[yagop]

Fixes #1390.

Problem

RateLimiter.acquire() took the global token, then waited on the chat's bucket. A request queued behind a busy chat spent its global token right away and was sent later, with no fresh check of the global rate. A chat backlog could then push sends over the global cap in short windows (the long-run average still can't exceed it), and it wasted the global burst on requests that wait for seconds.

The swap suggested in the issue (chat first, global last) moves the same problem to the per-chat cap. While the global bucket is contended, a second request for the same chat takes its chat token and waits too. When the global bucket frees up, both can go out within milliseconds, because take() is a wake-and-recheck loop, not a queue.

Fix

TokenBucket.takeAll(buckets, signal) waits max(waitMs) over all buckets, rechecks after waking, and takes one token from each only when all have one. acquire() passes it the global bucket and the chat bucket. take() is now takeAll([this]), so single-bucket behavior is unchanged.

• Neither token is spent while a request waits on the other, so both limits are checked at send time.
• An aborted wait spends no token from either bucket. The swap would have lost one chat token.
• TokenBucket is exported, so TokenBucket.takeAll is new public API (CHANGELOG and doc/api.md updated). Nothing existing changes signature.

Tests

Two new cases in test/unit/ratelimiter.test.ts, both on the fake clock:

• A request waiting on its chat holds no global token, and another chat gets it immediately. This fails on the old order.
• A request waiting on the global cap holds no chat token, and an abort spends neither. This fails on the swapped order (checked locally).

npm run check passes (305 unit tests).

[Copilot]

🟡 Changes recommended

The new takeAll method checks duplicate bucket references before decrementing, allowing requests to exceed available tokens and drive the token count negative.

1 open issue

• Deduplicate buckets before checking and decrementing tokens · New

What changed in this PR

Addresses #1390 by acquiring global and per-chat rate-limit tokens together, preventing queued requests from spending one token while waiting for the other.

Changes:

• Adds public TokenBucket.takeAll() and uses it for single-bucket and combined acquisition.
• Adds regression tests for token preservation during waits and cancellation.
• Updates the API reference and changelog.

File	Description
test/​unit/​ratelimiter.test.ts	Tests token preservation while either limit blocks acquisition.
src/​core/​ratelimiter.ts	Implements combined token acquisition and integrates it into RateLimiter.
doc/​api.md	Documents takeAll() and the revised rate-limiting behavior.
CHANGELOG.md	Records the rate-limiting fix and new public method.

_{Review effort: Balanced}

[yagop]

Good catch, fixed in 4683eb3. takeAll now deduplicates with a Set before the check and decrement loops, so a repeated bucket is checked and charged once. RateLimiter itself never passes the same bucket twice (global and chat buckets are distinct objects), so this only affected direct callers of the public method. Added a test: takeAll([b, b]) on a 1-token bucket leaves it at 0, and a second call waits instead of driving it negative.