Summary

Claude runs can still fail when session history replays a signed thinking block that the provider rejects.

Observed with github-copilot/claude-opus-4.7:

LLM request rejected: messages.7.content.13: Invalid `signature` in `thinking` block

This appears to be a Claude signed-thinking replay issue, not a generic GitHub Copilot issue. Copilot is one affected route because it exposes Claude models and rejects the replayed signed thinking block.

Actual behavior

Continuing an existing session can fail before producing a reply. The logs show the provider rejects the outbound request payload because a historical assistant thinking block contains a signature that is invalid for this replay path.

This poisons the session for that model/provider path: subsequent turns keep replaying the incompatible block unless history is sanitized.

Expected behavior

OpenClaw should not replay Claude signed thinking blocks to provider paths that cannot safely accept them.

If a provider/model route cannot preserve Claude thinking continuity safely, the outbound replay history should strip those blocks completely before the request is sent.

Notes

The current policy shape also looks suspicious: extensions/github-copilot/replay-policy.ts sets dropThinkingBlocks: true for Claude models, but the shared thinking sanitizer preserves the latest assistant thinking block for native Anthropic continuity. That may be correct for direct Anthropic paths, but it is unsafe for provider routes that reject replayed signed thinking.

Relevant areas:

• extensions/github-copilot/replay-policy.ts
• src/agents/pi-embedded-runner/replay-history.ts
• src/agents/pi-embedded-runner/thinking.ts
• src/agents/pi-embedded-runner/run/attempt.ts

Acceptance criteria

• Existing sessions containing Claude signed thinking blocks can continue on affected provider/model routes instead of failing with Invalid signature in thinking block.
• The replay policy distinguishes paths that may preserve latest thinking from paths that must strip it completely.
• Direct Anthropic behavior is not regressed if it still requires latest thinking continuity.
• Add a targeted regression test for the affected Claude replay sanitization path.