skill-registry（技能登记处）

Purpose

You generate or update the skill registry — a catalog of all available skills with compact rules (pre-digested, 5-15 line summaries) that any delegator injects directly into sub-agent prompts. Sub-agents do NOT read the registry or individual SKILL.md files — they receive compact rules pre-resolved in their launch prompt.

This is the foundation of the Skill Resolver Protocol (see _shared/skill-resolver.md). The registry is built ONCE (expensive), then read cheaply at every delegation.

When to Run

• After installing or removing skills
• After setting up a new project
• When the user explicitly asks to update the registry
• As part of sdd-init (it calls this same logic)

What to Do

Step 1: Scan User Skills

1. Glob for */SKILL.md files across ALL known skill directories. Check every path below — scan ALL that exist, not just the first match: User-level (global skills): Project-level (workspace skills):

- ~/.claude/skills/ — Claude Code - ~/.config/opencode/skills/ — OpenCode - ~/.gemini/skills/ — Gemini CLI - ~/.cursor/skills/ — Cursor - ~/.copilot/skills/ — VS Code Copilot - The parent directory of this skill file (catch-all for any tool) - {project-root}/.claude/skills/ — Claude Code - {project-root}/.gemini/skills/ — Gemini CLI - {project-root}/.agent/skills/ — Antigravity (workspace) - {project-root}/skills/ — Generic

1. **SKIP sdd-* and _shared** — those are SDD workflow skills, not coding/task skills
2. Also SKIP skill-registry — that's this skill
3. Deduplicate — if the same skill name appears in multiple locations, keep the project-level version (more specific). If both are user-level, keep the first found.
4. For each skill found, read the full SKILL.md (if a SKILL.md exceeds 200 lines, focus on the frontmatter and Critical Patterns / Rules sections only) to extract:

- name field (from frontmatter) - description field → extract the trigger text (after "Trigger:" in the description) - Compact rules — the actionable patterns and constraints (see Step 1b)

1. Build a table of: Trigger | Skill Name | Full Path

Step 1b: Generate Compact Rules

For each skill found in Step 1, generate a compact rules block (5-15 lines max) containing ONLY:

• Actionable rules and constraints ("do X", "never Y", "prefer Z over W")
• Key patterns with one-line examples where critical
• Breaking changes or gotchas that would cause bugs if missed

DO NOT include: purpose/motivation, when-to-use, full code examples, installation steps, or anything the sub-agent doesn't need to APPLY the skill.

Format per skill:

### {skill-name}
- Rule 1
- Rule 2
- ...

Example — compact rules for a React 19 skill:

### react-19
- No useMemo/useCallback — React Compiler handles memoization automatically
- use() hook for promises/context, replaces useEffect for data fetching
- Server Components by default, add 'use client' only for interactivity/hooks
- ref is a regular prop — no forwardRef needed
- Actions: use useActionState for form mutations, useOptimistic for optimistic UI
- Metadata: export metadata object from page/layout, no <Head> component

The compact rules are the MOST IMPORTANT output of this skill. They are what sub-agents actually receive. Invest time making them accurate and concise.

Step 2: Scan Project Conventions

1. Check the project root for convention files. Look for:

- agents.md or AGENTS.md - CLAUDE.md (only project-level, not ~/.claude/CLAUDE.md) - .cursorrules - GEMINI.md - copilot-instructions.md

1. If an index file is found (e.g., agents.md, AGENTS.md): READ its contents and extract all referenced file paths. These index files typically list project conventions with paths — extract every referenced path and include it in the registry table alongside the index file itself.
2. For non-index files (.cursorrules, CLAUDE.md, etc.): record the file directly.
3. The final table should include the index file AND all paths it references — zero extra hops for sub-agents.

Step 3: Write the Registry

Build the registry markdown:

# Skill Registry

**Delegator use only.** Any agent that launches sub-agents reads this registry to resolve compact rules, then injects them directly into sub-agent prompts. Sub-agents do NOT read this registry or individual SKILL.md files.

See `_shared/skill-resolver.md` for the full resolution protocol.

## User Skills

| Trigger | Skill | Path |
|---------|-------|------|
| {trigger from frontmatter} | {skill name} | {full path to SKILL.md} |
| ... | ... | ... |

## Compact Rules

Pre-digested rules per skill. Delegators copy matching blocks into sub-agent prompts as `## Project Standards (auto-resolved)`.

### {skill-name-1}
- Rule 1
- Rule 2
- ...

### {skill-name-2}
- Rule 1
- Rule 2
- ...

{repeat for each skill}

## Project Conventions

| File | Path | Notes |
|------|------|-------|
| {index file} | {path} | Index — references files below |
| {referenced file} | {extracted path} | Referenced by {index file} |
| {standalone file} | {path} | |

Read the convention files listed above for project-specific patterns and rules. All referenced paths have been extracted — no need to read index files to discover more.

Step 4: Persist the Registry

This step is MANDATORY — do NOT skip it.

A. Always write the file (guaranteed availability):

Create the .atl/ directory in the project root if it doesn't exist, then write:

.atl/skill-registry.md

B. If engram is available, also save to engram (cross-session bonus):

mem_save(
  title: "skill-registry",
  topic_key: "skill-registry",
  type: "config",
  project: "{project}",
  content: "{registry markdown from Step 3}"
)

topic_key ensures upserts — running again updates the same observation.

Step 5: Return Summary

## Skill Registry Updated

**Project**: {project name}
**Location**: .atl/skill-registry.md
**Engram**: {saved / not available}

### User Skills Found
| Skill | Trigger |
|-------|---------|
| {name} | {trigger} |
| ... | ... |

### Project Conventions Found
| File | Path |
|------|------|
| {file} | {path} |

### Next Steps
The orchestrator reads this registry once per session and passes pre-resolved skill paths to sub-agents via their launch prompts.
To update after installing/removing skills, run this again.

Rules

• ALWAYS write .atl/skill-registry.md regardless of any SDD persistence mode
• ALWAYS save to engram if the mem_save tool is available
• SKIP sdd-*, _shared, and skill-registry directories when scanning
• Read SKILL.md files (respecting the 200-line guard in Step 1) to generate accurate compact rules — this is a build-time cost, not a runtime cost
• Compact rules MUST be 5-15 lines per skill — concise, actionable, no fluff
• Include ALL convention index files found (not just the first)
• If no skills or conventions are found, write an empty registry (so sub-agents don't waste time searching)
• Add .atl/ to the project's .gitignore if it exists and .atl is not already listed