fix(responses): raise failed results with their error code for retry routing

[Lucifus613]

Summary

OpenAIResponsesClient mishandles a Responses result whose status is failed:

• Non-streaming: _parse_responses_output parses a failed result as if it were a normal answer. The partial output is returned to the caller, including any function_call items, so the agent loop can execute tool calls from a response that the server itself reported as failed.
• Streaming: a response.failed event matches no branch and is silently ignored.
• Retry routing: even after a failure is raised, a plain LLMError carries no status code. The shared retry classifier therefore cannot tell a rate limit from a permanent 400-class error. Rate limits get generic backoff, and errors such as invalid_prompt are retried pointlessly.

This PR raises on failed results and keeps the provider's error code on the error. The code is mapped to the equivalent HTTP status so that the existing call_llm ladder routes it:

Code	Status	Routing
rate_limit_exceeded	429	rate-limit backoff
server_error	500	transient backoff
vector_store_timeout	504	transient backoff
other SDK codes (e.g. invalid_prompt, bio_policy, invalid_image*)	400	stops immediately (non_transient)
unknown or missing	none	unchanged generic retry

The table's key set is exactly the ResponseError.code literal in openai 3.7.0. The existing text-based rules still take precedence: safety refusals advance the fallback chain, and upstream-timeout wording stays transient. For example, image_content_policy_violation advances the chain through the existing content_policy_violation rule.

Changes

• agent_core/providers/openai_responses.py:
  • A failed result raises _ResponsesError(LLMError) with code and status_code. Its message is "<code>: <provider message>", falling back to Responses request failed.
  • The streaming response.failed event raises the same error.
• tests/test_responses_error_routing.py (new):
  • Every code, as both a dict and an SDK-shaped object, keeps its code and status.
  • Each failure routes through the real call_llm: reason, call count and backoff band.
  • A streamed response.failed event raises the same error as the parsed result.
• changes/responses-failed-error-routing.fix.md: changelog fragment.

Test plan

Run on this branch, which is main plus this commit:

• python -m pytest tests -q: 1715 passed (openai 3.7.0, anthropic 1.3.0).
• ruff check on the changed files: clean.
• The new test file copied onto unmodified main: 99 failed. With this commit: 99 passed.
• Live OpenAI call returning a failed result: not run (paid).

pyright agent_core reports 6 errors, all in agent_core/runtime/loop/agent_loop.py. Unmodified main reports the same 6, and this PR does not touch that file.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, add credits to your account and enable them for code reviews in your settings.

[zhanghanduo]

Thanks for the fix. Very helpful. Cannot reproduce LLM errors easily. Should be correct.