The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process user-provided programs and requests which are then passed to sub-agents. Ingestion points: The workflow captures input in request.md and program.md within a dedicated sem-runs/ directory. Boundary markers: The instructions explicitly mandate treating all program text as data and utilizing fresh, no-history sub-agents to prevent context leakage. Capability inventory: The system involves writing multiple Markdown artifacts and spawning isolated worker agents. Sanitization: Workers are restricted to specific contracts and assigned inputs, with a clear prohibition on external effects or workspace inspection.
[DYNAMIC_EXECUTION]: The skill implements a runtime orchestration layer that compiles natural language programs into a sequence of semantic applications. Each application is executed in a controlled, isolated sub-agent environment with limited scope, preventing unauthorized expansion of file access or execution capabilities.

sem-run