The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill employs the WebFetch tool to retrieve content from external URLs for evaluation. This is standard behavior for its intended use case.
[COMMAND_EXECUTION]: It executes the local script bun meta-skills/scripts/manifest-sync.ts to maintain the agent's internal state, which is a safe project-level operation.
[SAFE]: Evaluation of indirect prompt injection surfaces: (1) Ingestion points: Data is fetched from user-provided post URLs. (2) Boundary markers: Content is passed as context to sub-agents without explicit sanitization headers. (3) Capability inventory: The agent has access to file writing and local command execution. (4) Sanitization: While explicit sanitization is absent, the rigid output schemas and the multi-agent validation process (Critic agent) provide inherent mitigation against adversarial content.

short-form-eval