dsh-cited
搜索 Cited 安装中的文档，并根据文档内容回答问题，同时提供引用。
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:RonnieGex/dsh-cited#5df1925c2a3762c7623d0b04d25a85bdf4e5be02Cited 以 dsh-cited 发布，当前版本为 0.1.0。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。
Cited inside DeepSeek Harness. Search the documents of a Cited installation and answer from them, with numbered citations, from any agent of the harness.
Cited already speaks the Model Context Protocol over POST <url>/api/mcp. This plugin is the native alternative: one
step to install, one card to configure, and two tools (cited_search and cited_ask) that behave like any other DSH
tool — no MCP client row in cordis.yml.
In DeepSeek Harness, open Plugins → Add plugin and paste either of these:
github:RonnieGex/dsh-cited
https://github.com/RonnieGex/dsh-cited
The bundle ships its compiled lib/ in the repository, so the installation compiles nothing and asks for no
allowBuilds permission. The card Cited appears with the description of the package.
Open the card's configuration and fill in two fields:
| Field | What it is |
|---|---|
url | The address of your Cited installation, like https://cited.example.com. The plugin calls <url>/api/mcp. |
token | The CITED_MCP_TOKEN of that installation. The field is marked as a secret, so the harness never sends it back to the browser. |
timeoutMs | How long one call may take. 30 seconds by default. |
An installation without a token has its MCP endpoint off. Start Cited with a long random value and give that same value to this plugin:
openssl rand -base64 32
CITED_MCP_TOKEN=<that value> npm start
Both fields start empty on purpose: an invalid configuration would fail the load of the plugin, and you install before you configure. Until you fill them in, the two tools answer an error that names the missing field and nothing else breaks.
cited_search{ query: string, limit?: 1..8 } → the passages of the documents that match, best first, each with its document, its
section, its position and its text. It never writes an answer and never calls a language model, so a search spends none
of your balance. The model receives the passages numbered and cites them as [1].
Search the price list of the business.
cited_ask{ question: string, sessionId?: string } → answered with the answer and its numbered citations, or refused with
the honest sentence when the documents do not hold the answer. With a sessionId the turn is stored like any other
conversation of the installation.
Ask Cited how much a bicycle tune-up costs and where you got that from.
Authorization: Bearer header of a request to the url you configured. The plugin
opens no other connection.401, a 403, a 404, a timeout and an unreachable host become short sentences the model can read. None of them
contains the token.The plugin is plain ESM JavaScript. src/ is the source and lib/ is the compiled artifact that ships in the
repository:
node scripts/link-host-deps.mjs # links the host packages of your DSH installation
npm run build # src/ -> lib/
npm test # unit tests and a real Loader composition
npm run gate # every claim of the contract, ending in GATE: GREEN
link-host-deps.mjs uses the DeepSeek Harness installation at $DSH_INSTALL, at ../deepseek-harness, or at
../../deepseek-harness. The tests import the host copy of @deepseek-ai/dsh-tools and @deepseek-ai/schemastery, so
they exercise the same revision your harness runs.
The gate also installs the plugin from its local path into a temporary DSH_HOME (never your own), boots the
headless profile there, reads the plugin card back from the plugin manager, and calls cited_search against a Cited
installation served from a sibling worktree on port 3231. Set CITED_REPO to that worktree and DSH_INSTALL to your
harness when the two do not sit where the defaults look for them.
Verified on Windows on 2026-10-09 against a local Cited with its sample documents:
| DeepSeek Harness | Result |
|---|---|
Desktop app 0.2.0-rc.2 (its bundled dsh command) | dsh plugin add github:RonnieGex/dsh-cited installs in 4 seconds with no build step and no version exemption, the profile composes the dsh-cited layer, and a run with the DeepSeek model called cited_search and answered "a bike tune-up costs 380 pesos [1]" citing cafe-la-horquilla.md |
| 0.1.6-alpha.2 built from source | installs, composes and lists the card; cited_search returns the passages |
The unit tests and a composition through the real Cordis loader run on every change. Not verified yet: the same install
clicked through the desktop app's Plugins page (it runs the same plugin manager as the command), and cited_ask
through the plugin with a chat model.
Apache-2.0. See LICENSE and NOTICE. Built by Katalis.