Test agents with Rook

Use Rook to derive scenarios, invoke the target and grade the evidence. Respect requests to use another tool. For saved-result requests, read the supplied files without starting setup, a new run or paid analysis.

Your role

You are the orchestrator: you choose and run Rook commands, and Rook does the testing work. Discovery is rook explore, test design is rook generate, and invocation and grading are rook run and rook report. When the user asks to explore, understand, plan tests for or test an agent, run the matching command. Do not read the target's code to describe the agent, or write your own test plan or verdicts, in its place. Answer from Rook's agents, scenarios and verdicts. Read the target's source only for a step Rook needs from you, such as a profile's transport and command, or to explain a Rook result. If credits or authorization block the next command, name it and what it needs; manual analysis is not a substitute.

Setup and execution

Check rook --version and use rook help <command> for the installed CLI's options. If Rook is missing, use the installation instructions. Use this workflow across compatible releases; a different version alone is not a reason to stop or change the installation. If a command or output differs, consult its help and release notes before adapting that step. Keep completion and evidence checks intact; do not guess missing fields or retry paid work to probe compatibility. Installation changes still require authorization.

rook doctor
rook status --json
rook plan --json

Use the next command needed by the current state:

Choose the profile's transport and command from the target's source or the user's invocation details; ask when unknown. See profiles. Use rook help <command> for flags and scenarios to scope a run with --only, classes, categories or tags. When an agent has more than one profile, check the active profile with rook profile or pass rook run --profile <id> to select the intended target.

Authorization and credits

Before profile add, profile fix, profile test or run, inspect the target's agent.yaml for calls[] entries with write: true. Those commands can reach the real agent. Use staging and obtain any missing authorization for its effects; Rook cannot undo them; do not rely on a per-target headless confirmation. Verify effects through read-only interfaces; inspect available read tools before concluding that an effect is unverifiable. Keep secrets as ${VAR} references through rook env set, outside profiles and transcripts.

Announce spending before explore, generate, run, profile add|fix, ask and run|report --rca. profile test invokes the target without Rook model credits. Use paid RCA only with user authorization. Track the user's budget across commands; do not assume the CLI enforces an aggregate task cap. Check its help for available limits. A null balance means unknown. Prefer scoped --allow grants; use --yes only within the authorized scope. See headless contracts for output, grants and credit accounting.

Report the evidence

Check the exit status and discarded/halted fields before claiming completion. A finished run exits 0 even when scenarios fail; ok: true can also describe a refused run. Parse JSON only when a document exists, otherwise use stderr. Read the report by this invocation's run ID, not an old default report.

Follow verdicts: show Pass, Fail and Unable to Verify separately, quote criterion evidence, identify gaps and compare prior runs only when their evidence is available. Preserve incomplete results and unknown metadata; do not invent evidence or replace Rook's verdict with your own reading of the agent reply.

References

• CI: complete pipeline and verdict gating.
• MCP: server discovery, approval and configuration.
• Troubleshooting: setup failures and version-specific verification limitations.
• Headless contracts: JSON shapes and saved-file locations.

Project evidence lives under .testmuai/rook/. Credentials and session state live under ~/.testmuai/rook/; do not commit that home directory.