dsh-codex-community 0.0.1 → 0.0.3

package/CHANGELOG.md

@@ -3,6 +3,84 @@
 本项目遵循 Keep a Changelog 的结构；`0.0.x` 为技术预览。
 This project follows the Keep a Changelog structure; `0.0.x` is a technical preview.
 
+## [0.0.3] - 2026-08-29
+
+### 中文
+
+#### 新增
+
+- OpenAI Codex 设置入口改用独立的代码括号图标，与其他 Provider 插件保持可辨识的视觉差异。
+- 额度卡新增带图标的刷新按钮、刷新中状态、响应式双列布局和更清晰的剩余额度层级；五小时额度始终显示精确重置时间，每周额度在不足 24 小时时显示精确时间，其余时间显示相对天数，完整时间保留在提示与无障碍标签中。
+- 模型设置按“已选择在前、未选择在后”排序；未选择模型默认收起并可按需展开，同时显示已选择、当前显示与总数摘要。
+
+#### 修复
+
+- 缩短设置页内容高度并避免插件覆盖 Harness 弹窗宽度或滚动容器，改善窄屏与较矮窗口中的滚动体验。
+- 模型能力改为与 Grok Build 一致的紧凑标签，使用 `272K`、`128K` 一类上下文与最大输出表示，不推测 provider catalog 未提供的能力。
+- 在收起未选择模型时取消勾选，键盘焦点会回到展开按钮，避免焦点停留在隐藏元素；重置时间解析也会拒绝无法构造有效日期的值。
+
+### English
+
+#### Added
+
+- The OpenAI Codex settings entry now uses a dedicated code-brackets icon that remains visually distinct from other provider plugins.
+- Usage cards now provide an icon button with a refreshing state, a responsive two-column layout, and a clearer remaining-usage hierarchy. Five-hour limits always show the exact reset time; weekly limits show the exact time within 24 hours and otherwise use relative days, while the complete timestamp remains available in tooltips and accessible labels.
+- Model settings now sort selected models before unselected models. Unselected models are collapsed by default and can be expanded on demand, with selected, visible, and total counts shown in the summary.
+
+#### Fixed
+
+- The settings content is shorter and no longer overrides the Harness dialog width or scroll container, improving scrolling in narrow or low-height windows.
+- Model capabilities use compact Grok Build-style labels such as `272K` and `128K` for context and maximum output, without inferring capabilities absent from the provider catalog.
+- Deselecting a model while unselected entries are collapsed returns keyboard focus to the expansion control instead of leaving it on a hidden element; reset-time parsing also rejects values that cannot produce a valid date.
+
+## [0.0.2] - 2026-08-29
+
+### 中文
+
+#### 新增
+
+- 设置页每次进入以及手动刷新时读取真实的五小时与每周额度；Host 通过官方 Codex 客户端使用的 Web 后端兼容接口获取并严格解析数据，Web 页面不会收到 OAuth 凭据或原始响应。
+- 会话模型选择器左侧新增闪电按钮，与 `/codex set fast on|off` 共用当前会话状态；GPT-5.4、GPT-5.5 和 GPT-5.6 系列可从下一次请求起切换官方 Fast 1.5 倍速。
+
+#### 变更
+
+- 额度刷新失败时不再伪装成零余额：保留最近一次已验证读数，或安全降级为最近请求观测/未知状态。
+- 模型展示元数据以当前安装的 provider catalog 为准；推理选择器改用逐模型核验的订阅 Codex 默认值与 Low 至 Max，隐藏通用 `Default`、`Off`、`Minimal`，且不把需要主动任务委派的 `Ultra` 伪装成普通 Provider 档位。
+- 模型卡片的上下文与最大输出改用 `272K`、`128K` 一类紧凑标签，不再展示冗长的原始 token 数字。
+- 保存了 `Off` 或 `Minimal` 的旧会话不再被静默迁移；只有用户点击带有默认模型影响说明的修复按钮后，才切换到当前模型默认档位。
+- Fast 仅在支持的 GPT-5.4、GPT-5.5 和 GPT-5.6 系列请求上发送 priority service tier，会消耗更多额度；关闭或开启均从下一次请求生效，进程重启后恢复关闭。
+- DSH profile smoke 改为核对新版模型设置标题，避免 UI 文案更新被误判为插件加载失败。
+
+### English
+
+#### Added
+
+- The settings page reads real five-hour and weekly usage limits whenever it opens or is refreshed manually. The Host obtains and strictly parses the data through the Web-backend compatibility endpoint used by the official Codex client; the Web page receives neither OAuth credentials nor raw responses.
+- A lightning button to the left of the conversation model selector shares current-session state with `/codex set fast on|off`. GPT-5.4, GPT-5.5, and the GPT-5.6 family can switch to official Fast at 1.5× starting with the next request.
+
+#### Changed
+
+- A failed usage refresh is no longer presented as zero remaining usage. The page retains the latest verified reading or safely falls back to recent-request observation/an unknown state.
+- Display metadata follows the installed provider catalog. The reasoning picker now uses model-specific subscription-Codex defaults and efforts from Low through Max, hides generic `Default`, `Off`, and `Minimal`, and does not disguise agent-level `Ultra` as a plain Provider effort.
+- Model cards format context and maximum output as compact labels such as `272K` and `128K` instead of long raw token counts.
+- Older conversations that saved `Off` or `Minimal` are no longer migrated silently. They switch to the current model default only after the user clicks a repair action that discloses its default-model effect.
+- Fast sends the priority service tier only for supported GPT-5.4, GPT-5.5, and GPT-5.6-family requests and consumes more usage. Enabling or disabling it applies to the next request, and a process restart restores off.
+- The DSH profile smoke now checks the current model-settings title so a UI copy update is not misreported as a plugin-load failure.
+
+## [Unreleased]
+
+### 中文
+
+#### 新增
+
+#### 变更
+
+### English
+
+#### Added
+
+#### Changed
+
 ## [0.0.1] - 2026-08-29
 
 ### 中文

package/README.en.md

@@ -4,7 +4,7 @@
 
 A Codex plugin for DeepSeek Harness with ChatGPT OAuth sign-in, Codex models, image input, and reliable streaming responses.
 
-> Current version: `0.0.1` technical preview, published under npm `latest` with a full GitHub Release. npm package: `dsh-codex-community`.
+> Current version: `0.0.3` technical preview, published under npm `latest` with a full GitHub Release. npm package: `dsh-codex-community`.
 
 ## Features
 
@@ -15,7 +15,10 @@ A Codex plugin for DeepSeek Harness with ChatGPT OAuth sign-in, Codex models, im
 - provide valid pixel and byte budgets for request images, preventing invalid `maxPixels` values;
 - classify `AccountQuotaExceeded` as non-retryable account quota and show a sanitized reset time;
 - preserve safe partial text after a stream failure, while incomplete tool calls stop without replaying the request;
-- use `/codex` to choose Fast and transport for the current session, with Fast off and transport automatic by default.
+- read the real five-hour and weekly usage limits whenever the settings page opens or is refreshed manually, with a safe fallback and no OAuth credentials exposed to the Web page;
+- present remaining usage in compact cards: five-hour limits show the exact reset time, while weekly limits show the exact time within 24 hours and otherwise use relative days;
+- list selected models first in settings and collapse unselected models by default, with compact `K` labels taken from the current provider catalog for context and maximum output;
+- use the lightning button to the left of the model selector or `/codex` to choose official Fast (1.5×) and transport for the current session, with Fast off and transport automatic by default.
 
 ## Requirements
 
@@ -27,7 +30,7 @@ A Codex plugin for DeepSeek Harness with ChatGPT OAuth sign-in, Codex models, im
 Install an exact version:
 
 ```sh
-dsh plugin --profile web add dsh-codex-community@0.0.1
+dsh plugin --profile web add dsh-codex-community@0.0.3
 dsh web
 ```
 
@@ -35,7 +38,7 @@ dsh web
 
 1. Open **Settings → OpenAI Codex** and complete the ChatGPT OAuth prompts.
 2. Enable models on the same page.
-3. Choose a Codex model from the conversation model selector.
+3. Choose a Codex model from the conversation model selector; click the lightning button to its left when Fast is needed.
 
 Available commands:
 
@@ -50,12 +53,12 @@ Available commands:
 /codex set transport auto|sse|websocket|websocket-cached
 ```
 
-Fast and transport preferences are kept only for the current session in the current process; they are not persisted. Fast requests the priority service tier, whose availability depends on the account and service; a failure is not automatically replayed on a lower tier.
+Fast and transport preferences are kept only for the current session in the current process; they are not persisted and return to their defaults after a process restart. For GPT-5.4, GPT-5.5, GPT-5.6 Luna, Sol, and Terra, Fast uses the official priority service tier starting with the next request, targets 1.5× speed, and consumes more usage. Turning it off restores standard speed on the next request. A failed request is not automatically replayed on a lower tier.
 
 Update:
 
 ```sh
-dsh plugin --profile web update dsh-codex-community@0.0.1
+dsh plugin --profile web update dsh-codex-community@0.0.3
 dsh web
 ```
 
@@ -78,10 +81,11 @@ See [Troubleshooting](docs/troubleshooting.en.md) for diagnostic steps.
 ## Support boundaries
 
 - The `0.0.x` line does not promise a stable API; pin the exact version.
-- Before publication, `0.0.1` must pass the Linux, macOS, and Windows platform gates, the complete supply-chain verification, and maintainer approval. It may be formally published with controlled-account validation at `0/13`, provided every unverified capability remains clearly disclosed and validation continues after release.
-- Problems found after publication are fixed and iterated in `0.0.2`; the published `0.0.1` is never overwritten.
-- Quota state comes from recent requests; it is not a live balance or proactive account query.
-- Fast is an explicitly enabled experimental capability; controlled post-release validation continues to confirm real-account entitlement.
+- Linux, macOS, and Windows CI/profile smoke for `0.0.3` reached `3/3`; the accepted commit and platform evidence are in the [acceptance record](docs/releases/v0.0.3.acceptance.json).
+- Live OAuth, conversation, image, transport, and Fast network acceptance for `0.0.3` starts in a fresh record at `0/13 pending`. These unverified capabilities remain clearly disclosed and will be completed individually after publication.
+- The settings page reads usage through the Web-backend compatibility endpoint used by the official Codex client. If that endpoint is unavailable or malformed, the page retains the latest safe reading and falls back to request observation or an unknown state instead of presenting the error as zero remaining usage.
+- Model names, context windows, and maximum outputs come from the installed provider catalog. The reasoning picker exposes only subscription-Codex levels through Max that the Provider can fully implement, plus each model's verified default. Generic `Default`, `Off`, and `Minimal` entries are hidden, and agent-level `Ultra` is not disguised as a plain reasoning effort.
+- Fast is limited to GPT-5.4, GPT-5.5, GPT-5.6 Luna, Sol, and Terra and consumes more usage. Other models never receive the Fast service tier.
 - Image generation and editing are not provided.
 - See [Compatibility](docs/compatibility.en.md) for Windows, Linux, real OAuth, and real Codex network acceptance status.
 

package/README.md

@@ -4,7 +4,7 @@
 
 面向 DeepSeek Harness 的 Codex 插件，提供 ChatGPT OAuth 登录、Codex 模型、图片输入和可靠的流式回复。
 
-> 当前版本：`0.0.1` 技术预览，使用 npm `latest` 与正式 GitHub Release 发布。npm 包名：`dsh-codex-community`。
+> 当前版本：`0.0.3` 技术预览，使用 npm `latest` 与正式 GitHub Release 发布。npm 包名：`dsh-codex-community`。
 
 ## 功能
 
@@ -15,7 +15,10 @@
 - 为请求图片补全有效的像素与字节预算，避免无效 `maxPixels`；
 - 将 `AccountQuotaExceeded` 识别为不可重试的账户配额错误，并显示脱敏后的重置时间；
 - 流中断时保留安全的半段纯文本；涉及未完成工具调用时停止，不自动重放请求；
-- 使用 `/codex` 为当前会话选择 Fast 和传输方式，默认关闭 Fast、自动选择传输。
+- 每次进入设置页或手动刷新时读取真实的 5 小时与每周额度；读取失败时安全降级，不向 Web 页面暴露 OAuth 凭据；
+- 使用紧凑的额度卡展示剩余比例：5 小时额度显示精确重置时间，每周额度在不足 24 小时时显示精确时间，其余时间显示相对天数；
+- 在设置页优先展示已选择模型，默认收起未选择模型；上下文和最大输出使用来自当前 provider catalog 的紧凑 `K` 标签；
+- 使用模型选择器左侧的闪电按钮或 `/codex` 为当前会话切换官方 Fast（1.5 倍速）和传输方式，默认关闭 Fast、自动选择传输。
 
 ## 运行要求
 
@@ -27,7 +30,7 @@
 请固定安装精确版本：
 
 ```sh
-dsh plugin --profile web add dsh-codex-community@0.0.1
+dsh plugin --profile web add dsh-codex-community@0.0.3
 dsh web
 ```
 
@@ -35,7 +38,7 @@ dsh web
 
 1. 打开 **设置 → OpenAI Codex**，按页面提示完成 ChatGPT OAuth；
 2. 在同一页面启用模型；
-3. 从会话模型选择器中选择 Codex 模型开始使用。
+3. 从会话模型选择器中选择 Codex 模型开始使用；需要时点击选择器左侧的闪电按钮开启 Fast。
 
 可用命令：
 
@@ -50,12 +53,12 @@ dsh web
 /codex set transport auto|sse|websocket|websocket-cached
 ```
 
-Fast 和传输偏好只保存在当前进程的当前会话中，不会持久化。Fast 请求 priority service tier，但是否可用取决于账号和服务端；失败后不会自动降级重放请求。
+Fast 和传输偏好只保存在当前进程的当前会话中，不会持久化，进程重启后恢复默认。Fast 在 GPT-5.4、GPT-5.5、GPT-5.6 Luna、Sol 和 Terra 上从下一次请求起使用官方 priority service tier，目标速度为 1.5 倍并消耗更多额度；关闭后，下一次请求恢复默认速度。失败请求不会自动降级重放。
 
 更新：
 
 ```sh
-dsh plugin --profile web update dsh-codex-community@0.0.1
+dsh plugin --profile web update dsh-codex-community@0.0.3
 dsh web
 ```
 
@@ -78,10 +81,11 @@ dsh web
 ## 支持边界
 
 - `0.0.x` 不承诺稳定 API，请固定版本；
-- `0.0.1` 发布前必须通过 Linux、macOS 和 Windows 三平台门禁、完整供应链校验与维护者批准；真实账号验证允许以 `0/13` 状态正式发布，未验证能力必须如实展示并在发布后持续补齐；
-- 发布后发现的问题在 `0.0.2` 中修复和迭代，不覆盖已发布的 `0.0.1`；
-- 额度状态来自最近请求，不是实时余额或主动账户查询；
-- Fast 是显式开启的实验能力，真实账号权限会在发布后的受控验证中继续确认；
+- `0.0.3` 的 Linux、macOS 和 Windows CI/profile smoke 已达到 `3/3`，候选提交与平台证据见[验收记录](docs/releases/v0.0.3.acceptance.json)；
+- `0.0.3` 的真实 OAuth、对话、图片、transport 和 Fast 网络验收从独立记录的 `0/13 pending` 开始；这些未验证能力继续如实展示，并在发布后逐项补齐；
+- 设置页额度通过官方 Codex 客户端使用的 Web 后端兼容接口读取；接口不可用或返回异常时保留最近的安全读数，并降级为请求观测或未知状态，不把错误伪装成零余额；
+- 模型名称、上下文和最大输出来自当前安装的 provider catalog；推理选择器只展示已按订阅 Codex 目录核验且 Provider 能完整实现的 Low 至 Max 与模型默认值，不展示通用的 `Default`、`Off`、`Minimal`，也不把需要主动任务委派的 `Ultra` 伪装成普通推理档位；
+- Fast 仅用于 GPT-5.4、GPT-5.5、GPT-5.6 Luna、Sol 和 Terra，并会提高额度消耗；其他模型不会发送 Fast service tier；
 - 不提供图片生成或编辑；
 - Windows、Linux、真实 OAuth 与真实 Codex 网络验收状态见[兼容性文档](docs/compatibility.md)。
 

package/THIRD_PARTY_NOTICES.md

@@ -11,6 +11,8 @@ This project composes the following independently published packages. Each relea
 | `@deepseek-ai/dsh-client-connection` | Web client connection injection | MIT |
 | `@deepseek-ai/dsh-client-locale` | Web client locale service | MIT |
 | `@deepseek-ai/dsh-client-runtime` | Web client plugin runtime | MIT |
+| `@deepseek-ai/dsh-client-ui-conversation` | Conversation composer slot integration | MIT |
+| `@deepseek-ai/dsh-client-ui-model-selection` | Current-session model directory integration | MIT |
 | `@deepseek-ai/dsh-client-ui-settings` | Web settings-page integration | MIT |
 | `@deepseek-ai/dsh-commands` | Harness command registration | MIT |
 | `@deepseek-ai/dsh-credentials` | Credential-store contracts | MIT |