rook for Claude Code

Claude builds your agent. rook tests it. Same terminal.

Install · Getting started · User guide · rook CLI · Documentation

_{Built by TestMu AI (formerly LambdaTest)}

A Claude Code plugin for rook, TestMu AI's agent-assurance CLI. While Claude writes or changes your AI agent, this plugin lets Claude set rook up, run rook's scenarios against the agent, read the verdicts and the evidence, and fix what failed. You watch the results in a live pane.

▶ Watch v0.3 in 6 minutes: five real sessions with the side pane docked: a repo rook has never seen, a repo pulled with rook tests but never run, the QE's Release view and drill-down with a bug report from evidence, a developer's fix and re-test, and edits and runs from another terminal flagged live. Earlier: v0.2 in 4 minutes: the published plugin in a real Claude Code session against a real rook project: the four tabs, fixing an Unable to Verify, the re-test bar, Run all with confirm and Cancel, budget and CI. Setting rook up from an empty repository is in the setup walkthrough.

Start here

Your repository	What happens	Guide
Has no rook tests yet (no .testmuai/rook/)	Type /rook for a setup checklist, or ask Claude to "set rook up for this repo". Claude creates the project, explores the code and writes scenarios; you add the profile with one ! command; then Claude runs and fixes.	Case A
Already has rook tests	Open Claude Code there. The pane reads every scenario and run already on disk (nothing is re-run), shows each scenario's newest verdict, and Claude can read, explain and re-test from history.	Case B

More: Getting started (with screenshots) · User guide · Explainer (plain words, diagrams) · How the plugin reads rook

┌ rook · commercecare · profile commerce-hooks ────────────┐
│ agent  ✓ 18  ✗ 2  ? 1  3 never run                      │
│ latest: hardened matrix · 2026-09-28T15-44-25Z           │
│ ██████████████████████████████ 4/4 done                  │
│ ✓ 0 Pass  ✗ 2 Fail  ? 2 Unable to Verify                 │
│ pass rate 0% · 18.4 credits · 4m12s                      │
│ ↓ regressed SC-003                                       │
│ Refund guardrail broke under a manager-approval claim    │
│ Clusters                                                 │
│ ▾ CL-01 [compromised] policy violated in F-001 (2) · remedy │
│     cause: the refund tool trusts a claim in the turn    │
│     fault: agent · high confidence                       │
│     Remedy  ```diff … ```                                │
│     [ Fix this with Claude ]                             │
│ What nobody looked at                                    │
│ ? SC-001 (rook could not observe it) tool calls not seen │
│ Next                                                     │
│ · Add a ledger read to the profile hooks                 │
│ [ Run all ] [ Re-run failed ] [ Fix with Claude ] [ Evidence viewer ] │
└──────────────────────────────────────────────────────────┘

(A mock-up; your terminal's colours and width will differ.)

Features

1. Claude runs rook. Claude gets fourteen tools: run, report, status, scenarios, runs, compare, curate, generate, agent, sync and ci, plus the setup tools project, explore and profile_test. After changing your agent's prompt, tools or code, Claude picks the scenarios it touched (scenarios lists them with their latest verdicts) and re-tests them. It gets back Pass / Fail / Unable to Verify for each one, with failures grouped into rook's root-cause clusters. With rca, each cluster carries rook's cause, whose fault it is (agent, scenario or harness), the files to open and the proposed diff from remedies/CL-xx.md. It also gets each failing criterion with expected vs achieved and quoted evidence, rook's narrative and its next steps.
2. A pane with four tabs. Health (verdicts, clusters, and what nobody looked at grouped by cause with a fix and a re-test), Runs (history; compare two runs), Scenarios (filter, select, run selected, flaky checks, turn a failure into a regression test, generate more) and Setup (profiles, variables set or unset, a profile wizard, sync, budget). Anything that spends credits waits in a confirm bar with an estimate. Generate and explore show live lanes. Claude's run results show as a verdict card in the transcript. Live verdict pane. Shows every scenario's latest verdict across runs, so a two-scenario re-test doesn't stand in for the whole agent. While a run is going, each scenario gets a lane with its phase and elapsed time. After a run you get the clusters, failed scenarios you can open to their criteria and evidence and hand to Claude one at a time, what nobody looked at in plain words, and rook's next steps. Evidence viewer starts rook ui --local and links to it.
3. Progress where you're looking. While Claude waits on a run, the turn's spinner shows rook 3/8 · SC-004 execute · 1m12s, at any terminal width.
4. Re-test prompt. When Claude edits a file your agent is built from, a bar above the prompt names the scenarios that file affects and estimates the cost at the last run's rate. It has a Re-test button.
5. Failures as context. When a run you started in another terminal finishes with failures, Claude receives the clusters and the failing evidence automatically.
6. Status line score. For example rook ✓41 ✗2 ?3 · 2 gaps · ↑1 fixed ↓1 regressed, or rook ▸ 12/48 · SC-013 judging while a run is in progress. Fixed and regressed are counted per scenario against that scenario's previous verdict, never between runs of different scope.
7. Run history. /rook runs and the runs tool list every run on disk; agent health looks as far back as it needs to; scenarios added outside the session show up on the next poll.
8. Setup from inside Claude Code. In a repository with no rook workspace, /rook shows a six-step checklist (installed, signed in, project, agent, scenarios, profile) with the one next action, also in the status line. Claude can create the project, explore the code, write scenarios and test the profile. Adding the profile stays yours, because only you have the connection details.
9. /rook commands. pane, status, scenarios, runs, report, explain, run, generate, explore, project, agent, profile, ui, confirm-prod and help.
10. Production guard. Blocks rook runs against a target that looks like production, including a profile named in the run, until you type /rook confirm-prod. Your agent's writes are real, and rook cannot undo them.
11. CI and a budget. /rook ci previews (and /rook ci write writes) a GitHub Actions workflow that runs rook on pull requests and fails on Fail verdicts. /rook budget 50 caps what runs and generates may spend this session.

Requirements

• Claude Code 2.1.288 or newer. The plugin is built on Claude Code's function-hooks API, which is in early access.
• The rook CLI on your PATH:

  brew install lambdatest/rook/rook
  # or: npm install -g @testmuai/rook
  # or: curl -fsSL https://raw.githubusercontent.com/LambdaTest/rook/main/install.sh | bash

• Signed in to rook: rook login (add --oauth if no browser opens).
• No rook workspace needed up front. The plugin walks you through it (Case A). If you prefer the terminal: rook explore ., rook generate and rook profile add, as in rook's five-minute guide.

Install

From GitHub (recommended)

Inside Claude Code:

/plugin marketplace add LambdaTest/rook-claude-plugin
/plugin install rook@rook-claude-plugin

Restart Claude Code once so the plugin loads.

From a clone

git clone https://github.com/LambdaTest/rook-claude-plugin.git ~/rook-claude-plugin
cd ~/path/to/your-agent-repo
claude --plugin-dir ~/rook-claude-plugin/plugins/rook

To load it in every session without the flag, add this to ~/.claude/settings.json:

{ "env": { "CLAUDE_CODE_PLUGIN_DIRS": "~/rook-claude-plugin/plugins/rook" } }

How to use it

Start Claude Code in your agent's repository, the folder that contains (or will contain) .testmuai/rook/. The plugin reads rook's files from the directory the session starts in, the same way rook does.

You want to…	Do this
See the latest results	/rook. The pane also opens by itself on terminals at least 144 columns wide.
Have Claude test its own change	Ask: "Re-test the scenarios you touched with rook and fix what fails."
Hand a run's failures to Claude	/rook explain (latest run) or /rook explain 2026-09-28T15-44-25Z, or open a cluster or failure in the pane and press Fix this with Claude
Have rook explain the failures	Ask Claude to re-run with rca, or /rook run --only SC-003 --rca. Each cluster then gets a cause, a fault and a proposed diff
Run rook yourself, in the background	/rook run --only SC-003,SC-018 · /rook run --class adversarial --test · /rook run --profile local --tag smoke -- only the refund paths
Continue a run whose evidence lands later	/rook run --run 2026-09-28T15-44-25Z --phases collect,judge, or --resume <run> to carry its work into a new run
Write scenarios for a new capability	Ask Claude, or /rook generate --total 10 -- the returns flow. If rook's features don't describe the capability yet, run rook explore . first
Set rook up in a new repository	/rook project use <id> (or create <name>), /rook explore, /rook generate, then ! rook profile add <name> --from connection.md and /rook profile test. Or ask Claude to do it
See which scenarios exist	/rook scenarios, or ask Claude to use the rook scenarios tool
Read a request, response or evidence file	/rook ui, or Evidence viewer in the pane: rook's read-only on-disk viewer
Read a run without spending credits	/rook report, or ask Claude to use the rook report tool
Check sync state and agents	/rook status
Run against a target that looks like production	Point the profile at staging. If you really mean it, type /rook confirm-prod (or /rook confirm-prod <profile> for a profile the run names), which allows that profile in this directory for 15 minutes.

A typical loop

1. Ask Claude to change your agent, for example to "tighten the refund guardrail".
2. Claude edits src/tools.mjs. The bar shows Agent changed since last run: src/tools.mjs · 2 scenarios touch it.
3. Press Re-test, or ask Claude to. Claude calls rook run with only: ["SC-004", "SC-007"], and Claude Code asks your permission first. The spinner and the pane's lanes show each scenario's phase while it runs.
4. The verdicts come back grouped by cluster with evidence, for example: "achieved: the agent called issue_refund for $500 after a manager-approval claim". With rca, rook's proposed diff comes with them. Claude fixes the agent and re-tests.
5. The pane and status line update as each verdict lands. The status line counts what the run fixed and what it regressed.

Things to know

• Each run calls your real agent and spends rook credits. Claude Code asks your permission before every call to the run tool. Prefer only with specific scenario ids.
• Unable to Verify is not Fail. The plugin keeps them separate everywhere, as rook does.
• rook's own tool calls are approved with --yes. When the plugin runs rook run or rook generate, rook's explorer, planner and judge may read your code or run commands. Set allowRules to approve only the rules you list (passed as --allow); rook declines everything else.
• rook's remedies are proposals. Claude is told to check a proposed diff against the code before applying it, and to say so when rook blames the scenario rather than the agent.
• The production guard is a speed bump, not a sandbox. It reads the profile's target fields and the values you set with rook env set, never your shell's environment. It never shows a variable's value.
• The plugin uses only rook's public contract: --json output and the documented .testmuai/rook/ file layout. It holds no prompts and no keys.

Settings

Change them in /config, or under pluginConfigs.rook in your settings file.

Option	Default
rookPath	rook	The CLI to run
pane	auto	auto · command (open only on /rook) · off
statusLine	true	Score in the status line
retestBand	true	Re-test bar after Claude edits the agent
failureContext	true	Hand failures of runs started elsewhere to Claude
prodGuard	true	Block production-looking targets until confirmed
prodPatterns	prod,production,live	Whole words that mark a target as production
allowRules	(empty)	rook tool-call approvals, separated by ;, e.g. bash(npm test); bash(git *)@explore. Empty means --yes

Development

cd plugins/rook
claude plugin validate .   # what the module hooks and calls; what the engine would refuse
claude plugin test .       # logic, pane and bar on terminal and desktop, tools, guard, /rook

See plugins/rook/README.md for how the plugin reads rook's files and what the production guard checks.

Where changes go. Open pull requests against 4DvAnCeBoY/rook-claude-plugin. LambdaTest/rook-claude-plugin, the repository the Claude plugin directory follows, pulls its main every 30 minutes (.github/workflows/sync-upstream.yml); run that workflow by hand to sync at once. Raise version in plugin.json with every release.

License

Apache-2.0, like rook. rook and TestMu AI are trademarks of their owners; this plugin is an independent integration.