The Agent Skills Directory

[SAFE]: The skill operates as an advisory routing mechanism that uses logic-based precedence rules to ensure the agent selects the narrowest and safest workflow appropriate for a given task.\n- [INDIRECT_PROMPT_INJECTION]: The skill manages an attack surface where untrusted user input and attachments are ingested to determine the execution route. Evidence chain: 1. Ingestion points: User prompts, attachments, and Git selectors as defined in SKILL.md. 2. Boundary markers: Explicit instruction to treat attachment content as data, not instructions. 3. Capability inventory: The skill identifies specialized runtimes (e.g., grilling, writing-plans, systematic-debugging) for the platform to load but contains no internal executable code or scripts. 4. Sanitization: It requires the resolution of scope locators through cheap repository signals before processing, reducing the risk of unintended context injection (Low severity finding).\n- [COMMAND_EXECUTION]: The skill implements a safety gate for high-risk operations by detecting patterns of destructive behavior (e.g., database deletions or force pushes) and placing them into an 'authorization hold' status, preventing execution until explicit user confirmation is received.

workflow-gate