[insoln]

Related issue

Fixes #7899

Summary

Anthropic can reject a historical extended-thinking block with 400 Invalid signature in thinking block. This change performs a bounded recovery after that exact validation failure instead of removing thinking eagerly from every request.

• sends normal requests unchanged;
• matches only the exact Anthropic signature-validation 400;
• removes thinking only from completed historical assistant turns;
• retries at most once;
• preserves the complete active tool-use cycle;
• completes recovery before connection cooldown and provider-breaker accounting.

Why recovery is conditional

Eager removal changes valid same-model requests, alters prompt-cache shape, and can modify Claude Code tool-use history unnecessarily. This patch leaves valid traffic untouched and changes the request only after Anthropic explicitly rejects a signature.

A trailing active cycle can contain several interleaved assistant[thinking, tool_use] → user[tool_result] pairs. Those assistant turns are preserved verbatim. If no completed historical thinking can be removed safely, no recovery retry is sent.

Recovery is not triggered for:

• generic 400 responses;
• 429 responses;
• non-Anthropic providers;
• Anthropic's separate "thinking blocks in the latest assistant message cannot be modified" error.

If the recovery attempt fails, the response continues through the normal error-classification path without another recovery attempt.

Tests

• node --import tsx/esm --test tests/unit/anthropic-thinking-signature-recovery.test.ts tests/unit/claude-passthrough-thinking-2454.test.ts
• node --import tsx/esm --test tests/unit/chatcore-translation-paths.test.ts
• npm run typecheck:core
• npm run lint

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[diegosouzapw]

Reviewed #7906 against release/v3.8.49.

This is a clean, well-scoped fix. The recovery is correctly bounded: it only fires on the exact Anthropic 400 Invalid signature in thinking block validation error (not generic 400s, not 429s, not the separate "latest message cannot be modified" error, not non-Anthropic providers), strips thinking only from completed historical assistant turns, and explicitly preserves the entire active tool_use↔tool_result cycle — including multi-hop interleaved chains, not just the latest turn. It retries at most once and completes before any connection-cooldown / provider-breaker accounting runs, so a successful recovery never pollutes resilience state with a transient, self-corrected error.

I checked out the branch and ran the new test suite plus the two adjacent suites named in the PR body:

• tests/unit/anthropic-thinking-signature-recovery.test.ts → 9/9 pass
• tests/unit/claude-passthrough-thinking-2454.test.ts + tests/unit/chatcore-translation-paths.test.ts → 68/68 pass, no regressions

The chatCore.ts wiring (providerFailure: … break providerFailure) mirrors the file's existing T5 intra-family-model-fallback retry pattern closely, which gives me good confidence it's correct even though there's no dedicated handleChatCore()-level integration test for this exact path — the extracted logic (isAnthropicThinkingSignatureError, stripHistoricalThinkingForSignatureRecovery, executeWithAnthropicThinkingSignatureRecovery) is thoroughly unit-tested instead, including the safety case where only active-cycle thinking exists and no retry is sent.

Two small optional suggestions, neither blocking:

1. isAnthropicThinkingSignatureError hardcodes "anthropic-compatible-" — could reuse the existing ANTHROPIC_COMPATIBLE_PREFIX constant from src/shared/constants/providers.ts.
2. A handleChatCore()-level integration test (mock fetch 400→200) would directly cover the chatCore.ts wiring itself, on top of the strong unit coverage already present.

CI is fully green and the PR is mergeable against release/v3.8.49. Nice work — thanks for both filing #7899 and fixing it with this level of care.

[diegosouzapw]

Merged — thanks! CI green on the current tip. (Merged on real GitHub CI; a parallel monaco 0.56 bump had left the local validation box with stale node_modules, unrelated to this PR.)