hitl-escalate

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:

Value	Shape
telegram	POST to https://api.telegram.org/bot$LOOPKIT_HITL_TELEGRAM_TOKEN/sendMessage with chat_id=$LOOPKIT_HITL_TELEGRAM_CHAT, text=<question + repo + short context>.
slack	POST JSON {"text": "..."} to $LOOPKIT_HITL_SLACK_WEBHOOK (incoming-webhook URL).
dial	Run $LOOPKIT_HITL_DIAL_CMD with the message on stdin. Whatever the operator wired up — SMS gateway, ntfy, phone call, pager.
none (or unset)	Skip the primary action. Go straight to fallback.

Message body — always these five lines, in this order:

[loopkit] blocked in <repo-name> on <ISO timestamp>
Q: <one-line question>
Context: <one-line what you were doing>
Attempted: <one-line what you tried>
Choices: <A | B | C>

Non-2xx from the channel is a soft failure. Log it, then fall through to the fallback so the loop still exits cleanly.

Fallback — write BLOCKED.md and exit

Whether the primary succeeds or fails, always write ./BLOCKED.md at the repo root with exactly four sections:

# Blocked

## Question:
<one sentence, answerable with a short reply>

## Context:
<what feature, which file, which commit, why now — 3-6 lines>

## Attempted:
<bulleted list of what you tried and why each fell short>

## Choices:
- A) <option> — <consequence>
- B) <option> — <consequence>
- C) <option> — <consequence>

Then exit 2. The loop runner (run.sh) checks for BLOCKED.md at the head of each iteration and exits with code 2 if it exists — that stops the loop until a human clears it.

Human unblocks

The human reads BLOCKED.md, edits PROMPT.md / drops the credential / approves the destructive step, then rm BLOCKED.md and restarts run.sh. The deleted BLOCKED.md is the unblock signal.

Anti-patterns — do not do these

• Do not fabricate an answer. "I'll assume the user meant X" is how divergence starts. Assume nothing under ambiguity — escalate.
• Do not loop endlessly. After the third consecutive /verify failure on the same task, escalate. The fourth attempt is not going to succeed by the same reasoning that failed thrice.
• Do not silently swallow the missing credential. Do not commit a placeholder, do not disable the feature, do not "TODO" the auth. Ask.
• Do not escalate for questions the repo already answers. Read AGENTS.md, PROMPT.md, and the last three commits first. Escalation is expensive human attention — spend it on real ambiguity.
• Do not skip writing BLOCKED.md because the primary channel succeeded. The file is the loop's exit contract. Without it, run.sh keeps spinning.

Pairs with

• adversarial-verify — the source of the 3-verify-failure trigger.
• shift-notes — after escalating, note the block in IMPLEMENTATION_PLAN.md so the next session (post-unblock) has full context.
• spec-first — an ambiguous PROMPT.md is a spec-first failure; the human's unblock should tighten the spec, not just answer the one-off.