Autopilot Scheduler

Put /speckit-autopilot-run on a recurring launchd timer so open issues get worked without a human starting each pass. This is strictly opt-in: nothing is scheduled until you run this command, and it only ever schedules this repo.

$ARGUMENTS may contain a subcommand and options:

• install (default if you just say "schedule autopilot") — schedule it.
  • --interval-hours N — cadence in hours (default 2, configurable to any positive integer).
• uninstall — stop the recurring runs and remove the agent.
• status — is it scheduled? show the interval, plist, and the tail of the log.
• run-now — trigger a single pass immediately (only if already scheduled).

All of these are thin wrappers over the script — run it, don't reimplement launchd by hand:

BIN="$CLAUDE_PROJECT_DIR/.specify/extensions/autopilot/scripts/bash/autopilot-schedule.sh"

(If $CLAUDE_PROJECT_DIR isn't set, resolve the repo root with git rev-parse --show-toplevel and use <root>/.specify/extensions/autopilot/scripts/bash/autopilot-schedule.sh.)

What to do

1. Parse $ARGUMENTS into a subcommand (default install) and, for install, an --interval-hours value if the user named one ("every 3 hours" → 3). If they gave no number, use the script's default of 2.

2. Run the script for this repo, passing the project explicitly so it works from a worktree too:

   # schedule every 2h (default)
   "$BIN" install --project "$CLAUDE_PROJECT_DIR"
   
   # schedule every N hours
   "$BIN" install --interval-hours N --project "$CLAUDE_PROJECT_DIR"
   
   # check / stop / run once
   "$BIN" status    --project "$CLAUDE_PROJECT_DIR"
   "$BIN" uninstall --project "$CLAUDE_PROJECT_DIR"
   "$BIN" run-now   --project "$CLAUDE_PROJECT_DIR"
   

3. Report back what the script printed — the label, cadence, plist path, and log path — and tell the user how to change or stop it (re-run install --interval-hours N to re-cadence; uninstall to stop). On install, remind them the first pass fires after the interval (RunAtLoad is off, so it won't surprise-run the moment it's scheduled); run-now triggers one immediately.

How it runs (so you can explain it)

• A per-repo launchd agent (~/Library/LaunchAgents/com.speckit.autopilot.<repo>.plist) fires every StartInterval seconds and runs autopilot-run.sh <repo>, which invokes claude -p "/speckit-autopilot-run" --dangerously-skip-permissions inside the repo. The permission bypass is what makes an unattended pass possible — the session can use git/gh/file tools without a human clicking approve. Flag that to the user when they schedule, so it's a conscious choice.
• Runs are single-flight: if a pass outlasts the interval, the next tick is skipped rather than stacking two autopilots on the same repo.
• Output goes to ~/Library/Logs/speckit-autopilot/<repo>.log (status tails it).

Stop conditions

• Not macOS — launchd is macOS-only; the script exits with a clear message. On Linux, a systemd --user timer or cron would be the equivalent (not built here).
• Not a git repo — the script needs a repo root to label the agent; it refuses outside one.