Dashboard

Open it

In the harness-demo checkout, run:

pnpm dashboard

Then open http://127.0.0.1:4777. The dashboard listens on 127.0.0.1 only. Its buttons write the same files as the CLI, so you can use both side by side. What Start, Pause, Resume and Stop do is on Running the loop.

Projects

The first screen lists every project the harness knows. A project is a local git repository the loop works in.

1234

1. 1The terminal does the same: harness projects add and harness example.
2. 2One project: its tasks, how many of them wait for you, and the unit in flight (T-001 in qa).
3. 3Open board goes to the project's board. Remove asks you to confirm.
4. 4Add a project takes any local git repository. The loop creates .harness/ in it and touches nothing else until you submit a task.

The Projects screen of the dashboard with the bookmarks project and the Add a project form.

Board

One board per project. Each column is a task status, and each card is a task. New task opens a two-line draft.

123456

1. 1The header counts the tasks that wait for you and says why each one waits. Drafts get a sentence of their own, because only Submit starts the loop on them.
2. 2Thirteen columns in five groups: Define, Plan, Build, Review, Finish. An empty column folds into a narrow rail.
3. 3A card shows its stage track. When the card waits for you, it shows one action: Submit, Answer, Approve, Review or Reopen.
4. 4The spinner marks the card a unit works on right now. At most one card has it.
5. 5A card in parked names the contract that stopped it: here a file, README.md, changed outside the task's scope.
6. 6The Live dock: the supervisor with Pause and Stop, the unit in flight, what comes next and the recent units.

The board of the bookmarks project during a qa unit, with the Live dock on the right.

A task: the interview

Click a card to open its drawer. A task in needs-input opens on the Interview tab, with the questions an agent asked.

1234567

1. 1The task you opened stays outlined on the board.
2. 2Who asked and when: the writer, in the spec stage, in unit 5.
3. 3The task file in tabs. Interview appears only when the task has questions.
4. 4Each question has lettered options. The recommended one comes first, with its reason.
5. 5An optional note for the agent goes with each answer.
6. 6One button writes all the answers into the task file.
7. 7Your own draft stays at the bottom of the drawer.

The task drawer of T-003 with the writer's two questions and their lettered options.

A task: the prototype

A task with kind: ui stops in prototype before it is built, and the drawer opens on the Prototype tab.

123456

1. 1T-001 is a UI task, so it stops in Prototype before it is built. The card shows the round and asks you to Review.
2. 2The Prototype tab appears for UI tasks.
3. 3Accept moves the task to ready.
4. 4Or write what the next round should change. The drawer shows how many rounds are left.
5. 5The round's picture from the task's artifacts folder. With an app, the designer screenshots the running route as prototype.png at 1440×900; without one it draws prototype.svg, a 1440 px wireframe with the page's real copy, layout and colours. Here the fake agent drew it.
6. 6Your draft, to compare the picture with what you asked for.

The task drawer of T-001 on the Prototype tab, with Accept, a comment box and the round's picture.

A task: delivered

With deliver.pr on, a done task gets one more runner step: a push and a pull request. The board shows both sides of it.

1234

1. 1T-002 is done. With deliver.pr on, the card says PR next, and the grey spinner shows the loop takes it next.
2. 2T-001 is delivered. The card links to its pull request, PR #1.
3. 3The drawer says it again: Pull request #1, merging is yours. This URL comes from the fake gh, not from a real repository.
4. 4The stage track is lit to the end: delivered is the last status.

The board with T-001 delivered and T-002 done, and the drawer of T-001 with its pull request link.

Live

Live trace shows what a unit does while it runs, and what every earlier unit did.

123456

1. 1Every unit the loop ran, newest first. The top one runs now.
2. 2Unit 8 ended in a hard stop, and T-004 went to parked.
3. 3The unit in flight: its task, its stage, how long it has run and the specialists it may call.
4. 4The coordinator's lane: it handed the work to qa.
5. 5The specialist's lane: each tool call with its time and the file name. Tool arguments and messages are never shown.
6. 6The supervisor's log, newest first: starts, exit codes, status moves and commits.

The Live trace screen: the unit list, the running qa unit with its agents, and the supervisor log.

Stats

Stats adds up the usage each unit reported, per stage and per task. Cost is the API-equivalent price claude -p reports: a subscription login is not billed per token, and usage limits apply instead. harness stats prints the same tables in the terminal.

123456

1. 1Totals: units, API-equivalent cost, tokens, agent time and denied tool calls. Cost is the price claude -p reports; a subscription login is not billed per token.
2. 2A unit whose agent printed no usage counts as unknown, not as zero.
3. 3Per stage, with its coordinator, the parked count and a bar for cost.
4. 4Per task, with its current status.
5. 5The verdicts the coordinators printed: qa PASS, judge CONFIRMED.
6. 6Tool calls the permission mode refused. Frequent ones point to a missing entry in agent.allowedTools.

The Stats screen: totals, costs per stage and per task, verdicts and permission denials.

Agents

Agents shows who works each stage and how a unit starts Claude Code, so you can check a setup before a run.

1234567

1. 1One coordinator per stage, each one claude -p on its model. The judge runs on opus.
2. 2The specialists each unit gets through --agents. A dot marks the stage that offers one.
3. 3The checks that refuse a run, summed up.
4. 4The checks: the judge's model differs from the implementer's and the tester's, and every agent carries the current standing rules.
5. 5How a unit starts Claude Code. This orange card appears only because AGENT_CMD points at the fake agent; a real project shows the plain claude command. With a real claude, a Billing row also says how the run is billed.
6. 6Delivery and Telegram, both off here. Delivery needs deliver.pr; Telegram needs TELEGRAM_BOT_TOKEN and TELEGRAM_CHAT_ID.
7. 7Permission mode, setting sources and allowed tools. Commits and pushes are absent on purpose: the loop commits only after the contracts pass.

The Agents screen: the stage coordinators, the specialist matrix, the checks and the invocation.