The Agent Skills Directory

[COMMAND_EXECUTION]: The skill makes extensive use of CLI tools such as git (to manage branches and commits), jq, and sed (to parse results). It also executes the codex CLI for model interaction.
[COMMAND_EXECUTION]: The skill automatically runs the repository's test command (e.g., pnpm test or make test) as a quality gate after applying fixes. This involves execution of scripts defined within the target repository.
[EXTERNAL_DOWNLOADS]: The skill depends on the @openai/codex NPM package, which is required to be installed globally on the system.
[DATA_EXFILTRATION]: Local source code and version control diffs are transmitted to OpenAI's servers via the codex CLI tool for the purpose of performing code analysis.
[PROMPT_INJECTION]: The skill contains an indirect prompt injection surface (Category 8).
Ingestion points: Local source code and diffs enter the agent's context and are processed by an external LLM (Codex).
Boundary markers: The response from the external model is structurally delimited by specific markers (), though these do not provide semantic protection against malicious instructions.
Capability inventory: The agent has permissions to modify local files, create git commits, and execute shell commands (tests).
Sanitization: The instructions explicitly mandate that the agent implement suggested mitigations "literally" and "mechanically," specifically forbidding the agent from applying its own reasoning or safety judgment to the external model's output.

self-review