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

Project	Deployment	Actions	Updated (UTC)
litellm	Error		Mar 5, 2026 3:23pm

Greptile Summary

This PR aligns translate_thinking_for_model with the default summary="detailed" injection behavior introduced in #22904, and adds documentation and tests for the opt-out flag.

• Core change: translate_thinking_for_model in adapters/transformation.py now injects {"effort": ..., "summary": "detailed"} by default when no user-provided summary is present and disable_default_reasoning_summary is not set, changing the return shape from {"reasoning_effort": "<string>"} to an object.
• Flag wiring: Correctly reuses the existing is_default_reasoning_summary_disabled() helper from experimental_pass_through/utils.py, consistent with the other two call sites in adapters/handler.py and responses_adapters/transformation.py.
• Tests: Two new mock-only tests added verifying both the default behavior (with summary injection) and the disabled-flag behavior (plain string return); existing assertions updated appropriately; global state is properly restored via try/finally.
• Docs: New section in reasoning_content.md documents the SDK flag, env var, and proxy config opt-out options clearly, including a note that user-provided summaries are always preserved.
• Backwards-compatibility concern: The default return shape change from {"reasoning_effort": "<string>"} to {"reasoning_effort": {"effort": "<string>", "summary": "detailed"}} is opt-out rather than opt-in. Downstream consumers expecting a string value may silently break unless they set the flag. However, the pattern aligns with changes already accepted in feat(anthropic): add opt-out flag for default reasoning summary #22904, and the new documentation makes the opt-out flag discoverable.

Confidence Score: 4/5

• Safe to merge with awareness of the opt-out backwards-compatibility pattern. The change is consistent with feat(anthropic): add opt-out flag for default reasoning summary #22904, tests are proper, and documentation is clear.
• The logic is correct, tests are proper mock-only unit tests with appropriate state cleanup, and the change is consistent with the two other code paths updated in feat(anthropic): add opt-out flag for default reasoning summary #22904. The primary concern is that the default behavior change from {"reasoning_effort": "<string>"} to {"reasoning_effort": {"effort": "<string>", "summary": "detailed"}} is opt-out rather than opt-in, which could silently affect downstream consumers expecting a plain string. However, this is mitigated by (1) clear documentation of the opt-out flag, (2) the pattern having been established and accepted in feat(anthropic): add opt-out flag for default reasoning summary #22904, and (3) a straightforward disable mechanism. This is a deliberate design choice aligned with the earlier PR rather than an oversight.
• litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py — the default return shape change may impact downstream consumers that expect reasoning_effort to be a plain string. The disable_default_reasoning_summary flag provides an opt-out, but existing users must discover and use it to avoid the new behavior.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[translate_thinking_for_model] --> B{is_anthropic_claude_model?}
    B -- Yes --> C[Return thinking param as-is]
    B -- No --> D[translate_anthropic_thinking_to_reasoning_effort]
    D --> E{reasoning_effort resolved?}
    E -- No --> F[Return empty dict]
    E -- Yes --> G{User provided summary<br/>in thinking dict?}
    G -- Yes --> H["Return {effort, summary: user_value}"]
    G -- No --> I{is_default_reasoning_summary_disabled?<br/>litellm flag OR env var}
    I -- True --> J["Return {reasoning_effort: string}"]
    I -- False --> K["Return {effort, summary: 'detailed'}"]

_{Last reviewed commit: 5b904f6}