Intent 0.5: Skills You Can Maintain

Intent 0.5 is trying to fix the part of agent skills that gets messy after the first release: keeping them accurate. A library changes, an export gets renamed, tests move on, and somewhere in a SKILL.md a stale instruction can still be telling an agent to use the old path. The new release is aimed squarely at that drift.

The big addition is a maintainer workflow that brings authoring, validation, and source review into the repository. You start with setup, which registers existing skills, adds the maintainer skills, and wires up CI. Existing skills are brought in without changing their contents. That matters, because Intent’s pitch has always been packaging guidance with the library version it describes, instead of leaving maintainers to assemble instructions from scattered places.

Validation got sharper too. Intent validate now checks fenced JavaScript and TypeScript examples in SKILL.md, including JSX and TSX, and it can also check JavaScript libraries using types inferred from source or declared through JSDoc. It uses the TypeScript compiler without running the code, and reports missing exports, invalid options, type mismatches, and syntax errors with file and line numbers. Deprecated imports show up as warnings. The check needs TypeScript 5.0 or newer, and if that isn’t available, or Intent can’t find the owning library’s entry point, it skips type checking instead.

The other half is review. intent maintainer review looks at source changes and flags guidance that may need updating, while also listing source changes that have no mapped guidance at all. You can confirm whether you updated the skill or explain why it still holds. Old decisions don’t automatically count for new changes, which is the right kind of annoying. A changed file isn’t the real question; whether someone checked the guidance against the change is.

There’s also a repair command for routine cleanup. It can fix clear frontmatter issues with --write, or produce a patch with --patch when metadata conflicts need a human hand. Suggested code-example changes still require review. On top of that, runtime libraries are now bundled, startup work is lighter, and agent hooks can use a local CLI instead of hitting the registry every session start.