Fix: add OpenAI-compatible API for Anthropic with modify_params=True

[Chesars]

Relevant issues

Related to #14194 and #9020

Pre-Submission checklist

Please complete all items before asking a LiteLLM maintainer to review your PR

• I have Added testing in the tests/litellm/ directory, Adding at least 1 test is a hard
  requirement - see details - N/A (documentation only)
• I have added a screenshot of my new test passing locally - N/A (documentation only)
• My PR passes all unit tests on make test-unit - N/A (documentation only)
• My PR's scope is as isolated as possible, it only solves 1 specific problem

Type

Documentation
Bug Fix

Changes

Summary

• Adds handling for OpenAI-Anthropic API incompatibility when using extended thinking with tool calling
• When modify_params=True, LiteLLM drops the thinking param if assistant messages with tool_calls are missing thinking_blocks
• Updates documentation explaining the architectural differences between OpenAI (stateful) and Anthropic (stateless) APIs

Details

OpenAI-compatible clients (LibreChat, Open WebUI, Vercel AI SDK, etc.) don't preserve thinking_blocks because it's not in the OpenAI spec. This causes Anthropic to reject requests with:
Expected thinking or redacted_thinking, but found tool_use

Solution

When modify_params=True:

1. Detect if thinking is enabled but thinking_blocks are missing from the last assistant message with tool_calls
2. Drop the thinking param for that turn
3. Log a warning so users are aware

[vercel]

@Chesars is attempting to deploy a commit to the CLERKIEAI Team on Vercel.

A member of the Team first needs to authorize it.

[ghost]

@Chesars for clients with litellm.modify_params enabled - can we not just drop the thinking flag, if the thinking_blocks are missing?

this would prevent a 400 right?

[Chesars]

Yes, that would prevent the 400, but the model would stop using extended thinking on that turn.

[ghost]

if litellm.modify_params is True, and the alternative is that it would fail the request - then it seems correct to handle the scenario to avoid failing the call

if the flag is not enabled, then we should just bubble up the error, and let them know this can be avoided with the flag.

this allows the anthropic model to work via the proxy when being called via openai clients - thoughts?

[Chesars]

Makes sense. For the proxy when modify_params=True, the fix would detect if the last assistant message has tool_calls but no thinking_blocks, drop the thinking param from the request. Typically on the tool result turn.