The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to dynamically build and execute bash commands using the codex CLI tool by substituting variables like PLAN_FILE and LOG_FILE into command strings, creating a potential command injection vector if inputs contain shell metacharacters.\n- [COMMAND_EXECUTION]: The skill uses complex shell features such as subshells ($(cat ...)), redirection (< /dev/null), and pipes, which can lead to unpredictable behavior in automated environments.\n- [CREDENTIALS_UNSAFE]: The skill directs the agent to read ~/.codex/config.toml to extract model information, which is a sensitive operation involving user configuration files that could potentially expose other settings or credentials.\n- [EXTERNAL_DOWNLOADS]: The skill requires the third-party codex CLI tool (version ≥ 0.130) to be installed and authenticated, creating a dependency on external software associated with the OpenAI service.\n- [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection. It ingests data from local files (PLAN.md, CONTEXT.md) and interpolates them into an adversarial review prompt. The evidence chain includes ingestion points in the review loop, a lack of explicit sanitization for ingested data, and the capability to execute shell commands based on the results of the review process.

grill-with-docs-codex