The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes external CLI binaries (e.g., cursor-agent, codex, copilot, agy) using node:child_process.spawnSync. The executables are resolved dynamically from environment variables (e.g., LISA_CURSOR_CLI), which allows the skill to run binaries from paths specified in the environment.
[PROMPT_INJECTION]: The skill processes output from external AI agents, which creates an indirect prompt injection surface.
Ingestion points: Output from the external CLI tools is captured in runtime-adapters.mjs and consumed by the advisory logic in first-round.mjs.
Boundary markers: The skill uses structured section headers and explicit instructions (e.g., "Operate in read-only advisory mode") in its prompts to Claude to isolate untrusted agent output.
Capability inventory: The skill has the ability to execute local commands and access specific environment variables.
Sanitization: It implements the sanitizeCouncilText function to redact common secrets (like GitHub tokens and API keys) and to annotate potentially destructive commands (like npm install or git push) with maintainer warnings.

harness-parity-council