The Agent Skills Directory

[DYNAMIC_EXECUTION]: The monitoring recipe detailed in references/codex-transport.md uses the node -e command to execute an inline JavaScript snippet that parses and validates progress data from a JSON file during the review loop. This represents dynamic execution of code generated or defined within the skill's instructions.\n- [COMMAND_EXECUTION]: The skill relies heavily on the Bash and node tools to run local scripts such as scripts/codex-exec.js (the transport adapter), scripts/resolve-feature.sh (for context detection), and scripts/validate-family-sentinel.js (for verdict validation). These scripts are invoked with arguments derived from the repository's state and file system.\n- [INDIRECT_PROMPT_INJECTION]: The skill's workflow involves reading and processing potentially untrusted data from the codebase, including git diffs and acceptance criteria from documentation files (Step 1.5). This data is subsequently interpolated into prompts sent to the Codex service and the secondary reviewer agents, creating a vector for indirect prompt injection.\n- [DATA_EXFILTRATION]: The primary purpose of the skill is to collect local code changes, diff statistics, and project metadata and transmit them to an external service ('Codex') for analysis. This inherent behavior constitutes a structured path for data exfiltration from the local development environment.\n- [PRIVILEGE_ESCALATION]: The transport adapter described in references/codex-transport.md performs high-complexity system-level operations, such as process liveness monitoring using kill(pid, 0) and atomic file management in shared temporary directories with strict permission enforcement (0600). The 'Locator' logic also describes a mechanism for automatically copying and installing scripts into the .claude/scripts/ directory, which can be used to modify the agent's executable environment.

codex-code-review