fix(client): honor Retry-After delays up to two minutes

[tescherm-openai]

Summary

• honor positive, finite Retry-After delays up to 120 seconds
• treat longer server-directed delays as non-retryable instead of substituting short exponential backoff
• keep exponential backoff for zero, negative, non-finite, missing, or malformed values
• cover sync and async clients with numeric seconds, milliseconds, and HTTP-date boundaries

Motivation

The Python SDK currently ignores Retry-After values above 60 seconds and falls back to a much shorter exponential delay. Historical reset-header data shows that common minute-scale waits can land just above 60 seconds, while extending the ceiling beyond two minutes adds little coverage before waits jump to hours or days.

Developer impact

Clients with automatic retries enabled will honor server-directed delays through two minutes. If Retry-After exceeds two minutes, the SDK surfaces the API error instead of blocking a synchronous worker for an extended period or retrying before the requested time. Retry attempts remain bounded by max_retries.

Validation

• .venv/bin/ruff format --check src/openai/_constants.py src/openai/_base_client.py tests/test_client.py
• .venv/bin/ruff check src/openai/_constants.py src/openai/_base_client.py tests/test_client.py
• .venv/bin/pytest -q tests/test_client.py (198 passed, 2 skipped)

[openai-sdks]

OkTest Summary

✅ 237/237 SDK tests passed in 11.956s for Python SDK PR #3555.

Test results — 42 files

Test	Result	Time
tests/chat-completions-complex-body.test.ts	✅ Passed	202ms
tests/chat-completions-create.test.ts	✅ Passed	272ms
tests/chat-completions-stream.test.ts	✅ Passed	155ms
tests/files-content-binary.test.ts	✅ Passed	248ms
tests/files-create-multipart.test.ts	✅ Passed	241ms
tests/files-list-pagination.test.ts	✅ Passed	203ms
tests/initialize-config.test.ts	✅ Passed	198ms
tests/instance-isolation.test.ts	✅ Passed	234ms
tests/models-list.test.ts	✅ Passed	141ms
tests/responses-background-lifecycle.test.ts	✅ Passed	222ms
tests/responses-body-method-errors.test.ts	✅ Passed	386ms
tests/responses-cancel-timeout.test.ts	✅ Passed	231ms
tests/responses-cancel.test.ts	✅ Passed	350ms
tests/responses-compact-retries.test.ts	✅ Passed	336ms
tests/responses-compact.test.ts	✅ Passed	250ms
tests/responses-create-advanced-stream.test.ts	✅ Passed	169ms
tests/responses-create-advanced.test.ts	✅ Passed	306ms
tests/responses-create-disconnect.test.ts	✅ Passed	1.279s
tests/responses-create-errors.test.ts	✅ Passed	206ms
tests/responses-create-malformed-api-responses.test.ts	✅ Passed	278ms
tests/responses-create-retries.test.ts	✅ Passed	371ms
tests/responses-create-stream-failures.test.ts	✅ Passed	133ms
tests/responses-create-stream-timeout.test.ts	✅ Passed	245ms
tests/responses-create-stream-wire.test.ts	✅ Passed	3.837s
tests/responses-create-stream.test.ts	✅ Passed	352ms
tests/responses-create-terminal-states.test.ts	✅ Passed	268ms
tests/responses-create-timeout.test.ts	✅ Passed	215ms
tests/responses-create.test.ts	✅ Passed	272ms
tests/responses-delete.test.ts	✅ Passed	240ms
tests/responses-input-items-errors.test.ts	✅ Passed	345ms
tests/responses-input-items-list.test.ts	✅ Passed	346ms
tests/responses-input-items-options.test.ts	✅ Passed	395ms
tests/responses-input-tokens-count-timeout.test.ts	✅ Passed	269ms
tests/responses-input-tokens-count.test.ts	✅ Passed	305ms
tests/responses-malformed-inputs.test.ts	✅ Passed	2.991s
tests/responses-not-found-errors.test.ts	✅ Passed	304ms
tests/responses-parse.test.ts	✅ Passed	413ms
tests/responses-retrieve-retries.test.ts	✅ Passed	298ms
tests/responses-retrieve.test.ts	✅ Passed	276ms
tests/responses-stored-method-errors.test.ts	✅ Passed	726ms
tests/retry-behavior.test.ts	✅ Passed	3.102s
tests/sdk-error-shape.test.ts	✅ Passed	357ms

View OkTest run #30584584865

_{SDK merge (f5f40e6ac3c9) · head (05b8ddf8b125) · base (3844843c277f) · OkTest (91635c6a2723)}

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 6e55b90fed

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Ignore unconvertible Retry-After dates before retry checks

When automatic retries are enabled, a non-retryable error such as HTTP 400 can now raise ValueError or OverflowError instead of APIStatusError if it carries a syntactically parseable but out-of-range date, such as Retry-After: Fri, 29 Sep 100000 16:26:57 GMT: parsedate_tz() accepts it, but mktime_tz() in _parse_retry_after_header() fails. Previously _should_retry() did not parse this header for statuses that would not be retried, so conversion failures should be treated as malformed header values rather than escaping from this new call.

Useful? React with 👍 / 👎.

[tescherm-openai]

Good catch. parsedate_tz() can accept dates that mktime_tz() cannot represent, so I updated _parse_retry_after_header() to treat TypeError, ValueError, OverflowError, and OSError as malformed header values. I also added sync and async regressions confirming that an HTTP 400 with this date still raises APIStatusError after one request, plus timeout-calculation coverage for the fallback behavior. The fix is committed locally as 8d149be8 and has not been pushed yet.

[HAYDEN-OAI]

Reviewed the current head and found no substantive correctness or compatibility issues. The sync and async retry paths consistently honor finite server-directed delays through 24 hours, reject longer delays without retrying early, and preserve fallback/status-error behavior for malformed or non-finite values.