OpenCode v2.3.0 runs subagents in the background while you keep working

OpenCode subagents now run in the background by default. A subagent hands control back right away, so you keep working in the parent session while several of them run at once.

Slide text

01 02 03 04 Background subagents

anomalyco/opencode · CHANGELOG.md · PR #63382

OpenCode keeps you working while subagents run

Background by default Several at once

A subagent returns its session ID the moment it starts. Then, whenever it goes idle, it sends its results back to the parent session, without the parent blocking on it.

Sometimes a subagent goes idle while its own background work keeps going. Its report now says it is still running, so the parent waits for the final answer instead of treating a partial one as done.

Part two: quieter transcripts in the TUI.

The TUI's transcript is quieter by default. Each run of tool calls and thoughts folds into one collapsible activity row, so the answer itself is easier to follow.

Verbosity now defaults to low. To see every step again, set Verbosity to medium or high in settings. If you already chose a verbosity, OpenCode keeps yours.

Three smaller changes in the TUI. A session ID in a message now opens directly. Activity tabs are clickable and open the right view. And shell mode gets its own prompt color.

Part three: the desktop app.

In the desktop app, Command Alt G, or Control Alt G on Windows and Linux, searches the session's prompts and jumps straight to one, even prompts the timeline hasn't loaded.

The desktop app has an experimental Recent sidebar. It orders your sessions across servers by recency, with local pins and reordering, and shows the last seven days, up to thirty rows.

Alt Enter puts a follow-up first in the queue, and Command or Control Enter steers the next queued prompt. Mermaid diagrams can now be maximized, for zooming and panning.

Part four: shell commands and models.

Shell commands now survive server handoffs, and an interrupted command keeps its output. Behind an experimental flag, a command whose timeout expires can move to the background instead.

Local model discovery in Ollama, LM Studio and vLLM now waits up to fifteen seconds per request instead of one. Remote or tunnelled endpoints list their models, and a slow one no longer holds up the rest.

Behind a gateway that requires mutual TLS? A provider can now present a client certificate. The certificate, key and passphrase load from a file or an environment variable, like an API key.

A spent rate limit no longer counts as a fatal error. A retryable provider stream that gets cut off keeps the work it finished. And GitHub Copilot requests use the right endpoint while models are still loading.

If you hand work to subagents, I'd upgrade today: you keep typing while they run, and each one reports back on its own. The rest of this release is in OpenCode's changelog on GitHub.