The Agent Skills Directory

[Indirect Prompt Injection] (LOW): The framework's core workflow (Gather-Plan-Apply) ingests untrusted content from a local codebase via precision_read and discover, then uses that information to drive actions through precision_write, precision_edit, and precision_exec. This pattern creates an architectural surface for indirect prompt injection. 1. Ingestion points: precision_read and discover tools target workspace files. 2. Boundary markers: Not evidenced in documentation; there are no instructions to the agent to treat file content as untrusted data. 3. Capability inventory: Includes precision_write, precision_edit, and precision_exec for system-level modifications. 4. Sanitization: No sanitization of ingested content is described before it influences the agent's plan.
[Command Execution] (SAFE): The precision_exec tool is used for standard development tasks in the provided examples, such as npm run lint and npm run build. No dangerous commands, such as sudo or remote script execution (e.g., curl | bash), were detected in the reference material.

gather-plan-apply