Skip to content
factory
Esc
↑↓navigate↵open⌘Jpreview
On this page

run

Drive the whole loop for one change at a chosen size, stop only at the human gates, and resume from disk.

What it does

/factory:run <size> <request or spec path> sequences the stages a change of that size needs, as sized on When to reach for factory: small is verify only; medium is spec, plan, implement, simplify, review, the review fixes, pr, and a review of the PR; large is the same with implement --all, simplify, review, and fixes per slice, then pr --stack. Every stage is an existing skill called through the Skill tool. The run skill adds no method, only the order and a state file.

It runs only when the user starts it, because it reaches /factory:pr. /factory:run with no arguments resumes the open run.

Reads

  • factory.toml for spec_dir.
  • <spec_dir>/<slug>/run.md, the only state it keeps.
  • Whatever each stage left on disk: spec.md status, plan.md and plan-check --stack, .verify/green, the review-NN.md boxes, the PR URLs. Nothing about the position lives in the conversation.

Writes

<spec_dir>/<slug>/run.md with Size, Stage, Slice, Waiting on, and one log line per stage with the evidence that proves it. The file is committed with the spec directory.

The human’s gates

  1. Spec answers. The run writes Waiting on: spec answers, asks, and ends the turn. Answer in that session or a later one; the spec skill applies the answers and the next /factory:run continues.
  2. Plan approval. The same, with Waiting on: plan approval.
  3. Starting the PRs. The same, with Waiting on: pr.

Everything between those gates is the agent’s. A failed stage keeps its Stage value and a log line saying what failed; the next /factory:run retries it.

Done when

run.md says Stage: done, every stage of the size has a log line with evidence, and every PR of the stack is open against its parent. The reply lists the spec path, the slices and their PRs, the review findings left unchecked, and the evidence paths under .verify/slices/.

Upstream

Wraps mattpocock-skills/implement-spec for the frontier and the sparse, pointer-based handoff between stages, and the cursor-plugins/pstack multi-phase plan playbook for one PR per unit with its own evidence. Upstream’s integration branch and merger agent become the stack that implement --all builds and pr --stack publishes; upstream’s single review fix becomes implement <NN> --from-review per slice.

Last updated on

Was this page helpful?