autopilot-triage

Autopilot Triage

Vet a pool of open issues against the autonomy rubric and, after your confirmation, tag the ones that qualify with the autopilot-queued label. /autopilot-batch later reads that label and runs the batch. This is the vet half of the triage → run split; the two are decoupled on purpose — triage once (cheap, start-of-day), run the batch whenever and from wherever.

The label is membership only: it means "vetted and pending," nothing more. Build model and tier are decided at run time (model by rubric, tier by your --merge opt-in), never encoded in the label.

Where this runs

Run from the target application repository — you need the issues and the code, because vetting scope means looking at what a fix would actually touch. Triage is read-only except for the label writes it makes after your confirm.

Arguments

• (no args) — vet all open issues.
• [label-or-filter] — narrow the candidate pool, e.g. a phase label or chore / fix. Passed through to gh issue list --label.

The autonomy rubric

An issue qualifies for autopilot when all hold:

• Complete spec — clear acceptance criteria; you could hand it off with no back-and-forth.
• Bounded scope — a few files, one subsystem, a known pattern to follow (the positive exemplar: a "localize these hardcoded strings" i18n issue).
• No product / UX / design decision — nothing that needs your judgment about what to build.
• Testable — its resolution can be proven by the suite (or a targeted inline check).

An issue is disqualified when it trips any of /autopilot's own escape-hatch triggers — the same guardrails, applied before the run instead of during it:

• a new DB column / migration, a changed model association, an altered public route, a new dependency, or an event-status-lifecycle change;
• a thin or ambiguous spec, or one that conflicts with the code;
• multiple subsystems, a real design choice, or anything you'd want to review mid-flight.

When a candidate is borderline, disqualify it — a false negative costs you one manual issue; a false positive burns an unattended run on something that should have stopped. State a one-line reason for every skip; those reasons are useful signal.

For each qualifying issue, also note — as a preview, not a commitment:

• Build model — Opus for data-model / security / ambiguous / cross-cutting; Sonnet for bounded pattern-work (i18n / views / config / a single test / CRUD / docs). /autopilot-batch re-derives this at run time from the same rubric; your preview is so you can eyeball and adjust before the run.
• Merge-tier candidate? — flag issues that look safe for --to merge (confined to config / locales / views / copy; no migration / route / dependency). These are suggestions you opt into with --merge at run time; the batch defaults everything to --to pr.

Your task

Step 1 — Reconcile (self-healing)

Before proposing anything, clean the existing queue so an interrupted prior run can't leave stale membership:

• gh issue list --label autopilot-queued --state open --json number,title.

• For each, check whether a PR already exists for it — a branch named for the issue:

  gh pr list --state all --json number,headRefName,state \
    --jq '.[] | select(.headRefName | test("gh-<n>-"))'
  

• If a PR exists (open or merged), the issue is in-flight or done, not pending — strip the label: gh issue edit <n> --remove-label autopilot-queued.

• Report what was reconciled (or "queue clean — nothing to reconcile").

Step 2 — Gather candidates

• gh issue list --state open [--label <filter>] --json number,title,labels,body.
• Drop any already labeled autopilot-queued (already vetted) or already carrying a PR (from Step 1's check).

Step 3 — Vet

Assess each candidate against the rubric — read the issue body and, where scope is unclear, the code a fix would touch. Assign a verdict (queue / skip), a one-line why, and for queue candidates the model preview + merge-tier flag.

Step 4 — Present + CONFIRM GATE

Present the triage table — proposed queue and skips together, so the exclusions are visible:

Queue:
  #851  Sonnet  pr     localize checkout flash messages — complete spec, i18n pattern
  #863  Sonnet  merge? copy tweak on the about page — config/copy only, no code paths
  #847  Opus    pr     recompute shipment weight rounding — bounded but touches money math

Skip:
  #870  needs a data-model decision (new column not in the schema)
  #872  spec ambiguous — two conflicting acceptance criteria

CONFIRM GATE: This is the human gate the lifecycle requires — the label is applied only after you confirm. Present the proposed queue and wait. You may trim, add, or adjust a model call. Nothing is labeled before you say go.

Step 5 — Apply the label

Only after confirmation:

• Ensure the label exists (first run bootstraps it): gh label list and look for autopilot-queued; if absent, create it:

  gh label create autopilot-queued \
    --description "Vetted by /autopilot-triage; pending an /autopilot-batch run" \
    --color 5319e7
  

• For each confirmed issue: gh issue edit <n> --add-label autopilot-queued.

Step 6 — Report + handoff

Summarize: how many were queued (list + model preview + which are merge-tier candidates), how many reconciled/removed, how many skipped (with reasons). Then the handoff:

Queue ready — N issues. Run it from this repo with /autopilot-batch. To authorize merge for the flagged candidates, add --merge <#,#>.

Important

• The confirm gate is a hard stop. The label is never applied speculatively — only issues you approve get queued (lifecycle step 1).
• Membership only. The label encodes neither model nor tier; those are run-time decisions. Do not add per-issue tier/model labels.
• Reconcile first, always. Every run self-heals the queue before proposing, so a crashed or interrupted prior run leaves no stale membership.
• Decoupled by design. Triage and run are separate sessions on purpose — vet cheaply now, run the batch later / from anywhere. Do not fold the batch run into this skill.
• When borderline, skip. Autopilot's value is unattended throughput on issues that genuinely don't need you; a wrongly-queued issue that should have stopped defeats that.