The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill defines primitives for processing external data via triggers (HTTP, pub/sub, streams) which can influence subsequent agent actions or worker execution.\n
Ingestion points: Triggers such as http (body, headers), subscribe (payload), and stream (event details) allow untrusted data into the system context.\n
Boundary markers: The skill documentation does not specify the use of delimiters or 'ignore' instructions for these payloads.\n
Capability inventory: The framework allows for invoking functions (iii.trigger), enqueuing tasks, and executing scripts defined in worker manifests (install, start).\n
Sanitization: While examples show basic validation logic, there are no specific instructions for sanitizing or escaping untrusted input before processing.\n- [EXTERNAL_DOWNLOADS]: The skill documents the installation of workers from external sources.\n
Evidence: Documentation for iii worker add which fetches from the vendor registry (workers.iii.dev) or container registries (ghcr.io).\n- [COMMAND_EXECUTION]: The worker manifest structure supports the definition of shell commands for lifecycle management.\n
Evidence: The iii.worker.yaml example includes install and start script definitions that execute shell commands like pip install and python.

iii-core-primitives