The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill directs the agent to fetch and use a relay script (relay.mjs) from a community-maintained GitHub repository (amElnagdy/delegate-skills). This source is external to the skill and not associated with a recognized trusted vendor.- [COMMAND_EXECUTION]: The orchestrator is instructed to execute shell commands involving node and the copilot CLI. These commands interact with the host system and the repository being modified.- [DYNAMIC_EXECUTION]: The skill executes a script at a runtime-determined path (relay.mjs). This script is not part of the skill's own package, increasing the risk of executing unverified code.- [INDIRECT_PROMPT_INJECTION]: The skill's architecture creates an attack surface where the orchestrator agent processes untrusted data produced by the GitHub Copilot CLI. A malicious implementer output could attempt to influence the orchestrator through the result.json report or the modified source code.
Ingestion points: The agent reads result.json, events.jsonl, and the git repository files after delegation.
Boundary markers: The skill lacks technical boundary markers or sanitization for these inputs, relying instead on narrative instructions for the agent to verify work.
Capability inventory: The orchestrator possesses shell execution and file system access capabilities.
Sanitization: No systematic sanitization of external tool outputs is described.

copilot-delegate