I kept running into the same bug in AI agent codebases: retry logic that ignores Retry-After under concurrency.

Looks fine at first. Under load it turns rate limits into request storms.

I wrote a small CLI to catch it:

  npx pitstop-check ./src

It scans TS/JS and flags things like:

  - 429 handled without Retry-After
  - blanket retry of all 429s (no CAP vs WAIT distinction)
  - unbounded retry loops (no max elapsed)

Example (ran against OpenClaw):

  [WARN] src/agents/venice-models.ts:24 — 429 handled without Retry-After
  [WARN] src/agents/venice-models.ts:24 — All 429s treated as retryable — CAP vs WAIT not distinguished

The retry primitive supports Retry-After. The callers just don’t wire it up.

So when the API returns Retry-After: 600, the client retries on its own schedule instead of backing off.

What’s going on is basically collapsing different failure modes into one:

  WAIT — respect Retry-After
  CAP  — limit retries / concurrency
  STOP — don’t retry

Most code just does:

  retry()

The tool is heuristic (will flag some test files), but it’s been useful for quickly spotting this in real repos.

https://github.com/SirBrenton/pitstop-check