A Guth Labs publication ↗
Back to News

OpenCode makes background delegation the default and fixes slow local-model discovery

OpenCode makes background delegation the default and fixes slow local-model discovery

OpenCode 2.3.0 changes how coding work is delegated: child agents run asynchronously instead of holding up the parent session. The release notes also describe restart fixes and support for LM Studio decision models.

The underlying November 11 implementation removes the foreground/background option. Calls return a session identifier; results arrive when the child becomes idle. Importantly, interrupting the parent does not cancel its children. Individual child interruption remains available. That distinction matters for anyone interpreting a stop button as a fleet-wide stop.

Local inference gets a practical compatibility improvement too. Discovery now allows 15 seconds instead of one for Ollama, LM Studio and vLLM endpoints. The release says a slow endpoint no longer holds up other endpoints or configuration reloads. This is discovery plumbing, not faster token generation.

Guth analysis: Asynchronous delegation makes a parent more responsive, but changes the meaning of completion and cancellation. A useful upgrade test should cover a child that finishes, one waiting on background work, and one that must be stopped explicitly. Local-model users should separately test endpoint discovery and actual inference.

Evidence limits: These are maintainer statements and merged implementation records, not our benchmark or independent safety audit. No performance or reliability gain is quantified. Release-page display gives November 13 at 08:55 without a timezone in the fetched text.

Sources and citations

Revision history

Published .