skill-radar

Cross-agent skill health radar for AI coding agents. Hook-free usage scoring from real prompt history, content-hash dedup, bulk cleanup TUI.

  Health 46/100   ~3090 tokens/turn wasted
  vs last scan: health ↑10  ·  waste ↓146631  ·  dead →  ·  skills →

  52 skills  46 dead  3 dups  0 broken

Why

Every skill's description ships in the system prompt every turn — even skills you never invoke. A folder of 50+ skills is a quiet tax on every session. Worse: when the listing overflows its budget, the agent silently drops descriptions starting with the skills you use least, so dead skills get deader (a ratchet).

skill-radar finds the waste and helps you kill it in seconds — across all your AI coding agents at once.

What's different

| | skills-janitor | skills-admin (skm) | skills-cleaner | skill-radar | |---|---|---|---|---| | Cross-agent | Claude + Codex | 53 agents | Claude only | 16+ agents | | Dedup | name + desc Jaccard | name only | AI similarity (costly) | content-hash + symlink-aware | | Usage scoring | transcript scan (Claude) | none | requires hooks | prompt_history (Devin+Claude+Codex, no hook) | | Trend / history | no | no | no | yes — compounding | | TUI bulk delete | swipe (one-by-one) | no | no | list + D nuke-all-dead |

The moat: history compounding. Each scan appends a snapshot to ~/.config/skill-radar/history.jsonl. Over time you see your hygiene trend — and that trend is the switching cost another tool can't reconstruct.

Install

npm install -g skill-radar
# or run once:
npx skill-radar scan

Usage

# default: interactive TUI (list view, bulk delete, search/filter)
skill-radar

# first-run setup wizard (check agents, verify history, explain output)
skill-radar init

# one-shot health report (persists a snapshot for trend)
skill-radar scan
skill-radar scan --json    # JSON output for scripting/CI

# full waste ranking (all skills)
skill-radar score
skill-radar score --json

# show duplicates (content clones, name collisions, symlink aliases)
skill-radar dedup
skill-radar dedup --json

# structural issues (broken symlinks, missing SKILL.md, empty dirs)
skill-radar doctor
skill-radar doctor --fix         # auto-fix: remove broken symlinks + dedup
skill-radar doctor --fix --dry-run
skill-radar doctor --json

# scan history (the compounding trend)
skill-radar history
skill-radar history --json

# auto-cleanup: remove broken symlinks + duplicate copies (keep canonical)
skill-radar clean --backup       # back up before deleting
skill-radar clean --dead --backup  # also remove skills with no manual invoke
skill-radar clean --dry-run      # preview without deleting

# generate markdown report (shareable, for teams/docs/GitHub)
skill-radar report
skill-radar report -o report.md  # write to file

# one-shot auto-fix everything: backup + clean broken + dedup + remove dead
skill-radar autofix              # safe by default (backs up before deleting)
skill-radar autofix --dry-run    # preview without changing
skill-radar autofix --no-dead    # only fix broken + duplicates, keep dead skills

TUI keybindings

| key | action | |---|---| | ↑ ↓ / k j | move cursor (vim-style) | | PgUp PgDn / u | scroll a page | | g / G | jump to top / bottom | | i | inspect full description + metadata of cursor row | | d | delete all copies of cursor skill across every agent (confirm y/n) | | D | delete ALL dead skills + all their copies (the one-keystroke win) | | space | toggle mark | | x | delete all marked | | / | search/filter by name or description (Esc to clear) | | r | re-scan (score jumps live) | | q | quit |

The list is windowed (20 rows visible) with a scroll indicator showing 1-20 of 52 ↑↓ so every skill is reachable. No swipe — browsing one-by-one is slow. Sort worst-first, nuke the top with D.

How usage scoring works (without a hook)

Other tools need you to install a PreToolUse/PostToolUse hook before they can score usage. skill-radar reads prompt history that agents already log for free:

• Devin: ~/.local/share/devin/cli/sessions.db → prompt_history table (via sqlite3 CLI, no native dep)
• Claude Code: ~/.claude/history.jsonl → display field
• Codex: ~/.codex/history.jsonl → text field

It extracts /skill-name invocations and counts them + tracks last-used date. No hook, no setup, no background process.

Limitation (honest): this catches manual /skill-name invocations, not model-auto-triggered skills. A skill the model invokes on its own will show "no manual invoke" even if it's active. This is a weaker signal than transcript scanning (which skills-janitor does for Claude), but it works cross-agent and requires zero setup.

Agents scanned

Global roots: ~/.agents/skills, ~/.claude/skills, ~/.config/devin/skills, ~/.codex/skills, ~/.cursor/skills, ~/.codeium/windsurf/skills, ~/.gemini/skills, ~/.cline/skills, ~/.continue/skills, ~/.goose/skills, ~/.opencode/skills, ~/.roo/skills, ~/.crush/skills, ~/.kiro/skills, ~/.kimi-code/skills, ~/.commandcode/skills, and more. Honors CLAUDE_CONFIG_DIR, CODEX_HOME, XDG_CONFIG_HOME.

Plugin/marketplace discovery: Recursively walks ~/.claude/plugins/ and ~/.kimi-code/plugins/ (up to 4 levels deep) to find nested skills/ directories installed by plugin managers and marketplaces. No manual configuration needed.

Project-level roots: Scans .devin/skills, .agents/skills, .windsurf/skills, and .claude/skills in the current working directory, so project-scoped skills show up too.

License

MIT