Summary

The anthropic-transport-stream.ts (used by the embedded agent runner) overwrites thinkingSignature on each signature_delta event instead of appending to it. Since Anthropic sends the thinking block signature across multiple signature_delta streaming chunks, only the last chunk survives. The truncated signature is persisted to the session JSONL, and all subsequent replay attempts fail with:

400: messages.N.content.M: thinking or redacted_thinking blocks in the latest assistant message cannot be modified

This permanently bricks the session with no user recovery path other than /reset.

Root Cause

File: src/agents/anthropic-transport-stream.ts:1289

// BUG: overwrites instead of appending
block.thinkingSignature = delta.signature;

Correct implementation (same repo): src/llm/providers/anthropic.ts:629-634

// Correct: initializes then appends all chunks
block.thinkingSignature = block.thinkingSignature || "";
block.thinkingSignature += event.delta.signature;

Failure Mechanism

1. Anthropic API sends the thinking block signature in N signature_delta chunks (base64 is long)
2. Transport stream overwrites on each chunk - only last chunk survives
3. Truncated signature stored in session JSONL
4. Next request replays the truncated signature - Anthropic rejects it with 400
5. Error repeats on every subsequent turn - session permanently bricked

Secondary Issue (Non-blocking follow-up)

src/agents/anthropic-transport-stream.ts:1125 does not set a display field when initializing thinking blocks, while src/llm/providers/anthropic.ts:987 sets display: "summarized".

Note: After analysis, this is a separate concern from the signature corruption bug:

• display is a request-level parameter (params.thinking.display), not a content block field
• The TransportContentBlock type does not include a display field
• Its absence does not affect signature integrity or replay correctness
• It may affect thinking behavior on newer models (Opus 4.7 / Mythos Preview) that support adaptive thinking
• This should be tracked as a separate follow-up, not a blocker for the signature fix

Recovery for Already-Corrupted Sessions

Existing sessions with truncated signatures are handled by wrapAnthropicStreamWithRecovery (thinking.ts:482):

1. Corrupted session replays truncated signature → Anthropic returns 400
2. THINKING_BLOCK_ERROR_PATTERN matches the error message
3. Recovery strips ALL thinking blocks from the message history
4. Retry succeeds without thinking blocks → session continues normally
5. New responses store correct signatures going forward

Users do NOT need to /reset — the recovery mechanism handles this automatically on the next request.

Real-World Observation

Observed on a production Mac Mini gateway (OpenClaw 0.6.0-dev) with claude-opus-4-6:

• Session with thinking enabled produces assistant message with thinking blocks
• After a few turns, every subsequent request fails with the 400 thinking block error
• Session permanently stuck until /reset
• Gateway logs show repeated embedded run failover decision: decision=surface_error reason=format

Related Issues

• Bug: thinking/redacted_thinking blocks cause permanent session corruption on cache miss #80625 - Same symptom (permanent session corruption from thinking blocks), suggests strip-on-persist as fix. Our finding shows the actual root cause is upstream in the streaming layer, not in persistence.
• Claude signed thinking blocks can still poison session replay #85781 - "Claude signed thinking blocks can still poison session replay" - P1 maintainer issue. The replay policy discussion is relevant but the root cause is the malformed signature from transport stream, not the replay policy.
• Filter noisy invalid_request_error subtypes (e.g. thinking-signature) out of user-facing assistant text #87475 - "Filter noisy invalid_request_error subtypes" - Reports the same Invalid signature in thinking block error reaching users. The wrapAnthropicStreamWithRecovery (Bug: Empty thinking-block signatures poison Anthropic Messages sessions (fix from #70054 not applied to standard Anthropic transport) #86886) single-retry cannot help because the signature is fundamentally truncated at storage time.
• Bug: Subagent announce-delivery echo messages inherit wrong provider/model metadata, causing persistent "thinking blocks cannot be modified" errors after gateway restart #87329 - Subagent echo messages with wrong provider metadata amplify the same thinking-block 400 error after gateway restart.

Why Existing Fixes Are Insufficient

• Bug: Empty thinking-block signatures poison Anthropic Messages sessions (fix from #70054 not applied to standard Anthropic transport) #86886 wrapAnthropicStreamWithRecovery: Provides one-shot recovery (strip thinking + retry), but cannot prevent corruption from recurring on every new thinking response. The root cause fix is needed to stop new signatures from being truncated.
• Session file repair (session file repaired: rewrote N assistant messages): Does not reconstruct the correct signature (impossible without re-streaming from Anthropic).
• dropThinkingBlocks replay policy: Only helps non-native routes (Copilot). Native Anthropic routes need valid signatures for continuity.

Fix

// src/agents/anthropic-transport-stream.ts:1289
// Before:
block.thinkingSignature = delta.signature;

// After:
block.thinkingSignature = (block.thinkingSignature || "") + delta.signature;

Acceptance Criteria

• signature_delta chunks are concatenated, not overwritten
• Sessions with thinking enabled can survive multiple turns without 400 errors
• Existing tests for thinking block handling pass
• Real-scenario proof demonstrates the patched code path correctly concatenates signatures

Non-blocking Follow-up

• Transport thinking display parameter alignment (separate from content block storage; does not affect signature integrity)