Practical patterns, starters & CLI tools for loop engineering with AI coding agents. Design systems that prompt and orchestrate agents (inspired by Addy Osmani and Boris Cherny). Includes loop-audit, loop-init, loop-cost.
Thirty weekday runs — what Loop Ready 100 hid
Pattern: Daily Triage (this repository)
Window: 2026-07-29 → 2026-08-27 (the run log prunes at 30 days)
Level: L1 report-only, GitHub Actions
Setup
.github/workflows/daily-triage.yml runs weekdays, writes STATE.md from open PRs and issues (scripts/github-triage.mjs), appends loop-run-log.md, and opens an automated PR. Humans merge. No auto-fix.
Loop Ready on this repo stayed 100 (L3) for the whole window.
What the log actually shows
| Metric | Value |
|---|---|
| Recorded runs | 22 (weekdays; log does not keep a full 60 days) |
| Outcome | 22 / 22 report-only |
| Escalations | 0 |
items_found sum |
28 (almost always 1 per morning) |
tokens_estimate sum |
1,144,000 |
The token column is a constant 52,000 written by the workflow, not a meter. Treat it as an upper bound we never measured. That is a loop-engineering own-goal: we tell everyone to log spend, then log a placeholder.
What worked
- The heartbeat did not die. Automated PRs for
STATE.mdkept landing. - Report-only held. Zero fix attempts, zero merges from the agent.
- The file that mattered to humans was the High Priority / Watch list, not the score.
What broke (or never started)
- Score ≠ work. L3 on a loop that never leaves L1 is verifier theater for the badge.
- Estimates are not observations.
tokens_estimate: 52000every row taught us nothing about cost. - Stale GFIs sat in Watch for 50+ days. The loop reported them; nobody closed or rewrote them. A report-only loop still needs a human who reads.
- We asked for 60 days of numbers and only kept 30. The log’s own prune rule deleted the evidence a Show HN needs.
Lesson
A loop that only writes STATE.md is still a cron job with a better filename — and that can be the right design. The failure is advertising L3 / 100 while the run log says report-only and the token field is hardcoded.
Changes that follow from this story: last-run badge on the README, Loop Ready weights fresh runs over files, and a thin loop for teams who should not pretend they have state files.
Metrics (copy for Show HN)
22 weekday Daily Triage runs
22 report-only, 0 escalations
Loop Ready 100 the whole time
token field: constant 52k, not measured
STATE.md last run: 2026-08-27