skill-map

argument-hint	[--skill NAME] [--root PATH | --portfolio-root PATH] [--format text|json|dot] [--include-catalog-sources] [--include-self] [--include-snippets] [--show-skipped]
coordination	exempt
disable-model-invocation	false
name	skill-map
user-invocable	true
description	Use to find agent skill installs, repository skill portfolios, duplicate skills, cross-dependencies, invocations, and cross-references across the local machine.

Skill Map

This skill is coordination-exempt: skip the ai-coord gate for its declared work.

Find skill installs and references across local files without scanning macOS protected home paths or obvious transcript, cache, dependency, and backup noise.

Arguments

• --root PATH: Scan this root. Repeatable. Default: ~.
• --portfolio-root PATH: Resolve PATH to its Git root, then scan that repository plus existing ~/.agents/skills and ~/.claude/skills. Mutually exclusive with --root.
• --skill NAME: Restrict the report to one skill name. Repeatable.
• --format text|json|dot: Select report format. Default: text.
• --include-catalog-sources: Include known local source checkouts such as ~/projects/agent-skills, ~/sablier/sablier-skills, and ~/sablier/agent-skills during broad scans. Explicit --root values inside those trees are always scanned.
• --include-self: Include self-references in dependency output.
• --include-snippets: Include matched reference text when exact lines materially improve the result. Default output omits snippets for concise, high-signal reports.
• --show-skipped: Include ignored path summaries in text or JSON output.

Workflow

1. Resolve the skill directory, then run the helper from that directory:

   uv run scripts/skill-map.py "$ARGUMENTS"
   

2. If no arguments were provided, run the default machine scan:

   uv run scripts/skill-map.py
   

3. Use --format json when another command or agent will consume the result.

4. Use --portfolio-root <repo> --format json when comparing repository skills with user-installed Codex and Claude Code exposures. Do not add broader home roots.

5. Use --format dot when the user asks for a graph, Graphviz input, or dependency visualization.

6. Use --include-snippets when exact matching lines materially improve the result, including when the user asks to see them.

7. Read references/ignore-policy.md only when explaining, auditing, or changing the ignore policy.

User-Facing Output

Keep JSON and DOT byte-valid and undecorated. For human output, lead with ### 🗺 Skill Map — <skills> skills · <duplicates> duplicates · <unresolved> unresolved and a compact skills/dependencies/duplicates/unresolved summary table. Always state the effective roots and material exclusions; for the default broad scan, explicitly say that standard agent homes and catalog source checkouts were excluded. Make a missing explicit --skill filter a visible ⚠️ Not found result rather than a clean-looking empty map. Use section labels sparingly and keep snippets, local paths, exact edges, commands, and diagnostics undecorated.

Output Semantics

• dependency: a skill declares another skill in skill-dependencies, or a skill file or support file references another discovered skill. Declared edges use their canonical bare or ORG/REPO#SKILL target and include declared: true; a declaration supersedes inferred prose evidence for the same source and target skill name.
• external-reference: a non-skill file references a discovered skill.
• duplicate-install: multiple discovered SKILL.md files declare or resolve to the same skill name.
• unresolved-like-reference: explicit $kebab-name or /kebab-name tokens that do not match a discovered skill.

JSON counts.declared_dependencies reports the number of emitted declared edges. Skill filters match declaration sources and target skill names, including the name after # for external identifiers. Invalid declaration fields stop the mapper with a path-specific parse error.

Portfolio JSON preserves those fields and adds:

• portfolio.repository_root and present/missing user roots, including the client exposed by each root.
• Per-skill lexical exposure_path beside resolved realpath; directory_name; repository/user location; install/catalog kind; and applicable clients.
• is_symlink and the lexical symlink_target for recognized skill-directory symlinks. Exposures that resolve to one real directory remain separate skill entries, while duplicate installs still require distinct real directories.
• skill_sha256 for SKILL.md bytes and tree_sha256 for the complete filtered skill tree. Tree hashes cover sorted relative paths, entry types, regular-file executable bits, streamed file bytes, and un-followed symlink targets.

The automatically selected user roots behave like explicit roots: the broad-home ignore policy does not suppress an explicit ~/.agents/skills or ~/.claude/skills root. Portfolio traversal follows symlinks only when the symlink is a direct child of a recognized repository or user skill root; other repository symlinks remain untraversed.

Related Skills

• skill-map only locates and cross-references skills; it does not validate them. To audit a catalog or installed root for metadata and doc-link issues, use the skill-doctor skill when it is installed.
• To evaluate the mapped repository-centered portfolio for conflicts, drift, consolidation, or missing workflows, use the manually invoked skill-harmonization skill.

Guard Rails

• Do not search transcript or backup directories manually after the helper excludes them unless the user explicitly requests transcript/history analysis.
• Do not broaden home-directory scans into macOS protected paths such as ~/Library or ~/.Trash; pass narrower explicit roots instead when a broad scan needs more coverage.
• Treat local skill catalog source checkouts as false positives during broad machine scans; pass them explicitly as --root when auditing catalog contents.
• Keep reports high-signal: include snippets only when exact evidence materially helps or the user explicitly requests it. Before copying output into a public or third-party artifact, perform an external-disclosure review.
• Prefer adding ignore rules in the helper and documenting the rationale in references/ignore-policy.md instead of ad hoc shell filters.