The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes a 'Local Codex Gate' to execute shell commands via the codex tool for applying verified fixes and running validation. This mechanism incorporates safety features such as sandboxing and configurable user approval modes, as documented in the usage and browser protocol references.
[EXTERNAL_DOWNLOADS]: The skill performs authorized network operations to well-known services including OpenAI and GitHub to facilitate external reviews and research. These interactions are managed through an 'External Action Gate' and are restricted to specific, user-authorized scopes.
[PROMPT_INJECTION]: The skill implements defenses against indirect prompt injection by treating data ingested from ChatGPT as untrusted. Ingestion points include external responses captured in review files. Boundary markers and explicit instructions mandate ignoring embedded instructions that attempt to bypass safeguards or reveal secrets. The skill also requires local verification of findings against repository ground truth before any actionable steps are taken.
[SAFE]: The skill contains comprehensive safety rules that prevent the exposure of sensitive data. It explicitly forbids the exfiltration of credentials, secrets, and private tokens, requiring their redaction from any outbound review-package artifacts.

ask-chatgpt