v1.3.0

v1.3.0

Added

• Model fallback page — New dedicated sidebar page "Model Fallback" between Providers and Settings. Select a configured model to automatically handle requests for unconfigured models (e.g. gpt5.4-mini), preventing Codex retry loops.

Changed

• LAN access toggle — Replaced listen address radio buttons with a single toggle switch and help-tip tooltip for a cleaner UI.

Fixed

• Provider edit/delete buttons unresponsive — Restored missing event bindings for edit and delete buttons on provider cards.

v1.2.0

v1.2.0

Added

• key/modelId routing — Provider-level routing key (auto-generated from name with ZH→EN mapping) enables key/modelId format routing (e.g. GLM/glm-5.1), eliminating model name conflicts across providers.
• Multi-model per provider — Each provider supports a models[] array with per-model maxOutputK and maxContextK settings, replacing the single model field.
• LAN access toggle — Settings page toggle switch to enable LAN access; Quick Start page shows hint about using machine IP for LAN devices.
• Quickstart dual-mode — Quick Start page shows two usage modes: cc-switch (API Key) and direct config.toml (ChatGPT account), with auto-generated config templates and clickable model selector.
• Markdown-aware tag parser — Thinking-tag parser tracks backtick code spans/blocks to avoid false positives on literal <thinking> text inside markdown.
• DeepSeek user_id passthrough — Provider-level userId field passes through as user_id (OpenAI Chat) or metadata.user_id (Anthropic) for account-level KVCache and safety isolation.

Changed

• Removed active provider concept — Routing is now purely request-based (key/modelId); no more "set as current" button or active badge.
• Provider data migration — loadProviders() auto-migrates old { model: "xxx" } format to { models: [{ id: "xxx" }] } and generates key for providers without one.
• Thinking parser conditional skip — OpenAI Chat adapter skips the thinking-tag parser when upstream sends reasoning_content field, preventing false-positive truncation.
• Anthropic adapter — max_tokens defaults to 65536 (from 8192) to prevent long response truncation.
• /v1/models endpoint — Returns models in key/modelId format.

Fixed

• Response truncation on OpenAI Chat — Thinking-tag parser no longer misidentifies literal <thinking> text in markdown content as thinking tags.
• CI pipeline — Fixed Prettier formatting inconsistencies and ESLint errors that caused all CI runs to fail.

v1.1.0 - Stability, Quality and Comprehensive Testing

Changes

Bug Fixes

• Fix duplicate anthropic-beta header in Anthropic non-streaming adapter
• Replace broken Atomics.wait hack in catalog.js with async execFile
• Fix port || 8629 to ?? 8629 so port 0 is handled correctly
• Fix Chinese character encoding corruption in catalog.js
• Add missing await on injectCodexConfig() in main process
• Remove conflicting ESLint quotes rule override

Code Quality

• Replace all console.warn calls with shared logger across adapters
• Change all var to const/let across all source files
• Move inline require() calls to top of files
• Remove unused getFeatures export from presets
• Lint: 0 errors, 0 warnings

Testing (402 tests)

• Add 289 comprehensive unit tests covering adapters, SSE bridge, events, cache, stream parser, presets, and edge cases
• Replace hardcoded ports with random port allocation in all integration tests
• Add npm run test:unit and npm run test:integration scripts

v1.0.0 — First Stable Release

Codex-Switch v1.0.0

将 Codex Desktop 桥接到任意 LLM 供应商。把 OpenAI Responses API 转换为 Anthropic/OpenAI Chat 协议，支持基于预设的供应商扩展。

---

亮点

• DeepSeek 思考透传 — 完整的多轮对话 thinking/reasoning 支持，适应 DeepSeek Anthropic 端点，并保留加密签名
• 模型别名系统 — 所有 Codex 原生模型标识（gpt-5.5、gpt-5.4、gpt-5.4-mini 等）自动别名到当前供应商模型，防止后台任务将原生模型名泄露到上游
• 混合 OAuth + 代理鉴权 — 双层架构：ChatGPT OAuth 处理 plugins/Mobile/quotas，模型请求则通过本地代理路由
• 预设注册表 — 供应商预设结合协议专用 Hook，无需修改核心适配器即可扩展

自 v0.1.1 以来的变更

修复

• DeepSeek 思考透传错误（content[].thinking must be passed back）在多轮工具调用对话中
  • 根因：function_call 类型的条目在 reasoning 到目标消息的搜索中被跳过，导致 thinking 块注入到错误的 assistant 消息
  • 新增签名缓存层（reasoning-cache.js），在 Codex 无状态协议中保留 DeepSeek 加密签名
  • 修复 SSE 桥接：新建 reasoning 项前先关闭之前打开的 message 项
  • Anthropic 适配器现在捕获 signature_delta 事件并在 block 完成时存入缓存
• Codex 启动卡死（无限加载）— 由 preferred_auth_method = "apikey" 与混合 OAuth 模式冲突所致，现已自动检测并禁用
• 原生模型名泄露 — Codex 后台任务（标题生成等）将原生模型名发送到上游；现已将所有已知标识别名化
• 原生 Catalog 读取增加带指数退避的重试机制，应对启动缓慢的场景

新增

• src/proxy/core/reasoning-cache.js — 内存级 reasoning 文本到签名缓存
• catalog.js 中基于原生 Catalog 的模型别名生成
• Thinking 注入和输入结构的详细诊断日志

安装

1. 下载下方的 Codex-Switch-v1.0.0-Windows_Setup.exe
2. 运行安装程序（已签名）
3. 启动 Codex-Switch，选择供应商预设，启动代理
4. Codex Desktop/CLI 会自动通过本地代理路由

校验

MD5: 9cac237aae5bc8f118c068363ce06419

环境要求

• Windows 10/11 x64
• 已安装 Codex Desktop 或 CLI
• 至少一个 LLM 供应商的 API Key（DeepSeek、OpenAI、Anthropic 等）