fix(fireworks): set finish_reason=tool_calls for content-embedded tool calls

[abhay-codes07]

Relevant issues

Fixes #33036

Pre-Submission checklist

• I have added meaningful tests
• My PR passes lint / format / unit tests locally
• My PR's scope is as isolated as possible; it only solves 1 specific problem

Screenshots / Proof of Fix

Reproducible offline (no keys/network).

Before (base 3d63eda) — tool call present but finish_reason wrong:

TOOL-CALL : tool_calls=['get_weather'] | finish_reason='stop'

After (this PR):

TOOL-CALL : tool_calls=['get_weather'] | finish_reason='tool_calls'
PLAIN-TEXT: content='Just a normal text answer.' | finish_reason='stop'   # unaffected

New regression tests (the tool-call one fails on base, passes here) + full Fireworks file green:

tests/test_litellm/llms/fireworks_ai/chat/test_fireworks_ai_chat_transformation.py  50 passed
  (incl. test_transform_response_sets_finish_reason_tool_calls_for_content_tool_call
         test_transform_response_keeps_stop_for_plain_text)

Type

🐛 Bug Fix

Changes

transform_response builds the ModelResponse directly (bypassing convert_to_model_response_object) and converts Fireworks' content-embedded tool call into message.tool_calls without updating finish_reason. Set finish_reason="tool_calls" after the conversion when a tool call is present and finish_reason is still "stop" — matching the shared response converter and providers like Ollama/xAI. Plain-text responses are unaffected.

---

cc @ishaan-jaff @krrish-berri-2 — isolated Fireworks tool-calling fix with regression tests. Thanks!

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[greptile-apps]

Greptile Summary

This PR fixes Fireworks AI's non-streaming transform_response path, which was leaving finish_reason=\"stop\" even after a content-embedded tool call was promoted into message.tool_calls, breaking agent loops that branch on finish_reason == \"tool_calls\".

• Adds a post-conversion check in transform_response: if finish_reason is still \"stop\" and message.tool_calls is populated, it is overwritten to \"tool_calls\", matching the contract of the shared OpenAI-style converter and providers like Ollama/xAI.
• Introduces two offline regression tests: one that asserts the flip occurs for a content-embedded tool call, and one that confirms plain-text responses remain "stop".

Confidence Score: 5/5

The change is a small, focused one-liner inside a Fireworks-specific code path with no impact on other providers.

The fix is correctly scoped to transform_response (the non-streaming path where the bug existed), does not touch shared infrastructure, and is covered by two new offline regression tests. The only finding is a minor redundant length guard.

No files require special attention.

Important Files Changed

Filename	Overview
litellm/llms/fireworks_ai/chat/transformation.py	Adds finish_reason="tool_calls" correction after content-embedded tool calls are promoted to message.tool_calls; logic is correct and well-scoped to the non-streaming path.
tests/test_litellm/llms/fireworks_ai/chat/test_fireworks_ai_chat_transformation.py	Adds two focused regression tests: one confirming finish_reason flips to "tool_calls" when content holds a tool call, one confirming plain-text responses remain "stop". Both are offline/mock-only.

_{Reviews (1): Last reviewed commit: "fix(fireworks): set finish_reason=tool_c..." | Re-trigger Greptile}

[greptile-apps]

The len(typed_choice.message.tool_calls) > 0 guard is redundant — the preceding truthiness check typed_choice.message.tool_calls already returns False for both None and an empty list, so the length check adds no new protection.

Suggested change

	if (
	typed_choice.finish_reason == "stop"
	and typed_choice.message.tool_calls
	and len(typed_choice.message.tool_calls) > 0
	):
	if (
	typed_choice.finish_reason == "stop"
	and typed_choice.message.tool_calls
	):

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

[abhay-codes07]

Heads up: the red auth-and-jwt check here is a pre-existing failure on the litellm_oss_daily_2026_07_10 base (the test_can_key_call_model_wildcard_access/test_can_team_access_model bedrockz/* cases), unrelated to this change. Root-caused and fixed separately in #33031; this PR's own changes are fully covered and green. Once #33031 lands in the base, this check goes green on re-run.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!