feat(anthropic): add opt-out flag for default reasoning summary

[Chesars]

Relevant issues

Allows users to disable the automatic "summary": "detailed" injection when LiteLLM translates Anthropic thinking parameters to OpenAI reasoning_effort.

Pre-Submission checklist

• I have Added testing in the tests/test_litellm/ directory
• My PR passes all unit tests on make test-unit
• My PR's scope is as isolated as possible, it only solves 1 specific problem
• I have requested a Greptile review by commenting @greptileai

Type

🆕 New Feature

Changes

• Add litellm.disable_default_reasoning_summary flag (default False) in litellm/__init__.py
• Support LITELLM_DISABLE_DEFAULT_REASONING_SUMMARY env var as alternative
• Update adapters/handler.py to skip injecting "summary": "detailed" when flag is active (both string and dict reasoning_effort)
• Update responses_adapters/transformation.py to respect the flag in translate_thinking_to_reasoning
• Add 6 unit tests covering: flag on/off, env var, user-provided summary preservation

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
litellm	Error		Mar 5, 2026 2:59pm

[greptile-apps]

Greptile Summary

This PR adds a litellm.disable_default_reasoning_summary flag (and LITELLM_DISABLE_DEFAULT_REASONING_SUMMARY env var) to let users suppress the automatic injection of "summary": "detailed" when Anthropic thinking parameters are translated to OpenAI reasoning_effort. The shared helper is_default_reasoning_summary_disabled() is correctly extracted into utils.py and the flag is applied in two of the three conversion code paths.

Critical issues found:

• Backwards-incompatible default: The flag defaults to False, which means summary: "detailed" is now injected for every request that previously produced no summary. This silently changes the API payload sent downstream for all existing users of the experimental pass-through. Following the project's backwards-compatibility guidelines, this behavioral change should be opt-in — either default the flag to True, or rename it to enable_default_reasoning_summary defaulting to False.
• translate_thinking_for_model not updated: adapters/transformation.py lines 692–700 is a third code path that converts thinking → reasoning_effort, but it was not modified to inject summary: "detailed" by default. This creates inconsistent behavior depending on which code path a request follows.

Confidence Score: 1/5

• Not safe to merge: breaks backwards compatibility for all existing users by changing default API payloads, and one code path is not updated creating inconsistent behavior.
• Two blocking issues prevent merging: (1) The flag defaults to False, which silently injects summary: "detailed" into requests that previously never included it. This breaks existing users without their consent and violates the backwards-compatibility rule. (2) The function translate_thinking_for_model in adapters/transformation.py was not updated to match the new default behavior, creating inconsistent outputs depending on which code path handles a request. Both issues require code changes to resolve.
• litellm/init.py (flag default must change), litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py (function not updated)

Comments Outside Diff (1)

1. General comment

   translate_thinking_for_model not updated — inconsistent default behavior

   translate_thinking_for_model is a third code path that converts Anthropic thinking to an OpenAI reasoning_effort value, but it does not check the disable_default_reasoning_summary flag. It only includes summary when the user explicitly provides one, unlike the two updated paths (_route_openai_thinking_to_responses_api_if_needed and translate_thinking_to_reasoning) which inject summary: "detailed" by default when the flag is False.

   This creates inconsistent behavior depending on which code path a request follows. To align the three code paths, translate_thinking_for_model should also consult is_default_reasoning_summary_disabled():

   This requires importing the helper at the top of the file and updating the corresponding test at line 224 to expect {"reasoning_effort": {"effort": "minimal", "summary": "detailed"}} by default.

_{Last reviewed commit: 3e9ea6f}

[greptile-apps]

Backwards-incompatible default behavior change

Setting disable_default_reasoning_summary to False (the default) means summary: "detailed" is now injected into every request that previously received no summary. Before this PR, translate_thinking_to_reasoning({"type": "enabled", "budget_tokens": 5000}) returned {"effort": "medium"}; after this PR it returns {"effort": "medium", "summary": "detailed"} by default, silently altering all existing callers' API payloads.

Per the repository guidelines on backwards compatibility (rule b48b7341), behavioral changes like this should be opt-in, not opt-out — meaning the old behavior should remain the default and users explicitly set the flag to enable the new behavior. The flag should either:

Option 1: Default to True to preserve old behavior:

Suggested change

	disable_default_reasoning_summary: bool = False
	disable_default_reasoning_summary: bool = True

Option 2: Rename to make the opt-in nature explicit:

enable_default_reasoning_summary: bool = False

and update the helper in utils.py accordingly.

Context Used: Rule from dashboard - What: avoid backwards-incompatible changes without user-controlled flags

Why: This breaks current ... (source)