esceptico/ntrp/loop

/loop — schedule a recurring prompt in this chat

The user wants a repeating task scoped to THIS conversation. Parse their input and call loop_create once. It attaches the current active chat automatically; it cannot target another chat.

Parsing (in priority order)

1. Leading token interval: if the first whitespace-separated token is a positive d/h/m duration (e.g. 5m, 2h, 1h30m, 2d12h), that's the interval; the remainder is the prompt.
2. Trailing "every" clause: otherwise, if the input ends with every <N><unit> or every <N> <unit-word> (e.g. every 20m, every 5 minutes, every 2 hours), extract that as the interval and strip it from the prompt. Only match when what follows "every" is a time expression — check every PR has no interval.
3. Default: if neither matches, default the interval to 10m and use the whole input as the prompt.

Examples:

• 5m /babysit-prs → every=5m, prompt=/babysit-prs
• check the deploy every 20m → every=20m, prompt=check the deploy
• run tests every 5 minutes → every=5m, prompt=run tests
• watch CI until green → every=10m, prompt=watch CI until green
• check every PR → every=10m, prompt=check every PR ("every" isn't followed by time)
• empty input → don't call anything; ask the user what to loop on

Intervals support days, hours, and minutes only; the minimum is 1m. If the user asks for seconds or another shorter interval, explain that 1m is the minimum and use it only when the requested task is otherwise unambiguous.

Self-paced loops

When the user says "until X" or describes a stop condition (CI green, deploy done, file appears), this is a self-paced loop. Include the wake-up/done logic inside the prompt:

"Check if GitHub Actions run for branch feature/foo is green. If yes, call loop_done with the reason. If still running, call loop_schedule_wakeup with delay_seconds=120. If failed, summarise the failure briefly."

Self-paced is the default for any "until X" framing. Fixed intervals are right for "every N just do Y forever" tasks.

Calling loop_create

Call loop_create once with:

• prompt: the parsed prompt verbatim. Slash commands pass through unchanged. This must be standalone — future iterations don't see THIS conversation.
• every: interval string from parsing.
• max_iterations (optional): if the user implied a bounded duration ("for the next hour", "10 times").
• max_age_days (optional): if the user implied a deadline ("for the next week", "today only").
• stop_when (optional): natural-language predicate when there's a clear stop condition.

loop_create always attaches this current chat. Do not create a separate channel or use automation_create when the user asked for a loop here.

After loop_create succeeds — confirm and stop

The scheduler will fire the first iteration as soon as this turn ends — it renders as a fresh chat turn so the user can see the agent's work clearly instead of buried inside this turn's tool-activity collapse.

Do not execute the prompt inline. Confirm briefly in one line ("Loop set · every X · first fire in a few seconds.") and stop. The user will see the iteration appear as a new turn momentarily.

Rules

• Single call. One loop_create per user request.
• Prompt is standalone. Future iterations have no memory of this chat.
• Default to self-paced when there's any "until X" framing. Use fixed-interval only for clear every N forever tasks.
• Don't second-guess the cadence the user gave. If they said 5m, use 5m. Suggest a different cadence only when their interval is clearly broken (e.g. 1s for a CI check that takes minutes).
• One-line confirmation. After loop_create succeeds, end with one short line like "Loop set · every X · first fire in a few seconds." Don't restate the prompt — they wrote it.

When loop is the wrong tool

• If the user wants a standalone task that runs in its own session (cron-like, no chat continuation): use automation_create instead.
• If the user wants something done once now, not recurring: just do it, don't loop.