dsh-teaching-material-workbenchv0.2.4
golden2002/dsh-teaching-material-workbench
dsh plugin --profile web add github:golden2002/dsh-teaching-material-workbench#2ed68442360001717e191df44dc0606cecbc1466
介绍
Evidence-gated teaching-material production plugin for DeepSeek Harness: layered gates, claim grounding, deterministic outline, rights audit, reproducibility.
标签
概述
dsh-teaching-material-workbench
Evidence-gated teaching-material production for DeepSeek Harness. It scaffolds a case, binds primary sources, runs layered gates, audits rights, and records artifact hashes. The core is subject- and setting-agnostic; subject content lives only in the case directory.
按证据链生产教学物料的 DeepSeek Harness 插件:建立案例、绑定一手材料、跑三层门禁、审计版权、记录产物哈希。核心与学科、学段、场景无关,学科内容只放在案例目录里。
What it adds
| Tool | Purpose |
|---|---|
teaching_material_init | Create a case scaffold: context profile, lesson data, source and claim registers, alignment, rubric, timeline, glossary, misconceptions, palette, teacher pack, verification templates. |
teaching_material_gate | Run E1 (self-consistency), E2 (external standards) and E3 (empirical evidence) gates; write gate-report.md. |
teaching_material_ground | Claim-level grounding audit: each claim needs a source and a locator; unsupported claims fail; the teacher pack needs [Sxx] markers. Writes verification/grounding-report.md. |
teaching_material_outline | Generate a deterministic slide-by-slide Markdown outline from the case data; writes build/outline.md. |
teaching_material_audit | Audit the source register and rights; write verification/rights-report.md. |
teaching_material_repro | SHA-256 every build artifact; write verification/repro.json. |
teaching_material_standards | Return the external standards register with an honest verified / partial / not-obtained status. |
A runtime skill teaching-material-workbench teaches the agent the workflow and the evidence rules.
Install
The plugin is a Host-only bundle. Do not edit the profile files by hand.
# From the Web UI: Plugins → Install → point at the package directory
# Or from an agent session:
plugin_manager action=install_bundle target=<absolute package directory>
CLI equivalent: dsh plugin --profile <name> add <absolute package directory>.
From the public repository (commit-pinned):
plugin_manager action=install_bundle target=git+https://github.com/Golden2002/dsh-teaching-material-workbench.git#7b805288da07a1985157948f35feb940179bb384
Repository: . Release tarball and checksum are under release/.
Requirements: DeepSeek Harness with Host tools service (current profiles), Node.js ≥ 20. The package declares engines.dsh >=0.2.2-rc.1, has no install scripts, no build step, no runtime dependencies, and makes no network requests.
Quick start
- Ask the agent to create a case:
teaching_material_init directory=D:workcase-001 title=... subject=... level=... audience=... setting=...
- Fill
profile.yml(authorities, thresholds) and the CSV files. Every source needs publisher, date, locator, access, licence and verification status. Secondary material may only locate primary material.claims.csvcarries one row per substantive claim (claim,source_id,locator,supported|partial|unsupported); optionalknowledge.csvmaps key concepts to the same source and locator. - Run
teaching_material_ground; remove or fix unsupported claims. - Generate
build/outline.mdwithteaching_material_outline, then producebuild/deck.pptx/build/handout.docxwith any tool and writeteacher-pack.mdwith[Sxx]citations. - Run
teaching_material_gate. Fix the data, never loosen the gate. - Run
teaching_material_auditandteaching_material_repro. - Add human review and pilot evidence when available: write
verification/trial-design.mdbefore fillingpilot.csv, and keep both marked "未执行" when not done. An empty pilot template fails the gate; pilot data without a written design also fails.
Gate layers
| Layer | Meaning | Examples |
|---|---|---|
| E1 | Self-consistency: structure, arithmetic, cross-file references | schema, timeline sum, rubric total, source ids exist |
| E2 | External standards and documents | source register rules, official mark scheme, WCAG 2.2 SC 1.1.1/1.4.3, GB/T 15834, licence terms |
| E3 | External empirical evidence | independent peer review, learner pilot data |
Each gate carries its layer and its basis. A skip means the check did not run; it never counts as a pass.
Verification
node --test tests/*.test.js
Scope and limits
- Data-level gates. Typography, render and overflow checks for the generated deck/handout still need the material tool (for example LibreOffice rendering) and are not claimed here.
- E3 gates only check that evidence exists; they do not judge its quality.
- No network, no telemetry, no student data collection. Case data stays in the directory you choose.
- The standards register states
not-obtainedwhere the text could not be obtained; do not cite those clauses.
Community package compliance
See docs/COMPLIANCE.md and community/README.md. The package follows the Host bundle conventions: dsh.bundle.patch, exported locale files with meta.title / meta.description in English and Chinese, an original SVG icon inside the package, an explicit files list, engines.dsh, no install scripts, and MIT licence.
Further reading
- docs/BENCHMARK.md — capability comparison and the four measurable indicators, with current values.
- docs/TSINGHUA-INTEGRATION.md — Tsinghua educational-agent research, licence resolution, and which capabilities were adopted.
- docs/REUSE.md — what was reused from the DSH Host bundle spec,
dsh-memory-evolve, the Python workbench, DSH skills and external standards. - docs/COMPLIANCE-REVIEW.md — 21-item compliance review against the official bundle and index rules.
verification/plugin-acceptance.mdandverification/final-check-0.2.3.md— independent acceptance and final install check (source repository only; not packaged).
Licence
MIT.
星级评价
基本信息
插件信息来自公开 GitHub 仓库与 SkillHub,版权归原作者所有;安装前请自行审查仓库代码。