fix: surface Anthropic stop_reason to detect truncation (#5148)

[devin-ai-integration]

Summary

Fixes #5148. Anthropic's Message response includes a stop_reason field that indicates why the API stopped generating (e.g. "max_tokens" means the output was truncated). Previously, AnthropicCompletion silently discarded this field, making it impossible for users to detect truncation via hooks or events.

Changes:

• Add stop_reason: str | None = None field to LLMCallCompletedEvent
• Add stop_reason parameter to BaseLLM._emit_call_completed_event
• Add _extract_stop_reason() static method on AnthropicCompletion that safely extracts stop_reason as str | None (guards against non-string values, e.g. from MagicMock in tests)
• Add _warn_if_truncated() helper on AnthropicCompletion that logs a warning when stop_reason == "max_tokens"
• Plumb stop_reason through all 6 Anthropic completion methods (sync + async × regular/streaming/tool-use)
• Add 7 unit tests covering warning behavior, event field propagation, and edge cases

Non-Anthropic providers are unaffected — they continue to emit stop_reason=None by default.

Review & Testing Checklist for Human

• Verify all 6 methods are consistently updated: The same pattern (extract → warn → pass to event) is applied across _handle_completion, _handle_streaming_completion, _handle_tool_use_conversation, and their async counterparts. Confirm no code path was missed, especially around early returns (structured output, tool use blocks).
• Streaming path correctness: In _handle_streaming_completion / _ahandle_streaming_completion, stop_reason is read from stream.get_final_message(). Verify that the reconstructed final_message actually carries the stop_reason from the stream (it should, per Anthropic SDK behavior).
• Consider whether from_agent.role access is safe: _warn_if_truncated accesses from_agent.role when from_agent is truthy. This should always work since from_agent is typed as Agent | None, but confirm no caller passes a non-Agent truthy value.
• Decide if other providers should also surface finish reasons: This PR is Anthropic-only. OpenAI has an analogous finish_reason field — consider whether a follow-up is needed.

Suggested manual test: Set max_tokens to a very small value (e.g. 50) on an Anthropic model, run a crew, and verify:

1. A warning is logged containing stop_reason='max_tokens'
2. The LLMCallCompletedEvent carries stop_reason="max_tokens" (observable via a custom event handler)

Notes

• The stop_reason field on LLMCallCompletedEvent and the BaseLLM parameter are additive and backwards-compatible (default None).
• _extract_stop_reason uses an isinstance(raw, str) guard so that non-string attribute values (e.g. auto-created MagicMock attributes in existing tests) safely become None rather than causing Pydantic validation errors.
• No async-specific tests for the warning (only sync), though the code is symmetrical. The event-bus tests do exercise the full emit→capture path.

Link to Devin session: https://app.devin.ai/sessions/7214a66c41b94b07803ad5faacf12270

[devin-ai-integration]

Prompt hidden (unlisted session)

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment and CI monitoring