fix(llm): use validated messages variable instead of raw params['messages'] access

[devin-ai-integration]

Summary

Fixes #5164. Replaces all direct params["messages"] accesses in _handle_emit_call_events calls with safe alternatives across three methods:

• _handle_non_streaming_response and _ahandle_non_streaming_response: In the response_model + is_litellm early-return branch, uses the already-extracted local messages variable (from params.get("messages", [])) instead of re-accessing params["messages"]. In all other branches, uses params.get("messages", []).
• _handle_streaming_response: Uses params.get("messages", []) for all emit calls (this method had no local messages variable).

12 occurrences of params["messages"] replaced total. _ahandle_streaming_response was already using params.get("messages") and was not changed.

Review & Testing Checklist for Human

• Verify the two uses of the bare messages local variable (lines ~1220 and ~1364) are actually in scope — they should be, since messages = params.get("messages", []) is assigned earlier in the same if response_model and self.is_litellm: block
• Confirm no behavioral regression for the normal path where params["messages"] does exist — params.get("messages", []) returns the same value as params["messages"] when the key is present
• Note: _ahandle_streaming_response uses params.get("messages") (returns None if missing) while the patched methods use params.get("messages", []) (returns []). Both are acceptable for _handle_emit_call_events which accepts None, but the inconsistency is worth being aware of

Notes

• In normal usage via call()/acall(), params["messages"] is always populated before these handlers are invoked. This fix is defensive — it prevents a confusing KeyError if the internal methods are ever called with incomplete params.
• 5 new unit tests added covering both sync and async non-streaming paths with and without response_model.

Link to Devin session: https://app.devin.ai/sessions/f9fa64e3c2b346e2b332c8b2e21c5512

---

Note

Low Risk
Low risk defensive change: swaps direct params["messages"] access for params.get(...)/validated messages, affecting only event emission metadata and adding tests to prevent regressions.

Overview
Prevents KeyError when internal LLM response handlers emit LLMCallCompletedEvent without a messages key in params (fix for #5164).

All _handle_emit_call_events call sites in streaming and (a)sync non-streaming paths now pass params.get("messages", []), and the LiteLLM response_model early-return branches pass the already-validated local messages variable. Adds unit tests covering sync/async non-streaming behavior with missing messages and with response_model to ensure proper errors (ValueError) and no KeyError.

^{Written by Cursor Bugbot for commit 25a13a0. This will update automatically on new commits. Configure here.}

[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

[greysonlalonde]

Duplicate of #5165. Closing.