feat(issue-5): auto-completion gate and durable pending notification

The gate does not re-decide anything. Phase 2a already decides, fail-closed and
against contract-pinned vectors; this supplies real evidence to that decision and
commits only if three independent gates agree: the global switch is on, the
backend has observed evidence that it reaches a quiescent state, and the job
itself opted in. Default is off.

The backend allowlist is evidence-based rather than optimistic. Inferring
"finished" from stillness requires stillness to have been observed, and only
three of six backends have that. An operator may narrow the list; a name that
was never observed is dropped rather than trusted, so it cannot be widened by
configuration.

Everything else holds. An unreadable manifest, a report rewritten after
finalization, missing evidence for a declared side effect, a policy that raises:
all refuse. "Could not decide" is never "done". A late explicit terminal wins,
because the boundary refuses the inferred one -- that is the boundary working,
not an error in the gate.

Notification intent is committed in the same atomic write as the state, so a
crash between committing and publishing leaves a record that still says the
world has not been told. Recovery re-sends; duplicates are already dropped by
sequence. A failed send deliberately leaves the intent pending, because an
unsent notification that looks sent is worse than one that is retried.

Two fixes this surfaced. Settling used a request that terminal records refuse,
which made an owed notification impossible to ever settle; settling is now a
status-free bookkeeping request, valid in any state. And the phase 2a guard
against shell references was matching hashlib.sha256 as a substring, so it now
requires a word boundary.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UrDNJkHmm18uu7n7JwV5Qd

.agents/skills/lib_py/delegation/engine.py

@@ -55,8 +55,16 @@ REQUEST_TO_STATUS = {
     # separate requests gated on authority rather than an exception in the table.
     "redispatch": "pending",
     "finalize": "error",
+    # Bookkeeping: records that an owed notification was delivered. It cannot
+    # change status, and it is *about* a terminal, so refusing it in a terminal
+    # state would make an owed notification impossible to ever settle.
+    "notification_settled": None,
 }
 
+# Requests that carry no status at all and are therefore safe in any state,
+# including a terminal one. Keep this list short and obviously status-free.
+BOOKKEEPING_REQUESTS = ("notification_settled",)
+
 # Only these sources may move a record out of a terminal state. An agent can
 # never walk back its own terminal; that is the whole point of the CAS table.
 OPERATOR_SOURCES = ("registry_cli", "operator")
@@ -160,6 +168,9 @@ def _decide(record: Dict[str, Any], request: str, expected_status, expected_atte
         return _Decision(False, current,
                          "{} requires operator authority (source={!r})".format(request, source))
 
+    if request in BOOKKEEPING_REQUESTS:
+        return _Decision(True, None, "bookkeeping ({})".format(request))
+
     if current in TERMINAL_STATUSES:
         # Idempotent re-receipt of the same terminal is a no-op ack.
         if requested == current:
@@ -207,7 +218,8 @@ def _notification(job_id, kind, **fields) -> Dict[str, Any]:
 
 
 def transition(job_id, request="merge", *, registry_dir=None, expected_status=None,
-               expected_attempt=None, fields=None, source="unknown") -> TransitionResult:
+               expected_attempt=None, fields=None, source="unknown",
+               notify=None) -> TransitionResult:
     """Decide and commit one state transition. The only writer of job status.
 
     Returns a result whose `notifications` have ALREADY been flushed; the field
@@ -231,7 +243,15 @@ def transition(job_id, request="merge", *, registry_dir=None, expected_status=No
             record["updated_at"] = mqtt_common._utcnow()
             # A monotonic counter so a late flush can still be ordered correctly
             # once the audit mirror happens outside the lock.
-            record["transition_seq"] = int(record.get("transition_seq", 0)) + 1
+            record["transition_seq"] = int(record.get("transition_seq", 0) or 0) + 1
+            if notify is not None:
+                # One atomic commit records both the new state and the fact that
+                # the world has not been told yet. If the process dies before the
+                # publish, the record still says a notification is owed.
+                from lib_py.delegation import notifier
+                record[notifier.FIELD] = notifier.intent_for(
+                    notify, record["transition_seq"], record["updated_at"],
+                    detail=str(fields.get("detail", "")))
             mqtt_common._atomic_write_record(job_id, registry_dir, record)
 
             to_status = record.get("status")

.agents/skills/lib_py/delegation/notifier.py

@@ -0,0 +1,97 @@
+"""Durable notification intent, carried in the job record.
+
+The contract forbids a separate outbox store, so the intent rides in the record
+the transition already writes: one atomic commit records both the new state and
+the fact that the world has not been told yet. Publishing then happens outside
+the registry lock, and settling the intent is itself an ordinary transition.
+
+That ordering is the point. If the process dies between the commit and the
+publish, the record still says a notification is owed, and startup recovery
+re-sends it. Duplicates are safe: the subscriber already drops any event whose
+sequence is not greater than the highest it has accepted.
+"""
+from __future__ import annotations
+
+import logging
+from typing import Any, Dict, List, Optional
+
+from lib_py.delegation import mqtt_common
+
+logger = logging.getLogger("delegate_job.notifier")
+
+FIELD = "pending_notification"
+
+
+def intent_for(event: str, seq: int, timestamp: str, detail: str = "") -> Dict[str, Any]:
+    return {"event": event, "seq": seq, "timestamp": timestamp, "detail": detail}
+
+
+def pending(job_id: str, registry_dir: str) -> Optional[Dict[str, Any]]:
+    try:
+        return mqtt_common.load_job(job_id, registry_dir).get(FIELD)
+    except FileNotFoundError:
+        return None
+
+
+def jobs_with_pending(registry_dir: str) -> List[str]:
+    """Records still owing a notification. The input to restart recovery."""
+    from lib_py.delegation import registry  # local: keeps module import order flat
+
+    out = []
+    for record in registry.list_jobs(registry_dir):
+        if record.get(FIELD):
+            out.append(record["job_id"])
+    return out
+
+
+def settle(job_id: str, registry_dir: str, source: str = "notifier") -> bool:
+    """Clear the intent once the event is on the wire.
+
+    Goes through the transition boundary like every other record write, using a
+    request that carries no status: settling is about a terminal, so a request
+    the terminal state would refuse could never be used to settle one.
+    """
+    from lib_py.delegation import engine
+
+    result = engine.transition(job_id, "notification_settled", registry_dir=registry_dir,
+                               fields={FIELD: None}, source=source)
+    return result.committed
+
+
+def flush(job_id: str, registry_dir: str, publish=None, attempts: int = 3) -> bool:
+    """Publish an owed notification and settle it. Safe to call repeatedly.
+
+    Returns True when nothing is owed afterwards. A failure leaves the intent in
+    place on purpose: an unsent notification that looks sent is worse than one
+    that is retried.
+    """
+    intent = pending(job_id, registry_dir)
+    if not intent:
+        return True
+
+    if publish is None:
+        from lib_py.delegation import publish_event as _pe
+
+        def publish(job, intent_, registry_dir_):
+            return _pe.main(["--registry-dir", registry_dir_, "--job", job,
+                             "--event", intent_["event"], "--detail",
+                             intent_.get("detail", ""), "--attempts", str(attempts)])
+
+    try:
+        rc = publish(job_id, intent, registry_dir)
+    except Exception as exc:
+        logger.warning("notification publish failed for %s: %s", job_id, exc)
+        return False
+    if rc != 0:
+        logger.warning("notification for %s not accepted (rc=%s); leaving it pending",
+                       job_id, rc)
+        return False
+    return settle(job_id, registry_dir)
+
+
+def recover(registry_dir: str, publish=None) -> Dict[str, bool]:
+    """Re-send every owed notification. Called at startup after a crash."""
+    results = {}
+    for job_id in jobs_with_pending(registry_dir):
+        results[job_id] = flush(job_id, registry_dir, publish=publish)
+    return results

.agents/skills/lib_py/loop/autocomplete.py

@@ -0,0 +1,137 @@
+"""The auto-completion gate.
+
+Phase 2a already decides, correctly and fail-closed. This does not re-decide;
+it supplies real evidence to that decision and, only if three independent gates
+agree, asks the transition boundary to commit.
+
+Three gates, all of which must hold:
+  * the global switch is on (default off),
+  * the backend has observed evidence that it reaches a quiescent state,
+  * the job itself opted in.
+
+Everything else is refusal. A failure anywhere -- an unreadable manifest, an
+exception in the policy, an unknown decision -- holds. "Could not decide" is
+never "done".
+"""
+from __future__ import annotations
+
+import logging
+import os
+from typing import Any, Dict, List, Optional
+
+from lib_py.loop import finalize, policy, store
+
+logger = logging.getLogger("loop.autocomplete")
+
+FLAG = "MAM_AUTO_COMPLETE"
+BACKENDS_FLAG = "MAM_AUTO_COMPLETE_BACKENDS"
+
+# Backends with recorded evidence that they reach a quiescent state. Auto
+# completion infers "finished" from stillness, so a backend whose stillness has
+# never been observed has no basis for the inference. agy has only been seen
+# working; hermes and opencode have no observations at all.
+EVIDENCE_BACKED_BACKENDS = ("claude", "grok", "codex")
+
+SOURCE = "auto_complete"
+
+
+def enabled() -> bool:
+    """Default off. Read per call so an operator can flip it without a restart."""
+    return os.environ.get(FLAG, "0") == "1"
+
+
+def allowed_backends() -> tuple:
+    """The allowlist may be narrowed by an operator, never widened.
+
+    Widening would need new observations, so an unknown name in the environment
+    is dropped rather than trusted.
+    """
+    raw = os.environ.get(BACKENDS_FLAG)
+    if not raw:
+        return EVIDENCE_BACKED_BACKENDS
+    requested = [name.strip() for name in raw.split(",") if name.strip()]
+    return tuple(n for n in requested if n in EVIDENCE_BACKED_BACKENDS)
+
+
+class GateResult(object):
+    __slots__ = ("committed", "decision", "reason")
+
+    def __init__(self, committed, decision, reason):
+        self.committed = committed
+        self.decision = decision
+        self.reason = reason
+
+    def __repr__(self):  # pragma: no cover - debugging aid
+        return "GateResult(committed={!r}, {!r}, {!r})".format(
+            self.committed, self.decision, self.reason)
+
+
+def _gates_open(record: Dict[str, Any]) -> Optional[str]:
+    """None when all three gates pass, else why they did not."""
+    if not enabled():
+        return "auto-completion is disabled ({}=0)".format(FLAG)
+    backend = record.get("agent") or ""
+    # `agent` may be a delegate key such as `claude-code`; match on the prefix.
+    if not any(backend == b or backend.startswith(b) for b in allowed_backends()):
+        return "backend {!r} has no observed quiescent-state evidence".format(backend)
+    if not record.get("auto_complete"):
+        return "job did not opt in to auto-completion"
+    return None
+
+
+def evaluate(job_id, scenario, registry_dir=None, workspace=None,
+             report_dir=None) -> GateResult:
+    """Decide, and commit only if every gate agrees.
+
+    `scenario` is the composite evidence the watcher assembled, in the same shape
+    the phase 0 fixtures use, so the decision is made by exactly the code the
+    scenario contracts already pin.
+    """
+    from lib_py.delegation import engine, mqtt_common
+
+    registry_dir = registry_dir or mqtt_common.DEFAULT_REGISTRY_DIR
+    try:
+        record = mqtt_common.load_job(job_id, registry_dir)
+    except FileNotFoundError:
+        return GateResult(False, "hold", "no such job")
+
+    closed = _gates_open(record)
+    if closed:
+        return GateResult(False, "hold", closed)
+
+    # The manifest is verified against the file on disk, not against itself, and
+    # against any side effects this job declared it would perform.
+    dispatch = store.load_dispatch(job_id, workspace) or {}
+    attempt = dispatch.get("attempt_id")
+    if report_dir is not None:
+        verdict = finalize.verify_manifest(
+            report_dir, job_id, expected_attempt=attempt,
+            expected_artifacts=record.get("expected_artifacts") or [])
+        if not verdict.ok:
+            return GateResult(False, "hold", verdict.reason)
+
+    try:
+        decision = policy.decide(scenario)
+    except Exception as exc:  # a policy that cannot decide has not approved
+        logger.warning("policy raised for %s: %s", job_id, exc)
+        return GateResult(False, "hold", "policy error: {}".format(exc))
+
+    if not decision.would_complete:
+        return GateResult(False, decision.decision, decision.reason)
+
+    # The attempt guard lives here, in the evidence, not in the record: a record
+    # carries a loop-iteration counter, which is a different thing.
+    result = engine.transition(
+        job_id, "completed", registry_dir=registry_dir,
+        expected_status="running", source=SOURCE,
+        fields={"completed_by": SOURCE, "auto_complete_reason": decision.reason},
+        notify="completed",
+    )
+    if not result.committed:
+        # A late explicit terminal, or a job that moved on. The boundary refused;
+        # that is the boundary doing its job, not an error here.
+        logger.info("auto-completion declined for %s: %s", job_id, result.reason)
+        return GateResult(False, "hold", result.reason)
+
+    logger.info("auto-completed %s (%s)", job_id, decision.reason)
+    return GateResult(True, decision.decision, decision.reason)