Enforcing AI coding patches with machine-readable contracts and linters

On November 7, 2026, an article appeared on DEV Community, in the publication's AI section, under the headline "Constraining AI Coding Patches with Small Machine-Readable Contracts and a Linter." The piece argues that a test strategy for agent patches dies the moment it meets a real diff. Such strategies live on a wiki page, the agent never reads them, and a reviewer working under time pressure approves a green build anyway.

The author's position is that the answer is not a better document. It is a small machine-readable contract that ships inside the patch, paired with a linter that rejects the patch when the contract is too thin. The mechanism works by forcing test case labels and a blast-radius rating to be recorded as data rather than prose, then letting an objective string comparison stand in for subjective review, so that test completeness is enforced mechanically instead of being negotiated during review.

As described, the workflow consists of one JSON contract per patch, four linter rules, and a blast-radius table that determines how much checking a given patch has earned. Each patch carries its own contract, the linter evaluates it against the rules, and the severity table decides the level of verification required. The article presents this as the workflow the author currently uses.

For now the proposal exists only as an individual author's write-up on DEV Community. No follow-up coverage has appeared describing adoption of the scheme in real projects, and the article itself offers no measured results. The status of the idea therefore rests at the level of a published proposal awaiting independent testing.