tool-auditlisted

# Tool audit One front door for every kind of tool the user has: skills, agents, MCP servers, plugins, CLIs, local services, API keys and their own scripts. `/skill-audit` is this skill scoped to skills only. There is no third tool, and there is **one ledger**: `~/.claude/skill-audit/ledger.jsonl`. Never start a second log. Neither runs on its own. The user types them. ## The four rules that hold for every run 1. **Nothing installs.** Parked tools are **borrowed** (read the file, follow it) or **session-launched** (CLI flags that die with the terminal). Promotion to live happens only when the user says so, in a separate step. 2. **Never write into `~/.claude/agents`, `~/.claude.json`, or `enabledPlugins` in `settings.json`.** the user runs many terminals at once and those are shared: an edit there changes every session they have open, including the ones mid-task. A loadout is a file in `~/toolbox/loadouts/` and a command they run in a new tab. 3. **Never print a secret value.** Key names and env var names only, and never open a key file "to check". Anything under `~/toolbox/mcp/` is read for server *names* and never echoed. 4. **Search everything, run few.** Every audit sweeps the whole library (section 2). The caps below limit what runs, never what is considered. ## Hard caps on a loadout | kind | cap | |---|---| | agents | 5 | | skills | 5 | | MCP servers | 3 | | plugins | 2 | Over the cap is not a loadout, it is the problem the toolbox was b