Raise on LiteLLM response.failed streams

[ousamabenyounes]

Fixes #719.

Summary

• Adds a Strix LiteLLM compatibility shim that raises when a Responses API stream yields response.failed.
• Preserves the provider failure message/code in the raised APIError, so SDK retry policy can treat the failed stream as an actual model failure instead of normal completion.
• Covers the real LiteLLM BaseResponsesAPIStreamingIterator._process_chunk path with a regression test.

Validation

• RED before fix: uv run pytest tests/test_litellm_response_failed.py -q -> Failed: DID NOT RAISE APIError
• GREEN after fix: uv run pytest tests/test_litellm_response_failed.py -q -> 1 passed
• Revert proof: temporarily reverted only strix/config/models.py patch -> Failed: DID NOT RAISE APIError
• uv run pytest tests/test_litellm_response_failed.py tests/test_model_retry.py tests/test_cost_tracking.py -q -> 17 passed
• uv run ruff check strix/config/models.py tests/test_litellm_response_failed.py -> pass
• uv run ruff format --check strix/config/models.py tests/test_litellm_response_failed.py -> pass
• uv run mypy strix/config/models.py tests/test_litellm_response_failed.py -> pass
• uv run bandit -q -c pyproject.toml strix/config/models.py -> pass

[greptile-apps]

Greptile Summary

Adds a LiteLLM compatibility shim that converts streamed response.failed events into APIError exceptions while preserving provider error details.

• Hooks LiteLLM’s Responses API streaming iterator during SDK model configuration.
• Handles dictionary and object-shaped response events and supplies fallback error text.
• Adds a regression test against the real LiteLLM streaming iterator path.

Confidence Score: 5/5

The PR appears safe to merge, with the failed-stream conversion covered through LiteLLM’s actual chunk-processing path.

The compatibility patch is idempotent across repeated SDK configuration, tolerates supported event representations, and turns the previously silent failed event into the retry-classifiable exception intended by the change.

Important Files Changed

Filename	Overview
strix/config/models.py	Adds an idempotent LiteLLM streaming compatibility patch that raises an API error when a transformed stream event reports response.failed.
tests/test_litellm_response_failed.py	Adds focused regression coverage proving failed response events preserve their provider message and code in the raised exception.

_{Reviews (1): Last reviewed commit: "fix: raise litellm failed response strea..." | Re-trigger Greptile}