fix: harden codex stale timeout and minimax overload retries

[Obrais-cloud]

Summary

• estimate Codex/Responses request size from input + instructions + tools, not just messages
• prevent false 300s stale aborts on large openai-codex non-streaming requests
• add retry wait-time selection so overloaded upstreams (503/529, e.g. MiniMax) back off longer and respect Retry-After
• add regression tests for both behaviors

Test Plan

• python -m pytest tests/test_retry_utils.py -q
• python -m pytest tests/run_agent/test_run_agent.py -q -k "RequestContextTokenEstimate"
• python -m pytest tests/run_agent/test_run_agent_codex_responses.py -q

[teknium1]

Thanks for the focused retry and timeout work. The Codex half has since landed through a broader refactor, but the generic overload retry gap remains.

Problems

• Current main already estimates Responses payloads from input, instructions, and tools in agent/chat_completion_helpers.py:66-112; it also applies the OpenAI-Codex large-payload stale floor at agent/chat_completion_helpers.py:131-147 and :387-393 (commit 2d422720b). The run_agent.py changes in this PR therefore need to be dropped during salvage.
• The retry gap is now in agent/conversation_loop.py: 503/529 classify as overloaded (agent/error_classifier.py:1033-1044), but overloaded is excluded from is_rate_limited (agent/conversation_loop.py:3141-3149), so Retry-After is not read (:4094-4113).
• The proposed 180-second cap conflicts with current main's documented 600-second Retry-After ceiling at agent/conversation_loop.py:4102-4107.

Suggested changes

• Re-scope the salvage to the current retry loop, preserve the 600-second cap, and add a conversation-loop test for a 503/529 overload with Retry-After.

Automated hermes-sweeper review.

[teknium1]

Please do not introduce a separate 180s cap when salvaging this. Current main intentionally caps Retry-After at 600s because a 120s cap retried before known reset windows (agent/conversation_loop.py:4102-4107). Reuse or centralize that policy and cover an overload response with a Retry-After header.

[GottZ]

This was generated by AI during triage.

Summary

Five PRs address overlapping retry and provider-handling failures: #47205/#52333 add named custom-provider matching and Z.AI-specific adaptive backoff, #59523/#60034 make that backoff reachable on the overloaded path, and #11462 combines an already-landed Codex timeout change with a still-distinct generic 503/529 Retry-After gap.

Related pull requests

• #11462 related — (+128/-10) — keep open with a salvage path: drop the Codex request-size/stale-timeout changes because commit 2d422720b already implements them in agent/chat_completion_helpers.py:66-112, :131-147, and :387-393; retain and rework the generic overload handling in agent/conversation_loop.py so 503/529 overloads read Retry-After while preserving the existing 600-second ceiling. This follows the contributor keep_open review rather than treating the current mixed diff as mergeable.
• #47205 [closed] related — (+237/-5) — already implemented on main: this closed PR added custom:<provider> extra_body matching and narrow adaptive backoff for Z.AI Coding Plan 429/1305 overloads; the discussion identifies equivalent upstream commit 6f2b2a1f34, and #52333 is the merged salvage/reference implementation.
• #52333 [merged] related — (+238/-5) — merged reference implementation: it salvaged #47205 by adding named custom-provider extra_body matching and the Z.AI-specific 30/60/90/120-second overload policy, but its original gate left that policy unreachable for errors classified as overloaded and under the default retry ceiling.
• #59523 [closed] related — (+87/-4) — canonical twin, merged through #60034: it identified and fixed both reachability defects by running Z.AI 429/1305 handling on the overloaded path and raising the retry ceiling so the long tier executes. The contributor discussion confirms its commits were cherry-picked into #60034 with authorship preserved.
• #60034 [merged] related — (+96/-5) — merged reference for the #59523 fix: its diff carries the same overloaded-path routing and retry-ceiling mechanism, plus a shared short-attempt constant. The visible contributor triage review calls #59523 the canonical twin by creation time; #60034 is retained here only as the merged implementation, not as a replacement for that canonical designation.

Duplicates

#47205 and #52333 are substantially the same provider-matching and Z.AI adaptive-backoff change, with #52333 the merged salvage. #59523 and #60034 are twins; #59523 remains the contributor-designated canonical PR, while #60034 is the merged vehicle containing its commits plus the shared-constant refinement.

Suggested consolidation

Keep #11462 open with a salvage path: remove the Codex half already covered by commit 2d422720b, retarget the remaining generic 503/529 Retry-After fix to agent/conversation_loop.py, preserve the documented 600-second ceiling, and add a conversation-loop regression test. Treat closed #47205 as superseded by merged #52333, and closed #59523 as the canonical source twin whose implementation landed through merged #60034; no reopen or merge action is needed for those completed duplicate chains.

Cross-PR triage: Reviewed 5 pull requests and 0 issues in this complex. Each diff was read against this issue; Assessment working set: 54 kB of PR diffs, 9 kB of issue/PR text, 6 kB of discussion (7 comments), 0 verify verdicts. verdicts reflect diff content, not PR titles. Part of an automated triage batch.