fix(anthropic): fix clear_thinking retry body after thinking strip

[mchtech]

Summary

• restore top-level thinking when invalid-signature retry keeps clear_thinking_20251015
• keep the fix scoped to Anthropic Messages passthrough retry handling
• avoid carrying manual budget_tokens into adaptive thinking retries
• add behavior coverage for enabled, adaptive, disabled, missing, and non-adaptive cases

Fixes #27512

Related: #25867 added the same clear_thinking_20251015 request-contract guard for Bedrock. This PR covers the Anthropic Messages passthrough retry path, where the retry cleanup can remove top-level thinking while leaving clear_thinking_20251015 in place.

Tests

.venv/bin/python -m pytest tests/test_litellm/llms/anthropic/test_anthropic_common_utils.py -q

Result: 72 passed

[CLAassistant]

All committers have signed the CLA.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[greptile-apps]

Greptile Summary

This PR overrides transform_anthropic_messages_request_on_http_error in AnthropicMessagesConfig to restore the caller's original top-level thinking parameter after the base class strips it during an invalid-signature retry, but only when a clear_thinking_20251015 context-management edit is present and the target model supports adaptive thinking.

• The new method captures thinking before the base call, then conditionally re-adds it post-strip, preserving budget_tokens for enabled mode and dropping it for adaptive mode (server-controlled).
• Five mock-only test cases are added covering the enabled, adaptive, disabled, missing, no-clear_thinking, and non-adaptive-model paths; all make no real network calls, satisfying the test-isolation rule.

Confidence Score: 4/5

The change is narrow and well-guarded: the restore only fires for adaptive-thinking models with a clear_thinking edit present, and the base-class strip still runs unconditionally before any restore attempt.

The logic correctly handles all edge cases tested (enabled, adaptive without budget, disabled, missing, no clear-thinking edit, non-adaptive model). The not has_active_thinking guard will always be True in the primary retry path because the base class always removes thinking for invalid-signature errors, making it a purely defensive check. The call site discards the return value and relies on in-place mutation, which is consistent with the base class but is an implicit contract.

No files require special attention; the transformation file is the only meaningful change and its scope is tightly controlled.

Important Files Changed

Filename	Overview
litellm/llms/anthropic/experimental_pass_through/messages/transformation.py	Adds transform_anthropic_messages_request_on_http_error override that restores top-level thinking config after the base class strips it, restricted to adaptive-thinking models with a clear_thinking edit present.
tests/test_litellm/llms/anthropic/test_anthropic_common_utils.py	Adds five new mock-only tests and a shared helper covering all branch outcomes of the new retry transform: enabled/adaptive restore, disabled/missing no-restore, no-clear-thinking no-restore, and non-adaptive-model no-restore.

_{Reviews (1): Last reviewed commit: "fix(anthropic): fix clear_thinking retry..." | Re-trigger Greptile}

[krrish-berri-2]

@mchtech — could you add a screenshot or short video showing that this change works as expected? It really helps reviewers verify the fix quickly. Thanks!

[mchtech]

[github-actions]

This pull request has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs.