Solution · Project 1 — A Watch Loop

Reference solution. Read the project first and build your own before comparing — the value is in the diff between your design and this one.

Worked example

Target repo: openai/openai-python (real, active, public — substitute any repo you don't maintain). Cadence: 15 minutes, capped at 20 runs/day.

watch-state.md (spine), after 3 beats:

# Spine: watch-loop (Project 1)

**Runs used:** 3 / 20

| Category | Last-seen ID | Last checked |
| --- | --- | --- |
| issues (label:bug) | #2184 | 2026-07-21T09:45Z |
| PRs → main | #2190 | 2026-07-21T09:45Z |
| release tags | v1.54.0 | 2026-07-21T09:30Z |

Prompt used (Claude Code):

> /loop 15m Check openai/openai-python 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. Stop after 20
  runs today.

Three run-log lines it produced:

{"run_id":"2026-07-21T09:15:00Z","pattern":"watch-loop","duration_s":40,"items_found":1,"actions_taken":0,"escalations":0,"tokens_estimate":3000,"outcome":"new PR #2190 opened against main"}
{"run_id":"2026-07-21T09:30:00Z","pattern":"watch-loop","duration_s":35,"items_found":1,"actions_taken":0,"escalations":0,"tokens_estimate":2800,"outcome":"new release v1.54.0"}
{"run_id":"2026-07-21T09:45:00Z","pattern":"watch-loop","duration_s":30,"items_found":0,"actions_taken":0,"escalations":0,"tokens_estimate":2500,"outcome":"nothing new — quiet beat, logged anyway"}

Why this satisfies "done when"

• The spine was committed before the first /loop invocation — not created by beat 1.
• The third beat found nothing and still logged a line — "silence is the loudest signal" (observability.md) applies to the loop's own heartbeat, not just what it's watching.
• gh issue list / gh pr list / gh release list are read-only subcommands; the harness permission grant for this loop excluded every gh issue comment, gh pr review, etc. outright, so the "never comment" rule didn't depend on the model choosing to obey it.

Common wrong turns (and why they fail the done-when)

• Storing "last checked" timestamp only, no per-category ID → re-reports everything created in a window if the loop restarts mid-cadence. IDs are the correct key, not time.
• Skipping the log line on a quiet beat → makes a genuinely stalled loop indistinguishable from one that's just watching a quiet repo.