Description

When a tool call is interrupted/terminated mid-execution, clawdbot inserts a "synthetic error result" to repair the transcript. However, when the session is later replayed for API calls, the tool_use/tool_result pairing becomes misaligned, causing Anthropic API to reject the request.

Error Message

LLM request rejected: messages.228.content.1: unexpected tool_use_id found in tool_result blocks: toolu_01TKLcHJJkbDgtJEvitNxEFg. Each tool_result block must have a corresponding tool_use block in the previous message.

Steps to Reproduce

1. Start a conversation with an agent via Telegram
2. Trigger a long-running tool call (e.g., exec command)
3. Interrupt/terminate the agent mid-tool-call (send "stop" or similar)
4. Continue the conversation
5. Eventually the session becomes corrupted and all requests fail

Session Analysis

Looking at the corrupted session JSONL:

Line 255: assistant message with toolCall id="toolu_01TKLcHJJkbDgtJEvitNxEFg", stopReason="error", errorMessage="terminated"
Line 256: synthetic toolResult with matching toolCallId, content="[clawdbot] missing tool result..."
Line 259+: API errors referencing the same tool_use_id

The synthetic result IS in the session file directly after the assistant message, but when repairToolUseResultPairing() runs in-memory before API calls, the pairing is not correctly reconstructed.

Root Cause Hypothesis

The issue is in src/agents/session-transcript-repair.js:

• The synthetic result is inserted into JSONL at the correct position
• But when building API messages from session entries, the message indexing differs
• The repair logic may be dropping the synthetic result as an "orphan" in edge cases

Environment

• Clawdbot version: 2026.1.24-3
• Provider: Anthropic (Claude Opus 4.5)
• Channel: Telegram
• Compaction mode: safeguard

Workaround

Delete the corrupted session file and restart:

rm ~/.clawdbot/agents/main/sessions/<session-id>.jsonl
systemctl --user restart clawdbot-gateway

Reported by: Nicolas Delrieu