The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external project source code and feature maps to generate recipes for live execution. This creates a surface where content in the analyzed code could potentially influence the agent's behavior during verification steps.
Ingestion points: Project source code and local verification skill files (SKILL.md, feature files).
Boundary markers: The skill implements a functional boundary by using "read-only" subagents that are prohibited from driving the app or editing files during the analysis phase.
Capability inventory: The skill can execute harness scripts, perform local file writes within the designated skill directory, and create pull requests.
Sanitization: No explicit sanitization of ingested source code is described; the skill relies on the subagent's summarization process to extract recipes.
[COMMAND_EXECUTION]: The "Live pass" phase executes verification harnesses and recipes defined in the target skill. The workflow includes several safety invariants to manage this execution risk.
The skill enforces health checks on the environment before and after command execution.
It requires cleanup of residue from failed iterations to prevent persistent side effects.
The edit scope is strictly limited to the verification skill's own directory, preventing unauthorized modification of product code.

maintain-verification-skill