The Agent Skills Directory

[COMMAND_EXECUTION]: The skill's Autoresearch_Mode is designed to execute the command omc autoresearch using arguments (mission and eval) derived directly from user responses during the interview loop. While this is the intended functionality of that mode, it lacks explicit security sanitization beyond checking for 'clarity' and 'safety' via LLM judgment before execution.
[PROMPT_INJECTION]: The skill implements internal behavior steering through 'Challenge Agents' (Contrarian, Simplifier, Ontologist) which use prompt injection techniques to modify the agent's reasoning perspective based on the number of interview rounds. These are benign, self-contained features to improve requirement quality.
[PROMPT_INJECTION]: There is an indirect prompt injection surface in the Autoresearch_Mode where user-controlled strings for the 'evaluator command' are interpolated into a shell command.
Ingestion points: User input provided in response to the question "What should autoresearch improve or prove for this repo?" and the evaluator command prompt in Autoresearch_Mode.
Boundary markers: The command uses double-quotes for the --mission and --eval arguments, which provides minimal protection against sophisticated command injection vectors like subshells or backticks.
Capability inventory: The skill uses the explore subagent for file system analysis, the Write tool for creating specification files, and has the capability to execute shell commands via the omc CLI.
Sanitization: The skill relies on the LLM's assessment of whether an evaluator is "explicit enough to launch safely," which is a qualitative assessment rather than a technical sanitization or validation of the command string.

deep-interview