| name | light-security-ClaudeSettingsAudit |
| description | Audit Claude Code config files (settings.json, settings.local.json, .mcp.json) across the computer for unsafe or malicious content — hooks that run shell commands, credential/proxy env vars, arbitrary MCP servers, permission-bypass modes, over-broad allow rules. Classifies each finding DANGEROUS / SUSPICIOUS / SAFE. Read-only. Use when the user invokes `/light-security-ClaudeSettingsAudit`, says "check my settings.local.json for malware", "are my Claude permission files safe", "scan my Claude config", or "audit settings.json across the computer". |
| disable-model-invocation | true |
/light-security-ClaudeSettingsAudit — Claude Code config safety scan
This is a thin launcher. Your only job is to resolve the scope and launch the light-security-ClaudeSettingsAudit agent (pinned to Sonnet — the scan is mostly mechanical). You do NOT scan the files yourself.
Step 1 — Resolve the scope
$args may name a folder or a single file to scan. If given, pass it as the scope. If empty, the default scope is the whole machine (the agent knows the roots to cover). If the user names extra trusted items or extra roots, carry them into the agent's prompt.
Step 2 — Launch the agent
Call the Agent tool with subagent_type: "light-security-ClaudeSettingsAudit". Pass the scope (or "whole machine, default roots") and any trust/extra-root notes in one prompt. The agent reads the classification at references/settings-danger.md, globs for config files, reads and classifies each, and returns the report.
Step 3 — Relay
Print the agent's report verbatim — the verdict line, the per-bucket tables, and the coverage line. Do not soften or drop findings. If anything is DANGEROUS, say so plainly at the top.
Notes
- Read-only. Removing or fixing a bad entry is a separate, explicit follow-up the user must request.
- The classification in
references/settings-danger.md is the source of truth. Update it (and its Sources date) when Claude Code adds fields or the machine's known-good baseline changes. The agent reads it by absolute path.
- Weekly run: this skill can be scheduled. See the audit note in the project CLAUDE.md for the cron/Task-Scheduler wiring.
