Skip to content

v1.6.0: Browse

Choose a tag to compare

View all tags
@xuzhougeng xuzhougeng released this 22 Aug 13:06
v1.6.0
c17ecdf

wisp-science v1.6.0

v1.6.0 is the Browse release. The agent can drive a real Chrome session: your daily shared browser and an isolated workspace profile, article scans, hashed asset harvest, ChatGPT one-shot on a logged-in tab, and auto-launch when the extension is disconnected. It also adds a trajectory inspector, live context-usage percent, MCP App server-tool callbacks, and tighter provenance for kernel-written files.

Downloads

  • macOS (.dmg, signed + notarized): choose the Apple Silicon (aarch64) or Intel (x86_64) build; compatible versions can update in-app.
  • Windows installers (MSI / NSIS): signed; if SmartScreen still triggers, choose "Run anyway".
  • Linux (deb / AppImage): available for both x86_64 and aarch64.

See Assets below.

Highlights

Real Browser Runtime

  • Dual sessions: shared uses your existing Chrome/Edge cookies and extensions; workspace launches a dedicated profile so research browsing stays separate.
  • Protocol v2 extension waits for the final URL, scans article figures and code blocks, and stages downloads with SHA-256 into the project.
  • ChatGPT one-shot send / wait / read on an already-logged-in chatgpt.com tab. Lookalike hosts are rejected before any prompt is filled.
  • Settings → Browser → Open browser automatically starts Chrome when a browser tool needs the session and the extension is disconnected.
  • Pause control from the extension popup stops later automations until you resume.
  • Workspace Chrome fails closed if the browser no longer loads unpacked extensions, instead of driving a blank window.
  • Save paths must stay inside the project. Blocked hosts from Settings cannot be opened or navigated to.

Trajectory Inspector

  • The timeline icon between Share and the inbox, or /trajectory, opens a modal inspector. The chat itself is unchanged.
  • A session Gantt sits above turn-grouped USER / ASSISTANT / TOOL / USAGE events.
  • Click a row or a Gantt segment for Summary, Preview, Raw, and Source.
  • The footer aggregates turns, LLM and tool time, tokens, and cache hit rate.

Context, MCP Apps & Provenance

  • The composer shows live context-usage percent with warn / danger colors, and offers compact or a new session when usage is critical.
  • Session options, including the proxy, live under Settings → Session.
  • While an MCP App is bound to the server that presented it, in-app actions can call that same server without starting a new agent turn.
  • Python and R kernels report the files a cell actually wrote, so computed names such as fig_{i}.png are credited without stealing another session's files.

Reliability

  • Chat stays pinned to the latest message after tool results; outline and review jumps no longer snap back to the bottom.
  • DOCX preview controls stay visible.
  • Remaining generated-file attribution gaps under parallel sessions are closed.
  • Side-chat classifies intent instead of matching keywords.
  • Custom theme CSS, including turning off the lead bar, is documented.

Full Changes

All commits since v1.5.0:
v1.5.0...v1.6.0

wisp-science v1.6.0

v1.6.0 是浏览版本。智能体可以驱动真实的 Chrome 会话：日常共享浏览器与隔离的工作区配置，文章扫描、带哈希的资源取回、已登录 ChatGPT 标签页的一次性发送/等待/读取，以及扩展断开时自动拉起浏览器。本版同时加入轨迹检查器、实时上下文占用百分比、MCP App 回呼同一服务器，以及内核写出文件的更精确归属。

下载

  • macOS（.dmg，已签名 + 公证）：请选择 Apple Silicon（aarch64）或 Intel（x86_64）对应架构；兼容版本可从应用内更新。
  • Windows 安装包（MSI / NSIS）：已签名；如仍触发 SmartScreen，可选择「仍要运行」。
  • Linux（deb / AppImage）：提供 x86_64 与 aarch64 两个架构。

见下方 Assets。

核心亮点

真实浏览器运行时

  • 双会话：共享沿用你现有的 Chrome/Edge Cookie 与扩展；工作区启动独立配置，让科研浏览与日常浏览器分开。
  • Protocol v2 扩展会等待最终 URL，扫描文章中的图与代码块，并把下载文件以 SHA-256 登记进项目。
  • 可在已登录的 chatgpt.com 标签页上一键发送 / 等待 / 读取回复。形似域名会在填写提示词前被拒绝。
  • 设置 → 浏览器 → 自动打开浏览器 会在浏览器工具需要会话且扩展断开时启动 Chrome。
  • 扩展弹窗的暂停控制会阻止后续自动化，直到你恢复。
  • 若浏览器不再允许加载未打包扩展，工作区 Chrome 会明确失败，而不是去操作空白窗口。
  • 保存路径必须落在项目内。设置中拉黑的主机无法打开或跳转。

轨迹检查器

  • Share 与收件铃之间的时间线图标，或 /trajectory，会打开模态检查器；对话本身不变。
  • 会话级甘特图下方按回合分组展示 USER / ASSISTANT / TOOL / USAGE 事件。
  • 点击行或甘特片段可查看 Summary、Preview、Raw、Source。
  • 页脚汇总回合数、LLM 与工具耗时、token 与缓存命中率。

上下文、MCP App 与归属

  • 输入框显示实时上下文占用百分比，接近上限时用警告/危险色，并提供压缩或新开会话。
  • 会话选项（含代理）移到 设置 → 会话。
  • MCP App 仍绑定到打开它的服务器时，界面内操作可直接调用该服务器，不必新开一轮对话。
  • Python / R 内核会报告 cell 真正写出的文件，因此 fig_{i}.png 这类运行时算出来的文件名也能正确记到本会话，而不会记到并行会话上。

稳定性

  • 工具结果返回后聊天仍钉在最新消息；大纲与审查跳转不再被拉回底部。
  • DOCX 预览控件保持可见。
  • 补上并行会话下剩余的生成文件归属缺口。
  • 侧边聊天按语义识别意图，不再只靠关键词。
  • 文档补充了自定义主题 CSS，包括如何关掉段首强调条。

完整变更

自 v1.5.0 以来的全部提交见：
v1.5.0...v1.6.0

Assets 15
Loading