PR #70437: fix(agents): preserve Anthropic thinking signatures on replay (#69579)

• Repository: openclaw/openclaw
• Author: ly85206559
• State: open | merged: False
• Link: https://github.com/openclaw/openclaw/pull/70437

Description (problem / solution / changelog)

Summary

• Problem: Multi-turn requests that replay assistant thinking blocks (extended thinking) failed with Invalid signature in thinking block because the transport layer ran sanitizeTransportPayloadText on stored thinking bytes, which can change character encoding (e.g. lone surrogates) and invalidates Anthropic’s cryptographic signature.
• Why it matters: Any channel (e.g. Telegram) that keeps conversation history and replays the prior turn’s assistant content hits this on turn 2+; the failure can look like a “silent” second reply.
• What changed: When a stored thinking block is signed, replay thinking verbatim; resolve signature via shared resolveAnthropicStoredThinkingSignature from thinkingSignature (if non-empty after trim) or signature. Use the same helper in transformTransportMessages and convertAnthropicMessages. Add regression tests (lone-surrogate, signature-only) plus a small unit test for the resolver (including thinkingSignature: "" falling back to signature).
• What did NOT change (scope boundary): No Telegram-specific code paths; no API contract or gateway protocol change; no CHANGELOG entry in this PR.

Change Type (select all)

• Bug fix
• Feature
• Refactor required for the fix
• Docs
• Security hardening
• Chore/infra

Scope (select all touched areas)

• Gateway / orchestration
• Skills / tool execution
• Auth / tokens
• Integrations
• API / contracts
• UI / DX
• CI/CD / infra

Linked Issue/PR

• Closes #69579
• Related
• This PR fixes a bug or regression

Root Cause (if applicable)

• Root cause: convertAnthropicMessages in src/agents/anthropic-transport-stream.ts always sanitized block.thinking in one branch; for signed thinking, bytes must match the server-issued signature exactly. Sanitization changed those bytes. The transform layer needed the same “has signature” resolution for both thinkingSignature and signature.
• Missing detection / guardrail: No unit test asserting that stored signed thinking is not passed through sanitizeTransportPayloadText, and no shared resolver between transform and stream paths.
• Contributing context (if known): Reported on Telegram + Opus extended thinking; the bug is in shared Anthropic transport, not Telegram.

Regression Test Plan (if applicable)

• Coverage level that should have caught this:
  • Unit test
  • Seam / integration test
  • End-to-end test
  • Existing coverage already sufficient
• Target test or file: src/agents/anthropic-transport-stream.test.ts, src/agents/anthropic-stored-thinking-signature.test.ts
• Scenario the test should lock in: A stored thinking block with a valid signature and a lone UTF-16 surrogate in thinking is emitted to the request payload with thinking unchanged; same for blocks that only store signature (no thinkingSignature).

User-visible / Behavior Changes

Multi-turn conversations with Anthropic models using extended thinking and prior assistant thinking blocks in history should no longer fail at the API with Invalid signature in thinking block when the prior turn’s thinking was stored with a signature.

Diagram (if applicable)

N/A

Security Impact (required)

• New permissions/capabilities? No
• Exposes or persists new secrets? No
• Network surface change? No

Testing

• pnpm test src/agents/anthropic-transport-stream.test.ts src/agents/anthropic-stored-thinking-signature.test.ts

Changed files

• src/agents/anthropic-stored-thinking-signature.test.ts (added, +29/-0)
• src/agents/anthropic-stored-thinking-signature.ts (added, +21/-0)
• src/agents/anthropic-transport-stream.test.ts (modified, +84/-0)
• src/agents/anthropic-transport-stream.ts (modified, +8/-4)
• src/agents/transport-message-transform.ts (modified, +3/-1)