v0.9.0

GPT-6 Sol and GPT-6 Luna

Both models now work. The Codex backend only offers them to client version 0.155.0 and above, and earlier versions of this plugin identified as 0.153.0, so both were missing from the picker.

Their context is capped at 372k (244k input), the same cost guard used for the GPT-5.6 family. The backend reports the same 872k window it does for GPT-6 Astra, but OpenAI's rate card exempts only Astra from the Codex long-context surcharge (2× input, 1.5× output above 272k input tokens), so running Sol or Luna past 272k would cost double without warning.

Changing reasoning effort mid-session now works on both, as it already did for Astra, without disturbing the prompt cache.

Multi-project sessions

When one OpenCode process hosted several project directories, opening the second project stopped the first project's command dialogs for the rest of that process. Each project now keeps its own. A related leak is fixed too: a background refresher could be left running each time the auth loader ran again, competing for the account-store lock until exit.

Quota

The sidebar now shows the account's credit budget next to its 5-hour and weekly windows. An account can have headroom on both windows and still be out of credits; that used to look like an unexplained failure. Routing now treats a spent credit budget like an exhausted window: it avoids placing new sessions there and moves a pinned session off it.

Safer failures

A request is now refused locally if it would go out with no access token. Previously this produced a provider 401, which looked like an expired or revoked account and sent debugging to the wrong place.

Vault custody (opt-in)

Fallback accounts, and the main account, can have their credentials held by a Claustrum vault instead of this plugin's store, so exactly one process owns each refresh token. It is off by default and has no effect unless you turn it on with /openai-account claustrum. API-key accounts stay local, because custody is about refresh tokens.

Retired models

GPT-5.4, GPT-5.4-mini and GPT-5.3-Codex-Spark now return "not supported when using Codex with a ChatGPT account" from the Codex backend. This release doesn't cause that, and they still appear in the picker for now.