feat(anthropic): retry /v1/messages after invalid thinking signature

[Sameerlite]

Summary

When Anthropic returns an invalid encrypted thinking signature error (common after switching API keys, model deployment, or provider), LiteLLM now strips thinking / redacted_thinking blocks and the top-level thinking param from the outbound Messages API body and retries the HTTP request once.

Implementation

• Helpers in common_utils.py: detect the error shape and strip thinking content from the request dict.
• BaseAnthropicMessagesConfig: hooks for max attempts, should-retry, and transform (lazy imports avoid circular imports).
• async_anthropic_messages_handler: retry loop with re-sign_request for Bedrock.

Coverage

Applies to any provider using BaseAnthropicMessagesConfig (direct Anthropic, Bedrock Claude messages invoke path, Vertex partner Claude, Azure AI Anthropic).

Tests

tests/test_litellm/llms/anthropic/test_anthropic_common_utils.py::TestAnthropicThinkingSignatureSelfHeal

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
litellm	Ready	Preview, Comment	Apr 14, 2026 4:34am

[codspeed-hq]

Merging this PR will not alter performance

✅ 16 untouched benchmarks

---

_{Comparing litellm_anthropic-messages-thinking-signature-retry (5670f6c) with main (e64d98f)}

[greptile-apps]

Greptile Summary

This PR adds a self-healing retry mechanism for the Anthropic /v1/messages path: when a 400 "invalid thinking signature" error is received (common after rotating API keys or switching model endpoints), LiteLLM now strips thinking/redacted_thinking blocks from the request history and re-signs and retries once. The previous threads' concerns about content: [] and the missing status-code guard have both been resolved in this version.

Confidence Score: 5/5

• Safe to merge; all remaining findings are non-blocking P2 style suggestions.
• The two P1 issues from prior review rounds (empty content arrays, missing status-code guard) are both fully addressed. The three remaining comments are minor — a redundant error check, two identical dict copies, and an edge case where a thinking-only assistant turn drops and leaves consecutive same-role user messages. None of these block correct operation for the common case.
• No files require special attention.

Important Files Changed

Filename	Overview
litellm/llms/anthropic/common_utils.py	Adds three helpers: error detector, message stripper (correctly drops messages that become empty after stripping), and request dict mutator. Logic is sound; the message-drop behavior when ALL messages contain only thinking blocks is documented but can leave messages: [].
litellm/llms/base_llm/anthropic_messages/transformation.py	Adds three hook methods to BaseAnthropicMessagesConfig: max attempts property (2), status-code-guarded should_retry, and transform_anthropic_messages_request_on_http_error. The transform method redundantly re-checks the error type that should_retry already confirmed.
litellm/llms/custom_httpx/llm_http_handler.py	Extracts the HTTP POST into a new private retry helper _async_post_anthropic_messages_with_http_error_retry. Loop structure, error handling, re-signing on retry, and the unreachable sentinel RuntimeError are all correct. Two identical dict(litellm_params) copies are created unnecessarily.
tests/test_litellm/llms/anthropic/test_anthropic_common_utils.py	Six new pure-mock unit tests covering the detector, stripper, and config hooks — no real network calls. Tests confirm the messages == [] edge case but don't cover the retry loop itself or the consecutive same-role message scenario.

Sequence Diagram

sequenceDiagram
    participant Caller as async_anthropic_messages_handler
    participant Retry as _async_post_anthropic_messages_with_http_error_retry
    participant HTTP as AsyncHTTPHandler
    participant Config as BaseAnthropicMessagesConfig

    Caller->>Retry: POST /v1/messages (with thinking blocks + signed body)
    Retry->>HTTP: post(url, headers, signed_json_body or json body, stream)
    HTTP-->>Retry: 400 HTTPStatusError (invalid thinking signature)
    Retry->>Config: should_retry_anthropic_messages_on_http_error(e, litellm_params)
    Config-->>Retry: "True (status==400 and error matches)"
    Retry->>Config: transform_anthropic_messages_request_on_http_error(e, request_body)
    Note over Config: strips thinking/redacted_thinking blocks<br/>removes top-level "thinking" param
    Config-->>Retry: mutated request_body
    Retry->>Config: sign_request(headers, optional_params, request_body, ...)
    Config-->>Retry: new headers + signed_json_body
    Retry->>HTTP: post(url, new_headers, new_signed_body, stream)
    HTTP-->>Retry: 200 OK response
    Retry-->>Caller: httpx.Response

Loading

_{Reviews (2): Last reviewed commit: "fix(anthropic): tighten thinking-signatu..." | Re-trigger Greptile}