fix(vertex/gemini): keep tool_call reference when a text-only assistant message follows

[1qh]

Title

fix(vertex/gemini): keep tool_call reference when a text-only assistant message follows a tool call

Relevant issues

Fixes #26965
Fixes #29560
Related: #28978, #22578 (same "Missing corresponding tool call" failure mode on the Gemini/Responses path)

Pre-Submission checklist

• I have Added testing in the tests/litellm/ directory (tests/test_litellm/llms/vertex_ai/gemini/test_tool_call_followed_by_text_assistant.py)
• My PR passes all unit tests on make test-unit
• My PR's scope is as isolated as possible (one branch, one fix)

Type

🐛 Bug Fix

Changes

Converting an OpenAI-style history to Gemini failed with "Missing corresponding tool call for tool response message" when an assistant message containing tool_calls was followed by a text-only assistant message before the tool result. This is a common shape — a model emits a short narration/"thinking out loud" turn after a tool call, before the tool result is appended (reported independently in #26965, #29560).

Root cause

In _gemini_convert_messages_with_history (litellm/llms/vertex_ai/gemini/transformation.py), last_message_with_tool_calls is assigned inside a branch guarded by:

assistant_msg.get("tool_calls", []) is not None

An empty list is not None, so this branch is entered for text-only assistant messages too, and the reference to the real tool-call message gets clobbered. The following tool result is then matched against a message with no tool_calls, so name lookup fails and conversion raises.

Fix

Only update last_message_with_tool_calls when the assistant message actually carries tool_calls (or a function_call). Minimal and behavior-preserving for the normal case.

if (
    assistant_msg.get("tool_calls")
    or assistant_msg.get("function_call") is not None
):
    last_message_with_tool_calls = assistant_msg

Test

tests/test_litellm/llms/vertex_ai/gemini/test_tool_call_followed_by_text_assistant.py builds the minimal [user, assistant+tool_call, assistant text, tool] history and asserts conversion succeeds and produces the correct functionResponse. It fails on main (raises) and passes with this change.

from litellm.llms.vertex_ai.gemini.transformation import _gemini_convert_messages_with_history

messages = [
    {"role": "user", "content": "list the files"},
    {"role": "assistant", "content": None,
     "tool_calls": [{"id": "call_abc123", "type": "function",
                     "function": {"name": "shell", "arguments": '{"command": ["ls"]}'}}]},
    {"role": "assistant", "content": "Running the command now."},
    {"role": "tool", "tool_call_id": "call_abc123", "content": "math.py"},
]
_gemini_convert_messages_with_history(messages=messages)  # raised before; succeeds now

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[codspeed-hq]

Congrats! CodSpeed is installed 🎉

🆕 16 new benchmarks were detected.

You will start to see performance impacts in the reports once the benchmarks are run from your default branch.

Detected benchmarks

• test_get_model_info_with_provider: 79.7 µs
• test_get_model_info_openai: 80.5 µs
• test_get_llm_provider_anthropic: 142.2 µs
• test_token_counter_long_content: 1.7 ms
• test_token_counter_with_tools: 412.8 µs
• test_token_counter_multi_turn: 567.6 µs
• test_get_llm_provider_with_prefix: 136.7 µs
• test_cost_per_token_anthropic: 557.4 µs
• test_cost_per_token_openai: 557.7 µs
• test_get_llm_provider_azure: 145 µs
• test_get_llm_provider_openai: 138.2 µs
• test_token_counter_raw_text: 181.8 µs
• test_get_model_cost_key_case_insensitive: 81.9 µs
• test_get_model_info_anthropic: 78.8 µs
• test_token_counter_simple_message: 233.7 µs
• test_get_model_cost_key_exact_match: 79.9 µs

---

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[veria-ai]

PR overview

All previously flagged issues have been addressed. No open security concerns remain on this pull request.

Security review

No open security issues remain on this pull request.

Fixed/addressed: 1 · PR risk: 0/10

[1qh]

Note: the only non-passing check is misc / Run tests (Unit Tests: MCP, Secrets, Containers & Misc), which failed with The job has exceeded the maximum execution time of 20m0s → operation canceled — an unrelated CI timeout, not a test failure. This PR only touches litellm/llms/vertex_ai/gemini/transformation.py and its new test; the relevant suites (Vertex AI, core-utils, lint, code-quality) pass. A re-run of that job should clear it.

[Sameerlite]

@greptileai

[greptile-apps]

Greptile Summary

Fixes a bug in _gemini_convert_messages_with_history where a text-only assistant message following a tool-call assistant message would overwrite last_message_with_tool_calls, causing the subsequent tool result to fail with "Missing corresponding tool call for tool response message". The fix adds a targeted inner guard so the reference is only updated when the message actually carries tool calls.

• transformation.py: Inner if check added to skip updating last_message_with_tool_calls for text-only assistant messages, preserving the reference to the real tool-call message across narration/thinking turns.
• New test: test_tool_call_followed_by_text_assistant.py builds the minimal failing message shape and asserts conversion succeeds with the correct functionResponse name, serving as a regression guard.

Confidence Score: 5/5

Safe to merge; the change is minimal, precisely targeted, and accompanied by a regression test.

The fix is a two-line inner guard that only narrows the condition under which a reference is updated — it cannot affect any path where an assistant message genuinely carries tool calls. The new test directly reproduces the failure shape and verifies the correct output. No other files are touched and there are no side effects to the broader conversion logic.

No files require special attention.

Important Files Changed

Filename	Overview
litellm/llms/vertex_ai/gemini/transformation.py	Adds an inner guard so last_message_with_tool_calls is only updated when the assistant message actually carries tool_calls or function_call, preventing a text-only follow-up assistant turn from clobbering the reference.
tests/test_litellm/llms/vertex_ai/gemini/test_tool_call_followed_by_text_assistant.py	New regression test covering the exact failure shape (user → assistant+tool_call → assistant text → tool result); no network calls, correctly placed in the mock-only test directory.

_{Reviews (1): Last reviewed commit: "fix(vertex/gemini): keep tool_call refer..." | Re-trigger Greptile}

[Sameerlite]

LGTM