The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill provides detailed instructions for the agent to generate standalone prototype applications within specific local directories (.agents/prototype/<name>/). These applications contain their own source code and dependency manifests, and the agent is instructed to execute them using commands such as pnpm start, pnpm dev, or bun run. Executing agent-generated code at runtime is a notable dynamic execution surface.
[COMMAND_EXECUTION]: The agent uses shell commands to manage external issue trackers (using tools like gg issue edit and gg api) and to run the generated prototypes. This capability allows for the execution of shell commands where the behavior could be influenced by data originating from external sources.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process tasks and decisions from external issue trackers (GitHub, GitLab) and local markdown files, creating an attack surface for indirect prompt injection.
Ingestion points: Data is read from external tracker issue titles, bodies, and comments, as well as from files in the .agents/plans/ directory.
Boundary markers: The instructions rely on markdown headers (e.g., ## Question, ## Answer) to structure and delimit data, which do not serve as secure boundaries against malicious instructions embedded in the ingested content.
Capability inventory: The agent possesses the ability to write to the filesystem, execute shell commands, and perform network requests via tracker APIs.
Sanitization: There is no explicit requirement for the agent to sanitize or validate the external content before processing it, potentially allowing malicious instructions in tickets to influence the agent's logic.

vibe-deep-plan