Claude AI Harness

最近更新 What's New

最新功能与改进一览。 Latest updates, features, and improvements.

2026-06-19
feat 🔍 PPT 现在「生成后会自己看一眼」：渲染 + 视觉质检闭环 🔍 PPT now "looks at itself" after building: a render + visual-QA loop
做 PPT 不再是「盲生成」。现在每份幻灯片在交付前都会被真正渲染成图、由 AI 逐页看一遍：检查有没有隐形/低对比文字、文字溢出被裁、装饰压字、元素重叠这类「一眼可见」的毛病。有硬伤就自动打回、列出每页问题并修正后重做，确认没问题才给你下载——所以基本不会再收到文字被裁掉或看不见的幻灯片了。主观的版面/留白只作建议、不会卡住一份能用的稿。 Building a deck is no longer "blind". Every presentation is now actually rendered to images and reviewed page-by-page by an AI before delivery — checking for invisible/low-contrast text, clipped overflow, decorations covering text, and overlapping elements. Hard defects are auto-rejected, listed per slide, fixed, and rebuilt; you only get the download once it's clean — so you should no longer receive slides with clipped or invisible text. Subjective layout/whitespace is advice only and won't block a usable deck.
2026-06-18
feat 📊 PPT 生成大升级：矢量图标 + 设计系统 + 自动防溢出 📊 Major PPT upgrade: vector icons + design system + auto overflow guard
让 AI 做 PPT 现在能出专业观感的幻灯片，对齐官方 pptx skill 水准：① 矢量图标——内置 react-icons，概念自动配上 crisp 的圆形图标徽章（不再靠糊掉的 AI 大图）；② 设计系统——统一调色板、衬线+无衬线字体配对、卡片阴影、去「AI 味」，并内置封面/2×3 网格/架构纵向流/双栏等版式范式；③ AI 生图——封面与氛围插图可一键生成嵌入；④ 自动防溢出卡口——生成前做确定性几何自检，元素越界会被拦下重排，不再出现被裁切的页面。只需说「做一份 XX 的 PPT」，下载即用、可二次编辑。 AI-built decks now reach a professional look, on par with the official pptx skill: ① Vector icons — built-in react-icons gives every concept a crisp circular icon badge (no more blurry AI images); ② Design system — unified palette, serif+sans pairing, card shadows, de-"AI-flavored" styling, plus built-in layout patterns (cover / 2×3 grid / vertical architecture flow / two-column); ③ AI image generation — one-shot covers and atmospheric illustrations embedded into slides; ④ Auto overflow guard — a deterministic geometry check before export rejects out-of-bounds elements and reflows them, so slides are never clipped. Just say "make a deck about X" — download-ready and fully editable.
2026-06-12
perf ⚡ 流畅模式跑通完整工具链：评测 21/30，与 Claude 持平 ⚡ Smooth Mode now runs the full tool chain: 21/30 on eval, on par with Claude
流畅模式（DeepSeek）现在跑的是同一套完整 harness——联网搜索、抓网页、多轮推理全部保留，不再只是纯聊天。在 30 题的 agent 评测数据集上拿到 21/30（70%），与 Claude 路径（20/30）基本持平，远高于原始基线（37%）。换句话说：日常用流畅模式又快又稳，能力也不打折。 Smooth Mode (DeepSeek) now drives the same full harness — web search, page fetching and multi-round reasoning all kept, no longer plain chat only. On our 30-task agent eval it scores 21/30 (70%), on par with the Claude path (20/30) and far above the original baseline (37%). In short: Smooth Mode is fast and stable for everyday use, with no drop in capability.
2026-06-11
feat ⚡ 流畅模式 —— DeepSeek 官方 API，快速稳定 ⚡ Smooth Mode — DeepSeek official API, fast and stable
顶栏点模型名字即可在「Claude Opus 4.7」和「⚡ 流畅模式」之间切换。流畅模式走 DeepSeek 官方 API，国内访问又快又稳，秒回、几乎不掉线，适合日常使用。它和 Claude 共用同一套完整 harness，联网搜索 / 抓网页 / 多轮推理等能力照常可用。选择会记住。 Tap the model name in the top bar to switch between "Claude Opus 4.7" and "⚡ Smooth Mode". Smooth Mode runs on DeepSeek's official API — fast and reliable in China, near-instant replies, rarely drops, great for everyday use. It shares the same full harness as Claude, so web search / page fetching / multi-round reasoning all keep working. Your choice is remembered.
2026-06-11
feat 等待回复时的动态状态提示 Live status while waiting for a reply
发出消息后、第一个字到达前不再是一片空白。现在会显示「✶ 正在思考…」的流光动画（短语会轮换），模型在联网 / 读网页 / 写文档时实时切换成「正在搜索…」「正在阅读…」「正在撰写…」，正文开始流式输出后自动消失。参考 Claude Code / Claude.ai 的等待反馈。 After you send a message, the gap before the first token is no longer dead air. A shimmering "✶ Thinking…" indicator now shows (the phrase cycles), switching live to "Searching…" / "Reading…" / "Writing…" while the agent runs a tool, and vanishing once prose starts streaming. Inspired by Claude Code / Claude.ai.
2026-06-11
fix 移动端 & iOS PWA 体验一揽子修复 Mobile & iOS PWA experience — a batch of fixes
围绕手机和"添加到主屏幕"的体验补齐：
  • 右侧"空气墙"空白：侧栏在手机上是抽屉，却仍占着一条 300px 网格轨道把内容挤到左边、右侧大片留白 —— 现已恢复单列满宽（720–1020px 宽度最明显）。
  • iOS 主屏 App 间歇性 "Load failed"：聊天请求遇到 WebKit 这个网络坑时自动重连最多 2 次，对用户透明。
  • 主屏图标满铺铜色：修了 Android 自适应图标被裁后缩水发灰、iOS 圆角留米色边框的问题。
  • iOS 安装图文引导页（/ios-install.html）：内嵌二维码 + 添加到主屏幕步骤，方便分享给 iPhone 用户。
 Polishing the phone and "Add to Home Screen" experience:
  • Right-side blank "wall": the sidebar is a drawer on phones but still reserved a 300px grid track, squishing content left with a big blank on the right — now single-column full width (most visible at 720–1020px).
  • Intermittent "Load failed" in the iOS home-screen app: chat requests now auto-retry up to twice on this WebKit network quirk, transparently.
  • Full-bleed home-screen icons: fixed Android adaptive-icon shrinkage/greying after masking and the iOS cream border after rounding.
  • iOS install guide page (/ios-install.html): embedded QR code + add-to-home-screen steps to share with iPhone users.
2026-06-10
perf 搜索 / 抓取管线重写 —— 又快又准 Search & fetch pipeline rewrite — faster and more accurate
把检索管线对照 Claude.ai 逐层重写：Brave 内联片段优先（够富就不抓页）、Exa highlights 做查询相关抽取（替掉整页字符截断）、9 引擎兜底链保覆盖、system/tools 前缀缓存 + 历史只追加撑住多轮（最多 12 轮）。在 30 道 GAIA+FRAMES 基准上，自建 harness 从 37% 提到 67%，对标 Claude.ai 的 80%。 The retrieval pipeline was rewritten layer by layer against Claude.ai: inline Brave snippets first (skip the fetch when rich enough), Exa highlights for query-relevant extraction (replacing whole-page truncation), a 9-engine fallback chain for coverage, and prefix caching + append-only history to sustain many rounds (up to 12). On a 30-task GAIA+FRAMES benchmark the harness went from 37% to 67%, versus Claude.ai's 80%.
2026-06-10
blog 技术博客：把搜索调到 Claude.ai 的水准 Tech blog: Tuning Search to Claude.ai Level
双语技术长文，拆解 Claude.ai 搜索"又快又准"的 5 个工程优化，再用自建 harness 的受控实验逐条验证（延迟、token、逐题 pass/fail，配框架图）。阅读全文 → A bilingual deep-dive: the 5 engineering optimizations behind Claude.ai's fast, accurate search, then validated one by one with controlled experiments on the hand-rolled harness (latency, tokens, per-task pass/fail, with diagrams). Read more →
2026-05-23
feat 设置统一入口 + 左栏布局优化 Unified settings entry + sidebar layout fix
左栏底部新增「设置」入口，把"长期记忆 / 我的数据 / 关于"合并成带 tab 的设置模态 —— 顶栏不再被账户级按钮挤占。同时修了左栏的老问题：「最近」对话多到一定数量就把「收藏」和「文档」挤到看不见 —— 现在三段各自有滚动边界，收藏和文档永远可见。 A new "Settings" entry now lives at the bottom of the left sidebar, consolidating "Memory / My Data / About" into a tabbed modal — the top bar is no longer crowded with account-level buttons. We also fixed a long-standing sidebar layout bug: when "Recent" filled up, "Starred" and "Documents" got pushed off-screen. Each section now has its own scroll boundary and stays visible.
2026-05-23
feat 隐私三件套 — 记忆 UI + 数据导出 + 一键彻底删除 Privacy triad — Memory UI + data export + one-click full deletion
围绕"数据主体权利"补齐三件事：
  • 长期记忆 UI：模型主动记下的事实现在能看 / 改 / 单条删除 / 清空全部
  • 导出我的数据：一个 JSON 包含全部对话 / 消息 / 文档 / 记忆 / 遥测 / 反馈
  • 彻底删除账号：两步确认（输邮箱 + 5s 倒计时），DB 里 threads / messages / documents / telemetry / user_memory / bug-reports 与 bug-images 全部级联清理
  • privacy.html 透明化：明确列出收集了什么字段、admin 可见范围、对应每条权利的操作入口
对应 GDPR/PIPL 的数据可携权 + 被遗忘权基本款。 Closing the loop on data-subject rights:
  • Memory UI: see, edit, delete individual entries, or wipe all of what the model has remembered about you
  • Export my data: one JSON containing all your threads / messages / documents / memory / telemetry / feedback
  • Full account deletion: two-step confirmation (type your email + 5s countdown), then cascading cleanup of every table plus bug-report files referencing your email
  • Transparent privacy.html: explicit field-level list of what's collected, admin visibility scope, and direct links to each rights-fulfillment entry
Covers the baseline GDPR/PIPL portability + erasure rights.
2026-05-23
feat 跨对话长期记忆 Cross-conversation long-term memory
新增 manage_memory 工具 + user_memory 表。告诉模型一次你的身份 / 项目 / 长期偏好（"我是软件工程师"、"请用中文"），新开任意对话都能延续 —— 不再每次重新自我介绍。模型自主决定记什么（read / append / replace 三种 action），不记录一次性问题或敏感信息。 New manage_memory tool + user_memory table. Tell the model once about your identity / project / persistent preferences ("I'm a software engineer", "reply in Chinese"), and every new conversation picks up where the last left off. The model decides autonomously what's worth remembering (read / append / replace actions); one-off questions and sensitive info are excluded by design.
2026-05-23
perf 启用 Prompt Caching，input 等效成本砍 ~50%、TTFT 显著下降 Prompt caching enabled — ~50% effective input cost reduction, faster TTFT
system prompt 和工具定义走 Anthropic 的 ephemeral cache，多轮对话里这部分基本只付一次"创建"成本，后续按 10% 成本读取。telemetry 新增 cache_creation_tokens / cache_read_tokens 两列可在管理后台追踪命中率。实测 8 轮新加坡新闻对话中 cache_read 累计 100K+ tokens，input 端等效成本约下降一半。 System prompt and tool definitions now ride Anthropic's ephemeral cache — paid once per ~5 min window, replayed at 10% of normal input cost. Telemetry exposes cache_creation_tokens / cache_read_tokens in the admin panel for hit-rate tracking. Measured on an 8-turn session: ~100K cached tokens replayed, effective input cost roughly halved.
2026-05-23
perf 主动上下文压缩 + Haiku 摘要长 tool_result Proactive context compression + Haiku-summarized tool results
长对话场景的压缩链路：超过预算阈值后用 Haiku 把过长的 tool_result（搜索全文等）摘成 1-2 句，按 content hash 缓存复用；早期工具轮整体摘要化；最后还有硬上限兜底。主对话仍走配置的好模型（默认 claude-opus-4-7），Haiku 只跑后台辅助任务 —— 不牺牲主回答质量。配套 telemetry 新增 compressed_from/to_tokens 和 haiku_calls 字段。 Long-conversation compression pipeline: above a token budget, oversized tool_results (search dumps, etc.) get Haiku-summarized to 1-2 sentences and cached by content hash; early tool rounds collapse to summaries; a hard cap is the last-line backstop. The main chat path still uses the configured strong model (default claude-opus-4-7), with Haiku only running offstage helper tasks — no quality trade-off on user-facing answers. Telemetry gains compressed_from/to_tokens and haiku_calls fields.
2026-05-22
feat 反馈带上对话 ID，bug 复现一步到位 Bug reports auto-attach conversation ID
反馈表单现在自动带上当前对话 ID 并落进通知邮件，省去用户手抄。管理员能直接点进对应对话看上下文。 The bug-report modal now auto-attaches the active thread ID and includes it in notification emails — admins can jump straight into the conversation context.
2026-05-22
fix Artifact 生成全链路体验改进 Artifact generation pipeline polish
一波连续修复：
  • 简单问题不再硬塞进 artifact，AI 先反问"要做成文档吗"再决定
  • docx 生成前先弹选项让用户确认风格 / 大纲，不再生成完才问
  • 预览面板支持 docx 直显，不再重复生成一份 html 副本
  • 进度条覆盖所有 artifact / Office 文件生成路径
  • 选项块从输出中段漂浮的 bug 修复，统一渲染在 message 末尾
  • 服务端拦截空 / 过短的 create_artifact，避免重试导致的空文档与重复
 A batch of artifact-related fixes:
  • Simple questions no longer get force-shunted into artifacts — model asks first
  • For docx whitepapers, style/outline options now appear BEFORE generation, not after
  • Preview pane renders docx directly — no more parallel HTML regeneration
  • Progress indicators now cover all artifact / Office generation paths
  • Option blocks no longer float mid-message — consistently rendered at the end
  • Server rejects empty / under-length create_artifact calls — no more empty-doc spam from retries
2026-05-22
feat Agent 可读回历史文档增量修改 Agents can now read & edit existing artifacts
新增 list_artifacts / get_artifact 工具。用户说"把第二章再扩两段"时，AI 先列出当前对话已有文档、读出原文，基于现有内容增量修改后 upsert —— 不再每次从零重新生成。 New list_artifacts / get_artifact tools. When users say "expand chapter two", the model now lists existing thread documents, reads the current body, and patches in place instead of regenerating from scratch.
2026-05-21
feat LaTeX 数学公式渲染 LaTeX math rendering
回答中的行内 $...$ 和块级 $$...$$ 公式现在直接渲染成排版好的数学表达式（基于 KaTeX）。反向传播推导、损失函数、积分号、矩阵这种密集公式不再原文露出。 Inline $...$ and block-level $$...$$ formulas in responses now render as properly typeset math (powered by KaTeX). Dense formulas — backprop derivations, loss functions, integrals, matrices — no longer leak raw source.
2026-05-21
feat UX 体验细节大复刻 claude.ai UX polish pass — closing the gap to claude.ai
Markdown 渲染补全（链接 / 有序列表 / 斜体 / 代码块语言标签 + 复制按钮）、代码块语法高亮（12 种语言，无外部依赖）、主题化弹层替换浏览器原生 prompt/confirm、滚动历史时浮现「跳到底部」按钮、空输入框时禁用发送按钮、文件拖拽上传、键盘快捷键（Cmd+N 新对话 / Esc 关面板 / Cmd+K 聚焦输入框）、侧边栏对话搜索与关键词高亮。 Full markdown rendering (links / ordered lists / italics / code-block language label + copy button), in-house syntax highlighting (12 languages, zero deps), themed inline modals replacing native prompt/confirm, scroll-to-bottom floating button, disabled send on empty input, file drag-and-drop, keyboard shortcuts (Cmd+N / Esc / Cmd+K), sidebar conversation search with keyword highlight.
2026-05-21
fix 修复代码块路由 + 点赞按钮无反馈 Code-block routing + thumbs-up no-feedback fixes
Python / JS / CSS 等代码块不再被错误推送到右侧文档面板，现在直接在聊天中渲染（仅 HTML / SVG 走文档预览）。点赞 / 踩按钮加上可见的橙色背景高亮（彩色 emoji 不响应 CSS color，原来的实现看上去毫无变化）。 Python / JS / CSS code blocks no longer hijacked into the artifact panel — they now render inline (only HTML / SVG continue to the preview pane). Thumbs-up / down buttons gain a visible orange background tint (color emojis ignore CSS color, so the previous attempt was a visual no-op).
2026-05-14
feat Office 文件生成 Office File Generation
在对话中直接生成 Word (.docx)、Excel (.xlsx) 和 PowerPoint (.pptx) 文件，在 Artifact 面板中一键下载。 Generate Word, Excel, and PowerPoint files directly from conversation — downloadable from the artifact panel.
2026-05-13
blog 技术博客：用 30 道题量化我和 Claude.ai 的差距 Tech blog: Benchmarking My Agent vs Claude.ai with 30 Tasks
固定模型、只变 harness 的对照评测：裸模型 0% → 自建 harness 37% → Claude.ai 80%。差距主要在检索管线而非模型——这也是上面"管线重写"的起点。阅读全文 → A controlled benchmark holding the model fixed and varying only the harness: naked model 0% → hand-rolled harness 37% → Claude.ai 80%. The gap is mostly in the retrieval pipeline, not the model — the starting point for the rewrite above. Read more →
2026-05-12
feat 用量追踪与管理后台 Usage Tracking & Admin Panel
新增按用户统计的 Token 用量面板，管理员可查看所有账户的月度 API 消耗。 Per-user token usage tracking with a dedicated admin panel tab and monthly breakdowns.
2026-05-11
fix 图表渲染修复 Matplotlib & Chart Rendering
  • 自动注册中文字体，matplotlib 图表正确显示中文Auto-register Chinese fonts for matplotlib code execution
  • 修复重复生成文档和图表显示截断的问题Fix duplicate documents and chart display cutoff
2026-05-10
feat 内联图表与 Bug 反馈增强 Inline Charts & Bug Report
Matplotlib 图表直接在聊天流中渲染；Bug 反馈支持图片上传、粘贴和拖拽。 Matplotlib charts render inline in chat. Bug reports support image upload, paste, and drag-drop.
2026-05-09
perf 缓解等待焦虑 Reduced Waiting Anxiety
新增进度条、计时器和轮播提示，减少 AI 响应时的等待焦虑。API Key 池支持轮询和自动冷却。 Progress bar, elapsed timer, and rotating tips during AI responses. API key pool with round-robin rotation.
2026-05-08
feat 分享功能与移动端文档访问 Sharing & Mobile Document Access
  • 生成临时分享链接（24 小时有效）Temporary share links with 24h expiry
  • 移动端浮动按钮快速访问文档Floating action button for mobile document access
  • Artifact 持久化到 SQLite，刷新页面不丢失Artifacts persist to SQLite — survive page refresh
  • 长文档生成断线可恢复Long document generation recoverable after disconnect
2026-05-07
feat SQLite 存储与多引擎搜索 SQLite Storage & Multi-Engine Search
从纯 localStorage 迁移到 SQLite 后端。搜索支持 5 引擎级联回退：Serper → Tavily → Brave → Google CSE → DuckDuckGo。 Migrated to SQLite backend. Web search supports 5-engine fallback chain.
  • 长文档并行分章生成Long document generation with parallel chapter writing
  • DOCX 导出DOCX export for generated documents
  • 移动端响应式布局优化Mobile-optimized responsive layout
2026-05-06
feat 多用户系统与 Bug 反馈 Multi-User Accounts & Bug Reports
用户注册与登录，Session 管理。内置 Bug 反馈功能，用户可直接在应用内提交问题。 User registration and login with session management. Built-in bug report feature.
2026-05-05
feat 双主题与代码解释器 Dual Theme & Code Interpreter
大版本 UX 升级：日/夜主题切换、沙盒 JavaScript 代码执行、文档管理侧栏与版本追踪。 Major UX overhaul — light/dark theme, sandboxed JS execution, document sidebar with versioning.
2026-05-04
feat v1 Agent 工具调用循环 Agentic Tool Use Loop
核心 Agent 架构上线：Claude 在循环中迭代调用工具（网络搜索、代码执行、Artifact 生成），直到任务完成。单次请求最多 8 轮，80K 上下文预算管理。 Core agentic architecture — Claude iteratively calls tools (web search, code execution, artifact creation) in a loop. Up to 8 rounds per request with 80K context budget.