factory

factory

Derive each unit's state from tracker issues, PRs, checks, reviews and deploy evidence; verify its exit gate and name the next skill. Keep no state file.

Inputs

• Mode — report (default) or run, from the first argument. Enter run only when the user's own request says it; a status question never does.
• Reference — one of: a spec (PRD) issue (PRWL-100, a GitHub issue URL or number), a ticket, a PR number or URL, or the word start for an idea with no spec yet. run also takes a label name (/factory run automate): every open issue carrying it. Missing → ask; never guess from the current branch.
• Tracker of record — Linear (via Linear MCP) or GitHub (via gh), read from the workspace AGENTS.md. Unclear → ask.
• Staging branch — default staging; confirm with git ls-remote --heads origin <branch> before any state that depends on it.

Rules

• Report mode is read only. It reads the tracker, gh, and git. It writes nothing remote and edits no file; every state change belongs to the skill it names.
• Report mode suggests, never auto-chains. Name the next skill with the exact invocation, then stop. The user runs it. Only run mode, started by the user, carries a suggestion out.
• Gates need evidence, not status words. A gate passes only on the observable signal in references/states.md: required CI evidence for head (or authoritative no-applicable-CI policy with mandatory local proof), head-bound QA/risk evidence, and deploy-watch proof for the merge commit. A label, a chat claim, or "CI is fine" does not pass a gate.
• Head SHA binds everything. Review, QA, and risk evidence counts only for the PR's current headRefOid; an approval counts only when its review's commit.oid is head. A new push sends the PR back to CI.
• Staging is the ceiling. READY_FOR_OWNER is terminal. Never suggest production deploys, production PRs/merges, or flag rollouts. Default branches are owner-only except explicitly authorized, non-production delivery local; name the owner otherwise.
• Unreadable signals stop. A signal the table needs can't be read (tracker access denied, no deploy workflow found) → report UNKNOWN with what is missing. Never guess the more advanced state. A failed or erroring gh query is unknown — never an empty result, a pass, or green; report the command and its error.
• Zero attribution. No co-author, AI, or tool attribution in any output, except Reviewer provenance in references/records.md.
• Emit Stage / Found / Next / Needs user at each phase transition — one line per field.

Workflow

1. Resolve the unit of work

From the reference, build the list of units to place: a spec expands to all its sub-issues, open and closed (state the count); a ticket or PR is one unit. For each ticket, find its PR: gh pr list --search "<issue-id>" --state all --json number,headRefName,baseRefName,state or the tracker's linked-PR field. Closed and merged tickets stay in the list — they may still be in STAGING. A qa-escape marker other than not-deployed places its ticket by the newest later PR.

2. Read the signals

Per unit, read only what the state table needs. Before interpreting checks, apply references/ci-policy.md; it distinguishes known no-CI policy from pending, failed, absent required checks and unreadable policy:

gh pr view <pr> --json state,isDraft,baseRefName,headRefOid,mergeCommit,mergedAt,reviewDecision,reviews,statusCheckRollup,autoMergeRequest,body,comments,commits
gh pr diff <pr> --name-only
gh run list --workflow <deploy-workflow> --branch <staging-branch> --json headSha,status,conclusion,url
gh issue view <issue> --json state,comments

Read unresolved review threads with a paginated GraphQL reviewThreads query; unreadable thread state is UNKNOWN before a merge.

Read QA/risk/deploy evidence on the PR and issue records through references/records.md. Linear comments use readable headings and fields; legacy records remain readable. Newest valid evidence for the required revision wins. The Acceptance criteria verdict clearing BUILDING must name head: read the body, otherwise the newest ## QA handoff whose Change: SHA is head. Old-head acceptance does not count.

3. Place each unit

Match the signals against the table in references/states.md, top row first — the first row whose condition holds is the state. States in order: OUTCOME, TICKETS, QA_RETURNED, BUILDING, CI_STUCK, CI, QA_STUCK, CHANGES, QA, REVIEW, HUMAN_REVIEW, MERGE, READY_FOR_OWNER, DEPLOY_FAILED, STAGING, plus UNKNOWN.

4. Check the gate and pick one next skill

For each unit, state the gate out of its current state and whether its evidence is present. Pick exactly one next skill per unit from the table's next column. When that skill runs or changes code, name the worktree it will use — <project-repo>/.worktrees/pr-<n> for a PR, .worktrees/qa-escape-<issue> for an escape. For a spec with several units, order the suggestions so a blocked unit (QA_RETURNED, DEPLOY_FAILED, CI_STUCK, QA_STUCK, CHANGES, UNKNOWN) comes before progress on healthy ones; healthy units follow oldest Since first, and a unit in one state for more than 3 days is flagged stalled. Units that share a gate and skill collapse into one suggestion (for example, /orchestrate-herdr for every BUILDING ticket of one spec).

5. Report and stop

Print the output below. Report mode ends here: do not run the suggested skill, even when the user is away; the next move is theirs.

Run mode

The user's /factory run <reference> approves the whole chain. Read references/run.md first; its Stage map names the late-stage files (references/review-and-delivery.md, references/returned-by-qa.md) to read when a unit first reaches those states.

1. Intake. Resolve units as in step 1 and apply Triage before builder/runtime work. A fully specified ticket may receive its narrow eligibility transition under run authority; ambiguous work parks. Label intake retains its reporter-follow-up path. Missing criteria restating existing outcomes are posted as a comment. Run the checks in Start, then ask its one batch:

   • "Which agent and model builds?"
   • "Which agent and model reviews the code? A different provider than the builder is recommended."

   User away → stop. After this batch the run asks nothing more.

2. Loop. Follow Dependency-aware scheduling: isolated independent units may overlap within capacity; the parent serializes integration. For each ready unit: place it (steps 2–4), carry out its Next per Stage map, and place it again; log it and apply Lessons. Per unit: status In Progress, its own worktree and Task stack, a builder thread, agentic QA in a QA worker, a PR, CI, the Review loop between the reviewer and the builder, risk review, Merge, Handover with status In Review, Cleanup. For setup or local-runtime blockers, read references/readiness-and-recovery.md; Loading a stage skill governs recovery and stops.

3. Parks. A unit whose next actor is a person — a high risk tier, a spent cap, a fact only the reporter holds — is parked with its evidence; the run continues. No tracked compose file or issue body is edited; no issue is closed.

4. Run report. The output table, each parked unit's reason, the criteria drafted, and Lessons report.

Output

Factory — <reference>  (<n> units)

| Unit | PR | State | Since | Gate evidence | Next |
| ---- | -- | ----- | ----- | ------------- | ---- |
| #418 | #87 | CI | 2026-10-02 14:10 | run 9912 failed on a1b2c3d (unit-tests) | /ci-loop 87 |
| #419 | #88 | HUMAN_REVIEW | 2026-09-28 09:02 (stalled) | risk-review tier=high on 4e5f6a7 (migration) | engineer review — no skill |
| #420 | #85 | READY_FOR_OWNER | 2026-10-01 17:45 | deploy run 9890 success on merge 7c8d9e0, smoke passed | owner promotes — factory ends |

Escapes: 2 qa-escape on this spec — 1 after agentic-qa verified (honesty gap: #418 missed-empty-state)
Next: /ci-loop 87

At most one table row per unit; Gate evidence names the run, SHA, or comment — never a bare "pass". Since is the timestamp of the signal that placed the unit (head push, marker, run), read this run. The closing Next: line carries the single most urgent suggestion. When a unit is UNKNOWN, its row says which signal is missing. The Escapes: line counts the spec's reproduced qa-escape markers, excluding class not-a-regression, and how many name a PR (pr=) whose final head carried a verified agentic-qa marker — that count is the honesty gap; name each one.

Completion criteria

• Every unit in the spec appears in the table, and the printed unit count matches the total sub-issue count, open and closed, read from the tracker
• Every Gate evidence cell names a run ID, SHA, or comment marker that was actually read this run
• Report mode: git status and the PRs show no change made by this run, and no next skill was invoked
• Run mode: readiness/repair evidence names checkout/runtime, causes, changed conditions, and rechecks within all caps
• Run mode: every unit ends merged with its delivery record read back, or parked with a named reason; cleanup read-back shows no task runtime resource or /tmp/<task-id>/ left, or names each one; no issue was closed and no tracked Compose file changed