Bug

During session compaction, assistant messages containing thinking or redacted_thinking blocks are modified (summarized/trimmed). The Anthropic API rejects this:

messages.131.content.7: thinking or redacted_thinking blocks in the latest assistant message cannot be modified. These blocks must remain as they were in the original response.

Root Cause

Compaction rewrites/summarizes assistant messages, but Anthropic requires thinking blocks to remain exactly as generated. The compactor needs to either:

1. Strip thinking blocks entirely before compaction
2. Preserve thinking blocks verbatim when summarizing surrounding content
3. Replace the entire message (including thinking) with a plain text summary

Impact

• Blocks LLM responses until compaction retries or session resets
• User sees raw error message in Telegram
• Happens more frequently with large sessions and adaptive thinking enabled

Environment

• OpenClaw 2026.3.1 (2a8ac97)
• Model: claude-opus-4-6 with thinking=adaptive
• Compaction mode: default