loop

name	loop
description	Use when the user types /loop or asks to repeat/poll/babysit something on a cadence inside THIS chat. Examples — "/loop 5m check the deploy", "watch CI until green", "every 30 minutes summarize my inbox". Sets up a recurring prompt that fires into this session via the create_loop tool.

/loop — schedule a recurring prompt in this chat

The user wants a repeating task scoped to THIS conversation. Parse their input, call create_loop once, then execute the parsed prompt inline yourself so the user sees a first iteration immediately instead of waiting for the next scheduler tick.

Parsing (in priority order)

1. Leading token interval: if the first whitespace-separated token matches ^\d+[smhd]$ (e.g. 5m, 2h, 1h30m), 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

Minimum interval is 1m. If parsing yields something shorter, round up to 1m and tell the user.

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 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 create_loop

Call create_loop 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.

After create_loop 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 create_loop 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 create_loop 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 create_automation instead.
• If the user wants something done once now, not recurring: just do it, don't loop.