Agent CLI
The commands agents call, hidden from uze --help
uze agent … is written for agents, not for you: it is hidden from
uze --help, and the agents learn it from the instructions uze gives every
agent it launches and from the Skills. You can run
context, artifacts, rules, market and plugin yourself; work answers only an
agent uze workspace launched.
uze agent work name <type>/<subject> # name the work: the tab's label, and the branch in a checkout of its own
uze agent work split <topic> # a checkout for a subagent, printed as a path
uze agent work join <topic> # bring the subagent's commits onto this branch
uze agent work list # this agent's subagent checkouts
uze agent memory remember <name> # keep a fact for later agents, read from stdin (-m: the machine's)
uze agent memory list [words] # the facts kept here, or those holding any of the words
uze agent memory show <name> # one fact in full
uze agent memory forget <name> # remove a fact
uze agent review show # the review comments sent to this agent and not yet answered
uze agent review reply <send>/<n> "<answer>" # answer one comment, by the number show printed
uze agent rules check [path] # validate every rule under .agents/rules, offline
uze agent artifacts check [path] # draw every declared diagram, fail on one that does not
uze agent context inspect [path] # what each agent currently sees of AGENTS.md
uze agent market create <name> --at <dir> # scaffold, register and link a marketplace (--description)
uze agent market create <name> --local # the project as its own marketplace (--plugins-dir, default plugins)
uze agent market check <path> # validate a marketplace, offline
uze agent plugin create <name> --market <m> # scaffold a plugin (--hook --mcp --agent --rules --extension <id> --category --description)
uze agent plugin check <path> # validate a plugin offline, and run its extensions' tests (--update-snapshots)
uze agent plugin lsp # completion, hover and diagnostics for extension files, over stdioThe commands above take --format json, except work split, work join,
memory forget, review show, review reply and plugin lsp; split prints
a path alone.
Naming work
uze agent work name only works inside an agent uze launched, and names that
agent's own work. It always sets the agent's label. In a checkout of its own it
also renames the branch; in your checkout only the label changes.
The types it accepts come from workspace.branch in agents.yaml: undeclared,
the Conventional Commits types (feat, fix, docs, …); a preset
(conventional, gitflow, flat, which takes a subject with no type); your
own list; or agent, which names no work. The subject is lowercase letters,
digits and single hyphens, at most three words. Work that reaches a commit unnamed is named from the commit's subject; a branch
you rename by hand is never renamed again.
Subagent checkouts
split gives the calling agent a checkout for a subagent, cut from its
current commit and taken from the same pool as an agent's. It prints the path
alone, so cd "$(uze agent work split parser)" works. Asking again for the
same topic answers with the same checkout.
join needs both checkouts committed. It replays the subagent's commits onto
the agent's current commit and moves the agent's branch forward, with no merge
commit. On a conflict the replay pauses in the subagent's checkout: resolve it
there, run git rebase --continue, and join again.
Only an agent in a checkout of its own has subagent checkouts. Every refusal says which condition held.
Checking diagrams
uze agent artifacts check exits non-zero when a diagram does not draw, draws
with a relation that found no path, or has a box whose link opens no file in
the project. Put it in your project's gate to
catch a diagram that broke silently.