Agent Skills: Test Runner

Test Runner

This project is a Claude Code plugins marketplace. "Tests" here means verifying the repository structure, version consistency, JSON syntax, and SKILL.md/agent frontmatter validity — equivalent to the /verify-plugins command logic.

Process

1. Determine test scope:
   • If $ARGUMENTS contains --base-commit <sha>: run git diff --name-only <sha> to get changed files (includes committed, staged, and unstaged changes)
   • Otherwise: run git diff --name-only HEAD to detect changed files (if HEAD is unavailable or no changed files detected, run full verification)
   • This marketplace is small and structural; any change touching .claude-plugin/marketplace.json, plugins/**, skills/**, or .claude-plugin/plugin.json should trigger full verification. If only unrelated files (e.g. README.md, CHANGELOG.md, docs/**) changed, still run full verification — it is fast
2. Spawn a subagent (Agent tool, subagent_type: general-purpose, model: sonnet) to execute verification — pass sonnet as the Agent tool's model parameter. The structural checks (jq / readlink / frontmatter validation) are mechanical, so sonnet is sufficient by default; this is a deliberate skill-side cost choice.
3. Return the subagent's structured summary to the caller

Subagent Instructions

Verify this Claude Code plugins marketplace repository. Perform the following checks yourself, directly, in this agent, using Read/Bash/Glob — do not spawn a nested Agent and do not dispatch any Skill(). Re-delegation is an observed failure mode: the executor stops to wait on its child and returns no verdict, which the caller then treats as a parse failure. Return a structured summary.

Checks

1. Load marketplace manifest: Read .claude-plugin/marketplace.json. Extract the plugins array. For each plugin, capture name, source, skills (optional), and version.

2. Skill entity existence: Verify each skills/*/ directory contains SKILL.md.

3. Plugin source directory structure: For each plugin in marketplace.json, dispatch by source prefix:

   • If source starts with ./skills/ (direct-skill plugin): verify <source>/SKILL.md exists and skills: ["./"] is present
   • If source starts with ./plugins/ (wrapper plugin): verify <source>/ exists; if <source>/skills/ exists, each entry under it must be a symlink (use readlink) resolving to an existing skills/<skill>/SKILL.md; if <source>/agents/ exists, each .md file must have YAML frontmatter; if <source>/.claude-plugin/plugin.json exists, its JSON must be valid
   • Additionally, for wrapper bundles (wrapper plugin with skills array of specific paths like ./skills/<name>): verify each path in skills array resolves to an existing skills/<name>/SKILL.md, AND verify the set of paths matches the set of symlinks under <source>/skills/ (each symlink has a corresponding skills entry, and vice versa — detect drift in either direction)

4. Version consistency: For each plugin, if <source>/.claude-plugin/plugin.json exists (only possible for ./plugins/ sources), verify its version matches marketplace.json.

5. JSON syntax: Validate .claude-plugin/marketplace.json and every plugins/*/.claude-plugin/plugin.json with jq empty.

6. Frontmatter presence: Verify each skills/*/SKILL.md and each agent file (plugins/*/agents/*.md) starts with --- on the first line (YAML frontmatter).

Return Format

Return a structured summary with one of three statuses:

Status: SUCCESS

• All checks passed
• Per-check summary: counts (e.g., "15 skills verified, 13 plugins, 0 version mismatches")

Status: TEST_FAILED

• Per-check results with failures highlighted
• For each failure:
  • The specific check that failed
  • File path and what was expected vs actual
  • Remediation hint (e.g., "recreate symlink", "bump version in plugin.json")
• Keep the summary concise but include enough detail to fix without re-running

Status: EXECUTION_ERROR

• Command that failed to execute (e.g., jq missing, marketplace.json unreadable)
• Error output
• This status is for infrastructure/environment errors, not verification failures