Can the dashboard and the CLI disagree about a run?

No. Both render one canonical projection, and the verdict resolver has a single call site in the whole product. The dashboard's JavaScript displays state; it never calculates one. A screen that can invent a second opinion is a screen you have to check, and then you own two sources of truth instead of none.

What do I see while a run is still going?

The run itself, immediately. The Workspace and Runs views show it the moment your prompt lands, marked as observed. Reviews shows the same incomplete run honestly while the trust review is pending: no invented verdict, no placeholder review ID. When the run stops, that same run finalises in place.

Is anything exposed to my network?

Nothing. It binds 127.0.0.1 and refuses any other bind address, ships its own assets and loads nothing from the internet. The port is isolated per worktree, so two projects never fight over one. If your preferred port is taken, it takes the next free one and prints the real URL rather than failing.

What happens if the dashboard will not start?

You keep working. A dashboard failure is recorded as a workspace limitation and never blocks the coding session, the run, or the review. The CLI prints the same facts, and the record on disk is the same record.

FeatureDashboard

Every run you have ever done, still open

The local AI coding dashboard runs on 127.0.0.1 and shows your whole project in one workspace: every run, every review, every executed check, every piece of evidence. It renders exactly what the CLI prints, so the two can never tell you different things about the same change.

Request access See how it works

Last updated: 2026-08-13

The workspace

Eight views, and the standards every run starts from

Everything a run produces lands in the same place and stays there. So does everything a run should have known beforehand, which is the half most tools leave in somebody's head.

View	What it is for
Workspace	Project Knowledge: the documents, decisions and conventions this project works to, in one local place instead of in six people's heads. It is what keeps a second run to the same standard as the first.
Runs	Every run, live and historical, resolved by exact ID and surviving restarts. The one from six weeks ago opens as readily as this morning's.
Reviews	The Trust Review for each run, and honestly nothing at all while one is still pending. It never invents a verdict or a review ID to fill the space.
Validation	Which checks are approved, which Reality Graph executed itself, and the exit code each one returned.
Evidence	The changed files, the diff and the receipts behind every claim the verdict makes. This is where you go when someone asks how you know.
Roadmap	The live board. A lane moves only when a real recorded event justifies it, so nothing on this screen animates to look busy.
Content	Content governance, for the projects where it applies. Everywhere else the tab stays away rather than showing you an empty box.
Settings	The project, its route, its approval mode and which agents are bound to it.

The dashboard views and what each one is for. Every one reads the same canonical projection the CLI prints, so no view can hold an opinion the others do not.

Project Knowledge is the part worth pausing on. Standards that live in a wiki nobody opens are standards the next run will not follow, so they live next to the work instead: copied into a local, git-ignored store, presented as metadata, never executed, and never sent anywhere. Files are stored rather than indexed, which is the deliberate first step and the honest description of it.

Why you can trust the screen

The dashboard displays. It does not decide.

This is the design decision the whole feature rests on, and it is worth being precise about.

run-4f2a, as the workspace shows it

Example – illustrative view, not real run data

run       4f2a                       observed · claude code
mission   fix the checkout validation bug, payments untouched
changed   3 files, 412 lines          read from Git
checks    pytest -q   exit 0          executed by Reality Graph
verdict   BLOCKED                     protected path changed
                                      src/payments/gateway.py
review    rv-91c                      opened 14:22, finalised 14:31

Every value on that screen was computed once, by the same resolver the CLI uses, and then rendered. The browser does no arithmetic on your evidence. That is why the dashboard cannot drift from the terminal, and why a screenshot of it is worth something in a conversation about what shipped.

The review summary keeps one distinction most tools blur: what was declared and what was proven. A requirement you wrote down and a requirement an executed check confirmed are different facts, and they are labelled as different facts.

By design

What this workspace is, and what it deliberately is not

A local-first workspace buys you specific things and gives up specific others. Both lists are short.

What you get

Every run and review in a local store that survives restarts and resolves by exact ID.
The same canonical projection the CLI prints, in every view, with no state computed in the browser.
127.0.0.1 only, its own bundled assets, and no request to the internet.
A port isolated per worktree, with an automatic free-port fallback that prints the real URL.
Your records as plain local files, exported on demand and scrubbed of home paths and secrets first.

What it is not

A team tool. One operator, one machine, no accounts, no authentication, no sharing.
A second opinion. It cannot produce a verdict the CLI would not produce.
A requirement. If it fails to start, that is a workspace limitation and your run continues.
A hosted service. There is no server to sign in to and nothing to keep online.

How the verdict on that screen is computed is covered on the verdict page, and what never leaves your machine is traced on the data-path page.

Questions people actually ask

Can the dashboard and the CLI disagree about a run?: No. Both render one canonical projection, and the verdict resolver has a single call site in the whole product. The dashboard's JavaScript displays state; it never calculates one. A screen that can invent a second opinion is a screen you have to check, and then you own two sources of truth instead of none.
What do I see while a run is still going?: The run itself, immediately. The Workspace and Runs views show it the moment your prompt lands, marked as observed. Reviews shows the same incomplete run honestly while the trust review is pending: no invented verdict, no placeholder review ID. When the run stops, that same run finalises in place.
Is anything exposed to my network?: Nothing. It binds 127.0.0.1 and refuses any other bind address, ships its own assets and loads nothing from the internet. The port is isolated per worktree, so two projects never fight over one. If your preferred port is taken, it takes the next free one and prints the real URL rather than failing.
What happens if the dashboard will not start?: You keep working. A dashboard failure is recorded as a workspace limitation and never blocks the coding session, the run, or the review. The CLI prints the same facts, and the record on disk is the same record.

Open the workspace and read last week's run

The one you already shipped. That is the moment this page is about: the record is still there, and it still says what happened.

Request access See how it works

Keep going

Step 7 of 9

BackApprovalsThe gate a green run cannot pass Next RoadmapEvery prompt becomes a ticket