ling/hermes-agent
1
0
Fork 0
mirror of https://github.com/NousResearch/hermes-agent.git synced 2026-04-28 06:51:16 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix(terminal): three-layer defense against watch_patterns notification spam (#15642)

Browse Source 
* fix(terminal): three-layer defense against watch_patterns notification spam

Background processes that stack notify_on_complete=True with watch_patterns
can flood the user with duplicate, delayed notifications — matches deliver
asynchronously via the completion queue and continue arriving minutes after
the process has exited. The docstring warning against this (PR #12113) has
proven insufficient; agents still misuse the combination.

Three layered defenses, each sufficient on its own:

1. Mutual exclusion (terminal_tool.py): When both flags are set on a
   background process, drop watch_patterns with a warning. notify_on_complete
   wins because 'let me know when it's done' is the more useful signal and
   fires exactly once. Extracted as _resolve_notification_flag_conflict() so
   the rule is testable in isolation.

2. Suppress-after-exit (process_registry.py): _check_watch_patterns() now
   bails the moment session.exited is True. Post-exit chunks (buffered reads
   draining after the process is gone) no longer produce notifications. This
   is the fix flagged as future work in session 20260418_020302_79881c.

3. Global circuit breaker (process_registry.py): Per-session rate limits don't
   catch the sibling-flood case — N concurrent processes can each stay under
   8/10s and still collectively spam. New WATCH_GLOBAL_MAX_PER_WINDOW=15 cap
   trips a 30-second cooldown across ALL sessions, emits a single
   watch_overflow_tripped event, silently counts dropped events, and emits a
   watch_overflow_released summary when the cooldown ends.

Also updates the tool schema + docstring to document the new behavior.

Tests: 8 new tests covering all three fixes (suppress-after-exit x2,
mutual-exclusion resolver x4, global breaker trip/cooldown/release x2).
All 60 tests across test_watch_patterns.py, test_notify_on_complete.py,
test_terminal_tool.py pass.

Real-world trigger: self-inflicted in session 20260425_051924 — three
concurrent hermes-sweeper review subprocesses each set watch_patterns=
['failed validation', 'errored'] AND notify_on_complete=True, then iterated
over multiple items, producing enough matches per process to defeat the
per-session cap while staying under the global cap that didn't yet exist.

* fix(terminal): aggressive 1-per-15s watch_patterns rate limit + strike-3 promotion

Per Teknium's direction, the watch_patterns rate limit is now much more
aggressive and self-healing.

## New rule — per session

- HARD cap: 1 watch-match notification per 15 seconds per process.
- Any match arriving inside the cooldown window is dropped and counts as
  ONE strike for that window (many drops in the same window still = 1 strike).
- After 3 consecutive strike windows, watch_patterns is permanently disabled
  for the session and the session is auto-promoted to notify_on_complete
  semantics — exactly one notification when the process actually exits.
- A cooldown window that expires with zero drops resets the consecutive
  strike counter — healthy cadence is forgiven.

## Schema + docstring rewritten

The tool schema description now gives the model explicit guidance:
- notify_on_complete is 'the right choice for almost every long-running task'
- watch_patterns is for RARE one-shot signals on LONG-LIVED processes
- Do NOT use watch_patterns with loops/batch jobs — error patterns fire every
  iteration and will hit the strike limit fast
- Mutual exclusion is stated on both parameter descriptions
- 1/15s cooldown and 3-strike promotion are stated in the watch_patterns
  description so the model sees the contract every turn

## Removed

- WATCH_MAX_PER_WINDOW (8/10s) and WATCH_OVERLOAD_KILL_SECONDS (45) — the
  new 1/15s limit subsumes both; keeping them would double-count.
- _watch_window_hits / _watch_window_start / _watch_overload_since fields
  on ProcessSession. Replaced by _watch_last_emit_at / _watch_cooldown_until
  / _watch_strike_candidate / _watch_consecutive_strikes.

## Kept

- Global circuit breaker across all sessions (15/10s → 30s cooldown) as a
  secondary safety net for concurrent siblings. Still valuable when 20
  short-lived processes each fire once — none individually violates the
  per-session limit.
- Suppress-after-exit guard.
- Mutual exclusion resolver at the tool entry point.

## Tests

- 6 new tests in TestPerSessionRateLimit covering: first match delivers,
  second in cooldown suppressed, multi-drop = single strike, 3 strikes
  disables + promotes, clean window resets counter, suppressed count
  carried to next emit.
- Global circuit breaker tests rewritten to use fresh sessions instead of
  hacking removed per-window fields.
- 50/50 watch_patterns + notify_on_complete tests pass.
- 60/60 including test_terminal_tool.py pass.
This commit is contained in:
Teknium
2026-04-25 06:41:58 -07:00
committed by GitHub
parent 6e561ffa6d
commit 97d54f0e4d
3 changed files with 492 additions and 131 deletions
Download Patch File Download Diff File Expand all files Collapse all files

250
tools/process_registry.py
View File

@@ -58,10 +58,20 @@ MAX_OUTPUT_CHARS = 200_000 # 200KB rolling output buffer
FINISHED_TTL_SECONDS = 1800 # Keep finished processes for 30 minutes
MAX_PROCESSES = 64 # Max concurrent tracked processes (LRU pruning)
# Watch pattern rate limiting
WATCH_MAX_PER_WINDOW = 8 # Max notifications delivered per window
WATCH_WINDOW_SECONDS = 10 # Rolling window length
WATCH_OVERLOAD_KILL_SECONDS = 45 # Sustained overload duration before disabling watch
# Watch pattern rate limiting — PER SESSION.
# Hard rule: at most ONE watch-match notification every WATCH_MIN_INTERVAL_SECONDS.
# Any match arriving inside that cooldown window is dropped and counted as a strike.
# After WATCH_STRIKE_LIMIT consecutive strike windows, watch_patterns for that
# session is permanently disabled and the session falls back to notify_on_complete
# semantics (one notification when the process actually exits).
WATCH_MIN_INTERVAL_SECONDS = 15 # Minimum spacing between consecutive watch matches
WATCH_STRIKE_LIMIT = 3 # Strikes in a row → disable watch + promote to notify_on_complete
# Global circuit breaker — across all sessions. Secondary safety net so concurrent
# siblings can't collectively flood the user even when each is under its own cap.
WATCH_GLOBAL_MAX_PER_WINDOW = 15
WATCH_GLOBAL_WINDOW_SECONDS = 10
WATCH_GLOBAL_COOLDOWN_SECONDS = 30
def format_uptime_short(seconds: int) -> str:
@@ -105,10 +115,18 @@ class ProcessSession:
watch_patterns: List[str] = field(default_factory=list)
_watch_hits: int = field(default=0, repr=False) # total matches delivered
_watch_suppressed: int = field(default=0, repr=False) # matches dropped by rate limit
_watch_overload_since: float = field(default=0.0, repr=False) # when sustained overload began
_watch_disabled: bool = field(default=False, repr=False) # permanently killed by overload
_watch_window_hits: int = field(default=0, repr=False) # hits in current rate window
_watch_window_start: float = field(default=0.0, repr=False)
_watch_disabled: bool = field(default=False, repr=False) # permanently killed after strike limit
# Per-session rate limit state: at most one match every WATCH_MIN_INTERVAL_SECONDS.
# When an emission happens, _watch_cooldown_until is set to now + interval and
# _watch_strike_candidate becomes True. The next match to arrive before that
# deadline counts as one strike (regardless of how many matches were dropped in
# between — a strike is a window, not a match). After WATCH_STRIKE_LIMIT strikes
# in a row, watch_patterns is disabled and the session promotes to
# notify_on_complete.
_watch_last_emit_at: float = field(default=0.0, repr=False)
_watch_cooldown_until: float = field(default=0.0, repr=False)
_watch_strike_candidate: bool = field(default=False, repr=False)
_watch_consecutive_strikes: int = field(default=0, repr=False)
_lock: threading.Lock = field(default_factory=threading.Lock)
_reader_thread: Optional[threading.Thread] = field(default=None, repr=False)
_pty: Any = field(default=None, repr=False) # ptyprocess handle (when use_pty=True)
@@ -151,6 +169,15 @@ class ProcessRegistry:
# via wait/poll/log. Drain loops skip notifications for these.
self._completion_consumed: set = set()
# Global watch-match circuit breaker — across all sessions.
# Prevents sibling processes from collectively flooding the user even
# when each stays under its own per-session cap.
self._global_watch_lock = threading.Lock()
self._global_watch_window_start: float = 0.0
self._global_watch_window_hits: int = 0
self._global_watch_tripped_until: float = 0.0
self._global_watch_suppressed_during_trip: int = 0
@staticmethod
def _clean_shell_noise(text: str) -> str:
"""Strip shell startup warnings from the beginning of output."""
@@ -163,12 +190,23 @@ class ProcessRegistry:
"""Scan new output for watch patterns and queue notifications.
Called from reader threads with new_text being the freshly-read chunk.
Rate-limited: max WATCH_MAX_PER_WINDOW notifications per WATCH_WINDOW_SECONDS.
If sustained overload exceeds WATCH_OVERLOAD_KILL_SECONDS, watching is
disabled permanently for this process.
Per-session rate limit: at most ONE watch-match notification per
WATCH_MIN_INTERVAL_SECONDS. Any match arriving inside the cooldown
window is dropped and counts as ONE strike for that window. After
WATCH_STRIKE_LIMIT consecutive strike windows, watch_patterns is
disabled for this session and the session is promoted to
notify_on_complete semantics — one notification when the process
actually exits, no more mid-process spam.
"""
if not session.watch_patterns or session._watch_disabled:
return
# Suppress-after-exit: once the reader loop has declared the process
# exited, any late chunk we still see is post-exit noise. Dropping these
# prevents the "stale notifications delivered minutes after the process
# ended" spam when completion_queue consumers run async.
if session.exited:
return
# Scan new text line-by-line for pattern matches
matched_lines = []
@@ -185,55 +223,80 @@ class ProcessRegistry:
return
now = time.time()
should_disable = False
with session._lock:
# Reset window if it's expired
if now - session._watch_window_start >= WATCH_WINDOW_SECONDS:
session._watch_window_hits = 0
session._watch_window_start = now
# Check rate limit
if session._watch_window_hits >= WATCH_MAX_PER_WINDOW:
# Case 1: still inside the cooldown from the last emission.
# Count this as a strike for the current window (only once per window)
# and drop the event. If we've hit the strike limit, disable watch
# and promote to notify_on_complete.
if session._watch_cooldown_until and now < session._watch_cooldown_until:
session._watch_suppressed += len(matched_lines)
if not session._watch_strike_candidate:
# First drop in this window — count one strike.
session._watch_strike_candidate = True
session._watch_consecutive_strikes += 1
if session._watch_consecutive_strikes >= WATCH_STRIKE_LIMIT:
session._watch_disabled = True
# Promote to notify_on_complete so the agent still gets
# exactly one notification when the process actually ends.
session.notify_on_complete = True
should_disable = True
return_early = True
else:
# Case 2: cooldown has expired.
# Decide whether this window was a "clean" one (no drops) or a
# strike window. If no strike candidate was set during the prior
# cooldown, reset the consecutive-strike counter — we're back to
# healthy emission cadence.
if (
session._watch_cooldown_until
and not session._watch_strike_candidate
):
session._watch_consecutive_strikes = 0
session._watch_strike_candidate = False
# Track sustained overload for kill switch
if session._watch_overload_since == 0.0:
session._watch_overload_since = now
elif now - session._watch_overload_since > WATCH_OVERLOAD_KILL_SECONDS:
session._watch_disabled = True
self.completion_queue.put({
"session_id": session.id,
"session_key": session.session_key,
"command": session.command,
"type": "watch_disabled",
"suppressed": session._watch_suppressed,
"platform": session.watcher_platform,
"chat_id": session.watcher_chat_id,
"user_id": session.watcher_user_id,
"user_name": session.watcher_user_name,
"thread_id": session.watcher_thread_id,
"message": (
f"Watch patterns disabled for process {session.id} — "
f"too many matches ({session._watch_suppressed} suppressed). "
f"Use process(action='poll') to check output manually."
),
})
return
# Emit the notification and start a new cooldown window.
session._watch_last_emit_at = now
session._watch_cooldown_until = now + WATCH_MIN_INTERVAL_SECONDS
session._watch_hits += 1
suppressed = session._watch_suppressed
session._watch_suppressed = 0
return_early = False
# Under the rate limit — deliver notification
session._watch_window_hits += 1
session._watch_hits += 1
# Clear overload tracker since we got a delivery through
session._watch_overload_since = 0.0
# Include suppressed count if any events were dropped
suppressed = session._watch_suppressed
session._watch_suppressed = 0
if return_early:
if should_disable:
# Emit exactly one "watch disabled, falling back to notify_on_complete"
# summary event so the agent/user sees why things went quiet.
self.completion_queue.put({
"session_id": session.id,
"session_key": session.session_key,
"command": session.command,
"type": "watch_disabled",
"suppressed": session._watch_suppressed,
"platform": session.watcher_platform,
"chat_id": session.watcher_chat_id,
"user_id": session.watcher_user_id,
"user_name": session.watcher_user_name,
"thread_id": session.watcher_thread_id,
"message": (
f"Watch patterns disabled for process {session.id} — "
f"{WATCH_STRIKE_LIMIT} consecutive rate-limit windows triggered "
f"(min spacing {WATCH_MIN_INTERVAL_SECONDS}s). "
f"Falling back to notify_on_complete semantics; you'll get "
f"exactly one notification when the process exits."
),
})
return
# Trim matched output to a reasonable size
output = "\n".join(matched_lines[:20])
if len(output) > 2000:
output = output[:2000] + "\n...(truncated)"
# Global circuit breaker — across all sessions (secondary safety net).
if not self._global_watch_admit(now):
return
self.completion_queue.put({
"session_id": session.id,
"session_key": session.session_key,
@@ -249,6 +312,93 @@ class ProcessRegistry:
"thread_id": session.watcher_thread_id,
})
def _global_watch_admit(self, now: float) -> bool:
"""Return True if this watch_match event is allowed through the global breaker.
Semantics:
- If we're currently in a cooldown period, drop the event and count it.
- Otherwise, slide the rolling window and check the global cap.
- If the cap is exceeded, trip the breaker for WATCH_GLOBAL_COOLDOWN_SECONDS
and emit ONE summary event so the agent/user sees "N notifications were
suppressed" instead of getting them individually.
- When the cooldown ends, emit a release summary and reset counters.
"""
with self._global_watch_lock:
# Handle cooldown expiry first so we can emit the release summary.
if self._global_watch_tripped_until and now >= self._global_watch_tripped_until:
suppressed = self._global_watch_suppressed_during_trip
self._global_watch_tripped_until = 0.0
self._global_watch_suppressed_during_trip = 0
self._global_watch_window_start = now
self._global_watch_window_hits = 0
if suppressed > 0:
# Queue a summary event outside the lock (below).
release_msg = {
"session_id": "",
"session_key": "",
"command": "",
"type": "watch_overflow_released",
"suppressed": suppressed,
"message": (
f"Watch-pattern notifications resumed. "
f"{suppressed} match event(s) were suppressed during the flood."
),
"platform": "",
"chat_id": "",
"user_id": "",
"user_name": "",
"thread_id": "",
}
else:
release_msg = None
else:
release_msg = None
# Still in cooldown — drop and count.
if self._global_watch_tripped_until and now < self._global_watch_tripped_until:
self._global_watch_suppressed_during_trip += 1
admit = False
trip_now = None
else:
# Slide the window.
if now - self._global_watch_window_start >= WATCH_GLOBAL_WINDOW_SECONDS:
self._global_watch_window_start = now
self._global_watch_window_hits = 0
if self._global_watch_window_hits >= WATCH_GLOBAL_MAX_PER_WINDOW:
# Trip the breaker.
self._global_watch_tripped_until = now + WATCH_GLOBAL_COOLDOWN_SECONDS
self._global_watch_suppressed_during_trip += 1
trip_now = now
admit = False
else:
self._global_watch_window_hits += 1
trip_now = None
admit = True
# Queue summary events outside the lock.
if release_msg is not None:
self.completion_queue.put(release_msg)
if trip_now is not None:
self.completion_queue.put({
"session_id": "",
"session_key": "",
"command": "",
"type": "watch_overflow_tripped",
"message": (
f"Watch-pattern overflow: >{WATCH_GLOBAL_MAX_PER_WINDOW} "
f"notifications in {WATCH_GLOBAL_WINDOW_SECONDS}s across all processes. "
f"Suppressing further watch_match events for "
f"{WATCH_GLOBAL_COOLDOWN_SECONDS}s."
),
"platform": "",
"chat_id": "",
"user_id": "",
"user_name": "",
"thread_id": "",
})
return admit
@staticmethod
def _is_host_pid_alive(pid: Optional[int]) -> bool:
"""Best-effort liveness check for host-visible PIDs."""