The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts/relay.mjs script executes several system commands including pi, git, and taskkill. On Windows, the script safely uses shell: true to resolve the pi.cmd shim, protecting against injection by validating all dynamic arguments (model, provider, and session IDs) against a restrictive alphanumeric whitelist. It also executes a local Node.js script from a sibling directory for configuration resolution.\n- [EXTERNAL_DOWNLOADS]: The skill instructions direct the user to install the @earendil-works/pi-coding-agent package via the NPM registry. This is an expected dependency for the skill's core functionality.\n- [INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion of task briefs which are then processed by an external agent. This creates a surface for indirect prompt injection if the brief or workspace context files contain malicious instructions.\n
Ingestion points: Task briefs are ingested from local files or stdin and passed to the sub-agent.\n
Boundary markers: Briefs are delivered to the sub-agent via stdin, ensuring they are not parsed as command-line arguments.\n
Capability inventory: The sub-agent (pi) has capabilities to read, write, and execute commands within the user's workspace.\n
Sanitization: The skill provides a robust mitigation strategy through the references/review-and-land.md guide, which enforces a mandatory review of all changes and a requirement to re-run project test gates before any code is committed.

pi-delegate