The Agent Skills Directory

[COMMAND_EXECUTION]: The skill initiates shell commands to launch agent environments using the Anthropic ant CLI (for managed agents) or a user-defined agent.command (for REST-based self-hosted agents). These executions are central to the skill's purpose of testing agent performance.
[EXTERNAL_DOWNLOADS]: The documentation references downloading the ant CLI from the official Anthropic repository and installing Pome-related packages and skills via standard package managers like npx and npm.
[INDIRECT_PROMPT_INJECTION]: The skill processes task prompts and event data which are subsequently passed to the agent under test, creating a potential surface for indirect injection.
Ingestion points: The skill ingests examinee_task.prompt and initial_events (sourced from the Pome control MCP) as described in SKILL.md.
Boundary markers: The architecture uses specific report structures and a ## Handoff (fix prompt) section to separate automated outputs from user-edited instructions.
Capability inventory: The skill possesses the capability to execute shell commands (via the ant CLI or local command scripts) and perform network-based tool calls to the Pome MCP server for session management.
Sanitization: The skill enforces strict security measures, including mandatory network egress clamping (restricting external access) and the use of secure vaults for sensitive session tokens to prevent credential exposure.

pome-run-task