cot-lint

Your coding agent writes great code, but it also drops a design-session number, a justification addressed to a reviewer who has already left, and a let-me-fix-it-later aside straight into your docs, comments, and JSDoc. None of that harms the code itself, yet it makes the text unreadable to anyone who was not in the room. cot-lint exists for exactly this: it scans repository text and applies a single test to every passage—can a reader who has only HEAD resolve every reference and verify every claim?

In terms of core capabilities, cot-lint covers eight leakage classes: dead design-session citations, PR vantage, change narration, review choreography, reviewer-addressed justification, control-flow narration, hedge and planning residue, and authoring-language slips. Both English and Chinese detectors are built in, with zero external dependencies and a Node 20 minimum. Exit codes 0, 1, and 2 map to clean, findings, and usage error respectively, so it drops straight into any CI pipeline. A built-in set of keep-rules mechanically exempts issue references, TODO/FIXME/XXX markers, and citations of external standards such as RFCs, so the linter does not destroy annotations that genuinely earn their place. The companion cot-trim skill goes beyond detection: it judges every hit against the one test, enumerates the passage propositions before deleting anything, and fixes owner-first—generated files via their source, model-visible strings via their owning snapshot.

Who is it for? Any team that relies on AI coding agents, any repository whose documentation should be self-sufficient at HEAD, and any project that wants a prose-hygiene gate in CI. It is language-agnostic, agent-agnostic, MIT-licensed, and ready to use the moment you run it.