fix(run_agent): skip OpenAI client rebuild on Anthropic-native sessions

[noodlemctwoodle]

What

_replace_primary_openai_client was called unconditionally by all three stream-cleanup paths in chat_completion_helpers (stale_stream_pool_cleanup, stream_retry_pool_cleanup, stream_mid_tool_retry_pool_cleanup). When api_mode=anthropic_messages the agent owns _anthropic_client and self.client is None; _client_kwargs is an empty dict. The unconditional OpenAI(**{}) call raised AuthenticationError with "The api_key client option must be set … OPENAI_API_KEY" even though the correct Anthropic key was present.

This produced a deterministic false-alarm WARNING in errors.log on every Anthropic stream stall for profiles that have ANTHROPIC_API_KEY but no OPENAI_API_KEY — which is correct configuration. 3 occurrences verified on a single session.

Root cause

All three cleanup sites call agent._replace_primary_openai_client(). That method calls _create_openai_client(self._client_kwargs, ...) → OpenAI(**{}). The OpenAI SDK then reads the env for OPENAI_API_KEY, finds nothing, and raises. The exception is caught by the except Exception: pass guard at each call site (so the agent continues), but the logger.warning inside _replace_primary_openai_client fires first and lands in errors.log with a misleading message.

Fix

Short-circuit on api_mode == "anthropic_messages" and call _rebuild_anthropic_client() instead — the existing method that already handles both direct Anthropic and Bedrock-hosted variants and honours _oauth_1m_beta_disabled. The OpenAI path is unchanged; the guard only fires on the Anthropic wire.

Tests

4 new regression tests in tests/run_agent/test_anthropic_stream_pool_cleanup.py:

• OpenAI() constructor never called for Anthropic-native sessions
• _rebuild_anthropic_client is invoked instead
• Failure emits the right warning text (Anthropic client, not OpenAI client)
• OpenAI path still fires for api_mode=chat_completions

All 4 pass; 19 adjacent lifecycle/switch-model tests unaffected.

[alt-glitch]

This was generated by AI during triage.

Duplicate of #36719 — same guard in the same file (run_agent.py _replace_primary_openai_client short-circuit for api_mode==anthropic_messages) fixing the misleading OPENAI_API_KEY warning on Anthropic stream stalls. #36719 no-ops; this rebuilds the Anthropic client (the approach #44076 takes in chat_completion_helpers). Related to the broader stale-stream cluster #44076/#28161/#36693.

[teknium1]

Closed in favor of #53926, now merged to main (commit a0b9663).

This was a popular bug — 12 independent PRs fixed the same issue (#28161): the three stream-cleanup paths in the streaming code rebuilt the OpenAI primary client unconditionally, which on Anthropic-native sessions both failed (no OPENAI_API_KEY) and left the wedged stream open, causing the ~15-minute hang.

We salvaged @EloquentBrush0x's #28240 (the earliest dedicated fix for this issue) onto current main, resolved the conflict against the newer request-client cleanup helper, repointed the bug-encoding regression test, and added coverage for the two reachable Anthropic cleanup sites. Verified live: stale stream torn down in <1s instead of ~900s, Anthropic client closed+rebuilt, OpenAI rebuild never called on the Anthropic path.

Thank you for the fix — closing as a duplicate of the merged work. Credit to everyone who reported and fixed this.

#53926