Health Skill from tw93 Waza: Audit Your Claude and Codex Setup

Makes Claude audit your agent setup and your project for AI coding problems, such as ignored instructions, broken hooks or MCP servers, and missing checks.

Installs: 16.2K; on skills.sh
Stars: 7.5K; on GitHub
Files: 20; besides SKILL.md
Updated: Nov 2026; last repo commit

Health is one of the eight Waza skills by Tw93. It makes Claude audit the current project's agent setup (instruction files, permissions, hooks, MCP servers, skills and memory) and how maintainable the code is for AI coding. The report lists each problem with its evidence and a command to fix or diagnose it.

The audit is report-only. By default it runs a summary pass, and it asks you before a deep audit because a deep run can use a lot of token quota. Health is not for application bugs or PR review.

Good for

Claude Code and Codex users whose agent ignores their instructions
Developers tidying CLAUDE.md, AGENTS.md, hooks and MCP settings
Teams checking whether a repo is easy for an AI agent to change safely

Install health

Claude Code

Copy this prompt.

Install the health skill in Claude Code. Run this command:

/plugin marketplace add tw93/Waza
/plugin install waza@waza

Then check that it loaded and tell me how to use it.

2
Paste it into Claude Code. It shows the change and asks before running it.

health works on its own. SKILL.md routes a bundled debugging or review request to a matching skill or built-in capability, as an optional step. The whole-set routes install all 8 skills (think, ui, check, hunt, write, learn, read, health). Plugin skills are namespaced, for example /waza:health. Pick one route, not both. The per-skill plugin line needs the marketplace added first (the first command in the all-skills plugin route).

From the project README. Source.

What's in the skill

When it loads

Audits agent config, instruction drift, hooks or MCP, and AI maintainability. Use when Claude, Codex, or Pi setup looks wrong. Not for application bugs or PR review.

Sections in SKILL.md

Outcome Contract
Durable Context Preflight
Hard Rules
Step 0: Establish the evidence basis
Step 1: Collect data
Step 1b: Runtime Inventory and MCP Check
Step 1c: Safety and security checks
Security Baseline Checks
Memory and Skill Supply Chain
Long-Running Agent Stop Conditions
Step 2: Analyze
Gotchas
Output
[PASS] Passing checks (table, max 5 rows)
Finding format
[!] Critical -- fix now

Key facts

Author: Tw93 (Waza)
License: MIT
Installs on skills.sh: 16,190 (Nov 2, 2026)
GitHub stars: 7,488 for the whole tw93/waza repo (Nov 2, 2026)
Last commit: Nov 1, 2026
What is inside: 21 files including SKILL.md: 3 inspector agents, 3 reference guides and 14 scripts (Python, shell and PowerShell).
Part of: Waza, a set of 8 skills: think, ui, check, hunt, write, learn, read, health
Audits: Claude Code, Codex and Pi setups
Works in: Claude Code, Codex, Cursor and other agents, Claude Desktop (ZIP) and Pi (Waza README)

What it includes

Risk, non-obvious constraints, failure evidence and verifier coverage
A large file, many skills or a high TODO count counts only when tied to one of these four kinds of evidence.
Read · Step 0: Evidence basis
Runs scripts/collect-data.sh (or the PowerShell launcher) in summary mode
Gathers agent config, project signals and recent sessions into one output. Scope can be limited to instructions only.
Try: “/health check why Claude keeps ignoring my CLAUDE.md”
Runs code · Step 1: Collect data
Lists hooks and MCP sources and flags missing executables
Compares static config with live probes. It only calls a harmless tool on servers that are already connected.
Read · Step 1b: Runtime and MCP check
Deny-list floor, permission gating, env overrides, memory and skill supply chain
Flags missing denies for credential folders and pipe-to-shell, base-URL overrides, wildcard tool allowlists, skip-permission flags and unpinned third-party skills.
Read · Step 1c: Safety checks
Summary by default, deep audit with three inspector agents when escalated
Inspectors cover context and security, control and behavior, and AI maintainability.
Read · Step 2: Analyze
Constraint reachability, hotspots, verifier coverage, broken doc links, stale logs
Uses scripts such as check_maintainability.py, check_doc_refs.py and check_verifier_output.py.
Runs code · Maintainability lane
Passing table, then Critical, Structural and Incremental findings with an action
Each finding names the layer, the evidence and a copy-pasteable action. A clean run says nothing needs fixing.
Read · Report

From the official docs

Permissions

Medium risk

Health reads your agent settings, project files and recent chat history for this project, and runs bundled scripts to do it. It is report-only and should not write to your project.

What it can do

Read your data
Create and change things
Delete things
Read-only mode

Watch out for

It reads your past chat sessions
SKILL.md says a summary audit scans up to three recent sessions for the current project across Claude and Codex, and a deep audit streams all of them. Session text can contain private details, so scope the audit to instructions only if that matters. github.com
A deep audit uses a lot of tokens
Deep runs launch up to three inspector subagents, and SKILL.md says to tell you before escalating. Stay on the summary pass unless you need more. github.com
The Windows launcher bypasses execution policy for one run
SKILL.md uses PowerShell with -ExecutionPolicy Bypass for that process only, and says not to change your machine policy. Read the launcher before you allow it. github.com
Report-only by design
SKILL.md says summary and deep audits run only Health-owned collectors and read-only probes, and do not authorize project tests, builds or installers. github.com
Bundled scripts read local state
collect-data.sh, the Python checkers and conversation_audit.py read your agent config, project files, git state and session logs. Their code runs git queries (ls-files, log, rev-parse, check-ignore, remote get-url) and local checkers. It imports no network library and contains no package install or file deletion. Credentials are masked in output. github.com

Technical details

Scopes	Reads Claude, Codex and Pi config, project instruction files, hooks and MCP settings Reads up to three recent session logs for this project in a summary audit, and all of them in a deep audit Runs bundled bash and Python scripts (on Windows, a PowerShell launcher)

Questions

What does the health skill do?

It audits your agent setup and your project's AI maintainability: instruction drift between CLAUDE.md and AGENTS.md, permissions, hooks, MCP servers, skills, memory, verifier coverage and stale docs. It returns a prioritized report.

How do I install health?

Use the install box at the top of this page. The Waza README also documents npx skills add tw93/Waza -a claude-code codex cursor -g -y for all eight skills, a ZIP upload for Claude Desktop, and pi install npm:@tw93/waza for Pi.

Does it change my files?

No fixes are made by default. SKILL.md says audits are report-only and a neutral health request does not authorize tests, builds, installers or snapshot updates.

Does it read my conversations?

Yes, for the current project. SKILL.md says a summary audit scans up to three recent sessions across Claude and Codex, and a deep audit goes through all of them. Ask for an instructions-only audit to skip session history.

Is it safe to run?

It runs bundled scripts that read your config and project files, and it prints credentials only as [REDACTED]. Read what the scripts collect first if your sessions or settings hold private details.

When should I use health instead of check?

Use health when Claude or Codex behaves wrongly because of setup. SKILL.md says it is not for application bugs or PR review, which is what the check skill covers.

Information

updated Nov 2, 2026

Repository: tw93/waza
Folder: skills/health

tw93/waza skills.sh

Sources

Docs, issue trackers and discussions this page is built from.

Show all 5

github.com · github.com
marketplace.json · github.com
github.com · github.com
collect-data.sh · github.com
check_maintainability.py · github.com