The Agent Skills Directory

[COMMAND_EXECUTION]: In the Autoresearch_Mode section, the skill constructs and runs a shell command: omc autoresearch --mission "<mission>" --eval "<evaluator>". The mission and evaluator components are derived directly from user input during the interview process. This presents a command injection risk if the user provides shell metacharacters or malicious scripts that the underlying tool does not properly sanitize.
[PROMPT_INJECTION]: The skill presents an Indirect Prompt Injection surface. It ingests untrusted user input during the Socratic interview loop and incorporates it into a final specification file (.omc/specs/deep-interview-{slug}.md). This specification is then used as the primary instruction source for downstream autonomous skills like omc-plan, autopilot, and ralph via the Skill() bridge.
Ingestion points: User input from {{ARGUMENTS}} and interview responses.
Boundary markers: None explicitly defined in the prompts to separate user content from system instructions.
Capability inventory: File writing (Write), shell command execution (omc autoresearch), and invocation of other high-capability agents (Skill()).
Sanitization: No sanitization or validation of the user's answers is mentioned before they are written to the specification file.
[COMMAND_EXECUTION]: The skill utilizes a sub-agent (oh-my-claudecode:explore) to systematically map the local codebase in 'brownfield' scenarios. While this is an intended feature for context gathering, it involves automated file system access and metadata extraction.

deep-interview