English
- If you have any questions or suggestions about the app, please contact us through our QQ group or GitHub Issues. We will respond as soon as possible.
Fixed
- Aligned the Pi configuration page with the plugin's model catalog, filtering hidden models and validating the default model against the same list to avoid selecting models unavailable in Pi. #527
- Installing or updating the Pi plugin, or applying its configuration, now clears the old model cache so Pi fetches the latest catalog on its next start. If the cache cannot be cleared, subsequent configuration writes stop with an error.
- Fixed a build failure caused by a missing reasoning-level field in Pi model initialization while retaining existing ZCode reasoning-level support.
Improved
- Models connected to DeepSeek Harness through CPA now default to text and image input, including custom model IDs and models without vision metadata. Reapplying the configuration updates previously generated settings while preserving manual overrides.
Usage Notes
- After updating, reapply the DeepSeek Harness configuration and restart its App or Web service. Restart Pi after changing its configuration to load the latest model list.
- Image input settings allow the client to send images; actual vision support still depends on the selected model and upstream service.
Selected Previous Updates
Added
- Added a Restart App button for DeepSeek Harness Desktop, matching the Codex App controls and supporting both automatically detected installations and manually configured executable paths.
- Restart App and the existing Restart Web control the desktop application and Web service separately. Restart App is disabled when no desktop installation is detected.
Fixed
- Fixed missing reasoning level options in ZCode configurations generated by the app. Options now reflect each model’s supported levels, including custom model IDs, while preserving existing manual settings.
- After updating, reapply the ZCode configuration, fully quit and restart ZCode, then start a new session to use the updated reasoning options.
简体中文
- 关于软件，如果您有任何问题或建议，请通过 QQ群 或 GitHub Issues 反馈，我们会尽快处理。
修复
- 修复 Pi 配置页面与 Pi 插件的模型列表不一致的问题，过滤隐藏模型，并按同一列表校验默认模型，避免选中 Pi 中不可用的模型。#527
- 安装、更新 Pi 插件或应用配置时清理旧模型缓存，让 Pi 下次启动获取最新列表；若缓存无法清理，则停止后续配置写入并提示错误。
- 修复 Pi 模型配置合并后缺少推理等级字段导致的编译失败，保留已有的 ZCode 思考级别支持。
优化
- CPA 接入 DeepSeek Harness 的模型默认启用文字和图片输入，包括自定义模型名及未返回视觉能力信息的模型。重新应用配置时会更新旧版自动生成的配置，已有的手动覆盖设置仍会保留。
使用提示
- 更新后请重新应用 DeepSeek Harness 配置，并重启对应的 App 或 Web 服务。Pi 配置变更后也需重启 Pi，才能加载最新模型列表。
- 图片输入配置允许客户端发送图片，实际识图能力仍取决于所选模型和上游服务。
部分历史更新
新增
- 为 DeepSeek Harness 桌面版新增“重启 App”按钮，交互与 Codex App 一致，支持自动识别安装位置和手动指定可执行文件路径。
- “重启 App”与现有“重启 Web”分别控制桌面应用和 Web 服务；未检测到桌面版时，“重启 App”按钮不可用。
修复
- 修复通过本软件配置 ZCode 后无法选择思考级别的问题。现在会根据模型支持的档位生成配置，支持自定义模型名，并保留已有的手动设置。
- 更新后，请重新应用 ZCode 配置，完全退出并重启 ZCode，再创建新会话以使用更新后的思考级别选项。