OpenCode v2.2.7 lets GitHub Copilot pick the model for each session

Pick Auto in the model picker, and GitHub Copilot's router chooses the model for each session at the Auto discount. Your account needs an Auto pool, and if routing fails, you get your default chat model.

Remote MCP servers that rotate refresh tokens no longer sign you out, even when the TUI, the desktop app and other projects refresh the same login at once, or the token endpoint fails for a moment.

The read tool takes a negative offset now, so minus fifty gives the agent the last fifty lines of a log. And when a file isn't there, read suggests the names it probably meant.

Every shell result now ends with its exit code. A command killed by a signal names the signal, and a timeout says how long it waited, so the agent can try a longer one.

In the desktop app, right-click a side panel tab to close every tab to its right, Command Shift T reopens the last one you closed, and Mark as unread keeps a session's dot until you come back.

Background compaction no longer stops a Claude Agent session in the middle of a tool call. Finished MCP actions aren't reported as failed. And the TUI follows your terminal's light or dark mode.

If you use OpenCode through Copilot, I'd upgrade today and switch to Auto: the router does the model picking, and you pay the Auto rate. If you build on the SDK, read the changelog first.