The Agent Skills Directory

[COMMAND_EXECUTION]: Bundled Python utility scripts (e.g., browser_capture_artifacts.py) utilize subprocess.run to execute standard git commands. These operations are used to resolve repository paths and verify that the .codex/reviews/ directory is correctly configured as an ignored path in version control, preventing the accidental tracking of review data.
[EXTERNAL_DOWNLOADS]: The skill coordinates interactions with official, well-known AI service provider domains (such as chatgpt.com, gemini.google.com, and claude.ai). These interactions are central to the skill's purpose and are conducted through verified transport routes defined in the adapter protocols.
[INDIRECT_PROMPT_INJECTION]: As the skill ingests third-party AI responses and webpage content, it establishes a dedicated 'Untrusted External Content' protocol. This includes a mandatory quarantine phase and redaction logic to ensure the agent does not interpret instructions, follow links, or execute tool calls embedded in external data.
[DATA_EXFILTRATION]: The skill implements a 'sanitized payload' requirement for all external data synchronization. It explicitly forbids the transmission of secrets, credentials, environment variables, or private browser profile data, employing automated redaction for relay and retention workflows.

ask-ai