void2anything/dsh-qingagent

dsh-qingagent

QingAgent inside DeepSeek Harness

A plugin built to the DSH spec: one command to install or remove, and you can write and review documents with QingAgent right inside DSH.

QingAgent repo · qingagent.com · Feature board · 中文

---

What is this

DeepSeek Harness (DSH) is DeepSeek's open-source "everything is a plugin" agent framework. Install dsh-qingagent and DSH gains a pen:

Describe what you want in the conversation, and the agent drafts, edits and submits reviews through the QingAgent engine. A paper panel grows on the right — rendering from the same source as the QingAgent desktop app: serif type, warm paper, square corners, with formulas, Mermaid, draw.io, tables, footnotes and a seal-stamped colophon.

The chat keeps a summary; the manuscript is the real thing.

What you can do

• Draft from one sentence: the agent writes the finished piece into the paper panel; while the page is still empty you get QingAgent's own "qing character diffusion" loading animation;
• Per-change verdicts: every AI edit is a candidate — walk through them on the page, accept or reject, commit to land. When the change ratio is high it switches to a full-document old/new comparison;
• Annotation carousel: review annotations sit in place on the page and can be flipped through one by one;
• Selection chips: select text on the page and it becomes a "selection" chip in the input box — multiple chips supported, hover for the full quote, and sent messages render them as chips too;
• Review and export live on the page: export to PDF / Word / HTML / Markdown / TXT and download directly;
• Several drafts per session: qing_list_docs / qing_focus_doc switch the preview on the right;
• "Open in QingAgent": a deep link launches the desktop client so you can keep editing — same local library.

In action: writing a piece with diagrams, tables and formulas inside DSH

Chat on the left, manuscript on the right — state the requirement once and the agent writes through the QingAgent engine while the paper panel fills in; Mermaid diagrams carry "visual edit / edit Mermaid" buttons and drawio drawings open on double-click	Full typesetting — tables, inline and block math (KaTeX), task lists and code blocks, rendered exactly as in the QingAgent desktop app
Per-change verdicts — the header shows "reviewing · N changes", additions and deletions are marked in the text, and the bottom bar offers previous / next / commit / discard all	Three-state onboarding — not installed, installed but not running, or handshake failed, each with the right guidance; once QingAgent is up the plugin recovers without restarting DSH

---

Relationship with the QingAgent repo

This plugin is the DSH frontend of QingAgent, not a standalone product.

Because of the plugin's own complexity — the paper rendering is deeply coupled with the manuscript engine, and documents and versions live in a local database — the QingAgent desktop client must be installed first for the plugin to work. The client hosts the engine and the local library; the plugin wires it into DSH's conversation and UI.

That coupling buys you something: a draft written in DSH can be opened and edited in the QingAgent client, and vice versa — one local library, not two copies.

---

Install (three steps)

① Start DeepSeek Harness (Node.js 20+)

npx @deepseek-ai/dsh web

② Install the QingAgent plugin

npx @deepseek-ai/dsh plugin --profile web add dsh-qingagent@latest

Restart dsh web afterwards.

③ Download and launch QingAgent once

Get it from qingagent.com or QingAgent Releases. Launching it once starts the engine and writes credentials to ~/.qingagent/instance.json; the plugin connects automatically.

If QingAgent is missing or not running, the panel shows an onboarding card: not installed → download instructions; installed but not running → a one-click "Launch QingAgent"; handshake failed → the specific reason (a corrupted instance.json, for example). Once QingAgent is up the plugin recovers on its own — no need to restart DSH.

Requirements: DSH 0.1.0-rc.6 as the baseline (peer dependency ^0.1.0-rc.6), with a profile that already composes the storage hub, storage-domain and a KV backend (usually @deepseek-ai/dsh-storage-json).

---

Supported setups

Officially supported: dsh running on Windows with the Windows QingAgent desktop client, same machine and same user.

The implementation also includes macOS client detection (mdfind with an Applications-directory fallback), but it has not been fully validated — issues welcome.

WSL and cross-OS setups are not supported: the plugin reads ~/.qingagent/instance.json from the current system's home directory, so dsh in WSL cannot see a client installed on Windows and cannot reach the engine.

---

Capabilities

Host tools (callable by the agent)

Tool	Parameters	Purpose
qing_write_draft	qingml required; title? requirements? docRef?	Directly submit the complete QingML authored by the main model. Length requirements only produce a status and gap; the tool never rewrites internally. Without docRef it creates a new draft; with a docRef bound to this session it rewrites the whole piece (requires explicit user authorisation)
qing_edit_draft	docRef?; ops[] required	Atomically submit a batch of structured local edits; refused while a review is in progress
qing_read_draft	docRef?; mode defaults to outline	Tiered reads: outline / full / base (committed baseline) / lines (numbered Markdown) / blocks (block ids)
qing_review_commit	docRef?; action: accept_all | reject_all	Accept or reject the pending review wholesale; hard-limited to one call per turn
qing_list_docs	scope: session | library	List drafts bound to this session; library lists recent documents from the whole QingAgent library (max 50)
qing_focus_doc	docRef required	Switch the paper panel; unbound documents can be adopted by engine id or an exact unique title

The eight qing_edit_draft operations:

kind	Semantics
strReplace	old → new, optionally the nth match
markText	Add / remove inline marks on matches: bold, italic, strike, underline, code, highlight(color), textColor(color), link(href,title?); supports all, isRegex, withinRef
insertAfterLine	Insert after line N of the committed Markdown
insertAfterBlock	Insert a block after a top-level block, or a sibling item after a list item
appendSection	Append a section
deleteBlock	Delete a whole top-level block
deleteListItem	Delete a list / task item; the engine cleans up the parent list after the last one
setTitle	Change the metadata title, leaving the body untouched

Request-level opId idempotency applies only to deleteBlock, deleteListItem and insertAfterBlock; every proposal also carries a random clientMutationId.

The paper panel (client)

QingAgent's web editor is compiled straight into the plugin from the vendor/qingagent submodule, so the look matches the desktop app:

• Per-change verdicts: DocumentSnapshotView takes the patch set; the PatchNav at the bottom handles previous/next, reject-all and settlement. After settlement a structured review-result message flows back only when something was rejected;
• Whole-document review: a high change ratio switches to old/new navigation with "apply new" / "revert to old";
• Annotation carousel: external annotations become the product's AnnotationGroup, decorated into ProseMirror and rendered with QingAgent's native carousel; annotations hide themselves while body patches are under review;
• Selection chips: the selected text plus block coordinates go through the bridge and become an input-box reference — multiple, de-duplicated, with full quotes on hover;
• Diagrams and export: double-click a draw.io block to open the offline editor and write back; export supports PDF / DOCX / HTML / Markdown / TXT;
• Deep link: qingjian://open?engineSessionId=<id> launches the desktop client.

Attribution: every external write from this plugin is tagged x-qa-client: deepseek and shows up as a "DeepSeek Harness" source inside QingAgent.

---

Connection & self-healing

1. Instance discovery: reads ~/.qingagent/instance.json for the current user, requires schemaVersion=2, and validates port / pid / version / attachProtocolVersion / token / startedAt;
2. Port authority: when an instance exists the configured engineUrl port is ignored and the plugin connects to http://127.0.0.1:<instance.port> (the desktop port defaults to 21823 and falls back to a random one when taken);
3. Handshake: checks the attach protocol and process liveness, then calls health with a Bearer token; on a 401 it re-reads the instance file and token once;
4. Four states: online / offline / starting / handshake-failed, each with a specific reason;
5. Backoff: retry intervals of 5s → 10s → 20s → 30s, then holding at 30s; back to a 5s health cadence once online;
6. Client detection: on Windows, the HKCU protocol registration plus HKCU/HKLM uninstall entries (including the /reg:64 view); on macOS, mdfind by bundle id with /Applications/青简.app and the user Applications directory as fallbacks. Detection results are cached for 30 seconds. The launch endpoint only accepts paths resolved and stat-ed by the detector — never a path submitted by the browser.

With autoLaunch and engineCommand configured, the wait budget for the engine to come up is 20 seconds.

---

Configuration

Field	Default	Notes
engineUrl	http://127.0.0.1:8080	Fallback only — used when no instance.json can be read; a live instance's port wins
engineCommand / engineCwd	unset	Optional launch command and working directory, executed only with autoLaunch
autoLaunch	false	Starts the engine detached when offline; removing the plugin never kills your engine
workspaceProjection	true	Reserved field — no runtime effect today

---

Security boundaries

• The token never reaches the browser: instance.json and its token are read only by the Node host, which attaches the Bearer header to health checks, external API calls and export requests. The bridge payload sent to the browser carries engine status, bindings, documents and selections — no token.
• The bridge is loopback-only: /qingagent-bridge/* and /drawio reject non-loopback addresses before any business logic (IPv4, IPv6 and IPv4-mapped loopback accepted).
• Session isolation: document reads/writes, assets, exports and reviews are all checked against the dshSessionId + engineSessionId binding. focus with adopt:true is the explicit adoption exception, which probes the engine document before joining it to the session.
• Style isolation: dynamically imported vendor CSS is wrapped in @scope, while the mechanically extracted qingdoc.css is rewritten with a [data-qingagent-doc-panel] selector prefix — neither leaks into the host UI.
• QingML rendering: production rendering goes through QingAgent's qingmlParse with an explicit tag allowlist, allowlists for links and images, and a final Zod schema check before anything enters ProseMirror; script and style are dropped.
• draw.io assets: GET/HEAD only, directory traversal blocked, CSP and SAMEORIGIN on HTML; iframe messages validate both event.source and a same-origin origin.

Binding data lives in the dsh_qingagent v1 domain of @deepseek-ai/dsh-storage-domain.

---

Developing from source

git clone --recursive https://github.com/void2anything/dsh-qingagent.git
cd dsh-qingagent
npm install
npm run check   # CSS pinning check + typecheck + tests + build

# POSIX shell
npx @deepseek-ai/dsh plugin --profile web add link:$(pwd)
# Windows PowerShell
npx @deepseek-ai/dsh plugin --profile web add link:${PWD}

Forgot --recursive? Run git submodule update --init. The build scripts use POSIX tools (rm -rf and friends), so on Windows run development builds in Git Bash or WSL.

The dsh.bundle.patch entry in package.json merges the in-repo cordis.patch.yml; do not keep a hand-written mount alongside the bundle mount or the plugin registers twice.

Build-time dependency: the vendor/qingagent submodule

The paper rendering reuses QingAgent's apps/web sources and CSS, read at build time from QING_ROOT:

• default vendor/qingagent (a submodule pinned to a verified commit);
• QING_ROOT=/path/to/qingagent overrides it (point it at your own QingAgent worktree during development);
• the offline draw.io runtime is published from there too, with QINGAGENT_DRAWIO_ROOT as a separate override.

CSS is extracted mechanically by "file + line range" (scripts/extract-qingdoc-css.mjs). npm run check:qingdoc-css does a byte-level comparison and is wired into check and prepack: run it right after upgrading the submodule — drifting line numbers cut the extraction in the wrong place and produce a broken build. A red check means do not publish; realign the pinned ranges, then run the full check.

Tests

npm run check   # everything: CSS pinning + typecheck + vitest + build
npm test        # unit tests only

Contract tests lock down: the 800px page width, serif type, square corners and warm palette applying only to the panel root; CSS extraction matching the pinned ranges; bridge loopback and session isolation; the QingML XSS allowlist; review-state interception and 401 token re-reads.

---

Community

Scan to join the WeChat user group — report problems, request features, follow updates:

---

Contact

• Usage questions and bugs: GitHub Issues
• Feature requests and upvotes: Feature board · DSH plugin — the most requested ones get built first
• Issues with QingAgent itself: QingAgent Issues | Feature board · desktop client

---

License

Apache-2.0 for this repository. The vendor/qingagent submodule is QingAgent, MIT.