Files
hermes-agent/gateway/platforms/signal_rate_limit.py

196 lines
10 KiB
Python

"""Signal attachment rate-limit scheduler: process-wide token-bucket simulator mirroring the per-account
attachment rate limit signal-cli/Signal-Server enforce. Producers (``SignalAdapter.send_multiple_images``
and the ``send_message`` tool's Signal path) call ``acquire(n)`` before an attachment send; on a 429 they
call ``feedback(retry_after, n)`` so the model recalibrates from the server's authoritative hint.
Concurrent calls serialize through an ``asyncio.Lock`` — FIFO fairness across sessions sharing one daemon."""
from __future__ import annotations
import asyncio
import logging
import re
import time
from typing import Any, Optional
from agent.retry_utils import parse_retry_after_seconds
logger = logging.getLogger(__name__)
SIGNAL_MAX_ATTACHMENTS_PER_MSG = 32 # per-message attachment cap (Signal-{Android,Desktop} source)
SIGNAL_RATE_LIMIT_BUCKET_CAPACITY = 50 # server-side token-bucket capacity for attachments
SIGNAL_RATE_LIMIT_DEFAULT_RETRY_AFTER = 4 # fallback token refill interval for signal-cli < v0.14.3
SIGNAL_RATE_LIMIT_MAX_ATTEMPTS = 2 # initial attempt + 1 retry
SIGNAL_BATCH_PACING_NOTICE_THRESHOLD = 10.0 # estimated wait above this → notify the user
SIGNAL_RPC_ERROR_RATELIMIT = -5 # signal-cli (v0.14.3+) JSON-RPC error code for RateLimitException
class SignalRateLimitError(Exception):
"""Raised by ``SignalAdapter._rpc`` for rate-limit responses when ``raise_on_rate_limit=True``.
``retry_after`` is the server's per-token Retry-After in seconds (signal-cli ≥ v0.14.3) or None."""
def __init__(self, message: str, retry_after: Optional[float] = None) -> None:
super().__init__(message)
self.retry_after = retry_after
class SignalSchedulerError(Exception):
pass
# "Retry after 4 seconds" — libsignal-net's RetryLaterException string form, surfaced when 429s hit during
# attachment upload (signal-cli wraps these as AttachmentInvalidException, so the typed path doesn't fire).
_RETRY_AFTER_RE = re.compile(r"Retry after (\d+(?:\.\d+)?)\s*second", re.IGNORECASE)
def _error_message(err: Any) -> str:
return str(err.get("message", "")) if isinstance(err, dict) else str(err)
def _extract_retry_after_seconds(err: Any) -> Optional[float]:
"""Per-token Retry-After from a signal-cli rate-limit error, or None. Sources, in order:
``error.data.response.results[*].retryAfterSeconds`` (signal-cli ≥ v0.14.3), then "Retry after N
seconds" parsed from the message (RetryLaterException wrapped as AttachmentInvalidException)."""
if isinstance(err, dict):
results = ((err.get("data") or {}).get("response") or {}).get("results") or []
candidates = [parse_retry_after_seconds(r.get("retryAfterSeconds")) for r in results
if isinstance(r, dict) and r.get("retryAfterSeconds")]
if candidates := [c for c in candidates if c is not None]:
return max(candidates)
match = _RETRY_AFTER_RE.search(_error_message(err))
return parse_retry_after_seconds(match.group(1)) if match else None
def _is_signal_rate_limit_error(err: Any) -> bool:
"""True if a signal-cli RPC error reflects a rate-limit failure: the typed ``RATELIMIT_ERROR`` code
(≥ v0.14.3), legacy ``[429]`` / ``RateLimitException`` substrings, or libsignal-net's
``RetryLaterException`` / "Retry after N seconds" leaked through AttachmentInvalidException."""
if isinstance(err, dict) and err.get("code") == SIGNAL_RPC_ERROR_RATELIMIT:
return True
message = _error_message(err)
msg_lower = message.lower()
return "[429]" in message or any(s in msg_lower for s in ("ratelimit", "retrylaterexception", "retry after"))
def _format_wait(seconds: float) -> str:
"""Human-friendly wait label for user-facing pacing notices."""
s = max(0.0, seconds)
return f"{int(round(s))}s" if s < 90 else f"{max(1, int(round(s / 60)))} min"
def _signal_send_timeout(num_attachments: int) -> float:
"""HTTP timeout for a Signal ``send`` RPC: signal-cli uploads attachments serially inside the call, so
the default 30s truncates large batches mid-upload (phantom failure). 5s/attachment, 60s floor."""
return 30.0 if num_attachments <= 0 else max(60.0, 5.0 * num_attachments)
class SignalAttachmentScheduler:
"""Process-wide token-bucket simulator for Signal attachment sends: up to ``capacity`` tokens (default
50 = Signal's server bucket), one per attachment, refilling at ``refill_rate``/s (calibrated from the
server's per-token Retry-After once a 429 has been observed; default 1 token / 4s). ``acquire(n)``
calls serialize through an ``asyncio.Lock`` — FIFO across sessions."""
def __init__(self, capacity: float = float(SIGNAL_RATE_LIMIT_BUCKET_CAPACITY),
default_retry_after: float = float(SIGNAL_RATE_LIMIT_DEFAULT_RETRY_AFTER)) -> None:
self.capacity = float(capacity)
self.tokens = float(capacity)
self.refill_rate = 1.0 / float(default_retry_after)
self.last_refill = time.monotonic()
self._lock = asyncio.Lock()
def _projected_tokens(self, now: Optional[float] = None) -> float:
"""Tokens the bucket would hold at ``now``, without mutating state."""
elapsed = (time.monotonic() if now is None else now) - self.last_refill
if elapsed > 0 and self.tokens < self.capacity:
return min(self.capacity, self.tokens + elapsed * self.refill_rate)
return self.tokens
def _refill(self) -> None:
now = time.monotonic()
self.tokens = self._projected_tokens(now)
self.last_refill = now
def estimate_wait(self, n: int) -> float:
"""Seconds until ``n`` tokens would be available (lock-free, informational — decides whether to emit
a pacing notice *before* a possibly-blocking ``acquire``; races vs. concurrent acquires are benign)."""
deficit = n - self._projected_tokens()
return 0.0 if deficit <= 0 else deficit / self.refill_rate
async def acquire(self, n: int) -> float:
"""Block until at least ``n`` tokens are available; return the seconds slept. Does **not** deduct
tokens — the bucket is a read-only model of server-side capacity; call ``report_rpc_duration()``
after the RPC to sync. The lock is released during ``asyncio.sleep`` so other callers interleave,
and the loop re-checks after each sleep in case the deadline was pessimistic. Signal's server is
ground truth and will 429 (→ requeue) if the model drifts."""
if n <= 0:
return 0.0
if n > self.capacity:
raise SignalSchedulerError(f"Signal scheduler was called requesting {n} tokens (max is {self.capacity})")
total_slept, first_pass = 0.0, True
while True:
async with self._lock:
self._refill()
if self.tokens >= n:
if not first_pass or total_slept > 0:
logger.debug("Signal scheduler: tokens sufficient for %d (remaining=%.1f, total_slept=%.1fs)",
n, self.tokens, total_slept)
return total_slept
deficit = n - self.tokens
wait = deficit / self.refill_rate
if first_pass:
logger.info("Signal scheduler: pausing %.1fs for %d tokens (available=%.1f, deficit=%.1f, "
"refill=%.4f/s ≈ %.1fs/token)", wait, n, self.tokens, deficit, self.refill_rate,
1.0 / self.refill_rate)
first_pass = False
await asyncio.sleep(wait)
total_slept += wait
async def report_rpc_duration(self, rpc_duration: float, n_attachments: int) -> None:
"""Deduct ``n_attachments`` tokens for a completed send RPC. No refill is credited for the upload
window: Signal's server checks the bucket at RPC start and resumes refill only after the response,
so crediting it causes cumulative drift that eventually triggers 429s. Advances ``last_refill``."""
if n_attachments <= 0:
return
async with self._lock:
token_before = self.tokens
self.tokens = max(0.0, token_before - float(n_attachments))
self.last_refill = time.monotonic()
logger.log(logging.INFO if rpc_duration > 10 and n_attachments > 5 else logging.DEBUG,
"Signal scheduler: RPC for %d att took %.1fs — tokens %.1f → %.1f (deducted=%d, no upload refill "
"credited, refill=%.4fs⁻¹)", n_attachments, rpc_duration, token_before, self.tokens, n_attachments,
self.refill_rate)
def feedback(self, retry_after: Optional[float], n_attempted: int) -> None:
"""Apply server feedback after a 429: empty the bucket and, when ``retry_after`` (per-token refill
window) is present, calibrate ``refill_rate`` from it."""
if retry_after and retry_after > 0 and (new_rate := 1.0 / float(retry_after)) != self.refill_rate:
logger.info("Signal scheduler: calibrating refill_rate to %.4f tokens/sec (server retry_after=%.1fs "
"per token)", new_rate, retry_after)
self.refill_rate = new_rate
self.tokens = 0.0
self.last_refill = time.monotonic()
def state(self) -> dict:
"""Current scheduler state for diagnostic logging (read-only, doesn't advance ``last_refill``)."""
return {"tokens": round(self._projected_tokens(), 1), "capacity": int(self.capacity),
"refill_rate": round(self.refill_rate, 4),
"refill_seconds_per_token": round(1.0 / self.refill_rate, 1) if self.refill_rate > 0 else float("inf")}
_scheduler: Optional[SignalAttachmentScheduler] = None
def get_scheduler() -> SignalAttachmentScheduler:
"""Return the process-wide scheduler, creating it on first access."""
global _scheduler
if _scheduler is None:
_scheduler = SignalAttachmentScheduler()
logger.info("Signal scheduler: created (capacity=%d tokens, refill=%.4f/s ≈ %.1fs/token)",
int(_scheduler.capacity), _scheduler.refill_rate, 1.0 / _scheduler.refill_rate)
return _scheduler
def _reset_scheduler() -> None:
"""Drop the cached scheduler so the next ``get_scheduler`` builds a fresh one. Test-only."""
global _scheduler
_scheduler = None