Summary: One package holds CLI, core, schemas, skill and comparison viewer. Private topics, originals and analysis stay outside public code; editors and adapters are excluded.
Current implementation comprises CLI/core, comparison server/UI and shared map. workspace.json owns card/project/review/relation connections; topic folders own discovery, selection, collection and analysis. Later editing/import proposals are not distributed features. See CLI, analysis and comparison.
danseo/
├─ README.md / AGENTS.md / TASKS.md
├─ docs/ # Usage guides, formats and evidence-check contracts
├─ src/
│ ├─ cli.ts # Commands, options, output, exit codes
│ ├─ core/ # Cards, topics, sources, analysis, reviews, relations
│ ├─ web/ # Read-only comparison server and browser UI
│ └─ views/ # Shared map and reading styles
├─ schemas/ # Card/workspace/relation JSON Schema
├─ scripts/ # TypeScript build/documentation tools
├─ skills/danseo/ # Research instructions, generators and assets
├─ skills/danseo-reading/ # Learning questions, source feedback and concept views
├─ skills/paper-to-project/ # Paper/project comparisons and application guidance
├─ examples/minimal/ # Runnable synthetic workspace
├─ evals/ # Evaluation plan; answers/runs remain private
├─ tests/ # TypeScript regressions, helpers, package checks
├─ workspaces/ # Private local data, Git-excluded
└─ work/ # Private work records and execution evidence
No separate package/permanent service is required. CLI owns options/output, core owns file/model rules, and the comparison server exposes a read-only loopback view. Build outputs: build-tools/, dist/, build/; development tools/tests are not distributed. See build contracts.
Docs/TASKS/evals are shareable product material/plans, not private logs, original papers or answer keys. The npm files allowlist includes built assets, schemas, skill, synthetic examples and public docs/brand/notices, rather than every repository file. Verify the actual tarball and installed CLI.
- This repository owns product code, common formats and distributed skill.
- Target projects own questions, source review, learning and experiment records.
- Existing global skills stay unchanged until an explicit installation/update; there is no automatic sync.
- Graphs/comparisons derive from cards; do not manually maintain parallel data.js/graph.json truth.
- Personal prototypes are references, not whole folders to copy with private paths/data.
workspace/
├─ papers/<paper-id>.md
├─ projects/<project-id>/project.md
├─ projects/<project-id>/applications/<card-id>.md
└─ relations.yaml
Frontmatter and body hold one record's structure/explanation. Round-trip preservation belongs to PTP-001. Any future cache/index is disposable derived data; paths/IDs/external references remain a draft compatibility contract.
The paper library is a local workspace, not a separate GitHub repository/database/server. Papers may be shared across projects. Registering a URL or existing PDF/Zotero attachment does not require downloading. Use explicit imports when offline access/edition retention matters; optional sources/<paper-id>/<version>.pdf is a layout suggestion, not a reason to create empty folders or move originals automatically.
Detailed source checks need actual access to relevant edition text/tables/figures. Successful registration or an existing PDF does not establish reading/review. URL-only unread sources remain unchecked; no-download workflows cannot promise offline access or link durability. Zotero's stored/linked attachments illustrate this separation; direct integration is unimplemented.
- Register title/source with a stable ID and draft status; metadata generation is a design direction, not an included card editor.
- Add relevant claims/locators/conditions/limits when a question needs them. Bulk summarizing the whole library is not a prerequisite.
- Link project questions, existing experiments and decisions through Applications without forcing one for every paper.
A proposed index derives IDs/titles/sources/status/references; people should not maintain a second index manually. Begin with scans if implemented, adding disposable caches only when needed. No indexer, embeddings/vector DB/full-text chunk store is included. Revisit when keyword/filter navigation actually fails.
Small Paper/Application Markdown templates can guide conditions, history, question, observations, rationale and next checks. Never invent facts to fill blanks or add a template engine. Current required fields belong to CARD_MODEL and schemas.
Keep full originals; summaries guide returns to sources rather than replacing them. Existing agents/Markdown suffice without adding a required automatic PDF-processing dependency.
- Identify edition/access coverage and locations of problem, method, experiments, results, limits and appendices. A location guide is not completed reading.
- Read meaningful section/experiment units; keep table headers, units, footnotes and metric definitions together. Inspect equations/figures on original pages. Preserve neighboring context when splitting sections.
- Record used claims/locators with conditions, comparators, limits and unknowns. Keep conflicts; do not smooth failed extraction or uncertain numbers into facts.
- Cross-check main text/appendix and table/prose values, samples, models and setup; distinguish condition differences from unresolved discrepancies.
- Write a short conclusion/limits guide with source locations. Avoid repeated summaries of summaries; return to evidence for numbers/methods and show read/unread scope.
An optional Paper body can use overview → conditions/evidence → limits/conflicts → check scope/history, without mandatory new overview/coverage fields. Registration-only records need not fill every section. A full-review request includes appendices and cannot be completed with unverified scope. Review states remain separate.
Agents choose relevant sections from the overview, then read evidence/conditions. Missing answers require source return. Split over-limit material with explicit omissions/remaining checks. Accuracy/cost/optimal length remain unmeasured. Lost in the Middle motivates distinguishing accepted context length from information use under its studied conditions, not general claims about all current models. Docling's model is a possible future importer reference; adoption/extraction accuracy is unverified.
File-writing proposals require explicit local workspace access; this does not imply cloud hosting. Cross-device access, authentication and public APIs are separate scope. Develop with native Windows PowerShell and supported CI; see runtime support. Never require a maintainer's private runtime/path. Standalone research UI supports Korean/English via candidate language; CLI/native comparison guidance remains English. Supplied/generated research prose follows the output-language rule.