Skip to content

Taber v0.3.0

Choose a tag to compare

View all tags
@github-actions github-actions released this 09 Jul 12:49
v0.3.0
d47d4ac

English

Highlights

  • Added the fs tool: a virtual file system exposing a per-session /workspace file area and the /skills site-skill library through one ls/read/write surface.
  • Added the site skills system: reusable, agent-authored site knowledge with lazy disclosure, host-change announcements, builtin seeds for 8 major sites, and a freshness loop.
  • Added fetch-first URL reading: getDocument source:"url" fetches public webpages and PDFs directly without opening a tab, making cross-page collection parallel.
  • Added a document workflow: upload PDF/Word/text files, generate Markdown/HTML/CSV/Word from tasks, download from the side panel, and export Markdown/HTML to PDF via the browser print dialog.
  • Added context economy: oversized document and REPL outputs are saved to /workspace and only a preview plus a stable file path enters the model context and event log.

Added

  • fs tool (ls/read/write): /workspace/<name> holds session uploads and outputs; /skills/<slug>.md exposes site skills as frontmatter Markdown files, matching the file-system prior LLMs are trained on.
  • File workspace: writing .md/.txt/.html/.csv/.json stores text; writing .docx converts Markdown to Word (docx, dynamically imported); uploaded .docx/.pdf/text files are readable as text through getDocument source:"file" (mammoth/pdfjs).
  • Side panel file UI: paperclip upload in the composer (with attachment chips and prompt annotation), a files strip with download, delete, and Export PDF actions, and a dedicated print page (print.html, marked + DOMPurify) for high-quality PDF export via the browser print dialog.
  • Site skills: task instructions carry a digest of skills matching the current site; navigate (all actions), browser, and browserRepl announce availableSkills whenever an action lands on a new host, and never repeat on the same host.
  • Skill authoring loop: the agent saves non-obvious reusable site flows as /skills/*.md files with structured frontmatter (name/hosts/description); skills are priors and live page state always wins.
  • Skill freshness loop: after two consecutive tool failures in a task that read skills, the error hints at revising the stale skill file (once per task).
  • Builtin seed skills (versioned, idempotent): Hacker News, Reddit, GitHub, Wikipedia, Stack Exchange, npm registry, PyPI, and arXiv fetch-first API endpoints with data shapes, quota pitfalls, and fallbacks.
  • Skills management UI in the side panel: list, enable/disable, and delete stored skills.
  • getDocument source:"url": direct fetch of http/https pages and PDFs (content-type routed) reusing the Readability/Markdown/table pipeline; failures explicitly point back to navigate + currentPage.
  • browser coordinate targets { x, y } (viewport CSS px): visual fallback for canvas/visual-only UIs, resolving through open shadow roots and failing clearly on iframes; viewport screenshots now include width/height in CSS px as the coordinate reference.
  • Context spill: getDocument content over 12k chars and browserRepl JSON values over the same threshold are written to /workspace/saved-<hash>.md|json with a 4k preview, savedTo, and a read-back hint; identical content maps to the same file.
  • Local database: skills table (v2), per-session files table pruned with sessions (v3), and a v4 repair migration that renames legacy skills whose names collide on the /skills/<slug>.md path (newest keeps its name, nothing deleted).
  • ADR 0015 (agent-authored site skills), ADR 0016 (fs tool and file workspace), ADR 0017 (fetch-first URL reading).

Changed

  • Fixed top-level tool set is now 6: getDocument, extractImage, navigate, browser, browserRepl, fs (debug builds still add debugger).
  • Agent instructions v5: goal-anchored efficient loop (define done → smallest evidence → shortest path → act once → verify), fetch-first URL policy, and file/skill strategy.
  • getDocument source:"pdf" merged into source:"url" (auto-detected); source:"file" now reads workspace files by name instead of accepting inline fileText.
  • Product narrative updated to a powerful, all-capable side panel agent across CONTEXT/PRODUCT/plan; stop/resume and tool-trail mechanics unchanged.
  • Store compliance disclosure now covers locally stored session files and site skills.
  • New dependencies: docx and mammoth, both pure JS and dynamically imported so they load only when converting documents.

Fixed

  • Builtin skill seeding no longer overwrites user- or agent-authored skills with the same name, and individual seed failures no longer abort the rest.
  • Skill names that collide on the /skills/<slug>.md path are rejected at save time, and renaming a skill file is atomic (save first, then delete the old entry).
  • Runtime smoke script asserted a hardcoded, outdated instructionsVersion; it now references the shared constant.

中文

重点

  • 新增 fs 工具：虚拟文件系统，用一套 ls/read/write 同时暴露每会话的 /workspace 文件区和 /skills 站点技能库。
  • 新增站点技能系统：Agent 自行沉淀的可复用站点知识，支持懒加载披露、跨站提示、8 个大站内置种子和时效闭环。
  • 新增 fetch-first URL 阅读：getDocument source:"url" 不开标签页直接拓取公开网页和 PDF，跨页收集可并行。
  • 新增文档工作流：上传 PDF/Word/文本，任务产出 Markdown/HTML/CSV/Word，侧边栏下载，Markdown/HTML 可通过浏览器打印导出 PDF。
  • 新增上下文经济：超大的文档和 REPL 输出自动落盘到 /workspace，模型上下文和事件日志只保留预览和稳定文件路径。

新增

  • fs 工具（ls/read/write）：/workspace/<name> 存会话上传与产出；/skills/<slug>.md 以 frontmatter Markdown 文件形态暴露站点技能，利用模型对文件系统操作的预训练先验。
  • 文件工作区：写 .md/.txt/.html/.csv/.json 存文本；写 .docx 自动把 Markdown 转 Word（docx 库，按需加载）；上传的 .docx/.pdf/文本通过 getDocument source:"file" 解析为文本（mammoth/pdfjs）。
  • 侧边栏文件 UI：输入框回形针上传（附件 chip + 提示词自动标注）；文件条支持下载、删除、导出 PDF；新增打印页（print.html，marked + DOMPurify）走浏览器原生排版导出高质量 PDF。
  • 站点技能：任务指令附当前站点匹配技能摘要；navigate（全部动作）、browser、browserRepl 落到新 host 时统一附 availableSkills 提示，同 host 不重复。
  • 技能沉淀闭环：Agent 把非显而易见、可复用的站点流程写成 /skills/*.md（结构化 frontmatter：name/hosts/description）；技能是先验，与页面实时状态冲突时以页面为准。
  • 技能时效闭环：任务内读过技能后若工具连续失败两次，错误信息提示修订可能过时的技能文件（每任务只提示一次）。
  • 内置种子技能（按版本幂等写入）：Hacker News、Reddit、GitHub、Wikipedia、Stack Exchange、npm、PyPI、arXiv 的 fetch-first API 端点，含数据结构、配额陷阱和回退方案。
  • 侧边栏技能管理 UI：列表、启用/停用、删除。
  • getDocument source:"url"：直接 fetch http/https 网页与 PDF（按 content-type 自动分流），复用 Readability/Markdown/表格解析管线；失败时明确引导回退 navigate + currentPage。
  • browser 坐标 target { x, y }（viewport CSS px）：canvas/纯视觉 UI 的视觉兜底，递归穿透 open shadow root，命中 iframe 时明确失败；视口截图结果附 CSS px 的 width/height 作为坐标基准。
  • 上下文落盘：getDocument 内容超 12k 字符、browserRepl JSON 结果超同阈值时，全文写入 /workspace/saved-<hash>.md|json，结果携带 4k 预览、savedTo 路径和回读提示；相同内容幂等落到同一文件。
  • 本地数据库：skills 表（v2）、随会话清理的 files 表（v3）、v4 修复迁移（对历史上 slug 碰撞的技能自动加后缀重命名，最新者保留原名，零数据丢失）。
  • 新增 ADR 0015（Agent 沉淀的站点技能）、ADR 0016（fs 工具与文件工作区）、ADR 0017（fetch-first URL 阅读）。

变更

  • 固定顶层工具变为 6 个：getDocument、extractImage、navigate、browser、browserRepl、fs（debug 构建仍额外提供 debugger）。
  • Agent 指令 v5：目标锚点高效循环（先明确完成标准 → 读最小证据 → 选最短路径 → 一次动作 → 验证），fetch-first URL 策略，文件与技能策略。
  • getDocument source:"pdf" 并入 source:"url"（自动识别）；source:"file" 改为按名读取工作区文件，不再接受内联 fileText。
  • 产品叙事更新为强大全能的浏览器侧边栏 Agent（CONTEXT/PRODUCT/plan）；停止/恢复与工具轨迹机制不变。
  • 商店合规披露补充本地存储的会话文件与站点技能。
  • 新增依赖 docx 与 mammoth：纯 JS，动态 import，仅在文档转换时加载。

修复

  • 内置种子不再覆盖用户/Agent 沉淀的同名技能；单个种子写入失败不再中断其余种子。
  • 技能名 slug 碰撞导致 /skills 路径重复：保存时拒绝冲突，技能文件重命名改为原子操作（先保存后删旧），失败不丢数据。
  • 运行时冒烟脚本硬编码的 instructionsVersion 已过时，改为引用共享常量。
Assets 3
Loading