The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill automates the execution of external AI agents (Claude and Codex) to apply code edits to the repository. In live-copy-edit-agent.mjs, it explicitly utilizes the --dangerously-bypass-approvals-and-sandbox flag for Codex and --permission-mode bypassPermissions for Claude, which circumvents standard security constraints during the code modification process.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection through its data ingestion surface.
Ingestion points: live-browser.js and live-server.mjs extract DOM structure, text content, and screenshots from the user's active browser tab.
Boundary markers: No explicit delimiters or instructions to ignore embedded commands were found in the prompt construction logic within live-copy-edit-agent.mjs or codex.md.
Capability inventory: The skill possesses extensive capabilities, including writing to source files (live-accept.mjs, live-wrap.mjs), modifying project configuration, and patching Content Security Policies (CSP).
Sanitization: External content from the browser is truncated but otherwise interpolated directly into prompts without robust escaping or filtering.
[COMMAND_EXECUTION]: Multiple scripts perform extensive file system operations and configuration changes. live-inject.mjs and live-wrap.mjs modify project source files (HTML, JSX, TSX, Svelte, Astro), while hook-admin.mjs and hook-lib.mjs modify git configuration (.git/info/exclude) and project-local config files.
[DATA_EXPOSURE]: The local helper server started by live-server.mjs sets Access-Control-Allow-Origin to *. While the server is bound to the loopback interface, this configuration increases the risk of cross-site interaction if local ports are identified by malicious external websites.

impeccable