The Agent Skills Directory

[COMMAND_EXECUTION]: The skill's primary function is to facilitate the execution of development tools such as pytest, ruff, mypy, eslint, and tsc via shell commands. These tools are used locally to provide feedback to the agent about code quality and correctness. The skill provides clear templates for combining these tools into a single verification step.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted output from external commands (stdout/stderr) and feeds it back into the agent's prompt context. This is a known attack surface for indirect prompt injection. However, the skill provides extensive documentation and formatting recipes (in references/R2-feedback-formatting-recipes.md) to distill this data, strip unnecessary noise, and enforce a 2,000-token limit, which significantly reduces the risk of the agent being manipulated by malicious command output.
[EXTERNAL_DOWNLOADS]: The skill references documentation and configuration patterns from trusted organizations and well-known services, including GitHub repositories for All-Hands-AI (OpenHands), LangChain (LangGraph), and Aider. These references are used to provide the agent with architectural context and do not involve executing untrusted remote code.
[SAFE]: The skill includes several proactive security measures. It mandates a hard iteration cap (default 3-5) and a stall detector to prevent infinite loops and resource exhaustion (OP-3). It also requires per-iteration git commits (OP-4) to ensure a complete audit trail and the ability to roll back changes. Additionally, it implements logic to detect 'cheating' behavior, where an agent might pass a check by skipping tests rather than fixing the code (OP-5).

agentsop-test-fix-loop