Step 13 · Build the Morning-Triage Loop — Twice

Everything from Steps 1–12 finally clicks together into one working loop: a morning triage that reads your repo's overnight noise and leaves a report next to your coffee. You'll build it twice — Claude Code and OpenCode — to prove that the shape is what matters.

The hook

Tomorrow at 7:00 am, a loop will comb through every issue, PR, and CI failure that arrived overnight. It will separate the genuinely urgent from the safely-ignorable and drop a five-line report where your eyes will land first. And it will do nothing else. That isn't a toy demo — it's the smallest possible loop that still has all six organs, and both walkthroughs in this part ship it for real.

The design (plain English)

The job: triage the repo, every morning. The shape falls straight out of the pattern-picker logic in the methods layer (09-methods/pattern-picker.md):

• The work comes back on a calendar, so the heartbeat is a schedule.
• The beat only reads, so it runs L1 report-only.
• The output is a judgment call, so a human gate reads the report.

Here are the six parts, filled in:

Both tools lean on two artifacts you write exactly once. A daily-triage/SKILL.md holds the procedure — what to read, how to rank it, the report template. A reviewer agent definition is the read-only grader both tools point at the finished report. The two walkthroughs part ways on only one thing: the heartbeat plumbing — a cloud Routine or claude -p cron (13a) versus cron/GitHub Actions (13b).

The minimum-safe checklist (before EITHER version runs)

Seven items. This repo's own loops clear this list before every first run, and so does yours — or it doesn't get to run:

1. Provable success condition (report file exists, matches template)
2. Run limit (daily cap: 1)
3. Spine written first, committed
4. Report-only (L1) — no write permissions beyond report + spine
5. Human gate placed (you, reading the report)
6. One log line per beat, no silent runs
7. Kill switch you've actually tested (schedule off / pause flag)

The skill both versions share

A single skill file carries the entire procedure, and one reviewer definition grades what it produces. Both walkthroughs use these two verbatim:

# daily-triage/SKILL.md — used verbatim by BOTH walkthroughs
# name: daily-triage
# description: Morning repo triage. Read-only. Produces the 5-line report.
# 1. Read: new/updated issues, PRs, CI runs since state's last-seen marks.
# 2. Rank: broken-main > failing-CI > stale-urgent-PRs > new-issues > rest.
# 3. Write triage-report.md: ≤5 lines, most urgent first, one line each:
#    [rank] what · why it matters · suggested (NOT taken) action.
# 4. Update triage-state.md last-seen marks. Take NO other action.

# The reviewer agent both tools run over the report (read-only grader):
# reviewer: "Check triage-report.md: ≤5 lines? ranked? each line has
#   what/why/suggested-action? last-seen marks updated? PASS/FAIL + why."
# In Claude Code this is a subagent definition; in OpenCode an agent
# config — see each walkthrough. Neither may edit the report.

[!NOTE] Going deeper: "one real morning" is the graduation rule — a loop earns trust by running for real, watched, at L1, at least once before you're allowed to stop checking it daily. The promotion ladder past L1 (labeling issues at L2, and what would ever justify L3) is Step 14's subject. Go build it now: 13a — Claude Code · 13b — OpenCode.

Check yourself

Q: While it's reading, the triage loop could trivially label the issues too — it would save you clicks. The design says don't. What justifies leaving obvious value on the table on day one?

Answer

Trust hasn't been earned yet — and labels are writes on a shared system other people see. Starting at L1 means the loop's judgment gets audited, through reports, across real mornings before it's ever handed a pair of hands. If the ranking logic is subtly wrong, you learn it from one wrong report, not from a hundred wrong labels. The saved clicks are the tuition. Promotion to labeling-at-L2 comes only after the reports have been boringly correct for a while.

Try With AI

Before you open either walkthrough, fill in the six-part table above for your repo: the real ranking rules (what actually counts as urgent here?), the real report destination, the real cap. Then ask your agent to red-team that design against the 7-item checklist. Every mismatch it turns up now is a 7 am surprise it just spared you.

When it goes wrong

---

Glossary terms used on this page: minimum-safe checklist, L1 (report-only), routine, human gate — see the glossary.

Sources: the morning-triage build and the minimum-safe checklist come from Panaversity's Loop Engineering: A Crash Course (S1). Full attribution: resources/sources.md.