fix(kanban): auto-block workers that exit without completing (NousResearch#20894) (NousResearch#21214)

When a kanban worker subprocess exits rc=0 but its task is still in
status='running', the agent almost certainly answered the task
conversationally without calling kanban_complete or kanban_block. The
dispatcher used to classify this as a generic crash and respawn, which
loops forever on small local models (gemma4-e2b q4 etc.) that keep
returning clean but unproductive output.

Dispatcher changes:
- The waitpid reap loop at the top of dispatch_once now records each
  reaped child's raw exit status in a bounded module registry
  (_recent_worker_exits, TTL 600s, size cap 4096).
- _classify_worker_exit distinguishes clean_exit / nonzero_exit /
  signaled / unknown using os.WIFEXITED / WIFSIGNALED.
- detect_crashed_workers consults the classification when a worker
  is found dead. clean_exit → protocol_violation event + immediate
  circuit-breaker trip (failure_limit=1). Everything else keeps the
  existing crashed-event + counter behavior.
- DispatchResult.auto_blocked now includes protocol-violation trips.

Gateway fix (Bug A in NousResearch#20894):
- gateway.run._notify_active_sessions_of_shutdown snapshots
  self.adapters with list(...) before iterating. adapter.send() can
  hit a fatal-error path that pops the adapter from the dict, which
  was raising 'RuntimeError: dictionary changed size during iteration'
  during shutdown.

Regression tests:
- test_detect_crashed_workers_protocol_violation_auto_blocks verifies
  rc=0 + still-running → status=blocked on first occurrence with
  protocol_violation + gave_up events and NO crashed event.
- test_detect_crashed_workers_nonzero_exit_uses_default_limit verifies
  non-zero exits keep the existing 2-strike behavior.

Closes NousResearch#20894.

Author: teknium1

gateway/run.py

@@ -2521,7 +2521,12 @@ async def _notify_active_sessions_of_shutdown(self) -> None:
                     platform_str, chat_id, e,
                 )
 
-        for platform, adapter in self.adapters.items():
+        # Snapshot adapters up front: adapter.send() can hit a fatal error
+        # path that pops the adapter from self.adapters (see _handle_fatal
+        # elsewhere), which would otherwise trigger
+        # ``RuntimeError: dictionary changed size during iteration`` —
+        # observed in a user report during gateway shutdown.
+        for platform, adapter in list(self.adapters.items()):
             home = self.config.get_home_channel(platform)
             if not home or not home.chat_id:
                 continue

hermes_cli/kanban_db.py

@@ -2618,6 +2618,77 @@ class DispatchResult:
     """Task ids whose workers exceeded ``max_runtime_seconds``."""
 
 
+# Bounded registry of recently-reaped worker child exits, populated by the
+# reap loop at the top of ``dispatch_once`` and consulted by
+# ``detect_crashed_workers`` to classify a dead-pid task.
+#
+# Entry: ``pid -> (raw_wait_status, reaped_at_epoch)``. We keep raw status
+# so both ``os.WIFEXITED`` / ``os.WEXITSTATUS`` and ``os.WIFSIGNALED`` can
+# be consulted. Entries are trimmed by age (and total size cap as a
+# belt-and-braces against unbounded growth on exotic platforms).
+_RECENT_WORKER_EXIT_TTL_SECONDS = 600
+_RECENT_WORKER_EXITS_MAX = 4096
+_recent_worker_exits: "dict[int, tuple[int, float]]" = {}
+
+
+def _record_worker_exit(pid: int, raw_status: int) -> None:
+    """Record a reaped child's exit status for later classification.
+
+    Called from the reap loop in ``dispatch_once``. Safe to call many
+    times; duplicate pids overwrite (pids can cycle, latest wins).
+    """
+    if not pid or pid <= 0:
+        return
+    now = time.time()
+    _recent_worker_exits[int(pid)] = (int(raw_status), now)
+    # Age-based trim: drop entries older than the TTL.
+    if len(_recent_worker_exits) > _RECENT_WORKER_EXITS_MAX // 2:
+        cutoff = now - _RECENT_WORKER_EXIT_TTL_SECONDS
+        for _pid in [p for p, (_s, t) in _recent_worker_exits.items() if t < cutoff]:
+            _recent_worker_exits.pop(_pid, None)
+    # Size cap as a final guard.
+    if len(_recent_worker_exits) > _RECENT_WORKER_EXITS_MAX:
+        # Drop oldest half.
+        ordered = sorted(_recent_worker_exits.items(), key=lambda kv: kv[1][1])
+        for _pid, _ in ordered[: len(ordered) // 2]:
+            _recent_worker_exits.pop(_pid, None)
+
+
+def _classify_worker_exit(pid: int) -> "tuple[str, Optional[int]]":
+    """Classify a recently-reaped worker by pid.
+
+    Returns ``(kind, code)`` where ``kind`` is one of:
+
+    * ``"clean_exit"`` — ``WIFEXITED`` with ``WEXITSTATUS == 0``. When the
+      task is still ``running`` in the DB, this is a protocol violation
+      (worker exited without calling ``kanban_complete`` / ``kanban_block``)
+      and should be auto-blocked immediately — retrying will just loop.
+    * ``"nonzero_exit"`` — ``WIFEXITED`` with non-zero status. Real error.
+    * ``"signaled"`` — ``WIFSIGNALED`` (OOM killer, SIGKILL, etc). Real crash.
+    * ``"unknown"`` — pid was not in the reap registry (either reaped by
+      something else, or died between reap tick and liveness check). Fall
+      back to existing crashed-counter behavior.
+
+    ``code`` is the exit status (for ``clean_exit`` / ``nonzero_exit``) or
+    the signal number (for ``signaled``), or ``None`` for ``unknown``.
+    """
+    entry = _recent_worker_exits.get(int(pid))
+    if entry is None:
+        return ("unknown", None)
+    raw, _ = entry
+    try:
+        if os.WIFEXITED(raw):
+            code = os.WEXITSTATUS(raw)
+            if code == 0:
+                return ("clean_exit", 0)
+            return ("nonzero_exit", code)
+        if os.WIFSIGNALED(raw):
+            return ("signaled", os.WTERMSIG(raw))
+    except Exception:
+        pass
+    return ("unknown", None)
+
+
 def _pid_alive(pid: Optional[int]) -> bool:
     """Return True if ``pid`` is still running on this host.
 
@@ -2924,12 +2995,22 @@ def detect_crashed_workers(conn: sqlite3.Connection) -> list[str]:
     are meaningless here. The host-local check is enough because
     ``_default_spawn`` always runs the worker on the same host as the
     dispatcher (the whole design is single-host).
+
+    When the reap registry shows the worker exited cleanly (rc=0) but
+    the task was still ``running`` in the DB, treat it as a protocol
+    violation (worker answered conversationally without calling
+    ``kanban_complete`` / ``kanban_block``) and trip the circuit breaker
+    on the first occurrence — retrying a worker whose CLI keeps
+    returning 0 without a terminal transition just loops forever.
     """
     crashed: list[str] = []
     # Per-crash details collected inside the main txn, used after it
     # closes to run ``_record_task_failure`` (which needs its own
-    # write_txn so can't nest).
-    crash_details: list[tuple[str, int, str]] = []  # (task_id, pid, claimer)
+    # write_txn so can't nest). ``protocol_violation`` flags the
+    # clean-exit-but-still-running case so we can trip the breaker
+    # immediately instead of incrementing by 1.
+    crash_details: list[tuple[str, int, str, bool, str]] = []
+    # (task_id, pid, claimer, protocol_violation, error_text)
     with write_txn(conn):
         rows = conn.execute(
             "SELECT id, worker_pid, claim_lock FROM tasks "
@@ -2943,6 +3024,39 @@ def detect_crashed_workers(conn: sqlite3.Connection) -> list[str]:
                 continue
             if _pid_alive(row["worker_pid"]):
                 continue
+
+            pid = int(row["worker_pid"])
+            kind, code = _classify_worker_exit(pid)
+            if kind == "clean_exit":
+                # Worker subprocess returned 0 but its task is still
+                # ``running`` in the DB — it exited without calling
+                # ``kanban_complete`` / ``kanban_block``. Retrying won't
+                # help.
+                protocol_violation = True
+                error_text = (
+                    "worker exited cleanly (rc=0) without calling "
+                    "kanban_complete or kanban_block — protocol violation"
+                )
+                event_kind = "protocol_violation"
+                event_payload = {
+                    "pid": pid,
+                    "claimer": row["claim_lock"],
+                    "exit_code": code,
+                }
+            else:
+                protocol_violation = False
+                if kind == "nonzero_exit":
+                    error_text = f"pid {pid} exited with code {code}"
+                elif kind == "signaled":
+                    error_text = f"pid {pid} killed by signal {code}"
+                else:
+                    error_text = f"pid {pid} not alive"
+                event_kind = "crashed"
+                event_payload = {"pid": pid, "claimer": row["claim_lock"]}
+                if code is not None and kind != "unknown":
+                    event_payload["exit_kind"] = kind
+                    event_payload["exit_code"] = code
+
             cur = conn.execute(
                 "UPDATE tasks SET status = 'ready', claim_lock = NULL, "
                 "claim_expires = NULL, worker_pid = NULL "
@@ -2953,34 +3067,47 @@ def detect_crashed_workers(conn: sqlite3.Connection) -> list[str]:
                 run_id = _end_run(
                     conn, row["id"],
                     outcome="crashed", status="crashed",
-                    error=f"pid {int(row['worker_pid'])} not alive",
-                    metadata={
-                        "pid": int(row["worker_pid"]),
-                        "claimer": row["claim_lock"],
-                    },
+                    error=error_text,
+                    metadata=dict(event_payload),
                 )
                 _append_event(
-                    conn, row["id"], "crashed",
-                    {"pid": int(row["worker_pid"]), "claimer": row["claim_lock"]},
+                    conn, row["id"], event_kind,
+                    event_payload,
                     run_id=run_id,
                 )
                 crashed.append(row["id"])
                 crash_details.append(
-                    (row["id"], int(row["worker_pid"]), row["claim_lock"])
+                    (row["id"], pid, row["claim_lock"],
+                     protocol_violation, error_text)
                 )
     # Outside the main txn: increment the unified failure counter for
     # each crashed task. If the breaker trips, the task transitions
     # ready → blocked with a ``gave_up`` event on top of the ``crashed``
     # event we already emitted.
-    for tid, pid, claimer in crash_details:
-        _record_task_failure(
+    #
+    # Protocol-violation crashes force an immediate trip (failure_limit=1)
+    # because clean-exit-without-transition is deterministic: the next
+    # respawn will do exactly the same thing. Better to surface to a
+    # human with a clear reason than to loop ``DEFAULT_FAILURE_LIMIT``
+    # times first.
+    auto_blocked: list[str] = []
+    for tid, pid, claimer, protocol_violation, error_text in crash_details:
+        tripped = _record_task_failure(
             conn, tid,
-            error=f"pid {pid} not alive",
+            error=error_text,
             outcome="crashed",
+            failure_limit=(1 if protocol_violation else None),
             release_claim=False,
             end_run=False,
             event_payload_extra={"pid": pid, "claimer": claimer},
         )
+        if tripped:
+            auto_blocked.append(tid)
+    # Stash auto-blocked ids on the function for the dispatch loop to pick up.
+    # Keeps the public return type (``list[str]``) stable for direct callers
+    # and tests that destructure the result; ``dispatch_once`` reads this
+    # side-channel attribute to populate ``DispatchResult.auto_blocked``.
+    detect_crashed_workers._last_auto_blocked = auto_blocked  # type: ignore[attr-defined]
     return crashed
 
 
@@ -3242,6 +3369,12 @@ def dispatch_once(
     # exit. WNOHANG keeps this non-blocking; ChildProcessError means no
     # children to reap. Bounded: at most one tick's worth of completions
     # can be in <defunct> at once.
+    #
+    # We also record the exit status keyed by pid, so
+    # ``detect_crashed_workers`` can distinguish a worker that exited
+    # cleanly without calling ``kanban_complete`` / ``kanban_block``
+    # (protocol violation — auto-block) from a real crash (OOM killer,
+    # SIGKILL, non-zero exit — existing counter behavior).
     try:
         while True:
             try:
@@ -3250,12 +3383,21 @@ def dispatch_once(
                 break
             if _pid == 0:
                 break
+            _record_worker_exit(_pid, _status)
     except Exception:
         pass
 
     result = DispatchResult()
     result.reclaimed = release_stale_claims(conn)
     result.crashed = detect_crashed_workers(conn)
+    # detect_crashed_workers stashes protocol-violation auto-blocks on
+    # itself so the public list-return stays stable. Pull them into the
+    # DispatchResult here so telemetry / tests see the trip.
+    _crash_auto_blocked = getattr(
+        detect_crashed_workers, "_last_auto_blocked", []
+    )
+    if _crash_auto_blocked:
+        result.auto_blocked.extend(_crash_auto_blocked)
     result.timed_out = enforce_max_runtime(conn)
     result.promoted = recompute_ready(conn)
 

tests/hermes_cli/test_kanban_core_functionality.py

@@ -3636,6 +3636,100 @@ def test_detect_crashed_workers_increments_counter(kanban_home):
         conn.close()
 
 
+def test_detect_crashed_workers_protocol_violation_auto_blocks(kanban_home):
+    """A worker that exited rc=0 while its task was still ``running``
+    is a protocol violation (agent answered conversationally without
+    calling kanban_complete / kanban_block). Retrying will just loop,
+    so auto-block immediately instead of waiting for the breaker to
+    trip at ``DEFAULT_FAILURE_LIMIT``.
+
+    Regression test for the respawn-loop-after-completion bug reported
+    against small local models (gemma4-e2b q4) where the model writes
+    the answer as plain text and the CLI exits rc=0 cleanly.
+    """
+    import hermes_cli.kanban_db as _kb
+    conn = kb.connect()
+    try:
+        tid = kb.create_task(conn, title="quiet", assignee="worker")
+        host_prefix = _kb._claimer_id().split(":", 1)[0]
+        lock = f"{host_prefix}:mock"
+        kb.claim_task(conn, tid, claimer=lock)
+        fake_pid = 999998
+        kb._set_worker_pid(conn, tid, fake_pid)
+
+        # Simulate the reap loop having recorded a clean exit for this pid.
+        # os.W_EXITCODE(status=0, signal=0) == 0 on POSIX.
+        _kb._record_worker_exit(fake_pid, 0)
+        # Force liveness check to say "dead" for the fake pid.
+        original_alive = _kb._pid_alive
+        _kb._pid_alive = lambda p: False
+        try:
+            result_crashed = kb.detect_crashed_workers(conn)
+        finally:
+            _kb._pid_alive = original_alive
+
+        assert tid in result_crashed, "should be detected as crashed"
+        task = kb.get_task(conn, tid)
+        assert task.status == "blocked", (
+            f"protocol violation should auto-block on first occurrence, "
+            f"got status={task.status}"
+        )
+        assert "kanban_complete" in (task.last_failure_error or ""), (
+            f"expected protocol-violation message, got {task.last_failure_error!r}"
+        )
+
+        events = kb.list_events(conn, tid)
+        kinds = [e.kind for e in events]
+        assert "protocol_violation" in kinds, (
+            f"expected 'protocol_violation' event, got {kinds}"
+        )
+        # The ``crashed`` event would be misleading here — the worker
+        # didn't crash, it returned 0.
+        assert "crashed" not in kinds, (
+            f"should NOT emit 'crashed' event on clean exit, got {kinds}"
+        )
+        assert "gave_up" in kinds, (
+            f"breaker should trip, expected 'gave_up' event, got {kinds}"
+        )
+    finally:
+        conn.close()
+
+
+def test_detect_crashed_workers_nonzero_exit_uses_default_limit(kanban_home):
+    """A worker that exited non-zero (real error / crash) uses the
+    normal counter path — one failure doesn't trip the breaker.
+    """
+    import hermes_cli.kanban_db as _kb
+    conn = kb.connect()
+    try:
+        tid = kb.create_task(conn, title="crashy", assignee="worker")
+        host_prefix = _kb._claimer_id().split(":", 1)[0]
+        kb.claim_task(conn, tid, claimer=f"{host_prefix}:mock")
+        fake_pid = 999997
+        kb._set_worker_pid(conn, tid, fake_pid)
+
+        # W_EXITCODE(1, 0) == 256 — WIFEXITED True, WEXITSTATUS == 1.
+        _kb._record_worker_exit(fake_pid, 256)
+        original_alive = _kb._pid_alive
+        _kb._pid_alive = lambda p: False
+        try:
+            kb.detect_crashed_workers(conn)
+        finally:
+            _kb._pid_alive = original_alive
+
+        task = kb.get_task(conn, tid)
+        assert task.status == "ready", (
+            f"single non-zero crash shouldn't auto-block, got {task.status}"
+        )
+        assert task.consecutive_failures == 1
+        events = kb.list_events(conn, tid)
+        kinds = [e.kind for e in events]
+        assert "crashed" in kinds
+        assert "protocol_violation" not in kinds
+    finally:
+        conn.close()
+
+
 def test_reclaim_task_clears_failure_counter(kanban_home):
     """Operator reclaim wipes the counter so the next retry gets a fresh
     budget."""