feat: skill registry + Claude Code native discovery

[rtack]

Summary

• Skill registry (REGISTRY.json): machine-readable index of all skills, auto-generated from SKILL.md frontmatter via generate-registry.sh. Agent-system agnostic — works with Claude Code, ampcode, Cursor, or anything that reads JSON.
• Claude Code native discovery: install.sh now detects .claude/ directories and creates per-skill symlinks (.claude/skills/adapt/SKILL.md → .agents/skills/adapt/SKILL.md) so skills appear in Claude Code's / slash command menu natively.
• Optional SKILL.md frontmatter fields: triggers, keywords, invocation — used by the registry for programmatic discovery. Backward-compatible: Claude Code ignores unknown fields (verified experimentally).

What changed

Area	Files	What
Registry	REGISTRY.json, generate-registry.sh	Machine-readable skill index + generator script
SKILL.md	adapt/, ralph/, research/, tmux/	Added optional triggers, keywords, invocation frontmatter
install.sh	install.sh	setup_claude_code_integration() — creates .claude/skills/ symlinks, cleanup on --uninstall
Tests	test/integration/install.bats	10 new tests (39 total, all pass)
Docs	docs/skills.md, CHANGELOG.md	Registry format docs + Claude Code native discovery section

How Claude Code integration works

1. install.sh checks if .claude/ directory exists (auto-detect)
2. If yes, creates per-skill symlinks: .claude/skills/*/SKILL.md → ../../../.agents/skills/*/SKILL.md
3. Generates .claude/skills/.gitignore with explicit skill names (user-created skills stay tracked)
4. Symlinks are recreated on --sync, cleaned up on --uninstall
5. Falls back to file copy if symlinks aren't supported (Windows)

Verification

Tested in a controlled environment (~/dev/test-skill-verify) across 4 experiments:

Experiment	Question	Result
exp-2	Is .agents/skills/ natively discovered?	No — Claude Code only scans .claude/skills/
exp-3	Do custom frontmatter fields break loading?	No — IDE warnings are cosmetic only
exp-2b	Do symlinked skills appear in / menu?	Yes — all 4 skills discovered, loaded, executed
exp-4	End-to-end local integration?	Yes — Claude explored repo, installed, set up symlinks

Test plan

• bash -n install.sh — syntax check passes
• 39 bats tests pass (29 existing + 10 new Claude Code integration tests)
• Symlinks resolve correctly (verified readlink + cat through symlink)
• --dry-run shows integration plan without creating files
• --uninstall removes only dot-agents symlinks, preserves user skills
• Idempotent on re-run (symlinks overwritten via ln -sf)
• Non-symlink user files in .claude/skills/ are skipped with warning
• Verify on a fresh project without .claude/ directory (should skip silently)
• Test on Linux (symlink behavior should be identical)

🤖 Generated with Claude Code

[colmarius]

Thanks Raphael for opening the PR with support for claude code. Will get to review and integrate it sometimes ~mid week.

[colmarius]

@rtack took longer to get to this PR opened.

I added a minimalistic approach to supporting the .agents/skills into .claude/skills by symlinking them. Done in PR: #5

Also docs updated.

To be honest my workflows have changed, for example not using ralph anymore, and have blend of research/planning skill something I call agent-work.

I will do a revamp of the dot-agents repo with my new workflow sometime in the next few weeks. Publish the latest version of the workflow that works for me on a day-to-day basis working on a larger monorepo codebase.

[colmarius]

Closing the PR, as addressed in #5