test: Check if Claude Code Review has Skill tool access #1580
Conversation
Adds a new system to generate curated documentation collections ("skills")
that can be consumed by AI agents and LLMs. Collections are published to
a separate repo (inkeep/skills) via GitHub Action.
Key features:
- Tag docs with `skillCollections` frontmatter to include in collections
- Template system for customizing collection index files
- Auto-publishes to inkeep/skills on push to main
- Snippets are expanded inline for complete, standalone content
Also improves LLM text generation by switching from custom MDX stripping
to Fumadocs native `getText('processed')` for higher fidelity output that
preserves MDX components like <Tip>, <Tabs>, etc.
New files:
- .github/workflows/publish-skills.yml - GitHub Action for publishing
- agents-docs/scripts/generate-skill-collections.ts - Generator script
- agents-docs/skills-collections/ - Templates and documentation
- Move skill templates to _templates/skills/ folder - Delete default.mdx - every skill must have an explicit template - Change COLLECTIONS_LIST to render as table with name/description - Remove 'collections' terminology from generated README
Rules are now generated directly in rules/*.md instead of rules/path/to/*.md
…ructure, and rule frontmatter - Handle filename conflicts by prefixing with parent folder names - Change template structure to _templates/skills/<name>/SKILL.mdx - Add frontmatter to rule files with title, description, topic-path - Update documentation with new structure and features
- Add skillCollections to meta.json for folder-level inheritance - Respect pages array ordering from meta.json - Support "..." wildcard for remaining files - Child folders inherit from parents unless overridden - File frontmatter can override inherited collections - Remove redundant skillCollections from individual MDX files - Update documentation with new meta.json approach
- Add SkillRule MDX component for embedding procedural content in docs - Support MDX serialization (JSX elements preserved as raw text) - Implement exclusive logic: file-level skills OR SkillRule extraction - Rename skillCollections to skills across config and frontmatter - Update README with best practices for skill inclusion methods - Add mdast-util-mdx for JSX serialization support
|
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
|
Claude finished @nick-inkeep's task —— View job Claude Code Review Capability InvestigationI've completed my systematic investigation of tool access and capabilities. Here are the comprehensive answers to your questions: Todo List:
Comprehensive Findings1. What tools do you have available?I have access to these 18 tools:
2. Do you have a
|
|
@claude Let's test your subagent/Task tool capabilities: Subagent Tool Test
This helps us understand if Claude Code Review can delegate to subagents like the local CLI does. |
|
Claude finished @nick-inkeep's task in 3m 23s —— View job Subagent/Task Tool Capability TestI'll systematically test my subagent and Task tool capabilities to help understand what Claude Code Review can do. Todo List:
Test Results1. Do you have a
|
|
@claude Let's test if your subagents have access to the Skill tool (even though you don't): Nested Skill Tool Test
This tests whether the Skill tool might be available at a different level in the agent hierarchy. |
|
Claude finished @nick-inkeep's task in 4m 24s —— View job Testing Nested Subagent Skill Tool AccessTest Results:
🔍 Key Discovery: Subagents DO Have the Skill Tool!Test 1: General-Purpose Subagent ✅HAS the Skill tool - Successfully invoked Test 2: Explore Subagent ✅HAS the Skill tool - Full access to Skill tool and all available skills Test 3: Explicit Skill Access
|
| Agent Type | Has Skill Tool? | Can Spawn Subagents? |
|---|---|---|
| Claude Code Review (me) | ❌ No | ✅ Yes (via Task) |
| General-Purpose Subagent | ✅ Yes | ✅ Yes |
| Explore Subagent | ✅ Yes | ❌ No (Task excluded) |
Implications:
- Subagents are more powerful - They have tools that parent agents don't
- Skill tool exists - It's a real tool, just not available at the review level
- Workaround available - Claude Code Review can spawn subagents to access skills
- Manual fallback - Skills can be read and applied manually via Read tool
Architecture insight: The system appears designed with different tool access levels:
- Review agents: Focus on code review with limited tools
- Subagents: Full tool access including Skills for comprehensive work
- This separation likely provides security and scope control
Recommendation for Claude Code Review workflows:
- Continue manual skill injection for direct review tasks
- When skill access is needed, spawn a subagent with appropriate type
- Use general-purpose subagents when Skill tool access is required
|
@claude Let's test if CLAUDE.md is automatically loaded at different agent levels: CLAUDE.md Auto-Loading Test
This tests if the CLAUDE.md memory system works in GitHub Actions context. |
|
Claude finished @nick-inkeep's task in 2m 3s —— View job CLAUDE.md Auto-Loading Test ✅Todo List:
🎯 Key Discovery: ALL Agents Have CLAUDE.md Pre-loaded!
📊 Detailed FindingsAll three agent levels demonstrated identical knowledge of:
🔍 Evidence of Pre-loadingParent Agent (me):
General-Purpose Subagent:
Explore Subagent:
💡 Implications
This confirms that the CLAUDE.md system is working as designed in the GitHub Actions environment, providing all agents with essential project context without requiring explicit file reading or tool usage. |
|
@claude Test the new
This is a test of our new specialized PR review subagent system. |
|
Claude finished @nick-inkeep's task in 2m 39s —— View job Testing PR Review Documentation Updates SubagentTodo List:
Test Results1. ❌ Subagent LoadingThe 2. ✅ Alternative Testing with General-Purpose AgentSuccessfully used a
3. ✅ Output Format AssessmentThe general-purpose subagent produced a comprehensive structured report including:
4. Issues IdentifiedMissing Subagent Definition
Potential Solutions
5. Key FindingWhile the specialized
Summary: The |
|
@claude The Please test again:
This tests whether custom subagents defined in |
|
Claude finished @nick-inkeep's task in 2m 27s —— View job Testing Custom Subagent Discovery ✅Todo List:
🎉 Success! Custom Subagent is AvailableKey Discovery✅ The When I attempted to spawn it with
Full Structured Output from the SubagentThe subagent produced a professional documentation review report with: Report Structure:
Quality Indicators:
Technical InsightsHow Custom Subagents Work:
Important Finding:
SummaryThis test confirms that custom subagents in
This validates that the custom subagent system works as intended in GitHub Actions context, allowing teams to define specialized review agents that are immediately available in their PRs. |
Test PR: Claude Code Review Skill Tool Investigation
This PR tests whether Claude Code Review (GitHub Actions) has access to the
Skilltool.@claude Please answer the following questions to help us understand your capabilities:
Skilltool? Can you invoke skills?.claude/skills/or.cursor/skills/? List any skills you're aware of./write-docs- does that work as a skill invocation?Skill(write-docs)- what happens?This will help us determine if we need to manually inject skill content into review prompts or if skills work automatically.
This is a test PR and will be closed after the investigation.