The Agent Skills Directory

Structured Prompt Interpolation: The skill implements a template system in scripts/interpolate.ts and scripts/batching.ts that dynamically builds prompts for subagents. It includes a specific security-minded feature that rewrites placeholders in batched instructions to help subagents distinguish between instructions and data, reducing the risk of indirect prompt injection.
Scoped File System Operations: Local state is managed within a dedicated .swarm/ directory as shown in scripts/table.ts. The skill includes logic to sanitize session identifiers and automatically evicts old table data to maintain a clean environment.
Controlled Subagent Dispatch: Task execution is delegated to subagents using the platform's swarm_task tool. This creates a clear execution boundary, as subagents operate independently of the main agent's context and are constrained by the provided responseSchema.
Input Validation: The scripts/index.ts and scripts/table.ts files include validation steps to ensure that data sources (globs, file paths, or custom tasks) are well-formed and that row identifiers are unique before processing begins.

swarm