fix(agent): recover Anthropic stale streams without OPENAI_API_KEY errors

[gislihaukur]

Summary

• Re-resolve Anthropic OAuth credentials when rebuilding the client after a stale stream kill, instead of falling back to an empty OPENAI_API_KEY
• Close the in-flight Anthropic streaming client before rebuild so the worker thread unblocks
• Abort after repeated stale-stream reconnect attempts instead of spinning forever

Context

Observed on Telegram profiles using Anthropic OAuth: stale SSE streams triggered client rebuild, which then failed with AuthenticationError: Missing OPENAI_API_KEY because rebuild reused stale/empty credentials.

Test plan

• scripts/run_tests.sh tests/run_agent/test_run_agent.py -q (new tests for rebuild key resolution + stale-stream abort path)
• Manual: trigger stale stream on Anthropic OAuth profile and confirm reconnect succeeds without OPENAI_API_KEY error

Made with Cursor

[alt-glitch]

This was generated by AI during triage.

The core fix here — routing Anthropic stale-stream cleanup to the Anthropic client instead of the OpenAI rebuild — has already landed on main via the merged #53926 (the stale-stream cleanup site now branches on api_mode == "anthropic_messages" and calls _rebuild_anthropic_client()). This PR re-resolves Anthropic credentials before still calling the OpenAI rebuild, which is superseded by that merge. The net-new piece worth reviewing on its own is the stale-stream retry-abort cap (HERMES_STREAM_RETRIES). Related: merged #53926, cluster #28161 / #51844.

[tonydwb]

Code Review Summary

Verdict: LGTM

Good fix for Anthropic stale stream recovery. The key improvement is closing the Anthropic client before rebuilding — this unblocks the for event in stream loop that would otherwise spin forever. The retry count tracking with configurable HERMES_STREAM_RETRIES is a nice touch for observability.

Correctness: The Anthropic-specific path correctly closes the client before rebuild, matching the non-streaming/interrupt paths.
Code Quality: Clean separation of Anthropic vs non-Anthropic recovery paths.

---

Reviewed by Hermes Agent

[tonydwb]

Code Review Summary

Verdict: LGTM

Fixes stale stream recovery for Anthropic. Re-resolves OAuth credentials when rebuilding the client, closes in-flight Anthropic client to unblock the worker thread, and adds retry logic with configurable max retries.

• Clean fix for a real production issue
• Good error handling and retry logic
• Well-scoped: 3 files, 178 additions

Reviewed 3 files, 178 additions. Approved.

[teknium1]

Thanks for the stale-stream recovery work. This is already implemented on current main; this automated hermes-sweeper review is closing the superseded PR.

• a0b9663c7cfed007acede4f1261212811aa623ef implemented the Anthropic-native cleanup path and shipped in v2026.7.1.
• agent/chat_completion_helpers.py:3080 now closes and rebuilds the Anthropic client for anthropic_messages; the OpenAI-client rebuild remains only in the non-Anthropic branch at :3087.
• tests/run_agent/test_28161_anthropic_stream_pool_cleanup.py:109 covers the stale-stream path and asserts that the OpenAI rebuild is not called.
• Current main also has a stale-stream circuit breaker at agent/chat_completion_helpers.py:2093 and :3075, addressing repeated unresponsive-provider attempts.

This matches the prior #53926 discussion identifying the Anthropic cleanup routing as already landed.