fix(gemini): deduplicate reasoning parts when assistant message has both reasoning_content and thinking_blocks

[Priyanka06081218]

Fixes #37973

TLDR

When a Gemini multi-turn conversation replays an assistant message that carries both reasoning_content and thinking_blocks, LiteLLM was sending the same reasoning text to Gemini twice. This PR makes the two branches mutually exclusive.

What happened

_gemini_convert_messages_with_history had two independent if blocks:

if reasoning_content is not None:
    assistant_content.append(PartType(thought=True, text=reasoning_content))
if thinking_blocks is not None:
    for block in thinking_blocks:
        ...
        assistant_content.append(PartType(thoughtSignature=block_signature, text=block_thinking_str))

Both ran when both fields were set. LiteLLM's own Gemini response transformation sets both fields from the same parts, so any plain multi-turn Gemini conversation hit this as soon as the model returned a thoughtSignature. Gemini billed the replayed reasoning twice.

This is the pre-existing bug explicitly listed in the Caveats of #37953:

"Gemini sends a prior turn's reasoning twice when a message carries both reasoning_content and a signed thinking block ... replaying an assistant turn with both fields produces the same three parts on this branch and on the merge base."

Fix

Changed if thinking_blocks / if reasoning_content to if thinking_blocks / elif reasoning_content. Signed blocks take priority — they are lossless (they carry the full reasoning text plus the signature). reasoning_content is emitted only when thinking_blocks is absent.

if thinking_blocks is not None:
    for block in thinking_blocks:
        ...
elif reasoning_content is not None:
    assistant_content.append(PartType(thought=True, text=reasoning_content))

Tests

Added two tests to tests/llm_translation/test_prompt_factory.py:

• test_gemini_no_duplicate_reasoning_parts_when_both_fields_present — asserts exactly one signed part is emitted and no plain thought part appears when both fields are set
• test_gemini_reasoning_content_fallback_when_no_thinking_blocks — asserts the fallback path still works correctly when only reasoning_content is present

[greptile-apps]

Greptile Summary

This PR prevents duplicate Gemini reasoning parts by preferring signed thinking blocks over the plain reasoning fallback.

• Makes signed thinking-block and reasoning_content serialization mutually exclusive.
• Adds regression coverage for messages containing both representations and for the reasoning-only fallback.

Confidence Score: 4/5

The PR needs the fallback condition corrected before merging because present-but-unusable thinking blocks can now discard valid reasoning content.

The new null-based branch suppresses reasoning_content even when the thinking-block loop emits nothing, causing assistant reasoning context to be lost during replay.

Files Needing Attention: litellm/llms/vertex_ai/gemini/transformation.py

Important Files Changed

Filename	Overview
litellm/llms/vertex_ai/gemini/transformation.py	Deduplicates reasoning serialization, but suppresses the fallback when the provided thinking-block list contains no replayable signed block.
tests/llm_translation/test_prompt_factory.py	Covers the signed-block and absent-block cases but omits empty or unusable thinking-block collections.

_{Reviews (1): Last reviewed commit: "fix(gemini): prefer signed thinking_bloc..." | Re-trigger Greptile}

[greptile-apps]

Unusable blocks suppress reasoning fallback

When thinking_blocks is empty or contains no block with both thinking text and a signature, the loop emits nothing but the non-null check still skips reasoning_content, causing valid reasoning context to be discarded during replay.

Knowledge Base Used: Provider adapters and capabilities

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[codspeed-hq]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing Priyanka06081218:fix/gemini-reasoning-duplicate-parts (1676c34) with litellm_internal_staging (aae36f4)}

[Priyanka06081218]

@mateo-berri — This fixes the Gemini deduplication bug you listed as a known remaining caveat in #37953. The change is one file (litellm/llms/vertex_ai/gemini/transformation.py), three targeted tests, and CI is green across all required checks. Would appreciate a review when you get a chance.

[Priyanka06081218]

@tin-berri, @yucheng-berri, @mateo-berri — This fixes the Gemini deduplication bug you listed as a known remaining caveat in #37953. The change is one file (litellm/llms/vertex_ai/gemini/transformation.py), three targeted tests, and CI is green across all required checks. Would appreciate a review when you get a chance.