~/projects / factory

The factory

The thing that builds the other things. A pipeline of AI agents that picks up tasks, writes tests first, builds, reviews and ships, with a human in the loop wherever it counts.

why

One person cannot build and maintain five products by hand. So the products are built by a factory: a control repository that knows every project, its tests and its risks, and a runner that turns a task on a board into a reviewed, tested change.

how a task flows

$ runner --next
✓ pick      next unblocked task from the board
✓ plan      explore the repo, propose seams and vertical slices
✓ red-1     a tester writes one failing test
✓ green-1   a builder that may not touch tests makes it pass
✓ review    a separate model reviews the diff
✓ gate      full test suite against the project baseline
✓ merge     task branch into main
→ deploy    waiting for a human

test-driven by construction

The tester and the builder are different agents. The runner proves the test fails before any code is written and proves it passes afterwards. A builder that edits a test to make it green is stopped.

gates you can switch

Four human gates per project: approve the seams, review, merge and deploy. Each one can be switched off once the factory has earned the trust, with the numbers from the last thirty runs next to every switch. Some things never switch off: a red quality gate, a blocking review finding, a database backup before a migration, and the blast radius, a list of files no agent may touch without explicit approval.

a model per phase

Planning, testing, building and reviewing each get the model that fits the job, from different providers. A cockpit shows every run live, and a morning report sums up what happened overnight.

it learns

A mistake that happens twice becomes a rule, and every change to the rules is checked against an eval suite before it lands.

stack

• Python
• Claude
• Codex
• Git worktrees
• Docker
• Floqify API
• SQLite trace