Factory

https://ai.codecot.com/factory-sheet.html ↗

AI-native control plane for personal projects. Specs in, PRs out — with an inspectable substrate of goals, specs, decisions, and accumulated context underneath.

Factory is the working pattern the rest of the practice grew from. An operator writes a spec; the system runs a coding agent against a clone of the target repo, opens a PR with the result, and surfaces the per-task log, artefacts, and goal-level progress in a project UI.

The substrate — what’s open, what’s decided, what’s shipped, what’s parked — is first-class. Specs reference goals; goals reference shipped commits; the project page shows live percent based on what acceptance criteria are met. The same primitive underlies AttentionRouter (inbound counterpart) and WhiteRabbit (generation studio).

What it is now. The numbers are published on a page of their own, the factory on one page, and read off it rather than restated here: 440 pull requests opened by the factory, 420 merged, and 418 of those 420 merged by a human clicking the button. 794 specifications written. 1,686 agent rounds, of which 410 were fix rounds after a review, a failing check or a conflict, and 35 were rounds that claimed the work was already done and were refused by a guard. Reviews ending in “merge”: 471 of 705.

The cockpit itself is private (it holds repository names and task logs), so the sheet is the public face, and it is built from the same database the cockpit reads.

Still dogfooded on itself: every Factory iteration is specced, PR’d, reviewed and merged through Factory. The review loop is its own current subject, measured first and respecified from what the measurement showed.