The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructs the agent to download and execute code from the npm registry using npx --yes @codedrobe/core@latest or bunx @codedrobe/core@latest. This pattern is used to run the vendor'specific CLI tool directly from a remote source.\n- [DATA_EXPOSURE_AND_EXFILTRATION]: The skill accesses local application configuration files, specifically ~/.codex/config.toml, to modify appearance settings. It also uses the Chrome DevTools Protocol to capture DOM snapshots of active applications. The instructions specify that the snapshot process excludes sensitive content like text, input values, and media sources by default to maintain privacy.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes external, potentially untrusted theme packages and CSS files.\n
Ingestion points: Untrusted .codedrobe-theme packages and CSS source files (SKILL.md).\n
Boundary markers: Instructions include explicit warnings to treat theme packages as untrusted input and to avoid evaluating embedded JavaScript (SKILL.md).\n
Capability inventory: The skill has the ability to write to application config files and interact with live renderers via CDP (references/codex.md).\n
Sanitization: The instructions rely on the Core CLI to safely parse packages without executing scripts.\n- [DYNAMIC_EXECUTION]: The skill performs runtime injection of CSS styles and theme markers into Electron/Chromium application renderers via the codedrobe apply and codedrobe probe commands.

codedrobe-theme