Skip to content

0.160.1-ma.1 (multi-account fork)

Choose a tag to compare

@github-actions github-actions released this 06 Oct 05:53
· 8 commits to feature/native-multi-account since this release

Multi-account improvements

  • Implemented automatic model and thinking-effort selection for new main tasks and eligible subagents. Off by default; choose Preview or Automatic, separate scopes, a 0–100 preference, exact model limits and relative roles.
  • Added local rules and an independently consented Jev/Clef task classifier. One configured service is used; tool ranking can stay off. Classification sends only a bounded task description, retains manual choices, respects catalog/context/media/tier constraints, and never enables Ultra, paid inference fallback or reset-credit spending.
  • Added direct browser controls, terminal manager M, codex account routing, and the native phone Model selection menu, in English and Chinese. /model auto and Use auto for this thread release a conversation's manual pin. Manual pins survive resume; retries, steering and full-history children keep their model.
  • Fixed interrupted reset recovery: explicit review retains the original owner/key/credit and a backup, without claiming success. Phone, terminal and browser managers expose review. A successful quota recovery reconciles stale automatic reset history.
  • Fixed browser recovery for multiple original reset operations, including changed or removed owners. Reviewed operations have an explicit browser-only clear action; no credit is spent by review/clear.
  • Bound quota-refresh receipts and reset-credit counts to the exact user/workspace identity so relogin cannot display another owner's cached result.
  • Preserved exact model allowlists during unrelated browser edits, added local simulation and protected unsaved policy drafts and stale confirmations.

Upstream

Includes official Codex 0.160.1: preserve Windows SYSTEMROOT, TEMP and TMP for remote stdio MCP startup with explicit remote environment variables.

使用说明

本版完整实现主任务/子代理自动选模型及思考强度,默认关闭;可以选择仅预览或自动应用,分别开关主任务与子代理,并设置 0–100 的续航/能力倾向。浏览器账号管理器、终端管理器 M 页、codex account routing 和手机 /account manage 的“模型选择”菜单均可设置,支持中英界面。

本地规则不请求外部服务。Jev 与 Clef 二选一,需要单独同意分享简短任务说明;不会自动打开付费 API、使用重置券或开启 Ultra。手动模型选择优先,/model auto 或手机“此会话恢复自动选择”可明确解除本会话固定。会话恢复、重试和任务中的补充消息保持正确选择来源。

同时修复中断兑券记录长期阻塞、多个原操作恢复、原账号移除后的复核,以及重新登录后刷新回执/券数串账号的问题。复核与清理保留未知结果及原始记录,不会兑券,也不冒充额度已恢复。

已跟进官方 0.160.1 的 Windows 远程 MCP 启动修复。

See the bilingual model-selection guide and account-manager guide. Real iOS rendering/reconnect and subscription savings have not been benchmarked; fixture and package checks do not establish a savings percentage.