The Agent Skills Directory

[COMMAND_EXECUTION]: The skill relies on the rook CLI to execute arbitrary command lines to invoke agents via profiles (e.g., rook profile add --command '<argv>') and MCP servers (e.g., rook mcp add <name> -- <command>). While this is central to its testing purpose, it provides a powerful capability for executing shell commands.
[INDIRECT_PROMPT_INJECTION]: The skill ingests and grades responses from external agents under test.
Ingestion points: Agent replies are retrieved through rook profile test, rook run, and rook report (as described in SKILL.md and verdicts.md).
Boundary markers: The instructions suggest quoting criterion evidence in reports, but do not specify robust delimiters for the agent's raw output when the orchestrator processes it.
Capability inventory: The skill uses bash and the rook CLI to perform network operations, file writes in .testmuai/, and process execution.
Sanitization: There are no explicit instructions for the orchestrator to sanitize or escape the agent's reply before evaluation or reporting.
[CREDENTIALS_UNSAFE]: The skill manages sensitive credentials like LT_ACCESS_KEY. While it advises using environment variables and rook env set to keep secrets out of code and commits, it notes that rook env show can print full secrets to the console, creating a risk of exposure in logs if handled improperly.
[EXTERNAL_DOWNLOADS]: The skill references an installation link for the Rook CLI (github.com/LambdaTest/rook). As LambdaTest is the skill's author, this is considered a vendor-owned resource and is documented neutrally.

rook-skill