OpenCode Desktop Shipped 5 Releases in 10 Days, Rewriting the v2 Input Box and Killing Ghost Tabs
OpenCode's pace signals that AI coding tool UX is still in flux, and the desktop client is catching up to the multi-session workflows developers already rely on in VS Code. The input-box rewrite and sub-agent depth cap directly address two failure modes—prompt collisions and runaway agent chains—that make these tools feel unreliable in daily use.
The OpenCode desktop client moved through five versions in ten days, from v1.18.0 to v1.18.4. The sprint finalized the v2 layout migration, added a toggle to switch back to the old UI, and introduced `Mod+N` for new tabs alongside automatic cleanup of stale session tabs. Sub-agent nesting is now off by default, reined in by a `subagent_depth` config.
v1.18.4 delivered the biggest single change: a ground-up rewrite of the v2 prompt input box. Commands, context, shell mode, attachments, and history no longer step on each other, and command drafts persist across sessions. The review panel also stopped jittering, stays aligned with open file tabs during diff changes, and loads full large-patch diffs without truncation.
The build/plan mode toggle disappeared from the default UI, which tripped up early updaters. It's still available as an opt-in setting, and OpenCode's own agent can surface the relevant GitHub issues and the toggle location when asked.
Five releases in ten days, including a full input-box rewrite, suggests the v2 desktop client was shipped early and is being hardened in public based on real usage friction.
Turning off sub-agent nesting by default is a tacit admission that unbounded agent chains create more problems than they solve for most workflows.
Preserving command drafts and cleaning up ghost tabs are small fixes that address the kind of state-loss and clutter that erode trust in AI coding tools over a session.