dsh-notebook-studio
dsh-notebook-studio 是 DeepSeek Harness（dsh）的插件，可将上传的 PDF 文献转换为带页码引用的报告（DOCX/PDF）或演示文稿（PPTX/PDF）
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:HaoKuo/dsh-notebook-studio#5d69780526a0bd0e4d526d3a5611e1d49d826ae3Notebook Studio 以 dsh-notebook-studio 发布，当前版本为 0.6.1。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
点击下方 Badge 复制 Markdown，粘贴到 README 即可。
验证 package.json 声明的 GitHub 仓库，即可管理这个公开页面。认领后，Hub 会优先安排当前版本的安全扫描，并在通过后公开展示结果。
认领这个 Plugin →English | 中文
Treat one dsh session as a literature project: import up to 30 text-based PDFs, choose which papers this run uses, search page-anchored evidence, and turn a prompt into a report (DOCX/PDF) or a deck (PPTX/PDF). Finished documents are saved automatically into the current session workspace and can be previewed as PDF inside Studio. Web background is optional and off by default; original figures, editable flowcharts and evidence tables are planned per section and per page, with AI concept illustrations an optional extra instead of one fixed image for every section.
This is a standalone Git repository; it does not contain dsh itself, your papers, or generated documents. Repository: https://github.com/HaoKuo/dsh-notebook-studio. The version history is in RELEASE_NOTES.md.
0.1.7-rc.2 (the development and acceptance baseline; dependencies are pinned to it), Node.js 24, Python 3.12/PyMuPDF. Text models can be any model already registered in dsh, or a standalone service compatible with OpenAI Chat Completions.STUDIO_CJK_FONT=/path/to/font.ttf. When no CJK glyphs can be found, the failure is explicit.@qithird/dsh-paste-input-plus@0.2.1 provides PDF upload from the chat input box. In-Studio file/folder upload does not depend on that plugin. An optional image service must return base64 PNG from images/generations.ctx.web.search and require the dsh deployment to configure a working search provider. When the service is unavailable, or no HTTPS result with an excerpt comes back, the task fails explicitly rather than quietly degrading into an offline document.Run from the repository root:
npm ci
python3 -m venv .venv
.venv/bin/python -m pip install -r requirements.txt
npm run check
dsh plugin --profile web add "$PWD"
mkdir -p "$HOME/.dsh/skills"
ln -sfn "$PWD/.agents/skills/research-studio" "$HOME/.dsh/skills/research-studio"
When dsh is not on PATH, use the absolute path to its executable instead. In this machine's web profile the plugin and the skill are already symlinked to this repository; do not let the install command stand in for confirmation. After installing or updating, the user schedules the restart of the running dsh web themselves so the Web client is rebundled. This task never stops or restarts a running dsh.
The Host installs git sources through pnpm, so no local clone is needed:
dsh plugin --profile web add github:HaoKuo/dsh-notebook-studio
# pin a release: github:HaoKuo/dsh-notebook-studio#v0.6.1
A git install ships no Python virtual environment, so PDF parsing resolves an interpreter in this order: STUDIO_PYTHON (which must be usable when set), the package's own .venv, then python3/python on PATH. The interpreter has to import pymupdf; when none qualifies, the error names every candidate it tried and how to fix it:
python3 -m venv ~/.venv-studio && ~/.venv-studio/bin/pip install "PyMuPDF>=1.26,<2"
export STUDIO_PYTHON=~/.venv-studio/bin/python # before starting dsh web
The plugin follows the dsh language setting (Settings → General → Language, zh/en). At load the Web client reads the active locale from the host locale service and fetches one catalog from the host (getMessages); catalog keys are the Chinese source strings, so a missing translation falls back to the Chinese text instead of an empty label. Switching the language re-renders an open workbench. The host-side studio_search text (tool description and the guidance returned to the model) follows an explicitly selected non-Chinese preference and otherwise stays in the source language.
v0.5.1 keeps a single sidebar entry: click NotebookStudio above "Plugins" on the left of dsh and the workbench opens in the main area while the host's left navigation stays. The session tab and the session-top Studio button are gone. Inside the workbench the library sits on the left, the body editor / PDF preview in the middle, and authoring settings on the right; both side columns collapse. The top bar gathers project, task progress, generation and export, while model configuration lives in the gear settings. Buttons, backgrounds, text, borders and status colors use dsh theme variables directly and follow the light/dark theme; the old standalone teal palette is gone. On narrow screens the layout stacks vertically with the content area first.
/v1/images/generations, leaving the API key empty for a local Qwen. A single image waits 600 seconds by default, configurable from 30 to 1800 seconds; raise it when a model is slow to cold-start.[Wn], URL and retrieval time, and cannot be passed off as [n] PDF p.x paper evidence.notebook-studio/<session-specific directory>/<report|deck-random suffix>/ in the current dsh session workspace; this is not the plugin repository directory. Re-typesetting saves into a new directory and never overwrites a copy the user has edited. Completed per-page deck drafts are cached for retry after a failure.
Illustrations are planned from the body: each report section may carry 0–2 visuals and each deck page one main visual area, with no fixed count for the whole document. Related results use original figures, procedures and mechanisms use flowcharts, multi-paper comparisons use tables, and abstract concepts may use AI illustrations; every image keeps the purpose and source of its section/page. Image generation runs serially; on failure the UI reports it and fills the slot with the flowchart for that section/page rather than passing it off as a paper result. Once the service recovers, re-typeset to retry image generation.This round is built on the public slot contract and the actual React runtime implementation of dsh 0.1.7-rc.2. The entry registers into sidebar.panellist (order -10, where the host's "Plugins" is 0), the workbench registers into main, reads the current session through the session-maybe sub-slot, and returns with layout.selectPanel(null). No host core layout, private route or simulated click is touched; navigation row styling and collapse hints are managed by the host. Isolated UI and host composition tests exist; the UI test loads the actual host navigation row component and the light/dark theme CSS.
Acceptance status on 0.1.7-rc.2: npm run check passes 40 JavaScript + 2 Python tests against the pinned 0.1.7-rc.2 packages; dsh --profile web --dump-config composes the profile with the plugin layer; and the running dsh web instance (0.1.7-rc.2) mounts the host side, where studio_search answers without error. Not yet verified: a full cloud generation with real papers under 0.1.7-rc.2, and deployment acceptance in a restarted instance that has picked up the new dependency tree — a restart is required for that, after which run one real report and one real deck. Host versions other than 0.1.7-rc.2 are outside this acceptance.
After an upgrade, opening Studio automatically back-fills the session's most recent completed documents into the workspace without rewriting the old content. To adopt the new content-planned illustrations, click generate report / full deck again; "Re-typeset" only reuses the saved body and visual plan. When the workspace is missing or not writable, internal documents, download and preview are kept and the UI explicitly reports the save failure.
src/: host, authenticated RPC, uploads, SQLite FTS5, model calls, DOCX/PPTX and standalone PDF typesetting; lib/client.js: the sidebar entry and the session-isolated Studio main panel; worker/: PyMuPDF parsing; tests/: synthetic material and client-registration contract regressions; .agents/skills/: the retrieval skill.~/.dsh/studio/v1/<session-specific directory>/; deliverable Office files, PDFs and source lists are additionally saved into the session workspace. Output paths come only from the host session cwd: neither the model nor the client can name arbitrary file paths, and symlinks in output subdirectories are rejected. Studio uploads in 512 KiB chunks through the authenticated dsh Connection; the server only handles staging files it created itself. The chat-box entry still verifies the session manifest and directory ownership. Downloads and previews are validated against record IDs belonging to the current session.Run npm run check to verify JS/Python; full cloud generation and real web retrieval with real papers have not been executed with permission. Older versions on this machine imported and extracted images from 12–30 real PDFs; this version's generation and web use are tested with synthetic material and mock interfaces, and never send those papers to the cloud automatically.
v0.4.0 split the checks for conclusion length, missing citations and excessive citations: the old code misreported a conclusion longer than 120 characters as a citation error. Retries hand the specific fields, constraints and previous JSON to the model for repair, and state which numbers are available on that page when refs are missing; fabricated citations are never inserted automatically and no unsupported document is accepted. Previously invalid per-page caches are regenerated. Regressions cover an over-long page-6 conclusion, a fix after omitted citations, and rejection of persistently fabricated citations.
v0.4.1 fixed the download route missing requestBody: 'buffered': in dsh 0.1.6-alpha.2 this made GET requests wrongly constructed with a request body and returned an empty HTTP 400. All six file routes now declare the correct mode and support HEAD preview checks; downloads check HTTP status, file size and transfer length first, and errors are no longer saved as empty files labelled "successful downloads".
This did not mean the server failed to generate the body, nor was it an image-model downgrade. After updating and reloading the plugin, download again from "Documents"; no new text-model call is needed. You can also open the original document in the workspace directly.
max-tokens means the model reached this call's output ceiling, not that PDF import failed. v0.3.1 disables unnecessary reasoning output for DeepSeek structured writing, raises the budgets for summaries, outlines, per-page drafts and reports, and retries a limited number of times with a higher budget when truncation is explicit; if it still fails, the task keeps the concrete stage and suggests a model with longer output or a narrower generation scope. After updating the plugin the user must restart the running dsh web themselves, then click generate again; completed per-paper summaries and per-page drafts continue to be reused.
MIT — see LICENSE.
studio_search in Chat for Q&A; answers should carry evidence IDs, paper names and PDF page numbers. On the left you can search evidence, view task records and retry failed papers.