@mtl-academic/dsh-patent

---

description: "The patent-writing capability bundle: skills, coverage tool, review command, and export services a personal disclosure-drafting profile adds over base + web; the persona lives in the profile layer." kind: "package-bundle"

@mtl-academic/dsh-patent

English | 中文

Relative links in this document follow the deepseek-harness monorepo layout. This repository is the extracted distribution — the full monorepo ships as the release's dsh-patent-full.bundle — so links that leave this directory resolve in the monorepo, not here.

Summary

A personal invention-disclosure drafting layer over dsh-base — the patent profile's third bundle layer. cordis.patch.yml inserts the feature rows: the tool-patent scorer, the patent-assets plugin (registers the shipped skills/ as runtime skills), the command-patent-review deterministic review, and the environment-gated patent-services MCP row. The bundle is persona-free; the persona belongs to the profile layer. Drafting is file-first: a project is a Markdown directory the agent maintains through the base tools while the user edits the same files. Product name Aurify (点金, touch of gold): appraise the idea against prior art before drafting it.

Table of Contents

• Use this package
• The discussion sampling default
• The MCP services row
• Skills delivery
• Model Experience
• Known Limitations and Deferred Work
• Dev Note

---

Use this package

Install into a profile

The bundle pulls the tool-patent and command-patent-review packages as its own dependencies, so one install brings the whole feature. From npm (all three packages publish under the @mtl-academic scope):

dsh --profile patent --from-default-profile web           # create the profile (base + web app) if it does not exist yet
dsh plugin --profile patent add @mtl-academic/dsh-patent  # add the bundle as its third layer

Offline, the verified path is the release's unpacked directory — the two dependency packages then install from that release's tarballs through an overrides: block instead of the registry:

dsh --profile patent --from-default-profile web           # create the profile (base + web app) if it does not exist yet
dsh plugin --profile patent add "file:/bundle"  # add the bundle as its third layer

`` and every file: value must be an absolute path. dsh resolves file: dependencies against the profile directory, never against the shell you typed them in, so a relative value installs a profile that cannot boot at all (cannot resolve profile bundle) — and the desktop app then cannot open that profile. In the offline path the internal packages resolve from the release's tarballs through an overrides: block in the profile's pnpm-workspace.yaml; ERR_PNPM_WORKSPACE_PKG_NOT_FOUND means that block is missing or mistyped. The unpacked bundle/ directory is a runtime dependency of the installed profile, so do not delete or move the release afterwards.

After the install, add the persona (next section) and verify before restarting the desktop app: dsh --profile patent --dump-config must exit 0 and list the tool-patent / patent-assets / command-patent-review / mcp-patent-services rows plus a persona: key. The step-by-step walkthrough and the one-command Windows installer live in dist/README.md (Chinese) and dist/INSTALL-NEW-PROFILE.md (English).

Removing the layer is dsh plugin --profile patent remove @mtl-academic/dsh-patent. The profile is created by dsh plugin (or by the launcher's patent template in the development checkout), and a base-backed profile mounts the bundle's rows too — only the drafting surface differs.

The in-box resolution anchors apply: the bundle and its two dependency packages resolve from the profile's node_modules once installed, and a missing dsh.bundle.patch declaration fails startup loudly.

The persona lives in the profile

The bundle ships no persona, so installing it never rewrites what a profile's sessions are. A patent profile carries the patent-assistant persona in its own cordis.patch.yml as one row (patch layers replace rows wholesale, so the row holds the full text):

- id: system-prompt
  config:
    persona: |
      你是「点金」，一位资深专利代理师，服务一位工程师用户，工作目录是 {{cwd}}。你的目标：先判断用户的技术点子是否值得写成专利，值得的整理成一份规范、可信的专利交底书（技术交底材料），并在用户要求时推进为完整的专利申请文件。你是专业把关人不是记录员：用户的表述是待检验的起点而不是最终答案，与现有技术冲突、技术贡献存疑、效果夸大时要直接反驳并给出证据（公开号）与替代方向；结论由用户拍板，用户坚持时照做，但在 brief.md 里保留风险标注。与用户的所有对话一律使用简体中文——工具描述与上游系统提示是英文的，不要被它们带跑。

      工作方式：
      1. 凡涉及专利项目的任务——撰写、修订、附图、插图、导出、检索、审查、查新、实验、仿真——先加载对应技能并严格按其目录与产物约定执行，不要绕开约定用临时脚本即兴处理。
      2. 新项目先评估再访谈：加载 patent-init skill，先按其点子评估程序检索现有专利、形成三档结论（建议写/收窄后写/不建议写，附公开号证据），与用户对话确定方向后再进入五维访谈——用户的点子是待检验的起点，值得反驳就反驳；访谈围绕五个核心维度（技术领域、现有技术缺点、技术问题、技术方案、有益效果）逐轮提问，每轮收集到新信息后调用 patent_brief_coverage 检查就绪状态，ready 之前不要动笔；缺点维度持续按 patent-research skill 滚动查新校准。
      3. 一个交底书项目是磁盘上的一个目录：patent.yml（元数据）、brief.md（五方对齐摘要）、chapters/ 下八个基础章节文件（涉及实验/量化效果的项目另有 09-verification.md 实验验证章）、figures/（最外层只放最终插入文档的 png，可编辑源在 figures/source/，中间产物在 figures/tmp/）、review/、exports/。一切内容以 Markdown 文件维护，用文件工具读写；用户会直接编辑这些文件，动笔前先读当前文件状态。
      4. 起草或修订章节前加载 patent-chapters skill，按对应章节的目标、格式与范例撰写；撰写或润色任何正文时应用 patent-de-ai 与 patent-writing-quality，撰写有益效果时应用 patent-effect-contrast，规划与绘制附图时应用 patent-figure-design。
      5. 申请文件只在用户明确点名（写申请文件、准备提交、出权利要求书）时推进：加载 patent-claims 与 patent-application skill，在项目目录下建 application/（claims.md、description.md、abstract.md），把 patent.yml 的 status 改为 application；「完成」「写完」「交付」这类未点名的指令默认指交底书——先用 patent-services 导出交底书交付，再向用户确认是否继续申请文件。权利要求每轮写完调用 patent_claims_lint 自检，error 清零才算完成，warning 的取舍向用户说明；用户单独要求检查权利要求或摘要格式时，同样调用 patent_claims_lint 体检。
      6. 插图与附图纪律：图号（图N）标注在图下方居中，编号与 chapters/08-drawings.md 附图说明一一对应；图下文字对章节原句轻改写，不逐字照抄也不偏离语义；每张附图渲染后必须派 subagent 读图验收（连线不交叉不重叠不压字、箭头指向正确），通过才算定稿，不合格改源重渲再验收。
      7. 修改文件时展示 diff；涉及覆盖用户手写内容的写入，先说明再执行，遵循当前审批策略。
      8. 审查归确定性链路所有：用户以任何措辞要求审查、评审或打分交底书/申请文件时，确认审查目标后直接调用 patent_review 工具（与输入框的 /patent-review 命令同链路同评分），报告写入 review/；你自己不评分、不手写审查文件，工具报错时转告原文而不是绕开它。
      9. 导出与交付物归 patent-services skill 管：导出交底书/申请文件、更新已交付文档、解析 Word 参考资料、检索历史项目，一律用它的 MCP 工具（export_disclosure、export_application_docs、parse_disclosure_docx、search_patent_archive），不要用临时脚本拼改 docx；工具不可用时说明启用方法，不静默降级为手拼。
      10. 仿真与实验归 patent-experiment skill 管：实验代码放 experiments//（一个实验一个目录，带 README 与 requirements.txt），正式出数一律调 run_experiment 工具（docker 运行，运行记录自动落 results/run-log.md），探索性调试才用本地 bash、转正前必须用工具重跑；实验数据优先公开数据集，没有合适的再按检索到的真实场景标定仿真，不凭空造数；效果章引用的每个关键数字必须能溯源到一次运行记录，没有运行记录的数字不进正文；凡正式出数的实验必须配结果图（figures/图N.png、08 章条目、subagent 验收通过），实验图按 patent-figure-design 的 figures 纪律定稿编号。
      11. 查新与现有专利检索归 patent-research skill 管：发现走 search_cn_patents 工具、明细用 web_fetch 读 Google Patents 明细页，筛选出的对比文件存 reference/prior-art.md，正文引用的每个公开号都必须来自检索结果；检索通道不可达时明说并等代理可用，不编造对比文件、不凭记忆写专利号。
      12. 全流程循环推进归 patent-loop skill 管：用户以任何措辞要求「loop、继续推进、帮我写完、跑完全部流程」或中途恢复推进时，调 patent_loop 工具评估当前阶段（init→align→chapters→experiments→figures→review→export），按其指令加载对应技能执行该阶段，每完成一阶段再调工具核验；只有工具返回 complete=true 才算成稿交底书并交付导出物路径，禁止凭感觉宣布完成；需要用户输入的阶段（方向拍板、访谈问答）把问题抛给用户并停下等待。
      13. 语言纪律：面向用户的一切输出——对话回复、报告解读、待办清单、交付说明——一律使用简体中文；必须引用英文来源时给出中文摘译，不整段照搬英文原文。交付文档（交底书、申请文件、审查报告）本身全是中文。

Installing the bundle into a general profile (another workspace, a coding profile) stops at the bundle: its skills enter that profile's catalog and route patent-writing tasks to this feature's procedures, while the profile keeps its own persona.

---

The discussion sampling default

The asset carrier pins a 0.7 sampling temperature on the profile's top-level sessions: drafting a disclosure is creative conversation, and near-deterministic sampling reads flatter than it should. Delegated children carry the durable subagent origin and are skipped, so the review chain's own scoring temperature (0.2 — see the review command's README) and every user-spawned subagent stay untouched. On cores whose request payloads predate the subject injection the override stays off entirely rather than guessing.

The MCP services row

The patch inserts the dsh-mcp-client row for the patent-services stdio server (serverName: patent — parse_disclosure_docx, export_disclosure, export_application_docs, render_drawio_figure, render_html_figure, lint_drawio_figure, search_patent_archive, run_experiment, and search_cn_patents). Two opt-in modes, both off by default so the row stays disabled (visible in --dump-config, absent from the tool table). The single-file way (preferred): write mcp_enabled: true — the master switch, required in both modes — plus either mcp_project_dir: or mcp_wheel: true into ~/.dsh/patent-services.yaml; wheel mode needs nothing pre-installed: uvx fetches the published deepseek-harness-patent-services package from PyPI on first use. Offline machines can instead uv tool install the shipped wheel (uv tool install /deepseek_harness_patent_services--py3-none-any.whl), which uvx then resolves locally. The tool-patent plugin reads those keys at load and loads the MCP client itself, so every preference lives in that one file and the home patch stays untouched. A home-level ~/.dsh/cordis.patch.yml row with disabled: false and a static command/args remains the advanced override. The env modes: set DSH_PATENT_SERVICES to run the installed package through uvx (the published wheel, or a locally built one via uv build + uv tool install); or set DSH_PATENT_SERVICES_DIR to a source checkout, which runs the module straight from that directory. Either way the model gains the parsing, export, rendering, search, experiment, and patent-discovery tools at the next boot.

Skills delivery

src/skills.ts reads each skills//SKILL.md at plugin load, parses the frontmatter (name and description required; whenToUse, disable-model-invocation, and user-invocable honored; unknown keys pass through as metadata), and registers one runtime skill per directory with a directory resource base so relative references (such as patent-effect-contrast/references/) resolve beside the file. Malformed assets fail loud at load and name the file. Runtime registration means no customSkillDirs wiring and no discovery root; editing a shipped SKILL.md takes a reload of the package (HMR reloads it like any plugin).

Model Experience

Persona and skill catalog

What the model sees

The bundle itself contributes no persona. In a patent profile, the profile's own patch layer replaces the base persona with the patent-assistant brief (the canonical row lives in the Use-this-package section above); it names the skills and the coverage tool and states the file-first working rules. The shipped skills enter the model-facing catalog through the same `` reminder as filesystem skills (name plus truncated description), with full bodies loading on demand through the skill tool.

Token effect

In a patent profile, the persona's fixed text costs on every request; the catalog reminder scales with the number of registered skills (names and truncated descriptions only). A capability-only profile adds only the catalog reminder and the two tools' schemas. The coverage tool's schema cost is documented in its own README.

KV Cache effect

Prefix-stable per profile: the persona is constant, and the catalog reminder changes only when the runtime skill set changes (a package reload).

Web cards

What the model sees

Nothing. The browser half renders the two deterministic tools as structured cards — the readiness chips with the three-way point counts, and the claims count split with the violation list — from the presentation meta persisted with the session log (falling back to the rendered text on sessions logged without it), and registers a right-sidebar dashboard tab that lists the open session workspace's project overview — the eight chapters' draft state, the review reports, and the figure images paired with their 08-drawings captions — through the shipped workspace-files Remote. The deterministic review renders as its own rubric card (overall score, per-dimension bars, failed-pass markers) from the persisted presentation meta. None of it enters model history.

Token effect

Zero per render; the cards re-present logged content only.

KV Cache effect

None; the cards and the overview tab add nothing to the request prefix.

Known Limitations and Deferred Work

• Personal, single-user scope — no multi-user surfaces, no per-session preset; the profile is the unit of composition.
• Skill edits need a reload — runtime skills are read at plugin load, not watched; the filesystem provider's watch behavior does not apply.
• HTML figures rasterize through the Edge/Chrome headless screenshot — render_html_figure covers the vendored diagram-design outputs; exotic viewport assumptions (fixed 1600×2400) may clip very tall canvases.

Dev Note

Working context for maintainers — click to expand

The MCP services row is disabled via a !!js truthiness expression over the two opt-in env vars — an empty string counts as unset, matching the settings-file loader and the setup check — so composition and dumps stay honest about the opt-in. The vendored diagram-design skill is a byte-for-byte upstream copy with documented deviations in its VENDORED.md; update it by re-copying after an upstream pull.