HanaAgent 1.1.111-beta

包含自 v1.1.35-beta 以来的更新。

这次更新为编程工作加入独立的 Code 页与快速动作，并重新整理了子代理协作、后台任务和审批的使用体验。工作目录、项目技能、中文搜索、模型设置、内置浏览器与日常界面也有一批改进。

建议使用完整安装包升级。 本次包含桌面运行时安全更新；退出确认、程序坞角标和部分更新检查修复也需要更新应用本体，单独安装 OTA 不包含这些变化。

1. Code：独立的编程工作空间

从项目开始，而不是从空白聊天开始

• 新增 Code 页，将编程会话与主页会话分开管理，并配有独立的文件树、改动视图、预览和终端。
• 新建会话时选择工作目录、Agent 和工作树；可以使用已有工作树，也可以从指定分支创建新的工作树，让不同任务在各自目录中进行。
• 尚未使用 Git 的文件夹可以在创建会话时选择初始化仓库。新建工作树是可选操作，不会因为开始一次 Code 对话就自动创建。
• 新工作树的目录与分支名称可以根据任务生成；只附带文件的首条消息也能进入创建流程。
• Code 会话列表支持按时间或项目整理、收起项目分组、置顶和拖动置顶项，并显示目录名与分支，方便区分同一项目里的不同任务。

直接查看文件和实际改动

• 新增改动卡，按文件列出新增、删除及修改情况，并查看带语法着色的差异。
• 改动视图根据卡片宽度调整文件列表与内容的布局，长代码支持横向滚动。
• 改善刷新时的闪动和重复读取；刷新失败时保留已有列表，不把正在查看的内容突然清空。
• 二进制和超大文件显示说明，不强行展开内容；优化未跟踪文件统计及大文件差异读取，减少不必要的等待。
• 文件树支持查看隐藏文件。Code 页使用独立的预览和终端位置，切回主页时不会把两个工作空间的内容混在一起。

为编码配置 Agent 与协作方式

• 内置「探索（Explorer）」与「手工（Artisan）」两个 Code 子代理，分别面向只读调查和实施工作；探索子代理的只读限制由系统执行。
• Code 会话不读取或写入个人记忆，也不参与后续的记忆整理，避免项目细节混入日常个人记忆。
• 设置新增 Code 分页，可调整默认 Agent、权限模式、是否默认新建工作树、起始分支、分支名前缀、隐藏文件显示和等宽字体。
• 可管理参与协作的子代理、编辑其身份信息，并设置同时运行的子代理上限。
• Code 界面采用更适合代码阅读的等宽排版与方角卡片；欢迎页、思考提示和新建任务的加载提示也作了相应调整。

2. 快速动作：把常用操作留在项目里

项目脚本、个人命令和提示词放在一处

• Code 会话标题栏新增「快速动作」，可搜索动作、查看详情，也可以将列表作为卡片打开。
• 集中显示仓库提供的动作、个人保存的动作，以及项目 package.json 中的 Scripts。
• 动作可以包含命令、提示词，或同时包含两者。提示词也可先填进输入框作为草稿，由你修改后再发送。
• 可以添加、编辑和删除个人动作，或打开仓库动作的来源文件查看定义。动作详情会说明来源、运行时机及确认状态。
• 项目脚本跟随当前会话所在的工作树；切换工作目录后，动作列表也随之更新。

看得见执行过程，也能随时停止

• 在列表中运行或停止动作，并直接跳到对应终端查看输出；正在运行的动作有单独分组。
• 聊天中显示动作变更和 Setup 执行回执；连续回执与文件条目可以折叠，减少长串记录占据对话。
• Agent 可以建议把重复操作保存为动作；建议以确认卡呈现，由你选择加入「我的」或仓库，不会仅因提出建议就自动保存或执行。

工作树准备与归档检查

• 创建工作树时可选择「创建后运行 Setup」，把项目准备命令接到新工作树流程中。
• 需要确认的 Setup 会在创建前显示；如果 Setup 内容发生变化，会要求重新确认。读取失败时保留草稿，不把未知情况当成已经同意。
• 支持在归档前运行检查命令。尚未确认或执行失败时会明确提示，由你决定继续、跳过或取消。
• 需要信任的命令在运行前先经确认；相关动作内容变化后，旧确认不会直接套用到新内容。

3. 子代理与后台任务：进展、审批和结果更连贯

在所属会话里看见协作进展

• 会话现在汇总它派出的子代理、工作流及相关后台任务，区分正在运行、正在停止、执行状态未知和等待审批。
• 会话列表显示待审批标记，不必逐个打开子代理寻找卡住的确认。
• 重连或切换会话后会重新补齐协作状态，减少任务仍在运行、界面却像已经结束的情况。

审批回到你能处理的地方

• 子代理发起的工具审批、交互询问和技能确认可以转交到上级会话处理。
• 审批卡标明实际请求者，便于判断是哪位子代理、哪项工作需要授权。
• 停止子代理或工作流时，一并作废已经转交但尚未处理的相关审批，避免任务结束后还留着可点击的旧请求。
• 等待审批的会话不会仅因聊天缓存腾出空间而被清出。

停止、续接与结果送达更可靠

• 停止父会话时，处理其下级任务和终端工作；停止过程不再提前伪装成已经结束。
• 保留子代理线程记录和任务摘要，避免停止后失去继续协作的上下文。
• 子代理的后台工作完成后，可以续接到对应子代理处理；原接收者已经不可用时，将结果交给可用的上级会话。
• 修复部分后台结果丢失、重复唤醒和重复送达问题；已经处理过的结果不会因为恢复或重连再次被当成新结果投递。
• 改善重启后遗留任务的状态恢复。远端设备失联时，保留“状态未知”的区别，而不是直接把仍无法确认的任务判为成功或失败。
• 修复子代理名称和头像错误显示为当前主 Agent 的问题。

4. 工作目录、项目技能与文件夹授权

• 聊天标题支持查看和切换工作目录，不必离开当前对话去寻找目录设置。
• 「本次对话」卡重新整理了项目技能入口、目录芯片和折叠控制；只展开与当前会话有关的补充信息，例如远端设备或 Code 分支状态。
• Code 分支信息可查看当前分支、跟踪分支、领先情况和未提交改动，方便判断正在操作哪个工作树。
• 工作目录改变后，会话列表、目录显示、分支信息和快速动作同步刷新，减少不同位置各自显示旧目录的问题。
• 项目技能按当前工作目录读取、安装和删除；进入项目子目录时，也能正确处理对应位置的技能。
• 工作目录与额外授权文件夹的选择更清楚；文件夹选择器中「用过的」条目可以正常加入或移出授权范围，嵌套文件夹也可分别控制。
• 修复跨空间切换时误用其他空间文件夹信息，以及部分会话工作目录显示不一致的问题。

5. 模型设置与服务商兼容

• 思考档位可以命名和排序。 为模型支持的档位设置更易懂的显示名称，并调整会话菜单中的顺序；实际发送给服务商的档位值仍与显示名称分开。
• 单个模型可以覆盖 API 地址。 同一服务商下的模型不必全部共用一个地址；清除覆盖后恢复跟随服务商配置。
• 单个模型可以设置受支持的缓存时间。 根据实际使用的协议检查可用设置，避免保存不适用的组合。
• 新增 Open Design 服务商预设与图标。 配置和选择时更容易辨认。
• 改进不同服务商提供的 DeepSeek 模型在角色扮演场景下的思考内容兼容，并完善 Claude 5 系列思考模式的识别。
• 修复自定义服务商修改 API 地址后没有及时保存、检测仍使用旧地址的问题；校验失败时提供可见的失败原因。
• 修复模型编辑面板超出设置区域的问题。

6. 中文搜索与长对话

• 改进会话搜索和会话内查找对中文词语的处理，短词、专有名词及不常见的组合更容易匹配。
• 记忆检索采用一致的中文匹配方式，减少同一个词在不同入口表现不一致的问题。
• 改进 Agent 查找工具时的中文检索，使中文任务描述更容易找到相关能力。
• 超大上下文模型的自动压缩触发阈值最高为 50 万 token，不再一直等到大窗口接近九成占用；原本触发点更低的模型保持原规则，手动压缩方式不变。

7. 内置浏览器、阅读与日常界面

• 改善内置浏览器在应用重启后的页面恢复；保存的页面无法直接恢复时，可按上次地址重新打开，减少「找不到浏览器标签页」的情况。
• 浏览器工具栏增加批注开关；前进、后退和刷新与地址栏重新整理，没有批注时不再留下多余空栏。
• 调整浏览器页面区域与外框的衔接、裁切和圆角，修复控件附近的白带及页面边缘不协调的问题。
• Markdown 编辑器改善段落之间的垂直留白，阅读长文和定位段落更清楚。
• 统一文件面板、工作区工具条与会话控制的布局，减少紧凑面板中的错位和多余留白。
• 改善页面切换动效，以及思考提示、输入状态条和消息发送时的排版一致性。
• 连续文件条目可以折叠显示，文件卡操作按钮也更紧凑。
• 修复草稿达到数量上限后无法继续保存的问题：保留新草稿，并淘汰最旧条目。

8. 通知、退出与更新体验

• 完成通知会考虑会话所在页面：只有你正在查看相应页面时才抑制打扰，避免停留在另一个页面却错过完成消息。
• 有任务正在运行时，退出应用会先提示；从系统托盘退出也采用相同检查。
• macOS 程序坞角标可显示等待审批的会话数量，方便在应用不在前台时发现待办确认。
• 更新检查失败时，更明确地区分网络、代理和服务器响应问题，不再只显示无法判断原因的笼统错误。
• 启动时网络尚未就绪而导致检查失败，会进行有限次数的短间隔重试；网络恢复并检查成功后，清除已经过时的检查失败状态。
• 更新下载或应用失败不会仅因一次检查成功就被误当成已经修复。

9. 权限、安全与兼容

• 「自动」权限模式下，系统能够确认安全的只读查看命令不再逐条弹出审批，例如符合只读条件的目录查看、文本搜索和 git status；无法确认或包含写入行为的命令不适用。
• 沙盒阻止了任务所需命令时，助手可以说明原因并重新申请在沙盒外执行；这仍须经过审批，不等于自动获得管理员权限，也不会解除系统明确禁止的操作。
• 自动审批使用的模型调用计入「用量统计」的「工具」类别，便于查看这部分消耗。统计以服务商返回的用量为准，实际费用仍以服务商账单为准。
• 更新随附运行时，接入上游针对受影响 eMudhra 中间证书密钥的拒绝验证修复。
• 桌面壳的安全修复需要通过完整安装包获得；OTA 不会替换旧版本的桌面运行时。
• 改善会话、App 与子代理的归属处理，减少 Code 会话出现在主页或错误参与个人记忆的问题，并保留既有 App 的兼容路径。

10. App 能力：输入栏图标与小精灵

以下能力需要相应 App 支持并启用，不会在所有用户界面中自动出现。

• App 放在输入栏中的按钮可以使用自己的 SVG、PNG 或 WebP 图标。App 更新或重载后，图标随之刷新，无需为此重启 Hana。
• App 可以在主窗口输入框上方放置可动画、可点击的小精灵，响应 Agent 工作、等待确认和完成等状态，也可以通过点击调用该 App 的工具。
• 同一窗口同时展示一只精灵；可右键暂时隐藏。精灵不会覆盖发送、停止和审批等受保护的操作区域，启用「减少动态效果」时使用静态画面。

11. 平台与升级说明

• 本次提供 macOS ARM64、Windows x64、Linux x64 安装包及 beta OTA，不包含手机端安装包。
• macOS 安装包须完成签名、公证和校验后发布。Windows 安装包未签名，系统可能提示「未知发布者」；Linux 仍为 Preview 平台。
• 推荐使用完整安装包升级，以获得桌面运行时安全更新、退出确认、程序坞角标和更新检查修复。
• OTA 只更新服务端与界面，不替换桌面壳；本次最低应用版本保持为 1.1.23-beta，更旧版本须先安装新版应用本体。
• 已明确选择 stable 通道的用户不会自动切换到 beta。
• 请从正式 Release 或官网下载，并核对 SHA256。GitHub 自动生成的「Source code」压缩包对应公开仓库的历史源码，并非本次安装包的源码。

HanaAgent 1.1.111-beta — English

Includes changes since v1.1.35-beta.

This release adds a dedicated Code page and Quick actions, and improves the experience of working with subagents, background tasks and approvals. It also updates working directories, project skills, Chinese search, model settings, the embedded browser and everyday interface details.

We recommend updating with the full installer. This release includes desktop runtime security updates. Quit confirmation, the Dock badge and some update-check fixes also require an application update; OTA alone does not include them.

1. Code: a dedicated coding workspace

Start with a project

• Adds a Code page with a conversation list separate from Home, plus its own file tree, changes view, previews and terminals.
• Choose a working directory, agent and worktree when starting a conversation. Use an existing worktree or create one from a selected branch to keep separate tasks in separate directories.
• Folders without Git can be initialized when creating a conversation. Creating a worktree is optional; starting a Code conversation does not automatically create one.
• New worktree directory and branch names can be generated from the task. A first message containing only attachments can also enter the creation flow.
• Organize Code conversations by time or project, collapse project groups, pin conversations and reorder pinned items. Directory and branch labels help distinguish tasks within the same project.

Inspect files and actual changes

• Adds a changes card with per-file additions, deletions and modifications, and syntax-highlighted diffs.
• The file list and diff layout adapt to the card width; long code lines can scroll horizontally.
• Reduces flicker and repeated reads during refresh. If a refresh fails, the existing list remains visible rather than disappearing.
• Binary and very large files show an explanation instead of expanding their contents. Improves untracked-file statistics and large-file diff reads to reduce unnecessary work.
• The file tree can show hidden files. Code previews and terminals have their own placement, keeping them separate from Home when switching pages.

Configure agents and collaboration for coding

• Includes Explorer and Artisan subagents for read-only investigation and implementation respectively. Explorer's read-only access is enforced by the system.
• Code conversations do not read or write personal memory and are excluded from later memory maintenance, keeping project details out of everyday personal memory.
• A new Code settings tab controls the default agent, permission mode, worktree creation preference, base branch, branch prefix, hidden-file visibility and monospaced font.
• Manage the subagents available for collaboration, edit their identities and set the concurrent subagent limit.
• Code uses monospaced typography and square cards, with corresponding updates to its welcome view, thinking indicators and task-creation loading feedback.

2. Quick actions: keep routine operations with your project

Project scripts, personal commands and prompts in one place

• The Code conversation titlebar gains Quick actions, with search, details and an option to open the list as a card.
• Lists repository-defined actions, personal actions and Scripts from the project's package.json.
• An action can contain a command, a prompt or both. Prompts can also be placed in the composer as drafts for editing before sending.
• Add, edit and delete personal actions, or open a repository action's source file. Details identify its source, trigger and confirmation status.
• Project scripts follow the current conversation's worktree. Changing the working directory refreshes the action list.

Follow execution and stop when needed

• Run or stop actions from the list, and jump to their terminals to inspect output. Running actions have their own group.
• Chat shows action-change and Setup receipts. Consecutive receipts and file entries can be collapsed to keep long runs of records compact.
• Agents can suggest saving repeated operations as actions. A confirmation card lets you add a suggestion to personal or repository actions; suggesting it does not automatically save or run it.

Worktree setup and pre-archive checks

• Worktree creation can run a Setup command afterward to prepare the new directory.
• Setup that requires confirmation is shown before creation. Changed content requires confirmation again. If Setup cannot be read, the draft is preserved rather than treating the unknown state as consent.
• Actions can run checks before archiving. Unconfirmed or failed commands produce an explicit prompt so you can continue, skip or cancel.
• Commands requiring trust ask for confirmation before execution. Confirmation for an earlier action does not silently authorize changed content.

3. Subagents and background tasks: clearer progress, approvals and delivery

See collaboration progress in the owning conversation

• Conversations summarize their subagents, workflows and related background tasks, distinguishing running, stopping, unknown execution state and pending approvals.
• Conversation lists mark pending approvals, reducing the need to open each subagent to find a blocked confirmation.
• Reconnecting or switching conversations refreshes collaboration state, reducing cases where the interface looks finished while work is still running.

Handle approvals where you can see them

• Subagent tool approvals, interactive questions and skill confirmations can be forwarded to a parent conversation.
• Approval cards identify the actual requester, making the source of each request clear.
• Stopping a subagent or workflow invalidates its pending forwarded approvals, preventing obsolete requests from remaining actionable.
• Conversations awaiting approval are protected from ordinary chat-cache eviction.

More reliable stopping, continuation and result delivery

• Stopping a parent handles descendant tasks and terminal work without prematurely reporting that stopping is complete.
• Preserves subagent thread records and task summaries so stopping work does not remove the context needed to continue.
• Completed background work can resume the corresponding subagent. If the original recipient is unavailable, the result can be delivered to an available parent conversation.
• Fixes some lost results, duplicate wake-ups and duplicate deliveries during recovery and reconnection.
• Improves recovery of tasks left over after restart. A disconnected remote device retains an unknown-state distinction rather than having unconfirmed work reported as successful or failed.
• Fixes subagents displaying the current main agent's name and avatar instead of their own.

4. Working directories, project skills and folder access

• View and change the working directory from the chat header without leaving the conversation.
• Reorganizes the conversation status card around project skills, the directory chip and collapse controls. Its expanded content shows relevant details such as a remote device or Code branch state.
• Code branch information includes the current and tracking branches, ahead count and uncommitted changes, helping identify the active worktree.
• Changing the working directory refreshes the conversation list, directory labels, branch information and Quick actions together.
• Project skills are read, installed and removed for the current working directory, including the appropriate location within project subdirectories.
• Clarifies working-directory and additional-folder selection. Previously used folders can be added to or removed from authorized access, and nested folders can be controlled separately.
• Fixes inconsistent working-directory displays and cases where switching spaces used folder information from another space.

5. Model settings and provider compatibility

• Name and reorder thinking levels. Assign readable labels to supported levels and arrange their order in the conversation menu. Display names remain separate from the values sent to providers.
• Override the API address per model. Models under one provider no longer need to share a single address. Clearing an override restores the provider setting.
• Set supported cache durations per model. Validation follows the effective protocol to reject unsupported combinations.
• Adds an Open Design provider preset and icon for easier configuration and recognition.
• Improves thinking-content compatibility for DeepSeek roleplay across providers, and thinking-mode detection for the Claude 5 family.
• Fixes changed custom-provider API addresses not being saved promptly or tests still using the old address. Validation failures now provide a visible reason.
• Fixes the model editor extending beyond the settings panel.

6. Chinese search and long conversations

• Improves Chinese handling in conversation search and in-conversation find, including short terms, proper nouns and uncommon combinations.
• Uses consistent Chinese matching for memory recall, reducing differences between search entry points.
• Improves Chinese tool discovery so Chinese task descriptions can find relevant capabilities more effectively.
• Caps the automatic compaction threshold for very large context windows at 500,000 tokens, rather than waiting until a large window approaches 90% occupancy. Models with lower existing thresholds retain their rules, and manual compaction is unchanged.

7. Embedded browser, reading and everyday interface

• Improves browser-page recovery after application restart. When a saved page cannot be restored directly, it can reopen at its last address, reducing browser-tab-not-found errors.
• Adds an annotation toggle to the browser toolbar and reorganizes navigation controls around the address bar. An empty annotation bar no longer occupies space.
• Refines browser-page clipping, corners and the transition to the surrounding frame, including fixes for white bands near controls.
• Improves vertical paragraph spacing in the Markdown editor for clearer long-form reading and editing.
• Aligns file panels, workspace toolbars and conversation controls, reducing misalignment and unnecessary space in compact panels.
• Improves page transitions and typography consistency in thinking indicators, input status and message-send animations.
• Consecutive file entries can be collapsed, and file-card controls are more compact.
• Fixes new drafts being rejected after the draft limit is reached: new drafts are retained while the oldest entries are evicted.

8. Notifications, quitting and updates

• Completion notifications account for the conversation's page. They are suppressed only when you are viewing that page, reducing missed notifications while working elsewhere.
• Quitting with tasks running prompts for confirmation, including quitting from the system tray.
• The macOS Dock badge can show the number of conversations awaiting approval when the application is not in front.
• Update-check failures better distinguish network, proxy and server-response problems instead of collapsing them into an opaque error.
• If a startup check fails before the network is ready, a bounded set of short-delay retries is attempted. A successful check clears obsolete check-failure state.
• A failed update download or application is not considered fixed merely because a later check succeeds.

9. Permissions, security and compatibility

• In Auto permission mode, inspection commands that the system recognizes as safely read-only no longer prompt individually. Examples include eligible directory listings, text searches and git status. Unrecognized commands or commands with write operations do not qualify.
• When the sandbox blocks a necessary command, an agent can explain why and request execution outside the sandbox. This still requires approval, does not automatically grant administrator privileges and does not override explicitly prohibited operations.
• Model calls used for automatic approval now appear in Usage under Tools. Statistics follow usage reported by the provider; the provider's bill remains authoritative for charges.
• Updates bundled runtimes with upstream fixes that reject affected eMudhra intermediate-certificate keys.
• Desktop-shell security fixes require the full installer; OTA does not replace an older desktop runtime.
• Improves conversation, App and subagent ownership handling, reducing Code conversations appearing on Home or incorrectly participating in personal memory while retaining existing App compatibility paths.

10. App capabilities: input-bar icons and companions

These capabilities require an enabled App that implements them. They do not automatically appear for every user.

• App buttons in the input bar can use their own SVG, PNG or WebP icons. Icons refresh when an App updates or reloads, without restarting Hana.
• Apps can place an animated, clickable companion above the main-window composer. It can react to the agent working, waiting for confirmation or finishing, and clicks can invoke the App's own tools.
• A window displays one companion at a time, with a context-menu option to hide it temporarily. Companions do not cover protected send, stop or approval controls, and use a static frame when reduced motion is enabled.

11. Platforms and upgrading

• Provides installers and beta OTA for macOS ARM64, Windows x64 and Linux x64. Mobile installers are not included.
• macOS installers must pass signing, notarization and verification before publication. The Windows installer is unsigned and may show an unknown-publisher warning. Linux remains a Preview platform.
• Use the full installer to receive desktop runtime security updates, quit confirmation, the Dock badge and update-check fixes.
• OTA updates only the server and interface, not the desktop shell. The minimum app version remains 1.1.23-beta; older installations must first update the application with a full installer.
• An explicit stable-channel preference is not changed to beta.
• Download from the official Release or website and verify SHA256. GitHub-generated Source code archives contain the historical public repository snapshot, not the source for these installers.

下载说明 / Download note

请下载本页提供的桌面安装包。GitHub 自动生成的 “Source code” 压缩包对应公开仓库的历史代码，并非本次 beta 安装包的源码。

Use the desktop installers attached to this release. GitHub-generated “Source code” archives contain the historical public repository snapshot, not the source for these beta binaries.