Software teams keep their conventions in prose — style guides, spec documents, the comment at the top of the tricky file, the shared understanding that this list stays in sync with that one. This works because a human developer reads the prose, remembers it, and feels the friction when a change violates it.
Coding agents break that mechanism. An agent session is stateless: it does not reliably re-read the convention it violated last week, because it was not there last week. It restates a spec in a comment and neither notices when the two diverge. It adds the ninth item to a list and forgets the registry that was supposed to grow with it. None of this is incompetence — it is the predictable result of asking a fresh reader, every time, to hold context that was never written down mechanically. The drift is silent, and silence is the whole problem: nobody is signalled.
Checkwright’s premise is a division of labour. Some consistency questions are mechanically decidable and cheap to check — does this link resolve, does this command exist, does every kit have a registry row, was this stage actually entered. Those belong to a machine. What remains is the irreducibly semantic judgment — is this the right design, does this prose actually explain the thing — and that is where human (or agent) attention should go, undiluted.
So every cheap, low-false-positive, mechanically-decidable axis becomes a gate: a small program that scans a surface and blocks the commit when it finds a violation, naming the finding and the fix. The residue — the judgment a gate cannot make — is left to a reviewer who is freed from also playing linter.
A gate earns its place only when it is cheap, rarely wrong, and guards a real failure mode. A gate that cries wolf trains its readers to bypass it, so a false positive is treated as a defect in the gate, not a cost of doing business.
Checkwright is not a monolith. It is a set of kits, each owning one axis of the problem — the lint framework, the iteration lifecycle, the task queue, the spec discipline, permission friction, delegation, context economics, drift, and evidence. A kit is vendored into your repository whole and governs it from the inside. You adopt the ones that pay for themselves and leave the rest.
This repository governs itself with the kits from day one — the same gates that ship to you run on the tree that builds them. A methodology that its own authors will not run is a slide deck; this one is dogfooded or it is nothing.