CLIDesktop

Changelog

Release notes for the Agent Relay desktop app. Full release artifacts are published on GitHub.

0.1.4

  • Quitting the app no longer leaves agents running. Quit stopped the runs it launched, but subagents, reviewers, handoffs, and race participants are separate processes and none of them were stopped. They kept working against your repository after the app was gone, and nothing on relaunch knew they existed. Races needed a different fix from the rest: their agents live in tmux, outside any process group the app can signal, so the app now asks its race helper to shut down and lets it close its own sessions. If you have used /subagent, /review, or a race, it is worth checking for leftover processes once before updating.
  • The menu bar tells you what your agents are doing. Closing the window has always kept your work running, but the menu said nothing about it. It now shows what is blocked on you, what is still working, and gets you back to that conversation in one click. See Menu bar.
  • A run waiting for permission can no longer go unnoticed. Approvals were only listened for while the Tasks view was on screen, so a request that arrived while you were in Settings or Review was dropped rather than queued, and the run waited forever on a question you were never shown.
  • A run that finished while you were elsewhere is no longer stuck. Run and orchestration events were handled only by the Tasks view, so leaving it during a run meant the conversation never learned it had ended and stayed "running" indefinitely.
  • Promoting a race winner starts a real conversation. It used to launch with no chat attached, which meant its turns had nowhere to appear and, worse, an approval request had nowhere to be answered.
  • "Recent" now means recently used. Chats were ordered by when they were created, so a conversation you worked on this morning could sit below one started last week and abandoned.
  • Bundles Agent Relay CLI 0.7.7.

0.1.3

  • Rate-limit handoffs now pick the right agent, and tell it the right thing. When an agent runs out mid-conversation, the app used to continue on whichever name came first in a fixed list. It had no idea which agents were installed, no memory of which ones were already spent, and nothing to say to the incoming agent beyond your last message.

    Now, once the running agent passes 90% of a usage window, Agent Relay asks it, while it still has quota to answer, what's left to finish and which of your installed, signed-in, not-already-limited agents is best suited to continue. You see one quiet line (claude at 91%: codex will continue if the limit hits), and when the limit arrives the switch is immediate, because the decision was already made.

  • The app remembers which agents are spent, and shares that with the CLI. Rate limits survive quitting the app, and are recorded with the provider's real reset time. Burn through Claude in the terminal and the desktop already knows not to send work there.

  • When a window resets you're offered the agent back, rather than switched to it automatically.

  • "No agent can take over" now says why. Instead of a bare notice, it names the obstacles: "codex is rate-limited, gemini is not installed".

  • Bundles Agent Relay CLI 0.7.7.

0.1.2

  • "Restart to update" now installs the update. On 0.1.0 and 0.1.1 the update downloaded and verified, but clicking restart did nothing and further clicks were ignored. Because the fix has to be in the version doing the updating, anyone on 0.1.0 or 0.1.1 needs to install 0.1.2 manually, once. See Updating. In-app updates work from 0.1.2 onward.
  • A real menu-bar icon. The tray showed an unlabelled dark shape; it now shows the Agent Relay mark, and adapts to light and dark menu bars.
  • Clicking the menu-bar icon opens the window, not just the menu.

0.1.1

  • Agent replies are no longer discarded as failures. A run was judged by scanning the agent's reply for words like "error" or "failed", so ordinary successful turns such as "I fixed the error in auth.py" were reported as provider failures, with the reply itself thrown away. The app now reads the structured error the CLI reports instead of guessing from prose.
  • Bundles Agent Relay CLI 0.7.6.

0.1.0

The first public release of Agent Relay Desktop.

  • Native macOS app (Apple Silicon), signed and notarized.
  • Launch and supervise agent runs for Claude Code, Codex, Gemini, and Kimi from one window.
  • Live chats: watch every turn, tool call, and file change as a run progresses.
  • Per-agent model selection with launch defaults for agent, model, command approval, and rate-limit behavior.
  • Automatic rate-limit failover between agents.
  • Signed, in-place automatic updates on the stable channel that never interrupt active work without confirmation.
  • Local-first: chats and session data stay on your machine.