AgentUse v0.20.0: revise an agent from the run that went wrong

AgentUse v0.20.0 is out. The headline is agent revision: a finished run, or one paused at an approval gate, can now diagnose itself against its agent file and propose a change you review before it applies. Around it, the dashboard gets a command palette and navigation that keeps running work one keystroke away.

Get it as AgentUse for Mac, which bundles the runtime and dashboard, or run npx -y agentuse@latest setup in a terminal with Node.js 22+. Existing installs update through the Mac app's Settings → About, or by running the same command again.

Revise the agent, not just the run

Until now the loop for fixing an agent was manual. You read the transcript, worked out which instruction was missing or wrong, opened the file, edited it, and hoped the next run confirmed the theory. A reviewer comment could steer the current run, but the agent file itself did not change.

v0.20.0 adds a visible internal revision session. From a completed run, or one waiting at a gate, choose Revise Agent. The reviser reads the originating transcript against the current agent source and a sanitized view of the project. If one material decision is missing, it pauses in the same session and asks. Otherwise it returns either a validated source proposal or an explanation of why the agent should not change.

A proposal shows its source diff and its capability changes before you apply it. Applying is atomic and guarded by a source-hash check, so a stale proposal cannot overwrite an edit you made in the meantime. The previous version can be restored, and the originating run keeps a durable link to the revision session that came from it.

Revision runs appear in both Sessions layouts and their detail view as AgentUse-owned internal sessions, so they never masquerade as your agent's own work. When the fix needs project code or a custom integration rather than a prompt change, Copy prompt for Coding Agent is still there inside the revision form.

A command palette and live-work navigation

⌘/Ctrl+K opens a command palette over pending approvals, recent sessions, agents, pages, and quick actions. Mobile gets a dedicated search button.

The navigation rail now shows the running-session count and recent sessions, stays usable as an icon rail when collapsed, and answers ⌘/Ctrl+1–6 for section shortcuts. Session detail presents a finished run's outcome, failure, timing, metrics, and artifacts without a reload. Logs filter to agent activity, tools, or errors, and reading position survives a trip into a child session and back.

Parent session logs keep delegated work visible, show nested running subagents with their live activity, and no longer show a working spinner while a run is actually waiting on a human. Approval gates render pick decisions, risks, source context, keyboard behaviour, scoped feedback, and payload-carrying actions consistently.

Recovery and the Mac app

Interrupted setup, onboarding, revision, continuation, and delegated runs recover more reliably across preparation, restart, retry, and request-changes boundaries, and they keep their session evidence while doing it.

AgentUse for Mac now reconnects to an external server after that server restarts, instead of starting a competing local one. The desktop app only replaces an external server after an explicit Start Server.

v0.20.0 closes a loop that has been open since the first release. An agent runs, something is off, and the fix now starts from the evidence rather than from memory of what you meant to write.