fix(gemini): correct streaming finish_reason for tool calls

[Chesars]

Relevant issues

Fixes #21041

Pre-Submission checklist

• I have Added testing in the tests/litellm/ directory, Adding at least 1 test is a hard requirement - see details
• My PR passes all unit tests on make test-unit
• My PR's scope is as isolated as possible, it only solves 1 specific problem
• I have requested a Greptile review by commenting @greptileai and received a Confidence Score of at least 4/5 before requesting a maintainer review

Type

🐛 Bug Fix

Changes

Gemini always returns finishReason: "STOP" regardless of whether the response contains tool calls. Per the OpenAI spec, finish_reason must be "tool_calls" when the model called a tool.

LiteLLM already maps this correctly for non-streaming responses (PR #10485), but in streaming mode the finishReason arrives in a separate chunk without content, which _process_candidates() skips entirely — so the mapping never runs.

Fix (in ModelResponseIterator, Gemini handler layer):

1. Track tool_calls across chunks — added self.has_seen_tool_calls flag that gets set when any streaming choice contains delta.tool_calls
2. Handle final chunk with no content — when _process_candidates produces empty choices but a candidate has finishReason, create a StreamingChoices with the correct finish_reason (respecting prior tool_calls)

Tests (4 new, all mock-only, no network calls):

• Tool call stream → finish_reason="tool_calls" ✓
• Text-only stream → finish_reason="stop" ✓
• Multiple parallel tool calls → finish_reason="tool_calls" ✓
• Content filter → finish_reason="content_filter" preserved ✓

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
litellm	Ready	Preview, Comment	Feb 19, 2026 6:05pm

[greptile-apps]

Greptile Summary

This PR fixes a streaming bug where Gemini's finishReason="STOP" was incorrectly mapped to finish_reason="stop" instead of finish_reason="tool_calls" when tool calls were present.

Root cause: Gemini sends tool calls and finishReason in separate chunks. The final chunk (with finishReason but no content) was skipped by _process_candidates(), causing the finish_reason to be lost entirely.

Solution:

• Track tool calls across chunks using has_seen_tool_calls flag
• Handle final chunks without content by creating a StreamingChoices with the correct finish_reason based on whether tool calls were seen earlier

Tests: 4 new comprehensive mock-based tests covering all scenarios (tool calls, text-only, multiple tools, content filter).

All changes are provider-specific and properly isolated within llms/vertex_ai/gemini/. No database queries, fastapi imports, or other violations of custom rules.

Confidence Score: 5/5

• This PR is safe to merge with minimal risk
• The fix is well-isolated to Gemini streaming logic, includes comprehensive mock-based tests (no network calls), doesn't modify critical paths or database queries, and follows OpenAI spec correctly. The solution is minimal and focused - only 35 lines added to track state and handle edge case. All existing tests pass.
• No files require special attention

Important Files Changed

Filename	Overview
litellm/llms/vertex_ai/gemini/vertex_and_google_ai_studio_gemini.py	Adds has_seen_tool_calls flag and final chunk handling logic to correctly map finish_reason="tool_calls" when tool calls arrive in separate chunks from finishReason
tests/test_litellm/llms/vertex_ai/gemini/test_gemini_streaming_tool_call_finish_reason.py	New test file with 4 comprehensive mock-based tests covering tool call streaming, text-only streaming, multiple tool calls, and content filtering scenarios

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[Chunk arrives in ModelResponseIterator.chunk_parser] --> B{Process candidates via _process_candidates}
    B --> C{Does candidate have 'content' field?}
    C -->|Yes| D[Extract tool_calls, text, etc]
    C -->|No| E[Skip candidate - returns empty choices]
    D --> F{Does delta have tool_calls?}
    F -->|Yes| G[Set has_seen_tool_calls = True]
    F -->|No| H[Keep flag unchanged]
    G --> I[_create_streaming_choice sets finish_reason via _check_finish_reason]
    H --> I
    E --> J{Is choices empty AND candidates exist?}
    J -->|Yes| K{Check has_seen_tool_calls flag}
    J -->|No| M[Return response]
    K -->|True| L[Create choice with finish_reason='tool_calls']
    K -->|False| N[Map finishReason normally via _check_finish_reason]
    L --> M
    N --> M

Loading

_{Last reviewed commit: 5596545}

[greptile-apps]

_{2 files reviewed, no comments}

_{Edit Code Review Agent Settings | Greptile}