Project 1 · A Watch Loop

⚠️ Throwaway repo. Set a limit first. Clone a disposable copy of whatever you point this at — never a production remote — and write the run cap into the prompt before beat one, per this course's own minimum-safe checklist.

The scenario

Pick a real, moderately active open-source repository you don't maintain — something with real traffic, like a mid-sized web framework or CLI tool with dozens of weekly commits (the reference-repo class of project — cobusgreyling/loop-engineering itself works fine, or any repo with 50+ stars and a live issue tracker). You want to know, every 15 minutes, whether anything changed that you'd care about — a new issue with the bug label, a PR opened against main, or a new release tag — without opening the browser yourself.

This is the shape every later project builds on: a heartbeat that fires on its own, a spine that remembers what it already saw, and nothing else. No writing, no fixing, no opinions — just a faithful watcher.

What you'll build

A schedule loop (Step 4/6 territory) with:

• Heartbeat: every 15 minutes, capped at 20 runs/day.
• Body: read the target repo's issues/PRs/tags via the SCM CLI (gh issue list, gh pr list, gh release list) — read-only, no gh write subcommands anywhere in the body.
• Spine: watch-state.md — the ID of the last issue/PR/release seen per category, so a restart doesn't re-report old news.
• Stop: none — this is a schedule loop; it runs until you cancel it. The provable target for this project is "one clean report emitted for at least 3 consecutive beats with no duplicate items and no missed new item you can independently verify."

# Claude Code
> /loop 15m Check <owner>/<repo> for anything new since watch-state.md's last-seen
  IDs: issues labeled bug, PRs against main, release tags. Update watch-state.md.
  Append exactly one line to loop-run-log.md whether or not anything was found.
  Never open, comment on, or close anything — report only. Stop after 20 runs today.

# OpenCode
for i in $(seq 1 20); do
  opencode run "Check <owner>/<repo> for anything new since watch-state.md's last-seen
    IDs (issues:bug, PRs:main, releases). Update watch-state.md. Append one line to
    loop-run-log.md. Read-only — never open/comment/close."
  sleep 900
done

Done when (provable stop for this exercise)

• watch-state.md exists and is committed before beat 1.
• The loop has completed at least 3 beats on the 15-minute cadence.
• loop-run-log.md has exactly one line per beat — including beats where nothing changed (a "nothing new" beat is not a skipped beat).
• You can point at one real item (an issue, PR, or tag) the loop caught, and show the spine line that recorded it.
• Zero writes landed anywhere outside watch-state.md and the log.

Going deeper

Once this is boring, this exact shape is pr-babysitter and daily-triage from the starter library — go read starters/daily-triage/ and notice it's the same six parts, just with a report file added. That's the whole trick of Track T2: you're not learning a new pattern per project, you're adding one part at a time to a shape you already trust.

When it goes wrong

---

Sources: project shape and the "throwaway repo, set a limit first" convention come from Panaversity's Loop Engineering: A Crash Course (S1), §18; schedule-loop mechanics from Scheduled Tasks: The Loop Skill & Cron Tools (S4). Full attribution: resources/sources.md.