v0.11.0

Added

• GPT-6.1 Sol (gpt-6.1-sol). The Codex backend only serves it to client version 0.159.0 and later, so the plugin now presents 0.159.0; before this release every request to it answered "not supported when using Codex with a ChatGPT account". Every model the plugin already supported was checked at the new version. Like GPT-6 Sol and Luna, it is held under 272K input tokens (a 372K window), because requests above that are charged 2x for input and 1.5x for output. Changing reasoning effort mid-session keeps the cached prefix, as on the other GPT-6 models. OpenCode lists the model once models.dev does.

Fixed

• Turns cut by an OpenAI service restart no longer end when nothing was lost. The Codex backend restarts regularly and closes open WebSocket streams. When only an unfinished tool call had streamed at that point, the turn is now retried. When a tool call had already finished, the response ends as completed, so that tool runs once and the session carries on. A cut after text or reasoning had been shown still ends the turn, because retrying would show it twice.
• A mid-stream error from the backend can no longer replay a turn. When Codex sent an error after output had started, the plugin passed the provider's wording on (for example "Rate limit exceeded"), and OpenCode could read that as retryable and repeat the turn: text shown twice, or a tool run twice. Errors after output now follow the same rules as a dropped connection.
• A rate limit that arrives while a tool call is still being written now moves the retry to another account instead of retrying the exhausted one.