The Agent Skills Directory

[COMMAND_EXECUTION]: The skill triggers the execution of a local command-line tool ao converge via the Bash tool. This is used to manage a bounded loop for automated code fixing and verification.
[DATA_EXPOSURE]: The skill interacts with local project state by reading from and writing to files within the .agents/ directory, specifically .agents/findings/registry.jsonl for persistent findings and .agents/rpi/KILL as a termination mechanism.
[PROMPT_INJECTION]: The skill presents an 'Indirect Prompt Injection' surface as part of its core functionality:
Ingestion points: Feedback ('reasons' and 'evidence') produced by external judge models (e.g., Codex or Claude) during the evaluation leg.
Boundary markers: None explicitly defined in the instructions to separate judge feedback from agent instructions.
Capability inventory: The agent utilizes Bash (via ao converge) and Read tools during the loop.
Sanitization: No specific sanitization or filtering of the judge's evidence is mentioned before the orchestrating agent processes it to apply fixes. This is the intended functional loop of the skill.

converge