Verify the fixture and command line.
The repository already contains the service, checks, workflow runner, evidence bundle and Round 0. Preparation confirms the current fixture behavior and shows the V1 target command contract. The current fixture still exposes diagnose, plan, validate and report; review, eval, define, tasks.json and run are the target handoffs to be implemented. No package install, external account or network access is required.
The change has to be reviewable first
The change has to be reviewable first
The course assumes a change you could read if you had to. It teaches you not to, and teaches you what to run instead, but the method needs a change with a shape.
- Why it matters
- A single build that produces 40 files and 3,000 lines is not a validation problem. It is a planning problem, and no workflow in this course fixes it. Oversized changes degrade every method the course teaches.
- Upstream
- Bound the work when the task is planned, not when the diff arrives. Split until each unit is independently reviewable and each has its own acceptance criteria.
- Rough guide
- If a change touches more than about 20 files or 400 lines across unrelated modules, look for the seam before you look for defects.
- Legitimate exceptions
- A mechanical rename across two hundred files is fine. Count alone cannot tell the two apart, which is why this is a judgment and not a gate.
Four steps, no installs
Use the sibling fixture
Use the sibling repository at ../payments-validation-fixture. The current verified revision is 836a75d.
cd ../payments-validation-fixturenode --versiongit rev-parse --short HEADnpm run verifynpm testnpm run workbench:doctor
- npm run verify
- Checks the fixture’s own schema, vocabulary, workflow and no-install rules.
- npm test
- Runs the payments service test suite. The verified baseline is 17 passing tests.
- npm run workbench:doctor
- Reports which hooks and commands are usable and which agent skills will report silence.
Verify the current fixture commands
The current command interface is vendored in bin/. Add it to this shell’s path; nothing is installed globally. These are the commands the fixture exposes today, not the V1 target sequence.
export PATH="$PWD/bin:$PATH"workbench diagnoseworkbench planworkbench validateworkbench report
- diagnose
- Creates a risk-surface worksheet for the learner to complete.
- plan
- Creates an editable validation workflow from the fixture baseline.
- validate
- Runs the workflow and writes a JSON evidence bundle.
- report
- Summarizes the most recent bundle, including classes not examined.
Run the verified baseline
The run exits 1 with one documented equivalent mutation finding in Ledger.net(). That exit means findings were produced, not that the runner failed. Agent skills report silence when no coding agent is connected, and F8 remains human-reserved.
git switch mainworkbench validateworkbench report
Read the V1 target command contract
The target stages replace the current validate/report loop with six explicit handoffs. Stage 2 chooses review methods before results are shown; Stage 4 audits defect files as a quality requirement, not as a separate command.
workbench diagnoseworkbench reviewworkbench evalworkbench defineworkbench planworkbench run
- YAML handoffs
- Diagnose, review and eval each write a YAML artifact.
- Defect specs
- Define writes defects/1.md, defects/2.md and any other confirmed defect files.
- Execution
- Plan writes tasks.json. Run consumes tasks.json and creates a new implementation diff and run output.
- Repeat
- After run, repeat diagnose → review → eval on the new diff.
Know how fresh-context review works
The fresh-context review is a provider-neutral skill contract and must begin in a session that has not seen the authoring transcript. The command prints the review contract. Invoke it in your coding agent with only the requirement and diff. If no agent is configured, the evidence bundle records silence rather than a pass.
workbench review --fresh-contextReady to start when
- The sibling fixture is at the expected revision or its newer state has been verified.
- npm run verify passes with 15 checks and 1 round.
- npm test passes all 17 tests and workbench:doctor reports every hook and command ready.
- The current fixture emits a bundle with workbench validate and names unexamined classes with workbench report.
- You can explain the V1 handoffs: YAML → defects/*.md → tasks.json → new diff, followed by Diagnose → Review → Eval again.
- You know how to start a genuinely fresh coding-agent session, or accept that agent skills will report silence.