Use the workbench

The workbench is organized around research objects, not a chat transcript. The navigator groups mode actions, project links, and the current run tree. The workspace shows the selected objective, result, notebook, artifact, or writeup. The inspector explains the selected project, run, task, attempt, claim, or artifact.

Move through the shell

The hosted workbench uses an email sign-in link. After the link restores the session, the account control in the top toolbar shows the authenticated email and provides Sign out. Project-loading errors keep that account control visible and say that you are signed in, so an unavailable API is not mistaken for a failed sign-in. A link that is invalid or expired returns to the email form with an explicit recovery message; request one new link and use the latest message.

The application menus sit above the location bar. File opens the project overview, its settings, closes the current tab, and reloads the saved projection. Edit copies the project, run, and artifact identifiers. View shows or hides the navigator, inspector, and research activity, splits the workspace, and sets the appearance. Go opens the command menu and every saved object in the current run, and steps between open tabs. Run reconnects live updates and refreshes the projection. Help opens the command menu and the documentation. Every menu entry names a command the workbench can actually run, and each one is also reachable by keyboard and by pointer elsewhere.

At wide viewports, the navigator, workspace, and inspector remain visible as three stable regions. Resize either side region by dragging its separator or using the separator's keyboard controls. The navigator can collapse to a 52px icon rail; its delayed tooltips preserve the same labels, and the current run's task, attempt, notebook, proposed claim, and artifact remain nested beneath that run. At intermediate widths, the navigator stays beside the workspace and the inspector opens as a sheet. At the narrowest widths, the workspace remains primary and both side regions open as labeled sheets. No research object or command disappears at a smaller viewport.

Open Commands with the toolbar button or Command/Control + K. The palette navigates to the project, settings, or current run and opens the same navigator, inspector, and activity controls available by pointer. Open Research activity from the toolbar, the command palette, or Command/Control + E. Activity opens in a focused bottom sheet instead of competing with the research workspace. The quiet status bar reports connection state and update count; reconnecting never changes the selected object's identity.

Work on more than one object at once

The workspace holds a tree of panes, and every pane holds a strip of open research objects. Opening an object from the navigator, the command menu, or a Go entry opens it as a tab in the active pane, or reveals the tab that already shows it — the same object never opens twice.

Split the workspace to compare objects side by side. Use View → Split pane right or Command/Control + \, Split pane down or Shift + Command/Control + \, or the split controls on a tab strip that holds more than one tab. You can also drag a tab: dropping it on another pane's tab strip moves it there at the exact position you drop it, and dropping it near a pane's edge splits that pane and opens the object in the new half. Dragging is never the only path; every split and move is also available from the menus and the keyboard.

Alt + Right and Alt + Left step through every open tab across every pane. Command/Control + W closes the active tab. Closing the last tab in a pane removes that pane and gives its space back to its neighbour; the workspace always keeps one pane to open the next object in.

The active tab of the focused pane is the route. Navigating changes that tab, and activating a tab changes the URL, so a shared link still restores exactly one selected object. The pane tree itself is a local workspace preference kept per project in this browser, not durable project state, so it never competes with the saved settings below.

Compose with saved context

Type @ in the objective composer to find saved projects, runs, tasks, documents, claims, and artifacts. Use the arrow keys to move through the grouped results, then press Enter or Tab to insert a reference. The draft keeps that reference structured while the submitted objective remains readable as @Label. Press Enter outside the open reference menu to start the run, or Shift + Enter to add a line. The send control repeats the same command for pointer users and reports its own progress while the run starts.

The completed story leads with the agent's proposed claim, its untested review state, and the action that opens exact evidence. Accepting or rejecting that proposal is an explicit human decision; an attempt cannot promote its own conclusion. Task, attempt, notebook, and writeup rows follow in one scan path. Source, output, stored markdown, immutable locators, and record identifiers stay in closed technical disclosures until you ask for them. The inspector uses the same rule for technical identifiers.

Restore a selection

Workbench URLs name persisted objects. Project and settings routes use /projects/{project_id} and /projects/{project_id}/settings. Run-qualified routes can select the run itself or one of its blocks, tasks, attempts, claims, or artifacts. A hard refresh, browser back/forward navigation, or shared URL restores the same route before the shell loads its projection. The browser never invents an object that is absent from the loaded story.

Theme, navigator width, inspector width, and collapsed states are project settings, not local browser preferences. Each change is saved with the observed project version. While an accepted change is refreshing from the canonical projection, the affected controls remain unavailable. If delivery is uncertain, retry the same command. If the version is stale, reload the project and reapply only the intended setting.

Core nouns

• A project owns durable research state.
• A run pursues one objective under an explicit plan and budget.
• A task is durable intent; an attempt is one immutable execution try.
• A block revision is exact source; an execution records what ran.
• An artifact is immutable byte content.
• A claim revision states a conclusion and links to immutable evidence.

The deterministic research story remains reachable inside this shell. Provider-backed attempts, schedules, extensions, and larger task graphs use the same objects and routes instead of replacing them with transient chat state.

An attempt's tools also preserve that distinction. Local file discovery, edits, bash, persistent scratch python, and bounded ephemeral subagents operate only inside the disposable attempt sandbox. Web research, connectors, remote MCP, canonical claims/evidence, messages, task controls, and workspace publication remain trusted Lumen operations. Scratch work becomes part of the project only after an explicit publication or research-recording step succeeds.