The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes external CLI tools (Codex and Claude) with security-bypassing flags.
Evidence: scripts/live-copy-edit-agent.mjs spawns codex with the --dangerously-bypass-approvals-and-sandbox flag and claude with --permission-mode bypassPermissions to apply batch source edits without user intervention.
[PERSISTENCE]: The skill can persistently modify the agent's operating environment by installing execution hooks.
Evidence: scripts/hook-admin.mjs is designed to update configuration files like .claude/settings.local.json, .cursor/hooks.json, and .github/hooks/impeccable.json to automatically trigger the design detector script after file edits or at the end of a session.
[INDIRECT_PROMPT_INJECTION]: The skill facilitates an attack surface where data from an external browser can influence source code modifications.
Evidence: The skill runs a local HTTP server (scripts/live-server.mjs) that accepts "manual edit" payloads from the browser via the /manual-edit-stash endpoint. These payloads are then applied to the project's source code by a local AI runner. Although protected by a token, this mechanism allows external input to drive file system writes.
[EXTERNAL_DOWNLOADS]: The skill is configured to download and execute code from the public npm registry.
Evidence: SKILL.md lists npx impeccable * in its allowed-tools, which fetches the latest version of the vendor's package at runtime.
[DYNAMIC_EXECUTION]: Several components use dynamic loading and execution of code or modules.
Evidence: scripts/font-match.mjs attempts to dynamically resolve and require playwright or puppeteer from various paths, including the user's node_modules and the current working directory.

impeccable