaClient
+        orig_token = os.environ.get("GITEA_TOKEN")
+        try:
+            sop.GiteaClient = fake_client_class
+            os.environ["GITEA_TOKEN"] = "fake-token"
+            return sop.main([
+                "--owner", "o", "--repo", "r", "--pr", "1",
+                "--config", CONFIG_PATH,
+                "--gitea-host", "git.example.com",
+            ])
+        finally:
+            sop.GiteaClient = orig_client
+            if orig_token is None:
+                os.environ.pop("GITEA_TOKEN", None)
+            else:
+                os.environ["GITEA_TOKEN"] = orig_token
+
+    def test_empty_na_descriptions_posts_success(self):
+        posted = []
+
+        class FakeClient(sop.GiteaClient):
+            def get_pr(self, owner, repo, pr):
+                return {
+                    "state": "open",
+                    "user": {"login": "alice"},
+                    "head": {"sha": "abc123"},
+                    "labels": [],
+                }
+
+            def get_issue_comments(self, owner, repo, issue, max_comments=None):
+                return []
+
+            def resolve_team_id(self, org, team_name):
+                return None
+
+            def is_team_member(self, team_id, login):
+                return False
+
+            def post_status(self, owner, repo, sha, state, context,
+                            description, target_url=""):
+                posted.append({
+                    "state": state,
+                    "context": context,
+                    "description": description,
+                })
+
+        rc = self._run_with_fake_client(FakeClient)
+        self.assertEqual(rc, 0)
+        na_posts = [p for p in posted if "na-declarations" in p["context"]]
+        self.assertEqual(len(na_posts), 1, f"expected one na-declarations post, got {posted}")
+        self.assertEqual(na_posts[0]["state"], "success")
+        self.assertEqual(na_posts[0]["description"], "N/A: (none)")
+
+    def test_populated_na_descriptions_posts_success(self):
+        posted = []
+
+        class FakeClient(sop.GiteaClient):
+            def get_pr(self, owner, repo, pr):
+                return {
+                    "state": "open",
+                    "user": {"login": "alice"},
+                    "head": {"sha": "abc123"},
+                    "labels": [],
+                }
+
+            def get_issue_comments(self, owner, repo, issue, max_comments=None):
+                return [
+                    {"user": {"login": "bob"}, "body": "/sop-n/a qa-review N/A: docs-only"},
+                ]
+
+            def resolve_team_id(self, org, team_name):
+                return 1
+
+            def is_team_member(self, team_id, login):
+                return True
+
+            def post_status(self, owner, repo, sha, state, context,
+                            description, target_url=""):
+                posted.append({
+                    "state": state,
+                    "context": context,
+                    "description": description,
+                })
+
+        rc = self._run_with_fake_client(FakeClient)
+        self.assertEqual(rc, 0)
+        na_posts = [p for p in posted if "na-declarations" in p["context"]]
+        self.assertEqual(len(na_posts), 1)
+        self.assertEqual(na_posts[0]["state"], "success")
+        self.assertIn("qa-review", na_posts[0]["description"])

.gitea/scripts/tests/test_sop_tier_check_stale_reviews.sh

@@ -0,0 +1,66 @@
+#!/usr/bin/env bash
+# Regression test for internal#816 — sop-tier-check must ignore APPROVED
+# reviews that were submitted against an old PR head SHA.
+#
+# Bug: the script collected approvers with
+#   jq '[.[] | select(.state=="APPROVED") | .user.login]'
+# without filtering on .commit_id == HEAD_SHA. After a PR head moved,
+# stale approvals looked valid to the tier gate.
+#
+# Fix: the jq filter now includes
+#   select(.state=="APPROVED" and .commit_id == $head_sha)
+# where $head_sha is the current PR head fetched from the API.
+
+set -euo pipefail
+
+# jq may not be on PATH in all environments (e.g. dev containers).
+PATH="/tmp/bin:$PATH"
+command -v jq >/dev/null 2>&1 || { echo "::error::jq required but not found"; exit 1; }
+
+PASS=0
+FAIL=0
+
+assert_eq() {
+  local label="$1"
+  local expected="$2"
+  local got="$3"
+  if [ "$expected" = "$got" ]; then
+    echo "  PASS  $label"
+    PASS=$((PASS + 1))
+  else
+    echo "  FAIL  $label"
+    echo "        expected: <$expected>"
+    echo "        got:      <$got>"
+    FAIL=$((FAIL + 1))
+  fi
+}
+
+# Sample reviews matching the shape from Gitea API
+REVIEWS_JSON='[
+  {"state":"APPROVED","commit_id":"abc123","user":{"login":"bob"}},
+  {"state":"APPROVED","commit_id":"old456","user":{"login":"alice"}},
+  {"state":"COMMENT","commit_id":"abc123","user":{"login":"carol"}},
+  {"state":"APPROVED","commit_id":"abc123","user":{"login":"dave"}},
+  {"state":"REQUEST_CHANGES","commit_id":"abc123","user":{"login":"eve"}}
+]'
+
+echo "test: jq filter keeps only APPROVED on current head"
+GOT=$(echo "$REVIEWS_JSON" | jq -r --arg head_sha "abc123" \
+  '[.[] | select(.state=="APPROVED" and .commit_id == $head_sha) | .user.login] | unique | .[]')
+assert_eq "current-head approvers" "bob dave" "$(echo "$GOT" | tr '\n' ' ' | sed 's/ $//')"
+
+echo "test: jq filter with all-stale reviews yields empty"
+GOT=$(echo "$REVIEWS_JSON" | jq -r --arg head_sha "new789" \
+  '[.[] | select(.state=="APPROVED" and .commit_id == $head_sha) | .user.login] | unique | .[]')
+assert_eq "all-stale yields empty" "" "$GOT"
+
+echo "test: jq filter handles null commit_id gracefully"
+NULL_JSON='[{"state":"APPROVED","commit_id":null,"user":{"login":"mallory"}}]'
+GOT=$(echo "$NULL_JSON" | jq -r --arg head_sha "abc123" \
+  '[.[] | select(.state=="APPROVED" and .commit_id == $head_sha) | .user.login] | unique | .[]')
+assert_eq "null commit_id excluded" "" "$GOT"
+
+echo
+echo "------"
+echo "PASS=$PASS FAIL=$FAIL"
+[ "$FAIL" -eq 0 ]

.gitea/workflows/ci-arm64-advisory.yml

@@ -96,6 +96,7 @@ env:
   GITHUB_SERVER_URL: https://git.moleculesai.app
 
 jobs:
+  # bp-exempt: advisory arm64 pilot, non-gating by design (internal#418).
   fast-checks:
     name: fast-checks
     # AND-set: only the Mac arm64 runner advertises macos-self-hosted.

.gitea/workflows/ci.yml

@@ -25,10 +25,9 @@
 #      sufficient for `actions/checkout` against this same repo.
 #
 #   4. Docs — no docs/scripts reference github.com URLs that need swapping.
-#      The canvas-deploy-reminder step writes a `ghcr.io/...` image
-#      reference into the step summary text — that's documentation prose
-#      pointing at the ECR-mirrored canvas image and stays unchanged for
-#      this port (a separate cleanup if ghcr→ECR sweep is in scope).
+#      The canvas-deploy-status step (core#2226, formerly canvas-deploy-reminder)
+#      writes the canvas ordered-deploy status into the step summary; it points
+#      at the ECR canvas image and the publish workflow, no ghcr.io prose.
 #
 # Cross-links:
 #   - RFC: internal#219 (CI/CD hard-gate hardening)
@@ -389,61 +388,61 @@ jobs:
 
   # mc#959 root-fix (sre)
 
-  canvas-deploy-reminder:
-    name: Canvas Deploy Reminder
+  canvas-deploy-status:
+    # core#2226: replaces the old advisory "Canvas Deploy Reminder". The canvas
+    # image now has a real ORDERED auto-deploy (publish-canvas-image.yml:
+    # build → push :staging-<sha> → wait green main CI → promote :latest by
+    # digest), and docker-compose pins via CANVAS_IMAGE_TAG. There is no longer
+    # a manual "go run docker compose pull by hand" step to remind operators
+    # about — so this job just records, on a canvas-touching main push, that the
+    # ordered deploy is handling it (and where to watch), instead of prescribing
+    # a manual action that determinism made obsolete.
+    name: Canvas Deploy Status
     runs-on: docker-host
-    # mc#1982 root-fix: added job-level `if:` so ci-required-drift.py's
-    # ci_job_names() detects this as github.ref-gated and skips it from F1.
-    # The step-level exit 0 handles the "not main push" case; the job-level
-    # `if:` makes the gating explicit so the drift script sees it.
-    # Runs on both main and staging pushes; step exits 0 when not applicable.
+    # Job-level `if:` so ci-required-drift.py's ci_job_names() detects this as
+    # github.ref-gated and skips it from the required-context F1 set (mc#1982).
+    # Step-level exit 0 handles the "not a canvas main push" case.
     if: ${{ github.ref == 'refs/heads/main' || github.ref == 'refs/heads/staging' }}
     needs: [changes, canvas-build]
     steps:
-      - name: Write deploy reminder to step summary
+      - name: Record canvas ordered-deploy status
         env:
           COMMIT_SHA: ${{ github.sha }}
           CANVAS_CHANGED: ${{ needs.changes.outputs.canvas }}
           EVENT_NAME: ${{ github.event_name }}
           REF_NAME: ${{ github.ref }}
-          # github.server_url resolves via the workflow-level env override
-          # to the Gitea instance, so the RUN_URL points at the Gitea run
-          # page (not github.com). See feedback_act_runner_github_server_url.
-          RUN_URL: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}
+          # github.server_url resolves via the workflow-level env override to the
+          # Gitea instance, so RUN_URL points at the Gitea run page (not github.com).
+          RUN_URL: ${{ github.server_url }}/${{ github.repository }}/actions
         run: |
           set -euo pipefail
           if [ "$CANVAS_CHANGED" != "true" ] || [ "$EVENT_NAME" != "push" ] || [ "$REF_NAME" != "refs/heads/main" ]; then
-            echo "Canvas deploy reminder not applicable for event=$EVENT_NAME ref=$REF_NAME canvas_changed=$CANVAS_CHANGED."
+            echo "Canvas deploy status not applicable for event=$EVENT_NAME ref=$REF_NAME canvas_changed=$CANVAS_CHANGED."
             exit 0
           fi
 
           # Write body to a temp file — avoids backtick escaping in shell.
-          cat > /tmp/deploy-reminder.md << 'BODY'
-          ## Canvas build passed — deploy required
+          cat > /tmp/deploy-status.md << 'BODY'
+          ## Canvas ordered deploy in progress — no manual action required
 
-          The `publish-canvas-image` workflow is now building a fresh Docker image
-          (`ghcr.io/molecule-ai/canvas:latest`) in the background.
+          This canvas-touching main push triggers `publish-canvas-image`, which now
+          runs an ORDERED, CI-gated deploy (core#2226) — the same shape as the
+          platform's deploy-production:
 
-          Once it completes (~3–5 min), apply on the host machine with:
-          ```bash
-          cd <runner-workspace>
-          git pull origin main
-          docker compose pull canvas && docker compose up -d canvas
-          ```
+          1. Build → push `molecule-ai/canvas:staging-<sha>` + `:staging-latest`.
+          2. Wait for green main CI on this SHA.
+          3. Promote `:latest` to the verified `:staging-<sha>` by digest.
 
-          If you need to rebuild from local source instead (e.g. testing unreleased
-          changes or a new `NEXT_PUBLIC_*` URL), use:
-          ```bash
-          docker compose build canvas && docker compose up -d canvas
-          ```
+          Tenants/hosts pin via `CANVAS_IMAGE_TAG` (default `latest` = the last
+          CI-green build), so a deploy is reproducible — no hand-run
+          `docker compose pull` needed. Watch the run in the canvas publish workflow.
           BODY
-          printf '\n> Posted automatically by CI · commit `%s` · [build log](%s)\n' \
-            "$COMMIT_SHA" "$RUN_URL" >> /tmp/deploy-reminder.md
+          printf '\n> Posted automatically by CI · commit `%s` · [publish workflow](%s)\n' \
+            "$COMMIT_SHA" "$RUN_URL" >> /tmp/deploy-status.md
 
-          # Gitea has no commit-comments API; write to GITHUB_STEP_SUMMARY,
-          # which both GitHub Actions and Gitea Actions render as the
-          # workflow run's summary page. (#75 / PR-D)
-          cat /tmp/deploy-reminder.md >> "$GITHUB_STEP_SUMMARY"
+          # Gitea has no commit-comments API; write to GITHUB_STEP_SUMMARY, which
+          # both GitHub and Gitea Actions render as the run's summary page.
+          cat /tmp/deploy-status.md >> "$GITHUB_STEP_SUMMARY"
 
   # Python Lint & Test — required check, always runs.
   # Runtime Python moved to molecule-ai-workspace-runtime. Keep this context as

.gitea/workflows/e2e-api.yml

@@ -327,7 +327,12 @@ jobs:
           # start-redis steps point at this run's per-run host ports.
           ./platform-server > platform.log 2>&1 &
           echo $! > platform.pid
-      - name: Wait for /health
+      - name: Wait for /health (with migration completion gate)
+        # Issue #2205: 30 one-second probes is insufficient when the migration
+        # chain is still running; /health can flip true before migrations
+        # finish, so subsequent steps that touch the DB fail. Hybrid fix:
+        # bump timeout to 300s AND gate exit on the same workspaces-table
+        # existence check the downstream "Assert migrations applied" uses.
         if: needs.detect-changes.outputs.api == 'true'
         run: |
           # Readiness signal: the platform binds /health only AFTER the full
@@ -343,13 +348,21 @@ jobs:
           # background platform-server process has exited (e.g. a broken
           # migration crashed it), we stop and fail loudly at once instead of
           # waiting out the whole budget.
-          DEADLINE_SECS=180          # cold-start + full migration chain headroom
+          #
+          # Issue #2205: /health can flip true before migrations finish on a
+          # growing chain, so we gate exit on the workspaces-table existence
+          # check the downstream "Assert migrations applied" uses.
+          DEADLINE_SECS=300          # cold-start + full migration chain headroom
           PLATFORM_PID="$(cat workspace-server/platform.pid 2>/dev/null || true)"
           start=$(date +%s)
           while :; do
             if curl -sf "$BASE/health" > /dev/null; then
-              echo "Platform healthy after $(( $(date +%s) - start ))s"
-              exit 0
+              tables=$(docker exec "$PG_CONTAINER" psql -U dev -d molecule -tAc \
+                "SELECT count(*) FROM information_schema.tables WHERE table_schema='public' AND table_name='workspaces'" 2>/dev/null || echo "0")
+              if [ "$tables" = "1" ]; then
+                echo "Platform healthy + migrations applied after $(( $(date +%s) - start ))s"
+                exit 0
+              fi
             fi
             # Fast-fail: if the platform process died, /health will never come.
             if [ -n "$PLATFORM_PID" ] && ! kill -0 "$PLATFORM_PID" 2>/dev/null; then
@@ -358,12 +371,13 @@ jobs:
               exit 1
             fi
             if [ "$(( $(date +%s) - start ))" -ge "$DEADLINE_SECS" ]; then
-              echo "::error::Platform did not become healthy within ${DEADLINE_SECS}s — see log below"
+              echo "::error::Platform did not become healthy with migrations applied within ${DEADLINE_SECS}s — see log below"
               cat workspace-server/platform.log || true
               exit 1
             fi
             sleep 1
           done
+
       - name: Assert migrations applied
         if: needs.detect-changes.outputs.api == 'true'
         run: |

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

@@ -126,6 +126,7 @@ jobs:
   # push/dispatch/cron only (30+ min). This is NOT a fake-green mask of
   # the real assertion — it validates the driving script's bash syntax
   # and inline-python so a broken test script fails at PR time.
+  # bp-required: pending #1296 — PR emitter, not yet required (tracked in #1296).
   pr-validate:
     name: E2E Peer Visibility
     runs-on: ubuntu-latest

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

@@ -167,16 +167,30 @@ jobs:
       - if: needs.detect-changes.outputs.canvas == 'true'
         uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
 
-      - name: Verify admin token present
+      # Skip-if-absent (core#2225), mirroring the serving-e2e gate's
+      # skip-if-secret-unset contract: a MISSING CI secret is an operator
+      # CONFIG gap, not a code regression, so it must not paint this E2E
+      # red. When CP_STAGING_ADMIN_API_TOKEN is unset we emit a LOUD
+      # ::warning:: + ::notice:: and skip the real provision/test steps (the
+      # job still completes green). When the secret IS present we run the
+      # full suite exactly as before. Operators: set
+      # CP_STAGING_ADMIN_API_TOKEN as a repo/org Actions secret on
+      # molecule-core to actually exercise this E2E.
+      - name: Check admin token (skip-if-absent)
+        id: token_check
         if: needs.detect-changes.outputs.canvas == 'true'
         run: |
           if [ -z "$MOLECULE_ADMIN_TOKEN" ]; then
-            echo "::error::Missing CP_STAGING_ADMIN_API_TOKEN"
-            exit 2
+            echo "::warning::CP_STAGING_ADMIN_API_TOKEN is not set on this runner — SKIPPING the staging canvas E2E (cannot auth to staging CP). This is an operator config gap, not a code failure; set the secret on molecule-core (repo or org Actions secrets) to run it. See core#2225."
+            echo "::notice::E2E Staging Canvas skipped: CP_STAGING_ADMIN_API_TOKEN absent."
+            echo "present=false" >> "$GITHUB_OUTPUT"
+          else
+            echo "CP_STAGING_ADMIN_API_TOKEN present ✓ — running staging canvas E2E."
+            echo "present=true" >> "$GITHUB_OUTPUT"
           fi
 
       - name: Set up Node
-        if: needs.detect-changes.outputs.canvas == 'true'
+        if: needs.detect-changes.outputs.canvas == 'true' && steps.token_check.outputs.present == 'true'
         uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e # v6.4.0
         with:
           node-version: '20'
@@ -184,11 +198,11 @@ jobs:
           cache-dependency-path: canvas/package-lock.json
 
       - name: Install canvas deps
-        if: needs.detect-changes.outputs.canvas == 'true'
+        if: needs.detect-changes.outputs.canvas == 'true' && steps.token_check.outputs.present == 'true'
         run: npm ci
 
       - name: Install Playwright browsers
-        if: needs.detect-changes.outputs.canvas == 'true'
+        if: needs.detect-changes.outputs.canvas == 'true' && steps.token_check.outputs.present == 'true'
         timeout-minutes: 10
         run: |
           PREBAKED_PLAYWRIGHT=/ms-playwright
@@ -200,7 +214,7 @@ jobs:
           npx playwright install --with-deps chromium
 
       - name: Run staging canvas E2E
-        if: needs.detect-changes.outputs.canvas == 'true'
+        if: needs.detect-changes.outputs.canvas == 'true' && steps.token_check.outputs.present == 'true'
         run: npx playwright test --config=playwright.staging.config.ts
 
       - name: Upload Playwright report on failure

.gitea/workflows/e2e-staging-reconciler.yml

@@ -0,0 +1,196 @@
+name: E2E Staging Reconciler (heals terminated EC2)
+
+# Live staging proof for the core#2261 instance-state reconciler
+# (workspace-server/internal/registry/cp_instance_reconciler.go). The
+# real-infra complement to the deterministic unit tests: provisions a real
+# staging workspace, TERMINATES its EC2, and asserts the reconciler flips it
+# off 'online' (PRIMARY gate) and auto-reprovisions on a new instance_id
+# (SECONDARY, best-effort). See
+# tests/e2e/test_reconciler_heals_terminated_instance.sh for the assertion
+# contract + timeouts.
+#
+# Modeled on e2e-staging-saas.yml. Same secrets + same Gitea-port caveats:
+#   - Dropped workflow_dispatch.inputs (Gitea 1.22.6 parser rejects them).
+#   - Dropped merge_group / environment (no Gitea equivalent).
+#   - Workflow-level env.GITHUB_SERVER_URL pinned per
+#     feedback_act_runner_github_server_url.
+#
+# NOT a required check (yet). This is a brand-new live E2E that provisions +
+# terminates real EC2 (costs money, shares the cp#245 cold-boot flake
+# surface). A new live e2e must NOT hard-gate every merge until it has a
+# green track record. continue-on-error: true surfaces failures without
+# blocking. PROMOTE to branch-required (flip continue-on-error → false AND
+# add "E2E Staging Reconciler" to branch protection) once it has run green on
+# main for several consecutive days — same de-flake discipline the
+# platform-boot job in e2e-staging-saas.yml documents.
+
+on:
+  # Run when the reconciler itself, the script, or the libs it depends on
+  # change — so a reconciler regression is caught on the PR that introduces
+  # it (paths filter), plus a daily schedule to catch infra/AMI drift.
+  push:
+    branches: [main]
+    paths:
+      - 'workspace-server/internal/registry/cp_instance_reconciler.go'
+      - 'tests/e2e/test_reconciler_heals_terminated_instance.sh'
+      - 'tests/e2e/lib/aws_leak_check.sh'
+      - 'tests/e2e/lib/model_slug.sh'
+      - '.gitea/workflows/e2e-staging-reconciler.yml'
+  pull_request:
+    branches: [main]
+    paths:
+      - 'workspace-server/internal/registry/cp_instance_reconciler.go'
+      - 'tests/e2e/test_reconciler_heals_terminated_instance.sh'
+      - 'tests/e2e/lib/aws_leak_check.sh'
+      - 'tests/e2e/lib/model_slug.sh'
+      - '.gitea/workflows/e2e-staging-reconciler.yml'
+  workflow_dispatch:
+  schedule:
+    # 08:00 UTC daily — offset from e2e-staging-saas (07:00) so the two live
+    # harnesses don't fight over staging's per-hour org-creation quota.
+    - cron: '0 8 * * *'
+
+# Serialize against itself: staging has a finite per-hour org-creation quota,
+# and a cancelled run mid-teardown leaks EC2. cancel-in-progress: false
+# mirrors e2e-staging-saas.yml.
+concurrency:
+  group: e2e-staging-reconciler
+  cancel-in-progress: false
+
+env:
+  GITHUB_SERVER_URL: https://git.moleculesai.app
+
+jobs:
+  # PR-validation path: always posts success so a workflow-only / script-only
+  # PR has a status check (this workflow's real job only fires on the paths
+  # filter). Mirrors the pr-validate job in e2e-staging-saas.yml.
+  pr-validate:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+        with:
+          fetch-depth: 1
+        continue-on-error: true
+      - name: YAML validation (best-effort)
+        run: |
+          echo "e2e-staging-reconciler.yml — PR validation: workflow YAML is valid."
+          echo "Live E2E step runs only when the reconciler / script / libs change."
+        continue-on-error: true
+
+  e2e-staging-reconciler:
+    name: E2E Staging Reconciler
+    runs-on: ubuntu-latest
+    # NOT required yet — surface failures without blocking merges. Flip to
+    # false + add to branch protection once green on main for a de-flake
+    # window (see the header note). mc#1982: do not renew this mask silently.
+    continue-on-error: true
+    timeout-minutes: 60
+    permissions:
+      contents: read
+
+    env:
+      MOLECULE_CP_URL: https://staging-api.moleculesai.app
+      # Single admin-bearer secret drives provision + tenant-token retrieval +
+      # teardown (= Railway staging CP_ADMIN_API_TOKEN). Same secret name the
+      # saas workflow canonicalised to under internal#322.
+      MOLECULE_ADMIN_TOKEN: ${{ secrets.CP_STAGING_ADMIN_API_TOKEN }}
+      AWS_ACCESS_KEY_ID: ${{ secrets.AWS_ACCESS_KEY_ID }}
+      AWS_SECRET_ACCESS_KEY: ${{ secrets.AWS_SECRET_ACCESS_KEY }}
+      AWS_DEFAULT_REGION: us-east-2
+      # Leak-check is REQUIRED here: this test deliberately terminates an EC2,
+      # so teardown MUST positively confirm no slug-tagged box survives.
+      E2E_AWS_LEAK_CHECK: required
+      E2E_AWS_TERMINATE_LEAKS: '1'
+      # claude-code + MiniMax is the cheapest boot-to-online path (same as the
+      # saas job). The reconciler test never makes a completion, but the key is
+      # wired so the first boot reaches online on the same path the saas
+      # harness uses. First non-empty wins in the script's priority chain.
+      E2E_MINIMAX_API_KEY: ${{ secrets.MOLECULE_STAGING_MINIMAX_API_KEY }}
+      E2E_ANTHROPIC_API_KEY: ${{ secrets.MOLECULE_STAGING_ANTHROPIC_API_KEY }}
+      E2E_OPENAI_API_KEY: ${{ secrets.MOLECULE_STAGING_OPENAI_API_KEY }}
+      E2E_RUNTIME: claude-code
+      E2E_MODEL_SLUG: MiniMax-M2
+      E2E_RUN_ID: "${{ github.run_id }}-${{ github.run_attempt }}"
+      E2E_KEEP_ORG: ${{ github.event.inputs.keep_org && '1' || '0' }}
+
+    steps:
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+
+      - name: Verify required secrets present
+        run: |
+          if [ -z "$MOLECULE_ADMIN_TOKEN" ]; then
+            echo "::error::CP_STAGING_ADMIN_API_TOKEN secret not set (Railway staging CP_ADMIN_API_TOKEN)"
+            exit 2
+          fi
+          for var in AWS_ACCESS_KEY_ID AWS_SECRET_ACCESS_KEY; do
+            if [ -z "${!var:-}" ]; then
+              echo "::error::$var not set — this test terminates an EC2 and verifies no leak; AWS creds are mandatory"
+              exit 2
+            fi
+          done
+          echo "Required secrets present ✓"
+
+      - name: CP staging health preflight
+        run: |
+          code=$(curl -sS -o /dev/null -w "%{http_code}" --max-time 10 "$MOLECULE_CP_URL/health")
+          if [ "$code" != "200" ]; then
+            echo "::error::Staging CP unhealthy (got HTTP $code). Skipping — not a reconciler bug."
+            exit 1
+          fi
+          echo "Staging CP healthy ✓"
+
+      - name: Run reconciler heal E2E
+        id: e2e
+        run: bash tests/e2e/test_reconciler_heals_terminated_instance.sh
+
+      # Belt-and-braces teardown: the script installs its own EXIT trap, but if
+      # the runner is cancelled the trap may not fire. This always() step
+      # double-deletes any e2e-rec-* org from THIS run. The admin DELETE is
+      # idempotent so double-invoking is safe.
+      - name: Teardown safety net (runs on cancel/failure)
+        if: always()
+        env:
+          ADMIN_TOKEN: ${{ secrets.CP_STAGING_ADMIN_API_TOKEN }}
+        run: |
+          set +e
+          orgs=$(curl -sS "$MOLECULE_CP_URL/cp/admin/orgs" \
+            -H "Authorization: Bearer $ADMIN_TOKEN" 2>/dev/null \
+            | python3 -c "
+          import json, sys, os, datetime
+          run_id = os.environ.get('GITHUB_RUN_ID', '')
+          d = json.load(sys.stdin)
+          today = datetime.date.today()
+          yesterday = today - datetime.timedelta(days=1)
+          dates = (today.strftime('%Y%m%d'), yesterday.strftime('%Y%m%d'))
+          # Slug shape: e2e-rec-YYYYMMDD-<run_id>-<attempt>-...
+          if run_id:
+              prefixes = tuple(f'e2e-rec-{d}-{run_id}-' for d in dates)
+          else:
+              prefixes = tuple(f'e2e-rec-{d}-' for d in dates)
+          candidates = [o['slug'] for o in d.get('orgs', [])
+                        if any(o.get('slug','').startswith(p) for p in prefixes)
+                        and o.get('instance_status') not in ('purged',)]
+          print('\n'.join(candidates))
+          " 2>/dev/null)
+          leaks=()
+          for slug in $orgs; do
+            echo "Safety-net teardown: $slug"
+            set +e
+            curl -sS -o /tmp/rec-cleanup.out -w "%{http_code}" \
+              -X DELETE "$MOLECULE_CP_URL/cp/admin/tenants/$slug" \
+              -H "Authorization: Bearer $ADMIN_TOKEN" \
+              -H "Content-Type: application/json" \
+              -d "{\"confirm\":\"$slug\"}" >/tmp/rec-cleanup.code
+            set -e
+            code=$(cat /tmp/rec-cleanup.code 2>/dev/null || echo "000")
+            if [ "$code" = "200" ] || [ "$code" = "204" ]; then
+              echo "[teardown] deleted $slug (HTTP $code)"
+            else
+              echo "::warning::reconciler teardown for $slug returned HTTP $code — sweep-stale-e2e-orgs will catch it within ~45 min. Body: $(head -c 300 /tmp/rec-cleanup.out 2>/dev/null)"
+              leaks+=("$slug")
+            fi
+          done
+          if [ ${#leaks[@]} -gt 0 ]; then
+            echo "::warning::reconciler teardown left ${#leaks[@]} leak(s): ${leaks[*]}"
+          fi
+          exit 0

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

@@ -243,7 +243,8 @@ jobs:
           # MUST exist for the integration tests to be meaningful. Hard-
           # fail if any didn't land — that would be a real regression we
           # want loud.
-          for tbl in delegations workspaces activity_logs pending_uploads; do
+          # workspace_schedules added for the #2149 scheduler integration tests.
+          for tbl in delegations workspaces activity_logs pending_uploads workspace_schedules; do
             if ! psql -h "${PG_HOST}" -U postgres -d molecule -tA \
                 -c "SELECT 1 FROM information_schema.tables WHERE table_name = '$tbl'" \
                 | grep -q 1; then
@@ -274,6 +275,16 @@ jobs:
           # workflow runs don't fight over a host-net 5432 port.
           go test -tags=integration -timeout 5m -v ./internal/handlers/ -run "^TestIntegration_"
 
+      - if: needs.detect-changes.outputs.handlers == 'true'
+        name: Run scheduler integration tests (#2149)
+        run: |
+          # #2149: real-PG regression coverage for the scheduler firing loop
+          # (tick → A2A fire → write-back of last_run_at/next_run_at/run_count/
+          # activity_logs jsonb incl. invalid-UTF-8 sanitization + sweepPhantomBusy).
+          # Reuses the same migrated Postgres (workspace_schedules / activity_logs
+          # / workspaces all landed by the migration replay step above).
+          go test -tags=integration -timeout 5m -v ./internal/scheduler/ -run "^TestIntegration_"
+
       - if: failure() && needs.detect-changes.outputs.handlers == 'true'
         name: Diagnostic dump on failure
         env:

.gitea/workflows/publish-canvas-image.yml

@@ -14,10 +14,37 @@ name: publish-canvas-image
 #     authenticate to ghcr.io.
 #
 
-# Builds and pushes the canvas Docker image to ECR whenever a commit lands
-# on main that touches canvas code. Previously canvas changes were visible in
-# CI (npm run build passed) but the live container was never updated —
-# operators had to manually run `docker compose build canvas` each time.
+# Builds, pushes, and (ordered) deploys the standalone canvas Docker image to
+# ECR whenever a commit lands on main that touches canvas code.
+#
+# Ordered deploy (core#2226) — mirrors publish-workspace-server-image.yml so the
+# standalone `molecule-ai/canvas` image is deterministic + verifiable, not a
+# side effect of the platform fleet pulling a mutable `:latest`:
+#
+#   build-and-push:  build → push :staging-<sha> + :staging-latest + :sha-<sha>
+#                    (does NOT move :latest — an unpromoted build must never
+#                    become the prod-blessed tag).
+#   promote-canvas:  waits for green main CI on this SHA, then re-points
+#                    :latest to the verified :staging-<sha> by digest
+#                    (imagetools create — no rebuild). So `:latest` == the
+#                    current prod-blessed canvas, byte-identical to staging-<sha>.
+#
+# Tag scheme produced (parallels platform-tenant):
+#   :staging-<sha> — per-commit immutable digest, what docker-compose pins to.
+#   :staging-latest — most recent BUILD on main (last-writer-wins, NOT gated).
+#   :sha-<sha>     — kept for back-compat with any consumer pinning the old tag.
+#   :latest        — most recent CI-GREEN build. Only moved by promote-canvas.
+#
+# WHY this is the canvas analogue of the platform's deploy-production, not a
+# literal copy: the standalone canvas co-deploys with the platform on the same
+# host via the root docker-compose.yml (`docker compose pull && up -d`). Gating
+# the canvas `:latest` promotion on the SAME green-main-CI signal the platform
+# deploy waits on makes platform + canvas roll together by the same SHA. The
+# canvas has no per-tenant fleet of its own and no /buildinfo endpoint, so there
+# is no fleet-rollout / per-tenant verify step to mirror here — CI-green +
+# digest-pin + immutable :staging-<sha> is the determinism contract. (A future
+# canvas /buildinfo would let this assert the served SHA like the platform does;
+# tracked in core#2226.)
 #
 # Mirror of publish-platform-image.yml, adapted for the Next.js canvas layer.
 # See that workflow for inline notes on macOS Keychain isolation and QEMU.
@@ -30,6 +57,7 @@ on:
       # platform-only / docs-only / MCP-only merges.
       - 'canvas/**'
       - '.gitea/workflows/publish-canvas-image.yml'
+  workflow_dispatch:
   # NOTE (Gitea port): the original GitHub workflow had a
   # `workflow_dispatch:` manual trigger for the
   # non-canvas-merge-but-need-fresh-image scenario. Dropped in the
@@ -69,6 +97,10 @@ jobs:
     # Phase 3 (RFC #219 §1): surface broken workflows without blocking.
     # mc#1982: pre-existing continue-on-error mask; root-fix and remove, do not renew silently.
     continue-on-error: true
+    outputs:
+      # Exposed so promote-canvas re-points :latest to the EXACT per-commit tag
+      # this build produced (digest-level), never a re-resolved mutable tag.
+      staging_sha: ${{ steps.tags.outputs.staging_sha }}
     steps:
       - name: Checkout
         uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
@@ -140,6 +172,7 @@ jobs:
         shell: bash
         run: |
           echo "sha=${GITHUB_SHA::7}" >> "$GITHUB_OUTPUT"
+          echo "staging_sha=staging-${GITHUB_SHA::7}" >> "$GITHUB_OUTPUT"
 
       - name: Resolve build args
         id: build_args
@@ -175,8 +208,19 @@ jobs:
           build-args: |
             NEXT_PUBLIC_PLATFORM_URL=${{ steps.build_args.outputs.platform_url }}
             NEXT_PUBLIC_WS_URL=${{ steps.build_args.outputs.ws_url }}
+            # Bake the merge SHA into the image so /api/buildinfo reports the
+            # served canvas SHA (core#2235). Mirrors how the platform image
+            # surfaces GIT_SHA at /buildinfo. Full 40-char SHA (not the
+            # 7-char tag) so the fleet redeploy verification can match exactly.
+            BUILD_SHA=${{ github.sha }}
+          # Ordered deploy (core#2226): the build job pushes the immutable
+          # per-commit tag + the build-tracking staging-latest + the legacy
+          # back-compat :sha-<sha> tag. It does NOT push :latest — :latest is
+          # the prod-blessed tag and is only re-pointed by promote-canvas after
+          # green main CI, so an unpromoted/red build can never become :latest.
           tags: |
-            ${{ env.IMAGE_NAME }}:latest
+            ${{ env.IMAGE_NAME }}:${{ steps.tags.outputs.staging_sha }}
+            ${{ env.IMAGE_NAME }}:staging-latest
             ${{ env.IMAGE_NAME }}:sha-${{ steps.tags.outputs.sha }}
           # Gitea artifact-cache reachability is best-effort on the operator
           # runner network. Do not let cache export fail an image that already
@@ -185,3 +229,107 @@ jobs:
             org.opencontainers.image.source=https://git.moleculesai.app/${{ github.repository }}
             org.opencontainers.image.revision=${{ github.sha }}
             org.opencontainers.image.description=Molecule AI canvas (Next.js 15 + React Flow)
+
+  # bp-exempt: post-merge canvas promote side-effect; merge is gated by CI /
+  # all-required and this job waits for green push CI on the SHA before acting.
+  promote-canvas:
+    name: Promote canvas :latest to CI-green build
+    needs: build-and-push
+    # Only on a real main push — workflow_dispatch / non-main never promotes.
+    if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/main' }}
+    # Side-effect deploy only; the image publish above is the durable artifact.
+    # mc#1982: do NOT renew this mask silently — it mirrors deploy-production's
+    # contract (a flaky promote must not red the ship lane), tracked in core#2226.
+    continue-on-error: true
+    runs-on: publish
+    timeout-minutes: 60
+    env:
+      # Same green-main-CI gate the platform deploy-production waits on, so
+      # platform + canvas advance :latest off the identical signal/SHA.
+      GITEA_HOST: git.moleculesai.app
+      GITEA_TOKEN: ${{ secrets.PROD_AUTO_DEPLOY_CONTROL_TOKEN || secrets.AUTO_SYNC_TOKEN }}
+      CI_STATUS_TIMEOUT_SECONDS: "3600"
+      # Re-uses the platform's disable kill-switch: when prod auto-deploy is
+      # paused, the canvas :latest promote pauses too (correct — an unpromoted
+      # build must not become :latest while the fleet is frozen).
+      PROD_AUTO_DEPLOY_DISABLED: ${{ vars.PROD_AUTO_DEPLOY_DISABLED || secrets.PROD_AUTO_DEPLOY_DISABLED || '' }}
+    steps:
+      # The publish runner's default HOME (/home/hongming) is not writable, so
+      # docker credential saves fail and halt the promote (#2193 on the platform
+      # side). Point HOME + DOCKER_CONFIG at the writable job temp dir.
+      - name: Prepare writable HOME + Docker config
+        run: |
+          set -euo pipefail
+          H="$RUNNER_TEMP/canvas-promote-home"
+          mkdir -p "$H/.docker"
+          echo "HOME=$H" >> "$GITHUB_ENV"
+          echo "DOCKER_CONFIG=$H/.docker" >> "$GITHUB_ENV"
+
+      - name: Checkout
+        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+
+      - name: Resolve promote gate
+        id: gate
+        env:
+          PROD_AUTO_DEPLOY_DISABLED: ${{ env.PROD_AUTO_DEPLOY_DISABLED }}
+        run: |
+          set -euo pipefail
+          if [ -n "${PROD_AUTO_DEPLOY_DISABLED:-}" ]; then
+            case "$(printf '%s' "$PROD_AUTO_DEPLOY_DISABLED" | tr '[:upper:]' '[:lower:]')" in
+              1|true|yes|on|disabled|disable)
+                echo "enabled=false" >> "$GITHUB_OUTPUT"
+                echo "::notice::Canvas :latest promote skipped: PROD_AUTO_DEPLOY_DISABLED=$PROD_AUTO_DEPLOY_DISABLED"
+                {
+                  echo "## Canvas :latest promote skipped"
+                  echo ""
+                  echo "Reason: \`PROD_AUTO_DEPLOY_DISABLED=$PROD_AUTO_DEPLOY_DISABLED\`. The CI-green build is published as \`:staging-${GITHUB_SHA::7}\`; \`:latest\` was left unchanged."
+                } >> "$GITHUB_STEP_SUMMARY"
+                exit 0 ;;
+            esac
+          fi
+          if [ -z "${GITEA_TOKEN:-}" ]; then
+            echo "::error::AUTO_SYNC_TOKEN/PROD_AUTO_DEPLOY_CONTROL_TOKEN is required so the canvas promote can wait for green CI."
+            exit 1
+          fi
+          echo "enabled=true" >> "$GITHUB_OUTPUT"
+
+      - name: Wait for green main CI on this SHA
+        if: ${{ steps.gate.outputs.enabled == 'true' }}
+        run: |
+          set -euo pipefail
+          # Same SSOT wait the platform deploy uses: blocks until the required
+          # push contexts (CI / all-required (push) + Secret scan) go green on
+          # THIS sha, and fails closed if any required context terminally fails.
+          python3 .gitea/scripts/prod-auto-deploy.py wait-ci
+
+      - name: Promote canvas :latest to the CI-green image
+        if: ${{ steps.gate.outputs.enabled == 'true' }}
+        env:
+          IMAGE_NAME: ${{ env.IMAGE_NAME }}
+          STAGING_SHA_TAG: ${{ needs.build-and-push.outputs.staging_sha }}
+          AWS_ACCESS_KEY_ID: ${{ secrets.AWS_ACCESS_KEY_ID }}
+          AWS_SECRET_ACCESS_KEY: ${{ secrets.AWS_SECRET_ACCESS_KEY }}
+          AWS_DEFAULT_REGION: us-east-2
+        run: |
+          set -euo pipefail
+          # Fail-safe: if the build job's output didn't propagate, recompute the
+          # immutable per-commit tag from the SHA so we never promote a guess.
+          SHA_TAG="${STAGING_SHA_TAG:-staging-${GITHUB_SHA::7}}"
+          ECR_REGISTRY="${IMAGE_NAME%%/*}"
+          aws ecr get-login-password --region us-east-2 | \
+            docker login --username AWS --password-stdin "${ECR_REGISTRY}"
+
+          # Digest-level re-tag (no pull/rebuild): :latest becomes byte-identical
+          # to the verified :staging-<sha> for this commit.
+          docker buildx imagetools create \
+            --tag "${IMAGE_NAME}:latest" \
+            "${IMAGE_NAME}:${SHA_TAG}"
+
+          {
+            echo "## Canvas :latest promoted"
+            echo ""
+            echo "Re-pointed \`molecule-ai/canvas:latest\` → \`${SHA_TAG}\` (by digest)."
+            echo ":latest now tracks the CI-green canvas build for commit \`${GITHUB_SHA::7}\`."
+            echo ""
+            echo "Tenants/hosts that \`docker compose pull canvas\` now get the same build the platform deploy rolled for this SHA."
+          } >> "$GITHUB_STEP_SUMMARY"

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

@@ -115,6 +115,19 @@ jobs:
           echo "Docker daemon OK"
           echo "::endgroup::"
 
+      # Pre-flight: verify every repo in manifest.json actually exists.
+      #
+      # Why: deleting a template repo without updating manifest.json breaks
+      # clone-manifest.sh with a generic git 404, which looks like a
+      # transient network error and wastes debug time. We catch it here
+      # with a per-entry ::error:: annotation naming the missing repo
+      # (issue #2192). This is the push-time complement to PR #2186's
+      # PR-time manifest-entry-existence gate.
+      - name: Validate manifest entries exist
+        run: |
+          set -euo pipefail
+          bash scripts/check-manifest-repos-exist.sh manifest.json
+
       # Pre-clone manifest deps before docker build.
       #
       # Why: workspace-template-* repos on Gitea are private. The pre-fix
@@ -262,6 +275,11 @@ jobs:
       PROD_AUTO_DEPLOY_BATCH_SIZE: ${{ vars.PROD_AUTO_DEPLOY_BATCH_SIZE || '3' }}
       PROD_AUTO_DEPLOY_DRY_RUN: ${{ vars.PROD_AUTO_DEPLOY_DRY_RUN || '' }}
       PROD_ALLOW_NON_PROD_CP_URL: ${{ vars.PROD_ALLOW_NON_PROD_CP_URL || '' }}
+      # #2213: per-tenant /buildinfo settle budget. A freshly-swapped tenant can
+      # keep serving the old image at the edge for a short drain window; the
+      # verify step polls each tenant up to this budget before declaring it stale.
+      PROD_AUTO_DEPLOY_VERIFY_BUDGET_SECONDS: ${{ vars.PROD_AUTO_DEPLOY_VERIFY_BUDGET_SECONDS || '240' }}
+      PROD_AUTO_DEPLOY_VERIFY_INTERVAL_SECONDS: ${{ vars.PROD_AUTO_DEPLOY_VERIFY_INTERVAL_SECONDS || '20' }}
     steps:
       # The publish runner's default HOME (/home/hongming) is not writable, so
       # git/docker credential saves fail (`Error saving credentials: mkdir
@@ -320,8 +338,50 @@ jobs:
           set -euo pipefail
           python3 .gitea/scripts/prod-auto-deploy.py wait-ci
 
-      - name: Call production CP redeploy-fleet
+      # Superseded-job guard — BEFORE any production side effect (#2213).
+      #
+      # This workflow has no `concurrency:` (see header: Gitea 1.22.6 cancels
+      # queued prod deploys). So two close main pushes run BOTH deploy-production
+      # jobs. The verify step already skips its strict /buildinfo check when this
+      # job is superseded (#2194) — but that guard was AFTER the redeploy and the
+      # :latest promote, so an OLDER job that started late still:
+      #   1. rolled the whole fleet BACKWARD to its older tag (canary hongming
+      #      was reverted from the newer SHA — the #2213 red), then
+      #   2. promoted :latest backward to the older image,
+      # and only THEN skipped verify and exited green. A superseded job must do
+      # NEITHER. We re-check the branch head here, immediately before the rollout,
+      # and skip every side effect when a newer commit already owns main.
+      #
+      # exit 0 + non-empty stdout => superseded (newer head printed); the redeploy
+      # and promote steps are gated off via this output. exit 10 => this job is
+      # still the latest, proceed to roll the fleet. Fail-safe: a head that can't
+      # be read returns NOT-superseded (exit 10), so a genuine deploy is never
+      # silently skipped. (Re-checked again at verify time to catch a newer job
+      # that lands DURING this rollout.)
+      - name: Check superseded before production side effects
+        id: supersede
         if: ${{ steps.plan.outputs.enabled == 'true' }}
+        run: |
+          set -euo pipefail
+          set +e
+          NEWER_HEAD="$(python3 .gitea/scripts/prod-auto-deploy.py check-superseded)"
+          SUPERSEDED_EXIT=$?
+          set -e
+          if [ "$SUPERSEDED_EXIT" -eq 0 ] && [ -n "$NEWER_HEAD" ]; then
+            echo "superseded=true" >> "$GITHUB_OUTPUT"
+            echo "::notice::Superseded before rollout: main head is now ${NEWER_HEAD:0:7} (this job deploys ${GITHUB_SHA:0:7}). Skipping redeploy + :latest promote so an older job never rolls the fleet backward."
+            {
+              echo "## Production auto-deploy skipped — superseded before rollout"
+              echo ""
+              echo "This deploy job's SHA \`${GITHUB_SHA:0:7}\` is no longer the head of \`main\` (now \`${NEWER_HEAD:0:7}\`)."
+              echo "A newer deploy job owns the fleet; rolling it backward to this older build would revert tenants and \`:latest\`. No side effects performed."
+            } >> "$GITHUB_STEP_SUMMARY"
+          else
+            echo "superseded=false" >> "$GITHUB_OUTPUT"
+          fi
+
+      - name: Call production CP redeploy-fleet
+        if: ${{ steps.plan.outputs.enabled == 'true' && steps.supersede.outputs.superseded != 'true' }}
         run: |
           set -euo pipefail
           python3 .gitea/scripts/prod-auto-deploy.py assert-enabled
@@ -380,7 +440,11 @@ jobs:
           fi
 
       - name: Verify reachable tenants report this SHA
-        if: ${{ steps.plan.outputs.enabled == 'true' }}
+        # Skip when superseded BEFORE rollout: the redeploy step did not run, so
+        # there is no redeploy-fleet response to verify against and the newer job
+        # owns verification (#2213). The in-step guard below still catches the
+        # case where a newer job lands DURING this job's rollout.
+        if: ${{ steps.plan.outputs.enabled == 'true' && steps.supersede.outputs.superseded != 'true' }}
         env:
           TENANT_DOMAIN: moleculesai.app
         run: |
@@ -422,6 +486,20 @@ jobs:
             exit 1
           fi
 
+          # Per-tenant settle/retry budget (#2213). A tenant whose container the
+          # CP just swapped can keep serving the OLD image at the edge for a short
+          # window while the old container drains — /buildinfo returns HTTP 200
+          # with the previous SHA, which `curl --retry` does NOT retry (it only
+          # retries connection/5xx failures, not a stale-but-200 body). Without a
+          # settle window a still-rolling tenant false-reds "stale" on the very
+          # first poll. So poll each tenant's /buildinfo until it reports the
+          # target SHA or the budget is exhausted; only THEN declare it stale or
+          # unreachable. This never masks a genuinely stuck tenant — a tenant that
+          # never reaches the target within the budget still fails loud (and the
+          # superseded-job revert class is already blocked before rollout above).
+          SETTLE_BUDGET_SECONDS="${PROD_AUTO_DEPLOY_VERIFY_BUDGET_SECONDS:-240}"
+          SETTLE_INTERVAL_SECONDS="${PROD_AUTO_DEPLOY_VERIFY_INTERVAL_SECONDS:-20}"
+
           STALE_COUNT=0
           UNREACHABLE_COUNT=0
           UNHEALTHY_COUNT=0
@@ -433,18 +511,36 @@ jobs:
               continue
             fi
             url="https://${slug}.${TENANT_DOMAIN}/buildinfo"
-            body="$(curl -sS --max-time 30 --retry 3 --retry-delay 5 --retry-connrefused "$url" || true)"
-            actual="$(echo "$body" | jq -r '.git_sha // ""' 2>/dev/null || echo "")"
-            if [ -z "$actual" ]; then
-              echo "::error::$slug did not return /buildinfo after deploy."
-              UNREACHABLE_COUNT=$((UNREACHABLE_COUNT + 1))
-              continue
-            fi
-            if [ "$actual" != "$GITHUB_SHA" ]; then
-              echo "::error::$slug is stale: actual=${actual:0:7}, expected=${GITHUB_SHA:0:7}"
-              STALE_COUNT=$((STALE_COUNT + 1))
-            else
+            deadline=$(( $(date +%s) + SETTLE_BUDGET_SECONDS ))
+            actual=""
+            last_actual=""
+            on_target=false
+            while :; do
+              body="$(curl -sS --max-time 30 --retry 3 --retry-delay 5 --retry-connrefused "$url" || true)"
+              actual="$(echo "$body" | jq -r '.git_sha // ""' 2>/dev/null || echo "")"
+              [ -n "$actual" ] && last_actual="$actual"
+              if [ "$actual" = "$GITHUB_SHA" ]; then
+                on_target=true
+                break
+              fi
+              now=$(date +%s)
+              if [ "$now" -ge "$deadline" ]; then
+                break
+              fi
+              # Still rolling (stale 200) or transiently unreachable — wait and
+              # re-poll within the settle budget rather than failing on first read.
+              remaining=$(( deadline - now ))
+              echo "$slug: waiting for target SHA (have '${actual:0:7}', want ${GITHUB_SHA:0:7}; ${remaining}s left)"
+              sleep "$SETTLE_INTERVAL_SECONDS"
+            done
+            if [ "$on_target" = true ]; then
               echo "$slug: ${actual:0:7}"
+            elif [ -z "$last_actual" ]; then
+              echo "::error::$slug did not return /buildinfo after deploy (waited ${SETTLE_BUDGET_SECONDS}s)."
+              UNREACHABLE_COUNT=$((UNREACHABLE_COUNT + 1))
+            else
+              echo "::error::$slug is stale: actual=${last_actual:0:7}, expected=${GITHUB_SHA:0:7} (waited ${SETTLE_BUDGET_SECONDS}s)"
+              STALE_COUNT=$((STALE_COUNT + 1))
             fi
           done
 
@@ -488,8 +584,12 @@ jobs:
       #
       #   Re-tag is digest-level (imagetools create), so no rebuild and
       #   :latest is byte-identical to :staging-<sha> for this commit.
+      # Gate on supersede: a superseded older job must NOT move :latest backward
+      # to its older image (#2213 — 275383 promoted :latest → the older
+      # staging-7a72516 after a newer job had already shipped). :latest must only
+      # ever advance under the job that owns main's head.
       - name: Promote :latest to the verified prod image
-        if: ${{ steps.plan.outputs.enabled == 'true' }}
+        if: ${{ steps.plan.outputs.enabled == 'true' && steps.supersede.outputs.superseded != 'true' }}
         env:
           TENANT_IMAGE_NAME: ${{ env.TENANT_IMAGE_NAME }}
           STAGING_TENANT_IMAGE_NAME: ${{ env.STAGING_TENANT_IMAGE_NAME }}

.gitea/workflows/sync-providers-yaml.yml

@@ -60,6 +60,7 @@ concurrency:
   cancel-in-progress: true
 
 jobs:
+  # bp-required: pending #718 — soak-then-promote, not in BP yet.
   compare:
     name: Compare synced providers.yaml against controlplane canonical
     runs-on: ubuntu-latest

.gitea/workflows/verify-providers-gen.yml

@@ -67,6 +67,7 @@ concurrency:
   cancel-in-progress: true
 
 jobs:
+  # bp-required: pending #718 — soak-then-promote, not in BP yet.
   verify:
     name: Regenerate providers artifact and fail on drift
     runs-on: ubuntu-latest

canvas/Dockerfile

@@ -24,6 +24,17 @@ COPY --from=builder /app/public ./public
 EXPOSE 3000
 ENV PORT=3000
 ENV HOSTNAME="0.0.0.0"
+# Git SHA the image was built from, surfaced at /api/buildinfo so canvas
+# deploys are verifiable by the served SHA the same way workspace-server's
+# /buildinfo is (core#2235). Wired from `${{ github.sha }}` in
+# publish-canvas-image.yml. Server-only (not NEXT_PUBLIC_) — the route
+# handler reads it at runtime on the standalone Node server, so it stays
+# out of the client bundle. Set on the final stage (not the builder) so it
+# lives in the runtime env that force-dynamic reads per request. Default
+# "dev" matches the route + workspace-server sentinel: an unwired build
+# fails the SHA comparison closed instead of looking deployed.
+ARG BUILD_SHA=dev
+ENV BUILD_SHA=$BUILD_SHA
 # Non-root runtime — use addgroup/adduser without fixed GID/UID to avoid conflicts with base image
 RUN addgroup canvas 2>/dev/null || true && adduser -G canvas -s /bin/sh -D canvas 2>/dev/null || true
 USER canvas

canvas/e2e/staging-setup.ts

@@ -241,7 +241,14 @@ export default async function globalSetup(_config: FullConfig): Promise<void> {
       name: "E2E Canvas Test",
       runtime: "hermes",
       tier: 2,
-      model: "gpt-4o",
+      // Provider-registry SSOT (internal#718) registers ONLY Kimi models for
+      // the hermes runtime — `moonshot/kimi-k2.6` is the platform-managed
+      // entry (workspace-server/internal/providers/providers.yaml, hermes ->
+      // platform). The old `gpt-4o` was never a registered hermes model and
+      // now 422s UNREGISTERED_MODEL_FOR_RUNTIME (core#2225). This workspace
+      // defaults closed to platform_managed (see the boot-shape note below),
+      // so a platform-namespaced model id is the registry-correct choice.
+      model: "moonshot/kimi-k2.6",
     }),
   });
   if (ws.status >= 400 || !ws.body?.id) {
@@ -264,7 +271,7 @@ export default async function globalSetup(_config: FullConfig): Promise<void> {
   // CP proxy env absent) made a platform_managed workspace ABORT AT BOOT
   // with MISSING_PLATFORM_PROXY when MOLECULE_LLM_BASE_URL /
   // MOLECULE_LLM_USAGE_TOKEN are not present in the tenant's env. The
-  // canvas E2E creates a bare hermes/gpt-4o workspace, which defaults
+  // canvas E2E creates a bare hermes/moonshot platform workspace, which defaults
   // closed to platform_managed (workspace_provision.go:~1009), and the
   // staging tenant does not carry the CP proxy env — so the agent never
   // starts. Pre-#2162 this same workspace booted credential-less (the bug

canvas/src/app/api/buildinfo/__tests__/route.test.ts

@@ -1,12 +1,17 @@
 /**
  * Canvas /api/buildinfo — version-display endpoint mirroring
  * workspace-server's /buildinfo. Lets `curl <url>/api/buildinfo`
- * confirm which git SHA is live on a canvas deployment.
+ * confirm which git SHA is live on a canvas deployment (core#2235).
  */
 import { describe, it, expect, beforeEach, afterEach } from "vitest";
 import { GET } from "../route";
 
-const ENV_KEYS = ["VERCEL_GIT_COMMIT_SHA", "VERCEL_GIT_COMMIT_REF", "VERCEL_ENV"];
+const ENV_KEYS = [
+  "BUILD_SHA",
+  "VERCEL_GIT_COMMIT_SHA",
+  "VERCEL_GIT_COMMIT_REF",
+  "VERCEL_ENV",
+];
 
 describe("GET /api/buildinfo", () => {
   let saved: Record<string, string | undefined>;
@@ -23,13 +28,24 @@ describe("GET /api/buildinfo", () => {
     }
   });
 
-  it("returns dev sentinel when Vercel env vars are unset", async () => {
+  it("returns dev sentinel when no SHA source is set", async () => {
     const res = await GET();
     const body = await res.json();
     expect(body).toEqual({ git_sha: "dev", git_ref: "", vercel_env: "local" });
   });
 
-  it("reports the SHA Vercel injected at build time", async () => {
+  it("reports BUILD_SHA baked into the Docker image (fleet deploy path)", async () => {
+    // BUILD_SHA is the authoritative source for the ECR-image fleet deploy,
+    // which never runs on Vercel. It must win even when a Vercel var is also
+    // present in the environment.
+    process.env.BUILD_SHA = "deadbeefcafe";
+    process.env.VERCEL_GIT_COMMIT_SHA = "should-not-win";
+    const res = await GET();
+    const body = await res.json();
+    expect(body.git_sha).toBe("deadbeefcafe");
+  });
+
+  it("falls back to the SHA Vercel injected when BUILD_SHA is unset", async () => {
     process.env.VERCEL_GIT_COMMIT_SHA = "abc1234567890";
     process.env.VERCEL_GIT_COMMIT_REF = "main";
     process.env.VERCEL_ENV = "production";

canvas/src/app/api/buildinfo/route.ts

@@ -1,17 +1,36 @@
 import { NextResponse } from "next/server";
 
 // Mirror of workspace-server's GET /buildinfo (PR #2398). Lets a developer
-// confirm which git SHA is live on a canvas deployment with the same
-// `curl <url>/buildinfo` flow they use against tenant workspaces.
+// or the fleet redeploy workflow confirm which git SHA is live on a canvas
+// deployment with the same `curl <url>/api/buildinfo` flow used against
+// tenant workspaces (core#2235; cross-ref core#2226).
 //
-// Vercel injects VERCEL_GIT_COMMIT_SHA / _REF / VERCEL_ENV at build time
-// from the deploying commit; outside Vercel (local `next dev`, harness)
-// these are unset and the endpoint reports `git_sha: "dev"`. Same sentinel
-// the workspace-server uses pre-ldflags-injection so both surfaces speak
-// the same vocabulary.
+// SHA source, in priority order:
+//   1. BUILD_SHA — server-only env baked into the canvas Docker image at
+//      build time (Dockerfile `ARG BUILD_SHA` → `ENV BUILD_SHA`, wired
+//      from `${{ github.sha }}` in publish-canvas-image.yml). This is the
+//      authoritative source for the fleet's ECR-image deploy path, which
+//      does NOT run on Vercel. Read server-side here (App Router route
+//      handler runs on the standalone Node server, `output: "standalone"`),
+//      so it is intentionally NOT a NEXT_PUBLIC_ var — keeping it out of
+//      the client bundle.
+//   2. VERCEL_GIT_COMMIT_SHA — Vercel injects this at build time when the
+//      canvas is deployed via Vercel rather than the Docker image.
+//   3. "dev" — local `next dev` / test harness, where neither is set. Same
+//      sentinel workspace-server uses pre-ldflags-injection, so both
+//      surfaces speak the same vocabulary and an unconfigured deploy
+//      fails the SHA comparison closed instead of round-tripping "".
+//
+// force-dynamic so the response is evaluated at request time against the
+// runtime env of the standalone server (where ENV BUILD_SHA lives), not
+// frozen into a static asset at `next build`.
+export const dynamic = "force-dynamic";
+
 export async function GET() {
+  const sha =
+    process.env.BUILD_SHA ?? process.env.VERCEL_GIT_COMMIT_SHA ?? "dev";
   return NextResponse.json({
-    git_sha: process.env.VERCEL_GIT_COMMIT_SHA ?? "dev",
+    git_sha: sha,
     git_ref: process.env.VERCEL_GIT_COMMIT_REF ?? "",
     vercel_env: process.env.VERCEL_ENV ?? "local",
   });

canvas/src/components/CreateWorkspaceDialog.tsx

@@ -10,6 +10,7 @@ import {
   buildProviderCatalog,
   buildProviderCatalogFromRegistry,
   findProviderForModel,
+  isPlatformManagedProvider,
   type SelectorModel,
   type SelectorValue,
   type RegistryProvider,
@@ -290,7 +291,15 @@ export function CreateWorkspaceButton() {
       setError("Model is required");
       return;
     }
-    if (!isExternal && selectedLLMProvider?.envVars.length && !llmSecret.trim()) {
+    // Platform-managed providers need NO user credential — the platform injects
+    // its own usage token (MOLECULE_LLM_USAGE_TOKEN = tenant admin_token) at
+    // provision time. Only BYOK providers require a user-supplied key. (#2245)
+    if (
+      !isExternal &&
+      !isPlatformManagedProvider(selectedLLMProvider) &&
+      selectedLLMProvider?.envVars.length &&
+      !llmSecret.trim()
+    ) {
       setError("Provider credential is required");
       return;
     }
@@ -325,7 +334,11 @@ export function CreateWorkspaceButton() {
           ? {
               model: llmSelection.model.trim(),
               llm_provider: nativeProvider.vendor,
-              ...(nativeProvider.envVars.length > 0
+              // Only BYOK providers carry a user secret. For platform-managed
+              // the token is provisioner-injected; sending an (empty) secret
+              // here would clobber it — so omit it entirely. (#2245)
+              ...(nativeProvider.envVars.length > 0 &&
+              !isPlatformManagedProvider(nativeProvider)
                 ? { secrets: { [nativeProvider.envVars[0]]: llmSecret.trim() } }
                 : {}),
             }
@@ -521,20 +534,26 @@ export function CreateWorkspaceButton() {
                   idPrefix="create-workspace-llm"
                   variant="stack"
                 />
-                {selectedLLMProvider.envVars.length > 0 && (
-                  <div>
-                    <label htmlFor="llm-secret-input" className="text-[11px] text-ink-mid block mb-1">
-                      {selectedLLMProvider.envVars[0]}
-                    </label>
-                    <input
-                      id="llm-secret-input"
-                      type="password"
-                      value={llmSecret}
-                      onChange={(e) => setLLMSecret(e.target.value)}
-                      autoComplete="off"
-                      className="w-full bg-surface-card/60 border border-line/50 rounded-lg px-3 py-2 text-sm text-ink placeholder-ink-soft focus:outline-none focus:border-accent/60 focus:ring-1 focus:ring-accent/20 transition-colors font-mono"
-                    />
+                {isPlatformManagedProvider(selectedLLMProvider) ? (
+                  <div className="text-[11px] text-ink-soft">
+                    Platform-managed — no API key required.
                   </div>
+                ) : (
+                  selectedLLMProvider.envVars.length > 0 && (
+                    <div>
+                      <label htmlFor="llm-secret-input" className="text-[11px] text-ink-mid block mb-1">
+                        {selectedLLMProvider.envVars[0]}
+                      </label>
+                      <input
+                        id="llm-secret-input"
+                        type="password"
+                        value={llmSecret}
+                        onChange={(e) => setLLMSecret(e.target.value)}
+                        autoComplete="off"
+                        className="w-full bg-surface-card/60 border border-line/50 rounded-lg px-3 py-2 text-sm text-ink placeholder-ink-soft focus:outline-none focus:border-accent/60 focus:ring-1 focus:ring-accent/20 transition-colors font-mono"
+                      />
+                    </div>
+                  )
                 )}
               </div>
             )}

canvas/src/components/ProviderModelSelector.tsx

@@ -55,6 +55,21 @@ export interface ProviderEntry {
   billingMode?: "platform_managed" | "byok";
 }
 
+/** A provider is "platform-managed" when the Molecule platform proxies the LLM
+ *  call and injects its own usage credential — the tenant admin_token, surfaced
+ *  to the workspace as MOLECULE_LLM_USAGE_TOKEN by the CP provisioner
+ *  (controlplane ec2.go: `MOLECULE_LLM_USAGE_TOKEN="$ADMIN_TOKEN"`). The user
+ *  supplies NO key for these: the credential is internal plumbing, not a user
+ *  input. Detected by vendor==="platform" (the platform proxy provider, which
+ *  declares MOLECULE_LLM_USAGE_TOKEN in its AuthEnv) OR
+ *  billingMode==="platform_managed" (registry-backed, internal#718 P3). BYOK
+ *  providers return false and DO require a user-supplied credential. */
+export function isPlatformManagedProvider(
+  p?: Pick<ProviderEntry, "vendor" | "billingMode"> | null,
+): boolean {
+  return p?.vendor === "platform" || p?.billingMode === "platform_managed";
+}
+
 /** RegistryProvider mirrors one entry of GET /templates `registry_providers`
  *  (workspace-server registryProviderView): the registry's native provider for
  *  a runtime, with its display label, auth-env NAMES, and billing mode. This is

canvas/src/components/__tests__/CreateWorkspaceDialog.test.tsx

@@ -2,6 +2,7 @@
 import { describe, it, expect, vi, beforeEach, afterEach } from "vitest";
 import { render, screen, fireEvent, waitFor, cleanup } from "@testing-library/react";
 import { CreateWorkspaceButton } from "../CreateWorkspaceDialog";
+import { isPlatformManagedProvider } from "../ProviderModelSelector";
 
 vi.mock("@/lib/api", () => ({
   api: {
@@ -65,6 +66,34 @@ const SAMPLE_TEMPLATES = [
       { id: "moonshot/kimi-k2.6", name: "Kimi K2.6", provider: "platform", required_env: [] },
     ],
   },
+  // #2245 fixtures. The real registry `platform` provider declares
+  // MOLECULE_LLM_USAGE_TOKEN in its auth_env — the default mock above masks the
+  // bug by using required_env:[]. This template gives the platform provider a
+  // non-empty auth env (matching production) so the credential-suppression
+  // logic is actually exercised.
+  {
+    id: "platform-managed-test",
+    name: "Platform Managed Test",
+    runtime: "claude-code",
+    model: "moonshot/kimi-k2.6",
+    providers: ["platform", "minimax"],
+    models: [
+      { id: "moonshot/kimi-k2.6", name: "Kimi K2.6", provider: "platform", required_env: ["MOLECULE_LLM_USAGE_TOKEN"] },
+      { id: "MiniMax-M2.7", name: "MiniMax M2.7", required_env: ["MINIMAX_API_KEY"] },
+    ],
+  },
+  // BYOK-only template (no platform provider) — the credential requirement
+  // MUST still hold for these (no-regression guard).
+  {
+    id: "byok-only-test",
+    name: "BYOK Only Test",
+    runtime: "claude-code",
+    model: "openai/gpt-4o",
+    providers: ["openai"],
+    models: [
+      { id: "openai/gpt-4o", name: "GPT-4o", required_env: ["OPENAI_API_KEY"] },
+    ],
+  },
 ];
 
 beforeEach(() => {
@@ -498,6 +527,25 @@ const REGISTRY_TEMPLATE = {
   ],
 };
 
+// Registry-backed platform provider WITH a non-empty auth_env — this matches
+// the PRODUCTION provider view, which ships the raw AuthEnv
+// ([MOLECULE_LLM_USAGE_TOKEN]). REGISTRY_TEMPLATE above uses auth_env:[] so it
+// never exercises suppression; this one drives the billingMode==="platform_
+// managed" branch end-to-end through buildProviderCatalogFromRegistry. (#2245)
+const REGISTRY_TEMPLATE_PLATFORM_AUTHENV = {
+  ...REGISTRY_TEMPLATE,
+  registry_providers: [
+    {
+      name: "platform",
+      display_name: "Platform",
+      auth_env: ["MOLECULE_LLM_USAGE_TOKEN"],
+      billing_mode: "platform_managed",
+    },
+    { name: "minimax", display_name: "MiniMax", auth_env: ["MINIMAX_API_KEY"], billing_mode: "byok" },
+    { name: "anthropic", display_name: "Anthropic API", auth_env: ["ANTHROPIC_API_KEY"], billing_mode: "byok" },
+  ],
+};
+
 describe("CreateWorkspaceDialog — registry-backed provider catalog (RFC#340 Fix C)", () => {
   beforeEach(() => {
     mockGet.mockImplementation(async (url: string) => {
@@ -574,6 +622,41 @@ describe("CreateWorkspaceDialog — registry-backed provider catalog (RFC#340 Fi
     expect(body.llm_provider).toBe("minimax");
     expect(body.secrets).toEqual({ MINIMAX_API_KEY: "sk-minimax-test" });
   });
+
+  it("suppresses the credential for a registry-backed platform provider that declares an auth_env — billingMode path (#2245)", async () => {
+    // Override the default REGISTRY_TEMPLATE (auth_env:[]) with the production-
+    // shaped one whose platform provider declares MOLECULE_LLM_USAGE_TOKEN.
+    mockGet.mockImplementation(async (url: string) => {
+      if (url === "/templates") {
+        // eslint-disable-next-line @typescript-eslint/no-explicit-any
+        return [REGISTRY_TEMPLATE_PLATFORM_AUTHENV] as any;
+      }
+      // eslint-disable-next-line @typescript-eslint/no-explicit-any
+      return SAMPLE_WORKSPACES as any;
+    });
+    await openDialog();
+    fireEvent.change(screen.getByPlaceholderText("e.g. SEO Agent"), {
+      target: { value: "Registry Platform Agent" },
+    });
+    // Platform is the default bucket; even with a non-empty auth_env the key
+    // field must NOT render (suppressed via billingMode==="platform_managed").
+    await waitFor(() => {
+      const sel = document.querySelector("[data-testid='provider-select']") as HTMLSelectElement;
+      expect(sel?.value).toBe("registry|platform");
+    });
+    expect(screen.getByText("Platform-managed — no API key required.")).toBeTruthy();
+    expect(document.getElementById("llm-secret-input")).toBeNull();
+
+    const createBtn = screen.getAllByRole("button").find((b) => b.textContent === "Create");
+    fireEvent.click(createBtn!);
+
+    await waitFor(() => expect(mockPost).toHaveBeenCalled());
+    expect(screen.queryByText("Provider credential is required")).toBeNull();
+    const body = mockPost.mock.calls[0][1] as Record<string, unknown>;
+    expect(body.llm_provider).toBe("platform");
+    // The provisioner-injected MOLECULE_LLM_USAGE_TOKEN must NOT be clobbered.
+    expect(body.secrets).toBeUndefined();
+  });
 });
 
 // ---------------------------------------------------------------------------
@@ -657,3 +740,70 @@ describe("CreateWorkspaceDialog — budget_limit field", () => {
     expect(budgetInput.value).toBe("");
   });
 });
+
+describe("CreateWorkspaceDialog — platform-managed credential suppression (#2245)", () => {
+  describe("isPlatformManagedProvider", () => {
+    it("is true for the platform proxy vendor", () => {
+      expect(isPlatformManagedProvider({ vendor: "platform" })).toBe(true);
+    });
+    it("is true for a registry billingMode of platform_managed", () => {
+      expect(
+        isPlatformManagedProvider({ vendor: "minimax", billingMode: "platform_managed" }),
+      ).toBe(true);
+    });
+    it("is false for a BYOK provider", () => {
+      expect(isPlatformManagedProvider({ vendor: "anthropic", billingMode: "byok" })).toBe(false);
+      expect(isPlatformManagedProvider({ vendor: "minimax" })).toBe(false);
+    });
+    it("is false for null/undefined", () => {
+      expect(isPlatformManagedProvider(null)).toBe(false);
+      expect(isPlatformManagedProvider(undefined)).toBe(false);
+    });
+  });
+
+  it("platform-managed provider with a declared auth env requires NO credential, hides the key field, and sends NO secret", async () => {
+    await openDialog();
+    await setTemplate("platform-managed-test");
+    fireEvent.change(screen.getByPlaceholderText("e.g. SEO Agent"), {
+      target: { value: "Platform Agent" },
+    });
+
+    // The credential input must NOT render for platform-managed; a "no key
+    // required" note appears instead.
+    await waitFor(() =>
+      expect(screen.getByText("Platform-managed — no API key required.")).toBeTruthy(),
+    );
+    expect(screen.queryByLabelText("MOLECULE_LLM_USAGE_TOKEN")).toBeNull();
+
+    const createBtn = screen.getAllByRole("button").find((b) => b.textContent === "Create");
+    fireEvent.click(createBtn!);
+
+    await waitFor(() => expect(mockPost).toHaveBeenCalled());
+    // No validation error, and the provisioner-injected token is NOT clobbered
+    // by an empty secret.
+    expect(screen.queryByText("Provider credential is required")).toBeNull();
+    const body = mockPost.mock.calls[0][1] as Record<string, unknown>;
+    expect(body.llm_provider).toBe("platform");
+    expect(body.secrets).toBeUndefined();
+  });
+
+  it("BYOK provider still requires a credential and renders the key field (no-regression)", async () => {
+    await openDialog();
+    await setTemplate("byok-only-test");
+    fireEvent.change(screen.getByPlaceholderText("e.g. SEO Agent"), {
+      target: { value: "BYOK Agent" },
+    });
+
+    // The credential field IS rendered for BYOK...
+    await waitFor(() => expect(screen.getByLabelText("OPENAI_API_KEY")).toBeTruthy());
+
+    const createBtn = screen.getAllByRole("button").find((b) => b.textContent === "Create");
+    fireEvent.click(createBtn!);
+
+    // ...and create stays blocked until it's filled.
+    await waitFor(() =>
+      expect(screen.getByText("Provider credential is required")).toBeTruthy(),
+    );
+    expect(mockPost).not.toHaveBeenCalled();
+  });
+});

canvas/src/components/tabs/DisplayTab.tsx

@@ -1,6 +1,6 @@
 "use client";
 
-import { useEffect, useRef, useState } from "react";
+import { useCallback, useEffect, useRef, useState } from "react";
 import { api } from "@/lib/api";
 import type RFB from "@novnc/novnc";
 
@@ -33,6 +33,11 @@ export function DisplayTab({ workspaceId }: Props) {
   const [controlBusy, setControlBusy] = useState(false);
   const [sessionUrl, setSessionUrl] = useState<string | null>(null);
   const requestGeneration = useRef(0);
+  // Freshest signed session URL (token bound to the lease's expires_at). The
+  // renewal timer keeps this current WITHOUT swapping the live stream's
+  // sessionUrl (which would needlessly reconnect the desktop); the stream uses
+  // it only when it has to reconnect after an unclean drop.
+  const latestSessionUrlRef = useRef<string | null>(null);
 
   useEffect(() => {
     const generation = requestGeneration.current + 1;
@@ -41,6 +46,7 @@ export function DisplayTab({ workspaceId }: Props) {
     setStatus(null);
     setControl(null);
     setSessionUrl(null);
+    latestSessionUrlRef.current = null;
     setError(null);
     setControlError(null);
     setControlBusy(false);
@@ -69,6 +75,41 @@ export function DisplayTab({ workspaceId }: Props) {
     };
   }, [workspaceId]);
 
+  // Acquire (or re-acquire) the display-control lease as the current holder.
+  // Re-acquiring extends the 300s server-side lock AND returns a freshly-signed
+  // session URL (token bound to the new expires_at). Used both to renew the
+  // lease on a timer and to mint a non-stale token for each reconnect — a
+  // cached URL can be past its ~300s expiry, which would make a reconnect 401.
+  const reacquireSession = useCallback(async (): Promise<string | null> => {
+    const generation = requestGeneration.current;
+    try {
+      const next = await api.post<DisplayControlStatus>(
+        `/workspaces/${workspaceId}/display/control/acquire`,
+        { controller: "user", ttl_seconds: 300 },
+      );
+      if (requestGeneration.current !== generation) return null;
+      setControl(next);
+      if (next.session_url) latestSessionUrlRef.current = next.session_url;
+      return next.session_url ?? null;
+    } catch {
+      // Transient failure, or another holder took over: the live stream keeps
+      // running on its existing connection; a reconnect re-evaluates control.
+      return null;
+    }
+  }, [workspaceId]);
+
+  // Renew the lease while we hold it. The lock is a 300s lease with no
+  // server-side auto-renewal, so without this the control (and the session
+  // token) silently expire mid-session — the user appears "kicked" every ~5
+  // minutes. We renew well inside the TTL and do not touch the live stream.
+  useEffect(() => {
+    if (!sessionUrl) return;
+    const timer = setInterval(() => {
+      void reacquireSession();
+    }, 120_000);
+    return () => clearInterval(timer);
+  }, [sessionUrl, reacquireSession]);
+
   const acquireControl = async () => {
     const generation = requestGeneration.current;
     const controlPath = `/workspaces/${workspaceId}/display/control`;
@@ -82,6 +123,7 @@ export function DisplayTab({ workspaceId }: Props) {
       if (requestGeneration.current !== generation) return;
       setControl(next);
       setSessionUrl(next.session_url || null);
+      latestSessionUrlRef.current = next.session_url || null;
     } catch (err) {
       if (requestGeneration.current !== generation) return;
       setControlError("Failed to take control");
@@ -108,6 +150,7 @@ export function DisplayTab({ workspaceId }: Props) {
       if (requestGeneration.current !== generation) return;
       setControl(next);
       setSessionUrl(null);
+      latestSessionUrlRef.current = null;
     } catch (err) {
       if (requestGeneration.current !== generation) return;
       setControlError("Failed to release control");
@@ -235,7 +278,11 @@ export function DisplayTab({ workspaceId }: Props) {
         />
       </div>
       {sessionUrl ? (
-        <DesktopStream sessionUrl={sessionUrl} />
+        <DesktopStream
+          sessionUrl={sessionUrl}
+          latestSessionUrlRef={latestSessionUrlRef}
+          reacquireSession={reacquireSession}
+        />
       ) : (
         <div className="flex flex-1 items-center justify-center p-8 text-center">
           <div>
@@ -311,7 +358,15 @@ function DisplayControlBar({
   );
 }
 
-function DesktopStream({ sessionUrl }: { sessionUrl: string }) {
+function DesktopStream({
+  sessionUrl,
+  latestSessionUrlRef,
+  reacquireSession,
+}: {
+  sessionUrl: string;
+  latestSessionUrlRef: { current: string | null };
+  reacquireSession: () => Promise<string | null>;
+}) {
   const containerRef = useRef<HTMLDivElement | null>(null);
   const rfbRef = useRef<RFB | null>(null);
   const [streamError, setStreamError] = useState<string | null>(null);
@@ -329,20 +384,37 @@ function DesktopStream({ sessionUrl }: { sessionUrl: string }) {
       clipboardTimer = setTimeout(() => setClipboardStatus(null), 2500);
     };
 
-    async function connect() {
+    let attempts = 0;
+    let retryTimer: ReturnType<typeof setTimeout> | null = null;
+    const maxAttempts = 10;
+
+    async function connect(reacquire = false) {
       setStreamError(null);
       try {
+        // On a reconnect, mint a fresh lease + token first — the original token
+        // is only ~300s, so a cached URL can be expired and would 401. The
+        // initial connect already holds a fresh token from acquireControl.
+        if (reacquire) await reacquireSession();
         const mod = await import("@novnc/novnc");
         if (cancelled || !containerRef.current) return;
-        const stream = displayWebSocketConnection(sessionUrl);
+        const stream = displayWebSocketConnection(latestSessionUrlRef.current || sessionUrl);
         rfb = new mod.default(containerRef.current, stream.url, {
           wsProtocols: ["binary", `molecule-display-token.${stream.token}`],
         });
         rfbRef.current = rfb;
         rfb.scaleViewport = true;
-        rfb.resizeSession = true;
+        // Do NOT request a server-side resize: the workspace display runs a
+        // fixed Xorg modeline and x11vnc rejects SetDesktopSize ("Resize is
+        // administratively prohibited"), which spams the console on every
+        // (re)connect. scaleViewport already fits the fixed framebuffer to the
+        // container client-side, so we don't need the server to resize.
+        rfb.resizeSession = false;
         rfb.focusOnClick = true;
         rfb.focus({ preventScroll: true });
+        rfb.addEventListener("connect", () => {
+          attempts = 0;
+          if (!cancelled) setStreamError(null);
+        });
         rfb.addEventListener("clipboard", (event: Event) => {
           const text = (event as CustomEvent<{ text?: string }>).detail?.text ?? "";
           if (!text) return;
@@ -353,7 +425,20 @@ function DesktopStream({ sessionUrl }: { sessionUrl: string }) {
         });
         rfb.addEventListener("disconnect", (event: Event) => {
           const detail = (event as CustomEvent<{ clean?: boolean }>).detail;
-          if (!cancelled && !detail?.clean) setStreamError("Desktop stream disconnected.");
+          rfbRef.current = null;
+          if (cancelled || detail?.clean) return;
+          // Auto-reconnect after an unclean drop (idle/network blip, brief
+          // agent hiccup); bounded backoff so a genuinely-dead session still
+          // surfaces an error instead of looping forever.
+          if (attempts < maxAttempts) {
+            attempts += 1;
+            setStreamError(`Reconnecting to desktop… (attempt ${attempts})`);
+            retryTimer = setTimeout(() => {
+              if (!cancelled) void connect(true);
+            }, Math.min(1000 * attempts, 5000));
+          } else {
+            setStreamError("Desktop stream disconnected.");
+          }
         });
       } catch {
         if (!cancelled) setStreamError("Desktop stream could not be opened.");
@@ -363,11 +448,12 @@ function DesktopStream({ sessionUrl }: { sessionUrl: string }) {
     connect();
     return () => {
       cancelled = true;
+      if (retryTimer) clearTimeout(retryTimer);
       if (clipboardTimer) clearTimeout(clipboardTimer);
       rfbRef.current = null;
       rfb?.disconnect();
     };
-  }, [sessionUrl]);
+  }, [sessionUrl, reacquireSession, latestSessionUrlRef]);
 
   useEffect(() => {
     const onPaste = (event: ClipboardEvent) => {

canvas/src/components/tabs/__tests__/DisplayTab.test.tsx

@@ -2,12 +2,13 @@
 import { describe, it, expect, vi, beforeEach } from "vitest";
 import { cleanup, fireEvent, render, screen, waitFor } from "@testing-library/react";
 
-const { mockGet, mockPost, mockRFBConstructor, mockRFBClipboardPasteFrom, mockRFBFocus } = vi.hoisted(() => ({
+const { mockGet, mockPost, mockRFBConstructor, mockRFBClipboardPasteFrom, mockRFBFocus, rfbInstances } = vi.hoisted(() => ({
   mockGet: vi.fn(),
   mockPost: vi.fn(),
   mockRFBConstructor: vi.fn(),
   mockRFBClipboardPasteFrom: vi.fn(),
   mockRFBFocus: vi.fn(),
+  rfbInstances: [] as EventTarget[],
 }));
 
 vi.mock("@/lib/api", () => ({
@@ -31,6 +32,7 @@ vi.mock("@novnc/novnc", () => ({
       this.url = url;
       this.options = options;
       mockRFBConstructor(target, url, options);
+      rfbInstances.push(this);
     }
     clipboardPasteFrom(text: string) {
       mockRFBClipboardPasteFrom(text);
@@ -52,6 +54,7 @@ describe("DisplayTab", () => {
     mockRFBConstructor.mockReset();
     mockRFBClipboardPasteFrom.mockReset();
     mockRFBFocus.mockReset();
+    rfbInstances.length = 0;
   });
 
   it("renders unavailable state for non-display workspaces", async () => {
@@ -400,6 +403,62 @@ describe("DisplayTab", () => {
     });
     expect(screen.getByRole("button", { name: "Take control" })).toBeTruthy();
   });
+
+  it("auto-reconnects the desktop stream after an unclean disconnect but not a clean one", async () => {
+    mockGet
+      .mockResolvedValueOnce({
+        available: true,
+        mode: "desktop-control",
+        protocol: "novnc",
+        width: 1920,
+        height: 1080,
+      })
+      .mockResolvedValueOnce({ controller: "none" });
+    // Initial acquire returns token "signed"; the reconnect re-acquire mints a
+    // FRESH token "signed2" (the lock/token is only ~300s — reconnecting with a
+    // cached, possibly-expired token would 401 and never recover).
+    mockPost
+      .mockResolvedValueOnce({
+        controller: "user",
+        controlled_by: "admin-token",
+        expires_at: "2026-05-23T08:48:27Z",
+        session_url: "/workspaces/ws-display/display/session/websockify#token=signed",
+      })
+      .mockResolvedValue({
+        controller: "user",
+        controlled_by: "admin-token",
+        expires_at: "2026-05-23T08:53:27Z",
+        session_url: "/workspaces/ws-display/display/session/websockify#token=signed2",
+      });
+
+    render(<DisplayTab workspaceId="ws-display" />);
+    await waitFor(() => {
+      expect(screen.getByRole("button", { name: "Take control" })).toBeTruthy();
+    });
+    fireEvent.click(screen.getByRole("button", { name: "Take control" }));
+    await waitFor(() => {
+      expect(rfbInstances.length).toBe(1);
+    });
+    expect(mockRFBConstructor.mock.calls[0][2].wsProtocols).toContain("molecule-display-token.signed");
+
+    // An idle/network drop closes the websocket uncleanly. The client must
+    // re-acquire a fresh token and reconnect instead of giving up — this is the
+    // "disconnects every ~5 min and stays dead" report.
+    rfbInstances[0].dispatchEvent(new CustomEvent("disconnect", { detail: { clean: false } }));
+    await waitFor(
+      () => {
+        expect(rfbInstances.length).toBe(2);
+      },
+      { timeout: 3000 },
+    );
+    // Reconnect dialed with the FRESH token, not the stale original.
+    expect(mockRFBConstructor.mock.calls[1][2].wsProtocols).toContain("molecule-display-token.signed2");
+
+    // A clean disconnect (the user released control) must NOT reconnect.
+    rfbInstances[1].dispatchEvent(new CustomEvent("disconnect", { detail: { clean: true } }));
+    await new Promise((resolve) => setTimeout(resolve, 1100));
+    expect(rfbInstances.length).toBe(2);
+  });
 });
 
 function deferred<T>() {

docker-compose.yml

@@ -159,15 +159,28 @@ services:
 
   # --- Canvas ---
   canvas:
-    # The publish-canvas-image CI workflow pushes a fresh image to GHCR on
-    # every canvas/** merge to main. To update the running container:
-    #   docker compose pull canvas && docker compose up -d canvas
-    # First-time local setup or testing unreleased changes — build from source:
-    #   docker compose build canvas && docker compose up -d canvas
+    # The publish-canvas-image CI workflow runs an ORDERED deploy (core#2226):
+    # build → push :staging-<sha> + :staging-latest → (after green main CI)
+    # re-point :latest to the verified :staging-<sha> by digest. So both tags
+    # below resolve to a CI-green, reproducible build, never a raw/red one.
+    #
+    # Reproducible deploy: pin CANVAS_IMAGE_TAG to the immutable per-commit tag
+    # the ordered deploy produced, e.g.
+    #   CANVAS_IMAGE_TAG=staging-<sha> docker compose pull canvas && docker compose up -d canvas
+    # This makes a tenant/host deploy reproducible (resolves the standing
+    # `TODO: pin canvas ECR image digest`). Unset it and the default `latest`
+    # is the prod-blessed tag the ordered deploy keeps pointed at the last
+    # green build — still deterministic vs. the old raw `:latest`.
+    #
+    # To pin by content digest instead of tag (fully immutable):
+    #   aws ecr describe-images --repository-name molecule-ai/canvas \
+    #     --image-tags staging-<sha> --region us-east-2 \
+    #     --query 'imageDetails[0].imageDigest' --output text
+    # then set CANVAS_IMAGE_TAG=staging-<sha>@<digest> (compose passes it through).
+    #
     # Note: ECR images require AWS auth — `aws ecr get-login-password --region us-east-2 | docker login --username AWS --password-stdin 153263036946.dkr.ecr.us-east-2.amazonaws.com` before pull.
-    # Digest-pin requires: aws ecr describe-images --repository-name molecule-ai/canvas --image-tags latest --query 'imageDetails[0].imageDigest'
-    # TODO: pin canvas ECR image digest once AWS creds are available in CI.
-    image: 153263036946.dkr.ecr.us-east-2.amazonaws.com/molecule-ai/canvas:latest
+    # Local dev keeps working via the `build:` context below (docker compose build canvas).
+    image: 153263036946.dkr.ecr.us-east-2.amazonaws.com/molecule-ai/canvas:${CANVAS_IMAGE_TAG:-latest}
     build:
       context: ./canvas
       dockerfile: Dockerfile
@@ -175,6 +188,10 @@ services:
         NEXT_PUBLIC_PLATFORM_URL: ${NEXT_PUBLIC_PLATFORM_URL:-http://localhost:${PLATFORM_PUBLISH_PORT:-8080}}
         NEXT_PUBLIC_WS_URL: ${NEXT_PUBLIC_WS_URL:-ws://localhost:${PLATFORM_PUBLISH_PORT:-8080}/ws}
         NEXT_PUBLIC_ADMIN_TOKEN: ${ADMIN_TOKEN:-}
+        # SHA surfaced at /api/buildinfo (core#2235). CI passes the real merge
+        # SHA via the publish-canvas-image workflow build-args; local compose
+        # builds default to "dev" (the route's unwired sentinel).
+        BUILD_SHA: ${BUILD_SHA:-dev}
     depends_on:
       platform:
         condition: service_healthy

manifest.json

@@ -28,9 +28,7 @@
     {"name": "claude-code-default", "repo": "molecule-ai/molecule-ai-workspace-template-claude-code", "ref": "main"},
     {"name": "hermes", "repo": "molecule-ai/molecule-ai-workspace-template-hermes", "ref": "main"},
     {"name": "openclaw", "repo": "molecule-ai/molecule-ai-workspace-template-openclaw", "ref": "main"},
-    {"name": "codex", "repo": "molecule-ai/molecule-ai-workspace-template-codex", "ref": "main"},
-    {"name": "google-adk", "repo": "molecule-ai/molecule-ai-workspace-template-google-adk", "ref": "main"},
-    {"name": "seo-agent", "repo": "molecule-ai/molecule-ai-workspace-template-seo-agent", "ref": "main"}
+    {"name": "codex", "repo": "molecule-ai/molecule-ai-workspace-template-codex", "ref": "main"}
   ],
   "org_templates": [
     {"name": "molecule-dev", "repo": "molecule-ai/molecule-ai-org-template-molecule-dev", "ref": "main"},

scripts/check-manifest-repos-exist.sh

@@ -0,0 +1,76 @@
+#!/usr/bin/env bash
+# check-manifest-repos-exist.sh — fail-fast guard: verify every repo listed in
+# manifest.json actually exists on Gitea before the expensive clone step runs.
+#
+# WHY: deleting an org-template/workspace-template repo that is still listed in
+# manifest.json breaks clone-manifest.sh with a generic git 404 error. The
+# failure is deep in the publish-workspace-server-image workflow and looks like
+# a transient network issue, wasting debug time. This script surfaces the
+# problem immediately with a per-entry ::error:: annotation naming the missing
+# repo (issue #2192).
+#
+# Usage:
+#   ./scripts/check-manifest-repos-exist.sh <manifest.json>
+#
+# Exit:
+#   0  all repos exist
+#   1  one or more repos 404 (printed to stderr)
+#   2  bad usage / missing inputs
+
+set -euo pipefail
+
+MANIFEST="${1:-manifest.json}"
+GITEA_API="${GITEA_API:-https://git.moleculesai.app/api/v1/repos}"
+
+if [ ! -f "$MANIFEST" ]; then
+    echo "::error::manifest not found: $MANIFEST" >&2
+    exit 2
+fi
+
+# Strip JSON5-style // comments before parsing (same as clone-manifest.sh)
+_strip_comments() {
+    sed 's/^[[:space:]]*\/\/.*//' "$MANIFEST"
+}
+
+MANIFEST_JSON="$(_strip_comments)"
+
+MISSING=0
+TOTAL=0
+
+# Categories to check — must match clone-manifest.sh categories
+check_category() {
+    local category="$1"
+    local count
+    count=$(echo "$MANIFEST_JSON" | jq -r ".${category} | length")
+
+    local i=0
+    while [ "$i" -lt "$count" ]; do
+        local name repo
+        name=$(echo "$MANIFEST_JSON" | jq -r ".${category}[$i].name")
+        repo=$(echo "$MANIFEST_JSON" | jq -r ".${category}[$i].repo")
+        TOTAL=$((TOTAL + 1))
+
+        # Check repo existence via Gitea API (public endpoint, no auth needed)
+        http_code=$(curl -sS -o /dev/null -w '%{http_code}' --max-time 10 "${GITEA_API}/${repo}" 2>/dev/null || true)
+
+        if [ "$http_code" != "200" ]; then
+            echo "::error::manifest.json ${category} entry '${name}' → repo '${repo}' returned HTTP ${http_code} (expected 200). Delete the manifest entry BEFORE deleting the repo." >&2
+            MISSING=$((MISSING + 1))
+        fi
+
+        i=$((i + 1))
+    done
+}
+
+echo "==> Checking manifest repo existence against ${GITEA_API} ..."
+check_category "plugins"
+check_category "workspace_templates"
+check_category "org_templates"
+
+if [ "$MISSING" -gt 0 ]; then
+    echo "::error::${MISSING}/${TOTAL} manifest entries are missing — fix manifest.json before publishing." >&2
+    exit 1
+fi
+
+echo "✓ All ${TOTAL} manifest entries resolved (HTTP 200)."
+exit 0

tests/e2e/test_peer_visibility_mcp_staging.sh

@@ -234,9 +234,30 @@ elif [ -n "${E2E_OPENAI_API_KEY:-}" ]; then
   SECRETS_JSON=$(python3 -c "import json,os;k=os.environ['E2E_OPENAI_API_KEY'];print(json.dumps({'OPENAI_API_KEY':k,'OPENAI_BASE_URL':'https://api.openai.com/v1','MODEL_PROVIDER':'openai:gpt-4o','HERMES_INFERENCE_PROVIDER':'custom','HERMES_CUSTOM_BASE_URL':'https://api.openai.com/v1','HERMES_CUSTOM_API_KEY':k,'HERMES_CUSTOM_API_MODE':'chat_completions'}))")
 fi
 
+# Workspace-create now enforces the MODEL_REQUIRED contract: there is NO
+# platform-side default model for a runtime (feedback_workspace_model_required_
+# no_platform_default). Every create MUST carry an explicit `model`, or the CP
+# rejects it with MODEL_REQUIRED before this gate's peer-visibility assertion
+# can run. We pick a PLATFORM-MANAGED id (Molecule owns billing — no tenant key
+# needed; this gate only needs the workspace to boot + list peers, not heavy
+# LLM work), validated against the controlplane providers SSOT
+# (internal/providers/providers.yaml runtimes.<rt>.providers[platform].models):
+#   claude-code → anthropic/claude-sonnet-4-6   (platform claude model)
+#   hermes/openclaw → moonshot/kimi-k2.6         (their only platform family)
+# E2E_MODEL_SLUG overrides for operator-dispatched runs.
+pv_platform_model_for_runtime() {
+  if [ -n "${E2E_MODEL_SLUG:-}" ]; then printf '%s' "$E2E_MODEL_SLUG"; return 0; fi
+  case "$1" in
+    claude-code) printf 'anthropic/claude-sonnet-4-6' ;;
+    hermes|openclaw) printf 'moonshot/kimi-k2.6' ;;
+    *) printf 'moonshot/kimi-k2.6' ;;
+  esac
+}
+
 log "4/6 provisioning parent (claude-code) + one sibling per runtime under test..."
+PARENT_MODEL=$(pv_platform_model_for_runtime claude-code)
 P_RESP=$(tenant_call POST /workspaces \
-  -d "{\"name\":\"pv-parent\",\"runtime\":\"claude-code\",\"tier\":3,\"secrets\":$SECRETS_JSON}")
+  -d "{\"name\":\"pv-parent\",\"runtime\":\"claude-code\",\"model\":\"$PARENT_MODEL\",\"tier\":3,\"secrets\":$SECRETS_JSON}")
 PARENT_ID=$(echo "$P_RESP" | python3 -c "import sys,json; print(json.load(sys.stdin).get('id',''))" 2>/dev/null)
 [ -n "$PARENT_ID" ] || fail "parent create failed: $(echo "$P_RESP" | head -c 300)"
 log "    PARENT_ID=$PARENT_ID"
@@ -245,8 +266,9 @@ log "    PARENT_ID=$PARENT_ID"
 declare -A WS_IDS WS_TOKENS
 ALL_WS_IDS="$PARENT_ID"
 for rt in $PV_RUNTIMES; do
+  RT_MODEL=$(pv_platform_model_for_runtime "$rt")
   R=$(tenant_call POST /workspaces \
-    -d "{\"name\":\"pv-$rt\",\"runtime\":\"$rt\",\"tier\":2,\"parent_id\":\"$PARENT_ID\",\"secrets\":$SECRETS_JSON}")
+    -d "{\"name\":\"pv-$rt\",\"runtime\":\"$rt\",\"model\":\"$RT_MODEL\",\"tier\":2,\"parent_id\":\"$PARENT_ID\",\"secrets\":$SECRETS_JSON}")
   WID=$(echo "$R" | python3 -c "import sys,json; print(json.load(sys.stdin).get('id',''))" 2>/dev/null)
   WTOK=$(echo "$R" | extract_auth_token)
   [ -n "$WID" ] || fail "$rt workspace create failed: $(printf '%s' "$R" | head -c 300)"

tests/e2e/test_reconciler_heals_terminated_instance.sh

@@ -0,0 +1,493 @@
+#!/usr/bin/env bash
+# Live staging E2E — the CP instance-state reconciler heals a terminated EC2.
+#
+# Real-infra complement to the deterministic unit tests for core#2261
+# (workspace-server/internal/registry/cp_instance_reconciler.go). Those unit
+# tests pin the reconcile logic against fakes; THIS script proves the loop
+# actually runs in a real tenant's workspace-server and drives the EXISTING
+# offline + auto-heal machinery against real AWS.
+#
+# Root regression (core#2247): a SaaS workspace whose EC2 is terminated out
+# from under the platform (manual AWS action, spot reclaim, CP reap) fell
+# through every existing liveness pass and kept reading status='online'
+# forever, pointing at a dead instance. The reconciler closes that gap with
+# CPProvisioner.IsRunning and feeds a clean "not running" into onOffline →
+# RestartByID (existing-volume reprovision).
+#
+# What this test does:
+#   1. Provision a fresh staging org + ONE workspace (same default
+#      runtime/model as the full-saas harness, so it actually boots).
+#   2. Poll the tenant API until the workspace is status=online; capture its
+#      instance_id.
+#   3. KILL it — terminate that exact EC2 via `aws ec2 terminate-instances`.
+#   4. Assert the reconciler heals it:
+#        PRIMARY (gate)      — within ~180s the workspace status LEAVES
+#                              'online' (the reconciler detected the dead
+#                              instance via IsRunning and flipped it). This
+#                              is the core regression guard: a dead instance
+#                              must NOT keep reading 'online'.
+#        SECONDARY (best-effort) — within ~10 min it auto-reprovisions:
+#                              status returns to 'online' with a NEW
+#                              instance_id (onOffline → RestartByID
+#                              existing-volume heal). If reprovision doesn't
+#                              finish in the bound we log it clearly but let
+#                              the PRIMARY assertion stand as the gate (see
+#                              the comment at the secondary block — a future
+#                              tightening that promotes this to a hard gate is
+#                              deliberately one edit away).
+#   5. Teardown ALWAYS (EXIT trap): delete the tenant + leak-sweep so no EC2
+#      is orphaned, even on a mid-test failure.
+#
+# Auth model + provisioning conventions are copied verbatim from
+# test_staging_full_saas.sh (single MOLECULE_ADMIN_TOKEN → CP admin; per-
+# tenant admin token + X-Molecule-Org-Id header for tenant API). The kill
+# primitive + leak sweep reuse lib/aws_leak_check.sh.
+#
+# Required env:
+#   MOLECULE_CP_URL        default: https://staging-api.moleculesai.app
+#   MOLECULE_ADMIN_TOKEN   CP admin bearer — Railway staging CP_ADMIN_API_TOKEN
+#
+# Optional env (mirrors the full-saas harness where they overlap):
+#   E2E_RUNTIME                        claude-code (default)
+#   E2E_PROVISION_TIMEOUT_SECS         default 900 (cold EC2 budget)
+#   E2E_WORKSPACE_ONLINE_TIMEOUT_SECS  default 3600 (cold-boot worst-case)
+#   E2E_RECONCILE_OFFLINE_TIMEOUT_SECS default 180 (PRIMARY: leave 'online'.
+#                                      Reconciler cadence is 60s — 3 cycles +
+#                                      AWS terminate-visibility slack.)
+#   E2E_REPROVISION_TIMEOUT_SECS       default 600 (SECONDARY: back to online
+#                                      with a NEW instance_id)
+#   E2E_MINIMAX_API_KEY / E2E_ANTHROPIC_API_KEY / E2E_OPENAI_API_KEY
+#                                      LLM key (same priority chain as
+#                                      full-saas; needed so the FIRST boot
+#                                      reaches online). Empty → '{}' (the
+#                                      workspace still boots online; the LLM
+#                                      key only matters for a completion,
+#                                      which this test never makes).
+#   E2E_KEEP_ORG                       1 → skip teardown (debugging only)
+#   E2E_RUN_ID                         Slug suffix; CI: ${GITHUB_RUN_ID}
+#   E2E_AWS_LEAK_CHECK                 auto (default) | required | off
+#   E2E_AWS_TERMINATE_LEAKS            1 → terminate slug-tagged leaked EC2 at
+#                                      teardown
+#
+# Exit codes:
+#   0  happy path (PRIMARY assertion held; SECONDARY logged either way)
+#   1  generic failure (incl. PRIMARY assertion failed = regression)
+#   2  missing required env
+#   3  provisioning timed out
+#   4  teardown left orphan resources
+
+set -euo pipefail
+
+CP_URL="${MOLECULE_CP_URL:-https://staging-api.moleculesai.app}"
+ADMIN_TOKEN="${MOLECULE_ADMIN_TOKEN:?MOLECULE_ADMIN_TOKEN required — Railway staging CP_ADMIN_API_TOKEN}"
+RUNTIME="${E2E_RUNTIME:-claude-code}"
+PROVISION_TIMEOUT_SECS="${E2E_PROVISION_TIMEOUT_SECS:-900}"
+WORKSPACE_ONLINE_TIMEOUT_SECS="${E2E_WORKSPACE_ONLINE_TIMEOUT_SECS:-3600}"
+# PRIMARY bound: the reconciler ticks every 60s; it needs one cycle to see
+# the dead instance after AWS makes the terminate visible to DescribeInstances
+# (typically seconds, but can lag). 180s = ~3 cycles + slack.
+RECONCILE_OFFLINE_TIMEOUT_SECS="${E2E_RECONCILE_OFFLINE_TIMEOUT_SECS:-180}"
+# SECONDARY bound: full existing-volume reprovision (new EC2 boot + agent
+# bootstrap) is a multi-minute cold path.
+REPROVISION_TIMEOUT_SECS="${E2E_REPROVISION_TIMEOUT_SECS:-600}"
+RUN_ID_SUFFIX="${E2E_RUN_ID:-$(date +%H%M%S)-$$}"
+
+# Slug MUST start with e2e- so sweep-stale-e2e-orgs.yml reaps any orphan this
+# run leaks (lint_cleanup_traps.sh enforces the e2e-/rt-e2e- prefix for any
+# staging tenant E2E; we honour it here too even though our filename isn't
+# *staging*).
+SLUG="e2e-rec-$(date +%Y%m%d)-${RUN_ID_SUFFIX}"
+SLUG=$(echo "$SLUG" | tr '[:upper:]' '[:lower:]' | tr -cd 'a-z0-9-' | head -c 32)
+
+log()  { echo "[$(date +%H:%M:%S)] $*"; }
+fail() { echo "[$(date +%H:%M:%S)] ❌ $*" >&2; exit 1; }
+ok()   { echo "[$(date +%H:%M:%S)] ✅ $*"; }
+
+# Per-runtime model slug dispatch — shared with the full-saas harness.
+# shellcheck disable=SC1091
+# shellcheck source=lib/model_slug.sh
+source "$(dirname "$0")/lib/model_slug.sh"
+# AWS kill primitive + leak sweep (e2e_aws_region / e2e_ec2_instances_for_slug /
+# e2e_terminate_instances / e2e_verify_no_ec2_leaks_for_slug).
+# shellcheck disable=SC1091
+# shellcheck source=lib/aws_leak_check.sh
+source "$(dirname "$0")/lib/aws_leak_check.sh"
+
+CURL_COMMON=(-sS --fail-with-body --max-time 30)
+
+# ─── cleanup trap ───────────────────────────────────────────────────────
+# Identical teardown contract to test_staging_full_saas.sh: delete the
+# tenant (synchronous GDPR cascade), poll for the org row to disappear, then
+# assert no slug-tagged EC2 survives. A leaked resource at teardown is a CI
+# failure (exit 4). The trap is installed UP-FRONT so a mid-test failure
+# (including a failed PRIMARY assertion) still cleans up.
+CLEANUP_DONE=0
+cleanup_org() {
+  # Capture upstream exit code IMMEDIATELY — must be the first statement in
+  # the trap, before any command (including the CLEANUP_DONE check) clobbers $?.
+  local entry_rc=$?
+
+  if [ "$CLEANUP_DONE" = "1" ]; then return 0; fi
+  CLEANUP_DONE=1
+
+  if [ "${E2E_KEEP_ORG:-0}" = "1" ]; then
+    log "E2E_KEEP_ORG=1 — skipping teardown. Manually delete $SLUG when done."
+    return 0
+  fi
+
+  log "🧹 Tearing down org $SLUG..."
+
+  # 120s curl budget for the synchronous DELETE cascade (EC2 terminate alone
+  # is 30-60s), then poll up to 60s for organizations.status='purged'/gone.
+  if curl "${CURL_COMMON[@]}" --max-time 120 -X DELETE "$CP_URL/cp/admin/tenants/$SLUG" \
+    -H "Authorization: Bearer $ADMIN_TOKEN" \
+    -H "Content-Type: application/json" \
+    -d "{\"confirm\":\"$SLUG\"}" >/dev/null 2>&1; then
+    ok "Teardown request accepted"
+  else
+    log "Teardown returned non-2xx (may already be gone)"
+  fi
+
+  local leak_count=1
+  local elapsed=0
+  while [ "$elapsed" -lt 60 ]; do
+    leak_count=$(curl "${CURL_COMMON[@]}" "$CP_URL/cp/admin/orgs" \
+      -H "Authorization: Bearer $ADMIN_TOKEN" 2>/dev/null \
+      | python3 -c "import json,sys; d=json.load(sys.stdin); print(sum(1 for o in d.get('orgs', []) if o.get('slug')=='$SLUG' and o.get('status') != 'purged'))" \
+      2>/dev/null || echo 1)
+    if [ "$leak_count" = "0" ]; then
+      break
+    fi
+    sleep 5
+    elapsed=$((elapsed + 5))
+  done
+
+  if [ "$leak_count" != "0" ]; then
+    echo "⚠️  LEAK: org $SLUG still present post-teardown after ${elapsed}s (count=$leak_count)" >&2
+    exit 4
+  fi
+  local aws_leak_rc=0
+  e2e_verify_no_ec2_leaks_for_slug "$SLUG" || aws_leak_rc=$?
+  if [ "$aws_leak_rc" != "0" ]; then
+    case "$aws_leak_rc" in
+      2) exit 2 ;;
+      *) exit 4 ;;
+    esac
+  fi
+  ok "Teardown clean — no orphan org or EC2 resources for $SLUG (${elapsed}s)"
+
+  # Normalize unexpected upstream exit codes to 1 — `set -e` propagates the
+  # raw exit code of the failing command (e.g. curl exits 22 under
+  # --fail-with-body), but this script's contract only emits {0,1,2,3,4}.
+  case "$entry_rc" in
+    0|1|2|3|4) ;;
+    *) exit 1 ;;
+  esac
+}
+trap cleanup_org EXIT INT TERM
+
+# ─── 0. Preflight ───────────────────────────────────────────────────────
+log "═══════════════════════════════════════════════════════════════════"
+log " Staging reconciler-heals-terminated-instance E2E (core#2261)"
+log "   CP:                 $CP_URL"
+log "   Slug:               $SLUG"
+log "   Runtime:            $RUNTIME"
+log "   Online timeout:     ${WORKSPACE_ONLINE_TIMEOUT_SECS}s"
+log "   PRIMARY (offline):  ${RECONCILE_OFFLINE_TIMEOUT_SECS}s"
+log "   SECONDARY (reprov): ${REPROVISION_TIMEOUT_SECS}s"
+log "═══════════════════════════════════════════════════════════════════"
+
+log "0/6 Preflight: CP reachable?"
+curl "${CURL_COMMON[@]}" "$CP_URL/health" >/dev/null || fail "CP health check failed"
+ok "CP reachable"
+
+admin_call() {
+  local method="$1"; shift
+  local path="$1"; shift
+  curl "${CURL_COMMON[@]}" -X "$method" "$CP_URL$path" \
+    -H "Authorization: Bearer $ADMIN_TOKEN" \
+    -H "Content-Type: application/json" \
+    "$@"
+}
+
+# ─── 1. Create org ──────────────────────────────────────────────────────
+log "1/6 Creating org $SLUG via /cp/admin/orgs..."
+CREATE_RESP=$(admin_call POST /cp/admin/orgs \
+  -d "{\"slug\":\"$SLUG\",\"name\":\"E2E $SLUG\",\"owner_user_id\":\"e2e-runner:$SLUG\"}")
+echo "$CREATE_RESP" | python3 -m json.tool >/dev/null || fail "Org create returned non-JSON: $CREATE_RESP"
+ORG_ID=$(echo "$CREATE_RESP" | python3 -c "import json,sys; print(json.load(sys.stdin).get('id',''))")
+[ -z "$ORG_ID" ] && fail "Org create response missing 'id': $CREATE_RESP"
+ok "Org created (id=$ORG_ID)"
+
+# ─── 2. Wait for tenant provisioning ────────────────────────────────────
+log "2/6 Waiting for tenant provisioning (up to ${PROVISION_TIMEOUT_SECS}s)..."
+DEADLINE=$(( $(date +%s) + PROVISION_TIMEOUT_SECS ))
+LAST_STATUS=""
+while true; do
+  if [ "$(date +%s)" -gt "$DEADLINE" ]; then
+    fail "Tenant provisioning timed out after ${PROVISION_TIMEOUT_SECS}s (last: $LAST_STATUS)"
+  fi
+  LIST_JSON=$(admin_call GET /cp/admin/orgs 2>/dev/null || echo '{"orgs":[]}')
+  # /cp/admin/orgs exposes 'instance_status' (org_instances.status), NOT 'status'.
+  STATUS=$(echo "$LIST_JSON" | python3 -c "
+import json, sys
+d = json.load(sys.stdin)
+for o in d.get('orgs', []):
+    if o.get('slug') == '$SLUG':
+        print(o.get('instance_status', ''))
+        sys.exit(0)
+print('')
+" 2>/dev/null || echo "")
+  if [ "$STATUS" != "$LAST_STATUS" ]; then
+    log "    status → $STATUS"
+    LAST_STATUS="$STATUS"
+  fi
+  case "$STATUS" in
+    running)  break ;;
+    failed)
+      log "── DIAGNOSTIC BURST (step 2 — tenant provisioning failed) ──"
+      echo "$LIST_JSON" | python3 -c "
+import json, sys
+d = json.load(sys.stdin)
+for o in d.get('orgs', []):
+    if o.get('slug') == '$SLUG':
+        print(json.dumps(o, indent=2))
+        sys.exit(0)
+print('(no org row found for slug=$SLUG — DB drift?)')
+" 2>&1 | sed 's/^/  /'
+      log "── END DIAGNOSTIC ──"
+      # Tenant provisioning failures are a CP-side fault, not a reconciler
+      # regression — exit 3 (provisioning) to keep the signal honest.
+      echo "[$(date +%H:%M:%S)] ❌ Tenant provisioning failed for $SLUG (see diagnostic above)" >&2
+      exit 3
+      ;;
+    *)        sleep 15 ;;
+  esac
+done
+ok "Tenant provisioning complete"
+
+# Derive tenant domain from CP hostname (same logic as the full-saas harness).
+CP_HOST=$(echo "$CP_URL" | sed -E 's#^https?://##; s#/.*$##')
+case "$CP_HOST" in
+  api.*)         DERIVED_DOMAIN="${CP_HOST#api.}" ;;
+  staging-api.*) DERIVED_DOMAIN="staging.${CP_HOST#staging-api.}" ;;
+  *)             DERIVED_DOMAIN="$CP_HOST" ;;
+esac
+TENANT_DOMAIN="${MOLECULE_TENANT_DOMAIN:-$DERIVED_DOMAIN}"
+TENANT_URL="https://$SLUG.$TENANT_DOMAIN"
+log "    TENANT_URL=$TENANT_URL"
+
+# ─── 3. Retrieve per-tenant admin token ────────────────────────────────
+log "3/6 Fetching per-tenant admin token..."
+TENANT_TOKEN_RESP=$(admin_call GET "/cp/admin/orgs/$SLUG/admin-token")
+TENANT_TOKEN=$(echo "$TENANT_TOKEN_RESP" | python3 -c "import json,sys; print(json.load(sys.stdin).get('admin_token',''))" 2>/dev/null || echo "")
+[ -z "$TENANT_TOKEN" ] && fail "Could not retrieve per-tenant admin token for $SLUG"
+ok "Tenant admin token retrieved (len=${#TENANT_TOKEN})"
+
+# Wait for tenant TLS / DNS propagation before any tenant API call.
+log "    Waiting for tenant TLS / DNS propagation..."
+TLS_DEADLINE=$(( $(date +%s) + 15 * 60 ))
+while true; do
+  if curl -sSfk --max-time 5 "$TENANT_URL/health" >/dev/null 2>&1; then
+    break
+  fi
+  if [ "$(date +%s)" -gt "$TLS_DEADLINE" ]; then
+    fail "Tenant URL never responded 2xx on /health within 15m"
+  fi
+  sleep 5
+done
+ok "Tenant reachable at $TENANT_URL"
+
+tenant_call() {
+  local method="$1"; shift
+  local path="$1"; shift
+  # X-Molecule-Org-Id is REQUIRED — the tenant guard 404s anything without it
+  # (it does NOT 403, to hide tenant existence from org scanners).
+  curl "${CURL_COMMON[@]}" -X "$method" "$TENANT_URL$path" \
+    -H "Authorization: Bearer $TENANT_TOKEN" \
+    -H "X-Molecule-Org-Id: $ORG_ID" \
+    "$@"
+}
+
+# Helper: read a single field off GET /workspaces/<id>. Echoes '' on any
+# error so callers can poll without `set -e` aborting on a transient blip.
+ws_field() {
+  local wid="$1"; local field="$2"
+  tenant_call GET "/workspaces/$wid" 2>/dev/null \
+    | python3 -c "import json,sys; print(json.load(sys.stdin).get('$field') or '')" 2>/dev/null \
+    || echo ""
+}
+
+# ─── 4. Provision ONE workspace ─────────────────────────────────────────
+# Same secrets-injection priority chain as the full-saas harness so the
+# FIRST boot reaches online. We never make a completion in this test (the
+# whole exercise is instance-state, not the LLM), so an absent key is
+# tolerable — but wiring the same keys keeps boot behaviour identical to the
+# sibling and avoids a config path that only this test would exercise.
+SECRETS_JSON='{}'
+if [ -n "${E2E_MINIMAX_API_KEY:-}" ]; then
+  SECRETS_JSON=$(python3 -c "import json,os; print(json.dumps({'MINIMAX_API_KEY': os.environ['E2E_MINIMAX_API_KEY']}))")
+elif [ -n "${E2E_ANTHROPIC_API_KEY:-}" ]; then
+  SECRETS_JSON=$(python3 -c "import json,os; print(json.dumps({'ANTHROPIC_API_KEY': os.environ['E2E_ANTHROPIC_API_KEY']}))")
+elif [ -n "${E2E_OPENAI_API_KEY:-}" ]; then
+  SECRETS_JSON=$(python3 -c "
+import json, os
+k = os.environ['E2E_OPENAI_API_KEY']
+print(json.dumps({
+    'OPENAI_API_KEY': k,
+    'OPENAI_BASE_URL': 'https://api.openai.com/v1',
+    'MODEL_PROVIDER': 'openai:gpt-4o',
+    'HERMES_INFERENCE_PROVIDER': 'custom',
+    'HERMES_CUSTOM_BASE_URL': 'https://api.openai.com/v1',
+    'HERMES_CUSTOM_API_KEY': k,
+    'HERMES_CUSTOM_API_MODE': 'chat_completions',
+}))
+")
+fi
+
+MODEL_SLUG=$(pick_model_slug "$RUNTIME")
+log "    MODEL_SLUG=$MODEL_SLUG"
+
+log "4/6 Provisioning workspace (runtime=$RUNTIME)..."
+WS_RESP=$(tenant_call POST /workspaces \
+  -H "Content-Type: application/json" \
+  -d "{\"name\":\"E2E Reconciler\",\"runtime\":\"$RUNTIME\",\"tier\":2,\"model\":\"$MODEL_SLUG\",\"secrets\":$SECRETS_JSON}")
+WS_ID=$(echo "$WS_RESP" | python3 -c "import json,sys; print(json.load(sys.stdin)['id'])")
+[ -z "$WS_ID" ] && fail "Workspace create response missing 'id': $WS_RESP"
+log "    WS_ID=$WS_ID"
+
+# Wait for the workspace to reach status=online and capture its instance_id.
+log "    Waiting for workspace to reach status=online (up to $((WORKSPACE_ONLINE_TIMEOUT_SECS/60)) min)..."
+ONLINE_DEADLINE=$(( $(date +%s) + WORKSPACE_ONLINE_TIMEOUT_SECS ))
+ORIGINAL_INSTANCE_ID=""
+WS_LAST_STATUS=""
+while true; do
+  if [ "$(date +%s)" -gt "$ONLINE_DEADLINE" ]; then
+    WS_LAST_ERR=$(ws_field "$WS_ID" "last_sample_error")
+    fail "Workspace $WS_ID never reached status=online within ${WORKSPACE_ONLINE_TIMEOUT_SECS}s (last status=$WS_LAST_STATUS, err=$WS_LAST_ERR)"
+  fi
+  WS_STATUS=$(ws_field "$WS_ID" "status")
+  if [ "$WS_STATUS" != "$WS_LAST_STATUS" ]; then
+    log "    $WS_ID → $WS_STATUS"
+    WS_LAST_STATUS="$WS_STATUS"
+  fi
+  if [ "$WS_STATUS" = "online" ]; then
+    ORIGINAL_INSTANCE_ID=$(ws_field "$WS_ID" "instance_id")
+    if [ -n "$ORIGINAL_INSTANCE_ID" ]; then
+      break
+    fi
+    # online but instance_id not surfaced yet — keep polling briefly.
+    log "    $WS_ID online but instance_id not populated yet — waiting"
+  fi
+  # 'failed' is transient on cold boot (bootstrap-watcher deadline vs heartbeat
+  # recovery, cp#245). Keep polling; only the deadline hard-fails.
+  sleep 10
+done
+ok "Workspace online (instance_id=$ORIGINAL_INSTANCE_ID)"
+
+# ─── 5. Kill the EC2 ────────────────────────────────────────────────────
+# Terminate the EXACT instance the workspace reported. Prefer the captured
+# instance_id (precise — kills only this workspace's box); fall back to the
+# slug-tag describe if the API didn't surface an id (shouldn't happen — we
+# only break out of the online-wait once instance_id is non-empty).
+log "5/6 KILLING the workspace EC2 to simulate an out-of-band termination..."
+if ! e2e_aws_creds_available; then
+  fail "AWS CLI/creds unavailable — cannot terminate the EC2 to exercise the reconciler. Set AWS_ACCESS_KEY_ID/AWS_SECRET_ACCESS_KEY (the CI workflow wires these)."
+fi
+AWS_REGION_RESOLVED=$(e2e_aws_region)
+if [ -n "$ORIGINAL_INSTANCE_ID" ]; then
+  log "    Terminating $ORIGINAL_INSTANCE_ID in $AWS_REGION_RESOLVED (aws ec2 terminate-instances)..."
+  aws ec2 terminate-instances --region "$AWS_REGION_RESOLVED" --instance-ids "$ORIGINAL_INSTANCE_ID" >/dev/null \
+    || fail "aws ec2 terminate-instances failed for $ORIGINAL_INSTANCE_ID"
+  KILLED_IDS="$ORIGINAL_INSTANCE_ID"
+else
+  # Fallback path — find by slug tag and terminate.
+  log "    instance_id was empty — falling back to slug-tag describe ($SLUG)..."
+  ROWS=$(e2e_ec2_instances_for_slug "$SLUG" 2>/dev/null || echo "")
+  KILLED_IDS=$(echo "$ROWS" | awk 'NF {print $1}' | sort -u | tr '\n' ' ')
+  [ -n "$KILLED_IDS" ] || fail "No slug-tagged EC2 found for $SLUG — nothing to terminate"
+  log "    Terminating $KILLED_IDS in $AWS_REGION_RESOLVED..."
+  e2e_terminate_instances "$KILLED_IDS" || fail "terminate-instances failed for $KILLED_IDS"
+fi
+ok "Terminated EC2: $KILLED_IDS — reconciler should now detect the dead instance"
+
+# ─── 6a. PRIMARY assertion — workspace leaves 'online' ─────────────────
+# This is THE regression gate for core#2261/#2247. The reconciler runs every
+# 60s in the tenant's workspace-server; when CPProvisioner.IsRunning returns a
+# clean "not running" for the terminated EC2, onOffline flips the row off
+# 'online'. A dead instance that keeps reading 'online' is exactly the bug.
+log "6a/6 PRIMARY: asserting workspace leaves 'online' within ${RECONCILE_OFFLINE_TIMEOUT_SECS}s (reconciler heal-detection)..."
+OFFLINE_DEADLINE=$(( $(date +%s) + RECONCILE_OFFLINE_TIMEOUT_SECS ))
+LEFT_ONLINE=0
+REC_LAST_STATUS=""
+while true; do
+  if [ "$(date +%s)" -gt "$OFFLINE_DEADLINE" ]; then
+    break
+  fi
+  REC_STATUS=$(ws_field "$WS_ID" "status")
+  if [ "$REC_STATUS" != "$REC_LAST_STATUS" ]; then
+    log "    $WS_ID status → ${REC_STATUS:-<empty>}"
+    REC_LAST_STATUS="$REC_STATUS"
+  fi
+  # Any non-online status (offline/provisioning/awaiting_agent/restarting/…)
+  # proves the reconciler acted. We deliberately don't pin the exact target
+  # status: onOffline flips offline AND kicks RestartByID, so the row may race
+  # straight into a provisioning/restarting state — all of which are "no longer
+  # falsely online".
+  if [ -n "$REC_STATUS" ] && [ "$REC_STATUS" != "online" ]; then
+    LEFT_ONLINE=1
+    ok "PRIMARY held — workspace left 'online' (now '$REC_STATUS') after EC2 termination"
+    break
+  fi
+  sleep 10
+done
+
+if [ "$LEFT_ONLINE" != "1" ]; then
+  fail "PRIMARY FAILED (core#2261 regression): workspace $WS_ID still reads status=online ${RECONCILE_OFFLINE_TIMEOUT_SECS}s after its EC2 ($KILLED_IDS) was terminated. The reconciler did NOT detect the dead instance — a terminated EC2 is masquerading as a healthy workspace."
+fi
+
+# ─── 6b. SECONDARY assertion — auto-reprovision (best-effort) ──────────
+# The onOffline → RestartByID existing-volume heal should bring the workspace
+# back to 'online' on a NEW instance_id. This is best-effort: a full EC2 cold
+# reprovision is a multi-minute path that shares the same boot-flake surface
+# as the initial provision. If it doesn't finish within the bound we LOG it
+# clearly but DO NOT fail — the PRIMARY assertion above is the gate.
+#
+# FUTURE TIGHTENING (deliberately one edit away): once this reprovision path
+# is proven reliable on staging, promote the `log "SECONDARY ..."` soft-miss
+# below to a `fail ...` so a stuck reprovision becomes a hard gate.
+log "6b/6 SECONDARY (best-effort): asserting auto-reprovision to online with a NEW instance_id within ${REPROVISION_TIMEOUT_SECS}s..."
+REPROV_DEADLINE=$(( $(date +%s) + REPROVISION_TIMEOUT_SECS ))
+REPROV_OK=0
+REPROV_LAST_STATUS=""
+NEW_INSTANCE_ID=""
+while true; do
+  if [ "$(date +%s)" -gt "$REPROV_DEADLINE" ]; then
+    break
+  fi
+  RP_STATUS=$(ws_field "$WS_ID" "status")
+  if [ "$RP_STATUS" != "$REPROV_LAST_STATUS" ]; then
+    log "    $WS_ID status → ${RP_STATUS:-<empty>}"
+    REPROV_LAST_STATUS="$RP_STATUS"
+  fi
+  if [ "$RP_STATUS" = "online" ]; then
+    NEW_INSTANCE_ID=$(ws_field "$WS_ID" "instance_id")
+    if [ -n "$NEW_INSTANCE_ID" ] && [ "$NEW_INSTANCE_ID" != "$ORIGINAL_INSTANCE_ID" ]; then
+      REPROV_OK=1
+      break
+    fi
+    # online again but instance_id either not surfaced yet or still the old
+    # (terminated) id — keep polling until the reprovision swaps it.
+  fi
+  sleep 15
+done
+
+if [ "$REPROV_OK" = "1" ]; then
+  ok "SECONDARY held — auto-reprovisioned to online on NEW instance_id=$NEW_INSTANCE_ID (was $ORIGINAL_INSTANCE_ID)"
+else
+  # Soft-miss — see FUTURE TIGHTENING note above. PRIMARY is the gate.
+  log "⚠️  SECONDARY not satisfied within ${REPROVISION_TIMEOUT_SECS}s (status=${REPROV_LAST_STATUS:-<empty>}, instance_id=${NEW_INSTANCE_ID:-<none>}, original=$ORIGINAL_INSTANCE_ID). NOT failing — the PRIMARY heal-detection assertion is the gate; reprovision is a slower, flakier cold path. Promote this to a hard fail once it's proven reliable."
+fi
+
+ok "Reconciler live E2E PASSED — PRIMARY heal-detection held (SECONDARY: $([ "$REPROV_OK" = "1" ] && echo "held" || echo "soft-miss, logged"))"
+# Teardown runs via the EXIT trap.

tests/e2e/test_staging_full_saas.sh

@@ -1041,7 +1041,7 @@ print(json.dumps({
             'messageId': f'e2e-{uuid.uuid4().hex[:8]}',
             'parts': [{'kind': 'text', 'text': 'Reply with exactly: ok'}],
         },
-        'configuration': {'max_tokens': 4}
+        'configuration': {'max_tokens': 32}
     }
 }))
 ")

tests/test_ci_required_drift.py

@@ -18,6 +18,7 @@ No network. No live Gitea calls.
 from __future__ import annotations
 
 import importlib.util
+import json
 import os
 import textwrap
 from pathlib import Path
@@ -117,6 +118,31 @@ def _write_audit_yaml(tmp_path: Path, required_checks: list[str]) -> Path:
     return p
 
 
+def _write_audit_yaml_json(tmp_path: Path, required_checks_json: dict) -> Path:
+    """Write a synthetic audit-force-merge.yml with REQUIRED_CHECKS_JSON env."""
+    block = json.dumps(required_checks_json, indent=2)
+    text = textwrap.dedent(
+        f"""\
+        name: audit-force-merge
+        on:
+          schedule:
+            - cron: '*/30 * * * *'
+        jobs:
+          audit:
+            runs-on: ubuntu-latest
+            steps:
+              - name: Run audit
+                env:
+                  REQUIRED_CHECKS_JSON: |
+                    {block.replace(chr(10), chr(10) + '                    ')}
+                run: bash .gitea/scripts/audit-force-merge.sh
+        """
+    )
+    p = tmp_path / "audit-force-merge.yml"
+    p.write_text(text, encoding="utf-8")
+    return p
+
+
 def _make_stub_api(responses: dict):
     """Build a fake `api()` callable.
 
@@ -363,6 +389,107 @@ def test_happy_path_no_drift(drift_module, tmp_path, monkeypatch):
     assert findings == [], findings
 
 
+# --------------------------------------------------------------------------
+# REQUIRED_CHECKS_JSON variant drift tests
+# --------------------------------------------------------------------------
+def test_f3a_env_wider_than_protection_json_variant(drift_module, tmp_path, monkeypatch):
+    """F3a: REQUIRED_CHECKS_JSON env has a context NOT in protection."""
+    ci = _write_ci_yaml(
+        tmp_path,
+        jobs={"build": {"runs-on": "ubuntu-latest"}},
+        sentinel_needs=["build"],
+    )
+    audit = _write_audit_yaml_json(
+        tmp_path,
+        {"main": ["ci / build (pull_request)", "ci / ghost (pull_request)"]},
+    )
+    _patch_paths(drift_module, monkeypatch, ci, audit)
+
+    stub = _make_stub_api({
+        ("GET", "/repos/owner/repo/branch_protections/main"): (
+            200,
+            {"status_check_contexts": ["ci / build (pull_request)"]},
+        ),
+    })
+    monkeypatch.setattr(drift_module, "api", stub)
+
+    findings, _ = drift_module.detect_drift("main")
+    assert any("F3a" in f and "ghost" in f for f in findings), findings
+
+
+def test_f3b_protection_wider_than_env_json_variant(drift_module, tmp_path, monkeypatch):
+    """F3b: protection has a context NOT in REQUIRED_CHECKS_JSON env."""
+    ci = _write_ci_yaml(
+        tmp_path,
+        jobs={
+            "build": {"runs-on": "ubuntu-latest"},
+            "test": {"runs-on": "ubuntu-latest"},
+        },
+        sentinel_needs=["build", "test"],
+    )
+    audit = _write_audit_yaml_json(
+        tmp_path,
+        {"main": ["ci / build (pull_request)"]},
+    )
+    _patch_paths(drift_module, monkeypatch, ci, audit)
+
+    stub = _make_stub_api({
+        ("GET", "/repos/owner/repo/branch_protections/main"): (
+            200,
+            {
+                "status_check_contexts": [
+                    "ci / build (pull_request)",
+                    "ci / test (pull_request)",
+                ]
+            },
+        ),
+    })
+    monkeypatch.setattr(drift_module, "api", stub)
+
+    findings, _ = drift_module.detect_drift("main")
+    assert any("F3b" in f and "ci / test (pull_request)" in f for f in findings), findings
+
+
+def test_happy_path_no_drift_json_variant(drift_module, tmp_path, monkeypatch):
+    """Happy path with REQUIRED_CHECKS_JSON: all aligned."""
+    ci = _write_ci_yaml(
+        tmp_path,
+        jobs={
+            "build": {"runs-on": "ubuntu-latest"},
+            "test": {"runs-on": "ubuntu-latest"},
+        },
+        sentinel_needs=["build", "test"],
+    )
+    audit = _write_audit_yaml_json(
+        tmp_path,
+        {
+            "main": [
+                "ci / build (pull_request)",
+                "ci / test (pull_request)",
+                "ci / all-required (pull_request)",
+            ]
+        },
+    )
+    _patch_paths(drift_module, monkeypatch, ci, audit)
+
+    stub = _make_stub_api({
+        ("GET", "/repos/owner/repo/branch_protections/main"): (
+            200,
+            {
+                "status_check_contexts": [
+                    "ci / build (pull_request)",
+                    "ci / test (pull_request)",
+                    "ci / all-required (pull_request)",
+                ]
+            },
+        ),
+    })
+    monkeypatch.setattr(drift_module, "api", stub)
+
+    findings, _ = drift_module.detect_drift("main")
+    assert findings == [], findings
+
+
 # --------------------------------------------------------------------------
 # MUST-FIX 1: find_open_issue must raise on transient HTTP errors
 # --------------------------------------------------------------------------

workspace-server/cmd/server/main.go

@@ -337,6 +337,25 @@ func main() {
 		})
 	}
 
+	// CP-mode instance-state reconciler — authoritative EC2-liveness pass
+	// for SaaS workspaces (core#2261). Every other liveness sweep keys off
+	// a PROXY (Redis TTL, agent heartbeat, local Docker, or
+	// runtime='external'); a SaaS claude-code workspace whose EC2 was
+	// terminated/stopped falls through ALL of them and stays status='online'
+	// pointing at a dead instance_id forever (root cause: core#2247). This
+	// loop asks the ONE authoritative question the others lack —
+	// cpProv.IsRunning (CP DescribeInstances-equivalent) — for each online
+	// SaaS row, and on a CLEAN "not running" feeds it into the SAME
+	// onWorkspaceOffline closure the other sweeps use (status flip +
+	// RestartByID reprovision, existing volume). Fail-safe: IsRunning is
+	// (true, err) on any transient error, so a CP blip never flips a healthy
+	// workspace.
+	if cpProv != nil {
+		go supervised.RunWithRecover(ctx, "cp-instance-reconciler", func(c context.Context) {
+			registry.StartCPInstanceReconciler(c, cpProv, onWorkspaceOffline, 60*time.Second)
+		})
+	}
+
 	// Pending-uploads GC sweep — deletes acked rows past their retention
 	// window plus unacked rows past expires_at. Without this the
 	// pending_uploads table grows unbounded; even with the 24h hard TTL,

workspace-server/internal/handlers/activity.go

@@ -152,7 +152,7 @@ func extractAttachmentsFromMessageParts(body map[string]interface{}) []map[strin
 		if kind == "" {
 			kind, _ = part["type"].(string)
 		}
-		if kind != "file" && kind != "image" && kind != "audio" {
+		if kind != "file" && kind != "image" && kind != "audio" && kind != "video" {
 			continue
 		}
 		// The file sub-object holds uri/mime_type/name. The a2a-sdk v1
@@ -380,12 +380,18 @@ func (h *ActivityHandler) List(c *gin.Context) {
 	// "row not found" — both indicate the cursor is no longer usable for
 	// this caller, no information leak.
 	var cursorTime time.Time
+	var cursorSeq int64
 	usingCursor := false
 	if sinceID != "" {
+		// Resolve BOTH ordering-key components of the cursor row. The feed is
+		// ordered by (created_at, seq), so the strictly-after filter below must
+		// compare the full tuple — comparing created_at alone silently drops a
+		// row written in the SAME microsecond as the cursor row (the boundary
+		// skip the since_id E2E intermittently tripped over).
 		err := db.DB.QueryRowContext(c.Request.Context(),
-			`SELECT created_at FROM activity_logs WHERE id = $1 AND workspace_id = $2`,
+			`SELECT created_at, seq FROM activity_logs WHERE id = $1 AND workspace_id = $2`,
 			sinceID, workspaceID,
-		).Scan(&cursorTime)
+		).Scan(&cursorTime, &cursorSeq)
 		if errors.Is(err, sql.ErrNoRows) {
 			c.JSON(http.StatusGone, gin.H{
 				"error": "since_id cursor not found (row may have been pruned or belongs to a different workspace); omit since_id to reset",
@@ -492,10 +498,20 @@ func (h *ActivityHandler) List(c *gin.Context) {
 		argIdx++
 	}
 	if usingCursor {
-		// Strictly after — never replay the cursor row itself.
-		query += fmt.Sprintf(" AND "+actCol+"created_at > $%d", argIdx)
-		args = append(args, cursorTime)
-		argIdx++
+		// Strictly after the cursor on the FULL ordering key (created_at, seq).
+		// Tuple comparison: a row is "after" the cursor if its created_at is
+		// later, OR it shares the cursor's created_at but has a higher seq.
+		// This (a) never replays the cursor row itself and (b) — unlike a bare
+		// `created_at > cursor` — never drops a row written in the same
+		// microsecond as the cursor row. Expressed as the expanded boolean
+		// rather than a row-value `(created_at, seq) > ($t, $s)` so it composes
+		// with the actCol qualifier prefix and the existing placeholder/arg
+		// builder cleanly.
+		query += fmt.Sprintf(
+			" AND ("+actCol+"created_at > $%d OR ("+actCol+"created_at = $%d AND "+actCol+"seq > $%d))",
+			argIdx, argIdx, argIdx+1)
+		args = append(args, cursorTime, cursorSeq)
+		argIdx += 2
 	}
 
 	// Polling clients (since_id) need oldest-first within the new window so
@@ -503,9 +519,13 @@ func (h *ActivityHandler) List(c *gin.Context) {
 	// since_id) keeps DESC — that's the canvas/UI shape and changing it
 	// would surprise existing callers.
 	if usingCursor {
-		query += fmt.Sprintf(" ORDER BY "+actCol+"created_at ASC LIMIT $%d", argIdx)
+		// (created_at, seq) ASC — seq is the deterministic tiebreaker for rows
+		// sharing a microsecond-collided created_at. Replays in recorded order.
+		query += fmt.Sprintf(" ORDER BY "+actCol+"created_at ASC, "+actCol+"seq ASC LIMIT $%d", argIdx)
 	} else {
-		query += fmt.Sprintf(" ORDER BY "+actCol+"created_at DESC LIMIT $%d", argIdx)
+		// (created_at, seq) DESC — same tiebreaker, newest-first for the
+		// canvas/recent-feed shape.
+		query += fmt.Sprintf(" ORDER BY "+actCol+"created_at DESC, "+actCol+"seq DESC LIMIT $%d", argIdx)
 	}
 	args = append(args, limit)
 
@@ -680,7 +700,8 @@ func buildSessionSearchQuery(workspaceID, query string, limit int) (string, []in
 				COALESCE(status, '') AS status,
 				request_body,
 				response_body,
-				created_at
+				created_at,
+				seq
 			FROM activity_logs
 			WHERE workspace_id = $1
 		)
@@ -702,7 +723,13 @@ func buildSessionSearchQuery(workspaceID, query string, limit int) (string, []in
 		args = append(args, "%"+query+"%")
 	}
 
-	sqlQuery += ` ORDER BY created_at DESC LIMIT $` + strconv.Itoa(len(args)+1)
+	// Deterministic order: created_at alone is not unique (same-microsecond
+	// rows), so tie-break on the monotonic seq — same fix as the since_id feed
+	// (§ No flakes: no unstable sorts, even on an unused surface). `seq` is
+	// projected through the session_items CTE above so this outer ORDER BY can
+	// reference it — the outer SELECT can only sort on the CTE's output columns,
+	// not on activity_logs directly.
+	sqlQuery += ` ORDER BY created_at DESC, seq DESC LIMIT $` + strconv.Itoa(len(args)+1)
 	args = append(args, limit)
 	return sqlQuery, args
 }

workspace-server/internal/handlers/activity_peer_info_test.go

@@ -118,6 +118,23 @@ func TestExtractAttachmentsFromRequestBody_ImageAndAudio(t *testing.T) {
 	}
 }
 
+func TestExtractAttachmentsFromRequestBody_VideoPart(t *testing.T) {
+	// Video parts are accepted in message-parts envelope (issue #2222).
+	body := []byte(`{"jsonrpc":"2.0","method":"message/send","params":{"message":{"parts":[
+		{"kind":"video","file":{"uri":"workspace:clip.mp4","mime_type":"video/mp4","name":"clip.mp4"}}
+	]}}}`)
+	atts := extractAttachmentsFromRequestBody(body)
+	if len(atts) != 1 {
+		t.Fatalf("want 1 attachment, got %d", len(atts))
+	}
+	if atts[0]["kind"] != "video" {
+		t.Errorf("kind: want video, got %v", atts[0]["kind"])
+	}
+	if atts[0]["uri"] != "workspace:clip.mp4" {
+		t.Errorf("uri mismatch: %v", atts[0]["uri"])
+	}
+}
+
 func TestExtractAttachmentsFromRequestBody_LegacyV0TypeDiscriminator(t *testing.T) {
 	// Legacy v0 shape: type=file (not kind), inlined fields (no nested .file)
 	body := []byte(`{"jsonrpc":"2.0","method":"message/send","params":{"message":{"parts":[

workspace-server/internal/handlers/activity_seq_backfill_integration_test.go

@@ -0,0 +1,211 @@
+//go:build integration
+// +build integration
+
+// activity_seq_backfill_integration_test.go — REAL Postgres proof of the
+// invariant the 20260604000000_activity_logs_seq.up.sql migration guarantees:
+// every activity_logs row carries a NON-NULL `seq`, both for rows that existed
+// before the migration ran (assigned during the ALTER TABLE rewrite) and for
+// rows created afterward via the normal INSERT path (assigned by the IDENTITY
+// default). This is the coverage CR2 (#2339 review) correctly flagged as
+// missing on PR #2258.
+//
+// WHY THIS IS A SEPARATE TEST from activity_since_id_ordering_integration_test.go:
+// that test pins the *ordering* contract (same-microsecond rows come back in a
+// deterministic (created_at, seq) order). THIS test pins the *backfill* contract
+// — that `seq` is never NULL — and the consequence the reviewer doubted: a
+// pre-existing/backfilled row is usable as a since_id cursor because its seq is
+// non-null, so the tuple cursor `(created_at, seq)` the handler builds is well
+// defined for it.
+//
+// EMPIRICAL BASIS (PostgreSQL 16.13, the prod PG version):
+//   - `ALTER TABLE activity_logs ADD COLUMN seq BIGINT GENERATED BY DEFAULT AS
+//     IDENTITY` rewrites the table and assigns seq to EXISTING rows in physical
+//     table-scan order — they are NON-NULL, not left NULL as the review claimed.
+//   - The identity sequence then advances ABOVE max(seq), so the next INSERT
+//     that omits seq gets max+1 with no collision.
+// Run against any Postgres 15/16 the integration harness boots — the property
+// holds on both.
+//
+// Run with (same harness as activity_delegation_a2a_integration_test.go):
+//
+//	docker run --rm -d --name pg-integration \
+//	  -e POSTGRES_PASSWORD=test -e POSTGRES_DB=molecule \
+//	  -p 55432:5432 postgres:15-alpine
+//	sleep 4
+//	# apply migrations (incl. 20260604000000_activity_logs_seq.up.sql) then:
+//	INTEGRATION_DB_URL="postgres://postgres:test@localhost:55432/molecule?sslmode=disable" \
+//	  go test -tags=integration ./internal/handlers/ -run Integration_ActivityLogs_Seq
+//
+// WATCH-IT-FAIL: if `seq` were left nullable / un-backfilled (the failure mode
+// the reviewer hypothesized), the NULL-count assertion in _NoNull trips, and
+// the since_id-on-a-backfilled-row case in _SinceIDOnBackfilledRow trips because
+// the handler cannot read a non-null seq for the cursor row. With the migration
+// as written both are green every run.
+
+package handlers
+
+import (
+	"context"
+	"encoding/json"
+	"net/http"
+	"testing"
+	"time"
+
+	"git.moleculesai.app/molecule-ai/molecule-core/workspace-server/internal/db"
+	"github.com/gin-gonic/gin"
+)
+
+// TestIntegration_ActivityLogs_SeqBackfill_NoNull pins the core migration
+// invariant: AFTER migrations have run, NO activity_logs row may have a NULL
+// seq — neither rows that the seedActivityRowAt path inserts (IDENTITY default)
+// nor any row the schema carries. It also proves the IDENTITY sequence keeps
+// producing distinct, non-null seq for fresh inserts (no collision, no NULL).
+//
+// This is the assertion that would FAIL if the ALTER had left existing rows
+// with NULL seq (the reviewer's claim) — table-scan backfill makes it pass.
+func TestIntegration_ActivityLogs_SeqBackfill_NoNull(t *testing.T) {
+	conn := integrationDB_ActivityDelegationA2A(t)
+	_ = conn
+	wsID := seedWorkspace(t, conn, "test-2151-seq-backfill-nonull")
+
+	// Insert several rows via the normal path. seq is left to the IDENTITY
+	// default — exactly how production writes activity_logs.
+	t0 := time.Date(2026, 6, 4, 9, 0, 0, 0, time.UTC)
+	const n = 5
+	ids := make([]string, 0, n)
+	for i := 0; i < n; i++ {
+		ids = append(ids, seedActivityRowAt(t, wsID, "backfill-row", t0.Add(time.Duration(i)*time.Second)))
+	}
+
+	// (a) No row in this workspace may have a NULL seq. If the column were
+	// un-backfilled / nullable this is > 0 and the test fails.
+	var nullCount int
+	if err := db.DB.QueryRowContext(context.Background(),
+		`SELECT COUNT(*) FROM activity_logs WHERE workspace_id = $1 AND seq IS NULL`,
+		wsID,
+	).Scan(&nullCount); err != nil {
+		t.Fatalf("null-seq count query: %v", err)
+	}
+	if nullCount != 0 {
+		t.Fatalf("found %d activity_logs rows with NULL seq — migration did NOT backfill/assign seq", nullCount)
+	}
+
+	// Belt-and-suspenders: the GLOBAL invariant (no NULL seq anywhere in the
+	// table) is what the migration actually guarantees. Assert it too, so a
+	// regression that nulls seq for rows written by some other path is caught.
+	var globalNull int
+	if err := db.DB.QueryRowContext(context.Background(),
+		`SELECT COUNT(*) FROM activity_logs WHERE seq IS NULL`,
+	).Scan(&globalNull); err != nil {
+		t.Fatalf("global null-seq count query: %v", err)
+	}
+	if globalNull != 0 {
+		t.Fatalf("found %d activity_logs rows table-wide with NULL seq — seq must be non-null for every row", globalNull)
+	}
+
+	// (b) The IDENTITY sequence yields DISTINCT, monotonic, non-null seq for
+	// the rows we just inserted (proves the normal insert path gets a real seq,
+	// and that the sequence advanced past any backfilled max instead of
+	// colliding). We read them back in insert order and require strictly
+	// increasing, all-non-null seq.
+	rows, err := db.DB.QueryContext(context.Background(),
+		`SELECT seq FROM activity_logs WHERE workspace_id = $1 ORDER BY created_at ASC, seq ASC`,
+		wsID,
+	)
+	if err != nil {
+		t.Fatalf("read-back seq query: %v", err)
+	}
+	defer rows.Close()
+	var seqs []int64
+	for rows.Next() {
+		var s *int64 // pointer so a NULL would scan as nil rather than 0
+		if err := rows.Scan(&s); err != nil {
+			t.Fatalf("scan seq: %v", err)
+		}
+		if s == nil {
+			t.Fatal("a freshly-inserted activity_logs row has NULL seq — IDENTITY default did not fire")
+		}
+		seqs = append(seqs, *s)
+	}
+	if err := rows.Err(); err != nil {
+		t.Fatalf("rows err: %v", err)
+	}
+	if len(seqs) != n {
+		t.Fatalf("expected %d rows, read back %d", n, len(seqs))
+	}
+	for i := 1; i < len(seqs); i++ {
+		if seqs[i] <= seqs[i-1] {
+			t.Fatalf("seq not strictly increasing in insert order: %v (IDENTITY collision / reuse)", seqs)
+		}
+	}
+}
+
+// TestIntegration_ActivityLogs_SeqBackfill_SinceIDOnBackfilledRow pins the
+// consequence the reviewer doubted: a row whose seq came from the migration /
+// IDENTITY (i.e. NOT explicitly set by the caller) is usable as a since_id
+// cursor, and a SECOND row sharing its exact created_at microsecond is returned
+// (not dropped). This proves the handler's (created_at, seq) tuple cursor
+// resolves a same-timestamp boundary that a created_at-only cursor would drop,
+// AND that the cursor row's seq is non-null (else the handler could not build
+// the tuple at all).
+//
+// Distinct from _BoundaryRowSameMicrosecondNotSkipped in the ordering test:
+// here the explicit angle under test is "the cursor row's seq is a
+// migration/IDENTITY-assigned (backfilled-style) value, non-null, and the
+// handler uses it" — i.e. the backfill behavior is what makes the boundary
+// resolution work, pinned head-on.
+func TestIntegration_ActivityLogs_SeqBackfill_SinceIDOnBackfilledRow(t *testing.T) {
+	conn := integrationDB_ActivityDelegationA2A(t)
+	_ = conn
+	wsID := seedWorkspace(t, conn, "test-2151-seq-backfill-sinceid")
+
+	tSame := time.Date(2026, 6, 4, 10, 0, 0, 0, time.UTC)
+	// Cursor row: seq comes purely from the IDENTITY default (never set by
+	// the caller) — the same assignment mechanism the migration uses to
+	// backfill pre-existing rows. The "next" row shares the exact created_at
+	// microsecond and is inserted afterward, so it gets a strictly higher seq.
+	cursorID := seedActivityRowAt(t, wsID, "sinceid-cursor", tSame)
+	nextID := seedActivityRowAt(t, wsID, "sinceid-next-same-us", tSame)
+
+	// Prove the precondition the reviewer doubted: the cursor row's seq is
+	// NON-NULL, so the handler can read it to build the (created_at, seq)
+	// tuple. If it were NULL the handler's cursor lookup would yield a NULL
+	// seq and the strictly-after tuple comparison would mis-behave.
+	var cursorSeq *int64
+	if err := db.DB.QueryRowContext(context.Background(),
+		`SELECT seq FROM activity_logs WHERE id = $1`, cursorID,
+	).Scan(&cursorSeq); err != nil {
+		t.Fatalf("read cursor seq: %v", err)
+	}
+	if cursorSeq == nil {
+		t.Fatal("cursor row has NULL seq — a since_id cursor on a backfilled-style row would be unusable")
+	}
+
+	h := NewActivityHandler(nil)
+	c, w := newTestGinContext()
+	c.Params = gin.Params{{Key: "id", Value: wsID}}
+	q := c.Request.URL.Query()
+	q.Set("since_id", cursorID)
+	q.Set("type", "a2a_receive")
+	q.Set("limit", "10")
+	c.Request.URL.RawQuery = q.Encode()
+
+	h.List(c)
+	if w.Code != http.StatusOK {
+		t.Fatalf("List returned %d, want 200: %s", w.Code, w.Body.String())
+	}
+	var resp []map[string]interface{}
+	if err := json.Unmarshal(w.Body.Bytes(), &resp); err != nil {
+		t.Fatalf("unmarshal: %v", err)
+	}
+	// Exactly the one same-microsecond row after the cursor — present (not
+	// dropped by a strict created_at-only filter) and the cursor itself
+	// excluded (strictly-after on the full tuple).
+	if len(resp) != 1 {
+		t.Fatalf("same-microsecond row after backfilled-style cursor dropped: expected 1 row, got %d: %+v",
+			len(resp), resp)
+	}
+	if got, _ := resp[0]["id"].(string); got != nextID {
+		t.Fatalf("expected boundary row id %s, got %s", nextID, got)
+	}
+}

workspace-server/internal/handlers/activity_since_id_ordering_integration_test.go

@@ -0,0 +1,162 @@
+//go:build integration
+// +build integration
+
+// activity_since_id_ordering_integration_test.go — REAL Postgres proof that
+// the poll-mode since_id activity feed (#2339) is DETERMINISTICALLY ordered
+// even when multiple rows collide on the same created_at microsecond.
+//
+// This is the test that the original bug report mis-labeled a "flake".
+// sqlmock cannot catch it: sqlmock returns rows in the order the test stuffs
+// them, so it can never reveal a non-deterministic ORDER BY. Only a real
+// planner over real same-created_at rows exposes it.
+//
+// Run with (same harness as activity_delegation_a2a_integration_test.go):
+//
+//	docker run --rm -d --name pg-integration \
+//	  -e POSTGRES_PASSWORD=test -e POSTGRES_DB=molecule \
+//	  -p 55432:5432 postgres:15-alpine
+//	sleep 4
+//	# apply migrations (incl. 20260604000000_activity_logs_seq.up.sql) then:
+//	INTEGRATION_DB_URL="postgres://postgres:test@localhost:55432/molecule?sslmode=disable" \
+//	  go test -tags=integration ./internal/handlers/ -run Integration_SinceID
+//
+// WATCH-IT-FAIL: against the pre-fix handler (ORDER BY created_at only, no
+// seq tiebreaker, and `created_at > cursor` strict) this test is unstable —
+// the equal-created_at rows come back in arbitrary planner order so the
+// ordered-id assertion fails intermittently, and the same-microsecond
+// boundary row is dropped so the count assertion fails. With the fix
+// (ORDER BY created_at, seq + tuple cursor) it is green every run.
+
+package handlers
+
+import (
+	"context"
+	"encoding/json"
+	"net/http"
+	"testing"
+	"time"
+
+	"git.moleculesai.app/molecule-ai/molecule-core/workspace-server/internal/db"
+	"github.com/gin-gonic/gin"
+)
+
+// seedActivityRowAt inserts one activity_logs row with an explicit created_at
+// (so the test can force microsecond-equal collisions) and a unique summary;
+// returns the generated id. seq is left to the IDENTITY default — Postgres
+// assigns it in INSERT order, which is the deterministic tiebreaker under test.
+// db.DB has been hot-swapped to the integration connection by
+// integrationDB_ActivityDelegationA2A(t) in the calling test.
+func seedActivityRowAt(t *testing.T, wsID, summary string, createdAt time.Time) string {
+	t.Helper()
+	var id string
+	err := db.DB.QueryRowContext(context.Background(), `
+		INSERT INTO activity_logs (workspace_id, activity_type, summary, status, created_at)
+		VALUES ($1, 'a2a_receive', $2, 'ok', $3)
+		RETURNING id
+	`, wsID, summary, createdAt).Scan(&id)
+	if err != nil {
+		t.Fatalf("seedActivityRowAt(%q): %v", summary, err)
+	}
+	return id
+}
+
+// TestIntegration_SinceID_StableOrderingSameMicrosecond proves the feed is
+// deterministic when rows share a created_at, AND that the same-microsecond
+// boundary row immediately after the cursor is NOT dropped.
+func TestIntegration_SinceID_StableOrderingSameMicrosecond(t *testing.T) {
+	conn := integrationDB_ActivityDelegationA2A(t)
+	_ = conn
+	wsID := seedWorkspace(t, conn, "test-2151-sinceid-ordering")
+
+	// One earlier row to serve as the cursor (the "last processed" row).
+	tCursor := time.Date(2026, 6, 4, 12, 0, 0, 0, time.UTC)
+	cursorID := seedActivityRowAt(t, wsID, "cursor-row", tCursor)
+
+	// Three rows that ALL collide on the exact same created_at microsecond,
+	// inserted in a known order. Pre-fix, ORDER BY created_at alone returns
+	// these in arbitrary planner order.
+	tEqual := time.Date(2026, 6, 4, 12, 0, 1, 0, time.UTC)
+	idA := seedActivityRowAt(t, wsID, "equal-A", tEqual)
+	idB := seedActivityRowAt(t, wsID, "equal-B", tEqual)
+	idCc := seedActivityRowAt(t, wsID, "equal-C", tEqual)
+	wantOrder := []string{idA, idB, idCc}
+
+	// Drive the handler exactly as a polling client would.
+	h := NewActivityHandler(nil)
+	c, w := newTestGinContext()
+	c.Params = gin.Params{{Key: "id", Value: wsID}}
+	q := c.Request.URL.Query()
+	q.Set("since_id", cursorID)
+	q.Set("type", "a2a_receive")
+	q.Set("limit", "10")
+	c.Request.URL.RawQuery = q.Encode()
+
+	h.List(c)
+	if w.Code != http.StatusOK {
+		t.Fatalf("List returned %d, want 200: %s", w.Code, w.Body.String())
+	}
+	var resp []map[string]interface{}
+	if err := json.Unmarshal(w.Body.Bytes(), &resp); err != nil {
+		t.Fatalf("unmarshal: %v", err)
+	}
+
+	// All three equal-created_at rows must be present (boundary not dropped)
+	// and the cursor row itself must be excluded (strictly-after).
+	if len(resp) != len(wantOrder) {
+		t.Fatalf("expected %d rows after cursor (the 3 equal-created_at rows), got %d: %+v",
+			len(wantOrder), len(resp), resp)
+	}
+
+	gotOrder := make([]string, len(resp))
+	for i, row := range resp {
+		idVal, _ := row["id"].(string)
+		gotOrder[i] = idVal
+	}
+	for i := range wantOrder {
+		if gotOrder[i] != wantOrder[i] {
+			t.Fatalf("non-deterministic ordering: got id order %v, want %v (seq tiebreaker not applied)",
+				gotOrder, wantOrder)
+		}
+	}
+}
+
+// TestIntegration_SinceID_BoundaryRowSameMicrosecondNotSkipped isolates the
+// cursor-boundary bug: a row written in the SAME microsecond as the cursor
+// row (but with a higher seq) must still be returned. Pre-fix the strict
+// `created_at > cursor` filter silently dropped it.
+func TestIntegration_SinceID_BoundaryRowSameMicrosecondNotSkipped(t *testing.T) {
+	conn := integrationDB_ActivityDelegationA2A(t)
+	_ = conn
+	wsID := seedWorkspace(t, conn, "test-2151-sinceid-boundary")
+
+	tSame := time.Date(2026, 6, 4, 13, 0, 0, 0, time.UTC)
+	// Cursor row and the next row share the exact same created_at; the next
+	// row is inserted afterwards so it gets a higher seq.
+	cursorID := seedActivityRowAt(t, wsID, "boundary-cursor", tSame)
+	nextID := seedActivityRowAt(t, wsID, "boundary-next-same-us", tSame)
+
+	h := NewActivityHandler(nil)
+	c, w := newTestGinContext()
+	c.Params = gin.Params{{Key: "id", Value: wsID}}
+	q := c.Request.URL.Query()
+	q.Set("since_id", cursorID)
+	q.Set("type", "a2a_receive")
+	q.Set("limit", "10")
+	c.Request.URL.RawQuery = q.Encode()
+
+	h.List(c)
+	if w.Code != http.StatusOK {
+		t.Fatalf("List returned %d, want 200: %s", w.Code, w.Body.String())
+	}
+	var resp []map[string]interface{}
+	if err := json.Unmarshal(w.Body.Bytes(), &resp); err != nil {
+		t.Fatalf("unmarshal: %v", err)
+	}
+	if len(resp) != 1 {
+		t.Fatalf("same-microsecond boundary row dropped: expected exactly the 1 next row, got %d rows: %+v",
+			len(resp), resp)
+	}
+	if got, _ := resp[0]["id"].(string); got != nextID {
+		t.Fatalf("expected boundary row id %s, got %s", nextID, got)
+	}
+}

workspace-server/internal/handlers/activity_since_id_test.go

@@ -26,17 +26,21 @@ func TestActivityHandler_SinceID_ReturnsNewerASC(t *testing.T) {
 
 	cursorID := "act-cursor-42"
 	cursorTime := time.Date(2026, 4, 30, 5, 0, 0, 0, time.UTC)
+	cursorSeq := int64(42)
 
 	// Step 1: cursor lookup — must include workspace_id scope so a UUID
-	// from another workspace can't be used.
-	mock.ExpectQuery(`SELECT created_at FROM activity_logs WHERE id = \$1 AND workspace_id = \$2`).
+	// from another workspace can't be used. Now resolves BOTH ordering-key
+	// components (created_at, seq) so the strictly-after filter can compare
+	// the full tuple.
+	mock.ExpectQuery(`SELECT created_at, seq FROM activity_logs WHERE id = \$1 AND workspace_id = \$2`).
 		WithArgs(cursorID, "ws-1").
-		WillReturnRows(sqlmock.NewRows([]string{"created_at"}).AddRow(cursorTime))
+		WillReturnRows(sqlmock.NewRows([]string{"created_at", "seq"}).AddRow(cursorTime, cursorSeq))
 
-	// Step 2: main query with the cursor's created_at as a > filter,
-	// ASC ordering. Args: workspace_id, cursorTime, limit.
+	// Step 2: main query with the cursor's (created_at, seq) as a tuple
+	// strictly-after filter, (created_at, seq) ASC ordering.
+	// Args: workspace_id, cursorTime, cursorSeq, limit.
 	mock.ExpectQuery("SELECT id, workspace_id, activity_type").
-		WithArgs("ws-1", cursorTime, 100).
+		WithArgs("ws-1", cursorTime, cursorSeq, 100).
 		WillReturnRows(newActivityRows())
 
 	broadcaster := newTestBroadcaster()
@@ -64,7 +68,7 @@ func TestActivityHandler_SinceID_ReturnsNewerASC(t *testing.T) {
 func TestActivityHandler_SinceID_CursorNotFound_410(t *testing.T) {
 	mock := setupTestDB(t)
 
-	mock.ExpectQuery(`SELECT created_at FROM activity_logs WHERE id = \$1 AND workspace_id = \$2`).
+	mock.ExpectQuery(`SELECT created_at, seq FROM activity_logs WHERE id = \$1 AND workspace_id = \$2`).
 		WithArgs("act-gone", "ws-1").
 		WillReturnError(sql.ErrNoRows)
 
@@ -96,7 +100,7 @@ func TestActivityHandler_SinceID_CrossWorkspaceCursor_410(t *testing.T) {
 
 	// Cursor exists in DB but the WHERE workspace_id = $2 filter excludes
 	// it — sqlmock returns no rows, which is what Postgres would do.
-	mock.ExpectQuery(`SELECT created_at FROM activity_logs WHERE id = \$1 AND workspace_id = \$2`).
+	mock.ExpectQuery(`SELECT created_at, seq FROM activity_logs WHERE id = \$1 AND workspace_id = \$2`).
 		WithArgs("act-other-ws", "ws-1").
 		WillReturnError(sql.ErrNoRows)
 
@@ -120,20 +124,23 @@ func TestActivityHandler_SinceID_CrossWorkspaceCursor_410(t *testing.T) {
 
 // TestActivityHandler_SinceID_CombinedWithSinceSecs: both filters apply
 // together (AND). Argument order in the main query: workspace_id,
-// since_secs, cursorTime, limit. Sanity-checks the placeholder index
-// arithmetic in the query builder.
+// since_secs, cursorTime, cursorSeq, limit. Sanity-checks the placeholder
+// index arithmetic in the query builder (the cursor now binds TWO args —
+// the (created_at, seq) tuple — so since_secs no longer shifts the tail by
+// one but by two).
 func TestActivityHandler_SinceID_CombinedWithSinceSecs(t *testing.T) {
 	mock := setupTestDB(t)
 
 	cursorID := "act-c"
 	cursorTime := time.Date(2026, 4, 30, 4, 0, 0, 0, time.UTC)
+	cursorSeq := int64(7)
 
-	mock.ExpectQuery(`SELECT created_at FROM activity_logs WHERE id = \$1 AND workspace_id = \$2`).
+	mock.ExpectQuery(`SELECT created_at, seq FROM activity_logs WHERE id = \$1 AND workspace_id = \$2`).
 		WithArgs(cursorID, "ws-1").
-		WillReturnRows(sqlmock.NewRows([]string{"created_at"}).AddRow(cursorTime))
+		WillReturnRows(sqlmock.NewRows([]string{"created_at", "seq"}).AddRow(cursorTime, cursorSeq))
 
 	mock.ExpectQuery("SELECT id, workspace_id, activity_type").
-		WithArgs("ws-1", 600, cursorTime, 100).
+		WithArgs("ws-1", 600, cursorTime, cursorSeq, 100).
 		WillReturnRows(newActivityRows())
 
 	broadcaster := newTestBroadcaster()

workspace-server/internal/handlers/mcp.go

@@ -126,6 +126,32 @@ var mcpAllTools = []mcpTool{
 					"type":        "string",
 					"description": "The task description to send to the target workspace",
 				},
+				"attachments": map[string]interface{}{
+					"type":        "array",
+					"description": "Optional files to send with the task. Each item must include uri and name; mimeType and size are optional.",
+					"items": map[string]interface{}{
+						"type": "object",
+						"properties": map[string]interface{}{
+							"uri": map[string]interface{}{
+								"type":        "string",
+								"description": "Workspace attachment URI, usually workspace:/absolute/path",
+							},
+							"name": map[string]interface{}{
+								"type":        "string",
+								"description": "Display filename",
+							},
+							"mimeType": map[string]interface{}{
+								"type":        "string",
+								"description": "Optional MIME type",
+							},
+							"size": map[string]interface{}{
+								"type":        "number",
+								"description": "Optional file size in bytes",
+							},
+						},
+						"required": []string{"uri", "name"},
+					},
+				},
 			},
 			"required": []string{"workspace_id", "task"},
 		},
@@ -144,6 +170,32 @@ var mcpAllTools = []mcpTool{
 					"type":        "string",
 					"description": "The task description to send to the target workspace",
 				},
+				"attachments": map[string]interface{}{
+					"type":        "array",
+					"description": "Optional files to send with the task. Each item must include uri and name; mimeType and size are optional.",
+					"items": map[string]interface{}{
+						"type": "object",
+						"properties": map[string]interface{}{
+							"uri": map[string]interface{}{
+								"type":        "string",
+								"description": "Workspace attachment URI, usually workspace:/absolute/path",
+							},
+							"name": map[string]interface{}{
+								"type":        "string",
+								"description": "Display filename",
+							},
+							"mimeType": map[string]interface{}{
+								"type":        "string",
+								"description": "Optional MIME type",
+							},
+							"size": map[string]interface{}{
+								"type":        "number",
+								"description": "Optional file size in bytes",
+							},
+						},
+						"required": []string{"uri", "name"},
+					},
+				},
 			},
 			"required": []string{"workspace_id", "task"},
 		},

workspace-server/internal/handlers/mcp_test.go

@@ -285,6 +285,121 @@ func TestMCPHandler_DelegateTaskAsync_RoutesThroughPlatformA2AProxy(t *testing.T
 // goroutine returns early and never calls proxyA2ARequest with a nil/empty
 // body. Before the fix the goroutine logged the error and fell through,
 // dispatching a malformed A2A request.
+
+func TestMCPHandler_DelegateTask_WithAttachments(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(), "pending").
+		WillReturnResult(sqlmock.NewResult(1, 1))
+	mock.ExpectExec(`UPDATE activity_logs`).
+		WithArgs("dispatched", "", callerID, sqlmock.AnyArg()).
+		WillReturnResult(sqlmock.NewResult(0, 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)
+		}
+		bodyStr := string(body)
+		if !strings.Contains(bodyStr, `"text":"review this video"`) {
+			t.Fatalf("A2A body missing task text: %s", bodyStr)
+		}
+		if !strings.Contains(bodyStr, `"kind":"video"`) {
+			t.Fatalf("A2A body missing video attachment kind: %s", bodyStr)
+		}
+		if !strings.Contains(bodyStr, `"uri":"workspace:/tmp/clip.mp4"`) {
+			t.Fatalf("A2A body missing attachment uri: %s", bodyStr)
+		}
+		if !strings.Contains(bodyStr, `"mime_type":"video/mp4"`) {
+			t.Fatalf("A2A body missing attachment mime_type: %s", bodyStr)
+		}
+		return 200, []byte(`{"result":{"message":{"parts":[{"text":"done"}]}}}`), nil
+	}
+
+	out, err := h.toolDelegateTask(context.Background(), callerID, map[string]interface{}{
+		"workspace_id": targetID,
+		"task":         "review this video",
+		"attachments": []interface{}{
+			map[string]interface{}{
+				"uri":      "workspace:/tmp/clip.mp4",
+				"name":     "clip.mp4",
+				"mimeType": "video/mp4",
+				"size":     12345,
+			},
+		},
+	}, 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 err := mock.ExpectationsWereMet(); err != nil {
+		t.Fatalf("unmet expectations: %v", err)
+	}
+}
+
+func TestMCPHandler_DelegateTaskAsync_WithAttachments(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(), "pending").
+		WillReturnResult(sqlmock.NewResult(1, 1))
+	mock.ExpectExec(`UPDATE activity_logs`).
+		WithArgs("dispatched", "", callerID, sqlmock.AnyArg()).
+		WillReturnResult(sqlmock.NewResult(0, 1))
+
+	called := make(chan []byte, 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)
+		}
+		called <- body
+		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 with image",
+		"attachments": []interface{}{
+			map[string]interface{}{
+				"uri":      "workspace:/tmp/screenshot.png",
+				"name":     "screenshot.png",
+				"mimeType": "image/png",
+			},
+		},
+	})
+	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 body := <-called:
+		bodyStr := string(body)
+		if !strings.Contains(bodyStr, `"kind":"image"`) {
+			t.Fatalf("A2A body missing image attachment kind: %s", bodyStr)
+		}
+		if !strings.Contains(bodyStr, `"uri":"workspace:/tmp/screenshot.png"`) {
+			t.Fatalf("A2A body missing attachment uri: %s", bodyStr)
+		}
+	default:
+		t.Fatal("async delegate did not call platform A2A proxy")
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Fatalf("unmet expectations: %v", err)
+	}
+}
 func TestMCPHandler_DelegateTaskAsync_MarshalFailureDoesNotCallProxy(t *testing.T) {
 	h, mock := newMCPHandler(t)
 	callerID := "11111111-1111-1111-1111-111111111111"

workspace-server/internal/handlers/mcp_tools.go

@@ -187,6 +187,28 @@ func (h *MCPHandler) toolGetWorkspaceInfo(ctx context.Context, workspaceID strin
 	return string(b), nil
 }
 
+// buildA2AMessageParts constructs the A2A message parts array from a task string
+// and optional attachments. The text part always comes first; attachment parts
+// follow in the order provided, with kind derived from MIME type.
+func buildA2AMessageParts(task string, attachments []AgentMessageAttachment) []map[string]interface{} {
+	parts := []map[string]interface{}{
+		{"type": "text", "text": task},
+	}
+	for _, att := range attachments {
+		kind := kindFromMimeType(att.MimeType)
+		filePart := map[string]interface{}{
+			"kind": kind,
+			"file": map[string]interface{}{
+				"uri":       att.URI,
+				"mime_type": att.MimeType,
+				"name":      att.Name,
+			},
+		}
+		parts = append(parts, filePart)
+	}
+	return parts
+}
+
 func (h *MCPHandler) toolDelegateTask(ctx context.Context, callerID string, args map[string]interface{}, timeout time.Duration) (string, error) {
 	targetID, _ := args["workspace_id"].(string)
 	task, _ := args["task"].(string)
@@ -208,6 +230,8 @@ func (h *MCPHandler) toolDelegateTask(ctx context.Context, callerID string, args
 		// Non-fatal: still make the A2A call even if activity log write fails.
 	}
 
+	attachments, _ := parseAgentMessageAttachments(args["attachments"])
+
 	a2aBody, err := json.Marshal(map[string]interface{}{
 		"jsonrpc": "2.0",
 		"id":      uuid.New().String(),
@@ -215,7 +239,7 @@ func (h *MCPHandler) toolDelegateTask(ctx context.Context, callerID string, args
 		"params": map[string]interface{}{
 			"message": map[string]interface{}{
 				"role":      "user",
-				"parts":     []map[string]interface{}{{"type": "text", "text": task}},
+				"parts":     buildA2AMessageParts(task, attachments),
 				"messageId": uuid.New().String(),
 			},
 		},
@@ -275,6 +299,8 @@ func (h *MCPHandler) toolDelegateTaskAsync(ctx context.Context, callerID string,
 		bgCtx, cancel := context.WithTimeout(context.Background(), mcpAsyncCallTimeout)
 		defer cancel()
 
+		attachments, _ := parseAgentMessageAttachments(args["attachments"])
+
 		a2aBody, marshalErr := marshalA2ABody(map[string]interface{}{
 			"jsonrpc": "2.0",
 			"id":      delegationID,
@@ -282,7 +308,7 @@ func (h *MCPHandler) toolDelegateTaskAsync(ctx context.Context, callerID string,
 			"params": map[string]interface{}{
 				"message": map[string]interface{}{
 					"role":      "user",
-					"parts":     []map[string]interface{}{{"type": "text", "text": task}},
+					"parts":     buildA2AMessageParts(task, attachments),
 					"messageId": uuid.New().String(),
 				},
 			},

workspace-server/internal/handlers/model_registry_validation.go

@@ -24,13 +24,23 @@ import (
 // validateRegisteredModelForRuntime reports whether (runtime, model) is
 // selectable per the provider registry. Returns:
 //
-//	(true,  "")     — allowed: model is registered for this runtime, OR the
-//	                  runtime is not in the registry (fail-open), OR model=="".
-//	(false, reason) — rejected: the runtime IS registered but the model is not
-//	                  in its native ModelsForRuntime set.
+//	(true,  "")     — allowed: model is on the runtime's platform menu
+//	                  (ModelsForRuntime) OR DeriveProvider(runtime, model)
+//	                  RESOLVES a native provider (the cp#529 routability-aware
+//	                  BYOK path), OR the runtime is not in the registry
+//	                  (fail-open), OR model=="".
+//	(false, reason) — rejected: the runtime IS registered, the model is not on
+//	                  its platform menu, AND no native provider prefix-owns it
+//	                  (genuinely unroutable).
 //
 // model=="" is allowed here: the MODEL_REQUIRED gate owns the empty-model case,
 // so this validator must not double-reject it.
+//
+// ROUTABILITY-AWARE (cp#529, CTO Option C): the final predicate is an OR —
+// `model ∈ ModelsForRuntime(runtime)` OR `DeriveProvider(runtime, model, nil)`
+// resolves. The platform menu carries platform-billed ids; the DeriveProvider
+// path covers BYOK ids that prefix-match a name-only native arm (no platform
+// billing). The drift checker in molecule-controlplane mirrors this exact OR.
 func validateRegisteredModelForRuntime(runtime, model string) (bool, string) {
 	model = strings.TrimSpace(model)
 	if model == "" {
@@ -52,6 +62,24 @@ func validateRegisteredModelForRuntime(runtime, model string) (bool, string) {
 			return true, ""
 		}
 	}
+	// ROUTABILITY-AWARE allow path (cp#529, CTO-approved Option C). The model is
+	// NOT on the runtime's platform menu (ModelsForRuntime) — but a model can be
+	// legitimately SELECTABLE without being a platform-menu id: a BYOK id whose
+	// prefix matches one of the runtime's NATIVE provider arms (a name-only arm
+	// added in providers.yaml) resolves to a concrete provider via DeriveProvider
+	// even though it carries no platform billing. Allow it iff DeriveProvider
+	// resolves a provider for (runtime, model). A genuinely-unroutable id (no
+	// native provider prefix-owns it) still falls through to the 422 below.
+	//
+	// BILLING GUARDRAIL: only CONFIRMED-NON-PLATFORM (BYOK) providers are wired as
+	// name-only arms in providers.yaml (never platform/anthropic-*/openai-*/
+	// moonshot/minimax/google/vertex), so a DeriveProvider-resolved id reached by
+	// THIS path can never bill the platform's key for a customer's model. The
+	// platform-menu ids that DO carry platform billing are already allowed by the
+	// exact-membership loop above; this path only ever resolves to a BYOK arm.
+	if _, derr := m.DeriveProvider(runtime, model, nil); derr == nil {
+		return true, ""
+	}
 	return false, fmt.Sprintf(
 		"model %q is not a registered model for runtime %q; pick one of the runtime's registered models (provider-registry SSOT, internal#718)",
 		model, runtime)

workspace-server/internal/handlers/model_registry_validation_test.go

@@ -79,6 +79,50 @@ func TestValidateRegisteredModelForRuntime(t *testing.T) {
 			model:   "",
 			wantOK:  true,
 		},
+		// ---- cp#529 routability-aware allow path -------------------------------
+		{
+			// BYOK passthrough id: NOT on hermes's platform menu, but the
+			// openrouter name-only native arm prefix-owns it → DeriveProvider
+			// resolves → ALLOWED (no platform billing — openrouter is BYOK).
+			name:    "byok_passthrough_routable_now_allowed",
+			runtime: "hermes",
+			model:   "openrouter/anthropic/claude-3.5-sonnet",
+			wantOK:  true,
+		},
+		{
+			// BYOK namespaced vendor id: deepseek's widened ^deepseek[-:/]
+			// matches the vendor/ form on a name-only hermes arm → allowed.
+			name:    "byok_namespaced_vendor_routable_now_allowed",
+			runtime: "hermes",
+			model:   "deepseek/deepseek-chat",
+			wantOK:  true,
+		},
+		{
+			// claude-code bare GLM- BYOK id: zai name-only arm + (?i)^(glm-|…)
+			// matches → DeriveProvider resolves → allowed.
+			name:    "claude_code_bare_glm_byok_routable_now_allowed",
+			runtime: "claude-code",
+			model:   "GLM-4.6",
+			wantOK:  true,
+		},
+		{
+			// Genuinely UNROUTABLE id: no native hermes arm prefix-owns bare
+			// gpt-4o (the platform-shared openai vendor is NOT wired into hermes
+			// — billing guardrail), so DeriveProvider errors → still 422.
+			name:    "genuinely_unroutable_still_rejected",
+			runtime: "hermes",
+			model:   "gpt-4o",
+			wantOK:  false,
+		},
+		{
+			// A namespaced vendor id NOW routable on hermes via the dedicated
+			// byok-openai provider (cp#529 BYOK-vendor arms): routes with the
+			// tenant's OPENAI_API_KEY → BYOK billing, never the platform key.
+			name:    "byok_openai_namespaced_routable_now_allowed",
+			runtime: "hermes",
+			model:   "openai/gpt-4o",
+			wantOK:  true,
+		},
 	}
 	for _, c := range cases {
 		t.Run(c.name, func(t *testing.T) {
@@ -109,58 +153,58 @@ func TestValidateDerivedProviderInRegistry(t *testing.T) {
 		// provider that IS in the providers list. These are the live corpus
 		// entries; the test pins the registry-consistency invariant.
 		{
-			name:   "claude_code_anthropic_api_native",
+			name:    "claude_code_anthropic_api_native",
 			runtime: "claude-code",
 			model:   "claude-sonnet-4-6",
-			wantOK: true,
+			wantOK:  true,
 		},
 		{
-			name:   "claude_code_kimi_coding_native",
+			name:    "claude_code_kimi_coding_native",
 			runtime: "claude-code",
 			model:   "kimi-for-coding",
-			wantOK: true,
+			wantOK:  true,
 		},
 		{
-			name:   "claude_code_minimax_native",
+			name:    "claude_code_minimax_native",
 			runtime: "claude-code",
 			model:   "MiniMax-M2.7",
-			wantOK: true,
+			wantOK:  true,
 		},
 		{
-			name:   "claude_code_platform_namespaced",
+			name:    "claude_code_platform_namespaced",
 			runtime: "claude-code",
 			model:   "moonshot/kimi-k2.6",
-			wantOK: true,
+			wantOK:  true,
 		},
 		{
-			name:   "codex_openai_subscription_default_arm",
+			name:    "codex_openai_subscription_default_arm",
 			runtime: "codex",
 			model:   "gpt-5.5",
-			wantOK: true,
+			wantOK:  true,
 		},
 		{
-			name:   "codex_platform_namespaced",
+			name:    "codex_platform_namespaced",
 			runtime: "codex",
 			model:   "openai/gpt-5.4-mini",
-			wantOK: true,
+			wantOK:  true,
 		},
 		{
-			name:   "hermes_kimi_coding",
+			name:    "hermes_kimi_coding",
 			runtime: "hermes",
 			model:   "kimi-coding/kimi-k2",
-			wantOK: true,
+			wantOK:  true,
 		},
 		{
-			name:   "hermes_platform_namespaced",
+			name:    "hermes_platform_namespaced",
 			runtime: "hermes",
 			model:   "moonshot/kimi-k2.6",
-			wantOK: true,
+			wantOK:  true,
 		},
 		{
-			name:   "openclaw_kimi_coding",
+			name:    "openclaw_kimi_coding",
 			runtime: "openclaw",
 			model:   "moonshot:kimi-k2.6",
-			wantOK: true,
+			wantOK:  true,
 		},
 		// FAIL — model-side validator catches this, but the provider-side
 		// gate is called AFTER it in Create and inherits the fail-open
@@ -168,30 +212,30 @@ func TestValidateDerivedProviderInRegistry(t *testing.T) {
 		// errors → allow, letting the model-side response own the message).
 		// This is the deliberate "don't double-reject" decision.
 		{
-			name:   "unregistered_model_pass_through_to_model_side",
+			name:    "unregistered_model_pass_through_to_model_side",
 			runtime: "claude-code",
 			model:   "totally-made-up-model-xyz",
-			wantOK: true, // pass-through: model-side validator owns the rejection
+			wantOK:  true, // pass-through: model-side validator owns the rejection
 		},
 		// Federation contract — mirror of the model-side test above.
 		{
-			name:   "langgraph_runtime_failopen",
+			name:    "langgraph_runtime_failopen",
 			runtime: "langgraph",
 			model:   "anything-goes",
-			wantOK: true,
+			wantOK:  true,
 		},
 		{
-			name:   "external_runtime_failopen",
+			name:    "external_runtime_failopen",
 			runtime: "external",
 			model:   "whatever",
-			wantOK: true,
+			wantOK:  true,
 		},
 		// Empty model — MODEL_REQUIRED owns it; allow.
 		{
-			name:   "empty_model_allowed_other_gate_owns_it",
+			name:    "empty_model_allowed_other_gate_owns_it",
 			runtime: "claude-code",
 			model:   "",
-			wantOK: true,
+			wantOK:  true,
 		},
 	}
 	for _, c := range cases {

workspace-server/internal/handlers/plugins_install.go

@@ -161,7 +161,7 @@ func (h *PluginsHandler) uninstallViaDocker(ctx context.Context, c *gin.Context,
 	// 1. Strip plugin's rule/fragment markers from CLAUDE.md (mirrors
 	//    AgentskillsAdaptor.uninstall lines 184-188). Best-effort: if
 	//    the user edited CLAUDE.md, our marker stays untouched.
-	h.stripPluginMarkersFromMemory(ctx, containerName, pluginName)
+	h.stripPluginMarkersFromMemory(ctx, workspaceID, containerName, pluginName)
 
 	// 2. Remove copied skill dirs declared in the plugin's plugin.yaml.
 	for _, skill := range skillNames {
@@ -171,9 +171,11 @@ func (h *PluginsHandler) uninstallViaDocker(ctx context.Context, c *gin.Context,
 			log.Printf("Plugin uninstall: skipping invalid skill name %q in %s: %v", skill, pluginName, err)
 			continue
 		}
-		_, _ = h.execAsRoot(ctx, containerName, []string{
+		if _, rmErr := h.execAsRoot(ctx, containerName, []string{
 			"rm", "-rf", "/configs/skills/" + skill,
-		})
+		}); rmErr != nil {
+			log.Printf("Plugin uninstall: failed to remove skill %s from %s: %v", skill, workspaceID, rmErr)
+		}
 	}
 
 	// 3. Delete the plugin directory itself (as root to handle file ownership).

workspace-server/internal/handlers/plugins_install_pipeline.go

@@ -393,7 +393,7 @@ func (h *PluginsHandler) readPluginSkillsFromContainer(ctx context.Context, cont
 // `# Plugin: <name> /` — mirrors AgentskillsAdaptor.uninstall's stripping
 // logic so install/uninstall are symmetric. Best-effort: silent on read or
 // write failure, since the rest of uninstall must still succeed.
-func (h *PluginsHandler) stripPluginMarkersFromMemory(ctx context.Context, containerName, pluginName string) {
+func (h *PluginsHandler) stripPluginMarkersFromMemory(ctx context.Context, workspaceID, containerName, pluginName string) {
 	// Use sed via bash -c for atomic in-place delete: drop the marker line
 	// and the blank line that follows it (install adds a leading blank line
 	// before the marker via append_to_memory). Three sed passes mirror the
@@ -417,7 +417,9 @@ func (h *PluginsHandler) stripPluginMarkersFromMemory(ctx context.Context, conta
 		`awk 'BEGIN{skip=0; blanks=0} /^%s/{skip=1; blanks=0; next} skip==1 && /^[[:space:]]*$/{blanks++; if(blanks>=2){skip=0; print; next} next} /^# Plugin: /{if(skip==1)skip=0} skip==1{next} {print}' /configs/CLAUDE.md > /tmp/claude.new && mv /tmp/claude.new /configs/CLAUDE.md`,
 		regexpEscapeForAwk(marker),
 	)
-	_, _ = h.execAsRoot(ctx, containerName, []string{"bash", "-c", script})
+	if _, awkErr := h.execAsRoot(ctx, containerName, []string{"bash", "-c", script}); awkErr != nil {
+		log.Printf("Plugin uninstall: failed to strip markers from CLAUDE.md for %s in %s: %v", pluginName, workspaceID, awkErr)
+	}
 }
 
 // regexpEscapeForAwk escapes characters that have special meaning inside an

workspace-server/internal/handlers/workspace_display_session_proxy_test.go

@@ -0,0 +1,331 @@
+package handlers
+
+import (
+	"context"
+	"database/sql"
+	"net/http"
+	"net/http/httptest"
+	"net/url"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/DATA-DOG/go-sqlmock"
+	"github.com/gin-gonic/gin"
+	"github.com/gorilla/websocket"
+)
+
+// rfbGreeting is the first frame a real websockify/RFB backend writes on
+// connect. The fake backend below sends these exact bytes so the positive
+// test can prove the upstream's first binary frame survives the reverse
+// proxy chain (the "WS 1006" regression surface from core#2247 was the
+// upgrade/handshake silently failing before any RFB byte reached the
+// browser).
+var rfbGreeting = []byte("RFB 003.008\n")
+
+// newFakeWebsockifyBackend stands up an httptest.NewServer that upgrades the
+// websocket, writes the RFB greeting as a binary frame, then echoes every
+// frame it receives back to the client. No EC2, noVNC, or SSH involved — it
+// is the stand-in for the on-instance :6080 websockify listener that
+// realDisplayForward would normally tunnel to.
+func newFakeWebsockifyBackend(t *testing.T) *httptest.Server {
+	t.Helper()
+	upgrader := websocket.Upgrader{
+		// The proxy rewrites Sec-WebSocket-Protocol to "binary"; accept any
+		// origin/subprotocol so the fake backend never rejects the handshake.
+		CheckOrigin:       func(*http.Request) bool { return true },
+		Subprotocols:      []string{"binary"},
+		HandshakeTimeout:  5 * time.Second,
+		EnableCompression: false,
+	}
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		conn, err := upgrader.Upgrade(w, r, nil)
+		if err != nil {
+			return
+		}
+		defer conn.Close()
+		if err := conn.WriteMessage(websocket.BinaryMessage, rfbGreeting); err != nil {
+			return
+		}
+		for {
+			mt, msg, err := conn.ReadMessage()
+			if err != nil {
+				return
+			}
+			if err := conn.WriteMessage(mt, msg); err != nil {
+				return
+			}
+		}
+	}))
+	t.Cleanup(srv.Close)
+	return srv
+}
+
+// wireDisplayForwardToBackend overrides the injectable displayForward package
+// var so DisplaySession proxies to the fake backend instead of opening an EIC
+// SSH tunnel. Restored via t.Cleanup. The returned *url.URL is the http://
+// backend address (the reverse proxy upgrades it to ws:// natively under
+// Go 1.25's ReverseProxy WebSocket support).
+func wireDisplayForwardToBackend(t *testing.T, backendURL string) {
+	t.Helper()
+	target, err := url.Parse(backendURL)
+	if err != nil {
+		t.Fatalf("parse backend URL %q: %v", backendURL, err)
+	}
+	prev := displayForward
+	displayForward = func(_ context.Context, _ string, fn func(target *url.URL) error) error {
+		return fn(target)
+	}
+	t.Cleanup(func() { displayForward = prev })
+}
+
+// newDisplaySessionTestServer mounts DisplaySession on a gin router behind an
+// httptest.NewServer so a real websocket client can dial the route end-to-end.
+// It returns the base ws:// URL for the websockify route.
+func newDisplaySessionTestServer(t *testing.T, handler *WorkspaceHandler) *httptest.Server {
+	t.Helper()
+	r := gin.New()
+	// Mirror the production registration in internal/router/router.go:
+	//   GET /workspaces/:id/display/session/*proxyPath -> wh.DisplaySession
+	r.GET("/workspaces/:id/display/session/*proxyPath", handler.DisplaySession)
+	srv := httptest.NewServer(r)
+	t.Cleanup(srv.Close)
+	return srv
+}
+
+const (
+	displayProxyWorkspaceID  = "ws-display"
+	displayProxyInstanceID   = "i-0fakedeadbeef00001"
+	displayProxyControlledBy = "admin-token"
+)
+
+// expectDisplaySessionTargetRow mocks loadWorkspaceDisplaySessionTarget's
+// workspaces SELECT. mode "desktop-control" + a non-empty instance_id is the
+// "display enabled, tunnel available" shape. (Note: the compute validator
+// accepts modes none/desktop-control/gpu-desktop-control and protocols
+// dcv/novnc — "novnc" is a *protocol*, not a mode, so the enabled rows use
+// mode=desktop-control,protocol=novnc.)
+func expectDisplaySessionTargetRow(mock sqlmock.Sqlmock, computeJSON, instanceID string) {
+	mock.ExpectQuery(`SELECT COALESCE\(compute, '\{\}'::jsonb\), COALESCE\(instance_id, ''\) FROM workspaces WHERE id = \$1`).
+		WithArgs(displayProxyWorkspaceID).
+		WillReturnRows(sqlmock.NewRows([]string{"compute", "instance_id"}).AddRow(computeJSON, instanceID))
+}
+
+// expectActiveDisplayControlRow mocks loadActiveDisplayControl's locks SELECT
+// returning an active lock owned by controlledBy expiring at expiresAt.
+func expectActiveDisplayControlRow(mock sqlmock.Sqlmock, controlledBy string, expiresAt time.Time) {
+	mock.ExpectQuery(`SELECT controller, controlled_by, expires_at FROM workspace_display_control_locks WHERE workspace_id = \$1 AND expires_at > now\(\)`).
+		WithArgs(displayProxyWorkspaceID).
+		WillReturnRows(sqlmock.NewRows([]string{"controller", "controlled_by", "expires_at"}).
+			AddRow("user", controlledBy, expiresAt))
+}
+
+const enabledComputeJSON = `{"display":{"mode":"desktop-control","protocol":"novnc","width":1280,"height":800}}`
+
+// dialDisplaySession dials the websockify route on the given test server with
+// the supplied Sec-WebSocket-Protocol values. It returns the conn (nil on
+// failure), the HTTP response, and the dial error.
+func dialDisplaySession(t *testing.T, srv *httptest.Server, subprotocols []string) (*websocket.Conn, *http.Response, error) {
+	t.Helper()
+	wsURL := "ws" + strings.TrimPrefix(srv.URL, "http") + "/workspaces/" + displayProxyWorkspaceID + "/display/session/websockify"
+	dialer := websocket.Dialer{
+		HandshakeTimeout: 5 * time.Second,
+		Subprotocols:     subprotocols,
+	}
+	return dialer.Dial(wsURL, nil)
+}
+
+// TestDisplaySessionProxy_Positive proves the full take-control WS-proxy path
+// without any network/EC2: a valid signed token + active lock + enabled
+// display upgrades successfully (HTTP 101), the backend's RFB greeting arrives
+// through the proxy, and a client->server byte round-trips back (bidirectional
+// proxy chain). This is the direct regression guard for the "WS 1006" failure
+// class in core#2247.
+func TestDisplaySessionProxy_Positive(t *testing.T) {
+	t.Setenv("DISPLAY_SESSION_SIGNING_SECRET", "test-secret")
+	mock := setupTestDB(t)
+	backend := newFakeWebsockifyBackend(t)
+	wireDisplayForwardToBackend(t, backend.URL)
+
+	handler := NewWorkspaceHandler(newTestBroadcaster(), nil, "http://localhost:8080", t.TempDir())
+	srv := newDisplaySessionTestServer(t, handler)
+
+	expiresAt := time.Now().Add(5 * time.Minute)
+	expectDisplaySessionTargetRow(mock, enabledComputeJSON, displayProxyInstanceID)
+	expectActiveDisplayControlRow(mock, displayProxyControlledBy, expiresAt)
+
+	token := signDisplaySessionToken(displayProxyWorkspaceID, displayProxyControlledBy, expiresAt)
+	if token == "" {
+		t.Fatal("signDisplaySessionToken returned empty token")
+	}
+
+	conn, resp, err := dialDisplaySession(t, srv, []string{"binary", displaySessionTokenProtocolPrefix + token})
+	if err != nil {
+		body := ""
+		if resp != nil {
+			body = resp.Status
+		}
+		t.Fatalf("websocket dial failed: %v (resp=%s)", err, body)
+	}
+	t.Cleanup(func() { conn.Close() })
+	if resp.StatusCode != http.StatusSwitchingProtocols {
+		t.Fatalf("expected 101 Switching Protocols, got %d", resp.StatusCode)
+	}
+
+	// 1. The backend's RFB greeting must arrive through the proxy.
+	conn.SetReadDeadline(time.Now().Add(5 * time.Second))
+	mt, msg, err := conn.ReadMessage()
+	if err != nil {
+		t.Fatalf("read greeting through proxy failed: %v", err)
+	}
+	if mt != websocket.BinaryMessage || string(msg) != string(rfbGreeting) {
+		t.Fatalf("greeting = %q (type %d), want %q binary", msg, mt, rfbGreeting)
+	}
+
+	// 2. A client->server byte must echo back (bidirectional chain).
+	probe := []byte{0x13, 0x37, 0x00, 0xff}
+	if err := conn.WriteMessage(websocket.BinaryMessage, probe); err != nil {
+		t.Fatalf("write probe through proxy failed: %v", err)
+	}
+	conn.SetReadDeadline(time.Now().Add(5 * time.Second))
+	_, echo, err := conn.ReadMessage()
+	if err != nil {
+		t.Fatalf("read echo through proxy failed: %v", err)
+	}
+	if string(echo) != string(probe) {
+		t.Fatalf("echo = %q, want %q", echo, probe)
+	}
+
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Errorf("unmet sqlmock expectations: %v", err)
+	}
+}
+
+// TestDisplaySessionProxy_Rejections is table-driven over the failure surface.
+// Each case asserts the WS upgrade does NOT happen (dial errors / no 101) and
+// the right HTTP status is returned, WITHOUT ever reaching the fake backend.
+func TestDisplaySessionProxy_Rejections(t *testing.T) {
+	t.Setenv("DISPLAY_SESSION_SIGNING_SECRET", "test-secret")
+	pastExpiry := time.Now().Add(-5 * time.Minute)
+	futureExpiry := time.Now().Add(5 * time.Minute)
+
+	cases := []struct {
+		name string
+		// expect wires the sqlmock rows that the handler will actually read
+		// for this case (the locks SELECT is only reached for token cases).
+		expect func(mock sqlmock.Sqlmock)
+		// subprotocols sent on the dial (token header, if any).
+		subprotocols []string
+		// proxyPath overrides the default "/websockify" route segment.
+		proxyPath  string
+		wantStatus int
+	}{
+		{
+			name: "missing token -> 403",
+			expect: func(m sqlmock.Sqlmock) {
+				expectDisplaySessionTargetRow(m, enabledComputeJSON, displayProxyInstanceID)
+				expectActiveDisplayControlRow(m, displayProxyControlledBy, futureExpiry)
+			},
+			subprotocols: []string{"binary"},
+			wantStatus:   http.StatusForbidden,
+		},
+		{
+			name: "tampered token -> 403",
+			expect: func(m sqlmock.Sqlmock) {
+				expectDisplaySessionTargetRow(m, enabledComputeJSON, displayProxyInstanceID)
+				expectActiveDisplayControlRow(m, displayProxyControlledBy, futureExpiry)
+			},
+			subprotocols: []string{"binary", displaySessionTokenProtocolPrefix + "garbage.not-a-valid-mac"},
+			wantStatus:   http.StatusForbidden,
+		},
+		{
+			name: "expired lock -> 403",
+			expect: func(m sqlmock.Sqlmock) {
+				expectDisplaySessionTargetRow(m, enabledComputeJSON, displayProxyInstanceID)
+				// Active-lock query filters expires_at > now(), so an
+				// expired lock returns no rows -> found=false -> 403.
+				m.ExpectQuery(`SELECT controller, controlled_by, expires_at FROM workspace_display_control_locks WHERE workspace_id = \$1 AND expires_at > now\(\)`).
+					WithArgs(displayProxyWorkspaceID).
+					WillReturnError(sql.ErrNoRows)
+			},
+			// Token signed against the past expiry would also fail validation
+			// even if a stale lock row were returned.
+			subprotocols: []string{"binary", displaySessionTokenProtocolPrefix +
+				signDisplaySessionToken(displayProxyWorkspaceID, displayProxyControlledBy, pastExpiry)},
+			wantStatus: http.StatusForbidden,
+		},
+		{
+			name: "display mode none -> 404",
+			expect: func(m sqlmock.Sqlmock) {
+				expectDisplaySessionTargetRow(m, `{"display":{"mode":"none"}}`, displayProxyInstanceID)
+			},
+			subprotocols: []string{"binary"},
+			wantStatus:   http.StatusNotFound,
+		},
+		{
+			name: "empty instance_id -> 503",
+			expect: func(m sqlmock.Sqlmock) {
+				expectDisplaySessionTargetRow(m, enabledComputeJSON, "")
+			},
+			subprotocols: []string{"binary"},
+			wantStatus:   http.StatusServiceUnavailable,
+		},
+		{
+			name: "wrong proxyPath -> 404",
+			expect: func(m sqlmock.Sqlmock) {
+				expectDisplaySessionTargetRow(m, enabledComputeJSON, displayProxyInstanceID)
+			},
+			subprotocols: []string{"binary"},
+			proxyPath:    "/frames",
+			wantStatus:   http.StatusNotFound,
+		},
+	}
+
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			mock := setupTestDB(t)
+			// A backend that fatals if it is ever reached — proves these
+			// rejections happen strictly before any proxy dial.
+			reached := false
+			backend := httptest.NewServer(http.HandlerFunc(func(http.ResponseWriter, *http.Request) {
+				reached = true
+			}))
+			t.Cleanup(backend.Close)
+			wireDisplayForwardToBackend(t, backend.URL)
+
+			handler := NewWorkspaceHandler(newTestBroadcaster(), nil, "http://localhost:8080", t.TempDir())
+			srv := newDisplaySessionTestServer(t, handler)
+			tc.expect(mock)
+
+			proxyPath := tc.proxyPath
+			if proxyPath == "" {
+				proxyPath = "/websockify"
+			}
+			wsURL := "ws" + strings.TrimPrefix(srv.URL, "http") +
+				"/workspaces/" + displayProxyWorkspaceID + "/display/session" + proxyPath
+			dialer := websocket.Dialer{HandshakeTimeout: 5 * time.Second, Subprotocols: tc.subprotocols}
+			conn, resp, err := dialer.Dial(wsURL, nil)
+			if conn != nil {
+				conn.Close()
+			}
+			if err == nil {
+				t.Fatalf("expected WS upgrade to fail, but dial succeeded")
+			}
+			if resp == nil {
+				t.Fatalf("expected an HTTP response on rejected upgrade, got nil (err=%v)", err)
+			}
+			if resp.StatusCode != tc.wantStatus {
+				t.Fatalf("status = %d, want %d", resp.StatusCode, tc.wantStatus)
+			}
+			if resp.StatusCode == http.StatusSwitchingProtocols {
+				t.Fatalf("upgrade unexpectedly succeeded (101)")
+			}
+			if reached {
+				t.Fatalf("rejection leaked to the upstream backend")
+			}
+			if err := mock.ExpectationsWereMet(); err != nil {
+				t.Errorf("unmet sqlmock expectations: %v", err)
+			}
+		})
+	}
+}

workspace-server/internal/providers/derive_provider_test.go

@@ -99,10 +99,16 @@ func TestDeriveProvider_UnregisteredErrors(t *testing.T) {
 		runtime string
 		model   string
 	}{
-		// gpt-* is OpenAI — not in claude-code's native set.
+		// gpt-* is OpenAI — not in claude-code's native set (no openai arm;
+		// the platform-shared openai vendor is never wired into a BYOK runtime).
 		{"claude-code", "gpt-5.5"},
-		// deepseek is a catalog provider but in NO runtime's native set.
-		{"claude-code", "deepseek-v4-pro"},
+		// qwen-* is alibaba — a catalog provider NOT wired into claude-code
+		// (cp#529 wires alibaba only into hermes; claude-code's name-only BYOK
+		// arms are zai/deepseek/xiaomi-mimo). So it stays unregistered here.
+		// (NB: deepseek-* IS now routable on claude-code via the deepseek
+		// name-only arm — see the routability tests — so it is no longer a valid
+		// "unregistered" example; qwen replaces it.)
+		{"claude-code", "qwen-max"},
 		// codex is OpenAI-only — a kimi id is unregistered for it.
 		{"codex", "kimi-for-coding"},
 		// a slug no provider in the manifest matches at all.

workspace-server/internal/providers/gen/registry_gen.go

@@ -16,7 +16,7 @@ const SchemaVersion = 1
 // Fingerprint is a stable content hash of the generated projection (schema
 // version + provider catalog + runtime native sets). It changes iff the
 // registry DATA changes (comment-only YAML edits do not churn it).
-const Fingerprint = "a491f5ff8a17ef59"
+const Fingerprint = "ec6b93409e7b9cf8"
 
 // GenProvider is the generated projection of one provider catalog entry —
 // the subset a downstream consumer needs to derive + display a provider.
@@ -51,26 +51,31 @@ var Providers = []GenProvider{
 	{Name: "moonshot", DisplayName: "Moonshot (Kimi)", Protocol: "openai", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"MOONSHOT_API_KEY", "KIMI_API_KEY"}, ModelPrefixMatch: "^moonshot[:/-]", IsPlatform: false, UpstreamVendor: "moonshot"},
 	{Name: "minimax", DisplayName: "MiniMax", Protocol: "openai", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"MINIMAX_API_KEY", "ANTHROPIC_AUTH_TOKEN", "ANTHROPIC_API_KEY"}, ModelPrefixMatch: "(?i)^minimax-m", IsPlatform: false, UpstreamVendor: "minimax"},
 	{Name: "platform", DisplayName: "Platform", Protocol: "anthropic", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"MOLECULE_LLM_USAGE_TOKEN"}, ModelPrefixMatch: "^platform/", IsPlatform: true},
-	{Name: "xiaomi-mimo", DisplayName: "Xiaomi MiMo", Protocol: "anthropic", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"ANTHROPIC_AUTH_TOKEN", "ANTHROPIC_API_KEY"}, ModelPrefixMatch: "^mimo-", IsPlatform: false},
-	{Name: "zai", DisplayName: "Z.ai (GLM)", Protocol: "anthropic", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"GLM_API_KEY", "ANTHROPIC_AUTH_TOKEN", "ANTHROPIC_API_KEY"}, ModelPrefixMatch: "(?i)^glm-", IsPlatform: false},
+	{Name: "xiaomi-mimo", DisplayName: "Xiaomi MiMo", Protocol: "anthropic", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"ANTHROPIC_AUTH_TOKEN", "ANTHROPIC_API_KEY"}, ModelPrefixMatch: "(?i)^(mimo-|xiaomi[:/])", IsPlatform: false},
+	{Name: "zai", DisplayName: "Z.ai (GLM)", Protocol: "anthropic", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"GLM_API_KEY", "ANTHROPIC_AUTH_TOKEN", "ANTHROPIC_API_KEY"}, ModelPrefixMatch: "(?i)^(glm-|zai[:/])", IsPlatform: false},
 	{Name: "kimi-coding", DisplayName: "Moonshot Kimi (coding-tuned)", Protocol: "anthropic", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"KIMI_API_KEY", "ANTHROPIC_API_KEY", "ANTHROPIC_AUTH_TOKEN"}, ModelPrefixMatch: "^kimi-", IsPlatform: false},
-	{Name: "deepseek", DisplayName: "DeepSeek", Protocol: "anthropic", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"DEEPSEEK_API_KEY", "ANTHROPIC_AUTH_TOKEN", "ANTHROPIC_API_KEY"}, ModelPrefixMatch: "^deepseek-", IsPlatform: false},
+	{Name: "deepseek", DisplayName: "DeepSeek", Protocol: "anthropic", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"DEEPSEEK_API_KEY", "ANTHROPIC_AUTH_TOKEN", "ANTHROPIC_API_KEY"}, ModelPrefixMatch: "^deepseek[-:/]", IsPlatform: false},
 	{Name: "google", DisplayName: "Google Gemini", Protocol: "openai", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"GEMINI_API_KEY", "GOOGLE_API_KEY"}, ModelPrefixMatch: "^gemini-", IsPlatform: false},
 	{Name: "vertex", DisplayName: "Google Vertex AI (keyless ADC)", Protocol: "openai", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"GOOGLE_APPLICATION_CREDENTIALS"}, ModelPrefixMatch: "^vertex:", IsPlatform: false},
-	{Name: "alibaba", DisplayName: "Alibaba Qwen (DashScope)", Protocol: "openai", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"DASHSCOPE_API_KEY", "ALIBABA_API_KEY"}, ModelPrefixMatch: "^qwen-", IsPlatform: false},
-	{Name: "nousresearch", DisplayName: "Nous Research (Hermes)", Protocol: "openai", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"NOUSRESEARCH_API_KEY"}, ModelPrefixMatch: "^nousresearch/", IsPlatform: false},
-	{Name: "openrouter", DisplayName: "OpenRouter (any model)", Protocol: "openai", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"OPENROUTER_API_KEY"}, ModelPrefixMatch: "^openrouter/", IsPlatform: false},
-	{Name: "huggingface", DisplayName: "Hugging Face Inference", Protocol: "openai", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"HUGGINGFACE_API_KEY", "HF_TOKEN"}, ModelPrefixMatch: "^huggingface/", IsPlatform: false},
-	{Name: "ai-gateway", DisplayName: "Vercel AI Gateway", Protocol: "openai", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"AI_GATEWAY_API_KEY"}, ModelPrefixMatch: "^ai-gateway/", IsPlatform: false},
-	{Name: "opencode-zen", DisplayName: "OpenCode Zen", Protocol: "openai", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"OPENCODE_ZEN_API_KEY"}, ModelPrefixMatch: "^opencode-zen/", IsPlatform: false},
-	{Name: "opencode-go", DisplayName: "OpenCode Go", Protocol: "openai", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"OPENCODE_GO_API_KEY"}, ModelPrefixMatch: "^opencode-go/", IsPlatform: false},
-	{Name: "kilocode", DisplayName: "Kilo Code", Protocol: "openai", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"KILOCODE_API_KEY"}, ModelPrefixMatch: "^kilocode/", IsPlatform: false},
-	{Name: "minimax-cn", DisplayName: "MiniMax China", Protocol: "openai", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"MINIMAX_API_KEY", "ANTHROPIC_AUTH_TOKEN"}, ModelPrefixMatch: "^minimax-cn/", IsPlatform: false},
-	{Name: "ollama-cloud", DisplayName: "Ollama Cloud", Protocol: "openai", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"OLLAMA_CLOUD_API_KEY"}, ModelPrefixMatch: "^ollama-cloud/", IsPlatform: false},
+	{Name: "alibaba", DisplayName: "Alibaba Qwen (DashScope)", Protocol: "openai", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"DASHSCOPE_API_KEY", "ALIBABA_API_KEY"}, ModelPrefixMatch: "(?i)^(qwen|alibaba[:/])", IsPlatform: false},
+	{Name: "nousresearch", DisplayName: "Nous Research (Hermes)", Protocol: "openai", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"NOUSRESEARCH_API_KEY"}, ModelPrefixMatch: "^nousresearch[:/]", IsPlatform: false},
+	{Name: "openrouter", DisplayName: "OpenRouter (any model)", Protocol: "openai", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"OPENROUTER_API_KEY"}, ModelPrefixMatch: "^openrouter[:/]", IsPlatform: false},
+	{Name: "huggingface", DisplayName: "Hugging Face Inference", Protocol: "openai", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"HUGGINGFACE_API_KEY", "HF_TOKEN"}, ModelPrefixMatch: "^huggingface[:/]", IsPlatform: false},
+	{Name: "ai-gateway", DisplayName: "Vercel AI Gateway", Protocol: "openai", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"AI_GATEWAY_API_KEY"}, ModelPrefixMatch: "^ai-gateway[:/]", IsPlatform: false},
+	{Name: "opencode-zen", DisplayName: "OpenCode Zen", Protocol: "openai", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"OPENCODE_ZEN_API_KEY"}, ModelPrefixMatch: "^opencode-zen[:/]", IsPlatform: false},
+	{Name: "opencode-go", DisplayName: "OpenCode Go", Protocol: "openai", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"OPENCODE_GO_API_KEY"}, ModelPrefixMatch: "^opencode-go[:/]", IsPlatform: false},
+	{Name: "kilocode", DisplayName: "Kilo Code", Protocol: "openai", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"KILOCODE_API_KEY"}, ModelPrefixMatch: "^kilocode[:/]", IsPlatform: false},
+	{Name: "minimax-cn", DisplayName: "MiniMax China", Protocol: "openai", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"MINIMAX_API_KEY", "ANTHROPIC_AUTH_TOKEN"}, ModelPrefixMatch: "^minimax-cn[:/]", IsPlatform: false},
+	{Name: "ollama-cloud", DisplayName: "Ollama Cloud", Protocol: "openai", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"OLLAMA_CLOUD_API_KEY"}, ModelPrefixMatch: "^ollama-cloud[:/]", IsPlatform: false},
 	{Name: "ollama", DisplayName: "Ollama (self-hosted)", Protocol: "openai", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"OLLAMA_HOST"}, ModelPrefixMatch: "^ollama/", IsPlatform: false},
-	{Name: "nvidia", DisplayName: "NVIDIA NIM", Protocol: "openai", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"NVIDIA_API_KEY"}, ModelPrefixMatch: "^nvidia/", IsPlatform: false},
-	{Name: "arcee", DisplayName: "Arcee", Protocol: "openai", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"ARCEE_API_KEY"}, ModelPrefixMatch: "^arcee/", IsPlatform: false},
-	{Name: "custom", DisplayName: "Custom OpenAI-compat endpoint", Protocol: "openai", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"CUSTOM_API_KEY", "OPENAI_API_KEY"}, ModelPrefixMatch: "^custom/", IsPlatform: false},
+	{Name: "nvidia", DisplayName: "NVIDIA NIM", Protocol: "openai", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"NVIDIA_API_KEY"}, ModelPrefixMatch: "^nvidia[:/]", IsPlatform: false},
+	{Name: "arcee", DisplayName: "Arcee", Protocol: "openai", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"ARCEE_API_KEY"}, ModelPrefixMatch: "^arcee[:/]", IsPlatform: false},
+	{Name: "custom", DisplayName: "Custom OpenAI-compat endpoint", Protocol: "openai", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"CUSTOM_API_KEY", "OPENAI_API_KEY"}, ModelPrefixMatch: "^custom[:/]", IsPlatform: false},
+	{Name: "byok-anthropic", DisplayName: "Anthropic (BYOK)", Protocol: "anthropic", AuthMode: "anthropic_api", AuthEnv: []string{"ANTHROPIC_API_KEY"}, ModelPrefixMatch: "^anthropic/", IsPlatform: false},
+	{Name: "byok-openai", DisplayName: "OpenAI (BYOK)", Protocol: "openai", AuthMode: "anthropic_api", AuthEnv: []string{"OPENAI_API_KEY"}, ModelPrefixMatch: "^openai[:/]", IsPlatform: false},
+	{Name: "byok-gemini", DisplayName: "Google Gemini (BYOK)", Protocol: "openai", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"GEMINI_API_KEY", "GOOGLE_API_KEY"}, ModelPrefixMatch: "^gemini/", IsPlatform: false},
+	{Name: "byok-minimax", DisplayName: "MiniMax (BYOK)", Protocol: "openai", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"MINIMAX_API_KEY"}, ModelPrefixMatch: "(?i)^(minimax[:/]|codex-minimax-)", IsPlatform: false},
+	{Name: "groq", DisplayName: "Groq", Protocol: "openai", AuthMode: "third_party_anthropic_compat", AuthEnv: []string{"GROQ_API_KEY"}, ModelPrefixMatch: "^groq:", IsPlatform: false},
 }
 
 // Runtimes maps each runtime to its native provider+model set, runtime names
@@ -82,23 +87,51 @@ var Runtimes = map[string][]GenRuntimeRef{
 		{Name: "kimi-coding", Models: []string{"kimi-for-coding", "kimi-k2.5", "kimi-k2", "moonshot:kimi-k2.6", "moonshot:kimi-k2.5"}},
 		{Name: "minimax", Models: []string{"MiniMax-M2", "MiniMax-M2.7", "MiniMax-M2.7-highspeed", "MiniMax-M3", "minimax:MiniMax-M2", "minimax:MiniMax-M2.7", "minimax:MiniMax-M2.7-highspeed", "minimax:MiniMax-M3"}},
 		{Name: "platform", Models: []string{"anthropic/claude-opus-4-7", "anthropic/claude-sonnet-4-6", "moonshot/kimi-k2.6", "moonshot/kimi-k2.5", "minimax/MiniMax-M2.7", "minimax/MiniMax-M2.7-highspeed", "minimax/MiniMax-M3"}},
+		{Name: "zai", Models: []string{}},
+		{Name: "deepseek", Models: []string{}},
+		{Name: "xiaomi-mimo", Models: []string{}},
 	},
 	"codex": {
 		{Name: "openai-subscription", Models: []string{"gpt-5.5", "gpt-5.4", "gpt-5.4-mini", "gpt-5.3-codex", "gpt-5.3-codex-spark", "gpt-5.2"}},
 		{Name: "openai-api", Models: []string{"gpt-5.5", "gpt-5.4", "gpt-5.4-mini", "gpt-5.3-codex", "gpt-5.3-codex-spark", "gpt-5.2"}},
 		{Name: "platform", Models: []string{"openai/gpt-5.4", "openai/gpt-5.4-mini"}},
+		{Name: "byok-minimax", Models: []string{}},
 	},
 	"google-adk": {
 		{Name: "platform", Models: []string{"platform:gemini-2.5-pro", "platform:gemini-2.5-flash"}},
 		{Name: "google", Models: []string{"gemini-2.5-pro", "gemini-2.5-flash"}},
-		{Name: "vertex", Models: []string{"vertex:gemini-2.5-pro"}},
 	},
 	"hermes": {
 		{Name: "kimi-coding", Models: []string{"kimi-coding/kimi-k2"}},
 		{Name: "platform", Models: []string{"moonshot/kimi-k2.6", "moonshot/kimi-k2.5"}},
+		{Name: "openrouter", Models: []string{}},
+		{Name: "huggingface", Models: []string{}},
+		{Name: "ai-gateway", Models: []string{}},
+		{Name: "opencode-zen", Models: []string{}},
+		{Name: "opencode-go", Models: []string{}},
+		{Name: "kilocode", Models: []string{}},
+		{Name: "custom", Models: []string{}},
+		{Name: "nvidia", Models: []string{}},
+		{Name: "arcee", Models: []string{}},
+		{Name: "ollama-cloud", Models: []string{}},
+		{Name: "minimax-cn", Models: []string{}},
+		{Name: "nousresearch", Models: []string{}},
+		{Name: "deepseek", Models: []string{}},
+		{Name: "zai", Models: []string{}},
+		{Name: "xiaomi-mimo", Models: []string{}},
+		{Name: "alibaba", Models: []string{}},
+		{Name: "byok-anthropic", Models: []string{}},
+		{Name: "byok-gemini", Models: []string{}},
+		{Name: "byok-openai", Models: []string{}},
+		{Name: "byok-minimax", Models: []string{}},
 	},
 	"openclaw": {
 		{Name: "kimi-coding", Models: []string{"moonshot:kimi-k2.6", "moonshot:kimi-k2.5"}},
 		{Name: "platform", Models: []string{"moonshot/kimi-k2.6", "moonshot/kimi-k2.5"}},
+		{Name: "openrouter", Models: []string{}},
+		{Name: "custom", Models: []string{}},
+		{Name: "byok-openai", Models: []string{}},
+		{Name: "byok-minimax", Models: []string{}},
+		{Name: "groq", Models: []string{}},
 	},
 }

workspace-server/internal/providers/providers.go

@@ -257,9 +257,20 @@ func parseManifest(raw []byte) (*Manifest, error) {
 				return nil, fmt.Errorf("providers: runtime %q references provider %q twice", rt, ref.Name)
 			}
 			refSeen[ref.Name] = struct{}{}
-			if len(ref.Models) == 0 {
-				return nil, fmt.Errorf("providers: runtime %q provider %q has no model ids", rt, ref.Name)
-			}
+			// A NAME-ONLY arm (zero model ids) is permitted (cp#529): it adds
+			// NOTHING to the runtime's platform menu (ModelsForRuntime only
+			// iterates ref.Models, so an empty Models contributes no selectable
+			// id — additive, zero platform-menu change) yet wires the provider
+			// into the runtime's NATIVE prefix-routing set, so a BYOK id the
+			// provider's model_prefix_match matches becomes routable via
+			// DeriveProvider step-4. This is the mechanism the cp#529
+			// routability-aware enforcer keys off: a name-only BYOK arm makes a
+			// passthrough id (openrouter/…, deepseek-…, etc.) resolve to a
+			// concrete provider without ever appearing on the platform menu.
+			// BILLING GUARDRAIL: only CONFIRMED-NON-PLATFORM (BYOK) providers
+			// are wired as name-only arms — never `platform`/anthropic-*/
+			// openai-*/moonshot/minimax/google/vertex — so a name-only arm can
+			// never route a customer model through the platform's key.
 		}
 	}
 

workspace-server/internal/providers/providers.yaml

@@ -317,7 +317,7 @@ providers:
     # Adapter prefix "mimo-"; canvas /^mimo-/i. proxy routing TBD (PR-3).
     # NOTE: canvas has a duplicate "xiaomi" VENDOR_LABELS key aliasing the
     # same vendor — collapsed into this one entry.
-    model_prefix_match: "^mimo-"
+    model_prefix_match: "(?i)^(mimo-|xiaomi[:/])"
     model_aliases: []
 
   # ===========================================================================
@@ -334,7 +334,7 @@ providers:
     auth_token_env: ANTHROPIC_AUTH_TOKEN
     # Adapter prefix "glm-" (lowercased match catches GLM-4.6); canvas /^GLM-/i.
     # canvas-only + adapter-only today; proxy routing TBD (PR-3).
-    model_prefix_match: "(?i)^glm-"
+    model_prefix_match: "(?i)^(glm-|zai[:/])"
     model_aliases: []
 
   # ===========================================================================
@@ -385,7 +385,7 @@ providers:
     auth_token_env: ANTHROPIC_AUTH_TOKEN
     # Adapter prefix "deepseek-"; canvas /^deepseek-/i. adapter+canvas only;
     # proxy routing TBD (PR-3).
-    model_prefix_match: "^deepseek-"
+    model_prefix_match: "^deepseek[-:/]"
     model_aliases: []
 
   # ===========================================================================
@@ -452,7 +452,7 @@ providers:
     auth_env: [DASHSCOPE_API_KEY, ALIBABA_API_KEY]
     auth_token_env: ANTHROPIC_AUTH_TOKEN
     # canvas-only today; proxy routing TBD. canvas /^qwen-/i.
-    model_prefix_match: "^qwen-"
+    model_prefix_match: "(?i)^(qwen|alibaba[:/])"
     model_aliases: []
 
   - name: nousresearch
@@ -466,7 +466,7 @@ providers:
     auth_token_env: ANTHROPIC_AUTH_TOKEN
     # canvas-only today; proxy routing TBD. Slash-prefix id
     # (e.g. nousresearch/hermes-4-70b).
-    model_prefix_match: "^nousresearch/"
+    model_prefix_match: "^nousresearch[:/]"
     model_aliases: []
 
   - name: openrouter
@@ -479,7 +479,7 @@ providers:
     auth_env: [OPENROUTER_API_KEY]
     auth_token_env: ANTHROPIC_AUTH_TOKEN
     # canvas-only today; proxy routing TBD. Wildcard: openrouter/<model>.
-    model_prefix_match: "^openrouter/"
+    model_prefix_match: "^openrouter[:/]"
     model_aliases: []
 
   - name: huggingface
@@ -492,7 +492,7 @@ providers:
     auth_env: [HUGGINGFACE_API_KEY, HF_TOKEN]
     auth_token_env: ANTHROPIC_AUTH_TOKEN
     # canvas-only today; proxy routing TBD. Wildcard: huggingface/<model>.
-    model_prefix_match: "^huggingface/"
+    model_prefix_match: "^huggingface[:/]"
     model_aliases: []
 
   - name: ai-gateway
@@ -505,7 +505,7 @@ providers:
     auth_env: [AI_GATEWAY_API_KEY]
     auth_token_env: ANTHROPIC_AUTH_TOKEN
     # canvas-only today; proxy routing TBD.
-    model_prefix_match: "^ai-gateway/"
+    model_prefix_match: "^ai-gateway[:/]"
     model_aliases: []
 
   - name: opencode-zen
@@ -518,7 +518,7 @@ providers:
     auth_env: [OPENCODE_ZEN_API_KEY]
     auth_token_env: ANTHROPIC_AUTH_TOKEN
     # canvas-only today; proxy routing TBD.
-    model_prefix_match: "^opencode-zen/"
+    model_prefix_match: "^opencode-zen[:/]"
     model_aliases: []
 
   - name: opencode-go
@@ -531,7 +531,7 @@ providers:
     auth_env: [OPENCODE_GO_API_KEY]
     auth_token_env: ANTHROPIC_AUTH_TOKEN
     # canvas-only today; proxy routing TBD.
-    model_prefix_match: "^opencode-go/"
+    model_prefix_match: "^opencode-go[:/]"
     model_aliases: []
 
   - name: kilocode
@@ -544,7 +544,7 @@ providers:
     auth_env: [KILOCODE_API_KEY]
     auth_token_env: ANTHROPIC_AUTH_TOKEN
     # canvas-only today; proxy routing TBD.
-    model_prefix_match: "^kilocode/"
+    model_prefix_match: "^kilocode[:/]"
     model_aliases: []
 
   - name: minimax-cn
@@ -559,7 +559,7 @@ providers:
     # canvas-only today; proxy routing TBD. China endpoint sibling of `minimax`
     # (api.minimaxi.com). Matched only by the explicit slash-prefix so it does
     # NOT collide with `minimax`'s (?i)^minimax- in the overlap guard.
-    model_prefix_match: "^minimax-cn/"
+    model_prefix_match: "^minimax-cn[:/]"
     model_aliases: []
 
   - name: ollama-cloud
@@ -572,7 +572,7 @@ providers:
     auth_env: [OLLAMA_CLOUD_API_KEY]
     auth_token_env: ANTHROPIC_AUTH_TOKEN
     # canvas-only today; proxy routing TBD.
-    model_prefix_match: "^ollama-cloud/"
+    model_prefix_match: "^ollama-cloud[:/]"
     model_aliases: []
 
   - name: ollama
@@ -598,7 +598,7 @@ providers:
     auth_env: [NVIDIA_API_KEY]
     auth_token_env: ANTHROPIC_AUTH_TOKEN
     # canvas-only today; proxy routing TBD.
-    model_prefix_match: "^nvidia/"
+    model_prefix_match: "^nvidia[:/]"
     model_aliases: []
 
   - name: arcee
@@ -611,7 +611,7 @@ providers:
     auth_env: [ARCEE_API_KEY]
     auth_token_env: ANTHROPIC_AUTH_TOKEN
     # canvas-only today; proxy routing TBD.
-    model_prefix_match: "^arcee/"
+    model_prefix_match: "^arcee[:/]"
     model_aliases: []
 
   - name: custom
@@ -624,7 +624,109 @@ providers:
     auth_env: [CUSTOM_API_KEY, OPENAI_API_KEY]
     auth_token_env: ANTHROPIC_AUTH_TOKEN
     # canvas-only today; proxy routing TBD. Wildcard free-text: custom/<model>.
-    model_prefix_match: "^custom/"
+    model_prefix_match: "^custom[:/]"
+    model_aliases: []
+
+  # ===========================================================================
+  # DEDICATED BYOK-VENDOR providers (cp#529). These exist so the NAMESPACED
+  # BYOK ids the hermes/openclaw/codex templates offer for the SHARED upstream
+  # vendors (anthropic, openai, gemini, minimax, groq) become routable with the
+  # TENANT's OWN vendor key — WITHOUT routing them through the platform-shared
+  # `platform` provider (which would bill the platform's key: a money bug).
+  #
+  # Each is NON-PLATFORM (name != "platform") -> IsPlatform()==false -> BYOK
+  # billing: the workspace env supplies the vendor key, never the platform key.
+  #
+  # COLLISION-FREE BY CONSTRUCTION: every matcher is NAMESPACED (anchored on the
+  # `vendor/` slash form or `vendor:` colon form) so it is DISJOINT from the
+  # platform vendors' BARE matchers (anthropic-api `^claude`, openai-subscription
+  # `^gpt-`, openai-api `^openai-api[:/]`, minimax `(?i)^minimax-m`,
+  # google `^gemini-`, minimax-cn `^minimax-cn[:/]`). DeriveProvider's overlap
+  # guard (no slug may match two native providers) stays green — verified for all
+  # 20 residual ids (cp#529).
+  #
+  # These siblings of the platform/upstream vendor entries point at the SAME
+  # PUBLIC upstream base URLs, but carry NO upstream_vendor (they are BYOK
+  # passthroughs, not proxy upstream targets — the proxy never dials a tenant's
+  # own key) and use the namespaced matchers above instead of the bare proxy
+  # prefixes.
+  # ===========================================================================
+  - name: byok-anthropic
+    display_name: "Anthropic (BYOK)"
+    vendor_logo: "anthropic"
+    protocol: anthropic
+    auth_mode: anthropic_api
+    base_url_template: "https://api.anthropic.com/v1"
+    base_url_anthropic: "https://api.anthropic.com/v1"
+    auth_env: [ANTHROPIC_API_KEY]
+    auth_token_env: ANTHROPIC_API_KEY
+    # Namespaced BYOK form `anthropic/<model>` (hermes). DISJOINT from
+    # anthropic-api's bare `^claude` and anthropic-oauth's alias set.
+    model_prefix_match: "^anthropic/"
+    model_aliases: []
+
+  - name: byok-openai
+    display_name: "OpenAI (BYOK)"
+    vendor_logo: "openai"
+    protocol: openai
+    auth_mode: anthropic_api  # openai-protocol; auth is a bearer API key.
+    base_url_template: "https://api.openai.com/v1"
+    base_url_anthropic: null
+    auth_env: [OPENAI_API_KEY]
+    auth_token_env: OPENAI_API_KEY
+    # Namespaced BYOK forms `openai/<model>` (hermes) + `openai:<model>`
+    # (openclaw). DISJOINT from openai-subscription's bare `^gpt-` and
+    # openai-api's `^openai-api[:/]` (the dash after `openai` keeps the two
+    # apart: `openai:` / `openai/` never start with `openai-api`).
+    model_prefix_match: "^openai[:/]"
+    model_aliases: []
+
+  - name: byok-gemini
+    display_name: "Google Gemini (BYOK)"
+    vendor_logo: "google"
+    protocol: openai
+    auth_mode: third_party_anthropic_compat
+    base_url_template: "https://generativelanguage.googleapis.com/v1beta/openai"
+    base_url_anthropic: null
+    auth_env: [GEMINI_API_KEY, GOOGLE_API_KEY]
+    auth_token_env: ANTHROPIC_AUTH_TOKEN
+    # Namespaced BYOK form `gemini/<model>` (hermes). DISJOINT from the `google`
+    # vendor's bare `^gemini-` and `vertex`'s `^vertex:`.
+    model_prefix_match: "^gemini/"
+    model_aliases: []
+
+  - name: byok-minimax
+    display_name: "MiniMax (BYOK)"
+    vendor_logo: "minimax"
+    protocol: openai
+    auth_mode: third_party_anthropic_compat
+    base_url_template: "https://api.minimax.io/v1"
+    base_url_anthropic: null
+    auth_env: [MINIMAX_API_KEY]
+    auth_token_env: ANTHROPIC_AUTH_TOKEN
+    # Namespaced BYOK forms `minimax:<model>` (openclaw) + `minimax/<model>`
+    # (hermes), PLUS the codex-runtime alias `codex-minimax-m2.7` (the codex
+    # template's `minimax-token-plan` route — same upstream api.minimax.io,
+    # tenant MINIMAX_API_KEY). The `codex-minimax-` leg is NARROWLY anchored so
+    # it resolves that one codex id WITHOUT a broad matcher: it is DISJOINT from
+    # `minimax` (?i)^minimax-m (which needs `minimax-m`, not `codex-`) and from
+    # `minimax-cn` ^minimax-cn[:/]. Verified collision-free for all 20 residual
+    # ids + codex-minimax-m2.7 (cp#529).
+    model_prefix_match: "(?i)^(minimax[:/]|codex-minimax-)"
+    model_aliases: []
+
+  - name: groq
+    display_name: "Groq"
+    vendor_logo: "groq"
+    protocol: openai
+    auth_mode: third_party_anthropic_compat
+    base_url_template: "https://api.groq.com/openai/v1"
+    base_url_anthropic: null
+    auth_env: [GROQ_API_KEY]
+    auth_token_env: ANTHROPIC_AUTH_TOKEN
+    # Namespaced BYOK form `groq:<model>` (openclaw). No other provider matches
+    # the `groq:` prefix.
+    model_prefix_match: "^groq:"
     model_aliases: []
 
 # =============================================================================
@@ -762,6 +864,16 @@ runtimes:
           - minimax/MiniMax-M2.7
           - minimax/MiniMax-M2.7-highspeed
           - minimax/MiniMax-M3
+      # NAME-ONLY BYOK arms (cp#529): zero model ids → add NOTHING to the
+      # platform menu (ModelsForRuntime), but wire these CONFIRMED-NON-PLATFORM
+      # providers into claude-code's NATIVE prefix-routing set so the bare BYOK
+      # ids the claude-code template offers (GLM-*, deepseek-*, mimo-*) resolve
+      # via DeriveProvider. BILLING-SAFE: zai/deepseek/xiaomi-mimo are tenant-key
+      # (BYOK) providers — never platform-shared — so routing through them bills
+      # the tenant's own key, never the platform's.
+      - name: zai
+      - name: deepseek
+      - name: xiaomi-mimo
 
   # hermes: native Kimi only (kimi-coding gateway). hermes-agent owns its own
   # broad provider matrix, but the CTO native matrix for the Molecule
@@ -777,6 +889,38 @@ runtimes:
         models:
           - moonshot/kimi-k2.6
           - moonshot/kimi-k2.5
+      # NAME-ONLY BYOK arms (cp#529): zero model ids → no addition to the
+      # platform menu, but wire hermes's CONFIRMED-NON-PLATFORM passthrough +
+      # bare-vendor providers into its NATIVE prefix-routing set so the BYOK
+      # ids the hermes template offers (openrouter/…, huggingface/…, deepseek/…,
+      # zai:…, etc.) resolve via DeriveProvider. ALL tenant-key (BYOK).
+      - name: openrouter
+      - name: huggingface
+      - name: ai-gateway
+      - name: opencode-zen
+      - name: opencode-go
+      - name: kilocode
+      - name: custom
+      - name: nvidia
+      - name: arcee
+      - name: ollama-cloud
+      - name: minimax-cn
+      - name: nousresearch
+      - name: deepseek
+      - name: zai
+      - name: xiaomi-mimo
+      - name: alibaba
+      # DEDICATED BYOK-VENDOR arms (cp#529): the namespaced ids hermes offers for
+      # the SHARED upstream vendors (anthropic/claude-*, gemini/*, openai/*,
+      # minimax/*) NOW resolve to these tenant-key BYOK-vendor providers — NOT
+      # the platform-shared `platform` provider (which would bill the platform's
+      # key). NAME-ONLY (no models) → no platform-menu change, prefix-routing
+      # only, BYOK-billed. This converts the last 12 hermes residual ids from
+      # cp#529 drift to routable.
+      - name: byok-anthropic
+      - name: byok-gemini
+      - name: byok-openai
+      - name: byok-minimax
 
   # codex: OpenAI — BYOK split across TWO native providers
   # (openai-subscription + openai-api), mirroring claude-code's anthropic
@@ -828,6 +972,14 @@ runtimes:
         models:
           - openai/gpt-5.4
           - openai/gpt-5.4-mini
+      # NAME-ONLY BYOK arm (cp#529): the codex template offers a BYOK MiniMax
+      # token-plan model `codex-minimax-m2.7` (its `minimax-token-plan` provider:
+      # base_url api.minimax.io, tenant MINIMAX_API_KEY, model_id_override
+      # codex-MiniMax-M2.7). It resolves to byok-minimax via the narrowly-anchored
+      # `codex-minimax-` leg of byok-minimax's matcher (same upstream, tenant key)
+      # — NOT a broad matcher. NAME-ONLY → no platform-menu change, BYOK-billed.
+      # Converts the last codex residual id from cp#529 drift to routable.
+      - name: byok-minimax
 
   # openclaw: native Kimi only. openclaw's moonshot: model prefix + a
   # KIMI_API_KEY (sk-kimi-*) routes to api.kimi.com/coding (kimi-for-coding),
@@ -846,6 +998,21 @@ runtimes:
         models:
           - moonshot/kimi-k2.6
           - moonshot/kimi-k2.5
+      # NAME-ONLY BYOK arms (cp#529): zero model ids → no platform-menu change,
+      # but wire openclaw's CONFIRMED-NON-PLATFORM passthroughs into its NATIVE
+      # prefix-routing set so the BYOK colon/slash ids the openclaw template
+      # offers (openrouter:…, custom:…) resolve via DeriveProvider. BYOK only.
+      - name: openrouter
+      - name: custom
+      # DEDICATED BYOK-VENDOR arms (cp#529): openclaw's default model is
+      # `minimax:MiniMax-M2.7`, plus it offers `openai:*` and `groq:*` BYOK ids.
+      # These NOW resolve to the tenant-key BYOK-vendor providers (NOT the
+      # platform key). NAME-ONLY → prefix-routing only, BYOK-billed. This converts
+      # the last 7 openclaw residual ids from cp#529 drift to routable AND makes
+      # the runtime's DEFAULT model (minimax:MiniMax-M2.7) resolve.
+      - name: byok-openai
+      - name: byok-minimax
+      - name: groq
 
 
   # google-adk: Gemini via Vertex AI, keyless ADC (Workload Identity
@@ -869,9 +1036,3 @@ runtimes:
         models:
           - gemini-2.5-pro
           - gemini-2.5-flash
-      # DEPRECATED transitional: vertex: ids stay registered until templates
-      # move to platform: (superseded by the platform arm above). Remove in a
-      # cleanup once no template references vertex:gemini-*.
-      - name: vertex
-        models:
-          - vertex:gemini-2.5-pro

workspace-server/internal/providers/providers_test.go

@@ -249,11 +249,17 @@ func TestGoogleADKRuntimeRegistered(t *testing.T) {
 }
 
 // TestVertexProviderRegistered locks the keyless Vertex provider variant in the
-// providers.yaml SSOT. google-adk serves Gemini via Vertex AI with ADC/WIF
-// (no API key); the registry must model that as a first-class "vertex" provider
-// (auth_env GOOGLE_APPLICATION_CREDENTIALS, ^vertex: namespace) distinct from
-// the API-key "google" vendor, and the google-adk runtime must offer both arms.
-// See project_canvas_runtime_dropdown_ssot_fix.
+// providers.yaml SSOT. google-adk serves platform-managed Gemini via the LLM
+// proxy -> Vertex AI with server-side WIF (no on-box key); the registry must
+// still model the keyless "vertex" provider (auth_env GOOGLE_APPLICATION_CREDENTIALS,
+// ^vertex: namespace) as a first-class entry distinct from the API-key "google"
+// vendor, so the proxy can still route/bill any Vertex-upstream request that
+// carries a `vertex:` id. The TRANSITIONAL `vertex:` arm on the google-adk
+// RUNTIME (the selectable model set) was removed in cp#514 now that templates
+// default to `platform:`; the runtime offers only the `platform` + API-key
+// `google` arms. A saved `vertex:gemini-*` model still RESOLVES harmlessly via
+// this standalone provider (it is just no longer offered as a new selection).
+// See project_canvas_runtime_dropdown_ssot_fix + cp#514.
 func TestVertexProviderRegistered(t *testing.T) {
 	ps, err := Load()
 	if err != nil {
@@ -298,20 +304,23 @@ func TestVertexProviderRegistered(t *testing.T) {
 	for _, p := range provs {
 		names[p.Name] = true
 	}
-	if !names["vertex"] {
-		t.Errorf("google-adk runtime should offer the keyless vertex arm; got %v", names)
+	// cp#514: the transitional `vertex` arm was dropped from the google-adk
+	// runtime. The runtime keeps the platform-managed default + the API-key
+	// google arm; the standalone `vertex` PROVIDER (asserted above) survives
+	// for ^vertex: resolution but is no longer a selectable runtime arm.
+	if names["vertex"] {
+		t.Errorf("google-adk runtime should NOT offer the transitional vertex arm (removed cp#514); got %v", names)
+	}
+	if !names["platform"] {
+		t.Errorf("google-adk runtime should keep the platform-managed arm; got %v", names)
 	}
 	if !names["google"] {
 		t.Errorf("google-adk runtime should keep the API-key google arm; got %v", names)
 	}
 	models, _ := m.ModelsForRuntime("google-adk")
-	hasVertexModel := false
 	for _, id := range models {
 		if id == "vertex:gemini-2.5-pro" {
-			hasVertexModel = true
+			t.Errorf("google-adk models should NOT include vertex:gemini-2.5-pro (removed cp#514); got %v", models)
 		}
 	}
-	if !hasVertexModel {
-		t.Errorf("google-adk models should include vertex:gemini-2.5-pro; got %v", models)
-	}
 }

workspace-server/internal/providers/runtimes_test.go

@@ -17,19 +17,39 @@ import (
 // of its native vendors the proxy can serve — kimi for hermes/openclaw,
 // openai for codex, anthropic+kimi+minimax for claude-code.
 //
+// cp#529 adds NAME-ONLY BYOK arms (zero model ids) to claude-code/hermes/
+// openclaw: they add NOTHING to the platform menu (ModelsForRuntime) but wire
+// CONFIRMED-NON-PLATFORM providers into the runtime's NATIVE prefix-routing set
+// so a matching BYOK id resolves via DeriveProvider. ProvidersForRuntime returns
+// the full native arm set (menu + name-only), so the expected sets below include
+// them. The platform-shared/denylist providers are NEVER wired into a BYOK arm.
+//
 //	claude-code -> anthropic (oauth+api), kimi (kimi-coding), minimax, platform
+//	               + BYOK name-only: zai, deepseek, xiaomi-mimo
 //	hermes      -> kimi (kimi-coding), platform
-//	codex       -> openai (subscription + api), platform
-//	openclaw    -> kimi (kimi-coding), platform
+//	               + BYOK name-only: openrouter, huggingface, ai-gateway,
+//	                 opencode-zen, opencode-go, kilocode, custom, nvidia, arcee,
+//	                 ollama-cloud, minimax-cn, nousresearch, deepseek, zai,
+//	                 xiaomi-mimo, alibaba
+//	codex       -> openai (subscription + api), platform   (no BYOK name-only)
+//	openclaw    -> kimi (kimi-coding), platform + BYOK name-only: openrouter, custom
 var runtimeNativeProviders = map[string][]string{
-	"claude-code": {"anthropic-api", "anthropic-oauth", "kimi-coding", "minimax", "platform"},
-	"hermes":      {"kimi-coding", "platform"},
+	"claude-code": {"anthropic-api", "anthropic-oauth", "kimi-coding", "minimax", "platform", "zai", "deepseek", "xiaomi-mimo"},
+	"hermes": {"kimi-coding", "platform",
+		"openrouter", "huggingface", "ai-gateway", "opencode-zen", "opencode-go",
+		"kilocode", "custom", "nvidia", "arcee", "ollama-cloud", "minimax-cn",
+		"nousresearch", "deepseek", "zai", "xiaomi-mimo", "alibaba",
+		// cp#529 dedicated BYOK-vendor name-only arms (shared-vendor namespaced ids).
+		"byok-anthropic", "byok-gemini", "byok-openai", "byok-minimax"},
 	// codex's OpenAI BYOK is split across the OAuth subscription arm
 	// (openai-subscription) and the direct-key arm (openai-api), mirroring
 	// claude-code's anthropic oauth+api split; platform openai via the proxy
-	// Responses surface.
-	"codex":    {"openai-subscription", "openai-api", "platform"},
-	"openclaw": {"kimi-coding", "platform"},
+	// Responses surface. cp#529 adds the byok-minimax name-only arm so the
+	// template's BYOK MiniMax token-plan id (codex-minimax-m2.7) resolves.
+	"codex":    {"openai-subscription", "openai-api", "platform", "byok-minimax"},
+	"openclaw": {"kimi-coding", "platform", "openrouter", "custom",
+		// cp#529 dedicated BYOK-vendor name-only arms (openai:/minimax:/groq:).
+		"byok-openai", "byok-minimax", "groq"},
 }
 
 func sortedCopy(in []string) []string {
@@ -253,6 +273,56 @@ func TestParseManifest_ValidBaseline(t *testing.T) {
 	}
 }
 
+// TestParseManifest_NameOnlyArm proves a NAME-ONLY runtime arm (zero model
+// ids) is PERMITTED (cp#529) and is additive: it contributes nothing to the
+// runtime's platform menu (ModelsForRuntime) yet wires the provider into the
+// runtime's NATIVE prefix-routing set so a matching BYOK id resolves via
+// DeriveProvider. This is the loader half of the cp#529 routability change.
+func TestParseManifest_NameOnlyArm(t *testing.T) {
+	const y = `
+schema_version: 1
+providers:
+  - name: openai
+    display_name: "OpenAI"
+    protocol: openai
+    auth_mode: anthropic_api
+    auth_env: [OPENAI_API_KEY]
+    model_prefix_match: "^gpt-"
+  - name: openrouter
+    display_name: "OpenRouter"
+    protocol: openai
+    auth_mode: third_party_anthropic_compat
+    auth_env: [OPENROUTER_API_KEY]
+    model_prefix_match: "^openrouter[:/]"
+runtimes:
+  codex:
+    providers:
+      - name: openai
+        models: [gpt-5.5]
+      - name: openrouter
+`
+	m, err := parseManifest([]byte(y))
+	if err != nil {
+		t.Fatalf("parseManifest(name-only arm) error = %v; want nil (name-only arms are permitted)", err)
+	}
+	// The name-only arm adds NOTHING to the platform menu.
+	models, err := m.ModelsForRuntime("codex")
+	if err != nil {
+		t.Fatalf("ModelsForRuntime(codex) error = %v", err)
+	}
+	if len(models) != 1 || models[0] != "gpt-5.5" {
+		t.Fatalf("ModelsForRuntime(codex) = %v; want [gpt-5.5] (name-only arm must not add a menu id)", models)
+	}
+	// …yet a BYOK id matching the name-only arm's prefix now ROUTES.
+	p, err := m.DeriveProvider("codex", "openrouter/anthropic/claude-3.5-sonnet", nil)
+	if err != nil {
+		t.Fatalf("DeriveProvider(codex, openrouter/…) error = %v; want it to resolve via the name-only arm", err)
+	}
+	if p.Name != "openrouter" {
+		t.Fatalf("DeriveProvider resolved to %q; want openrouter", p.Name)
+	}
+}
+
 // TestParseManifest_FailDirection is the load-bearing-guard proof: each case
 // breaks the manifest in one way and asserts the matching error fires. If a
 // future edit removes a guard, the corresponding case flips red.
@@ -287,19 +357,6 @@ runtimes:
 `,
 			wantErr: "empty native provider set",
 		},
-		{
-			name: "provider ref with no models",
-			yaml: `
-schema_version: 1
-providers:
-  - {name: openai, display_name: "OpenAI", protocol: openai, auth_mode: anthropic_api, auth_env: [OPENAI_API_KEY], model_prefix_match: "^gpt-"}
-runtimes:
-  codex:
-    providers:
-      - {name: openai, models: []}
-`,
-			wantErr: "no model ids",
-		},
 		{
 			name: "duplicate provider ref",
 			yaml: `

workspace-server/internal/providers/sync_canonical_test.go

@@ -29,7 +29,7 @@ import (
 // canonicalProvidersYAMLSHA256 is the sha256 of the canonical providers.yaml as
 // synced from molecule-controlplane. Bumped deliberately on each re-sync (see
 // file doc). Cross-checked live by the sync-providers-yaml CI workflow.
-const canonicalProvidersYAMLSHA256 = "021ae082c2bbbbb61c406cae03205ac6b7fff160ae5976cfc64de3de676d02b2"
+const canonicalProvidersYAMLSHA256 = "846ddef11ec423ebf2e96b5da21bd89129dbc3f0a2d14ac086940e005c079387"
 
 func TestSyncedYAMLMatchesCanonicalSHA(t *testing.T) {
 	sum := sha256.Sum256(embeddedYAML)

workspace-server/internal/registry/cp_instance_reconciler.go

@@ -0,0 +1,178 @@
+package registry
+
+// cp_instance_reconciler.go — authoritative EC2-state reconcile for
+// SaaS workspaces (core#2261).
+//
+// Root cause (core#2247): every existing liveness pass keys off a PROXY
+// for "is this workspace alive?":
+//
+//   - StartLivenessMonitor   — Redis TTL expiry (agent stopped heartbeating).
+//   - StartHealthSweep (Docker pass) — local Docker daemon (prov != nil only).
+//   - StartHealthSweep (remote pass) — last_heartbeat_at freshness for
+//     runtime='external' rows.
+//   - StartCPOrphanSweeper    — status='removed' rows with a stray instance_id.
+//
+// A SaaS claude-code workspace whose EC2 was terminated/stopped out from
+// under us (manual AWS action, spot reclaim, CP-side reap, etc.) falls
+// through ALL of them: it's not 'removed' (so the orphan sweeper skips
+// it), it's not runtime='external' (so the heartbeat pass skips it), and
+// on a pure-SaaS front-door prov == nil so the Docker pass never runs.
+// The registry kept status='online' pointing at a dead instance forever.
+//
+// This sweeper closes that gap with the ONE authoritative check the
+// others lack: CPProvisioner.IsRunning, which ultimately asks the
+// control-plane "is this EC2 actually running?" (DescribeInstances-
+// equivalent). When the answer is a CLEAN "no" it feeds the workspace
+// into the EXISTING offline/auto-heal machinery (onOffline → status flip
+// + RestartByID reprovision with the existing volume) — no new healing
+// path, just real ground truth driving the one we already have.
+//
+// Guardrails:
+//   - FAIL-SAFE: IsRunning is (true, err) on any transient DB/transport
+//     error and (false, nil) ONLY when CP genuinely reports the instance
+//     is not running. We act ONLY on (false, nil); any err short-circuits
+//     to "leave it alone" so a CP blip never flips a healthy workspace.
+//   - ONLINE + SaaS ONLY: status='online', instance_id present, and
+//     runtime <> 'external'. Paused/hibernated/removed/provisioning/
+//     awaiting_agent rows are out of scope; external rows are covered by
+//     the remote-heartbeat pass.
+//   - Per-cycle row cap + per-workspace timeout so one slow CP call can't
+//     stall the sweep.
+
+import (
+	"context"
+	"log"
+	"time"
+
+	"git.moleculesai.app/molecule-ai/molecule-core/workspace-server/internal/db"
+)
+
+// InstanceRunningChecker is the narrow dependency the reconciler takes
+// from the CP provisioner. *provisioner.CPProvisioner satisfies this
+// naturally; tests inject fakes.
+//
+// Contract (load-bearing): IsRunning is FAIL-SAFE — it returns
+// (true, err) on transient DB/transport errors and (false, nil) ONLY
+// when CP reports the instance is genuinely not running. The reconciler
+// flips a workspace offline strictly on (false, nil).
+type InstanceRunningChecker interface {
+	IsRunning(ctx context.Context, workspaceID string) (bool, error)
+}
+
+// CPInstanceReconcileLimit caps the per-cycle row count so a sustained
+// CP slowdown can't make a single sweep cycle run unbounded. With a 60s
+// cadence and a per-workspace timeout below, this bounds worst-case
+// cycle wall-time and lets subsequent cycles drain any backlog.
+const CPInstanceReconcileLimit = 200
+
+// cpInstanceCheckTimeout bounds a single IsRunning call so one slow CP
+// round-trip can't stall the whole sweep. Each workspace gets its own
+// timeout context derived from the cycle context.
+const cpInstanceCheckTimeout = 10 * time.Second
+
+// StartCPInstanceReconciler runs the authoritative EC2-state reconcile
+// loop until ctx is cancelled. A nil checker makes the loop a no-op
+// (matches the nil-tolerant pattern of the sibling CP sweeper).
+//
+// Caller is expected to gate on `cpProv != nil` (matching how
+// StartCPOrphanSweeper is gated at the wiring site in cmd/server/main.go)
+// — passing a nil *CPProvisioner here would also short-circuit, but the
+// gate at the call site keeps the call shape symmetric across sweepers.
+//
+// interval <= 0 falls back to the default 60s cadence so a misconfigured
+// caller can't spin a zero-duration ticker (which panics).
+func StartCPInstanceReconciler(ctx context.Context, checker InstanceRunningChecker, onOffline OfflineHandler, interval time.Duration) {
+	if checker == nil {
+		log.Println("cp-instance-reconciler: checker is nil — reconciler disabled")
+		return
+	}
+	if interval <= 0 {
+		interval = 60 * time.Second
+	}
+	log.Printf("cp-instance-reconciler started — reconciling online SaaS workspaces against real EC2 state every %s", interval)
+	ticker := time.NewTicker(interval)
+	defer ticker.Stop()
+	// Kick once at boot so a platform restart starts healing immediately
+	// rather than waiting a full interval.
+	reconcileOnce(ctx, checker, onOffline)
+	for {
+		select {
+		case <-ctx.Done():
+			log.Println("cp-instance-reconciler: shutdown")
+			return
+		case <-ticker.C:
+			reconcileOnce(ctx, checker, onOffline)
+		}
+	}
+}
+
+// reconcileOnce executes one reconcile pass. Defensive against db.DB
+// being nil so a misconfigured boot doesn't panic.
+//
+// Scope: online + SaaS-EC2 workspaces only. runtime='external' rows are
+// excluded (covered by the remote-heartbeat pass); paused/hibernated/
+// removed/provisioning/awaiting_agent are excluded by the status filter.
+func reconcileOnce(ctx context.Context, checker InstanceRunningChecker, onOffline OfflineHandler) {
+	if db.DB == nil {
+		return
+	}
+
+	rows, err := db.DB.QueryContext(ctx, `
+		SELECT id::text
+		  FROM workspaces
+		 WHERE status = 'online'
+		   AND instance_id IS NOT NULL
+		   AND instance_id != ''
+		   AND COALESCE(runtime, '') <> 'external'
+		 ORDER BY updated_at DESC
+		 LIMIT $1
+	`, CPInstanceReconcileLimit)
+	if err != nil {
+		log.Printf("cp-instance-reconciler: DB query failed: %v", err)
+		return
+	}
+	defer rows.Close()
+
+	var ids []string
+	for rows.Next() {
+		var id string
+		if scanErr := rows.Scan(&id); scanErr != nil {
+			log.Printf("cp-instance-reconciler: row scan failed: %v", scanErr)
+			continue
+		}
+		ids = append(ids, id)
+	}
+	if iterErr := rows.Err(); iterErr != nil {
+		log.Printf("cp-instance-reconciler: rows iteration failed: %v", iterErr)
+		return
+	}
+
+	for _, id := range ids {
+		// Per-workspace timeout so one slow CP round-trip can't stall
+		// the whole sweep.
+		checkCtx, cancel := context.WithTimeout(ctx, cpInstanceCheckTimeout)
+		running, checkErr := checker.IsRunning(checkCtx, id)
+		cancel()
+
+		if checkErr != nil {
+			// FAIL-SAFE: transient DB/transport error (or a no-backend
+			// signal). IsRunning returns (true, err) on these, so never
+			// flip — leave the row online and retry next cycle.
+			log.Printf("cp-instance-reconciler: IsRunning(%s) errored, leaving online (fail-safe): %v", id, checkErr)
+			continue
+		}
+		if running {
+			continue
+		}
+
+		// CLEAN "not running" — CP authoritatively reports the EC2 is
+		// terminated/stopped/absent. Feed it into the existing offline +
+		// auto-heal machinery: onOffline flips the row offline and
+		// triggers RestartByID, which reprovisions with the existing
+		// volume.
+		log.Printf("cp-instance-reconciler: workspace %s is status=online but its EC2 is not running (terminated/stopped) — flipping offline + triggering reprovision", id)
+		if onOffline != nil {
+			onOffline(ctx, id)
+		}
+	}
+}

workspace-server/internal/registry/cp_instance_reconciler_test.go

@@ -0,0 +1,282 @@
+package registry
+
+import (
+	"context"
+	"errors"
+	"sync"
+	"testing"
+	"time"
+
+	"github.com/DATA-DOG/go-sqlmock"
+
+	"git.moleculesai.app/molecule-ai/molecule-core/workspace-server/internal/db"
+)
+
+// fakeRunningChecker implements InstanceRunningChecker for the
+// instance-reconciler tests. Records every IsRunning call so tests can
+// assert which workspace IDs were probed, and returns a per-id
+// (running, err) pair so we can model CP's three answers:
+//
+//	(true,  nil) — instance is running.
+//	(false, nil) — CLEAN "not running" (terminated/stopped/absent).
+//	(true,  err) — transient DB/transport error (FAIL-SAFE path).
+type fakeRunningChecker struct {
+	mu      sync.Mutex
+	running map[string]bool
+	errs    map[string]error
+	calls   []string
+}
+
+func (f *fakeRunningChecker) IsRunning(_ context.Context, wsID string) (bool, error) {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	f.calls = append(f.calls, wsID)
+	if err, ok := f.errs[wsID]; ok {
+		// Mirror CPProvisioner.IsRunning: (true, err) on transient errors
+		// so callers stay on the alive path.
+		return true, err
+	}
+	return f.running[wsID], nil
+}
+
+// recordingOffline is an OfflineHandler that records the workspace IDs
+// it was invoked with.
+type recordingOffline struct {
+	mu    sync.Mutex
+	calls []string
+}
+
+func (r *recordingOffline) handler() OfflineHandler {
+	return func(_ context.Context, wsID string) {
+		r.mu.Lock()
+		defer r.mu.Unlock()
+		r.calls = append(r.calls, wsID)
+	}
+}
+
+func (r *recordingOffline) got() []string {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	out := make([]string, len(r.calls))
+	copy(out, r.calls)
+	return out
+}
+
+// expectReconcileQuery registers the reconciler's SELECT, pinning the
+// scope-critical predicates: status='online', instance_id present, and
+// runtime <> 'external'. A future widening that drops any of these (e.g.
+// sweeping paused rows, or external rows the heartbeat pass owns) fails
+// every test that uses this helper.
+func expectReconcileQuery(mock sqlmock.Sqlmock, rows *sqlmock.Rows) {
+	mock.ExpectQuery(`(?s)^\s*SELECT id::text\s+FROM workspaces\s+WHERE status = 'online'\s+AND instance_id IS NOT NULL\s+AND instance_id != ''\s+AND COALESCE\(runtime, ''\) <> 'external'\s+ORDER BY updated_at DESC\s+LIMIT \$1`).
+		WithArgs(CPInstanceReconcileLimit).
+		WillReturnRows(rows)
+}
+
+// TestReconcileOnce_NotRunning_FlipsOffline — the core bug (core#2247):
+// an online SaaS workspace whose EC2 is terminated. CP reports a CLEAN
+// (false, nil); onOffline MUST be called with that id so the existing
+// auto-heal (status flip + RestartByID reprovision) kicks in.
+func TestReconcileOnce_NotRunning_FlipsOffline(t *testing.T) {
+	mock := setupTestDB(t)
+	checker := &fakeRunningChecker{running: map[string]bool{"ws-dead": false}}
+	off := &recordingOffline{}
+
+	expectReconcileQuery(mock, sqlmock.NewRows([]string{"id"}).AddRow("ws-dead"))
+
+	reconcileOnce(context.Background(), checker, off.handler())
+
+	if got := off.got(); len(got) != 1 || got[0] != "ws-dead" {
+		t.Fatalf("expected onOffline(ws-dead), got %v", got)
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Fatalf("unmet expectations: %v", err)
+	}
+}
+
+// TestReconcileOnce_Running_DoesNotFlip — healthy steady state. CP
+// reports (true, nil); the workspace stays online, onOffline is NOT
+// called.
+func TestReconcileOnce_Running_DoesNotFlip(t *testing.T) {
+	mock := setupTestDB(t)
+	checker := &fakeRunningChecker{running: map[string]bool{"ws-alive": true}}
+	off := &recordingOffline{}
+
+	expectReconcileQuery(mock, sqlmock.NewRows([]string{"id"}).AddRow("ws-alive"))
+
+	reconcileOnce(context.Background(), checker, off.handler())
+
+	if got := off.got(); len(got) != 0 {
+		t.Fatalf("running workspace must NOT be flipped offline, got %v", got)
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Fatalf("unmet expectations: %v", err)
+	}
+}
+
+// TestReconcileOnce_TransientError_DoesNotFlip — FAIL-SAFE contract.
+// IsRunning returns (true, err) on a transient DB/transport blip; the
+// reconciler MUST NOT flip the workspace offline. This is the guardrail
+// that stops a CP outage from cascading every healthy workspace through
+// reprovision.
+func TestReconcileOnce_TransientError_DoesNotFlip(t *testing.T) {
+	mock := setupTestDB(t)
+	checker := &fakeRunningChecker{
+		errs: map[string]error{"ws-blip": errors.New("cp provisioner: status: connection reset")},
+	}
+	off := &recordingOffline{}
+
+	expectReconcileQuery(mock, sqlmock.NewRows([]string{"id"}).AddRow("ws-blip"))
+
+	reconcileOnce(context.Background(), checker, off.handler())
+
+	if got := off.got(); len(got) != 0 {
+		t.Fatalf("fail-safe violated: transient IsRunning error must NOT flip offline, got %v", got)
+	}
+	if calls := checker.calls; len(calls) != 1 || calls[0] != "ws-blip" {
+		t.Fatalf("expected IsRunning(ws-blip), got %v", checker.calls)
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Fatalf("unmet expectations: %v", err)
+	}
+}
+
+// TestReconcileOnce_QueryScopeExcludesExternalAndNonOnline — pins the
+// SELECT predicate. The regex in expectReconcileQuery requires
+// status='online' AND runtime <> 'external'; if a future edit widens the
+// scope to include paused/hibernated/removed rows or external rows (owned
+// by the heartbeat pass), this query no longer matches and sqlmock fails
+// the test. With the predicate intact, a DB that has only out-of-scope
+// rows returns empty → no IsRunning, no flip.
+func TestReconcileOnce_QueryScopeExcludesExternalAndNonOnline(t *testing.T) {
+	mock := setupTestDB(t)
+	checker := &fakeRunningChecker{}
+	off := &recordingOffline{}
+
+	// The predicate filters out external + non-online rows server-side,
+	// modelled as the empty result those filters produce.
+	expectReconcileQuery(mock, sqlmock.NewRows([]string{"id"}))
+
+	reconcileOnce(context.Background(), checker, off.handler())
+
+	if len(checker.calls) != 0 {
+		t.Fatalf("out-of-scope rows must never reach IsRunning, got %v", checker.calls)
+	}
+	if got := off.got(); len(got) != 0 {
+		t.Fatalf("expected no offline flips for out-of-scope rows, got %v", got)
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Fatalf("unmet expectations: %v", err)
+	}
+}
+
+// TestReconcileOnce_MixedBatch — each row is judged independently: the
+// dead one flips, the alive one and the transient-error one don't.
+func TestReconcileOnce_MixedBatch(t *testing.T) {
+	mock := setupTestDB(t)
+	checker := &fakeRunningChecker{
+		running: map[string]bool{"ws-dead": false, "ws-alive": true},
+		errs:    map[string]error{"ws-blip": errors.New("503")},
+	}
+	off := &recordingOffline{}
+
+	expectReconcileQuery(mock, sqlmock.NewRows([]string{"id"}).
+		AddRow("ws-dead").
+		AddRow("ws-alive").
+		AddRow("ws-blip"))
+
+	reconcileOnce(context.Background(), checker, off.handler())
+
+	if got := off.got(); len(got) != 1 || got[0] != "ws-dead" {
+		t.Fatalf("expected only ws-dead flipped, got %v", got)
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Fatalf("unmet expectations: %v", err)
+	}
+}
+
+// TestReconcileOnce_QueryError — DB transient failure. Reconcile returns
+// without panicking and never probes IsRunning or flips anything.
+func TestReconcileOnce_QueryError(t *testing.T) {
+	mock := setupTestDB(t)
+	checker := &fakeRunningChecker{}
+	off := &recordingOffline{}
+
+	mock.ExpectQuery(`(?s)^\s*SELECT id::text\s+FROM workspaces`).
+		WithArgs(CPInstanceReconcileLimit).
+		WillReturnError(errors.New("connection refused"))
+
+	reconcileOnce(context.Background(), checker, off.handler())
+
+	if len(checker.calls) != 0 || len(off.got()) != 0 {
+		t.Fatalf("query error must short-circuit; calls=%v offline=%v", checker.calls, off.got())
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Fatalf("unmet expectations: %v", err)
+	}
+}
+
+// TestReconcileOnce_NilDB — defensive against db.DB being nil. Must not
+// panic, must not probe, must not flip.
+func TestReconcileOnce_NilDB(t *testing.T) {
+	saved := db.DB
+	db.DB = nil
+	t.Cleanup(func() { db.DB = saved })
+
+	checker := &fakeRunningChecker{}
+	off := &recordingOffline{}
+	reconcileOnce(context.Background(), checker, off.handler())
+
+	if len(checker.calls) != 0 || len(off.got()) != 0 {
+		t.Fatalf("nil db.DB must short-circuit; calls=%v offline=%v", checker.calls, off.got())
+	}
+}
+
+// TestStartCPInstanceReconciler_NilCheckerDisabled — boot-safety: a SaaS
+// CP without cpProv configured must not start the loop (immediate return,
+// no goroutine leak).
+func TestStartCPInstanceReconciler_NilCheckerDisabled(t *testing.T) {
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+
+	done := make(chan struct{})
+	go func() {
+		StartCPInstanceReconciler(ctx, nil, nil, 60*time.Second)
+		close(done)
+	}()
+	select {
+	case <-done:
+		// expected — nil checker short-circuits.
+	case <-time.After(500 * time.Millisecond):
+		t.Fatal("StartCPInstanceReconciler(nil) did not return immediately")
+	}
+}
+
+// TestStartCPInstanceReconciler_RunsOnceImmediatelyAndExitsOnCancel —
+// cadence contract: one sweep at boot (so a restart starts healing
+// immediately), and the loop terminates on ctx cancel.
+func TestStartCPInstanceReconciler_RunsOnceImmediatelyAndExitsOnCancel(t *testing.T) {
+	mock := setupTestDB(t)
+	checker := &fakeRunningChecker{}
+	off := &recordingOffline{}
+
+	// Boot sweep query. The 60s ticker won't fire inside the test window;
+	// register a second optional expectation so a stray tick can't fail.
+	expectReconcileQuery(mock, sqlmock.NewRows([]string{"id"}))
+	expectReconcileQuery(mock, sqlmock.NewRows([]string{"id"}))
+
+	ctx, cancel := context.WithCancel(context.Background())
+	done := make(chan struct{})
+	go func() {
+		StartCPInstanceReconciler(ctx, checker, off.handler(), 60*time.Second)
+		close(done)
+	}()
+	time.Sleep(100 * time.Millisecond)
+	cancel()
+	select {
+	case <-done:
+		// expected
+	case <-time.After(2 * time.Second):
+		t.Fatal("StartCPInstanceReconciler did not exit on ctx cancel")
+	}
+}

workspace-server/internal/scheduler/scheduler_integration_test.go

@@ -0,0 +1,558 @@
+//go:build integration
+// +build integration
+
+// scheduler_integration_test.go — REAL Postgres integration tests for the
+// workspace-server cron scheduler firing loop. Regression coverage for
+// molecule-core issue #2149 (filed under SOP rule internal#765).
+//
+// Run with:
+//
+//	docker run --rm -d --name pg-integration \
+//	  -e POSTGRES_PASSWORD=test -e POSTGRES_DB=molecule \
+//	  -p 55432:5432 postgres:15-alpine
+//	sleep 4
+//	# apply every migration up.sql / legacy .sql in lexicographic order
+//	for f in $(ls workspace-server/migrations/*.sql | grep -v '\.down\.sql$' | sort); do \
+//	  psql "postgres://postgres:test@localhost:55432/molecule?sslmode=disable" -f "$f"; done
+//	cd workspace-server
+//	INTEGRATION_DB_URL="postgres://postgres:test@localhost:55432/molecule?sslmode=disable" \
+//	  go test -tags=integration ./internal/scheduler/ -run '^TestIntegration_'
+//
+// CI: .gitea/workflows/handlers-postgres-integration.yml runs these on every
+// PR/push that touches workspace-server/internal/scheduler/ (the
+// `handlers-postgres` detect-changes profile was extended to include the
+// scheduler package + this workflow file).
+//
+// Why these are NOT the existing sqlmock unit tests (scheduler_test.go)
+// --------------------------------------------------------------------
+// The strict-sqlmock unit tests pin which SQL statements fire — fast, no DB.
+// But sqlmock CANNOT validate:
+//   - the activity_logs `$3::jsonb` cast (#2026 wedge) — sqlmock never parses
+//     the payload, so an invalid-UTF-8 jsonb body that wedges a real INSERT
+//     looks "green" under mock.ExpectExec(`INSERT INTO activity_logs`).
+//   - the ROW STATE after tick()/fireSchedule run: that last_run_at,
+//     next_run_at, run_count, last_status actually landed on the row.
+//   - sweepPhantomBusy's NOT IN (SELECT … activity_logs) subquery semantics
+//     against real rows — it has no unit test at all (#2149).
+//
+// A SQL regression here = a fleet-wide silent cron outage (#85 ran 12h before
+// detection). These tests boot a real Postgres, insert real rows, run the
+// production tick()/sweepPhantomBusy, and SELECT the rows back to assert the
+// observable end state — the gap sqlmock structurally cannot cover.
+//
+// Watch-fail intent: each test is written to FAIL on a regression of the
+// behavior under test (e.g. drop the activity_logs INSERT, drop the
+// write-back UPDATE, drop the UTF-8 sanitize, or break the phantom-busy
+// subquery) and to PASS against the current-correct scheduler.go.
+
+package scheduler
+
+import (
+	"context"
+	"database/sql"
+	"os"
+	"testing"
+	"time"
+
+	mdb "git.moleculesai.app/molecule-ai/molecule-core/workspace-server/internal/db"
+	_ "github.com/lib/pq"
+)
+
+// ── test doubles ──────────────────────────────────────────────────────────
+
+// recordingProxy is an A2AProxy that records each fire and returns a
+// configurable response. Used to assert that tick()/fireSchedule actually
+// reached the A2A boundary for the due schedule.
+type recordingProxy struct {
+	status int
+	body   []byte
+	err    error
+
+	fires       int
+	lastBody    []byte
+	lastCaller  string
+	lastLogFlag bool
+	lastWSID    string
+}
+
+func (p *recordingProxy) ProxyA2ARequest(
+	_ context.Context, workspaceID string, body []byte, callerID string, logActivity bool,
+) (int, []byte, error) {
+	p.fires++
+	p.lastWSID = workspaceID
+	p.lastBody = body
+	p.lastCaller = callerID
+	p.lastLogFlag = logActivity
+	if p.err != nil {
+		return 0, nil, p.err
+	}
+	return p.status, p.body, nil
+}
+
+// ── connection + fixture helpers ──────────────────────────────────────────
+
+// integrationDB returns the configured integration-test connection or skips
+// the test if INTEGRATION_DB_URL is unset. Hot-swaps the package-level
+// mdb.DB so the production scheduler helpers (tick, fireSchedule,
+// sweepPhantomBusy) operate on this connection; restores it via t.Cleanup.
+//
+// NOT SAFE FOR t.Parallel(): the package-global swap races across tests.
+func integrationDB(t *testing.T) *sql.DB {
+	t.Helper()
+	url := os.Getenv("INTEGRATION_DB_URL")
+	if url == "" {
+		t.Skip("INTEGRATION_DB_URL not set; skipping (local devs: see file header)")
+	}
+	conn, err := sql.Open("postgres", url)
+	if err != nil {
+		t.Fatalf("open: %v", err)
+	}
+	ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
+	defer cancel()
+	if err := conn.PingContext(ctx); err != nil {
+		t.Fatalf("ping: %v", err)
+	}
+	// Clean slate. activity_logs + workspace_schedules cascade off workspaces,
+	// but we DELETE explicitly (and in FK order) so a partial prior run can't
+	// leave orphan rows that perturb the next test's assertions.
+	cctx, ccancel := context.WithTimeout(context.Background(), 10*time.Second)
+	defer ccancel()
+	for _, q := range []string{
+		`DELETE FROM activity_logs`,
+		`DELETE FROM workspace_schedules`,
+		`DELETE FROM workspaces`,
+	} {
+		if _, err := conn.ExecContext(cctx, q); err != nil {
+			t.Fatalf("cleanup %q: %v", q, err)
+		}
+	}
+	prev := mdb.DB
+	mdb.DB = conn
+	t.Cleanup(func() {
+		mdb.DB = prev
+		conn.Close()
+	})
+	return conn
+}
+
+// insertWorkspace inserts a workspace row and returns its UUID. active is the
+// initial active_tasks value; status defaults to 'online' (valid workspace_status enum).
+func insertWorkspace(t *testing.T, conn *sql.DB, name string, active int) string {
+	t.Helper()
+	var id string
+	err := conn.QueryRowContext(context.Background(), `
+		INSERT INTO workspaces (name, status, active_tasks, max_concurrent_tasks)
+		VALUES ($1, 'online', $2, 1)
+		RETURNING id
+	`, name, active).Scan(&id)
+	if err != nil {
+		t.Fatalf("insertWorkspace(%s): %v", name, err)
+	}
+	return id
+}
+
+// insertSchedule inserts an enabled workspace_schedules row whose next_run_at
+// is in the past (so tick() picks it up immediately) and returns its UUID.
+func insertSchedule(t *testing.T, conn *sql.DB, wsID, name, cronExpr, prompt string) string {
+	t.Helper()
+	var id string
+	err := conn.QueryRowContext(context.Background(), `
+		INSERT INTO workspace_schedules
+		    (workspace_id, name, cron_expr, timezone, prompt, enabled, next_run_at, source)
+		VALUES ($1, $2, $3, 'UTC', $4, true, now() - interval '1 minute', 'runtime')
+		RETURNING id
+	`, wsID, name, cronExpr, prompt).Scan(&id)
+	if err != nil {
+		t.Fatalf("insertSchedule(%s): %v", name, err)
+	}
+	return id
+}
+
+type scheduleState struct {
+	lastRunAt  sql.NullTime
+	nextRunAt  sql.NullTime
+	runCount   int
+	lastStatus string
+	lastError  string
+}
+
+func readScheduleState(t *testing.T, conn *sql.DB, id string) scheduleState {
+	t.Helper()
+	var st scheduleState
+	var status, errStr sql.NullString
+	err := conn.QueryRowContext(context.Background(), `
+		SELECT last_run_at, next_run_at, run_count, last_status, last_error
+		FROM workspace_schedules WHERE id = $1
+	`, id).Scan(&st.lastRunAt, &st.nextRunAt, &st.runCount, &status, &errStr)
+	if err != nil {
+		t.Fatalf("readScheduleState(%s): %v", id, err)
+	}
+	st.lastStatus = status.String
+	st.lastError = errStr.String
+	return st
+}
+
+// ── TestIntegration_TickFiresAndWritesBack (#2149 core) ───────────────────
+//
+// Insert one due schedule, run tick() once, and assert the full firing
+// loop landed against a REAL Postgres:
+//   - the A2A proxy was invoked exactly once for the schedule's workspace
+//   - the post-fire UPDATE wrote last_run_at (was NULL), advanced next_run_at
+//     into the future, bumped run_count to 1, set last_status='ok'
+//   - a cron_run activity_logs row was inserted with VALID jsonb request_body
+//     (the `$3::jsonb` cast #2026 path) carrying the schedule metadata
+//
+// Regression watch-fail: if a refactor drops the write-back UPDATE, the
+// activity_logs INSERT, or breaks the jsonb cast, this test fails where every
+// sqlmock unit test stays green.
+func TestIntegration_TickFiresAndWritesBack(t *testing.T) {
+	conn := integrationDB(t)
+
+	wsID := insertWorkspace(t, conn, "cron-fire-ws", 0)
+	schedID := insertSchedule(t, conn, wsID, "hourly-audit", "0 * * * *", "run the hourly audit")
+
+	proxy := &recordingProxy{
+		status: 200,
+		body:   []byte(`{"jsonrpc":"2.0","result":{"kind":"message","parts":[{"kind":"text","text":"done"}]},"id":"1"}`),
+	}
+	s := New(proxy, nil)
+	s.tick(context.Background())
+
+	// 1. A2A boundary reached exactly once for the right workspace.
+	if proxy.fires != 1 {
+		t.Fatalf("proxy fires = %d, want 1 (tick must fire the one due schedule)", proxy.fires)
+	}
+	if proxy.lastWSID != wsID {
+		t.Errorf("proxy fired for workspace %q, want %q", proxy.lastWSID, wsID)
+	}
+	// Empty callerID = canvas-style (bypasses access control); logActivity=true.
+	if proxy.lastCaller != "" {
+		t.Errorf("callerID = %q, want empty (canvas-style scheduler fire)", proxy.lastCaller)
+	}
+	if !proxy.lastLogFlag {
+		t.Error("logActivity flag = false, want true")
+	}
+
+	// 2. Row write-back.
+	st := readScheduleState(t, conn, schedID)
+	if !st.lastRunAt.Valid {
+		t.Error("last_run_at is NULL after fire, want set (write-back UPDATE did not land)")
+	}
+	if !st.nextRunAt.Valid {
+		t.Fatal("next_run_at is NULL after fire, want a future timestamp")
+	}
+	if !st.nextRunAt.Time.After(time.Now()) {
+		t.Errorf("next_run_at = %v, want a time in the future (schedule would tight-loop otherwise)", st.nextRunAt.Time)
+	}
+	if st.runCount != 1 {
+		t.Errorf("run_count = %d, want 1", st.runCount)
+	}
+	if st.lastStatus != "ok" {
+		t.Errorf("last_status = %q, want \"ok\"", st.lastStatus)
+	}
+
+	// 3. activity_logs cron_run row with valid jsonb request_body.
+	var actCount int
+	var summary, status string
+	var reqBody []byte
+	err := conn.QueryRowContext(context.Background(), `
+		SELECT count(*) OVER (), summary, status, request_body
+		FROM activity_logs
+		WHERE workspace_id = $1 AND activity_type = 'cron_run'
+		LIMIT 1
+	`, wsID).Scan(&actCount, &summary, &status, &reqBody)
+	if err == sql.ErrNoRows {
+		t.Fatal("no cron_run activity_logs row inserted after fire (#152/#2026 path missing)")
+	}
+	if err != nil {
+		t.Fatalf("read activity_logs: %v", err)
+	}
+	if actCount != 1 {
+		t.Errorf("cron_run activity_logs rows = %d, want 1", actCount)
+	}
+	if status != "ok" {
+		t.Errorf("activity_logs.status = %q, want \"ok\"", status)
+	}
+	// request_body must be valid jsonb carrying the schedule_id — proves the
+	// `$3::jsonb` cast accepted the payload (the #2026 wedge surface).
+	var sid string
+	if err := conn.QueryRowContext(context.Background(), `
+		SELECT request_body->>'schedule_id'
+		FROM activity_logs WHERE workspace_id = $1 AND activity_type = 'cron_run' LIMIT 1
+	`, wsID).Scan(&sid); err != nil {
+		t.Fatalf("request_body is not queryable jsonb: %v", err)
+	}
+	if sid != schedID {
+		t.Errorf("activity_logs request_body->>'schedule_id' = %q, want %q", sid, schedID)
+	}
+}
+
+// ── TestIntegration_InvalidUTF8PromptSanitizedIntoJsonb (#2026 / #2149) ────
+//
+// The agent-editable prompt can carry raw invalid-UTF-8 bytes. Postgres jsonb
+// columns REJECT invalid UTF-8, which (pre-#2026) wedged the activity_logs
+// INSERT and held the transaction open — stalling the whole scheduler.
+// fireSchedule now sanitizeUTF8()s every string before the `$3::jsonb` insert.
+//
+// Postgres TEXT columns (workspace_schedules.prompt) also reject invalid UTF-8
+// in a UTF-8 database, so we cannot INSERT the bad bytes through the fixture.
+// Instead we insert a valid prompt, then call fireSchedule directly with a
+// scheduleRow whose Prompt field contains the invalid bytes — this simulates
+// the real regression path (e.g. truncation splitting a multi-byte rune, or
+// an agent-edited template arriving via a path that bypasses DB validation).
+//
+// Assertions:
+//   - the fire still completed (write-back UPDATE landed)
+//   - the cron_run activity_logs row was inserted (the jsonb cast accepted
+//     the SANITIZED payload — the INSERT did not wedge)
+//   - the stored request_body is queryable jsonb (valid UTF-8 on disk)
+//
+// Watch-fail: remove the sanitizeUTF8() wrapping around the jsonb payload and
+// this test fails on a real Postgres (INSERT errors / row absent), while the
+// sqlmock unit test that only checks "an INSERT fired" stays green.
+func TestIntegration_InvalidUTF8PromptSanitizedIntoJsonb(t *testing.T) {
+	conn := integrationDB(t)
+
+	wsID := insertWorkspace(t, conn, "utf8-ws", 0)
+	// Insert with valid UTF-8 — Postgres TEXT rejects 0x80/0xff.
+	schedID := insertSchedule(t, conn, wsID, "utf8-job", "0 * * * *", "valid prompt")
+
+	// Prompt with invalid UTF-8: orphan continuation byte + bare 0xff.
+	badPrompt := "audit \x80 report \xff end"
+	row := scheduleRow{
+		ID:          schedID,
+		WorkspaceID: wsID,
+		Name:        "utf8-job",
+		CronExpr:    "0 * * * *",
+		Timezone:    "UTC",
+		Prompt:      badPrompt,
+	}
+
+	proxy := &recordingProxy{
+		status: 200,
+		body:   []byte(`{"result":{"kind":"message","parts":[{"kind":"text","text":"ok"}]}}`),
+	}
+	s := New(proxy, nil)
+	s.fireSchedule(context.Background(), row)
+
+	if proxy.fires != 1 {
+		t.Fatalf("proxy fires = %d, want 1", proxy.fires)
+	}
+
+	// Write-back must have landed despite the bad prompt bytes.
+	st := readScheduleState(t, conn, schedID)
+	if st.runCount != 1 || st.lastStatus != "ok" {
+		t.Errorf("post-fire state run_count=%d last_status=%q, want 1/\"ok\" "+
+			"(invalid-UTF-8 prompt must not block the fire)", st.runCount, st.lastStatus)
+	}
+
+	// The cron_run activity_logs row MUST exist — proving the `$3::jsonb`
+	// INSERT accepted the sanitized payload (did not wedge on invalid UTF-8).
+	var n int
+	if err := conn.QueryRowContext(context.Background(), `
+		SELECT count(*) FROM activity_logs
+		WHERE workspace_id = $1 AND activity_type = 'cron_run'
+	`, wsID).Scan(&n); err != nil {
+		t.Fatalf("count cron_run rows: %v", err)
+	}
+	if n != 1 {
+		t.Fatalf("cron_run activity_logs rows = %d, want 1 — the jsonb INSERT wedged "+
+			"on invalid UTF-8 (the #2026 regression)", n)
+	}
+
+	// The stored prompt inside request_body must be queryable + valid UTF-8.
+	var storedPrompt string
+	if err := conn.QueryRowContext(context.Background(), `
+		SELECT request_body->>'prompt'
+		FROM activity_logs WHERE workspace_id = $1 AND activity_type = 'cron_run' LIMIT 1
+	`, wsID).Scan(&storedPrompt); err != nil {
+		t.Fatalf("request_body->>'prompt' not queryable jsonb: %v", err)
+	}
+	if storedPrompt == "" {
+		t.Error("stored prompt is empty, want the sanitized prompt text")
+	}
+	// Round-trip through Postgres jsonb guarantees valid UTF-8; assert the
+	// replacement character replaced the bad bytes rather than them surviving.
+	for i := 0; i < len(storedPrompt); i++ {
+		if storedPrompt[i] == 0x80 || storedPrompt[i] == 0xff {
+			t.Fatalf("stored prompt still contains raw invalid byte 0x%x at %d", storedPrompt[i], i)
+		}
+	}
+}
+
+// ── TestIntegration_TickErrorStatusWriteBack (#2149) ──────────────────────
+//
+// When the A2A proxy returns a transport error, fireSchedule must still write
+// back: last_status='error', last_error populated, next_run_at advanced (so
+// the schedule does not get stuck re-firing), run_count bumped. Verifies the
+// error path persists to a real row, not just that "an UPDATE fired".
+func TestIntegration_TickErrorStatusWriteBack(t *testing.T) {
+	conn := integrationDB(t)
+
+	wsID := insertWorkspace(t, conn, "err-ws", 0)
+	schedID := insertSchedule(t, conn, wsID, "err-job", "0 * * * *", "do work")
+
+	proxy := &recordingProxy{err: context.DeadlineExceeded}
+	s := New(proxy, nil)
+	s.tick(context.Background())
+
+	st := readScheduleState(t, conn, schedID)
+	if st.lastStatus != "error" {
+		t.Errorf("last_status = %q, want \"error\"", st.lastStatus)
+	}
+	if st.lastError == "" {
+		t.Error("last_error is empty, want the proxy error text persisted (#152)")
+	}
+	if st.runCount != 1 {
+		t.Errorf("run_count = %d, want 1 (run still counted on error)", st.runCount)
+	}
+	if !st.nextRunAt.Valid || !st.nextRunAt.Time.After(time.Now()) {
+		t.Errorf("next_run_at not advanced to future on error path (= %v) — schedule would tight-loop", st.nextRunAt)
+	}
+	// The error activity_logs row must carry status='error' + error_detail.
+	var status, errDetail string
+	if err := conn.QueryRowContext(context.Background(), `
+		SELECT status, COALESCE(error_detail,'') FROM activity_logs
+		WHERE workspace_id = $1 AND activity_type = 'cron_run' LIMIT 1
+	`, wsID).Scan(&status, &errDetail); err != nil {
+		t.Fatalf("read error activity_logs: %v", err)
+	}
+	if status != "error" {
+		t.Errorf("activity_logs.status = %q, want \"error\"", status)
+	}
+	if errDetail == "" {
+		t.Error("activity_logs.error_detail empty on error fire, want the error message (#152)")
+	}
+}
+
+// ── TestIntegration_SweepPhantomBusy (#2149 — no prior test) ──────────────
+//
+// sweepPhantomBusy resets active_tasks=0 for workspaces stuck busy with NO
+// activity_logs row in the last phantomStaleThreshold window, and must LEAVE
+// ALONE workspaces that have recent activity. The NOT IN (SELECT DISTINCT
+// workspace_id FROM activity_logs WHERE created_at > now() - interval) subquery
+// is exactly the kind of set-semantics that sqlmock cannot validate — there is
+// no unit test for this method at all (#2149).
+//
+// Fixture:
+//   - phantomWS: active_tasks=3, NO recent activity_log     → must reset to 0
+//   - recentWS:  active_tasks=2, activity_log 1 min ago      → must stay at 2
+//   - staleWS:   active_tasks=1, activity_log 30 min ago     → must reset to 0
+//   - removedWS: active_tasks=4, status='removed', no activity → must stay (status guard)
+//   - idleWS:    active_tasks=0                               → untouched (not >0)
+//
+// Watch-fail: break the subquery (e.g. drop the status!='removed' guard, or
+// invert the NOT IN), and the asserted end-state diverges on a real Postgres.
+func TestIntegration_SweepPhantomBusy(t *testing.T) {
+	conn := integrationDB(t)
+
+	phantomWS := insertWorkspace(t, conn, "phantom-ws", 3)
+	recentWS := insertWorkspace(t, conn, "recent-ws", 2)
+	staleWS := insertWorkspace(t, conn, "stale-ws", 1)
+	idleWS := insertWorkspace(t, conn, "idle-ws", 0)
+
+	// removedWS: busy but status='removed' — the sweep must skip it.
+	var removedWS string
+	if err := conn.QueryRowContext(context.Background(), `
+		INSERT INTO workspaces (name, status, active_tasks, max_concurrent_tasks)
+		VALUES ('removed-ws', 'removed', 4, 1) RETURNING id
+	`).Scan(&removedWS); err != nil {
+		t.Fatalf("insert removedWS: %v", err)
+	}
+
+	// recentWS has a fresh activity_log (1 min ago → inside the 10-min window).
+	if _, err := conn.ExecContext(context.Background(), `
+		INSERT INTO activity_logs (workspace_id, activity_type, status, created_at)
+		VALUES ($1, 'a2a_receive', 'ok', now() - interval '1 minute')
+	`, recentWS); err != nil {
+		t.Fatalf("insert recent activity_log: %v", err)
+	}
+	// staleWS has only an OLD activity_log (30 min ago → outside the window).
+	if _, err := conn.ExecContext(context.Background(), `
+		INSERT INTO activity_logs (workspace_id, activity_type, status, created_at)
+		VALUES ($1, 'a2a_receive', 'ok', now() - interval '30 minutes')
+	`, staleWS); err != nil {
+		t.Fatalf("insert stale activity_log: %v", err)
+	}
+
+	s := New(nil, nil)
+	s.sweepPhantomBusy(context.Background())
+
+	active := func(id string) int {
+		var n int
+		if err := conn.QueryRowContext(context.Background(),
+			`SELECT active_tasks FROM workspaces WHERE id = $1`, id).Scan(&n); err != nil {
+			t.Fatalf("read active_tasks(%s): %v", id, err)
+		}
+		return n
+	}
+
+	if got := active(phantomWS); got != 0 {
+		t.Errorf("phantomWS active_tasks = %d, want 0 (busy + no recent activity → must be swept)", got)
+	}
+	if got := active(staleWS); got != 0 {
+		t.Errorf("staleWS active_tasks = %d, want 0 (only stale activity → must be swept)", got)
+	}
+	if got := active(recentWS); got != 2 {
+		t.Errorf("recentWS active_tasks = %d, want 2 (recent activity → must NOT be swept)", got)
+	}
+	if got := active(removedWS); got != 4 {
+		t.Errorf("removedWS active_tasks = %d, want 4 (status='removed' → sweep must skip it)", got)
+	}
+	if got := active(idleWS); got != 0 {
+		t.Errorf("idleWS active_tasks = %d, want 0 (was never busy)", got)
+	}
+
+	// The swept rows must also have current_task cleared.
+	var ct string
+	if err := conn.QueryRowContext(context.Background(),
+		`SELECT COALESCE(current_task,'') FROM workspaces WHERE id = $1`, phantomWS).Scan(&ct); err != nil {
+		t.Fatalf("read current_task: %v", err)
+	}
+	if ct != "" {
+		t.Errorf("phantomWS current_task = %q, want empty after sweep", ct)
+	}
+}
+
+// ── TestIntegration_NativeSchedulerSkipAdvancesNextRunAt (#2149) ──────────
+//
+// When a workspace's adapter owns scheduling natively, tick() must SKIP the
+// fire but still advance next_run_at (so the row doesn't tight-loop on every
+// poll) — observability (next_run_at) is preserved while the fire is dropped.
+// Asserts the native-skip UPDATE landed on a real row and the proxy was NOT
+// invoked. This is the native-skip UPDATE path #2149 calls out — sqlmock can
+// only assert an UPDATE fired, not that next_run_at moved forward.
+func TestIntegration_NativeSchedulerSkipAdvancesNextRunAt(t *testing.T) {
+	conn := integrationDB(t)
+
+	wsID := insertWorkspace(t, conn, "native-ws", 0)
+	schedID := insertSchedule(t, conn, wsID, "native-job", "0 * * * *", "native run")
+
+	// Capture the pre-tick next_run_at (it is in the past by construction).
+	before := readScheduleState(t, conn, schedID)
+	if !before.nextRunAt.Valid || before.nextRunAt.Time.After(time.Now()) {
+		t.Fatalf("precondition: next_run_at should start in the past, got %v", before.nextRunAt)
+	}
+
+	proxy := &recordingProxy{status: 200, body: []byte(`{}`)}
+	s := New(proxy, nil)
+	// Every workspace reports native scheduling → fire must be skipped.
+	s.SetNativeSchedulerCheck(func(string) bool { return true })
+	s.tick(context.Background())
+
+	if proxy.fires != 0 {
+		t.Errorf("proxy fires = %d, want 0 (native-scheduler workspace must NOT fire)", proxy.fires)
+	}
+
+	after := readScheduleState(t, conn, schedID)
+	if !after.nextRunAt.Valid || !after.nextRunAt.Time.After(time.Now()) {
+		t.Errorf("next_run_at = %v, want advanced into the future (native-skip UPDATE must still run)", after.nextRunAt)
+	}
+	// Skip path does NOT bump run_count or write last_run_at (no fire happened).
+	if after.runCount != 0 {
+		t.Errorf("run_count = %d, want 0 (skip must not count as a run)", after.runCount)
+	}
+	if after.lastRunAt.Valid {
+		t.Error("last_run_at set on native-skip, want NULL (no fire occurred)")
+	}
+}

workspace-server/migrations/20260604000000_activity_logs_seq.down.sql

@@ -0,0 +1,9 @@
+-- Rollback for 20260604000000_activity_logs_seq.up.sql.
+-- Drops the feed-ordering index and the monotonic seq column.
+-- Run manually by an operator via psql; the boot-time runner never applies
+-- *.down.sql (see RunMigrations in internal/db/postgres.go, issue #211).
+
+DROP INDEX IF EXISTS idx_activity_ws_created_seq;
+
+ALTER TABLE activity_logs
+    DROP COLUMN IF EXISTS seq;

workspace-server/migrations/20260604000000_activity_logs_seq.up.sql

@@ -0,0 +1,54 @@
+-- Add a monotonic `seq` tiebreaker to activity_logs to make the poll-mode
+-- since_id activity feed (#2339) deterministically ordered.
+--
+-- ROOT CAUSE this fixes: the feed orders by created_at ASC/DESC with NO
+-- tiebreaker, and activity_logs.id is a random gen_random_uuid() — there is
+-- no monotonic column to break ties. Two rows inserted in the same
+-- microsecond (back-to-back A2A logging) share a created_at and come back in
+-- arbitrary planner order, so the E2E intermittently sees
+-- hello-from-e2e-3 before hello-from-e2e-2. Not a flake — a missing
+-- tiebreaker. (Second, related bug fixed in the handler: the since_id cursor
+-- filtered `created_at > cursor` strictly, silently dropping a row written in
+-- the same microsecond as the cursor row. The composite key below lets the
+-- handler compare the full (created_at, seq) tuple.)
+--
+-- `seq` is a GENERATED BY DEFAULT AS IDENTITY BIGINT — a UNIQUE,
+-- monotonic-once-assigned tiebreaker. Precisely (verified on PostgreSQL
+-- 16.13, the prod version):
+--   * Backfill: adding the IDENTITY column to a populated table REWRITES the
+--     table and assigns `seq` to every EXISTING row during the ALTER, in
+--     PHYSICAL TABLE-SCAN order (NOT NULL — existing rows do get a value).
+--     That order is not guaranteed to equal historical insertion order.
+--   * The identity sequence then advances ABOVE max(seq), so every subsequent
+--     INSERT that omits `seq` gets a fresh value strictly greater than the
+--     backfilled max — collision-free with the backfilled rows.
+--   * GENERATED BY DEFAULT (not ALWAYS) so existing INSERTs that don't name
+--     `seq` keep working and a caller may still override it if ever needed.
+--
+-- What `seq` is NOT, and why that's fine:
+--   * NOT guaranteed gap-free — rolled-back transactions burn sequence values.
+--   * NOT a strict commit-order guarantee under concurrency — two concurrent
+--     INSERTs may commit in the opposite order to the `seq` values they drew.
+-- Neither property is needed. The feed only requires a TOTAL, STABLE
+-- tiebreaker so that (created_at, seq) is a deterministic order: for any two
+-- rows it always sorts them the same way and never ties. `seq` being unique
+-- and non-null on every row delivers exactly that. Same-created_at rows were
+-- returned in ARBITRARY order before this migration; afterward they have a
+-- fixed, repeatable order — strictly better, never worse. New traffic is fully
+-- deterministic; the backfill makes historical rows deterministic too.
+--
+-- Idempotent: ADD COLUMN IF NOT EXISTS + CREATE INDEX IF NOT EXISTS so the
+-- boot-time runner (and the CI migrate-replay step) can re-apply this safely.
+
+ALTER TABLE activity_logs
+    ADD COLUMN IF NOT EXISTS seq BIGINT GENERATED BY DEFAULT AS IDENTITY;
+
+-- Composite index supporting the feed query: WHERE workspace_id = $1
+-- AND created_at <cmp> $t ORDER BY created_at, seq. The (workspace_id,
+-- created_at, seq) prefix serves both the ASC cursor path and the DESC recent
+-- path (Postgres reads the same btree backwards for DESC). This is distinct
+-- from migration 009's idx_activity_ws_type_time (workspace_id, activity_type,
+-- created_at) — that one is type-prefixed and can't drive a type-agnostic feed
+-- scan — and from 048's per-peer source_id/target_id indexes.
+CREATE INDEX IF NOT EXISTS idx_activity_ws_created_seq
+    ON activity_logs (workspace_id, created_at, seq);