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

Project	Deployment	Actions	Updated (UTC)
litellm	Ready	Preview, Comment	Mar 21, 2026 10:04pm

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, you can upgrade your account or add credits to your account and enable them for code reviews in your settings.

Greptile Summary

This PR introduces a reasoning_auto_summary opt-in flag that controls whether "summary": "detailed" is automatically injected when translating Anthropic thinking parameters to OpenAI reasoning_effort for non-Claude model routing. It also fixes Gemini/Vertex AI usage token counting to accumulate across multiple modality entries instead of overwriting, adds case-insensitive modality normalization, and refactors translate_anthropic_to_openai into focused helper methods.

Key changes:

• New litellm/llms/anthropic/experimental_pass_through/utils.py with is_reasoning_auto_summary_enabled() helper reading from litellm.reasoning_auto_summary (defaults to False) or LITELLM_REASONING_AUTO_SUMMARY env var
• translate_thinking_for_model, _translate_thinking_to_openai, translate_thinking_to_reasoning, and _route_openai_thinking_to_responses_api_if_needed all updated consistently to use the new flag
• Gemini promptTokensDetails, responseTokensDetails, candidatesTokensDetails, and cacheTokensDetails parsing now accumulates += instead of overwriting =, fixing double-entry modality responses
• translate_anthropic_to_openai decomposed into _translate_metadata_to_openai, _translate_tool_choice_to_openai, _translate_tools_to_openai, _translate_thinking_to_openai, _translate_output_format_to_openai, and _copy_untranslated_anthropic_params

Concerns:

• The change from always-on summary: detailed injection to opt-in False default is backwards-incompatible for users routing Anthropic thinking requests through non-Claude models: they will silently stop receiving reasoning content in responses until they set reasoning_auto_summary: true. Per the project policy, the safer default would be True with an opt-out path.
• "litellm_metadata" is excluded from translatable_anthropic_params() but is removed via .pop() inside _translate_metadata_to_openai. This implicit ordering dependency between private helpers is fragile if the methods are ever called independently.

Confidence Score: 3/5

• Merging as-is will silently break existing users routing Anthropic thinking requests to non-Claude models, who will stop receiving reasoning content by default.
• The Gemini token accumulation fix and code refactoring are solid and well-tested. However, flipping the reasoning_auto_summary default from effectively True to False is a backwards-incompatible behavioral change — existing users relying on automatic summary: detailed injection will silently lose reasoning content in their responses without any migration guidance in the PR description or changelog entry. Additionally, the implicit ordering dependency created by the litellm_metadata pop in _translate_metadata_to_openai introduces fragility in the new helper-method structure.
• litellm/llms/anthropic/experimental_pass_through/utils.py and litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py need attention due to the backwards-incompatible default change and the litellm_metadata ordering dependency.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["Anthropic /v1/messages with thinking"] --> B{Is Claude model?}
    B -- Yes --> C["Pass thinking as-is to litellm.completion"]
    B -- No --> D["translate_anthropic_thinking_to_reasoning_effort\nbudget_tokens maps to effort string"]
    D --> E{User-provided summary in thinking?}
    E -- Yes --> F["reasoning_effort = effort plus summary"]
    E -- No --> G{is_reasoning_auto_summary_enabled}
    G -- True --> H["reasoning_effort = effort plus summary:detailed"]
    G -- False --> I["reasoning_effort = plain effort string\nNo summary - OpenAI skips reasoning text"]
    F --> J["litellm.completion or litellm.responses"]
    H --> J
    I --> J
    J --> K["OpenAI or non-Claude model"]

_{Last reviewed commit: "fix: apply Black for..."}

Merging this PR will not alter performance

✅ 16 untouched benchmarks

_{Comparing litellm_oss_staging_03_05_2026 (e3b62c0) with main (2b889f1)}

Codecov Report

❌ Patch coverage is 39.72603% with 88 lines in your changes missing coverage. Please review.

📢 Thoughts on this report? Let us know!