Rumplestiltskin, [Jan 9, 2026 at 1:58:45 PM]:
...Summary

When an agent goes off the rails during a long multi-tool execution, there's currently no way to interrupt it mid-run. Stop commands (/stop, stop, abort, etc.) sit in the message queue and aren't processed until the current turn completes.

What Happened

User sent a task. Agent misunderstood and started working on the wrong thing. User sent:

• "stop" (ignored)
• "STOP STOP STOP STOP" (ignored)
• "/stop" x4 (ignored)
• "FUCKING STOP ALREADY" (ignored)
• "/new" (ignored)
Agent continued for 10+ minutes making changes across 9 different files, completely deaf to all abort attempts.

Current Architecture (per investigation)

1. Default queue mode is "collect" — messages queue until current run finishes

2. Abort trigger check only happens when messages are processed, not when they arrive

3. No polling for abort signals between tool calls during execution

4. Even with "queue": { "mode": "interrupt" } added to config, this may not help mid-execution
   Kill Procedure Required

5. pkill -f clawdbot (kill gateway)

6. Clear sessions: rm -f ~/.clawdbot/agents/main/sessions/*

7. Clear Telegram queue: curl "https://api.telegram.org/bot/getUpdates?offset=<last+1>"

8. Restart gateway
   Requested Enhancement

Add ability to check for abort signals between tool calls during a long-running agent turn.

Possible approaches:

1. Poll message queue between tool invocations
2. Use a signal file or IPC mechanism
3. Add a check_abort() call in the tool execution loop
4. WebSocket/event-driven

abort notification

Impact

Without this, a runaway agent can:

• Make many unwanted changes before user can stop it
• Waste tokens/compute on wrong tasks
• Erode user trust in the system
• Require manual intervention (kill process, clear queues)

Environment

• Clawdbot version: 2026.1.8-2
• Surfaces: Telegram, WhatsApp
• Model: claude-opus-4-5

Daniel Fries, [Jan 9, 2026 at 1:58:57 PM]:
dev is same as oracle dev - so might mention that

Rumplestiltskin, [Jan 9, 2026 at 1:58:58 PM]:
🧙

Good point! Add this to the top or bottom:

Note: This was discovered while using the Oracle CLI integration. Same developer — hi! 👋 The runaway agent was attempting Oracle evaluations of agent files when it went off the rails. The Oracle tool itself worked fine; the issue is Clawdbot's inability to interrupt mid-execution.