The Agent Skills Directory

[PERSISTENCE]: The installation scripts (scripts/setup-codex-hook.sh and scripts/setup-gemini-hook.sh) modify CLI configuration files in the user's home directory (~/.codex/config.toml and ~/.gemini/settings.json) to install persistent developer instructions and autonomous loop hooks that survive session restarts.
[DYNAMIC_EXECUTION]: The skill dynamically generates shell scripts (e.g., ralph-ooo-check.sh) and writes them to the filesystem, configuring them as executable hooks that re-trigger the agent until specified verification criteria are met.
[INDIRECT_PROMPT_INJECTION]: The skill ingests and processes potentially untrusted data such as user requirements, existing codebases (in brownfield mode), and tool outputs during the specification and execution phases.
Ingestion points: Requirements provided to ooo interview and ooo ralph commands; file content read during the ooo run (Double Diamond) phase.
Boundary markers: The skill uses an immutable YAML 'seed' as a ground truth to bound behavior, but does not implement explicit delimiters to prevent the agent from following instructions embedded in the project files it analyzes.
Capability inventory: Access to Bash, Write, and parallel Agent tools allows the skill to perform complex system operations and spawning sub-agents.
Sanitization: Scripts use standard JSON and YAML parsing (via Python's json module) for state management, but there is no explicit sanitization of user-provided task descriptions before they are interpolated into the continuation prompts.
[COMMAND_EXECUTION]: The skill makes extensive use of the Bash tool to perform state management, configuration patching, and project-specific tasks (like running tests or linting).

ralph