What problem does it solve?
Long-running autonomous agent loops tend to guess when they hit ambiguous specs, missing credentials, or repeated failures, causing divergence that compounds across sessions. This Skill stops the loop cleanly and routes a structured question to a human before bad assumptions ship.
Core Features & Use Cases
- Multi-channel escalation: Sends a structured five-line message via Telegram bot API, Slack incoming webhook, or a custom dial command configured through environment variables.
- Guaranteed loop exit: Always writes a BLOCKED.md file with Question, Context, Attempted, and Choices sections, then exits with code 2 so run.sh halts until a human clears the file.
- Strict trigger discipline: Escalates only on ambiguous specs, missing credentials, destructive actions needing approval, external dependency outages, or 3+ consecutive verify failures on the same task.
- Use Case: An agent loop fails verification three times on a database migration that requires dropping a table. The Skill posts the question to Slack, writes BLOCKED.md with options A/B/C, and exits so the human can approve the destructive step and restart the loop.
Quick Start
Ask the agent to escalate the current blocked task to a human using the configured LOOPKIT_HITL_CHANNEL and write BLOCKED.md before exiting the loop.
