Bug Description

When using Anthropic models with extended thinking enabled, the following error occurs intermittently:

LLM request rejected: messages.7.content.1: `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 Analysis

Anthropic's API requires that thinking / redacted_thinking blocks in the latest assistant message remain exactly as they were in the original response. Any modification triggers a rejection.

OpenClaw's compaction and context pruning pipelines do not currently have special protection for thinking blocks in the latest assistant message:

1. compaction.ts: No thinking-aware logic. When compaction replaces old messages with a summary, the latest assistant message's thinking blocks may be lost or altered.

2. context-pruning/pruner.ts: Only touches toolResult messages (not assistant messages), so unlikely the direct cause. However, it counts thinking block chars for window estimation.

3. transcript-policy.ts: dropThinkingBlocks is only enabled for GitHub Copilot Claude (isCopilotClaude), not for direct Anthropic API usage — correct behavior. But there's no complementary "preserve latest assistant thinking blocks" guard for Anthropic.

Expected Behavior

The latest assistant message's thinking / redacted_thinking blocks should be preserved exactly as-is through compaction, pruning, and session restore operations when targeting Anthropic's API.

Suggested Fix

Add a guard in the compaction and/or session sanitization pipeline that:

• Identifies the last assistant message in the transcript
• Ensures its thinking / redacted_thinking blocks are never modified or dropped
• Applies this guard specifically when the target provider is Anthropic

Environment

• OpenClaw version: 2026.2.27 (dev channel)
• Model: anthropic/claude-opus-4-6
• Thinking: enabled (low)
• Compaction mode: default
• Context pruning: cache-ttl (20m)

Reproduction

1. Use Anthropic model with thinking enabled
2. Have a long session that triggers compaction or context pruning
3. Send a follow-up message after compaction runs
4. Error occurs when the compacted transcript is sent to Anthropic with modified thinking blocks