DeepSeek Harness Plugin Hub

发布与管理完整 Harness Profiles,发现适合你的插件。

探索

插件目录环境预设文档中心动态

社区

发布插件联系我们报告问题

相关链接

Plugin Hub GitHubDeepSeek Harness 官方项目系统状态隐私说明
© 2026 DeepSeek Harness Plugin HubPowered byPaxTech

独立、非官方社区项目,与 DeepSeek 官方无隶属、授权或背书关系。

Pharma Product Facts — DeepSeek Harness 插件(DSH Plugin)
DeepSeek Harness Plugin Hub
ProfilesPlugins分类动态文档登录管理 Profiles
ProfilesPlugins分类动态文档登录
← Plugins
P

dsh-pharma-product-facts

Pharma Product Facts

用于经核实药品产品事实的自包含 DSH 捆绑包,包含其技能、工具和软路由

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:Alberssssss/dsh-pharma-product-facts-plugin#c95d6dfef71625abfa03beeee787b2c9b90eea6c
README兼容性版本

兼容性与来源证明

Pharma Product Facts 以 dsh-pharma-product-facts 发布,当前版本为 0.2.2。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。

DSH 兼容范围
*
运行环境
any
发布来源
github
Registry 更新时间
2026/8/24

版本

0.2.2stable
2026/8/24
0.2.1stable
2026/8/21

相关插件

正在加载相关插件…

最新版
0.2.2
DSH
*
HMR
重启进程
Tree shaking
未声明可安全裁剪
解包体积
未提供
文件数
未提供
Surface
any
许可证
MIT
发布源
github
GitHub
★ 0
周下载
0
最近提交
2026/8/24
查看源码 ↗
README Badge

点击下方 Badge 复制 Markdown,粘贴到 README 即可。

这是你的 Plugin?认领权益 · 优先安全扫描

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 →
报告问题

相关插件

继续浏览 search-research 分类下经过校验的插件。

Weknora@wxg-prc-cpg/dsh-weknora适用于 DeepSeek Harness (dsh) 的 WeKnora 知识检索工具:通过自有知识库进行语义搜索、文档阅读以及 RAG/代理回答。Anysearch Dsh@anysearch/anysearch-dsh适用于 DeepSeek Harness 的 AnySearch 网络搜索和获取提供程序及高级工具Research Reportdsh-research-report适用于 DeepSeek Harness 的可验证研究报告引擎:基于内容寻址的证据账本(声明↔快照绑定,可检测篡改),以及带版本的封存报告,其中每项声明都带有验证结论,清单哈希用于封存目录Industry Researchdsh-industry-research适用于 DeepSeek Harness 的行业与公司研究领域包:方法论技能、行业链结构模型(industry_map)、通过 ctx.web 进行的公开来源政策/新闻跟踪(industry_track)、公司扫描卡片(company_scan)以及可审计的研究

README

dsh-pharma-product-facts

English | 中文

Standalone DSH bundle containing one pharma-product-facts skill provider, its narrow agent/pre-step router, and two DSH-native evidence tools. One Cordis row owns all four contributions, so installation, disablement, reload, and removal are atomic.

The runtime workflow is self-contained inside DSH. It uses the shipped web_search tool for URL discovery, a package-owned restricted CDE/NMPA HTML/PDF retriever, and a package-owned deterministic answer finalizer. It does not execute external medical skills, local Python scripts, or user-home resources.

Install

In the DSH Web UI plugin field, paste only:

github:Alberssssss/dsh-pharma-product-facts-plugin

From a terminal:

dsh plugin --profile web add github:Alberssssss/dsh-pharma-product-facts-plugin

Restart the web profile Host and create a new session. Existing sessions retain the composition and skill body with which they were created.

This repository commits verified prebuilt lib/ files and declares no install-time scripts, so a clean DSH profile can install the Git package without a pnpm allowBuilds exception. Source checkouts remain independently buildable with pnpm run build. Pin a reviewed commit when appropriate:

github:Alberssssss/dsh-pharma-product-facts-plugin#<commit>

DSH may report peer packages as absent from the profile directory because the Host supplies its core packages through the installation-owned module fallback. Do not install a second Cordis or DSH core copy merely to suppress that warning. Verify the resolved row with:

dsh --profile web --dump-config

Model visibility

Use a preset with @deepseek-ai/dsh-tool-skill, such as standard or code. The shared skill tool loads this package with skill({ name: "pharma-product-facts" }); there is no tool named after the GitHub repository.

Loading the plugin also registers these model-visible tools:

  • pharma_product_facts_fetch_source accepts only HTTPS URLs on CDE/NMPA hosts, follows only same-origin redirects, bounds response size and time, extracts HTML/text/PDF content, and requires the exact requested product identity to occur in the document.
  • pharma_product_facts_finalize accepts exact quotations and same-session evidence ids, rejects cross-session or altered evidence, derives source URLs itself, and renders the canonical public answer.

The skill uses DSH web_search only to discover candidate official URLs. Search snippets are not accepted as label evidence. The standard DSH base profile already exposes web_search; its configured search provider must be usable for live discovery.

The packaged workflow caps each turn at three web_search calls and four official-source candidates, counting timeouts and rejections; a lower user-supplied cap wins. Once search returns an official candidate, the skill validates it before spending another search call.

hcp_focus_card normally accepts only clinical_focus. To prevent a redundant model field from causing a long retry loop, the finalizer also tolerates a listed label_boundary only when its evidence id and exact scope quotation duplicate one focus item. The duplicate is validated but not rendered; not_listed, unrelated boundaries, and other mixed-mode inputs still fail. A mode-field correction reuses existing evidence and does not restart search or source retrieval.

For non-boundary modes, a tool-schema placeholder is treated as absent only when questioned_use, scope_quote, and evidence_id are all empty; partially populated objects still fail closed. PDF matching ignores CJK layout line wraps but preserves English word spacing and number/unit spacing. Canonical answers have no trailing newline so Web chat can reproduce them byte-for-byte.

Optional OpenAI search companion

This bundle does not package or select a web-search backend. It consumes the stable DSH web_search tool and works with any usable configured provider. To route that tool through OpenAI Responses native Web Search, install the companion bundle:

dsh plugin --profile web add github:Alberssssss/dsh-web-search-openai

The companion changes discovery only. This pharma bundle still fetches the cited CDE/NMPA page or PDF itself, checks the requested product identity and exact quotations, and owns the final answer. Installing either bundle does not force the conversation model to call the skill or web_search.

Runtime safety

  • Evidence is isolated by DSH agent/session and held only in a bounded in-memory store.
  • The retriever sends no cookies or credentials and rejects non-regulator hosts, HTTP, embedded credentials, non-default ports, cross-origin redirects, unsupported media, oversized responses, and product-identity mismatches.
  • PDF text extraction is bundled through the maintained unpdf dependency; no external document service is needed.
  • The finalizer verifies every public quotation against fetched text and prevents a truncated document from supporting an “absence” conclusion.
  • Public output rejects local paths, credential-like strings, tool names, and execution narration.

The package does not make the network or public regulator sites infallible. When no complete matching official source can be obtained, the skill returns a bounded “not verified” answer instead of filling facts from memory.

Configuration

All deployment-varying resource limits are fields on the same plugin row. Override the bundle row by id and restate its name:

- id: pharma-product-facts
  name: dsh-pharma-product-facts
  config:
    fetchTimeoutMs: 30000
    sourceToolTimeoutMs: 35000
    maxResponseBytes: 12000000
    maxSourceChars: 180000

The remaining defaults are maxUrlChars: 4096, maxRedirects: 3, maxEvidenceScopes: 64, and maxEvidenceRecordsPerScope: 24. userAgent is also configurable for deployments that require an operator-specific public identifier. Invalid numbers, a source-tool timeout shorter than the fetch timeout, and unsafe User-Agent text fail while the plugin loads. The HTTPS requirement and CDE/NMPA hostname allowlist are fixed security rules and cannot be relaxed through configuration.

Routing boundary

The router recommends this skill for product identity, static label facts, label safety fields, approval boundaries, and non-individualized HCP focus cards. It excludes patient-specific dosing or administration, eGFR/renal-function dose decisions, adverse-event management, combinations, competitor comparisons, commercial promotion, registration-document download tasks, and evidence reviews. The skill catalog description carries the same exclusions so an out-of-scope product mention is less likely to trigger model-initiated loading. A router hint is not proof of skill use; inspect the later skill call and evidence-tool calls in the session log.

Lifecycle

Disable every contribution together:

- id: pharma-product-facts
  disabled: true

Remove the installed bundle:

dsh plugin --profile web remove dsh-pharma-product-facts

Development and verification

The Git repository is independent: package versions are ordinary registry ranges, there are no workspace: dependencies, and pnpm run build builds from this checkout alone. Git installs use the committed lib/ artifacts and execute no dependency lifecycle script.

pnpm install
pnpm run typecheck
pnpm run test:coverage
pnpm run build
pnpm run pack:check

The package remains an external plugin rather than an official DeepSeek Harness release component. Medical correctness still depends on the exact public source selected and quoted; deterministic validation reduces source substitution and transcript leakage but does not replace clinical or regulatory review.