claude-health

name	claude-health
description	Audit the health of your Claude Code configuration across all six layers — permissions, hooks, MCP servers, installed skills, agents, settings conflicts. Use when asked to "/health", "check my Claude Code config", "audit my setup", or before debugging unexpected behavior (silent permission denials, MCP tool not appearing, hooks not firing).
license	MIT
metadata	{"author":"tw93","source":"https://github.com/tw93/claude-health","source-version":"2026-05-15","maintainer":"izo","argument-hint":"[--layer <permissions|hooks|mcp|skills|agents|settings>]"}

Claude Health

Six-layer config audit for Claude Code. Detects:

1. Permissions — over-broad or contradictory allow / deny rules in ~/.claude/settings.json and .claude/settings.local.json.
2. Hooks — registered but missing scripts, scripts without execute bit, conflicting matchers.
3. MCP servers — declared but unreachable, schema mismatches, expired tokens.
4. Skills — installed but never invoked, stale skills (no recent activity), duplicate skill names.
5. Agents — frontmatter issues (missing name, invalid model), orphaned references.
6. Settings — env var conflicts, deprecated keys, version mismatches.

Output: prioritized fix list with file:line references.

Invocation

/health — full six-layer audit. /health --layer permissions — single layer.

Coexistence with ulk

ulk has two adjacent capabilities:

• /context-audit (ulk bundled skill) — health score 0-100, focused on token waste (MCP inutiles, CLAUDE.md bloaté, skills surdimensionnées, settings manquants).
• gandalf (34) — session-time hygiene watcher (5 context rules, drift detection).

/claude-health is complementary:

• gandalf (34) — runtime, during a session, watches the 5 hygiene rules + drift.
• /context-audit — token-cost focused, gives a score and ranking.
• /claude-health — broader config audit across six layers, surfaces wiring bugs invisible to the runtime watcher (a hook that never fires because the script is missing, an MCP that returns 401, a skill that's installed but never matched).

Routing rule:

• "Why isn't this hook firing?" → /claude-health --layer hooks
• "My session feels slow" → gandalf + /context-audit
• "Full pre-mission diagnostic" → run all three in sequence

Installation

Bundled in framework/community-skills/claude-health/. Upstream ships with a binary detection script:

# Upstream restructuré en suite multi-skills (Waza) — la skill /health vit dans skills/health/ :
curl.md https://raw.githubusercontent.com/tw93/claude-health/main/skills/health/SKILL.md

Hint d'invocation depuis godspeed (00) en phase reconnaissance et gandalf (34) en complément du health check session.