Human-in-the-Loop Escalate

Long-running agent loops fail in a specific way when they meet a question they can't answer: they guess. The guess ships, the next session builds on it, and by the time a human looks the divergence is three commits deep. This skill is the release valve — when the agent is stuck, it stops and asks, cleanly, in a shape the human can act on.

Trigger — escalate when any of these are true

• Ambiguous spec. PROMPT.md admits two implementations and the choice changes user-visible behavior.
• Missing credential. A required env var, API key, or config file is absent, and there is no sanctioned way for the agent to create one.
• Destructive action needing approval. Dropping a table, force-pushing, deleting user data, spending money, sending mail to real recipients.
• 3+ consecutive /verify failures on the same task. The loop is thrashing. Stop and get a human eye before the fourth attempt.
• External dependency down. A third-party service the feature needs is unreachable and no offline path exists.

If none of the above hold, do not escalate. Guessing is bad; escalating on a solvable problem is also bad (it trains humans to ignore the channel).

Primary action — call the human via the configured channel

Read LOOPKIT_HITL_CHANNEL. Supported values: