Speak the task without sending your voice away
Some instructions are easier to say than to type: a bug you have just reproduced, the shape of a release check, or a list of small changes you can already see on screen. Onevium 1.1.22 adds voice dictation directly to the composer so those instructions can start as speech without leaving the workspace.
Recognition runs on your device on macOS, Windows, and Linux. On first use, Onevium downloads the speech model you choose; the available models are roughly 75–180 MB. After that, transcription works locally, including when the network is unavailable. Your microphone audio is not uploaded for recognition.
The transcript remains editable before it becomes a message. If you enable optional AI polish, only the text transcript is sent through the provider you already configured; the audio still stays local. Dictation history makes it possible to recover or reuse a recent transcript without recording it again.
- Private by default — speech recognition runs locally on your machine.
- Optional polish — clean up wording with your configured provider, or leave the raw transcript untouched.
- Designed for the composer — switch naturally between speaking, editing, and typing in the same input surface.
A dev server should be a service, not an orphaned terminal
Agent work often depends on something that must remain alive after the command that launched it: a Vite preview, a local API, a database, or a worker. Treating those processes as ordinary terminal output makes their real state hard to answer. Is the process alive? Did it bind the expected port? Is it ready to accept a connection? Will the agent find it after the app restarts?
Managed services in Onevium 1.1.22 make that lifecycle explicit. Onevium records the real process ID, checks operating-system port ownership, waits for TCP readiness, and restores the service record when the desktop app reopens. The Activity Panel shows the current state and exposes deliberate stop and restart actions instead of asking you to hunt for a stale shell.
Logs are retained with a cap, so a service can run for hours without turning its history into an unbounded file. The result is a small but important change in the agent model: long-running infrastructure becomes visible session state rather than invisible background luck.
Browser checks and tool output stay readable
The built-in browser now exposes clearer console and network inspection and more reliable typed input. Split-screen actions route to the intended conversation more consistently, and the refreshed chat minimap makes long sessions easier to navigate without losing the current work.
Tool activity is also easier to scan in the conversation. Long output can move into a focused floating surface while the underlying session remains visible, and the Activity Panel is grounded in the process state rather than a decorative status. These details matter most during verification, when a green label is only useful if it corresponds to something the system actually checked.
Channel agents understand more of the room
A useful team bot needs more than the final message that mentioned it. Feishu and DingTalk agents in 1.1.22 can use recent group conversation. DingTalk can also inspect member details and attached files, while scheduled channel pushes can address real @mentions. That gives the agent enough context to continue an active discussion instead of making the team restate it in a private prompt.
The permission boundary does not disappear when the conversation moves into a channel. Onevium still keeps the session, tool calls, and resulting delivery visible on the desktop, so a request that began in a team room can be inspected where the work actually ran.
Fable 5.1 and the current Claude Agent SDK
Onevium 1.1.22 adds Claude Fable 5.1 and updates the Claude Agent SDK to 0.3.251. The release also improves macOS microphone permission recovery and reduces the prompt overhead used by Autonomy, leaving more of the context window for the task itself.
The update is available for Apple silicon and Intel Macs, Windows x64, and Linux x64 and arm64. Onevium's update prompt takes existing installations to the official release; on macOS, download and install the current build manually from that page. The release assets and checksums are listed with the v1.1.22 notes.