Repository navigation
Description
What happens
Two things, checked on smolagents==1.26.0 and on current main (96f33fa).
-
Retry-Afteris never read.grep -ri "retry-after" src/returns nothing. When a provider answers429and says "come back in 3 seconds", the model-level retry sleeps for whatever the exponential schedule says instead. -
The first retry waits 2–3× the configured
wait_seconds. Is multiplying before the first sleep intended? InRetrying.__call__the delay is multiplied byexponential_base(plus jitter) and then slept, so the first wait iswait_seconds × 2 × (1 + jitter)rather thanwait_seconds. Measured below withRETRY_WAIT=2: the first gap between retry groups is 5.16 s, where the configured value was 2 s.
With the ApiModel defaults (max_attempts=3, wait_seconds=60, exponential_base=2, jitter on) this means a single rate-limited call sleeps 120–240 s and then 240–960 s — 6–20 minutes before it gives up.
Reproduction
A local OpenAI-compatible server that answers the first N /v1/chat/completions requests with 429 and a Retry-After: 3 header, then a normal response; one OpenAIModel call through a ToolCallingAgent.
RETRY_WAIT=2 python r429.py # RETRY_WAIT only scales the waits down so the run finishes quicklyRequests the server received for one logical call, with Retry-After: 3 on every 429:
| # | t (s) | x-stainless-retry-count |
note |
|---|---|---|---|
| 1 | 2.25 | 0 | first request |
| 2 | 5.25 | 1 | 3.00 s later — the openai client does honour the header |
| 3 | 8.25 | 2 | 3.00 s later |
| 4 | 13.41 | 0 | model-level retry: 5.16 s after the previous attempt |
| 5 | 16.41 | 1 | 3.00 s |
| 6 | 19.41 | 2 | 3.00 s |
| 7 | 34.23 | 0 | 14.82 s later |
| 8 | 37.23 | 1 | 3.00 s |
| 9 | 40.23 | 2 | 3.00 s |
So: 9 HTTP requests for one logical call. The openai client's own retries honour Retry-After (3.00 s apart, the counter going 0→1→2), but the two waits between retry groups are 5.16 s and 14.82 s — with RETRY_WAIT=2 those should have been 2 s and 4 s if the schedule started at wait_seconds, or 3 s if the header had been used. Neither wait relates to the Retry-After: 3 the server sent.
Raw data (public, reproducible): evidence/e0-retry-after/ra3_smol.jsonl, r429.py (the server) and NOTE.md in https://github.com/simin-yuan/agent-double-exec-bench.
Why it matters
On a 429 with a short Retry-After, the agent stays blocked for minutes when the provider was ready in seconds. With max_attempts=3 and the default 60 s base, the worst case is 20 minutes of wall-clock for one call, which is longer than most CI jobs and than most people's patience with a CLI.
Related but different: #2586 reports that wait_seconds defaults to 0.0 (so backoff is off) and that reraise has no effect — that is about the default value and the error path. This one is about the schedule when a wait is configured, and about ignoring a server-declared cooldown.
What I'd expect
- If the server sends
Retry-After, sleep that long — and never sleep less than the server asked for. - If the doubling before the first sleep is intentional, a comment saying so would help: as written,
wait_secondsmeans "half of the first wait", which is easy to misread. If it isn't intentional, the first retry should waitwait_secondsand doubling should start after that.
Status
A patch with five tests is ready locally (first-retry delay, jitter bounds, Retry-After honoured, opt-out flag, fallback when the header is absent). I have not opened a PR: I have one open PR (#3036) and I'll follow the one-open-PR rule until it lands. Happy to post the diff here, or to open the PR as soon as the issue is accepted and #3036 is resolved — whichever the maintainers prefer.
Disclosure: I used an AI assistant to help draft this issue and to write and run the patch. I have read the change, run the tests, and will answer review comments.
Activity
simin-yuan commentedon Oct 20, 2026
Prior art on one of the two points, which I should have checked before opening this.
#2596 (opened 2026-08-02, still open) fixes exactly my point 2: it moves the sleep before the
multiplication, so the first retry waits
wait_secondsand the doubling starts after that. Its testsassert sleeps of 1, 2, 4 with the default. I re-derived the same change without knowing that PR existed;
that finding is theirs, not mine.
#2662 (2026-08-20) and #2617 (2026-08-11) touch the same class but only bring the
wait_secondsdefault to 1.0 and make
reraisewrap intoRetryError. Neither moves the sleep, so neither coverspoint 2, and neither has had a maintainer reply either. All three edit the same lines of
Retrying.__call__, and#2586— which all three say they fix — has been open since 2026-07-29 withouta label.
What is left that none of them has:
Retry-Afteris never read (grep -ri "retry-after" src/returnsnothing), and the cost of that is measurable — with
RETRY_WAIT=2the waits between retry groups were5.16 s and 14.82 s while the server asked for 3 s, 40 s of wall clock for one logical call, and at the
shipped defaults (
wait_seconds=60) 6–20 minutes.So the honest state of this issue: if #2596 lands, point 2 is closed and only the header work is left. I
would rather say that than leave four competing changes over the same two lines.
Disclosure: I used an AI assistant to help draft this comment. I will answer replies myself.