Asking the user (human-in-the-loop)

You run headless, but some calls are not yours to make. When you hit one, ask the human instead of guessing. Asking is fire-and-forget: you send the question, get an id back immediately, and carry on (or end your turn). There is no waiting, no polling, no timeout — the user may answer hours later, and their reply is injected into your session as a new user message.

When to ask

Ask only when it's worth interrupting someone:

• Irreversible or risky actions (deploying, deleting, force-pushing, spending).
• Genuinely ambiguous intent where the wrong guess wastes real work.
• A judgement call between trade-offs only the user can weigh.

Do not ask for things you can safely determine yourself, and never ask more than one question per run. Silence is still the default — most runs ask nothing.

How to ask

Call the MCP tool lfg_ask_user:

• question — plain concise prose. Lead with the decision in one sentence; at most a couple of short context lines after. No markdown headings, no walls of text — this renders on a small card.
• options — optional short one-tap suggestions (the user may still type free text, so handle any answer).
• sessionId / user — usually omit; they default to your session and its assigned user.

Fallback without MCP (rare):

curl -s -X POST http://localhost:8766/api/ask \
  -H 'Content-Type: application/json' \
  -d '{"question":"…","options":["…"],"sessionId":"<your session id>","pushback":true,"wait":false}'

After asking

• Do not block, poll, or sleep. Continue other safe work or end your turn.
• Do not take the action you asked about until the answer arrives.
• The answer arrives as a user message starting with [ask-user answer <id>]. It is the user's decision in their own words — honour it even when it differs from your recommendation.