[cikowski]

The Gemini agent's chat-completions bridge, used for provider-prefixed models (vertex_ai/...), reads each tool call's thought signature off the LiteLLM response and hands it to the CLI as thoughtSignature, but drops it when it turns the CLI's history back into OpenAI messages. Gemini 3 and newer require the signature on the next request, so Vertex answers Function call is missing a thought_signature in functionCall parts and the turn fails after its first tool call. LiteLLM hides this only for model names its _is_gemini_3_or_newer recognizes, where it fills in a placeholder signature; for any other Gemini 3+ name the call goes out unsigned and fails every time.

This change carries each part's thoughtSignature back as the tool call's provider_specific_fields.thought_signature, the field LiteLLM reads first on input. Because every signature now goes back to Vertex, the response side also stops copying a message-level signature onto every call of a parallel batch: Gemini signs only the first call, and LiteLLM's Vertex transformation notes that Gemini 3+ bills each replayed copy as the previous turn's reasoning tokens (litellm/llms/vertex_ai/gemini/transformation.py).

Rollout

• Changes only the request the bridge sends upstream, and which parallel call carries a message-level signature in the response it hands the CLI. A history with no signatures produces the same request as before. The native /gemini passthrough (bare model names) is untouched.
• Ships in the next agentenv-gemini image (src-<hex> changes with this folder). Anyone running a registered copy of the agent needs to register the new image to pick it up. Until then, a bare model name avoids the bridge.
• Revert: revert this commit. No data or config migration.

Validation

Head c6bfb22 (same tree as the reviewed 5cb0252).

Done when	Target	Status	Evidence	Limits
Every tool call's signature reaches the next upstream request, and calls without one stay unsigned	the /litellm-gemini proxy, two tool turns end to end	MET	test_each_tool_calls_thought_signature_survives_the_round_trip fails on main (no signatures sent), passes on head	upstream LiteLLM is a mock transport
A message-level signature goes only to the first parallel call	the proxy's response to the CLI	MET	test_a_message_level_signature_goes_only_to_the_first_parallel_call fails on main, passes on head	mock transport
Vertex accepts the round-tripped signatures	a Gemini 3+ model through a live LiteLLM proxy	MET	6-turn tool loop through the bridge functions (1, 3 parallel, 1, 2 parallel, 1 calls, then a final answer): every request 200 on head	that proxy fills placeholder signatures, so main passes there too: shows acceptance, not the fix
The 400 no longer happens on a model LiteLLM does not recognize	live run	UNPROVEN		needs the released image

• pytest tst/unit/agents/gemini: 146 passed, 6 skipped. Mutants that drop the signature, attach it unconditionally, or keep only the first turn's fail the suite.
• make unit-test: 11896 passed, 66 skipped, 2 failed, neither in code this PR touches: tst/unit/agents/hermes/test_hermes_bridge.py::test_the_event_pipe_is_not_left_open_in_the_server fails the same way on a clean origin/main locally (macOS), and tst/unit/cli/submit_runs_test.py::test_a_first_ctrl_c_lets_the_runner_finish_stopping passes 3 of 3 alone and passed in an earlier full run (timing under the parallel run).
• ruff check .: clean. Not run: integration tiers (nothing outside the Gemini agent folder changed).

Not in this PR

• Thought signatures on text parts are still dropped in both directions. Gemini treats them as optional, so it costs quality rather than a 400.

🤖 Generated with Claude Code

Confidence Score: 5/5

The PR appears safe to merge; no actionable issues were found.

Summary

The Gemini bridge now sends tool-call thought signatures back to LiteLLM.

• Tool-call signatures travel back to LiteLLM with their calls.
• Parallel replies keep a message-level signature on the first call.

No actionable issues were found. Tests were inspected, not run during this review.

Acknowledged scope: cikowski states that signatures on text parts remain dropped because they are optional and outside this fix.

Diagram

sequenceDiagram
    participant CLI as Gemini CLI
    participant Bridge as Gemini bridge
    participant LiteLLM
    LiteLLM-->>Bridge: Tool calls with thought signatures
    Bridge-->>CLI: Function-call parts with thoughtSignature
    CLI->>Bridge: History containing those parts
    Bridge->>LiteLLM: Tool calls with provider_specific_fields.thought_signature

Loading

_{Reviews (1) · Last reviewed commit: "fix(agents): send Gemini thought signatu..." · Reviewed by Greptile}