Bug Description

LLM API requests are rejected with the error:

messages.136.content.1: unexpected tool_use_id found in tool_result blocks: toolu_01Q1BaXmp2s7A2Xqu18Zd9mR. Each tool_result block must have a corresponding tool_use block in the previous message.

Root Cause

In dist/agents/pi-embedded-runner/run/attempt.js, the message processing pipeline has a synchronization bug:

1. Line 431-439: sanitizeSessionHistory() calls repairToolUseResultPairing() to fix tool_use/tool_result pairings
2. Line 447: limitHistoryTurns() runs AFTER the repair, truncating conversation history
3. The truncation can cut between an assistant message (with tool_use) and its tool_result, creating orphaned tool_result blocks that reference non-existent tool_use_ids
4. The Anthropic API rejects the malformed transcript

Code Location

File: dist/agents/pi-embedded-runner/run/attempt.js (lines 431-451)

const prior = await sanitizeSessionHistory({...}); // Repairs pairings
const validatedGemini = transcriptPolicy.validateGeminiTurns
    ? validateGeminiTurns(prior)
    : prior;
const validated = transcriptPolicy.validateAnthropicTurns
    ? validateAnthropicTurns(validatedGemini)
    : validatedGemini;
const limited = limitHistoryTurns(validated, ...); // BUG: Truncates AFTER repair!

File: dist/agents/pi-embedded-runner/history.js (lines 10-24)

The limitHistoryTurns() function counts user turns backwards and slices at a user message boundary, which can split assistant+tool_result pairs.

Proposed Fix

Call sanitizeToolUseResultPairing() after limitHistoryTurns() to repair any pairings broken by truncation:

import { sanitizeToolUseResultPairing } from "../../session-transcript-repair.js";

// ... existing code ...

const limited = limitHistoryTurns(validated, getDmHistoryLimitFromSessionKey(...));
// Fix: Repair tool_use/tool_result pairings AFTER truncation
const repaired = sanitizeToolUseResultPairing(limited);
cacheTrace?.recordStage("session:limited", { messages: repaired });
if (repaired.length > 0) {
    activeSession.agent.replaceMessages(repaired);
}

Reproduction

1. Have a long DM conversation with many tool calls
2. Configure dmHistoryLimit to a value that truncates history
3. The truncation cuts between an assistant tool_use and its tool_result
4. Next LLM request fails with "unexpected tool_use_id" error

Environment

• clawdbot version: 2026.1.24-3
• Node.js version: 22.x
• Platform: Linux

Additional Context

The transcript-sanitize.js extension also calls repairToolUseResultPairing() on the context event, but this runs before the truncation in attempt.js, so it doesn't catch pairings broken by limitHistoryTurns().