Workbench

A board of tickets. Each ticket is a piece of work you want done. Agents do the work; you decide what gets done and whether it is good enough.

Every ticket goes through four stages — plan, implement, review, verify — each one a separate agent with its own tools. You approve the plan before any code is written, and accept or reject the result as a pull request.

Each ticket gets its own git branch and worktree, so tickets running at the same time cannot collide and your own working tree is never touched.

What you need

• Node 26+ — it uses Node's built-in SQLite and its TypeScript support, so there is no database to install.
• git, and gh authenticated, for pull requests.
• A Claude credential. claude setup-token gives a long-lived one from a subscription; ANTHROPIC_API_KEY uses a Console key, billed per token.

Getting started

In the repository you want work done in:

npm install github:peterbjohnson/workbench
npx wb init      # writes .workbench/ — config, and somewhere for your own skills
npx wb auth      # spends a fraction of a penny proving the credential works
npx wb serve     # the board, on http://127.0.0.1:4600

Then, from anywhere in that repository:

wb new "Add a retry" "It gives up on the first failure."
wb queue t1        # commit to it; nothing starts before this
wb list            # every ticket and where it is
wb approve t1      # let implementation start
wb show t1         # the plan, and everything that happened

There is nothing to release and no version to track: npm install resolves your dependency to a commit and writes it into your lock file, and wb update asks the same question again later. It predicts nothing — it installs, then says what actually moved, so a tag, a branch or a commit all behave the way you told npm they should.

wb serve says at startup when something newer is waiting, and never does anything about it. A dependency pinned to a tag is silent, because its tag has not moved.

wb update          # fetch whatever has been pushed since; then restart wb serve

wb --help lists every command. Everything the command line does, the board does too — they are both clients of wb serve, which is the only process that touches the database or calls an agent.

⚠️ wb serve calls real agents and spends real money. WB_RUNNER=fake wb serve runs a scripted agent that makes no external calls but does real work on disk, so worktrees, commits and diffs are genuine. It is the way to try this without spending anything.

Where it stops for you

A ticket waits for you at three points, and guesses at none of them:

• The backlog — until you commit to it. The workbench never touches this column.
• The plan gate — the plan is written and waiting for your approval. This is the one place a person sees the work before any money goes on building it.
• Anything an agent cannot decide — it stops and asks rather than choosing for you.

Finished work arrives as a pull request. Merging it accepts the work; requesting changes sends it back, and what you write becomes what the next attempt is briefed with.

Documentation

• Using it — the board, the ticket panel, and how work moves
• Reference — commands, configuration, limits, the HTTP API, tools
• Design — why four stages, and why they are allowed to say what they say
• Archive — superseded documents, kept for how this got here

Working on the workbench itself

npm install
npm test
npm run typecheck

The tests need no credential, no network and no agents: the rules, the state machine, the guard and the orchestrator are tested with the outside world faked, and the git tests run against throwaway repositories.