Space

The overview is one screen about the agent in front: what it was asked, its settings, and its project's plugins, memory and worktrees. Press Alt+O, or click the ⋯ button at the right end of the tab strip, to open it. Alt+O or Esc puts it away.

Its five sections are Agent, Settings, Plugins, Memory and Worktrees. 1 to 5 jump to one, and ← → move between them. Select a row to read its detail in the drawer on the right.

Agent

Every prompt this agent was given, newest first and grouped by age: the harness, when it was asked, and the prompt itself. Enter copies the selected prompt. X clears this agent's prompts, and no other's, after asking, with the keyboard on cancel.

The drawer shows the agent: its harness, its checkout, and what of the project reaches it (AGENTS.md, project skills and agents) and how it receives skills, MCP servers, agents and hooks.

Settings

The workspace: section of agents.yaml as a form, in four groups.

Enter edits the key you are on, - and + step a number, a adds an entry to a list and d removes one or unsets the key. Pickers offer only what makes sense: the repository's branches for the target, files Git ignores for linked files, tracked directories for artifacts. An agent action is a text box of its own, where Enter starts a new line and Ctrl+S finishes.

Each change is written to agents.yaml as soon as you make it, and the key shows saved. It is the file of the checkout the agent works in: for an agent in the project's own tree, the project's, which places every new agent; for an agent in a worktree, that worktree's, so the change is part of its work, committed and delivered with it, and new agents follow the project's file until then. The file is read again before every write, its comments are kept, and the result is checked by the same parser the workspace reads it with: a change that would not parse is refused, shown at its key, and nothing is written. x discards refused changes. When the file uses something the form cannot edit safely, such as a YAML anchor, the section is read-only and says why.

Setup and Gate are commands from the repository, so a change to either waits for your approval again before anything runs. The drawer's Review button, or v, opens the same approval as uze workspace allow; see Approving the commands.

The drawer explains the key you are on, its default and the value it has now.

Plugins

The plugins the project declares, with their marketplace, the commit installed and where each stands: installed, linked, behind or with an update available, different from agents.lock, or missing. Below the list, anything that has drifted between agents.yaml, agents.lock and the store is spelled out. uze install puts it right; see Plugins.

Memory

Agents keep facts for the agents that come after them, on any harness: how you want work done, a constraint of the project, where something lives. A fact is a short Markdown file with a name, a one-line description and a kind (feedback, project, reference or user). uze keeps them outside every checkout, readable by you alone, at two scopes: the project, shared by every agent launched in it, and the machine, shared by every project.

Every agent uze launches is given an index of that memory in its context, whatever its harness: every fact by name, then as many with their description as there is room for, feedback first, and how to read one in full. A fact that cites a file this checkout no longer has is marked, so an agent checks it before relying on it. Agents write facts through uze, on any harness (Agent CLI). A harness that keeps memory of its own, like Claude Code, goes on writing it there, and uze reads it where it is rather than copying it. A machine fact an agent proposes waits for your approval before it reaches any agent. Facts are notes, never instructions, and memory holds no secrets.

This section lists the project's facts and the machine's together, newest first: their name, kind, scope, origin (the harness that wrote it, or uze for one written outside an agent) and when they changed. / searches by name, description, kind, scope and origin as you type, and Esc clears the search. Enter reads the selected fact in the drawer, which also shows the file it lives in; the wheel or PageUp/PageDown scrolls a long one. Nothing is changed from here: a fact is a file, and you can edit it by hand.

Worktrees

The checkouts uze made for the project's agents and their subagents, as cards: the branch, its size on disk, how long since it changed, and how far ahead it is. They are sorted by what they need from you: ready to deliver, to join into the agent that split it, kept work nobody finished, then those an agent is working in, idle, and free for the next agent. A checkout with uncommitted work and no agent is flagged, since nothing but you will pick it up. Checkouts you or a tool made by hand are not listed.

A free checkout no agent holds can be removed with d. uze asks first, deletes its directory with its build output, and keeps its branch and every commit on it. It refuses if anything is working inside it, if an agent holds it or if it has uncommitted work. See Where an agent works for how checkouts are reused.

Moving around

The overview is there only for an agent and the shells opened beside it: a shell of the space works in no project, so it has none. It stands where the pane is, over the tab in front. Selecting another agent turns it to that agent; selecting the space's row or one of its shells, or opening Spec, Architect or Code, puts it away. When you open it again it is on the section you left, and it reads the project again.

The top line names the agent and its directory. A section with something to look at carries a mark, and its first line says what it is. The foot shows the keys for the section you are on and, on the right, how many things are left to check. Click that count for the whole list.

Drag the drawer's edge to make it wider. On a narrow screen it opens over the list instead, with Enter, and closes with Esc.