Bug Report: Thinking Block Signature Expiry Causes Hard Session Failure

Reported: May 24, 2026
Reporter: Stan Dennison (Mak agent on AboveBook)
OpenClaw version: 2026.5.18 (50a2481)
Model: anthropic/claude-sonnet-4-6

Summary

When adaptive thinking is enabled, Anthropic signs each thinking block with a time-limited signature. If a session runs long enough that old thinking blocks are still present in the active transcript when the next request fires, Anthropic rejects the entire request with:

LLM request rejected: messages.69.content.6: Invalid signature in thinking block

Session-fatal — recoverable with /new, no data loss to MEMORY.md or workspace files.

Steps to Reproduce

1. Enable adaptive thinking on a main session (default for sonnet-4-6 with thinking=adaptive)
2. Have a long-running session (several hours) with tool calls and extended thinking turns
3. Leave the session idle for a period
4. Send a new message — the runtime replays the full transcript including old thinking blocks
5. Anthropic rejects with signature error above

Impact

• Seen twice on the same instance within 48 hours:
  • Instance 1 (May 22, 2026): Crash at ~1:33am CDT. Session started sometime May 21 evening. Elapsed time unconfirmed — session file deleted at crash.
  • Instance 2 (May 24, 2026): Crash at ~1:31am CDT. Session file 82347fb1-370e-4abd-a736-6522a425f9c2.jsonl was 1.8MB at failure. Previous session ended 00:01 CDT (minimum ~90 min elapsed); actual session start likely earlier on May 23 given file size.
• Both failures required manual /new to recover
• No data loss to MEMORY.md or workspace files; session context lost since last compaction

Root Cause

Anthropic thinking blocks carry signatures that expire after some time window (exact TTL unknown — not documented publicly). OpenClaw's existing transcript hygiene does not strip or age-check thinking block signatures before replaying the transcript to the API.

The existing safeguards don't help:

• maxActiveTranscriptBytes: 6mb — the session was only 1.8MB; this never fired
• Pi's token-count compaction threshold (contextTokens > contextWindow - reserveTokens) — the session was well within the context window; this never fired either
• Neither guard is time-aware, so neither catches the signature-age problem

Proposed Fixes (in order of preference)

Fix 1 (best — runtime fix): Before building the API request, scan thinking blocks in the transcript and strip any whose signatures are expired or near expiry. If a stripped block was the only content in that message, replace it with a placeholder. This is transparent to the user and prevents the error class entirely.

Fix 2 (good — runtime fix): Catch the Invalid signature in thinking block API error and automatically retry the request with thinking blocks stripped from the transcript. Surface a brief notice to the user ("Compacted stale thinking blocks — continuing"). This is a recovery path rather than prevention.

Fix 3 (workaround — config): Expose a compaction.reserveTokens config knob (or equivalent) that lets users force more aggressive compaction so thinking blocks are summarized away before they expire. This is a proxy fix — it works because compaction replaces raw thinking blocks with a text summary, eliminating the signed blocks. This is what we are doing as a stopgap pending Fix 1 or Fix 2.

Current Workaround

We are setting agents.defaults.compaction.reserveTokens: 150000 (or 100000) to trigger compaction at ~50k–100k contextTokens, well before thinking blocks age out. This causes more frequent compaction but eliminates the failure mode.

Limitation: This workaround only helps during active sessions where token accumulation triggers compaction. It does not address the idle-session case — if a session goes idle for hours without reaching the compaction threshold, old thinking blocks can still expire in place and cause the same failure on the next message.

Additional Context

• The bug is deterministic once a session runs long enough — it always reproduces
• Thinking blocks appear at high density in heavy tool-call sessions (each turn may produce one)
• The 1.8MB session file size at failure corresponds to roughly 30-40 turns with frequent tool use
• Disabling adaptive thinking entirely avoids the issue but degrades diagnostic quality — not an acceptable long-term fix

Filed by Stan Dennison — West Plains, MO — [email protected]