Terminal
The server behind the workspace, resuming after a reboot, prompt history and logs
Closing the workspace never stops an agent. Its panes belong to a local
server, not to the window you look at, and the next uze workspace attaches
to it again.
uze workspace # attach the workspace, starting its server when needed
uze workspace stop # stop the workspace's server and every process in it
uze workspace allow # approve the commands this project's agents.yaml runs
uze workspace revoke # take that approval backRunning uze workspace inside one of the workspace's own panes does not nest a
second workspace: it selects the space already working in that project in the
one already running, or opens one for it.
Resuming after a reboot
The server does not survive a reboot, but your work does. The next
uze workspace brings back the spaces, tabs and agents, each relaunched in its
checkout and resuming the conversation it was in.
That resume is done by the shim uze setup creates for each agent in
~/.uze/shims/. The workspace launches agents through it, and puts
~/.uze/shims first on PATH in every pane, so an agent you type into a pane
goes through it too. Your own shell is never edited: an agent you start in any
other terminal is the agent's own binary, exactly as it would be without uze.
Besides the resume, the shim carries what some agents do not read on their own:
the project's .agents/skills and .agents/agents for Claude Code, and the
project's agents for Codex. Those reach them inside the workspace only; the
plugins you install reach every agent everywhere. Whether each agent supports
the resume is the Session column in Harnesses.
If a pane's shell puts another copy of an agent ahead of the shims (a version
manager such as mise, asdf or nvm can), the workspace says so the first
time that agent runs there: its conversation will not resume after a restart,
and the project's own skills may not reach it.
Prompt history
What you send to an agent is recorded per project, and read and cleared per
agent. The Agent section of the overview
(Alt+O, or the ⋯ mark at the end of the tab strip) lists that agent's
prompts, newest first, and Enter copies one. It is a convenience, not a
transcript: uze forwards keystrokes without interpreting them, so entries can
be missing. The file is readable only by you, and X clears an agent's
prompts after asking.
Tab marks
A shell tab running a program shows ▸. When the program ends while you are
on another tab, the shell shows ✓ until you look at it. A return to the
prompt counts only after two seconds, so the gap between two commands of
make && ./run is not taken for an end.
Opening a path
Hold Ctrl over a file or directory path a program printed in a pane, whichever
program it was, and the path is underlined once it is found on disk. Ctrl+click
opens it in Code: a file selected, a directory unfolded.
On macOS Alt works too. A click on anything else goes to the program as usual.
Agent panes are sealed
On Linux 6.12 or later, the panes agents run in are sealed from the
workspace's own controls, and nothing in them can gain privileges: sudo and
su do not work there. Run them from a shell tab. See
Security.
Notifications
uze config notification takes three states, and test:
| State | Rings for |
|---|---|
silent | nothing (the default) |
off | a finished agent in a tab you are not looking at |
on | every finished agent |
test | once, now, so you hear what it sounds like |
It is your terminal's own bell, rung after an agent has stayed finished for about ten seconds; agents finishing together ring once.
Logs
The workspace and its server keep a journal without being asked, in
~/.uze/cache/logs/, rolled daily and kept for a week. For a trace of a single
command, set UZE_LOG=info.