The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses the codex command-line tool to implement code changes and run validations. This behavior is managed through a 'Local Codex Gate' described in references/usage.md, which requires explicit user approval and uses a sandboxed environment.
[EXTERNAL_DOWNLOADS]: The skill utilizes an external browser operator (ops-browser) to navigate to ChatGPT and potentially other live web pages. This capability allows for the inspection and capture of remote content as detailed in references/live-browser-review.md and references/browser-operation-protocol.md.
[DATA_EXFILTRATION]: The skill's primary function involves sending repository data (files, diffs, and metadata) to ChatGPT. While this is the intended purpose, it constitutes a data transfer. The skill includes 'Hard Rules' in SKILL.md and references/usage.md specifically designed to exclude secrets, tokens, and private keys from these transfers.
[PROMPT_INJECTION]: The skill is subject to indirect prompt injection because it processes untrusted data from external sources (ChatGPT responses and live web pages) which can then influence local command execution.
Ingestion points: External assistant responses are captured into review.md and live page content is retrieved via the transport/reviewer browsers.
Boundary markers: The skill implements the 'External Action Gate' and 'Local Codex Gate' to ensure user authorization at critical steps. It explicitly instructs the agent to treat external findings as 'untrusted review input' in SKILL.md.
Capability inventory: The skill can perform local command execution (codex exec, git, gh), file system operations (writing review-package.md and review.md), and browser automation via ops-browser.
Sanitization: SKILL.md and references/usage.md mandate the redaction of credentials and the sanitization of URLs and workspace identifiers to protect privacy.

chatgpt-review