v2.4.4

Codex App Transfer v2.4.4

中文

接入 QoderWork CN 供应商 + QoderWork / WorkBuddy 模型选择器显示 Credit 折扣倍率

本版本接入阿里 QoderWork CN 供应商(账号登录 + 模型路由,无需另装 Qoder),并在 Codex 模型选择器里为 QoderWork 与 WorkBuddy 的模型显示各自的 Credit 消耗倍率(如 GLM-5.2 · 0.6×)。

• QoderWork CN 接入(#581):纯客户端 PKCE 账号登录 + 自动 refresh;模型请求走网关 gateway.qoder.com.cn 的 Cosy 签名协议(内嵌官方 qoder_auth_wasm);多账号池与 WorkBuddy 复用同一 account_pool。
• QoderWork 细节优化(#582):per-account 设备隔离、模型显示名、context window、reasoning 档位、额度面板与官方图标;修复 auto 等通用别名撞名导致 reasoning / context 误套的回归。
• QoderWork 模型倍率显示(#583):模型选择器显示 GLM-5.2 · 0.6× 这类倍率,优先取官方文档 docs.qoder.com/zh/cli/model.md;Auto 无固定倍率不显示。
• WorkBuddy 模型倍率显示(#584):同口径显示 GLM-5.2 · 0.79× 等,倍率取 WorkBuddy 客户端下发的 product 配置;与 QoderWork 共用同一渲染逻辑。
• 倍率漂移检测(#585):新增每周定时 provider-rate-drift workflow,拉官方文档对比仓库倍率表,发现漂移或检测器失效即开 / 更新 issue。

验证:registry / codex_integration / src-tauri 相关单测全绿,cargo fmt --check 干净,rate-drift 脚本本地实测无漂移退 0、篡改倍率退 1。

English

QoderWork CN provider integration + Credit discount multipliers in the QoderWork / WorkBuddy model picker

This release integrates the Alibaba QoderWork CN provider (account login + model routing, no separate Qoder install), and shows each QoderWork / WorkBuddy model's Credit consumption multiplier in the Codex model picker (e.g. GLM-5.2 · 0.6×).

• QoderWork CN integration (#581): pure-client PKCE login + auto refresh; model requests use the gateway.qoder.com.cn Cosy-signature protocol (embedded official qoder_auth_wasm); the multi-account pool reuses the same account_pool as WorkBuddy.
• QoderWork refinements (#582): per-account device isolation, model display names, context window, reasoning tiers, quota panel and official icon; fixes a regression where generic aliases like auto collided with other providers and mis-applied reasoning / context.
• QoderWork multipliers (#583): the model picker shows multipliers like GLM-5.2 · 0.6×, sourced primarily from the official docs.qoder.com/zh/cli/model.md; Auto has no fixed multiplier and shows none.
• WorkBuddy multipliers (#584): the same styling shows GLM-5.2 · 0.79× and so on, with multipliers from WorkBuddy's client-delivered product config; QoderWork and WorkBuddy share one rendering path.
• Rate-drift detection (#585): a new weekly provider-rate-drift workflow pulls the official docs and diffs them against the in-repo multiplier tables, opening / updating an issue on drift or detector failure.

Verification: registry / codex_integration / src-tauri unit tests pass, cargo fmt --check is clean, and the rate-drift script exits 0 on no drift and 1 on a tampered rate locally.

v2.4.3

Codex App Transfer v2.4.3

中文

第三方 prompt cache 命中塌方修复 + Codex 内逐轮命中率趋势折线

修复第三方上游(GLM / WorkBuddy 等 OpenAI Chat 兼容网关)长会话 prompt cache 命中率从约 95% 塌到约 33% 的问题:每轮把大量本可命中的内容按未缓存重算,既慢又多耗额度。同时给 Codex 顶栏用量面板加了逐轮命中率趋势可视化。

• 根因与修复:tool 输出折叠每轮给同一条历史工具输出重新生成随机 artifact_id,折叠进历史的消息跨轮字节不稳定,打断上游按 messages 前缀做的缓存。改为按内容确定性生成(call_id + kind + 原文 SHA-256)+ 落盘 UPSERT 幂等:同一输出跨轮 / 跨进程恒得同一 id、同内容不再每轮插新行(顺带治 tool_artifacts.db 膨胀),折叠消息跨轮字节稳定,命中率恢复到接近原生。原生 ChatGPT 走 Responses 透传不受影响。
• 命中率趋势:用量面板底部 缓存命中 X% 可点,正下方弹固定宽度浮层显示当前对话逐轮命中率折线(左轴刻度 + 折线);轮数变化只改折线疏密、宽度恒定,点外部 / Esc / 滚动即关。
• 稳健性:关面板 / 升级重装注入脚本时一并回收浮层监听与残留节点,不在 Codex 页面泄漏。

验证:adapters / usage_tracker / 注入器单测全绿,fmt / clippy 干净,3 个审查 agent 复核。

English

Third-party prompt cache hit-rate collapse fix + per-turn hit-rate trend line in Codex

Fixes a bug where the prompt cache hit rate on third-party upstreams (GLM / WorkBuddy and other OpenAI-Chat-compatible gateways) collapsed from ~95% to ~33% in long conversations, re-billing large amounts of cacheable content as uncached (slower and costlier). Also adds a per-turn hit-rate trend to the Codex usage panel.

• Root cause and fix: tool-output folding regenerated a random artifact_id for the same historical output every turn, making the folded history message byte-unstable and breaking the upstream prefix-based cache. It is now derived deterministically from content (call_id + kind + SHA-256 of the raw output) with an idempotent UPSERT on persistence: the same output always maps to the same id across turns/processes and identical content no longer inserts a new row each turn (also curbing tool_artifacts.db growth). Folded messages are byte-stable across turns and hit rates recover to near-native. Native ChatGPT (Responses passthrough) is unaffected.
• Hit-rate trend: the Cache hit X% entry at the bottom of the usage panel is now clickable and opens a fixed-width popover below it with the conversation's per-turn hit-rate line (left-axis ticks + polyline); more turns only make the line denser while width stays constant, closing on outside-click / Esc / scroll.
• Robustness: turning the panel off or reinstalling the injection on upgrade now cleans up the popover's global listeners and leftover nodes so nothing leaks into the Codex page; cumulative total and trend series are read in a single disk pass, adding no per-tick cost.

Verification: adapters / usage_tracker / injector unit tests all green, fmt / clippy clean, reviewed by 3 audit agents.

v2.4.2

Codex App Transfer v2.4.2

中文

WorkBuddy 多账号池 + 配额守护自动切换 + 反应式失败转移

为 workbuddy-login provider 增加多账号池能力:单 provider 下登录多个账号,配额守护自动切换,每账号独立设备指纹隔离风控,对 Codex 透明。

• 多账号池:同一 workbuddy-login provider 下登录多个账号组成池,每账号独立 X-Device-Id(避免网关风控关联多账号同设备);账号列表显示脱敏手机号(如 183****5600)替代不可读 uid,可手动设为当前 / 移除。
• 配额守护:后台每 ≥60s 查各账号 get-user-resource 剩余积分,某账号总剩余低于阈值(默认 20,设置项可调)即标记耗尽、自动 sticky 切到下一个有额度账号;独立于额度面板开关,面板关也跑;全耗尽兜底选最快解禁账号。
• 反应式失败转移:选中账号 refresh token 被网关拒或凭证丢失时,标记 5min 退避并重试池里下一个账号;配额守护遇 401/403 标记 1h 退避。两条路径闭环,单账号失效不再撞死整个池。
• 迁移:老单文件 workbuddy-oauth.json + 全局 workbuddy-device-id 首次启动自动迁移为池首账号,无缝。

验证:cargo test gemini_oauth 147 + src-tauri 292 全绿;vue-tsc --noEmit 0 错;CI 全绿;3 轮 code review 意见全部 resolved。

English

WorkBuddy multi-account pool + quota-guard auto-switch + reactive failover

Adds multi-account pool support for workbuddy-login: multiple accounts under one provider, automatic quota-guard switching, per-account device-fingerprint isolation, transparent to Codex.

• Multi-account pool: log into multiple accounts under one workbuddy-login provider, each with an independent X-Device-Id (avoids gateway risk-control correlating accounts on the same device); the account list shows masked phone numbers (e.g. 183****5600) instead of unreadable uids, with manual set-active / remove.
• Quota guard: a background task (≥60s throttle) queries each account's get-user-resource remaining credits; when an account's total remaining drops below the threshold (default 20, configurable) it is marked exhausted and the sticky selector auto-switches to the next account with quota; independent of the quota-panel toggle, runs even when the panel is off; when all are exhausted it falls back to the soonest-unblocking account.
• Reactive failover: when the selected account's refresh token is rejected or its credential is lost, it is marked with a 5min backoff and the next pool account is tried; the quota guard marks 1h backoff on 401/403. Both paths close the loop so a single dead account no longer stalls the whole pool.
• Migration: the legacy single-file workbuddy-oauth.json + global workbuddy-device-id auto-migrate into the pool's first account on first launch, seamlessly.

Verification: cargo test gemini_oauth 147 + src-tauri 292 all green; vue-tsc --noEmit 0 errors; CI all green; 3 rounds of code review all resolved.

v2.4.1

Codex App Transfer v2.4.1

中文

Antigravity 接入 superpowers 强约束插件 + GLM 等 reasoning 档位在新版 Codex 恢复显示 + Codex 草稿暂存 + Trae 登录

本版接入 Antigravity superpowers 强约束插件,修复新版 Codex 砍掉 GLM 等思考档显示的回归,新增草稿暂存与 Trae 登录。

• Antigravity superpowers 强约束插件:把 obra/superpowers 完整接入 Antigravity(Gemini),约束其 coding 时超范围 / 假完成;隐藏开关 antigravitySuperpowersEnabled 默认开,检测到自装则关。
• GLM 等 reasoning 档位恢复显示(MOC-285):Codex 26.623 新增 enabled-reasoning-efforts(默认不含 none / max),把 GLM 等的 none / max 砍成 Medium。app 启动 Codex 前把隐藏档并入持久 atom(只增不删),还原显示。
• Codex 草稿暂存:输入框旁 push / pop + Usage 面板下方暂存面板,临时存取未发送输入。
• Trae 账号登录:Trae provider OAuth 登录 + 额度查看,每账号设备指纹隔离。
• 修复:GLM 文本款(glm-5.2)漏判视觉致含图会话(view_image 截图)卡死,现按版本谓词剥图(#563);小米 MiMo 预设 apiFormat 改 responses;cas_ 校验消除间歇 401;依赖升级。

验证

• cargo test --workspace / cargo fmt --all -- --check / cargo clippy 全绿;CI 全绿
• MOC-285 真机手测:重启 Codex 后 glm-5.2 选择器恢复 None / Max
• GLM 含图会话真机手测:#563 修复后卡死会话恢复可继续

English

Antigravity gains the superpowers constraint plugin + GLM reasoning tiers display again on new Codex + Codex draft stash + Trae login

This release adds the superpowers constraint plugin for Antigravity, fixes the new Codex regression hiding GLM-style thinking tiers, and adds a Codex draft stash and Trae login.

• Antigravity superpowers plugin: integrates obra/superpowers into the Antigravity (Gemini) provider to rein in scope-exceeding / fake-completion behavior; hidden toggle antigravitySuperpowersEnabled on by default, off if a user install is detected.
• GLM reasoning tiers display again (MOC-285): Codex 26.623 added enabled-reasoning-efforts (default excludes none / max), collapsing the none / max tiers of GLM and others to Medium. Before each Codex launch the app unions those hidden tiers into the persisted atom (adds only), restoring the native display.
• Codex draft stash: a push / pop stash next to the composer plus a stash panel below the Usage panel.
• Trae login: OAuth login and quota viewing for the Trae provider, with per-account device-fingerprint isolation.
• Fixes: GLM text models (glm-5.2) were misclassified as vision, hanging image-bearing conversations (view_image screenshots) — now stripped via a version predicate (#563); Xiaomi MiMo preset apiFormat switched to responses; the cas_ format is validated defensively, eliminating intermittent 401s; dependency bumps.

Verification

• cargo test --workspace / cargo fmt --all -- --check / cargo clippy all green; CI green
• MOC-285 manual: after restarting Codex the glm-5.2 picker shows None / Max again
• GLM image conversation manual test: the stuck conversation continues again after #563

v2.4.0

Codex App Transfer v2.4.0

中文

新增 OpenCode Go 供应商 + OpenCode / Kimi Code 网页套餐额度 + 额度框架重构

本版新增 OpenCode Go 预设供应商,为 OpenCode Go 与 Kimi Code 接入网页控制台套餐额度,把多 provider 额度统一成三槽模型与共用 web-session 框架,并打磨供应商编辑页。

• 新增 OpenCode Go 供应商:opencode.ai/zen/go/v1(openai_chat),涵盖 GLM / Kimi / DeepSeek / MiMo 系编程模型;附官方图标。
• OpenCode Go 套餐额度:用量面板显 5 小时 / 每周 / 每月三档剩余%(编辑页「登录 OpenCode」抓控制台 session + workspace id,daemon 解析 /workspace/<id>/go SSR)。
• Kimi Code 套餐额度:kimi-code(api.kimi.com/coding)显套餐用量,鉴权用控制台 localStorage access_token(「登录 Kimi」+ 注入 JS cookie 桥抓取),daemon 调 connect-RPC MembershipService/GetSubscriptionStat。
• 额度框架重构:统一成 5h/周/月三槽 RollingWindows;MiMo / OpenCode / Kimi 共用 web_session_quota(网页登录抓 session → 查控制台用量),新增同类 provider 只需填规格 + 写 fetcher。
• 供应商编辑页:获取模型时过滤走 /messages(Anthropic)端点的模型(OpenCode Go 的 MiniMax / Qwen 系)+ 持久提示;API Key 框加显示 / 隐藏按钮。

验证

• cargo test --workspace(249 通过)、cargo fmt --all -- --check
• 多 agent code-review(xhigh):修复 cookie 误删 / daemon panic / 残缺 session / MiMo cookie 回归 4 项
• 真机手测:OpenCode Go 三档 + Kimi Code 套餐用量面板;两者内嵌登录抓取

English

New OpenCode Go provider + OpenCode / Kimi Code web-console quota + quota framework refactor

Adds the OpenCode Go preset provider, wires web-console plan quota for OpenCode Go and Kimi Code, and unifies multi-provider quota into a three-slot model with a shared web-session framework.

• New OpenCode Go provider: opencode.ai/zen/go/v1 (openai_chat), covering the GLM / Kimi / DeepSeek / MiMo coding models; ships its official icon.
• OpenCode Go quota: the panel shows 5-hour / weekly / monthly remaining % ("Sign in to OpenCode" captures the console session + workspace id; the daemon parses the /workspace/<id>/go SSR).
• Kimi Code quota: kimi-code (api.kimi.com/coding) shows plan usage, authed via the console's localStorage access_token ("Sign in to Kimi" + an injected-JS cookie bridge), with the daemon calling connect-RPC MembershipService/GetSubscriptionStat.
• Quota framework refactor: unified into a 5h/weekly/monthly three-slot RollingWindows; MiMo / OpenCode / Kimi share a web_session_quota framework, so a new same-kind provider only needs a spec + a fetcher.
• Provider edit page: fetching models filters out /messages (Anthropic) endpoint models (OpenCode Go's MiniMax / Qwen) with a persistent notice; the API Key field gains a show / hide button.

Verification

• cargo test --workspace (249 passing), cargo fmt --all -- --check
• Multi-agent code-review (xhigh): fixed 4 issues — cookie-wipe, daemon panic, partial session, MiMo cookie regression
• Manual: OpenCode Go three-window + Kimi Code plan-usage panels; embedded sign-in for both

v2.3.4

Codex App Transfer v2.3.4

中文

多 Provider 额度与推理打磨 + 协议 adapter 收敛 + 联网统一

本版扩展多 Provider 额度 / 推理档,把 responses 透传纳入统一 adapter 框架,统一联网工具,并修复 archive 背景异常。

• Provider 额度与推理档 (MOC-211 / MOC-244 / MOC-241):用量面板新增多 Provider 额度(GLM / MiMo / DeepSeek / anyrouter);Kimi(Moonshot)PAYG 账户余额(对齐 DeepSeek);GLM 等思考模型在 reasoning 选择器显示原生两档 none / max。
• 协议 adapter 收敛 (MOC-234 / MOC-243 / MOC-11):responses ↔ responses 透传纳入统一 adapter 框架(1:1 直透,移除 direct);Gemini compact 摘要排除 thought + V2 prev_id 重建历史;compact 路径剥离历史 reasoning items 省 input budget。
• 联网统一 (MOC-208):移除所有第三方 Provider 原生 web_search,统一走自研联网工具。
• 工具 / 性能 / 兼容 (MOC-235 / MOC-232 / MOC-214 / MOC-96):新增通用工具输出回取 read_tool_artifact;上下文 breakdown 的 o200k 计算移出转发关键路径;rollout reader 支持 .jsonl.zst 压缩检测;app 内 HTTP client 补 connect_timeout 基线。
• UX 与修复 (MOC-106 / MOC-247):MCP server / 文档写配置前二次确认;修复 archive 页背景注入显示异常(选择器漂移致主面板半透明、hover 撕裂)。

验证

• cargo test --workspace
• cargo fmt --all -- --check
• 真机手测:额度面板 / archive 背景;chat、compact、联网、antigravity(gemini)

English

Multi-provider quota & reasoning polish + protocol adapter consolidation + unified web search

Extends multi-provider quota / reasoning, brings responses passthrough into the unified adapter framework, unifies web search, and fixes the archive-page background anomaly.

• Provider quota & reasoning tiers (MOC-211 / MOC-244 / MOC-241): the usage panel adds multi-provider quota (GLM / MiMo / DeepSeek / anyrouter); Kimi (Moonshot) PAYG account balance (aligned with DeepSeek); GLM and other thinking models show their native none / max tiers in the picker.
• Protocol adapter consolidation (MOC-234 / MOC-243 / MOC-11): responses ↔ responses passthrough joins the unified adapter framework (1:1 direct, direct mode removed); Gemini compact excludes thought + rebuilds history via V2 prev_id; the compact path strips historical reasoning items to save budget.
• Unified web search (MOC-208): removes every third-party provider's native web_search in favor of the in-house web tool.
• Tools / perf / compat (MOC-235 / MOC-232 / MOC-214 / MOC-96): adds a tool-output retrieval tool read_tool_artifact; moves the context-breakdown o200k compute off the forwarding hot path; the rollout reader detects .jsonl.zst compression; the in-app HTTP client gains a baseline connect_timeout.
• UX & fixes (MOC-106 / MOC-247): confirm before MCP server / doc config writes; fix the archive-page background injection anomaly (selector drift left the main panel translucent, tearing on hover).

Verification

• cargo test --workspace
• cargo fmt --all -- --check
• Manual: quota panel / archive background; chat, compact, web search, antigravity (gemini)

v2.3.3

Codex App Transfer v2.3.3

中文

新增 Codex 内置 Usage 用量面板 + 上游兼容打磨

本版新增 Codex Desktop 内置 Usage 用量面板与 Claude 风格上下文 by-source 明细下拉,并修正 gemini 工具重打包精度、回退冗余的 chat preamble 注入。

• Usage 面板 (MOC-204 / MOC-230):Codex Toggle pinned summary 弹窗底部注入可折叠 Usage 分区(5 小时 / 每周额度、上下文、Tokens 实时速率与累计、缓存命中率,设置默认关);累计 / 缓存 / 速率按 React fiber 当前 conversationId 取数,切对话不串号、解析不到显 —。
• 上下文 by-source 明细下拉 (MOC-231):上下文行 caret 展开分类明细(工具调用与输出 / 推理 / 开发者指令 / 对话消息 / 工具定义 / 系统提示),proxy 用 o200k_base 逐 item 计数、按对话 uuid 持久化重启即用,占比按权威 used 缩放对齐。
• gemini 工具重打包精度 (MOC-88):custom_tool_call 重打包从 name-only 改为按请求注册的 type:custom 判定,防撞名误 repack 并支持 apply_patch 外的 custom 工具(latent,无可见变化)。
• 回退 preamble + 杂项 (MOC-219 / MOC-229):Codex 26.609.41114 自行恢复工具轮间思考渲染,回退 PR #458 chat preamble 注入;移除失效的上下文圆环开关。

验证

• cargo test --workspace
• cargo fmt --all -- --check
• 真机手测:Usage 面板额度 / 明细下拉 / 切对话隔离;chat 与 antigravity(gemini)路径

English

New built-in Usage panel inside Codex + upstream compatibility polish

Adds a built-in Usage panel inside Codex Desktop's pinned-summary popup with a Claude-style per-source context breakdown dropdown, and fixes gemini tool-repack precision while reverting the now-redundant chat preamble injection.

• Usage panel (MOC-204 / MOC-230): injects a collapsible Usage section at the bottom of Codex's Toggle pinned summary popup (5-hour / weekly quota, context, real-time + cumulative tokens, cache hit rate; off by default); cumulative / cache / rate key to the active conversationId from the React fiber, so switching conversations no longer cross-contaminates and shows — when unresolved.
• Per-source context breakdown dropdown (MOC-231): a caret on the context row expands a breakdown (tool calls & output / reasoning / developer instructions / conversation messages / tool definitions / system prompt), counted by the proxy with o200k_base, persisted per conversation uuid for instant reload, scaled to the authoritative used.
• gemini tool-repack precision (MOC-88): custom_tool_call repack moves from name-only to the request-registered type:custom check, preventing mis-repack on name collisions and supporting custom tools beyond apply_patch (latent, no visible change).
• Revert preamble + misc (MOC-219 / MOC-229): Codex 26.609.41114 restored inter-tool reasoning rendering, so PR #458's chat preamble injection is reverted; the dead context-ring toggle is removed.

Verification

• cargo test --workspace
• cargo fmt --all -- --check
• Manual: Usage panel quota / breakdown dropdown / conversation isolation; chat and antigravity (gemini) paths

v2.3.2

Codex App Transfer v2.3.2

中文

联网搜索与 antigravity 增强 + 新版 Codex 兼容修复

新增智谱 GLM Coding preset 与 antigravity 原生出图,联网搜索开箱可用并升级双引擎,修复新版 Codex 多项兼容问题。

• 智谱 GLM Coding 内置 preset:适配 GLM Coding Plan 订阅专属端点 /api/coding/paas/v4,注入 Claude Code UA;原开放平台 preset 并存。
• antigravity 增强 (MOC-210 / MOC-221 / MOC-217):Codex 内置 image_gen 原生出图;gemini context window 兜底 258k 修正为真实 1M,auto-compact 不再过早触发;适配 tool_search builtin,恢复 MCP 工具暴露。
• 联网搜索增强 (MOC-215 / MOC-227):web_fetch_backend 默认 auto,内置联网工具开箱可用;web_search 改 DDG+Bing 并行合并去重,新增 web_search_more 翻页工具;移除 web_fetch 摘要兜底及 summaryModel 配置整链。
• 新版 Codex 兼容 (MOC-219 / MOC-218):修 26.609 工具轮间无文本全折叠,chat 路径节流注入 preamble(纯文本回答改为整段一次性出现);gemini thoughtSignature 改存本地缓存,reasoning item 出 wire 不带 content,修切真 GPT 两连 400。
• 主题对话区可读性 (MOC-225):#root 升 3 层复合 scrim,11 套主题按壁纸亮度逐套校准。
• 杂项:Usage 统计时区/精度/缓存打磨 (MOC-19);重复 system 指令块 wire 级去重 (MOC-193);release notes 文件发版门禁 (MOC-66)。

验证

• cargo test --workspace
• 真机手测:web_fetch 全文返回、antigravity image_gen 出图、GLM Coding preset 连通
• Codex 26.609 真机验证工具轮间文本渲染

English

Web search and antigravity upgrades + new Codex compatibility fixes

Adds the Zhipu GLM Coding preset and native antigravity image generation, makes built-in web search work out of the box, and fixes newer-Codex issues.

• Zhipu GLM Coding built-in preset: targets the GLM Coding Plan subscription endpoint /api/coding/paas/v4 with a Claude Code UA; the open-platform preset coexists.
• antigravity upgrades (MOC-210 / MOC-221 / MOC-217): built-in image_gen now generates images natively; gemini context window fallback corrected from 258k to 1M; tool_search builtin adapted, restoring MCP tool visibility.
• Web search upgrades (MOC-215 / MOC-227): web_fetch_backend defaults to auto; web_search merges DDG and Bing in parallel with dedup, plus web_search_more pagination; the web_fetch summary fallback and summaryModel chain are removed.
• New Codex compatibility (MOC-219 / MOC-218): fixes 26.609 collapsing tool-only turns via throttled preamble injection; gemini thoughtSignature moves to a local cache and reasoning items drop wire content, fixing 400s when switching to a real GPT account.
• Theme readability (MOC-225): 3-layer composite #root scrim, all 11 themes calibrated per wallpaper brightness.
• Misc: usage stats timezone/precision/caching polish (MOC-19); wire-level dedup of repeated system instruction blocks (MOC-193); release-notes file gate (MOC-66).

Verification

• cargo test --workspace
• Manual: web_fetch full text, antigravity image_gen, GLM Coding preset connectivity
• Tool-turn text rendering verified on Codex 26.609

v2.3.1

Codex App Transfer v2.3.1

中文

新版 Codex Desktop 思考显示修复 + 多项稳定性增强

修复新版 Codex Desktop(v26.608+)升级后第三方模型思考过程整段不显示,并补齐 v2.3.0 之后合入的多项功能。

• 思考过程不显示 (MOC-203):新版渲染读 reasoning 的 content 通道(reasoning_text)而非旧 summary 通道。gemini_native / chat 改为双发两通道兼容新旧版、剥 **Thinking** 头;开工具前先闭合 reasoning 修事件交错丢思考,交错思考重开新 item,gemini 修同轮工具折叠空白行。
• remote compaction v2 (MOC-198):新版 autocompact 改发普通 /responses + compaction_trigger 标记,此前代理不识别导致 expected exactly one compaction output item 挂死;新增双轨检测、按协议包装 compaction SSE 流,四大 mapper 适配。
• apply_patch 中间层 (MOC-194):第三方模型畸形 V4A patch 按白名单恢复成合法格式再发 Codex,非破坏性、未知原样透过。
• 强杀自愈 (MOC-197):kill -9 后残留 sandbox_mode 致 GPT 账号无法对话,重启自动还原 stale 快照、拍快照过滤残留防污染基线。
• antigravity 指纹 (MOC-67) + 残留面板新增显示重复字段的只读入口 + macOS 15+ Gatekeeper 文档修正。
• 依赖升级:windows 0.62、tokio-tungstenite 0.29、dom_query 0.28、dom_smoothie 0.18。

验证

• cargo test -p codex-app-transfer-adapters
• cargo test -p codex-app-transfer --bin codex-app-transfer
• 运行中 Codex v26.608 真机验证 reasoning 双通道显示

English

Reasoning display fix for new Codex Desktop + stability improvements

Fixes third-party reasoning blocks disappearing after upgrading to Codex Desktop v26.608+, plus features merged after v2.3.0.

• Reasoning not shown (MOC-203): new Codex reads the reasoning content channel (reasoning_text), not the old summary channel. gemini_native / chat now dual-emit both channels and strip the **Thinking** header; reasoning is closed before tool calls to fix interleaving, a new item reopens for interleaved thinking, and gemini fixes tool-folding blank lines.
• remote compaction v2 (MOC-198): new autocompact sends a normal /responses with a compaction_trigger marker; the proxy previously hung with expected exactly one compaction output item. Added dual-track detection wrapping the compaction SSE per Codex v2.
• apply_patch middle layer (MOC-194): malformed V4A patches from third-party models are recovered to valid format before reaching Codex, non-destructive, unknown forms passed through.
• kill-9 self-heal (MOC-197): leftover sandbox_mode after a hard kill broke GPT accounts; restore now heals stale snapshots and filters residue when snapshotting.
• antigravity fingerprint (MOC-67) + residual-scan read-only show-duplicate-fields entry + macOS 15+ Gatekeeper doc fix.
• Dependency bumps: windows 0.62, tokio-tungstenite 0.29, dom_query 0.28, dom_smoothie 0.18.

Verification

• cargo test -p codex-app-transfer-adapters
• cargo test -p codex-app-transfer --bin codex-app-transfer
• Real-device Codex v26.608 check of dual-channel reasoning display

v2.3.0

Codex App Transfer v2.3.0

中文

web_fetch 综合能力重做:默认返回全文 + 本地回看 + 历史压缩

web_fetch 此前强制把抓到的正文喂第三方总结子模型摘要,综合能力堪忧:精确信息被弱模型改写、前置条件脆弱、与 web_search 两段式割裂。

本版改为默认直接返回完整正文(主模型自己读),摘要降为 summarize=true 显式 opt-in;adapter 配套当前轮全文、历史压缩,新增 read_url_local 让模型从本地缓存回看历史抓过的 URL。

本版改动:

• web_fetch 默认返回完整正文,不再分页 / offset 翻页 / query 选块;summarize=true 才走摘要(opt-in)。
• adapter 保留当前轮(本次新产生那组,一轮可调多工具)所有 tool 输出全文,更早历史压缩;current 与 cached 按合并后末尾连续 tool 组区分。
• 新增 read_url_local(url),从进程内缓存取回历史 URL 完整正文;initialize 用 server 级 instructions 引导优先回看而非重抓。
• 修 curl / wreq 未解压 gzip / brotli 响应导致的乱码,补 gzip / brotli / deflate / zstd 解压 feature。
• web_search max_results 默认 8 到 15、上限 20 到 30;cat-webfetch serverInfo.icons 用本应用图标。

验证

• cargo test -p codex-app-transfer-adapters
• cargo test -p codex-app-transfer --bin codex-app-transfer mcp_webfetch_server
• cargo test -p codex-app-transfer-http
• CI cf-canary real network 验证解压 feature 不破 ClientHello 指纹

English

web_fetch overhaul: full text by default, local re-read, history compaction

web_fetch used to force every fetched page through a third-party summary model, which hurt quality: precise content got rewritten, preconditions were fragile, and it broke the web_search to web_fetch flow.

This release returns the full extracted text by default (the main model reads it), with summarization demoted to an explicit summarize=true option; the adapter keeps the current turn full and compacts history, and a new read_url_local tool re-reads a previously fetched URL from the local cache.

Changes in this release:

• web_fetch returns full text by default, no pagination / offset / query chunk selection; summarize=true opts into the summary model.
• The adapter keeps the whole current-turn group of tool outputs full (a turn may call multiple tools) and compacts earlier history; current vs cached is split by the trailing consecutive tool group after merge.
• New read_url_local(url) reads a previously fetched URL from the in-process cache; initialize ships server-level instructions to prefer re-reading over re-fetching.
• Fixed garbled output from curl / wreq not decompressing gzip / brotli responses, by enabling the gzip / brotli / deflate / zstd decode features.
• web_search max_results default 8 to 15, cap 20 to 30; cat-webfetch serverInfo.icons uses the app icon.

Verification

• cargo test -p codex-app-transfer-adapters
• cargo test -p codex-app-transfer --bin codex-app-transfer mcp_webfetch_server
• CI cf-canary real network for decode-feature fingerprint safety