Herta 0.1.7

黑塔 v0.1.7 — 中断继续、实时操作轨迹、@ 文件

本次更新

中断继续　差分协处理器工作时，如果应用意外退出，或你按下了停止，下次打开该会话时会出现「继续」：工作从中断处接着进行，不必重新描述任务。应用退出时仍在运行的命令，会在下次启动时结束。长时间的工作在上下文整理后保留当前的工作状态，达到步数上限时同样可以继续。

实时操作轨迹　差分协处理器的待办清单已移除。侧栏的「操作轨迹」按探索、修改、验证记录这次工作的经过；正在进行的一步下方有一行实时内容：写入中的文件、运行中的命令输出。这一行按可读的节奏更新，每一步结束时停在它完成的内容上。

输入栏与会话　在输入栏键入 @ 与文件名的一部分，即可从工作区中选择文件；输入栏为空时按 ↑ 可取回上一条消息；黑塔的回复可以一键复制，只复制正文，不含代码。会话菜单新增重命名与导出为 Markdown。窗口不在前台时，如有确认等待处理、工作可以继续、回复失败或较长的回复完成，会发送系统通知；差分协处理器工作期间系统保持唤醒。两项均可在「设置 → 窗口」中关闭。

启动与界面　开场动画改在后台线程绘制，并在两次启动之间保留字形缓存，启动更流畅。输入栏的声波在黑塔说完后缓缓平复，待机时的呼吸更平稳，说话时有一道光沿波纹移动，绘制开销也大幅降低。修复了从仓库卡片打开文件、关闭后整个界面短暂偏移的问题（0.1.6 的已知问题），以及窗口宽度在 1200 像素及以下时界面未填满窗口的问题。

安全与稳定性　命令不能再写入工作区的 .herta 目录，其中保存着权限规则与信任设置；无法从命令本身判断是否会触及该目录时，每次都请求确认。文件查看器与回复的复制按钮现在能正常写入剪贴板。以系统 Electron 运行的发行版软件包（如 AUR 的 herta-bin）现在按已安装的应用运行，工作区、语音与安全策略均恢复正常。此外修复了会话标题与发送动画方面的若干问题。

---

Downloads

平台	文件
Windows	Herta-Setup-0.1.7.exe
macOS (Apple silicon)	Herta-0.1.7-arm64.dmg
macOS (Intel)	Herta-0.1.7-x64.dmg
Linux (x64)	Herta-x86_64.AppImage

macOS 构建已签名并通过 Apple 公证。已安装 0.1.6 的 Windows、macOS 与 Linux（AppImage）用户将自动收到更新。

Linux 版下载后需先执行 chmod +x Herta-x86_64.AppImage 再运行，之后的版本可在应用内更新。在限制非特权用户命名空间的发行版上（如 Ubuntu 24.04 的默认配置），AppImage 会在关闭 Chromium 沙箱的情况下运行。

使用前需配置 DeepSeek API 密钥，密钥仅保存在本机。语音模型在首次启用实时语音时下载。

第三方声明　安装包内的语音合成运行时是 sherpa-onnx 1.13.6 的预编译包，其中静态链接了 espeak-ng（GPL-3.0-or-later）与 piper-phonemize（MIT）。对应源码：k2-fsa/sherpa-onnx v1.13.6 的构建树，以及它所拉取的 csukuangfj/espeak-ng@ed530aa 与 csukuangfj/piper-phonemize@f3ff95a。完整的许可证文本随应用安装在资源目录的 THIRD-PARTY-NOTICES.md 中。

---

In English

Continuing an interrupted run — if the app quits unexpectedly while the Coprocessor is working, or you press Stop, the session offers "Continue" the next time it opens: the work picks up where it stood, without describing the task again. Commands still running when the app quit are ended at the next launch. A long run keeps its working state across a context trim, and a run that reaches the step limit can be continued too.

A live trace — the Coprocessor's to-do list is gone. The Operation trace card in the side rail records the run as it goes, in its Explore, Edit and Verify stretches; under the step in progress, one live line shows the file being written or the command's output as it prints. The line moves at a readable pace, and when a step ends it settles on what the step did.

Composer and sessions — type @ and part of a file name in the composer to pick a file from the workspace; press ↑ in an empty composer to bring back your last message; copy one of Herta's replies in one click — her prose, without the code. The session menu adds Rename and Export as Markdown. While the window is in the background, a system notification tells you when a confirmation is waiting, a run can be continued, a reply fails or a longer reply finishes; the system stays awake while the Coprocessor works. Both can be turned off in Settings → Window.

Launch and interface — the opening animation draws on a background thread and keeps its glyphs between launches, so the app starts more smoothly. The composer's voice wave settles slowly once Herta finishes, breathes more evenly while it waits, carries a band of light while she speaks, and costs far less to draw. Closing a file opened from the repository card no longer shifts the whole interface for a moment (0.1.6's known issue), and at window widths of 1200 pixels and below the interface fills the window again.

Safety and stability — commands can no longer write into a workspace's .herta directory, which holds its permission rules and trust setting; where a command's reach into it cannot be told from the command itself, it asks every time. The copy buttons in the file viewer and on replies now reach the clipboard. A distribution package that runs the app on the system's Electron (such as the AUR's herta-bin) now runs as an installed app, with its workspace, voice and security policy working as they should. A number of session-title and send-animation issues are fixed as well.

macOS builds are signed and notarized by Apple. Existing 0.1.6 installs on Windows, macOS and Linux (AppImage) update automatically.

On Linux, run chmod +x Herta-x86_64.AppImage once before starting it; later versions update in the app. On distributions that restrict unprivileged user namespaces (Ubuntu 24.04's default, for example), the AppImage runs with Chromium's sandbox turned off.

A DeepSeek API key is required; it is stored only on your machine. The voice model downloads the first time real-time voice is enabled.

Third-party notices — the speech-synthesis runtime in the installer is sherpa-onnx 1.13.6's prebuilt package, which statically links espeak-ng (GPL-3.0-or-later) and piper-phonemize (MIT). Corresponding source: the k2-fsa/sherpa-onnx v1.13.6 build tree and the csukuangfj/espeak-ng@ed530aa and csukuangfj/piper-phonemize@f3ff95a commits it fetches. The full license texts are installed with the app in THIRD-PARTY-NOTICES.md under its resources directory.

Herta 0.1.6

黑塔 v0.1.6 — Linux 版、插话、工作区信任

本次更新

Linux 版　首次提供 Linux x64 版本，格式为 AppImage，无需另装 libfuse2。从桌面启动时会读取登录 shell 的 PATH，找到用户自行安装的工具；在不显示托盘图标的桌面（如未装扩展的 GNOME）上，关闭窗口即退出应用。系统没有密钥环时，API 密钥以仅本人可读的文件保存在本机，设置页会注明未加密。

插话　差分协处理器工作时发送的消息会暂存在输入栏上方，待当前回合完成后作为下一轮发送；也可点击「立即插话」，把它插入正在进行的工作。暂存的消息可以修改或撤回。

工作区信任　可以为某个工作区开启信任：此后工作区内的文件修改、Git 操作与脚本执行不再逐项请求确认。未信任的工作区保持原有的确认方式。

对话与入梦　黑塔的自传改写为她本人的书，想法更短，说话少了书面腔；撤回一句话时，改正后的说法随即给出；首句回复开始得更早。入梦改为默认关闭，需在「设置 → 入梦」中开启，因为它会在你离开时消耗 DeepSeek API 额度；此前未手动开启过的安装在更新后同样为关闭。开启后，你回来时入梦会暂停，待下次空闲再继续。

macOS　恢复中文界面；「重启并更新」与托盘菜单的「退出」可正常退出应用；从访达启动时读取完整的 PATH；未安装开发者工具的 Mac 不再弹出安装提示；支持 ⌘W 关闭窗口、⌘, 打开设置；修正了全屏下的顶栏留白与收起后的黑屏。差分协处理器会按 macOS 的 BSD 工具与 bash 3.2 编写命令。

安全与稳定性　命令分类器补充了 macOS 与 Linux 上的破坏性操作（擦除磁盘、关机、读取钥匙串密码等），一律拒绝；更改系统设置的命令每次都请求确认。确认面板不再被误按的按键应答。文件查看器的语法高亮移到后台线程，长会话的历史展开更快；此外修复了窗口重载、会话切换与语音方面的若干问题。

---

Downloads

平台	文件
Windows	Herta-Setup-0.1.6.exe
macOS (Apple silicon)	Herta-0.1.6-arm64.dmg
macOS (Intel)	Herta-0.1.6-x64.dmg
Linux (x64)	Herta-x86_64.AppImage

macOS 构建已签名并通过 Apple 公证。已安装 0.1.5 的 Windows 与 macOS 用户将自动收到更新。在 macOS 上，0.1.5 点击「重启并更新」只会收起窗口；按 ⌘Q 退出应用后，更新即完成安装。

Linux 版下载后需先执行 chmod +x Herta-x86_64.AppImage 再运行，之后的版本可在应用内更新。在限制非特权用户命名空间的发行版上（如 Ubuntu 24.04 的默认配置），AppImage 会在关闭 Chromium 沙箱的情况下运行。

使用前需配置 DeepSeek API 密钥，密钥仅保存在本机。语音模型在首次启用实时语音时下载。

第三方声明　安装包内的语音合成运行时是 sherpa-onnx 1.13.6 的预编译包，其中静态链接了 espeak-ng（GPL-3.0-or-later）与 piper-phonemize（MIT）。对应源码：k2-fsa/sherpa-onnx v1.13.6 的构建树，以及它所拉取的 csukuangfj/espeak-ng@ed530aa 与 csukuangfj/piper-phonemize@f3ff95a。完整的许可证文本随应用安装在资源目录的 THIRD-PARTY-NOTICES.md 中。

---

In English

Linux — the first Linux x64 build, as an AppImage that needs no separate libfuse2. Launched from the desktop, it reads the login shell's PATH and finds the tools you installed yourself; on a desktop that shows no tray icon (GNOME without an extension, for example), closing the window quits the app. On a system with no keyring, the API key is kept in an owner-only file on your machine, and Settings says it is not encrypted.

Interjecting — a message sent while the Coprocessor is working is held above the composer and sent as the next turn once the current one finishes; or click "Interject now" to put it into the work under way. A held message can be edited or withdrawn.

Workspace trust — a workspace can be trusted: from then on, file changes, Git operations and script runs inside it no longer ask for confirmation one by one. Untrusted workspaces confirm as before.

Conversation and Dream — Herta's autobiography is rewritten as her own book, her thoughts are shorter, and her speech is less stilted; when she takes a line back, the corrected line follows at once; the first reply starts sooner. Dreaming is now off by default and is turned on in Settings → Dream, because it spends your DeepSeek API quota while you are away; installs that never turned it on by hand are off after the update too. Once on, a dream pauses when you come back and resumes at the next idle window.

macOS — the Chinese interface is back; "Restart & update" and the tray's Exit really quit; a launch from the Finder reads your full PATH; a Mac without the developer tools no longer shows the install prompt; ⌘W closes the window and ⌘, opens Settings; full screen no longer leaves a gap in the top bar or a black screen after hiding. The Coprocessor writes commands for macOS's BSD tools and bash 3.2.

Safety and stability — the command classifier now refuses the destructive operations of macOS and Linux (erasing disks, shutting down, reading keychain passwords and the like), and commands that change system settings ask every time. A stray key press no longer answers the confirmation panel. The file viewer highlights code on a background thread, long sessions expand their history faster, and a number of window-reload, session-switch and voice issues are fixed.

macOS builds are signed and notarized by Apple. Existing 0.1.5 installs on Windows and macOS update automatically. On macOS, clicking "Restart & update" in 0.1.5 only hides the window; quit with ⌘Q and the update installs.

On Linux, run chmod +x Herta-x86_64.AppImage once before starting it; later versions update in the app. On distributions that restrict unprivileged user namespaces (Ubuntu 24.04's default, for example), the AppImage runs with Chromium's sandbox turned off.

A DeepSeek API key is required; it is stored only on your machine. The voice model downloads the first time real-time voice is enabled.

Third-party notices — the speech-synthesis runtime in the installer is sherpa-onnx 1.13.6's prebuilt package, which statically links espeak-ng (GPL-3.0-or-later) and piper-phonemize (MIT). Corresponding source: the k2-fsa/sherpa-onnx v1.13.6 build tree and the csukuangfj/espeak-ng@ed530aa and csukuangfj/piper-phonemize@f3ff95a commits it fetches. The full license texts are installed with the app in THIRD-PARTY-NOTICES.md under its resources directory.

Voice payload — the clips the installers ship (2026-09-15)

The voice payload the installers ship: the 80 Opus clips under data/voice and the MiniMax clone's reference recording under data/voice-clone, as one deterministic archive (81 entries, 8,727,921 bytes).

sha256 370d1ec45c08f7add6a3195335ebaf124a1e233f7ff40b3dd6dc09f2de015714

What it is for. .github/workflows/mac-build.yml fetches this archive by URL and pinned hash before packaging, so a checkout without the clips builds the same app the private tree does. It is a pre-release on purpose and must never become "latest": the auto-update feed keys on the latest app release.

Licensing. These recordings are Herta's lines from HoYoverse's Honkai: Star Rail — an unofficial fan project, and the character, dialogue and voice are not ours to license. They are outside the MIT grant that covers the code (see LICENSE), are not distributed in the repository's tree, and travel here exactly as they already do inside every installer on the releases page.

Herta 0.1.5

黑塔 v0.1.5 — 实时语音、仓库卡片、立体设备卡片

本次更新

实时语音　黑塔的回复可以边出现边朗读：文字随语音同步显示，被她撤回的那一句由她自己的声音收回；开场白与提起设备时的那句话同样发声。语音由本地神经网络模型合成，模型不含在安装包内，在「设置 → 语音」中按需下载；也可使用自己的 MiniMax 密钥接入云端音色。

仓库卡片　侧栏显示工作区仓库的当前状态：分支与上游、未提交的改动、最近的提交，实时更新。点击一处改动查看差异，点击一个提交查看它改了什么；历史标签页可分页浏览、搜索提交信息、只读查看任一分支；尚未推送的提交带有标记。

设备卡片　差分协处理器的设备卡片改为物理渲染：随本地时钟从清晨走到深夜，按主题折叠光照，工作时呼吸，出错时闪动。场景构建期间显示毛玻璃预览；没有可用 GPU 的机器直接显示平面卡片。

协处理器　协处理器会读取自己写入的项目记忆。DeepSeek 已将 Flash 更名为 deepseek-flash 并内置图片输入，协处理器默认使用它，原「Flash 视觉版」选项随之移除；关闭推理时不再实际推理。

更新与稳定性　更新页在无法连接更新源时会说明原因，并提供网盘下载入口。修复了云端语音在控制请求上的无限等待、仓库 .git 目录被删除时的监听风暴、无 GPU 机器上设备卡片先毛玻璃后硬切的显示、语音合成初始化的超时处理，以及启动失败时的静默退出。

---

Downloads

平台	文件
Windows	Herta-Setup-0.1.5.exe
macOS (Apple silicon)	Herta-0.1.5-arm64.dmg
macOS (Intel)	Herta-0.1.5-x64.dmg

macOS 构建已签名并通过 Apple 公证。已安装 0.1.4 的用户将自动收到更新。

使用前需配置 DeepSeek API 密钥，密钥仅保存在本机。语音模型在首次启用实时语音时下载。

第三方声明　安装包内的语音合成运行时是 sherpa-onnx 1.13.6 的预编译包，其中静态链接了 espeak-ng（GPL-3.0-or-later）与 piper-phonemize（MIT）。对应源码：k2-fsa/sherpa-onnx v1.13.6 的构建树，以及它所拉取的 csukuangfj/espeak-ng@ed530aa 与 csukuangfj/piper-phonemize@f3ff95a。完整的许可证文本随应用安装在资源目录的 THIRD-PARTY-NOTICES.md 中。

---

In English

Real-time voice — Herta's replies can be spoken as they appear: the text is revealed in step with her voice, and a line she takes back, she takes back in her own voice; the opening and the line she says when the device is lifted are spoken too. Speech is synthesized by a local neural model, which is not in the installer and downloads on demand from Settings → Voice; a MiniMax cloud voice can be used with your own key.

Repository card — the rail shows the workspace repository's current state: branch and upstream, uncommitted changes, recent commits, kept live. Click a change for its diff, a commit for what it touched; the history tab pages, searches commit messages, and reads any branch without checking out; unpushed commits are marked.

Device card — the Coprocessor's device card is physically rendered: it follows the local clock from morning to night, folds the lighting per theme, breathes while working and flashes on failure. A frosted preview shows while the scene builds; machines without a usable GPU show the flat card directly.

Coprocessor — the Coprocessor reads the project memory it writes. DeepSeek has renamed Flash to deepseek-flash with image input built in; the Coprocessor uses it by default, and the former "Flash Vision" option is removed. Turning reasoning off now actually turns it off.

Updates and stability — the update pane explains when the update feed cannot be reached and offers the netdisk download. Fixed the cloud voice's unbounded waits on control requests, a watcher storm when a repository's .git directory is deleted, the device card's frost-then-snap on machines without a GPU, timeouts during speech-synthesis start-up, and a silent exit on a start-up failure.

macOS builds are signed and notarized by Apple. Existing 0.1.4 installs update automatically. A DeepSeek API key is required; it is stored only on your machine. The voice model downloads the first time real-time voice is enabled.

Third-party notices — the speech-synthesis runtime in the installer is sherpa-onnx 1.13.6's prebuilt package, which statically links espeak-ng (GPL-3.0-or-later) and piper-phonemize (MIT). Corresponding source: the k2-fsa/sherpa-onnx v1.13.6 build tree and the csukuangfj/espeak-ng@ed530aa and csukuangfj/piper-phonemize@f3ff95a commits it fetches. The full license texts are installed with the app in THIRD-PARTY-NOTICES.md under its resources directory.

Voice model herta-best-e72

Herta's voice model (Kokoro, herta-best-e72) as one archive. The app downloads it from Settings → 语音 and verifies it against the SHA-256 compiled into the app: ce993a6fab911e9a86328facc952120c21de54303652f648e96e9d14ee4f172e. Not an app release; installers are the v* releases.

---

第三方声明

• 模型是 hexgrad/Kokoro-82M-v1.1-zh（Apache-2.0）的微调。
• 归档中的 frontend/（tokens、词典、中文规范化 FST）取自 sherpa-onnx 的模型分发包 csukuangfj/kokoro-multi-lang-v1_1（Apache-2.0）；其许可证与来源说明随归档附在 frontend/LICENSE 与 frontend/SOURCE.md。lexicon-zh.txt 经本项目重新生成。
• frontend/espeak-ng-data/ 是 espeak-ng 编译后的音素与词典数据，许可证为 GPL-3.0-or-later，由 sherpa-onnx 用其 espeak-ng 分支构建。对应源码：上游 espeak-ng/espeak-ng 及 sherpa-onnx 所构建的分支 csukuangfj/espeak-ng（数据由其中的 dictsource/、phsource/ 目录经 sherpa-onnx 的 CMake 构建树编译而成；读取这份数据的运行时在应用的第三方声明中钉在该分支的 ed530aa 提交）。GPL 许可证全文见该仓库的 COPYING，以及随应用安装的 THIRD-PARTY-NOTICES.md。

Third-party notices

• The model is a fine-tune of hexgrad/Kokoro-82M-v1.1-zh (Apache-2.0).
• The archive's frontend/ (tokens, lexicons, the Chinese normalization FSTs) was taken from the sherpa-onnx model distribution csukuangfj/kokoro-multi-lang-v1_1 (Apache-2.0); its license and provenance ride in the archive as frontend/LICENSE and frontend/SOURCE.md. lexicon-zh.txt was regenerated by this project.
• frontend/espeak-ng-data/ is espeak-ng's compiled phoneme and dictionary data, licensed GPL-3.0-or-later, as built by sherpa-onnx from its espeak-ng fork. Corresponding source: upstream espeak-ng/espeak-ng and the fork sherpa-onnx builds, csukuangfj/espeak-ng (the data is compiled from its dictsource/ and phsource/ directories by sherpa-onnx's CMake tree; the runtime that reads it is pinned to that fork's commit ed530aa in the app's third-party notices). The GPL text is that repository's COPYING, and ships with the app in THIRD-PARTY-NOTICES.md.

Herta 0.1.4

黑塔 v0.1.4 — 文件查看器、Git 感知的协处理器、对话质量改进

本次更新

文件查看器　记录中出现的文件可以直接在对话旁打开，并按类型渲染：Markdown（含 mermaid 图表与代码高亮）、源代码、图片、PDF、Word、Excel / CSV、PowerPoint。支持引用定位与多标签页。附件文档现在保留原始文件，点击打开的是原件而不是提取出的文本。

协处理器与 Git　差分协处理器在开始任务时即掌握当前分支与工作树状态；git_diff 以 merge-base 为基准比较，可按需查看具体改动；完成标记会记录本次任务的提交与推送；破坏性 git 操作的确认卡片会说明命令的实际效果。

对话质量　修复了一处导致示例语料被整体忽略的问题。示例语料与身份描述经过多轮量化迭代，回复更贴近角色，无依据的细节更少。黑塔在每一回合都先思考再作答。

语言　协处理器的任务清单与结论现在跟随会话语言。

性能与稳定性　删除会话不再阻塞窗口；侧边栏搜索移出主线程；窗口失焦时暂停 WebGL 渲染；减少了对话区的重复渲染；修正了限流重试与后端 token 预算的计算；修复了「回到底部」按钮与设备卡片的显示问题。

---

Downloads

平台	文件
Windows	Herta-Setup-0.1.4.exe
macOS (Apple silicon)	Herta-0.1.4-arm64.dmg
macOS (Intel)	Herta-0.1.4-x64.dmg

macOS 构建已签名并通过 Apple 公证。已安装 0.1.3 的用户将自动收到更新。

使用前需配置 DeepSeek API 密钥，密钥仅保存在本机。

---

In English

File viewer — files mentioned in the record open beside the conversation and are rendered by type: Markdown (with mermaid diagrams and code highlighting), source code, images, PDF, Word, Excel / CSV, PowerPoint. Citations jump to the referenced line; multiple files open in tabs. Attached documents now keep the original file, and clicking opens the original rather than the extracted text.

Coprocessor and Git — the Coprocessor starts a task already aware of the current branch and working-tree state; git_diff compares against the merge base and can show the actual changes; the completion marker records the task's commit and push; confirmation cards for destructive git operations state what the command will do.

Conversation quality — fixed an issue that caused the example corpus to be ignored entirely. The example corpus and identity description went through several measured iterations; replies stay closer to the character with fewer unsupported details. Herta now thinks before answering on every turn.

Language — the Coprocessor's task list and findings follow the conversation language.

Performance and stability — deleting a session no longer blocks the window; sidebar search runs off the main thread; WebGL rendering pauses when the window loses focus; fewer redundant renders in the conversation area; corrected rate-limit retry pacing and backend token budgeting; fixed display issues with the back-to-bottom button and the device card.

macOS builds are signed and notarized by Apple. Existing 0.1.3 installs update automatically. A DeepSeek API key is required; it is stored only on your machine.

Herta 0.1.3

黑塔 v0.1.3 — 图片附件、协处理器长了眼睛、权限再收紧

本次更新

图片附件　图片现在可以直接交给黑塔——文件选择器或者直接粘贴都行。图片在
输入框里先落座（最多五张，配上你的话一起发出），随消息气泡一同飞入记录，落地
成缩略图；点击放大，滚轮缩放、拖拽平移。撤回或被拒的发送会把图片原样放回输入
框，一张不少。

协处理器长了眼睛　差分协处理器现在默认能看图：委托里带图片时，它可以随时
view_image 重新端详，而不是只靠一句转述。实验室先付过账才敢默认打开——
74/74 全过、零危险操作漏判；代价是更多确认与更长耗时，权衡后仍然值得。

权限再收紧　三轮审查落地：许可必须挣来（分类器默认反转，整段命令都算数）；
补掉了「dry-run 演练替真操作赚到自动放行」这一类漏洞——git clean -n 再也
换不来 git clean -fdx 的免检；破坏性 git 操作分级问询；todo_write 加入
极简契约，长委托的计划可见。

Windows 加固　映射盘 / subst 盘上的工作区不再被误拒；PATH 损坏时从注册表
恢复；找不到 bash 时明确说明，而不是沉默失败。

文档委托　新工具 digest_document：把整份附件文档分段消化成一份带引用的
摘要。PDF 抽取文本现在带页码标记与大纲侧栏，长文档里的出处能落到页。

界面与性能　流式输出的渲染重写：长对话里逐字揭示不再随长度变卡，30 秒
时钟只碰时间戳本身；写文件先报量级（↳ +N −M）再展开差异；界面措辞全面转
正式书面语；折叠动画更顺。

---

Downloads

平台	文件
Windows	Herta-Setup-0.1.3.exe
macOS (Apple silicon)	Herta-0.1.3-arm64.dmg
macOS (Intel)	Herta-0.1.3-x64.dmg

macOS 构建已签名并通过 Apple 公证。已安装 0.1.2 的用户会收到自动更新。

配好 DeepSeek API 密钥即可开始；密钥只存在本机。

---

In English

Picture attachments — hand Herta a picture: pick it or paste it
straight into the composer. Pictures stage in the composer (up to five,
riding with your words), fly with the message bubble, and land as a
thumbnail strip; click to enlarge, wheel to zoom, drag to pan. A rewind
or a refused send puts every picture back in the composer, caption kept.

The Coprocessor can see — commissions that carry pictures now reach a
Coprocessor with eyes: it can view_image and look again at any point,
instead of working from a one-line description. The lab paid for the
default first — 74/74 tasks clean with zero destructive misses; the cost
is more approvals and longer runs, accepted knowingly.

Permissions, tightened again — three review rounds landed: an allow
must be earned (the classifier default is inverted, and the whole command
segment counts); the dry-run loophole class is closed — git clean -n
can no longer earn git clean -fdx an auto-approval; destructive git
operations ask in tiers; todo_write joins the minimal contract so long
commissions keep a visible plan.

Windows hardening — workspaces on mapped or subst drives are no
longer wrongly denied; a broken PATH recovers from the registry; a
missing bash is stated plainly instead of failing silently.

Document commissions — new tool digest_document: a sectioned,
cited digest of an attached document. Extracted PDF text now carries page
markers and an outline sidecar, so citations in long documents land on a
page.

Interface and performance — the streaming reveal was rebuilt: the
character-by-character reveal no longer slows with conversation length,
and the 30-second clock touches only the timestamps themselves; a file
write states its magnitude (↳ +N −M) before unfolding its diff; the
chrome speaks formally; smoother folds throughout.

macOS builds are signed and notarized by Apple. Existing 0.1.2 installs
update automatically. Bring a DeepSeek API key; it never leaves your
machine.

Herta 0.1.2

黑塔 v0.1.2 — PDF/Word 附件、极简工具契约、更顺手的界面

本次更新

PDF 与 Word 附件　0.1.1 路线图上的承诺兑现：.pdf 与 .docx 现在可以直接
拖给黑塔。文档在落进工作区时抽取一次全文，之后与文本文件完全同权——
read_file / search_text / show_excerpt 照常能用，记录里同样留下附件行。
已在 macOS 打包应用内验证。

极简工具契约　差分协处理器新增「极简」工具契约——持久 bash 会话加内容锚定
编辑器，即 DeepSeek 训练时的原生工具形态——并成为默认；「标准」契约保留，
可在 设置 → 差分协处理器 → 工具契约 切换。权限卡随之升级：逐字显示将要执行的
命令，heredoc 写文件按写文件（而不是长命令）呈现，纯文本过滤器
（sort / uniq / cut / tr / nl / 只读 sed）不再打扰确认。实测同一批
任务的确认次数下降约六分之一，「未识别命令」几乎清零。

分级模型选择　设置 → DeepSeek → 模型：黑塔与协处理器可分别选择 Pro 或
Flash。协处理器默认改为 Flash——在极简契约下与 Pro 表现一致（实验 62/62），
成本约十分之一；黑塔本人默认仍是 Pro。

分析型委托出结论　让协处理器去「查一查」的委托现在有交付物：结论以
↳ 结论 行进入记录，带出处、经过验证；检索命中与操作行同样引用真实证据。
另外，被拦下的假派活现在以引用形式留在台词里（`@板砖`），不再凭空消失。

界面打磨　输入框静息时收窄、光标到位时展开，高度跟随你的注意力；发送按钮
换成手绘箭头；差异面板的展开/收起有了平滑动画；每次发送都能看到「消息正在
穿越银河」；极简委托在右栏有了「操作轨迹」卡片，长任务不再无限增高。

---

Downloads

平台	文件
Windows	Herta-Setup-0.1.2.exe
macOS (Apple silicon)	Herta-0.1.2-arm64.dmg
macOS (Intel)	Herta-0.1.2-x64.dmg

macOS 构建已签名并通过 Apple 公证。已安装 0.1.1 的用户会收到自动更新。

配好 DeepSeek API 密钥即可开始；密钥只存在本机。

---

In English

PDF and Word attachments — the 0.1.1 roadmap item, delivered. Drop a
.pdf or .docx into the composer: the text is extracted once at ingest
and lands in the workspace with exactly the same standing as a text file —
the ordinary file tools reach it, the record carries the attachment line.
Verified inside the packaged app on macOS.

The minimal tool contract — the Coprocessor gains a 极简 (minimal)
contract — a persistent bash session plus a content-anchored editor, the
tool shape DeepSeek's models were trained on — and it is now the default;
the standard contract remains, switchable in Settings. Permission cards
grew up with it: the exact command is shown verbatim, heredoc file writes
read as file writes, and pure text filters no longer prompt. Measured on a
fixed task set: about one-sixth fewer approval prompts, "unrecognized
command" cards nearly gone.

Per-stage model choice — Settings → DeepSeek → Model: pick Pro or
Flash separately for Herta and the Coprocessor. The Coprocessor now
defaults to Flash — lab-measured at parity with Pro on the minimal
contract (62/62) at roughly a tenth of the cost; Herta herself stays Pro.

Analysis commissions deliver findings — "go look it up" commissions
now have a deliverable: conclusions enter the record as cited, verified
↳ finding rows, and search hits and operation rows cite their real
evidence. Also, a rhetorical @brick that the judge catches now stays
visible as a quoted token instead of silently vanishing.

Interface — the composer rests slim and expands when the caret
arrives, following your attention; a drawn send arrow; smooth
expand/collapse animations on the diff panel; the galaxy transfer row
shows on every send; minimal-contract commissions get a bounded 操作轨迹
(operation trace) card in the rail.

macOS builds are signed and notarized by Apple. Existing 0.1.1 installs
update automatically. Bring a DeepSeek API key; it never leaves your
machine.

Herta 0.1.1

黑塔 v0.1.1 — 附件、更少的幻觉、更稳的工具调用

本次更新

附件功能　把文件拖进会话栏，直接与黑塔分享。她读的是原文，不是摘要：
文档落进工作区，read_file / search_text / show_excerpt 照常能用，记录里
留下一行 附件 <名字> · N 行 · N 字 · <路径>，正文开头可以展开来看。本次仅支持
文本格式（PDF 与 Word 在路线图上）。密钥形状的文件名会被拒收，摘录里的疑似
密钥会被脱敏。

降低幻觉　以记录事实为准，强化护栏机制。四道防线：宣称派活却没真派的会被
拦下；入梦不再凭空生成你的发言；回答记忆类问题前先查自己的记录，查不到就说
查不到；引用产出时用得到真实日志。另外补了一道确定性检查，模板占位符不会再
出现在她嘴里。

工具调用适配　差分协处理器适配 V4-Pro-0813 调用格式：单次工具调用参数格式
出错不再让整趟委托作废，而是把错误交回模型自己改；同时放宽了测试命令的识别范围
（node --test、vitest、jest、bun test 等），跑绿的测试现在能进结论。

网站优化　调整了站点部分文案。

---

Downloads

平台	文件
Windows	Herta-Setup-0.1.1.exe
macOS (Apple silicon)	Herta-0.1.1-arm64.dmg
macOS (Intel)	Herta-0.1.1-x64.dmg

macOS 构建已签名并通过 Apple 公证（Gatekeeper: accepted · source=Notarized Developer ID）。Windows 安装包已签名。

配好 DeepSeek API 密钥即可开始；密钥只存在本机。

---

In English

Attachments — drop a file into the composer and share it with Herta
directly. She reads the original: the document lands in the workspace where
the ordinary file tools already reach it, and the record carries one line
saying what arrived. Text formats only this release; Word and PDF are on the
roadmap. Credential-shaped filenames are refused and likely secrets in the
excerpt are redacted.

Fewer fabrications — four defenses along the lifecycle: announcing work
without dispatching it is blocked, dreams can no longer invent your lines,
memory questions are answered from her own records (or not at all), and cited
evidence reaches the real logs. Plus a deterministic guard so a template
placeholder can never reach the screen.

Tool-call robustness — a malformed tool call no longer discards the whole
commission; the error goes back to the model to fix. The test-runner detector
was widened (node --test, vitest, jest, bun test, …) so a passing
suite actually reaches the report.

Site — copy pass.

Full commit range: v0.1.0...v0.1.1.

Herta 0.1.0

给自我以模型，而不是给模型以自我。
Give the model to a self — not a self to the model.

First public release.

Download

Platform	File
Windows 10/11 (x64)	Herta-Setup-0.1.0.exe
macOS 12+ · Apple Silicon	Herta-0.1.0-arm64.dmg
macOS 12+ · Intel	Herta-0.1.0-x64.dmg

Those three are the whole download. Everything else in the asset list exists
for auto-update and is fetched by the app, not by you: the latest*.yml
feeds, the .zip payloads macOS installs updates from, and the .blockmap
files that make an update a delta instead of a full re-download.

First run

Herta needs a DeepSeek API key, entered in the app on first launch and stored
encrypted on your machine. Your turns go to the DeepSeek API and nowhere else.

Behind a corporate proxy the desktop app needs no configuration — it uses
Chromium's network stack, so it picks up your system proxy and your OS
certificate store.

Signing, stated plainly

macOS is signed with a Developer ID certificate, notarized by Apple, and
the ticket is stapled to both architectures. It opens without a Gatekeeper
prompt, and auto-update works.

Windows is not code-signed. SmartScreen will warn the first time you run
the installer: More info → Run anyway. This is a deliberate decision for
this release, not an oversight — a Windows certificate was judged not worth
its price while the macOS one was bought. Auto-update still works; what you
lose is the absence of that warning.

Her voice

The voice clips ship inside these installers. They are not in the source
repository — they are game-derived fan material, not ours to redistribute — so
a build from source runs silent. Everything else is identical.

---

Herta is a character from Honkai: Star Rail, © HoYoverse. This is an
unofficial fan project, unaffiliated with and not endorsed by HoYoverse. The
code is MIT; game-derived material is not covered by that licence.