dsh-plugin-deep-paper-reading

A DeepSeek Harness plugin that ships the 文献精读模式 (Deep Paper Reading) agent preset: install it and the preset picker gains a mode for figure-by-figure, experiment-by-experiment close reading of a single biomedical paper.

Since v0.2.0 the report is diagram-first. The paper's logic chain is drawn as a flowchart instead of described in prose, and the judgement sections (figure/text consistency, experiment strength, claim-evidence gap, limitation risk, term relations) each get a figure. A Word conversion standard ships with the skill, so the report exports to a compliant .docx.

• Keywords: dsh-plugin · deepseek-harness · agent-preset · paper reading · diagrams · DOCX
• License: MIT

What changed in 0.2.0

What it can do

Full introduction (Chinese, with four capability figures): docs/skill-intro.md.

The eight-phase workflow — arrow direction is reading order, and every phase has an artifact:

Which required figure answers which question — structure, evidence, gap, and stage delivery are separate jobs:

Also: deliverables and the gates they must pass · what a real run looks like

What the mode does

1. Acquire and first-pass scan the paper; persist the full text to sources/paper-fulltext.txt
2. Read the main text line by line
3. Treat supplementary material with the same rigor; label anything unobtainable as missing
4. Build the figure-legend ↔ main-text correspondence table plus V3 (text↔figure map) and V4 (consistency matrix)
5. Decompose every experiment into purpose / design & tools / results / interpretation, plus V5 (strength overview)
6. Reconstruct the logic chain, classifying each transition as elimination, extension, mechanism, application, or validation — unclassifiable transitions become dashed edges
7. Render V1–V8, then check each PNG for existence, width, real Chinese glyphs, and the three-line caption
8. Assemble the report; export to Word following the bundled standard

Three hard constraints live in the persona:

• Evidence boundary — never invent figure panels, legends, numbers, p-values, sample sizes, catalog numbers, or a claim of full-text access.
• Figures must actually be looked at — if a judgement rests on a legend rather than the image, say so.
• Figures must actually be drawn — an ASCII sketch, a table drawing, or an empty placeholder is not a figure; if nothing renders, say "unrendered (reason)" and keep the Mermaid source.

Figure inventory

Rendering ladder (record the tier in every caption): A mmdc → B npx -y @mermaid-js/mermaid-cli → C python scripts/render-diagram.py spec.json --out fig.png --mermaid fig.mmd (offline, Pillow only) → D unrendered, labelled.

python presets/deep-paper-reading/skills/deep-paper-reading/scripts/render-diagram.py \
  --selftest --tmp /tmp/diagram-check

See references/visualization-spec.md and assets/diagram-specs.md.

Install

1. As a bundle (needs the dsh CLI)

dsh plugin --profile web add github:1283011852/dsh-plugin-deep-paper-reading
# DSH Desktop: pnpm dsh plugin --profile web add github:1283011852/dsh-plugin-deep-paper-reading

No prepare script and no npm dependencies, so pnpm's git-build permission prompt does not apply. Restart DSH, then pick 文献精读模式 in the preset picker. Remove with dsh plugin --profile web remove dsh-plugin-deep-paper-reading.

2. As a preset package (.dshpreset) — no CLI

Download dsh-plugin-deep-paper-reading-<version>.dshpreset from Releases and import it in DSH Desktop. The desktop validates, previews, and atomically installs into the user preset root; an existing preset id is never overwritten. This path touches no profile configuration.

3. Skill only

cp -r presets/deep-paper-reading/skills/deep-paper-reading ~/.agents/skills/

Then deep-paper-reading is callable through the skill tool in any mode, without the mode persona. The skill directory carries scripts/render-diagram.py, so the rendering capability travels with it. (DSH does not scan ~/.claude/skills/.)

Usage

Start a session on 文献精读模式 and point it at a PDF or DOI. The report lands in paper-reading/<short-slug>-精读报告.md, extracted text and page renders under sources/, diagram PNGs plus their specs under paper-reading/figures/ — unless the workspace already has an established layout.

The skill assumes a working PDF toolchain (python + PyMuPDF). If it is missing the agent says so rather than pretending it read the full text.

Layout

├── package.json                    # dsh.bundle.patch → cordis.patch.yml
├── cordis.patch.yml                # registers this package's presets/ as a roster root
├── index.js                        # boot-time self-check row
├── presets/deep-paper-reading/
│   ├── agent.cordis.yml            # persona (incl. VISUALS) + bundled skill mount
│   ├── preset.yml
│   ├── SOURCE.md
│   └── skills/deep-paper-reading/
│       ├── SKILL.md                            # 8-phase workflow
│       ├── assets/report-template.md           # report template (figure slots + captions)
│       ├── assets/diagram-specs.md             # V1–V11 specs (JSON + Mermaid)
│       ├── references/reading-checklist.md
│       ├── references/visualization-spec.md    # diagram contract
│       ├── references/docx-conversion-standard.md
│       ├── references/biomedical-tools.md
│       └── scripts/render-diagram.py           # offline renderer (Pillow)
├── scripts/build-dshpreset.mjs     # dependency-free .dshpreset packer
├── scripts/check-composition.py    # composition + bundled-file + V1–V8 checks (CI)
└── .github/workflows/              # ci.yml (incl. renderer selftest), release.yml

⚠️ Known impact

This plugin's patch replaces the whole config of the agent-presets row, because DSH patches replace a targeted row's config value rather than deep-merging its keys. Every key that row requires is therefore restated in cordis.patch.yml, which means:

1. The roots list belongs to this plugin. If another bundle or your own cordis.patch.yml also configures agent-presets, the later layer wins outright and the earlier roots disappear. To keep both, merge the two roots arrays by hand in your profile patch instead of mounting both bundles.
2. default is restated as standard. Your agent-presets.default in $DSH_HOME/settings.yaml lives in the settings layer, outranks the config, and is unaffected.

Prefer the .dshpreset install if you would rather not take either consequence — it changes no profile configuration.

Verification

What was actually checked before publishing, rather than assumed:

• Preset mounts: agentPresets.standingKeyFor('deep-paper-reading') → mounted OK, which composes the plugin subtree for real.
• Bundled skill travels with the preset: customSkillDirs uses new URL('skills/', baseUrl), and the preset Include sets baseUrl to the composition's own directory, so the skill root resolves from the user root, this package's presets/, or an imported .dshpreset.
• Bundle install and patch application: in a throwaway profile, dsh plugin add this package, add the dsh-web-app layer, then --dump-config — the insert row lands and the agent-presets override restates all four keys.
• The !!js expression evaluates: dsh --profile <test profile> --help loads every row and evaluates its config — exit 0, with the plugin's self-check row printing registered from …\presets\deep-paper-reading.
• --dump-config proves the patch applied, not that it evaluated (it prints the !!js source), which is why the step above is the one that matters.
• Package format: the packer mirrors the DSH Desktop importer's constraints and CI re-reads the archive with unzip. v0.2.0 run: 12 files, 172.5 KiB → 71.2 KiB, with the renderer, spec and template files all inside.
• Static composition checks: python3 scripts/check-composition.py (every push). Since 0.2.0 it also verifies that every backticked references/ · assets/ · scripts/ path named in SKILL.md is really packaged, and that V1–V8 are declared in both SKILL.md and visualization-spec.md.
• Renderer self-test: --selftest renders all four kinds and asserts that a spec without source is rejected. Locally the four PNGs came out 1020×1640 / 1360×816 / 1760×1232 / 1984×876 px, all ≥1200 px, with real Chinese glyphs.

Not verified:

• That the preset appears in your picker — that requires one real install, the only step that must happen in a live environment.
• On a Linux box with no CJK font, the renderer falls back to box glyphs (with a stderr warning); switch to mermaid-cli there. CI installs fonts-noto-cjk to cover that path.
• The actual Word layout: this repo publishes the standard and template only; reflow in specific Word/WPS versions needs the visual acceptance pass in section 13 of the standard.

License and provenance

MIT — see LICENSE.

• The bundled skill is the author's own (frontmatter records MIT and skill-author: user). Since 0.2.0 it is no longer a byte-identical upstream copy: Phase 7, the V1–V8 contract, the visualization spec, the diagram templates, the offline renderer and the Word standard were added, and the report template and checklist were rewritten to match.
• The Cordis composition derives from a locally authored scientific-evidence-en-mode preset: its tool surface is kept as-is, with the persona row rewritten (VISUALS paragraph added, OUTPUT updated), the bundled skill directory added, and display metadata written.
• Mermaid conventions (accTitle/accDescr, classDef only, snake_case ids, node caps) follow the K-Dense markdown-mermaid-writing style guides.
• Orientation evidence that diagrams aid comprehension (not a validation of this design): Infographics Improve Comprehension and Recall at the Orthopaedic Journal Club (J Surg Educ, 2020) and Let Your Ideas Flow: Using Flowcharts to Convey Methods and Implications of the Results (J Microbiol Biol Educ, 2018, PMC5969414).
• See presets/deep-paper-reading/SOURCE.md.