The Agent Skills Directory

[COMMAND_EXECUTION]: Documentation describes an API (resolveAgentHarness) used to spawn external agent binaries as local child processes communicating via stdio JSON-RPC.
[COMMAND_EXECUTION]: Details a run-code tool that executes code modules through a pluggable SandboxAdapter, with the default implementation using a local Node.js child process.
[EXTERNAL_DOWNLOADS]: Provides instructions to install and run optional dependencies from well-known technology organizations, including @google/gemini-cli and @zed-industries/agent-client-protocol.
[CREDENTIALS_UNSAFE]: Specifies that spawned child processes inherit the host's environment variables, which is a standard feature allowing agents to reuse existing local CLI login sessions for authentication.
[PROMPT_INJECTION]: The skill documents a surface for indirect prompt injection by processing untrusted user prompts through agents with file system and command execution capabilities.
Ingestion points: Untrusted data enters the system via the prompt field in the startAgentHarnessRun function (SKILL.md).
Boundary markers: The documentation does not specify explicit boundary markers or delimiters for the prompt input.
Capability inventory: The system enables child process spawning, arbitrary code execution via sandboxes, and local file system access (SKILL.md).
Sanitization: Includes workspace path validation to prevent file system traversal and uses isolation through SandboxAdapter for code execution.

harness-agents