Skip to content

Releases

v0.3.0 — real installation regressions and plugin context

Choose a tag to compare

@oh79 oh79 released this 21 Oct 03:06
d57d4d8

Three reported installations exposed resources that the comparison previously skipped. This release adds concrete inline docs/ and templates/ paths and makes undeclared ${CLAUDE_PLUGIN_ROOT} context a finding.

npx --yes agent-context-check@0.3.0 compare --source ./skill --artifact ./installed/skill

For complete plugin trees, declare a relative plugin location inside both roots:

npx --yes agent-context-check@0.3.0 compare \
  --source ./plugin-source --artifact ./plugin-copy \
  --plugin-root . skills/example/SKILL.md

The API accepts pluginRoot; the GitHub Action accepts plugin-root. Use oh79/agent-context-check@v0.3.0.

Validation replayed a documented copy and two actual pinned skills@1.5.25 GitHub installations, without deleting files from the outputs:

  • everyday-causal-skills: template coverage increases from 6 to 10 recognized references. Both versions fail with source-missing because shared resources are not skill-local.
  • Salesforce: five selected skills now report 10 unresolved plugin-context references instead of silently checking none. This does not inspect the runtime environment or fabricate missing-file flags.
  • ai-dotfiles: the real empty-docs installation now reports 75 artifact-missing references instead of a false clean.
  • Separate resource-complete controls pass for all three cases. They are local data copies, not upstream fixes or native plugin installations.

65 offline tests pass. The public report records exact SHAs, commands, source lines and limits; the attached JSON captures before/after reports and the final scanner source hash.

Full validation report.

The scanner remains offline, read-only and dependency-free. It does not run skill commands or agents, infer a shared parent resource root, evaluate dynamic shell paths, or prove successful agent behavior.

v0.2.0 — source/artifact comparison and GitHub Action

Choose a tag to compare

@oh79 oh79 released this 21 Oct 02:35
93a6445

Skill references can work in a repository and disappear from the package users receive. This release adds a static source/artifact comparison to catch that before release.

npx --yes agent-context-check@0.2.0 compare --source ./skill --artifact ./dist/skill
- uses: oh79/agent-context-check@v0.2.0
  with:
    source: ./skill
    artifact: ./dist/skill
  • Checks instruction files and supported Markdown, literal resource paths, and ${CLAUDE_SKILL_DIR} references. Reports file/line evidence and distinguishes source errors from artifact omissions.
  • GitHub Action uses Node.js 24 with no npm install; CLI supports Node.js 20+ and retains the original local link check.
  • Zero dependencies; scanner stays offline and read-only. Comparison checks existence and file/directory type, not content equality or dynamic shell behavior.
  • Validation used two actual npm skill trees and one pinned GitHub source archive: 3 instruction files, 41 supported references, no findings in the original artifacts. All 3 separate, deliberately damaged copies failed as expected. These are controlled checks, not upstream defect or adoption claims.

See the validation report for exact versions and provenance limits.

Skill maintainers: please try your own prepared package and report false positives or missed references in Issues.

v0.1.0 — Offline agent instruction link checks

Choose a tag to compare

@oh79 oh79 released this 21 Oct 02:10

Initial public release of agent-context-check.

  • Detect missing local Markdown link targets in AGENTS.md, CLAUDE.md, and SKILL.md, or explicitly selected Markdown files.
  • Report source files, line numbers, missing targets, and project-boundary escapes.
  • Support CLI text/JSON output and a JavaScript API.
  • Run offline and read-only on Node.js 20+, with no runtime dependencies or API keys.
  • Include English/Korean documentation, MIT license, examples, and 20 automated tests.

Validation: the initial CI matrix passed on Ubuntu, Windows, and macOS with Node.js 20, 22, and 24. A packed-tarball installation was also exercised locally.

Known limits: this is a focused Markdown-link parser, not a full CommonMark implementation. It does not check remote URLs, fragment IDs, bare paths, or whether an agent actually follows the instructions. See README for details.