v1.8.2

##v1.8.2

• 重大更新，新增第 13 种泰案画布「命理画布」，四门术数在同一画布内起课、推演、断课与落档。
  • 六爻：京房纳甲、八宫世应、六亲六神、动爻变卦、互错综、伏神；月建日辰旺衰、卦级六冲六合、三合局、归魂游魂、用神取舍与应期。
  • 大六壬：天地盘、四课、九宗门、三传、课体与神煞，盘面关系网；提供 2D 经典神阁盘与 3D 浑天栻盘仪两种形态，舞台内可随时切换。
  • 紫微斗数：十二宫安星、十四主星、四化、大限与流年，另有 3D 仰观星穹。
  • 四柱八字：四柱藏干十神、纳音、十二长生、大运顺逆与起运；含全盘关系图谱（刑冲合害破）、流年推演、大运画像、从格专旺与调候用神。
  • 亲手起卦演法：大衍筮法 3D 案上蓍草推演（五十蓍十八变，支持键盘操作、自动推演与进度提示），火珠林 3D 案上掷钱起卦。
  • 聊天入口全流程编排：说「算个卦」「排个八字」即可起课，泰深负责打开画布、代填参数、推进演法；盘面由画布引擎确定性生成，泰深只做断语。
  • 每门流派统一走「录入 → 仪轨 → 六步推演 → 断课」，断语与盘面双视图，落档导出 A4 竖版 PDF，卷末附免责声明。
  • 三档主题（玄天 / 青碧 / 素壁），默认青碧，选择本地持久化。
• 新增 DeepSeek Search / Fetch 按次计费登记与消费统计。
  • 预算面板新增「按次调用」细分行，设置页可按次编辑单价表。
• 新增文件预览的原生打印，PDF / HTML / Markdown 均可唤起系统打印对话框。
• 新增代码变更预览与 HTML 预览的关键词搜索，支持高亮与上下跳，Ctrl+F 唤出。
• 优化了泰案画布的启动速度：画布与面板改为按需加载，首屏模块从 234 降到 29。
• 优化了画布窗口的默认尺寸：改为 1600×900，并按屏幕工作区 88% 自适应居中。
• 优化了设计画布元素库，新增 8 个范式元素。
• 修复了设置页侧边栏版本号固定显示 v1.7 的问题。
• 修复了文本画布重启恢复后停在空屏的问题。
• 修复了对话中发送引导时长时间无响应的问题。
• 修复了内置浏览器与画布窗口截图空白的问题。
• 修复了网页获取失败时仍返回正文的问题。

##v1.8.2 (English)

• Major update: added the 13th Tai An canvas, the "Metaphysics canvas", bringing four divination systems into one canvas for casting, deriving, reading and archiving.
  • Six Lines (Liu Yao): Jing Fang na-jia, eight-palace generation and response, six relations and six spirits, moving lines and changed hexagram, mutual and opposite hexagrams, hidden spirits; strength by month and day, hexagram-level six clashes and six harmonies, three-harmony formations, returning-soul and wandering-soul hexagrams, useful spirit selection and timing.
  • Da Liu Ren: heaven and earth plates, four lessons, nine gates, three transmissions, lesson bodies and deities, plus relations on the board; provides a 2D classic shen-ge board and a 3D armillary board, switchable inside the stage at any time.
  • Zi Wei Dou Shu: twelve palaces with star placement, fourteen major stars, four transformations, major limits and annual cycles, plus a 3D star dome view.
  • Four Pillars (Ba Zi): four pillars with hidden stems and ten gods, na-yin, twelve life stages, forward and reverse major cycles with starting age; includes a full-chart relation map (clash, punishment, harmony, harm, break), annual-cycle derivation, major-cycle profile, dominant-pattern charts and climate-balancing useful gods.
  • Hands-on casting: the Great Evolution yarrow-stalk method rendered in 3D on a desk (fifty stalks, eighteen changes, with keyboard control, automatic derivation and progress prompts), and the Huo Zhu Lin coin method cast in 3D on a desk.
  • Full workflow from the chat entry: say "cast a hexagram" or "read my Ba Zi" to start; Taishen opens the canvas, fills in the parameters and advances the derivation, while the board is generated deterministically by the canvas engine and Taishen only provides the reading.
  • Every system follows the same steps "enter data to ritual to six-step derivation to reading", with dual views for the reading and the board, export to A4 portrait PDF, and a disclaimer at the end of the reading.
  • Three themes (Xuantian, Qingbi, Subi) with Qingbi as the default; the choice is persisted locally.
• Added per-call billing registration and cost tracking for DeepSeek Search / Fetch.
  • The budget panel gains a "per-call" breakdown line, and the unit price table is editable in Settings.
• Added native printing in the file preview; PDF, HTML and Markdown can all invoke the system print dialog.
• Added keyword search to the code-change preview and HTML preview, with highlighting, jump to the previous or next match and Ctrl+F.
• Improved Tai An canvas startup speed: canvases and panels now load on demand, cutting first-screen modules from 234 to 29.
• Improved the default canvas window size: now 1600x900, centered and scaled to 88% of the screen work area.
• Improved the Design canvas element library with 8 new paradigm elements.
• Fixed the Settings sidebar version label being stuck at v1.7.
• Fixed text canvases being stuck on a blank screen after a restore on restart.
• Fixed long unresponsiveness when sending guidance during a conversation.
• Fixed blank screenshots of the built-in browser and canvas windows.
• Fixed web fetch returning body content for failed requests.

---

文件校验（SHA256）

文件	SHA256
taishen_setup_1.8.2.exe	AD69030F04D2F48175AF89ECDCAA25C073A9430C73A522B96E464C1C98DE87F2
taishen_1.8.2.zip	4B4B03A9551E87AFA7C14D9EC7DA9658CC9757A2F2A888C27221DC1E69BC83BA
taishen_1.8.2_macOS_arm64.dmg	950525ADB79A9B9E0982B255E581A69ABBEA088ACD6D407DB60E53E682260C3C

---

v1.8.1

##v1.8.1

• 重大更新，新增内置浏览器受限站点登录态导入与「认证中枢」。
  • 更新前：Google、Microsoft 等网站判定内置浏览器不安全，直接拒绝登录，页面无法使用。
  • 更新后：受限站点地址栏常驻「导入 Cookie 登录」按钮，点击即呼叫泰深协助导入；剪贴板已有 Cookie 时会自动附带。泰深也可从本机 Chrome 直接提取登录态完成导入，无需多轮对话。
  • 导入的登录态经系统加密存入凭据保险箱，再次访问同站点自动恢复；失效时自动检测并呼叫泰深重新导入。
  • 设置页新增「认证中枢」Tab，可查看已导入站点台账，支持只清除记录或连同 Cookie 一并注销。
  • 跨浏览器复制登录态会触发 Google 的会话复制检测，导入时已自动过滤对应签名字段；本机 Chrome 存在多个配置且默认配置被占用时，会请求用户确认，不再静默改用其他账号的登录态。
• 新增「界面动效与物理质感」设置。
  • 弹窗（洞察卡片、提问弹窗）加入物理阻尼过冲与 3D 微视差；待办清单完成态加入路径勾勒与辉光；会话列表加入悬停平移与点击下陷；首页背景加入随鼠标移动的点阵微光。
  • 开关位置：设置 - 通用设置 -「界面动效与物理质感」，三档可选「完整物理质感 / 柔和减弱 / 完全关闭」，默认完整物理质感；系统开启「减弱动态效果」时自动降档。
• 新增生图参考图支持。
  • 生图可附带 1-5 张参考图，用于锁定人物形象与场景风格；超出 5 张会被拒绝并在提示中说明原因。
  • 用法：让泰深生图时提供参考图路径，或由「生图专员」按配置通道生成。
• 升级了 AI 视频前期筹备能力。
  • 分镜阶段按景别自动挑选参考图（特写 / 中近景 / 双人 / 道具 / 群像），并强制 5 张配额上限。
  • 生成提示词自动消歧，避免多角色参考图互相串用。
  • 支持导出 Visual DNA 交付包（含清单与 SHA-256 校验），可交付后续视频生成环节。
• 新增用户工具安全防护。
  • 创建用户工具时若使用同步子进程调用（execSync / execFileSync / spawnSync），创建会被拦截并给出异步改法，可强制跳过。
  • 启动时体检已有用户工具，命中同一份内容只提示一次。
• 优化了记忆搜索对中文查询的召回。
  • 更新前：中文自然语句查询（如「为什么我提交代码把别人的改动带进去了」）按整串匹配，基本搜不到；短关键词也常漏召回。
  • 更新后：中文查询补充降权切分兜底，并单独提取查询中的英文与数字词。实测 40 条真实查询首位命中从 17 条提升到 26 条，自然语句从 0/20 提升到 7/20。
  • 另为超长查询增加词元上界，避免长文本查询变慢。
• 优化了设计画布主题面板的 Token 引用展示，去掉多余的前缀符号。
• 修复了部分功能执行时界面假死的问题。
  • 更新前：Visio 绘图、性能监控、隧道管理、MCP、语音等功能的同步调用会锁死主进程，实测最长 400 秒界面无响应，且 30 秒超时兜底同时失效。
  • 更新后：上述调用全部改为异步，长任务执行期间界面保持可操作。

##v1.8.1 (English)

• Major update: added sign-in state import for restricted sites in the built-in browser, plus an "Auth Hub".
  • Before: Sites such as Google and Microsoft flagged the built-in browser as insecure and refused sign-in, leaving the page unusable.
  • Now: A persistent "Import Cookies to sign in" button appears in the address bar on restricted sites; clicking it asks Taishen to import the sign-in state, and cookies already on the clipboard are attached automatically. Taishen can also extract the sign-in state directly from your local Chrome, without extra conversation.
  • Imported sessions are encrypted and stored in the credential vault, and are restored automatically on the next visit to the same site; when they expire, verification runs automatically and Taishen is called to re-import.
  • A new "Auth Hub" tab in Settings lists imported sites, and supports clearing only the record or logging out along with the cookies.
  • Copying sessions across browsers triggers Google's session-copy detection; the related signature field is filtered out automatically on import. When multiple Chrome profiles exist and the default profile is in use, Taishen asks for confirmation instead of silently switching to another account's session.
• Added "UI Motion & Dynamics" settings.
  • Popups (insight cards, question dialogs) gain spring damping overshoot and subtle 3D parallax; the todo list gains stroke drawing and a glow on completion; the session list gains hover shift and press-down; the home background gains a dot-grid glow that follows the mouse.
  • Location: Settings - General - "UI Motion & Dynamics", with three levels "Full physics / Reduced / Off", defaulting to Full physics; it steps down automatically when the system enables "reduce motion".
• Added reference image support for image generation.
  • Generation accepts 1-5 reference images to lock a character's look and scene style; more than 5 is rejected with the reason stated.
  • Usage: provide reference image paths when asking Taishen to generate, or let the image specialist generate through its configured channel.
• Upgraded AI video pre-production.
  • The storyboard stage selects reference images automatically by shot type (close-up / medium close-up / two-shot / prop / crowd), enforcing a hard limit of 5.
  • Prompts are disambiguated automatically so that reference images of multiple characters are not mixed up.
  • Supports exporting a Visual DNA delivery package (with manifest and SHA-256 verification) for downstream video generation.
• Added safety checks for user tools.
  • Creating a user tool that uses synchronous subprocess calls (execSync / execFileSync / spawnSync) is blocked with an async alternative, and can be forced through.
  • Existing user tools are checked at startup, and each unique finding is reported only once.
• Improved memory search recall for Chinese queries.
  • Before: Chinese natural-language queries (for example, "why did my commit include someone else's changes") were matched as a whole string and almost never returned results; short keywords were frequently missed as well.
  • Now: Chinese queries fall back to weighted term splitting and separately extract English and numeric words. In a test of 40 real queries, top-1 hits rose from 17 to 26, and natural-language queries from 0/20 to 7/20.
  • An upper bound on query terms was added to keep very long queries fast.
• Improved the Token reference display in the Design canvas theme panel by removing redundant prefix symbols.
• Fixed UI freezes while some features were running.
  • Before: Synchronous calls in Visio drawing, performance monitoring, tunnel management, MCP and voice features locked the main process, leaving the UI unresponsive for up to 400 seconds in testing, while the 30-second timeout fallback also stopped working.
  • Now: All those calls run asynchronously, and the UI stays responsive while long tasks execute.

---

文件校验（SHA256）

文件	SHA256
taishen_setup_1.8.1.exe	2AD63DD9BABEDA7EC068F5CE38EE3C68E4FF476A9E4D879783062180580A42D3
taishen_1.8.1.zip	56B830B83C2EEC2A4DF2EF75857BDD9F6E549CEDAA26AA1A5F8967A79B2CC7D3
taishen_1.8.1_macOS_arm64.dmg	5CB097F43D704FB550303B65C3798FDC2012AB5322D2A9F267267F5428A50510

---

v1.8.0

##v1.8.0

• 升级了泰案设计画布的工作台架构与组件表现力。
  • 重构为三栏工作台布局（左侧整合图层与资产库、中间画布舞台、右侧属性面板），优化顶部操作菜单，画板默认自适应匹配设备外框与质感边框。
  • 新增动效卡片与现代交互组件资产系统，内置极光流光、微光呼吸、3D 视差等 8 款高保真现代动效卡片，支持真机扫码预览与动效参数可视化微调。
• 修复了旧版本数据库升级后历史会话可能报错的问题。
• 修复了会话搜索全文索引重建偶发中断导致搜索结果不全的问题。

##v1.8.0 (English)

• Upgraded the workspace architecture and component expressiveness of the Tai An Design canvas.
  • Restructured into a three-column studio layout (layers and asset library integrated on the left, canvas stage in the center, properties inspector on the right), optimized the top action menu, and made boards adaptively match device frames and styled borders by default.
  • Added an animated card and modern interactive component asset system with 8 high-fidelity animated cards (aurora glow, subtle shimmer breathing, 3D parallax, etc.), supporting mobile QR scan preview and visual fine-tuning of motion parameters.
• Fixed an issue where historical sessions could encounter errors after upgrading from legacy database versions.
• Fixed an issue where occasional interruptions during full-text search index rebuilds caused incomplete search results.

---

文件校验（SHA256）

文件	SHA256
taishen_setup_1.8.0.exe	8E3BD1B0217EBFA35D2CE55084224198A6E90A9D82DE513CD76127B0995A80A4
taishen_1.8.0.zip	C85F9B57E34CD07292EB9B871AFC8B469DB5E7812241436BAFAB077ED68D3312
taishen_1.8.0_macOS_arm64.dmg	30092B0882A286F92B7B3F80BCE2AEC21A6F188EDDE30C2CAF091D2CAE44E5B8

---

v1.7.9

##v1.7.9

• 新增设计画布「手机预览」。
  • 更新前：设计画布原型只能在桌面窗口内查看，无法还原手机触屏的点按与滑动体验。
  • 更新后：点击顶部操作栏「手机预览」生成二维码，手机连接同一 Wi-Fi 扫码即可打开原型，支持页面跳转、弹层与转场动画；预览服务仅监听局域网地址，非内网来源请求直接拒绝。
  • 入口位于设计画布顶部操作栏。
• 新增内置浏览器下载能力与保存位置设置。
  • 更新前：网页下载缺少统一的保存流程与位置控制。
  • 更新后：下载接入系统原生保存对话框，默认指向系统下载目录，文件名自动清洗，不自动打开或执行下载文件；浏览器工具栏新增下载设置按钮，可选择「每次询问 / 上次保存位置 / 指定文件夹」，设置跨重启保留。
  • 首次默认每次询问；浏览器面板与画布窗口共用同一份配置。
• 优化了设计画布素材库。
  • 更新前：保存的素材与模板不携带 Design Tokens 依赖，跨文档复用会丢失主题级联；整页模板不记录设备外框。
  • 更新后：保存时自动收集并打包引用的语义 Tokens 与设备预设/外框，复用后主题切换与设备外壳还原保持完整；用户素材在面板内置顶展示，并支持重命名。
• 优化了上下文压缩机制。
  • 更新前：/compact 以轮数作为门槛，工具调用密集的少轮会话即使已塞满上下文也会被拒绝压缩；设置页的压缩水位滑块不生效。
  • 更新后：压缩判据改为按 token 水位判断，压缩触发线、字节线与阈值全部随当前模型的上下文窗口派生；设置页「全量压缩水位」滑块恢复生效，并移除了三项从不被读取的压缩参数控件。
• 修复了泰深安装目录无法读写的问题。
  • 更新前：安装目录下的内置技能、模板与资源被访问拦截，无法查看或落盘自定义内容。
  • 更新后：安装目录读写整体放行；安装到 Program Files 时仍受系统写权限限制。

##v1.7.9 (English)

• Added "Mobile Preview" to Design Studio.
  • Before: Design prototypes could only be viewed within desktop windows, failing to replicate touch-screen tap and swipe experiences on mobile devices.
  • Now: Click "Mobile Preview" on the top action bar to generate a QR code, connect your phone to the same Wi-Fi, and scan to open the prototype, supporting page transitions, overlays, and animations; the preview server listens only on local network addresses and rejects non-LAN requests.
  • Located on the top action bar of Design Canvas.
• Added native browser download handling and save location settings.
  • Before: Web downloads lacked a unified save flow and location controls.
  • Now: Downloads connect to the system native save dialog, defaulting to the system Downloads folder with sanitized filenames; does not automatically open or execute downloaded files; the browser toolbar adds a download settings menu with three modes: "Ask every time / Last used location / Specified folder", persisting across restarts.
  • Defaults to "Ask every time" initially; browser panels and canvas windows share the same configuration.
• Improved the Design Studio asset library.
  • Before: Saved assets and templates did not package Design Tokens dependencies, losing theme cascades across documents; page templates did not preserve device frames.
  • Now: Automatically collects and packages referenced semantic tokens and device presets/frames upon saving, keeping theme switching and device shells intact; user-created assets are pinned to the top of the panel and support renaming.
• Improved the context compaction mechanism.
  • Before: /compact used conversation round count as a prerequisite, rejecting compaction even when tool-heavy sessions filled the context window; the compaction watermark slider in settings had no effect.
  • Now: Compaction decisions are based on token watermarks, with triggers, byte limits, and thresholds derived dynamically from the current model's context window; the "Full compaction watermark" slider in settings is functional again, and three unused compaction controls have been removed.
• Fixed an issue where Taishen's installation directory could not be read or written.
  • Before: Built-in skills, templates, and resources under the installation directory were intercepted by access guards, preventing users from viewing or persisting custom content.
  • Now: Both read and write operations in Taishen's installation directory are allowed; installations in Program Files remain subject to operating system write permissions.

---

文件校验（SHA256）

文件	SHA256
taishen_setup_1.7.9.exe	F4D15575D5F1812E72E739300CE95BE5EB5419690A4BD2494ED695A4639081A8
taishen_1.7.9.zip	7E46BC477F0B0B0A3A9C61BB5709CCE0DB5E1B19821E5CF3E4CB0C5C8F85DBB3
taishen_1.7.9_macOS_arm64.dmg	DCDB15FA84D3F6E8A555C75989768483E01004B732FB474EFFA8851776CDBE12

---

v1.7.8

##v1.7.8

• 重大更新，泰案画布家族迎来自主升维的 Design 设计画布 V2（Design Studio V2）。
  • 更新前：设计画布主要偏向轻量线框生成，缺少专业矢量图形表现力、布尔运算与工程交付能力。
  • 更新后：全面升级为专业级矢量设计工作台。新增遮罩与蒙版系统、背景磨砂玻璃、内阴影与单边描边；支持矢量布尔运算（并集/减去/交集/排除）、智能参考线与 Alt 间距测距；引入三层类型化 Design Tokens 设计系统；支持导出真实矢量 SVG、Tailwind CSS 与 React JSX 代码，以及 Play Mode 原型交互演示；打通视口截帧多模态审美自检、原位主体一键抠图与截图逆向转译设计稿。
  • 入口位于泰案画布「新建-设计画布」，交互与样式面板随画布默认启用。
• 优化了记忆系统的配额管理与索引注入机制。
  • 更新前：记忆索引缺少类型配额约束，条目增多后易被次要记录挤占，长索引偶发截断残缺。
  • 更新后：记忆注入采用类型配额制，优先保障用户约定与核心身份信息，全面打通用户偏好（preference）注入链路；大幅放宽索引视图容量上限，彻底解决长索引截断；优化了中文分词与记忆检索排序。
• 优化了用户自建工具的签名校验与防护。
  • 更新前：用户自建工具若误用单参数签名，调用 ID 会被误认成参数，导致入参静默丢失且无报错。
  • 更新后：加载工具时增加严格的签名校验，发现不合规的单参写法主动拦截并提示标准格式，消除静默失效隐患。
• 优化了底层网络请求的长连接保活机制。
  • 网络请求保活超时从 4 秒提升至 60 秒，显著降低高频请求与长任务交互时的重复建连开销。

##v1.7.8 (English)

• Major update: Tai An canvas family welcomes the upgraded Design Studio V2.
  • Before: The Design canvas focused primarily on lightweight wireframe generation, lacking professional vector graphic styling, boolean operations, and production code delivery.
  • Now: Fully upgraded to a professional vector design studio. Added mask and clip path systems, background backdrop blur, inner shadow, and individual single-side strokes; supports vector boolean operations (union/subtract/intersect/exclude), smart guides, and Alt-distance measurement; introduces a three-tier typed Design Tokens system; supports exporting real vector SVG, Tailwind CSS, and React JSX code, alongside Play Mode prototype interaction previews; integrates viewport capture multimodal visual inspection, in-place subject matting, and vision-to-design screenshot reverse translation.
  • Located in Tai An Canvas → New → Design Canvas; interaction and property panels are enabled by default with the canvas.
• Improved quota management and index injection in the memory system.
  • Before: Memory index injection lacked type quota constraints, easily becoming crowded by secondary records as items grew, with occasional truncation on long indexes.
  • Now: Memory injection adopts a type quota system, prioritizing user conventions and core identity facts while fully connecting the user preference injection pipeline; significantly expands the index view capacity limit, resolving long index truncation; optimizes Chinese word segmentation and memory retrieval ranking.
• Improved signature validation and protection for custom user tools.
  • Before: If a custom tool mistakenly used a single-parameter signature, the runtime treated the tool call ID as the parameters object, causing arguments to be silently lost without errors.
  • Now: Enforces strict signature validation when loading tools, actively intercepting non-compliant single-parameter signatures and displaying standard format hints to prevent silent parameter loss.
• Improved keep-alive timeout for underlying network requests.
  • Network request keep-alive timeout increased from 4 seconds to 60 seconds, significantly reducing connection handshake overhead and network jitter during high-frequency calls and long-running tasks.

---

文件校验（SHA256）

文件	SHA256
taishen_setup_1.7.8.exe	2B92B38C85FCF011B890BC7D02C1C31B3AA23DE9E4F5E9D1E930D092C0F96CFE
taishen_1.7.8.zip	866EE57E7E8612BB0E5A26683D8A53218AD09BCAB0DBFB45C97FBC6513A4167C
taishen_1.7.8_macOS_arm64.dmg	17C8F5EFA24C3208EB436954508586700E1DCB63754A1B8E8D27AD7106381148

---

v1.7.7

##v1.7.7

• 重大更新，Google 订阅支持多账号管理与一键切换。
  • 更新前：Google 订阅只能保持单个账号登录，额度用尽或切换账号需要先退出再重新登录。
  • 更新后：支持多账号凭据存档与一键切换，切号即时生效无需重启；设置页提供卡片化账号列表，展示各账号的额度快照与批次进度，并支持单个账号移除。
  • 入口位于「设置-模型配置-Google 订阅」。
• 重大更新，泰案画布家族新增 Image Studio 图像创作工作台。
  • 更新前：AI 生图主要依赖对话流中的文字与图片卡片交互，缺乏专门的生图操作与版本管理界面。
  • 更新后：提供集参数配置、生成画布、参考图输入、历史版本对比于一体的三栏工作台；支持文生图与图生图参考图输入，生图任务全生命周期打通，结果跨重启持久化保存。
  • 入口位于泰案画布「新建-生图画布」。
• 重大更新，系统提示词与工具结构实施全量手册化减重。
  • 更新前：大量工具使用规范与画布操作说明常驻在系统提示词与工具说明中，挤占上下文并增加每轮 Token 消耗。
  • 更新后：建立核心工作手册体系，15 个领域专有工具与系统工具说明全面精简，转为精细化手册索引；首轮冷启动及后续每轮对话净节省约 15,000 tokens，提升了模型聚焦度与响应效率。
• 新增内置技能「ai-video-preproduction」（AI 视频视听前期筹备）。
  • 为 AI 视频与短片创作提供前期工作流底座，支持剧本节拍拆解、六层分镜设计、空间调度、资产审阅返修以及可重现的视听筹备包导出。
• 优化了调色画布本地视觉模型的下载与管理。
  • 模型管理浮层新增「打开模型目录」入口与离线放置说明，模型清单接入 GitHub Release 第三级镜像兜底源；重构了下载超时与任务排队保护，避免首次大模型下载被误杀中断。
• 优化了部分界面主题的视觉质感。
  • 调整了护眼绿（eyecare）与深海（deepsea）等主题的层级色彩，优化了亮色主题的高饱和度表现，提升长时间阅读舒适度。
• 修复了第三方模型在工具调用返回后偶发中断停住（空 turn 假死）的问题。
• 修复了 Windows 平台下文件搜索（glob）路径模式偶尔失效的问题。
• 修复了冷启动与历史继承会话中偶发出现伪造工具调用格式文本的问题。
• 修复了后台任务监控（monitor）在 Windows 环境下工作目录绑定异常的问题。

##v1.7.7 (English)

• Major update: Google subscription now supports multi-account management with one-click switching.
  • Before: Google subscription only allowed a single account sign-in; switching accounts or running out of quota required signing out and signing back in.
  • Now: supports multi-account credential archiving and instant one-click switching without restarting; the settings page provides a card-based account list showing quota snapshots and batch progress, along with individual account removal.
  • Located in Settings → Model Configuration → Google.
• Major update: Tai An canvas welcomes a new member — Image Studio creative workspace.
  • Before: AI image generation relied mainly on text and image cards in chat streams, lacking a dedicated creation and version management workspace.
  • Now: provides a three-column workspace integrating parameter setup, generation canvas, reference image input, and version comparison; supports text-to-image and image-to-image reference routing with persistent cross-restart storage.
  • Located in Tai An Canvas → New → Image Studio Canvas.
• Major update: system prompt and tool schema slimming via comprehensive manual architecture.
  • Before: extensive tool usage rules and canvas guidelines resided continuously in the system prompt and tool schemas, consuming context budget and increasing per-turn token usage.
  • Now: established a core workspace manual system, condensing 15 domain and system tool schemas into fine-grained manual indexes; saves ~15,000 tokens across cold start and subsequent conversation turns, improving model focus and responsiveness.
• Added built-in skill "ai-video-preproduction" (AI Video Audiovisual Pre-production).
  • Provides a pre-production workflow foundation for AI video and short film creation, covering screenplay beat breakdown, six-layer storyboard design, spatial staging, asset review/repair, and reproducible audiovisual pre-production package export.
• Improved local vision model downloading and management in the Color canvas.
  • The model management modal adds an "Open model directory" button and offline placement guide, with model manifests connected to a third-tier GitHub Release mirror fallback; reworked download timeout and queue protection to prevent first-time large model downloads from being killed prematurely.
• Improved visual quality across UI themes.
  • Tuned hierarchical contrast for eyecare and deepsea themes, and smoothed high-saturation colors in light themes for enhanced long-term reading comfort.
• Fixed an issue where third-party models occasionally stalled after tool calls (empty turn freeze).
• Fixed an issue where file search (glob) path patterns occasionally failed on Windows.
• Fixed occasional format poisoning of pseudo tool-call syntax during cold start and session inheritance.
• Fixed working directory binding anomalies for background task monitoring (monitor) on Windows.

---

文件校验（SHA256）

文件	SHA256
taishen_setup_1.7.7.exe	3790EFD94D2B20C1A38F581DE493C283A137FAB99FD835E7633DCE3D1D24EED8
taishen_1.7.7.zip	C4BEB1D4B90222709C27E865F57280D7157AB8D68A8828603D7DE908CF4F9DFE
taishen_1.7.7_macOS_arm64.dmg	0995041711348071ADA6C46B954AAA756A74C65BA4930A53E1E097D9EB2DE5D0

---

v1.7.6

##v1.7.6

• 重大更新，调色画布升级到 V3 版本。

  • 图层体系：更新前调整是一条全局链加若干区域批注，区域之间不能独立成层，也不能单独开关或调顺序；更新后采用图层栈，每层可有自己的调整链、曲线、蒙版、不透明度与混合模式，层可启用/停用、可调顺序，原图成为最底层可直接调色。
  • 图层蒙版：用矩形、椭圆、套索等形状建立，支持多部件加/减/交运算、反选、形状手柄拖动与旋转；层行显示蒙版缩略图，选中层可编辑蒙版形状、羽化或移除。
  • 本地视觉模型：新增提示分割（框选或点选目标出候选）、天空/水体/人像分割、相对深度分层、人脸检测与人脸关键点；模型输出的选区可直接转为图层蒙版。
  • 模型按需下载：模型不随安装包分发，首次使用按需下载，支持多源镜像、断点续传与字节级进度；下载后离线可用，图片不出本机。
  • 智能选取面板：在图上框选或点选 → 出候选（带缩略图与「没圈住」评分）→ 选定后建蒙版。
  • 旧文档升级：v1/v2 文档可一键升级到 v3，原有调整链与区域批注迁移为对应图层；关闭画布后可从 outputs 目录的 .color.json 重新打开。
  • 泰深侧同步打通：读写都走 v3 图层语义，区域调色改为新建带蒙版的层。

• 重大更新，新增 AI 生图能力与图片消息链路。

  • 新增 img_creator 生图子代理与 generate_image 工具，泰深可直接按描述出图。
  • 生成结果落盘到会话 outputs，在会话流里以缩略图展示，点击可在右侧面板全屏放大。
  • 支持两条订阅通道：Google（gemini-3.1-flash-image）与 Codex（gpt-image-2），都走订阅额度，不需要额外 API Key。
  • 设置页新增「生图专员」配置：选择 provider 会自动联动预设协议与模型。
  • 图片跨重启保留。

• 新增 Google 订阅 Provider，用 Google 账号订阅额度直连。

  • 该通道使用 Antigravity 的鉴权与额度接口，订阅账号需先通过 Antigravity 验证才能使用。
  • 泰深自持 OAuth 登录 Google 账号，不用 API Key；登录可取消、可重试。
  • 模型按账号可用清单展示，Gemini 系与 Claude 系分别显示图标，只提供实际存在的强度档位。
  • 额度卡显示 5 小时与 7 天两个池子的用量，临近阈值提醒。

• 新增内置浏览器登录态导入，并修复登录被拒问题。

  • 新增 browser_import_cookies：把系统浏览器的 cookie 导入内置浏览器，绕开嵌入式浏览器登录被站点拒绝的问题；支持多种导出格式自动归一化，可先 dry-run 确认。
  • 修复内置浏览器被 Google 判定为「不安全浏览器」而无法登录（补齐 UA-CH 品牌与 window.chrome 对象）。
  • 内置浏览器增加主路径广告拦截。
  • 新增内置技能 browser-login-fallback，登录被拒时给出处置路径。

• 优化了设置页的界面布局与视觉细节（表单行排版重排、控件宽度收敛、关闭入口统一为 X 与 ESC）。

• 优化了会话页的工具执行提示（执行中显示具体动作，如「阅读文件中：xxx」）。

• 优化了 Fork 子代理的上下文接力（携带完整历史与占位工具结果，避免提示词前缀缓存断裂；修正用量统计虚高）。

• 修复了以下问题。

  • Claude 通道工具调用返回 400。
  • 设置页定价与悬浮时钟的 provider 不一致。
  • 调色画布换图时若扩展名相同不刷新（v2 文档同样存在）。
  • 调色画布反选后无法取消，只能撤销。
  • 自建工具产出的文件没被收录进会话文件卡片。
  • 跨会话写文件落到错误目录。

##v1.7.6 (English)

• Major update: the color canvas is upgraded to V3.

  • Layer system: previously adjustments were one global chain plus several region annotations — regions could not be layers of their own, and could not be toggled or reordered individually. Now it uses a layer stack: each layer has its own adjustment chain, curves, mask, opacity and blend mode; layers can be enabled/disabled and reordered, and the source image is the bottom layer that can be adjusted directly.
  • Layer masks: built from rectangle, ellipse, lasso and other shapes, supporting multi-part add/subtract/intersect operations, inversion, and shape-handle dragging and rotation; the layer row shows a mask thumbnail, and the selected layer lets you edit the mask shape, feather or remove it.
  • Local vision models: added prompt segmentation (box or point on a target to get candidates), sky/water/person segmentation, relative depth layering, face detection and face landmarks; model output selections can be turned into layer masks directly.
  • On-demand model download: models are not shipped in the installer, they download on first use with multi-source mirrors, resumable transfer and byte-level progress; after download they work offline and images never leave the machine.
  • Smart selection panel: box or click on the image → candidates (with thumbnails and a "not captured" score) → pick one to create a mask.
  • Document upgrade: v1/v2 documents upgrade to v3 in one click, with existing adjustment chains and region annotations migrated into layers; after closing the canvas you can reopen the .color.json from the outputs folder.
  • Taishen side wired up: reads and writes both go through v3 layer semantics, and region color grading becomes creating a layer with a mask.

• Major update: added AI image generation and the image message pipeline.

  • Added the img_creator image-generation subagent and the generate_image tool — Taishen can generate images directly from a description.
  • Results are saved to the session outputs and shown as thumbnails in the session stream; click to view full screen in the right panel.
  • Two subscription channels are supported: Google (gemini-3.1-flash-image) and Codex (gpt-image-2), both using subscription quota with no extra API key.
  • The settings page has a new "image creator" configuration: picking a provider automatically syncs the preset protocol and model.
  • Images persist across restarts.

• Added a Google subscription provider that connects directly with a Google account's subscription quota.

  • This channel uses Antigravity's authentication and quota APIs; the subscription account must pass Antigravity verification first.
  • Taishen runs its own OAuth login for the Google account, no API key needed; login can be cancelled and retried.
  • Models are listed per what the account actually has available, with separate icons for the Gemini and Claude families, showing only the effort levels that really exist.
  • The quota card shows both the 5-hour and 7-day pools, with near-threshold reminders.

• Added browser login-state import for the built-in browser, and fixed login rejections.

  • New browser_import_cookies: import cookies from your system browser into the built-in browser, working around sign-in rejection in the embedded browser; multiple export formats are normalized automatically, with a dry run available first.
  • Fixed the built-in browser being flagged by Google as an "unsupported browser" that could not sign in (added UA-CH brands and the window.chrome object).
  • Added main-path ad blocking to the built-in browser.
  • Added the built-in browser-login-fallback skill, which lays out the diagnosis and handling path when sign-in is rejected.

• Improved the settings page layout and visual details (reworked form row layout, tightened control widths, unified the close entry as X and ESC).

• Improved tool execution hints on the session page (in-progress actions now show the specific operation, e.g. "Reading file: xxx").

• Improved context hand-off for Fork subagents (carries full history plus placeholder tool results to avoid prompt prefix cache breaks; fixed inflated usage stats).

• Fixed the following issues.

  • Tool calls on the Claude channel returned 400.
  • The provider shown in settings-page pricing and the floating clock did not match.
  • The color canvas did not refresh when swapping to an image with the same file extension (same in v2 documents).
  • Inversion in the color canvas could not be turned back off — only undo worked.
  • Files produced by custom tools were not collected into the session file card.
  • Files written across sessions landed in the wrong directory.

---

文件校验（SHA256）

文件	SHA256
taishen_setup_1.7.6.exe	2A275C6E5B7D52FE2A07D9C0BCBD58898BA1FC4D147C86F03E6F89CC02029BCB
taishen_1.7.6.zip	A77A786E28603F43F7B79A3A4F8B91907CB37283E87C009F28D93A112C432DD5
taishen_1.7.6_macOS_arm64.dmg	70464180DD7B8DB6555EB309B0F4A0649837688DFF17BAC1654E29043073ACA1

---

泰深 Color 画布本地模型（兜底源）

泰深 Color 画布 · 本地视觉模型（第三级兜底源）

本 Release 存放泰深调色画布（Color 画布 V3）本地推理所用的模型文件，作为模型按需下载的第三级源。

下载源顺序

泰深客户端下载模型时按以下顺序尝试，前一个不可用则自动换下一个（保留已下字节续传）：

1. HuggingFace 官方 —— huggingface.co
2. hf-mirror 国内镜像 —— hf-mirror.com
3. 本 Release —— github.com/EricXu20266/taishen/releases/tag/taishen_colors_model

三级源的同一文件逐字节一致（同 revision / 同字节 / 同 SHA256），可安全互换。

为什么会有这个 Release

上游模型仓由第三方维护，存在被删除、改名或变更的风险。本 Release 提供一份由泰深控制的镜像，确保模型下载链路不会因第三方变动而中断。

资产命名规则

资产名采用 模型标识__原文件名 的扁平形式（GitHub Release 资产名不支持目录分隔符）：

模型标识	对应模型
dav2-small	Depth Anything V2 Small（相对深度）
sam21-tiny	SAM 2.1 Hiera-Tiny（提示分割）
skywater-b2	SkyWater-Seg SegFormer MiT-B2（天空 / 水体 / 人像）
yunet	YuNet 2023mar（人脸检测）
facemesh	MediaPipe Face Mesh V2（478 点关键点）
withoutbg	withoutBG Open Weights v10.8.0 matting 分支（主体 / 前景）

许可与归属

模型	许可	上游
Depth Anything V2 Small	Apache-2.0	depth-anything/Depth-Anything-V2-Small
SAM 2.1 Hiera-Tiny	Apache-2.0	facebook/sam2.1-hiera-tiny
SkyWater-Seg	MIT	Vincentqyw/skywater
YuNet 2023mar	MIT	ShiqiYu/libfacedetection.train
MediaPipe Face Mesh V2	Apache-2.0	google-ai-edge/mediapipe
withoutBG Open Weights	Apache-2.0 AND LicenseRef-DINOv3	withoutbg/withoutbg-openweights-onnx

本 Release 中 withoutBG 相关资产的骨干网络使用 Meta DINOv3。按 DINOv3 License 第 1.b.i 条「分发时须随附协议副本」的要求，协议全文作为资产 withoutbg__LICENSE-DINOv3.md 一并提供。

Built with DINOv3

使用

泰深客户端会自动从本 Release 下载所需模型，无需手动操作。模型下载后落于用户本机 ~/.taishen/models/<modelId>/<revision>/。

所有模型推理均在本地完成，不上传任何用户图片或数据。

v1.7.4

##v1.7.4

• 增加了调色画布的图层混合模式与曲线端点自由拖动。

  • 混合模式：全局调整与每个区域调整都能单独设置，可选 正常 / 柔光 / 叠加 / 正片叠底 / 滤色 / 强光；侧栏下拉选择，效果强度滑杆对混合模式同样生效；区域之间互不影响；导出 PSD 时保留混合模式，可在 Photoshop 里继续调整。
  • 曲线端点：黑白场端点支持横纵拖动、方向键微调与输入/输出数值编辑；图表留白保证四角端点完整可见，端点拖出输入范围时保持对应输出值、不外推。
  • 泰深侧同步打通：可通过 set-blend 设置全局或区域混合模式。

• 增加了 PPT 画布三套能力展示型内置模板（打开即改、导出沿用自身骨架）。

  • motion-keynote.pptx：25 条原生元素动画 + 8 个转场，覆盖位移 / 淡入 / 擦除 / 旋转 / 缩放与路径位移。
  • data-narrative.pptx：5 张原生图表——柱 + 折线次坐标轴组合图、堆积柱、雷达、散点、气泡。
  • visual-lab.pptx：178 形状取样、三维格式与柔化边缘、完整轮廓、渐变与图案填充、39 种艺术字取样、图片圆形 / 六边形裁剪。

• 增加了会话页对自建工具文件改动的 diff 展示。

  • 更新前：只有内置的 write_file / edit_file 会在会话页显示 diff，自建工具改了文件看不到变化。
  • 更新后：编辑类自建工具也会显示 diff 块；超大的 diff 有展示防线，不阻塞会话。

##v1.7.4 (English)

• Added layer blend modes and free-draggable curve endpoints to the color canvas.

  • Blend modes: global adjustments and each region adjustment can set their own mode — normal / soft light / overlay / multiply / screen / hard light; picked from a dropdown in the sidebar, the effect-strength slider applies to blend modes as well, regions do not affect each other, and the mode is preserved on PSD export so it can be adjusted further in Photoshop.
  • Curve endpoints: the black and white point endpoints support horizontal and vertical dragging, arrow-key nudging, and input/output value editing; chart padding keeps all four corner endpoints fully visible, and an endpoint dragged outside the input range holds the matching output value instead of extrapolating.
  • The Taishen side is wired up too: blend modes for global or region scope can be set via set-blend.

• Added three capability-showcase built-in templates to the PPT canvas (open and edit directly; export keeps the template’s own skeleton).

  • motion-keynote.pptx: 25 native element animations + 8 transitions, covering motion, fade, wipe, spin, scale and path movement.
  • data-narrative.pptx: 5 native charts — bar + line combo with a secondary axis, stacked bar, radar, scatter and bubble.
  • visual-lab.pptx: 178 shape samples, 3D format and soft edges, full outlines, gradients and pattern fills, 39 WordArt samples, and circular / hexagonal picture cropping.

• Added diff display for file changes made by custom tools on the session page.

  • Before: only the built-in write_file / edit_file showed diffs on the session page, so file changes made by custom tools were invisible.
  • Now: editing-type custom tools show diff blocks as well; oversized diffs have a display guard so they do not block the session.

---

文件校验（SHA256）

文件	SHA256
taishen_setup_1.7.4.exe	32D0AFE57F95FFDFEE0E7057CDA7820A16FB442983C2D29E08AD37F23A793060
taishen_1.7.4.zip	C091DC0487ABD674C6CD7A36AB508B2DF10075A6F2BA2CCEE972435E1EFC0BAB
taishen_1.7.4_macOS_arm64.dmg	A2DA87C0358A50B48F21979BCE6239DE059D02DF56573A80CBFE87CCB4C292D4

---

v1.7.3

##v1.7.3

• 重大更新，PPT 画布支持原生转场与动画，同时补齐了一大批形状、效果与图表能力。

  • 转场与动画：打开已有的 PPT，转场和动画会被原样识别，可在画布上编辑并跟着导出。右侧新增动画面板，转场与元素动画按页列出，可增删、调顺序、改时长与触发方式；支持预览放映（含进入页转场、失焦自动暂停、自动换页与点击换页开关）。
    • 更新前：动画是看不见也改不了的盲区，动过内容的页面导出后效果可能丢失。
    • 更新后：导出按「保持 / 补丁 / 重写」分别处理，无法保真的页面会明确提示，而不是静默丢弃。
    • 支持范围：转场 cut / cover / pull、遮罩族（blinds / checker / comb / diamond / split / wheel）、几何变换（zoom / doors）、window / conveyor；元素动画支持路径位移、缩放、颜色 / 字体 / 字号与复合。逐字逐段构建与 Morph 未开放。
  • 形状与文本效果：预设形状开放面从 30 种扩到 178 种；新增三维格式（旋转与厚度）、柔化边缘、完整轮廓（虚线 / 端点 / 连接 / 箭头）；渐变支持变体与路径族（线性 / 径向 / 矩形 / 形状），新增 54 种图案填充，图片支持拉伸 / 平铺两种填充方式。
  • 艺术字：39 种文字包络预设，属性面板逐项标注「可预览 / 近似 / 需导出后查看」，不隐藏也不误导。图片艺术效果链（模糊 / 纹理化等）能读取并在面板展示，改效果链仍需在 PowerPoint 完成。
  • 图表：可插入类型扩到 13 种（雷达 / 散点 / 气泡 / 曲面 / 组合图 + 次坐标轴等），多系列图可在面板直接「加系列」；改动数据后导出保持原生图表（可在 PowerPoint 里继续改数据），不再降级成图片；支持瀑布图 / 漏斗图（chartEx）的解析、渲染与回写。
  • 编辑交互：连接线端点可拖拽吸附到形状连接点（拖到空白处解绑）、对象对齐分布工具条、元素级与文本级超链接、多级段落编号与缩进、把本页图片一键设为页面背景。

• 重大更新，调色画布升级到 v2 蒙版引擎——区域选区从单一形状变成可组合的复合蒙版，并新增曲线调色。

  • 更新前：一个区域批注只能是一种形状，想做「框选后挖掉一角」没有入口。
  • 更新后：一个区域可叠加多个形状（并集 / 相减 / 相交），画布上按住 Alt 拖动即可挖空；侧栏可查看每个形状、单独改羽化；叠加预览支持黑白灰度与过渡带轮廓两种显示。
  • 新增曲线调色：全局曲线 + 每个区域的独立曲线，控制点拖动实时预览。
  • 新增全局效果强度滑杆与「原图比例」叠加对比（第 4 种对比模式）。
  • 新建调色文档直接产出 v2，旧文档可在画布内一键升级，升级失败会给出具体原因。
  • GPU 加速覆盖组合蒙版 / 曲线 / 局部光；GPU 不可用时自动回落 CPU 并在界面说明。
  • 泰深侧通道同步补齐：可自建 / 删除选区、撤销与重做、导入图片、导出 PNG / PSD；批量调色也支持全局曲线。

• 增加了 A-Stocks 市场全景页的实时消息区。

  • 更新前：市场消息要靠对话里主动拉取，盘中即时性丢失。
  • 更新后：全景页内嵌消息流，60 分钟窗口持续滚动，重复消息按指纹去重，条目可点击「问 AI」。
  • 采集源为财联社 / 东财 / 腾讯三路；泰深可自行启停采集，并写入关心的订阅条件。

• 增加了泰深内置的 Python 与 uv 运行时——打包版自带，开箱即可跑 Python 系 MCP。

  • 更新前：uvx 托管的 MCP 和 Python 脚本依赖用户自己装环境，没装就用不了。
  • 更新后：安装包自带 uv 与 Python 3.12，MCP 的 uv / uvx / python 命令自动改写为内置绝对路径，首次冷启动有超时预算与提示。
  • 用户自己的环境优先，内置只作兜底：不写 PATH、不设环境变量，终端画布零注入。

• 增加了定时任务的批量管理与泰深自主清理。

  • 定时任务页新增「进行中 / 全部 / 已完成」三个页签，一次性任务执行完自动归档；批量模式可多选删除。
  • 泰深可通过 purge 清理已完成的定时任务，支持先预演清单再删除。

• 增加了 Google Gemini 原生协议支持（直接走 Gemini API，修复工具调用回喂 400）。

• 增加了 AI 约定文件的用户级全局来源：除项目根目录外，~/.claude 与 ~/.codex 下的约定文件也会作为最低优先级上下文注入。

• 优化了图片载荷：超大图片进入对话前先压缩，超出预算的会话图片会被裁剪，遇到 413 按会话恢复而不是整轮失败。

• 优化了悬浮时钟的高峰 / 低谷判断，改为跟随设置里的配置，周末不再算作高峰。

• 修复了界面里 70 处占位符原样显示的问题（回收站时间 / 条数等）。

• 修复了 monitor 在 Windows 下带引号的命令可能失败或静默错执行。

• 修复了设置页删除模型后聊天框里仍显示的问题。

• 修复了经验封装任务每轮重复生成、后台唤醒完成通知重复推送。

• 修复了定时任务页「已完成」页签恒空、MCP 详情页禁用状态显示原始占位符。

• 修复了调色画布 GPU 开关关闭再打开后永久退回 CPU。

• 修复了滚动条鼠标靠近变粗时的内容重排抖动。

##v1.7.3 (English)

• Major update: the PPT canvas now supports native transitions and animations, along with a large batch of new shape, effect and chart capabilities.

  • Transitions and animations: transitions and animations in existing decks are recognized as-is, editable on the canvas, and preserved on export. A new animation panel on the right lists transitions and per-slide element animations, where you can add, remove, reorder, and change duration and trigger; preview playback is supported (including the entry transition, auto-pause on focus loss, and auto-advance / click-to-advance toggles).
    • Before: animations were an invisible, uneditable blind spot, and pages whose content you edited could lose their effects on export.
    • Now: export handles each page as keep / patch / rewrite, and pages that cannot be preserved faithfully are reported explicitly instead of being silently dropped.
    • Supported: transitions cut / cover / pull, mask family (blinds / checker / comb / diamond / split / wheel), geometric transforms (zoom / doors), window / conveyor; element animations include path motion, scaling, color / font / font-size, and combinations. Per-character / per-paragraph build and Morph are not open.
  • Shapes and text effects: the open set of preset shapes grew from 30 to 178; added 3D format (rotation and depth), soft edges, and full outlines (dash / caps / joins / arrowheads); gradients support variants and path families (linear / radial / rectangular / shape), 54 pattern fills are available, and pictures support both stretch and tile fills.
  • WordArt: 39 text-warp presets, each labeled in the properties panel as "previewable / approximate / visible after export" — nothing hidden and nothing misleading. Picture artistic-effect chains (blur / texturize and others) can be read and shown in the panel; editing the chain still happens in PowerPoint.
  • Charts: insertable chart types expanded to 13 (radar / scatter / bubble / surface / combo with secondary axis, and more), and multi-series charts can add a series directly in the panel; after editing data the export stays a native chart (still editable in PowerPoint) instead of degrading to an image; waterfall and funnel (chartEx) are parsed, rendered and written back.
  • Editing: connection lines can be dragged to snap onto shape connection sites (drag to empty space to unbind), an align-and-distribute toolbar, element-level and text-level hyperlinks, multi-level paragraph numbering and indentation, and one-click "set this page's picture as the page background".

• Major update: the color canvas moves to a v2 mask engine — a region goes from a single shape to a composable composite mask — and adds curve grading.

  • Before: a region annotation could only be one shape, and there was no way to "select an area and carve a corner out of it".
  • Now: one region can stack multiple shapes (union / subtract / intersect); hold Alt and drag on the canvas to carve a hole; the sidebar lists each shape and lets you adjust its feather individually; the mask overlay previews in monochrome grayscale or transition contours.
  • New curve grading: a global curve plus an independent curve per region, with live preview while dragging control points.
  • New global effect-strength slider and an "original opacity" overlay compare (the 4th compare mode).
  • New color documents are created as v2 directly; existing documents can be upgraded in one click inside the canvas, and a failed upgrade states the specific reason.
  • GPU acceleration now covers composite masks / curves / local light; when the GPU is unavailable it falls back to CPU and says so in the UI.
  • The Taishen-side channels are filled in as well: create / delete selections, undo and redo, import an image, export PNG / PSD; batch color also supports global curves.

• Added a live news feed to the A-Stocks market panorama page.

  • Before: market news had to be pulled on demand from the conversation, losing intraday immediacy.
  • Now: the panorama page embeds a message stream with a rolling 60-minute window, duplicates removed by fingerprint, and entries clickable to "ask AI".
  • Sources are CLS / Eastmoney / Tencent; Taishen can start and stop collection on its own and record the conditions you care about.

• Added bundled Python and uv runtimes — the packaged app ships with them, so Python-based MCP servers work out of the box.

  • Before: uvx-hosted MCP servers and Python scripts depended on your own environment; without one they simply did not run.
  • Now: the installer bundles uv and Python 3.12; the uv / uvx / python commands used by MCP are rewritten to the bundled absolute paths, with a timeout budget and a notice on the first cold start.
  • Your own environment takes priority and the bundled one is only a fallback: no PATH writes, no environment variables, and zero injection in the terminal canvas.

• Added batch management and self-cleanup for scheduled tasks.

  • The scheduled-tasks page now has "Active / All / Completed" tabs, one-off tasks are archived automatically after they run, and batch mode lets you delete several at once.
  • Taishen can clean up completed scheduled tasks through purge, with a dry-run listing before deletion.

• Added native Google Gemini protocol support (calling the Gemini API directly, fixing the 400 error when tool results are fed back).

• Added a user-level global source for AI convention files: besides the project root, convention files under ~/.claude and ~/.codex are injected as the lowest-priority context.

• Improved image payloads: oversized images are compressed before entering a conversation, images beyond the session budget are cropped, and a 413 is recovered per session instead of failing the whole turn.

• Improved the floating clock's peak / off-peak judgement to follow the settings, and weekends no longer count as peak.

• Fixed 70 places in the UI where raw placeholders were displayed instead of values (recycle-bin times, counts, and more).

• Fixed monitor commands with quotes possibly failing or being silently mis-executed on Windows.

• Fixed models deleted on the settings page still showing up in the chat box.

• Fixed experience-encapsulation tasks being regenerated every round, and duplicate wake-up completion notices.

• Fixed the scheduled-tasks "Completed" tab always being empty, and the MCP details page showing raw placeholders for disabled states.

• Fixed the color canvas permanently falling back to CPU after turning the GPU toggle off and on again.

• Fixed content reflow jitter when the scrollbar thickens as the mouse approaches.

---

文件校验（SHA256）

文件	SHA256
taishen_setup_1.7.3.exe	13322C4B913C7E87C3B4C6771B0FBD7DC4DEADEF1E8BFCF47E8F939FDC7CAFA3
taishen_1.7.3.zip	C9E3BF05EA68AA46AEA022D1F10A4AA8491255011AD25EBAE1572716BF88979D
taishen_1.7.3_macOS_arm64.dmg	03BCB87104EF76EFAA6CCF52129A5A351AD4FE19F196CBCE5CDB45402BB49213

---