v0.1.1-alpha.9

中文

alpha.7 和 alpha.8 用户请注意： 在 Windows 上，旧版本的应用内更新会报 msiexec 1619 错误，因此无法升级到本版本。请手动下载并运行一次下面的 MSI 安装包。alpha.9 已修复这个问题，以后的版本都可以在应用内直接更新。

alpha.9 候选版。 CI 已在 Linux、Windows、macOS 上全部通过。Windows 实机验收还在进行，已通过的项目有：从 alpha.9 应用内更新、从 alpha.7 升级、WSL 实际中转、统计行、关闭路由、窗口尺寸。遇到问题请到 Issues 反馈。

新功能

• 概览支持所有工具：每个编程工具单独一行，显示从工具经路由或直连、到供应商和线路、再到模型的完整路径；可按工具查看今日用量。Claude Code 的会话记录也计入统计；没有价格的请求显示「未定价」，不显示 $0。
• 会话浏览器：
  • 搜索、筛选、多选会话；
  • 右键可以：
    • 复制 codex://threads/… 链接（远程主机的链接带 hostId）；
    • 在 ChatGPT 中打开；
    • 复制为 Markdown；
    • 导出为 Claude Code 或 pi 会话。
  • 会话迁移移到这个页面的右上角。
• 停用单条线路：在供应商页或命令行（spb profile edit --disable-line）中操作。停用后，路由和直连都不会再用这条线路。
• 混合登录：远程控制、Codex Cloud 等远程功能继续使用官方 ChatGPT 账号，模型请求则走本地路由；官方 token 不会发送给中转。开关在供应商页的「官方 ChatGPT 登录」中。
• 故障转移规则可设置：可以设置冷却时间、哪些状态码会切换线路、会话的空闲时长，以及首选线路恢复后是否切回。修改立即生效。
• 用量 → 路由：每个请求的每一次尝试、状态码、耗时都会列出，并标明协议转换和上游 fallback。
• Codex 设置 → 新版本新增：按版本列出最新 Codex（含 alpha）新增的配置项，并显示目录收录到了哪个版本。
• 工具检测：
  • 能找到 PATH 以外的工具，包括 Homebrew、npm/pnpm/bun 全局目录和登录 shell 的 PATH；
  • 可以「指定路径…」；
  • macOS 上默认用 Homebrew 安装；
  • 已安装的工具会出现在侧边栏，dsh 和 opencode 显示官方图标。
• 价格移到模型页：
  • 支持超长上下文分档、5 分钟/1 小时缓存写入、峰谷时段价；
  • 内置 Claude、Gemini、DeepSeek、Qwen、GLM、Kimi、Grok 等 68 个模型的官方价格，每条都标明来源页和日期。

改动

• 路由页改为平铺，不再折叠。后台服务信息移到「路由已开启」状态下方，「最近的路由事件」和 Claude Code 部分已去掉。
• 去掉 Cmd/Ctrl+K 命令面板和概览的 ⋯ 菜单。Superba 设置改为页内筛选，并去掉了「跳到」标签。
• 所有搜索框都带 × 清除按钮，按 Esc 也可以清空。
• 写入文件前的确认框会完整显示每一项改动（旧值 → 新值），密钥只显示末 4 位。

修复

• 测试通过的供应商模型会显示在 Codex 中。切换模型列表模式后，Codex 内核会自动重启并刷新列表；直连时页面会提示需要开启路由。
• Windows：
  • 应用内更新不再报 msiexec 1619；
  • 路径不再带 \\?\ 前缀；
  • Superba 打不开的进程不再阻碍退出，提示里会写出进程名。
• WSL：
  • 切换计划会写明保留哪个模型；
  • 冷启动时会多等一会儿再验证。
• macOS：不再把 Cursor 等其他 Electron 应用误认为 ChatGPT。
• 用 spb router disable 关闭路由后，界面会立即更新。
• 当前版本比最新发布版还新时，更新页不再显示指向旧版本的箭头。
• 读屏和辅助功能方面的若干修复。

已知问题

• 安装包没有签名，和之前的 alpha 版本一样。macOS 第一次打开时请右键选择「打开」；Windows 弹出 SmartScreen 时请点「仍要运行」。
• 「官方账号作为路由成员」在路由层已经实现，但还没有界面入口。混合登录还需要用真实账号验收。
• DeepSeek 的峰谷价没有处理中国法定节假日。部分 Gemini 价格是限时价，没有记录到期日。
• 导出到 Claude Code 或 pi 的会话还没有用真实客户端验证过能否恢复；dsh 导出暂不支持。
• macOS 只经过 CI 测试和 DMG 检查，Homebrew 安装只在模拟环境中验证过。

English

alpha.7 and alpha.8 users: on Windows, the in-app update in those versions fails with msiexec error 1619, so they cannot update to this version. Download and run the MSI below once. alpha.9 fixes this, so later versions will update in-app.

alpha.9 release candidate. CI passes on Linux, Windows and macOS. Acceptance testing on real Windows machines is still in progress. These checks have passed: the in-app update from alpha.9, the upgrade from alpha.7, WSL over a real relay, the stats line, turning the router off, and window sizes. Please report problems in Issues.

New

• Overview for every tool: each coding tool gets its own row showing its full path: through the router or direct, then the provider and line, then the model. Today's usage can be broken down by tool. Claude Code sessions now count too; requests without a price show "unpriced", never $0.
• Session browser:
  • search, filter and select several sessions;
  • right-click to:
    • copy a codex://threads/… link (links to remote hosts include the hostId);
    • open the session in ChatGPT;
    • copy it as Markdown;
    • export it as a Claude Code or pi session.
  • Session migration moves to the top right of this page.
• Disable a single line, on the Providers page or from the CLI (spb profile edit --disable-line). Once disabled, neither the router nor direct mode uses it.
• Mixed login: remote features such as remote control and Codex Cloud keep using your official ChatGPT account, while model requests go through the local router. The official token is never sent to the relay. The switch is under Providers → Official ChatGPT login.
• Configurable failover rules: set the cooldowns, which status codes switch to another line, how long a conversation can sit idle, and whether to return to the preferred line once it recovers. Changes apply immediately.
• Usage → Routing: lists every attempt for each request, with its status code and time taken, and marks protocol translation and upstream fallback.
• Codex Settings → New in latest: config keys added by recent Codex releases (alphas included), grouped by version, with the newest Codex version the catalog covers.
• Tool detection:
  • finds tools outside PATH: Homebrew, the npm/pnpm/bun global directories, and the login shell's PATH;
  • "Choose path…" lets you point at a tool yourself;
  • on macOS, Homebrew is the default installer;
  • installed tools appear in the sidebar, and dsh and opencode show their official icons.
• Prices moved to the Models page:
  • supports long-context tiers, 5-minute/1-hour cache writes and peak/off-peak windows;
  • built-in official prices for 68 models from Claude, Gemini, DeepSeek, Qwen, GLM, Kimi, Grok and others, each with its source page and date.

Changes

• The Router page is now flat, with nothing collapsed. The background service info sits under the "router on" status line. "Recent router events" and the Claude Code section are gone.
• The Cmd/Ctrl+K palette and the Overview ⋯ menu are removed. Superba Settings has an inline filter instead, and the "Jump to" label is gone.
• Every search field has a × clear button, and Esc clears it too.
• The confirmation before writing a file shows every change in full (old → new); secrets show only their last 4 characters.

Fixes

• Provider models that passed their test now show up in Codex. Switching the model list mode restarts the Codex core so the list refreshes; in direct mode, the page says the router needs to be on.
• Windows:
  • the in-app update no longer fails with msiexec 1619;
  • paths no longer carry the \\?\ prefix;
  • processes Superba cannot open no longer block quitting, and the message names them.
• WSL:
  • the switch plan says which model is kept;
  • verification waits a little longer when the distro starts cold.
• macOS: Cursor and other Electron apps are no longer mistaken for ChatGPT.
• The GUI updates immediately after spb router disable.
• When the running version is newer than the latest release, the Updates page no longer shows an arrow pointing to the older version.
• Several screen reader and accessibility fixes.

Known issues

• The installers are unsigned, as in earlier alphas. On macOS, right-click and choose Open the first time; on Windows, click "Run anyway" when SmartScreen appears.
• The official account as a router member works in the router, but there is no UI for it yet. Mixed login still needs acceptance testing with a real account.
• DeepSeek's peak/off-peak prices ignore Chinese public holidays. Some Gemini prices are promotional, and their end dates are not recorded.
• Sessions exported to Claude Code or pi have not yet been resumed in the real clients; dsh export is not supported yet.
• macOS was tested by CI and DMG checks only; Homebrew installs were verified against a fake brew only.

v0.1.1-alpha.8

中文

alpha.8 候选版。 CI 已在 Linux、Windows、macOS 上全部通过。Windows 实机验收还在进行，遇到问题请到 Issues 反馈。

界面

• 模型页：新增「路由模型列表」，有三种模式。
  • 仅 Codex 模型（默认）；
  • Codex 模型 + 供应商的其他模型；
  • 再给每个模型加上供应商前缀，例如「GPT-5.5 · Krill」。
  • 每个模型只占一行，旁边的「供应商」列列出哪几家提供它。
  • 已知的 GPT 模型沿用 Codex 目录里的上下文长度和速度档。
  • 修复：「测试」按钮被截断；前缀显示成 provider/。
• 路由页：开关和状态合并到一行；成员的线路改成表格；不常用的设置收进「高级」。
• 工具页：改为单栏，分为「连接 / 模型 / 选项 / 配置文件」四节。
• Claude Code 连接与模型：分为「连接 / 模型槽位 / /model 列表 / 高级」四节，每个槽位可以直接选模型。
• 供应商页：直连时可以直接「切换到此线路」；开着路由时改为「设为首选」，立即生效。
• 对话框：高度不超出窗口，内容区可以滚动，按钮固定在底部；同类提示合并成一条。

功能与修复

• 开着路由时，ChatGPT app 的旁路请求（插件、用量、统计、个人资料等）由本地路由直接回答，不再发到上游，避免 429。
• WSL 里的 Codex 也走 PAT 和本地 TLS，WSL 中也能看到小闪电。
• 修复：配置了 WSL 远程主机并开着路由时，app 停在登录页。
• 修复：Windows 上路径含空格时，通知钩子执行失败。
• 修复：强制退出 ChatGPT 沙盒进程时报 Access denied。
• 修复：本机其他服务（比如 Ollama）被误判为「路由已开启」。
• 修复：关闭路由后，模型被降级成 gpt-5。
• 只能经过协议转换才能使用的模型，现在也会出现在模型目录里；需要时由路由代做压缩。

已知问题

• 安装包没有签名，和之前的 alpha 版本一样。macOS 第一次打开时请右键选择「打开」；Windows 弹出 SmartScreen 时点「仍要运行」。
• macOS 只经过 CI 测试和 DMG 检查，没有做真机全矩阵测试。

English

alpha.8 release candidate. CI passes on Linux, Windows and macOS. Acceptance testing on real Windows machines is still in progress; please report problems in Issues.

Interface

• Models: a new router model list with three modes.
  • Codex models only (the default);
  • Codex models plus the provider's other models;
  • all of the above, plus a provider prefix on each model, e.g. "GPT-5.5 · Krill".
  • Each model takes one row, with a Providers column showing which providers offer it.
  • Known GPT models keep the context length and speed tiers from the Codex catalog.
  • Fixed: the Test button was clipped; prefixes showed as provider/.
• Router: the switch and status share one row, member lines are shown as a table, and rarely used settings are under Advanced.
• Tools: a single column with four sections: Connection, Model, Options and Config file.
• Claude Code connection and models: four sections (Connection, Model slots, /model list, Advanced), and you can pick a model directly in each slot.
• Providers: "Use this line" switches the line directly in direct mode; with the router on, the button is "Set as preferred" and takes effect immediately.
• Dialogs: they never grow taller than the window, the content scrolls, and the buttons stay pinned at the bottom. Repeated notices are grouped into one.

Features and fixes

• With the router on, the ChatGPT app's side calls (plugins, usage, analytics, profile and so on) are answered by the local router and never reach the upstream, which avoids 429s.
• Codex inside WSL now also uses PAT and local TLS, so the lightning icon shows in WSL too.
• Fixed: the app stuck on the login page when a WSL remote host was configured and the router was on.
• Fixed: Windows notification hooks failing when the path contains spaces.
• Fixed: "Access denied" when force-quitting ChatGPT sandbox processes.
• Fixed: other local services (such as Ollama) wrongly detected as "router on".
• Fixed: the model downgrading to gpt-5 after turning the router off.
• Models that are only usable through protocol translation now appear in the model catalog; when needed, the router handles compaction for them.

Known issues

• The installers are unsigned, as in earlier alphas. On macOS, right-click and choose Open the first time; on Windows, click "Run anyway" when SmartScreen appears.
• macOS was tested by CI and DMG checks only, not by a full matrix on real machines.

v0.1.1-alpha.7

v0.1.1-alpha.7

预览版 / Prerelease

中文

本次更新

alpha.7 将 Claude Code 和更多编程工具纳入 Superba，并加入通知、协议转换和更集中的维护入口。

• Claude Code 与工具配置。 Claude Code 有独立的连接、模型、设置和诊断页面；工具页支持 dsh、opencode、pi、ZCode、WorkBuddy / WorkBuddy AI，并加入 Gemini CLI 的路由配置。应用前可查看变更，配置有备份和还原；已应用的工具保持绑定的供应商，切换 Codex 供应商后需要明确重新应用。
• 统一安装和更新。 在「Superba → 更新」查看 Superba、桌面应用、Codex 和各编程工具的版本与可用操作。工具操作提供进度、取消和日志。新装 opencode 默认使用 OpenCode 2，已有安装保留原主版本。
• 通知到手机和聊天工具。 在「Superba → 设置 → 通知」配置任务完成、需要批准、任务失败，以及 Superba 切换、路由和更新事件。渠道包括 Telegram、飞书 / Lark、企业微信、钉钉、ServerChan、PushPlus、Bark、ntfy 和带签名的 Webhook；支持免打扰、限流、去重和测试通知。任务通知需要安装对应工具的钩子，回复摘要默认不发送。对应 #12。
• 协议转换，包含 Gemini。 本地路由可在 Anthropic Messages、OpenAI Responses、Chat Completions 和 Gemini 之间转换，并提供 /gemini 入口。优先使用原生端点，转换可按方向关闭；工具页会说明是否需要转换。/openai/v1 同时保留 Responses 和 Chat Completions 原样转发。
• 模型目录与供应商前缀。 合并内置目录、供应商 Codex 目录、标准模型列表和用户覆盖，展示来源、错误和能力探测状态。路由模式下可用 <供应商>/<模型> 固定供应商；裸模型名保留故障切换顺序。兼容 krill/、krill-usage/、krill-balance/ 前缀，对应 #16。
• PAT 按模型设置与压缩。 基于密钥的 PAT 配置也可通过模型目录应用单个模型的上下文窗口和自动压缩阈值，对应 #14。对通过能力检查的标准列表模型，路由可代做上下文压缩并保留加密摘要；原生与合成目标在故障切换时分别使用自己的压缩方式。
• 桌面应用的小闪电。 路由 PAT 保留上游模型目录的速度档信息，使支持该能力的模型可显示并使用 Fast 控件。是否可用仍由上游能力决定。
• 数据位置与磁盘占用。「Superba → 数据」可以移动 Superba 数据、查看占用、设置请求记录保留期和清理缓存；移动过程有进度、校验和失败回滚。命令行提供 spb data where|move|usage。对应 #15，迁移边界见下方。
• 登录后自动启动与按工具用量。 可设置当前用户登录后启动和仅托盘启动；没有可用托盘时打开窗口。用量与请求记录可按工具筛选，路由页展示各入口，并提供按工具设置的统计行。

界面与修复

底部「Superba」页改为「概况 · 更新 · 设置 · 诊断 · 数据」标签；设置集中在一张可滚动页面，所有安装和更新都在「更新」。主侧栏更紧凑，移除了重复标题，更新清单使用完整内容宽度；概览菜单统一叫「操作」。

同时修复恢复原始配置时遗留路由设置的问题（默认恢复原路由并反向迁移会话供应商 ID）、首选线路修改后的路由热加载、路由状态误判、WSL 冷启动等待及失败信息保留、会话锁持有者诊断，以及 macOS 登录启动、隐藏后重开、关闭设置向导和中文字体显示。另修复迁移/回滚后的会话自动刷新、卸载后的认证端点还原，以及窄窗下 Claude 设置和网络控件的可读性。

使用范围与限制

• 数据迁移只移动 Superba 自身的数据。 不自动迁移 Codex 的原始会话、会话数据库或 Codex 主目录，也不改变以后 Codex 会话的保存位置。界面单独显示 Codex 占用；如果系统盘空间主要由 Codex 会话占用，此功能不能直接解决这一部分。
• 手机回复继续工作仅完成设计。 alpha.7 不提供双向聊天、手机批准后继续任务或原生个人微信机器人。个人微信通知通过 ServerChan / PushPlus 第三方中转。
• 协议转换并不等于模型目录全部放开。 合并 Codex 目录中新合成的标准列表模型需要工具调用探测通过；标为「需要协议转换」的条目仍不能通过这条目录路径使用。探测结果保存在内存中，路由重启后可能需要重新测试。
• 合并目录和加密摘要路由使用 HTTP 回退。原生供应商的加密历史不能由合成模型解密；需返回原供应商，或明确新建会话。PAT 按模型配置也不等于官方 ChatGPT 账号与 API 在同一配置中并发共存，Fast / Ultrafast 不保证适用于所有模型。
• ZCode、WorkBuddy 的安装器管理限 Windows；macOS / Linux 提供下载链接。WorkBuddy 保留自身更新器。Gemini CLI 本轮仅接入配置，不接管安装更新；工具的 WSL 内安装和 Windows ARM64 安装包不在本轮范围。

安装包

提供 Windows x86_64 MSI、macOS universal DMG / ZIP、Linux x86_64 / aarch64 压缩包、Linux x86_64 musl 命令行包，以及 SHA256SUMS.txt。macOS 和 Windows 安装包沿用此前 alpha 版本的未签名发布方式，首次打开方法见「已知问题」。

English

What's new

alpha.7 brings Claude Code and more coding tools into Superba, with notifications, protocol conversion and a consolidated maintenance page.

• Claude Code and tool configuration. Claude Code has dedicated connection, model, settings and diagnostics pages. The Tools page configures dsh, opencode, pi, ZCode and WorkBuddy / WorkBuddy AI, plus Gemini CLI through the router. Review changes before applying, with configuration backups and restore. An applied tool keeps its chosen provider until explicitly reapplied, even after switching Codex's provider.
• One place for installs and updates. Superba → Updates lists Superba, the desktop app, Codex and the coding tools, with versions and available actions. Tool operations show progress, cancellation and logs. New opencode installs use OpenCode 2; existing installations retain their major version.
• Notifications to your phone or chat. Superba → Settings → Notifications covers task completion, approval requests and failures, as well as Superba switch, router and update events. Channels include Telegram, Feishu / Lark, WeCom, DingTalk, ServerChan, PushPlus, Bark, ntfy and signed webhooks, with quiet hours, rate limits, deduplication and a test action. Task notifications require the corresponding tool hooks; reply summaries are opt-in. See #12.
• Protocol conversion, including Gemini. The local router converts between Anthropic Messages, OpenAI Responses, Chat Completions and Gemini, with a /gemini ingress. Native endpoints take precedence, conversions can be disabled by direction, and the tool pages show when conversion is needed. /openai/v1 also retains native Responses and Chat Completions forwarding.
• Model catalogs and provider prefixes. Merge built-in models, provider Codex catalogs, standard model lists and user overrides, with source errors and capability checks visible. In router mode, <provider>/<model> pins a provider; bare model names retain failover order. The legacy krill/, krill-usage/ and krill-balance/ prefixes are accepted. See #16.
• PAT per-model settings and compaction. Key-based PAT profiles can also apply per-model context windows and auto-compaction thresholds through a model catalog, addressing that part of #14. For eligible standard-list models, the router can compact history and retain an encrypted summary. Native and synthetic targets keep their own compaction behavior during failover.
• The desktop app's lightning control. Routed PAT connections preserve upstream speed-tier metadata so models that support it can expose and use Fast mode. Availability still depends on the upstream.
• Data location and disk usage. Superba → Data can relocate Superba's data, show storage use, configure request-record retention and clean caches. Moves report progress, verify copied files and roll back on failure. CLI: spb data where|move|usage. See #15 and the migration boundary below.
• Launch at login and usage by tool. Enable startup for the current user, optionally in the tray; if no usable tray is available, the window opens. Filter usage and requests by tool, inspect router ingress cards and configure per-tool statistics lines.

Interface and fixes

The bottom Superba page now has Overview, Updates, Settings, Diagnostics and Data tabs. Settings is a single scrolling page with section links, and all installs and updates live in Updates. The sidebar is more compact, duplicate headings are removed, and the update list uses the available width. The main Overview menu is named Actions.

Fixes cover Restore Original leaving routing settings behind (the default restores original routing and reverses session provider-ID migration), preferred-line changes reloading the router, incorrect routing status, bounded WSL cold starts and retained failure results, session-lock holder diagnostics, and macOS login startup, reopening hidden windows, setup dismissal and CJK fonts. Further fixes refresh sessions after migration and rollback, restore the owned authentication endpoint on uninstall, and keep Claude settings and network controls readable in narrow windows.

Scope and limitations

• Relocation moves Superba data only. It does not move Codex's raw conversations, session database or home directory, or change where future Codex conversations are stored. Codex disk usage is shown separately. This does not directly free space occupied by Codex conversations on the system drive.
• Continuing work by replying from a phone is design-only. alpha.7 does not include two-way chat, remote approval and continuation, or a native personal-WeChat bot. Personal WeChat notifications use the third-party ServerChan / PushPlus relays.
• Protocol conversion does not unlock every catalog model. Newly synthesized standard-list models in the merged Codex catalog require a successful tool probe. Entries marked Needs protocol conversion remain unavailable through that catalog path. Probe results are held in memory and may need to be repeated after a router restart.
• Merged-catalog and encrypted-summary routing use HTTP fallback. Native provider-encrypted history cannot be decoded for a synthetic model; return to the original provider or explicitly start a new conversation. PAT per-model settings do not provide concurrent official ChatGPT and API authentication in one profile, and Fast / Ultrafast is not guaranteed for every model.
• ZCode and WorkBuddy installer management is Windows-only; macOS / Linux use download links. WorkBuddy retains its own updater. Gemini CLI configuration is included, but its installation and updates are not managed here. Tool installs inside WSL and Windows ARM64 packages are outside this release's scope.

Packages

Downloads include a Windows x86_64 MSI, macOS universal DMG / ZIP, Linux x86_64 / aarch64 archives, a Linux x86_64 musl CLI archive, and SHA256SUMS.txt. The macOS and Windows packages are unsigned, as in previous alpha releases; see Known issues for first-launch instructions.

完整变更 / Full changelog: alpha.6 → alpha.7

已知问题 / Known issues

中文

• Windows（W90）： 配置了 WSL 远程主机、又开着本地路由时，ChatGPT 桌面 app 可能停在登录页。修复会在 alpha.8。临时办法：关闭本地路由，或者先移除 WSL 远程主机再切换。

• Windows（W95）： 开启通知时，如果用户目录或路径里有空格，PowerShell 通知钩子可能执行失败。修复会在 alpha.8。

• macOS： 这一版的 macOS 只经过 CI 测试和 DMG/Finder 检查，没有做真机全矩阵测试。

• macOS / Windows 安装包： 这一版的 macOS 和 Windows 安装包都没有签名或公证，和之前的 alpha 版本一样。macOS 首次打开要右键选「打开」，或者执行 xattr -dr com.apple.quarantine /Applications/Superba.app；Windows 的 SmartScreen 要点「仍要运行」。

English

• Windows (W90): With a WSL remote host configured and local routing enabled, the ChatGPT desktop app may remain on the login screen. A fix is planned for alpha.8. Workaround: turn off local routing, or remove the WSL remote host before switching.
• Windows (W95): When notifications are enabled, the PowerShell notification hook may fail if the user directory or another path contains spaces. A fix is planned for alpha.8....

v0.1.1-alpha.6

Full Changelog: v0.1.0...v0.1.1-alpha.6

v0.1.1-alpha.5

Full Changelog: v0.1.0...v0.1.1-alpha.5

v0.1.1-alpha.4

Full Changelog: v0.1.0...v0.1.1-alpha.4

v0.1.1-alpha.3

Full Changelog: v0.1.0...v0.1.1-alpha.3

v0.1.1-alpha.2

Full Changelog: v0.1.0...v0.1.1-alpha.2

v0.1.1-alpha.1

Full Changelog: v0.1.0...v0.1.1-alpha.1