paper-wiki-manager
Maintain an OKF paper wiki under `paper-wiki/` or another user-named wiki root, including optional Chinese mirror notes under `papers_zh/`. Use when asked to add arXiv or research-paper URLs, capture non-paper sources such as blog posts, docs, or talks, update or translate paper notes, maintain paper/topic/concept/source indexes, automatically create or update topic summary pages for important new themes, maintain concept entity pages for methods, datasets, benchmarks, metrics, terms, or tools, link papers to research projects, normalize paper metadata, track reading status, compare papers, generate required `viz.html` visualizations, or validate the paper wiki.
- License 许可
- MIT
- 最近更新
- 2026-11-14
- Stars
- ★ 3
- Forks
- 0
01 前言 Front Matter
- name
- paper-wiki-manager
- description
- Maintain an OKF paper wiki under `paper-wiki/` or another user-named wiki root, including optional Chinese mirror notes under `papers_zh/`. Use when asked to add arXiv or research-paper URLs, capture non-paper sources such as blog posts, docs, or talks, update or translate paper notes, maintain paper/topic/concept/source indexes, automatically create or update topic summary pages for important new themes, maintain concept entity pages for methods, datasets, benchmarks, metrics, terms, or tools, link papers to research projects, normalize paper metadata, track reading status, compare papers, generate required `viz.html` visualizations, or validate the paper wiki.
02 资源文件 14
- 参考文档agents/openai.yaml242 B
- 资产文件assets/paper-wiki.toml8.1 KB
- 参考文档references/repeated-ingestion.md7.1 KB
- 参考文档references/schema.md14.3 KB
- 参考文档references/SPEC.md14.9 KB
- 脚本scripts/create_paper_body_profile.py17.4 KB
- 脚本scripts/generate_viz.py11.7 KB
- 脚本scripts/set_status.py3.8 KB
- 脚本scripts/static/viz.css11.4 KB
- 脚本scripts/static/viz.js33.7 KB
- 脚本scripts/templates/viz.html4.2 KB
- 脚本scripts/test_set_status.py2.8 KB
- 脚本scripts/test_validate_paper_wiki.py2.2 KB
- 脚本scripts/validate_paper_wiki.py23.4 KB
03 SKILL.md 正文
Paper Wiki Manager
Overview
Maintain an OKF paper wiki as Markdown files with YAML frontmatter: one canonical page per paper, one page per topic, and one page per recurring named entity (method, dataset, benchmark, metric, term, or tool), all cross-linked in both directions. Optionally maintain Chinese paper-note mirrors under papers_zh/; these are auxiliary reading artifacts, not OKF graph concepts. Keep wiki content in paper-wiki/ unless the user names another wiki root; keep <wiki-root>/viz.html as the required generated graph. Treat references/SPEC.md as the bundled OKF format snapshot and references/schema.md as the stricter paper-wiki profile.
This skill supersedes paper-library-manager, which managed the same kind of bundle under the historical root name paper-library/.
Scope
Use paper-wiki/ as the default wiki root unless the user names a different path. Treat every non-reserved .md file in that tree as an OKF concept except files under the auxiliary papers_zh/ collection. The four canonical page collections are papers/ (one page per arXiv paper), topics/ (thematic synthesis pages), concepts/ (entity pages for named methods, datasets, benchmarks, metrics, terms, and tools), and sources/ (non-paper reading — blogs, docs, talks — as type: Reference). papers_zh/ stores zh-CN mirrors of canonical paper notes and does not participate in indexes, topic/concept backlink requirements, or graph nodes and edges. Treat the papers_zh/ directory name and zh-CN language tag as fixed parts of this skill's contract; [localized_notes].enabled is the only localized-note configuration switch.
Workflow
When adding or updating a paper:
- Parse the arXiv ID from the URL, user input, or PDF filename/path. For a PDF, check the filename first (e.g.,
2401.00001.pdf); if the ID is not in the filename, extract it from the PDF header orarxiv.orgURL embedded in the document. If no ID can be extracted (user provided only a title or keyword): resolve it withpaper-searchas described in Fetching Papers. - Read the existing paper file if
paper-wiki/papers/<arxiv_id>.mdalready exists. If it exists, read and followreferences/repeated-ingestion.mdbefore editing; never append a second generated note or template wholesale. - Read
assets/paper-wiki.tomlfrom this skill and follow its paper body profile settings. - Fetch metadata and paper content through
paper-searchas described in Fetching Papers. Prefer arXiv for bibliographic facts; use project pages, GitHub, Hugging Face paper pages, or Semantic Scholar only as additional sources, and only through a source the human has switched on. - Create or update one paper concept under
paper-wiki/papers/. - Complete the fidelity pass in Paper Documents against the retrieved full paper and every additional source used.
- Read
[localized_notes].enabledinassets/paper-wiki.toml. When true, create or update thezh-CNmirror underpaper-wiki/papers_zh/<arxiv_id>.mdfollowing Chinese Paper Notes. - Identify and update 1 to 3 important themes following Topic Documents.
- Identify named entities that meet the concept-page criteria in Concept Documents; create or update pages under
paper-wiki/concepts/and cross-link them with the paper. - Add concise topic and concept links under the paper body; add the paper to each affected topic's and concept's
# Paperssection. - If the paper is used by a project in this repository, add or update the paper's
# Used In Projectssection following Project Links. - Update
paper-wiki/papers/index.md,paper-wiki/topics/index.md,paper-wiki/concepts/index.md(when concepts exist), and every affected topic or concept page; update the wiki homepaper-wiki/index.mdif a new canonical collection appeared. Do not create apapers_zh/index.mdentry. - Preserve user-curated fields and existing body layout following Metadata Rules, Paper Documents, and Chinese Paper Notes.
- Run the Finishing Commands after content edits.
- Cite only sources that were actually used.
Paper Documents
Use references/SPEC.md as the base OKF format reference and references/schema.md as the stricter paper-wiki profile.
Paper bodies are user-customizable Markdown. Do not treat any one summarization template as part of the OKF contract. Use assets/paper-wiki.toml as the default paper body configuration for this skill. When reading the config, enumerate all paper_body.profiles entries. Use paper_body.default_profile by default, but choose another configured profile when the user requests it or when the paper type clearly matches it. Use the selected paper_body.profiles.<name>.sections list for new paper bodies. Treat paper_body.required_sections as validation requirements only when the user configures them.
Use paper_body.section_descriptions as drafting guidance for what each section should contain. Do not copy those descriptions into generated paper notes unless the user explicitly asks for visible prompts.
If a specific paper needs a different summarization style, derive a temporary profile or template from the asset config for that paper, and persist the new profile only when the user asks. Keep generated summaries concise and distinguish paper claims from personal notes. If a paper has not been read in full, avoid presenting speculative critique as established fact. Preserve existing paper body layout when updating a paper unless the user explicitly asks to reorganize it.
For every repeated paper input, treat the existing canonical body as the current best synthesis. Merge only evidence-supported changes into its existing sections, protect human-owned notes, and use the no-op and review gates in references/repeated-ingestion.md.
Before finalizing a drafted note, run a fidelity pass: locate every load-bearing number, taxonomy branch, named exemplar, and comparative claim in the retrieved full paper or an additional source listed under # Citations. Mark reader-constructed examples and inferences explicitly, distinguish the source's qualitative synthesis from comparable empirical evidence, and remove claims that cannot be traced to a used source. The pass is complete only when every such claim has an identified source and the note's evidence boundaries match what that source can establish.
Chinese Paper Notes
Use [localized_notes].enabled in assets/paper-wiki.toml as the Chinese-mirror on/off switch. The mirror directory is always <wiki-root>/papers_zh/ and the language is always simplified Chinese with the zh-CN tag; do not treat either value as configurable. When enabled = true, create or update <wiki-root>/papers_zh/<arxiv_id>.md after writing the canonical English note. Use the same arXiv ID as the filename and distil from the same retrieved full paper; do not translate from the abstract alone.
Treat papers/<arxiv_id>.md as the sole source of truth for bibliographic metadata, status, tags, topic/concept links, project links, and graph identity. A Chinese mirror is an auxiliary reading artifact and must not be added to papers/index.md, topic or concept # Papers sections, or viz.html as a node or edge. The viewer may embed its title and body under the canonical paper ID for an English/Chinese detail-pane toggle. It may link one-way to its canonical English note and to external citations without creating backlink obligations.
Use this lean frontmatter:
---
type: LocalizedPaperNote
title: 中文标题或原论文标题
language: zh-CN
arxiv_id: "2401.00001"
source_note: ../papers/2401.00001.md
timestamp: YYYY-MM-DDTHH:MM:SSZ
---
Translate the selected canonical paper-body profile semantically rather than mechanically. Preserve the same reasoning coverage and evidence boundaries while using natural Chinese headings, explanations, tables, and terminology. Keep important English technical terms in parentheses on first use when that improves precision. Preserve # 个人笔记 and any user-authored or custom sections when updating an existing Chinese note. If the canonical note changes only in metadata, links, or formatting, do not rewrite the Chinese body unnecessarily. Keep # Reading History only in the canonical English paper; do not duplicate it in the Chinese mirror.
Custom Paper Body Profiles
When the user asks for their own paper-note template, treat it as a custom paper_body.profiles.<name> profile, not a change to the OKF contract.
Use scripts/create_paper_body_profile.py as the profile-generation interface. The user may describe the desired reading style in natural language; convert that request into a concise profile name, an ordered section list, and optional section descriptions. If the user's description explicitly names sections, preserve those names. If it does not, either let the script infer a draft from the description or pass agent-chosen sections with repeated --section flags. By default, the script saves the finished profile to this skill's bundled assets/paper-wiki.toml so it is reusable in later paper-wiki tasks.
If the user asks for an English-only template, pass --language english and provide English --description, --section, and --section-description values. The script rejects non-ASCII profile text in that mode, so fix translated section names before saving or previewing the TOML.
Create and save a reusable profile without changing the default paper body:
python scripts/create_paper_body_profile.py \
--name critique-card \
--description "Short review template focused on assumptions, evidence gaps, and follow-up questions" \
--section "Summary" \
--section "Assumptions" \
--section "Missing Evidence" \
--section "Questions"
Preview a generated TOML template without changing the config:
python scripts/create_paper_body_profile.py \
--name critique-card \
--description "Short review template focused on assumptions, evidence gaps, and follow-up questions" \
--section "Summary" \
--section "Assumptions" \
--section "Missing Evidence" \
--section "Questions" \
--preview
Do not change paper_body.default_profile after creating a custom template unless the user explicitly asks to make that template the default. If asked, rerun the script with --set-default for the intended profile.
For a wiki-specific template, write to a separate config and pass it to validation with --config <path/to/paper-wiki.toml>. Avoid overwriting global section_descriptions for common section names unless the user explicitly asks for a different meaning; prefer unique section names for specialized templates.
Fetching Papers
Reach papers through the paper-search skill ($PS is its scripts/paper_search.py), which honours the human's per-source switches and read order in memory/paper-sources.json. A source switched off there stays off on every route, including web fetches and MCP connectors; when the note needs one, say so and let the human decide.
Input resolution:
- arXiv, alphaXiv, or Hugging Face paper URL, or a bare ID (with or without a version suffix): pass it to
fetchas is; it extracts the versionless ID. - PDF file path: check the filename first; if no numeric ID is present, run
grep -a 'arxiv.org' "$PDF" | head -5to extract an embedded URL. - Natural-language title or description only: run
uv run $PS search "TITLE" --mode keyword --limit 5. Proceed without asking only when exactly one candidate's title equals the given title once case and punctuation are ignored, and name that candidate in your reply. Otherwise show the top candidates to the user and wait for confirmation before proceeding; never guess the ID.
Fetch: run uv run $PS fetch ARXIV_ID. It writes the full text from the first source in the read order that returns a complete text to the file its JSON report names, and fills metadata (title, authors, date, abstract) from arXiv, else Hugging Face. Read the file as the primary source for the paper body summary; distil it into the configured body sections rather than reproducing it. Cite the report's source and url under # Citations, and note in the body which source the text came from. On exit 1, no source returned a complete text: report the tried outcomes and the pdf link to the user rather than fetching through another route.
Topic Documents
Use topic concepts for stable themes such as multi-agent-systems, llm-agents, ai-for-science, agent-self-evolution, or benchmarks. A topic page should list related papers and open questions, not duplicate each paper's full summary.
Create or update topic pages proactively whenever a new paper introduces an important theme. Important themes usually satisfy at least two of these conditions:
- The theme appears in the title, abstract, method name, benchmark name, or central contribution.
- The theme could group multiple current or future papers.
- The theme is useful for retrieval, comparison, or literature-review synthesis.
- The theme is more specific than a broad field label such as
aiormachine-learning.
Do not create a topic for every tag. Tags can be granular; topic pages should represent durable synthesis nodes. Prefer reusing an existing topic when the new theme is semantically equivalent to one already present.
Every topic page should include # Scope, # Papers, and # Open Questions. When a topic has two or more papers, optionally add # Synthesis with concise cross-paper observations.
When adding a new topic, update paper-wiki/topics/index.md. Ensure links are bidirectional: the paper links to the topic, and the topic links back to the paper.
Concept Documents
Concept pages live in paper-wiki/concepts/<slug>.md and are the wiki's entity pages: one page per concrete named thing, as opposed to a topic's thematic grouping. Use one of these type values: Method, Dataset, Benchmark, Metric, Term, or Tool.
Create a concept page when a named entity satisfies at least one of:
- Two or more papers in the wiki reference it — the strongest signal (e.g., a benchmark that several papers report results on).
- It is a durable, field-level entity likely to recur in future reading (e.g., a widely used technique such as LLM-as-a-judge).
- The user asks to track it.
Do not create a concept page for a method that only its own paper describes — the paper page already covers it. Add that method page later, when a second paper builds on or compares against the method.
Every concept page should include # Definition (what the entity is, in a few sentences) and # Papers (links to wiki papers that use or reference it, each with a one-line note on how). Optional sections: # Notes, # Related.
Ensure links are bidirectional: the paper links to the concept (typically in its # Related section), and the concept lists the paper under # Papers. When adding a new concept, update paper-wiki/concepts/index.md.
Source Documents
When the input is a non-paper source (a blog post, documentation page, or talk) with no arXiv ID, add it under paper-wiki/sources/<slug>.md instead of papers/. Fetch it with a web request (paper-search fetch does not apply). Use type: Reference and the source frontmatter in references/schema.md: resource is the source URL (its identity), authors/published/medium are optional, and the filename is a lowercase hyphenated slug derived from the title.
For a blog, survey, or tutorial that organizes others' work rather than presenting one contribution, use the synthesis-source body profile from assets/paper-wiki.toml; read that profile for its current section list and drafting guidance rather than duplicating the list here. The profile keys on the content type (synthesis), so an arXiv survey paper can use it too — do not treat it as blog-specific.
Sources are a lighter tier than papers. Link a source to the topics and concepts it covers, but these links are one-way: do not add the source to a topic's # Papers list, and the validator does not require a backlink. When a new source is added, update paper-wiki/sources/index.md and, if sources/ is new, add a Sources line to the wiki home. A source that references many methods is a natural hub — the techniques it surveys may later become their own paper or concept pages.
Project Links
Papers can record where they are used in this repository's research projects, keeping the two reference layers distinct: shared paper understanding lives in the wiki; project-specific use of a paper (BibTeX entries, claims, experiments) stays inside the project.
When a repo project cites or builds on a paper, add a # Used In Projects section to the paper body with a relative link to the project folder and a one-line note on how the project uses it:
# Used In Projects
* [my-project](../../projects-folder/my-project/index.md) - baseline method for the packing experiments.
Link a Markdown file such as the project's index.md, not the bare directory — the validator only checks link targets ending in .md. These links point outside the wiki bundle; the validator checks that the target exists in the repository but does not require a backlink from the project. If the wiki bundle is copied to another repository, these links may break — that is acceptable per OKF's tolerance for broken links; prune or update them in the copy.
Indexes
Maintain index files as plain Markdown without frontmatter. The wiki root index.md is the wiki home: it lists the papers/, topics/, and concepts/ collections and may carry a short current-focus section. Sort paper entries by arXiv ID or submitted date when the user does not specify a preference. Keep descriptions one sentence.
Do not create or maintain an index for papers_zh/. Chinese mirrors are discovered by their shared arXiv-ID filenames and their source_note pointers.
Metadata Rules
Preserve these user-curated fields when updating a paper:
statusprioritytags# Notes# 个人笔记- any custom frontmatter keys not defined in
references/schema.md
status is the human's reading state — unread, skimmed, or read — and changes on the human's word alone. Write status: unread on every new paper and source, including ones an agent read in full: the note body is the agent's memory of a paper, while read tells later sessions the human already knows it. Change it only when the human does so in the viewer's Status row or asks you to, then run scripts/set_status.py <papers/ID | sources/slug> <status> --root <wiki-root>, which rewrites only that frontmatter line. Find what the human has read with grep -l '^status: read' <wiki-root>/papers/*.md <wiki-root>/sources/*.md.
Finishing Commands
Before finishing paper-wiki edits:
- Confirm that the manual fidelity pass in Paper Documents is complete; the structural validator cannot verify factual faithfulness to the paper.
- Check required metadata fields from
references/schema.md; the bundled validator is the executable check. - Check configured paper body sections only when
assets/paper-wiki.tomlsetspaper_body.required_sections. - Check that internal Markdown links resolve within the wiki root, and that project links resolve within the repository.
- Check that index entries point to existing files.
- When
papers_zh/exists, check each Chinese mirror's identity and canonicalsource_note; the validator does not require every canonical paper to have a mirror. - Run bundled scripts from this skill's root directory, using paths relative to the directory that contains this
SKILL.md. Pass the wiki root as an absolute path when the target repo is not the current working directory. Generateviz.htmlbefore validation because the validator reads the generated graph file.
uv run scripts/generate_viz.py /absolute/path/to/paper-wiki
uv run scripts/validate_paper_wiki.py /absolute/path/to/paper-wiki
If uv is unavailable or blocked by sandbox/cache access and the environment already has Python 3.11+, run the scripts directly with python:
python scripts/generate_viz.py /absolute/path/to/paper-wiki
python scripts/validate_paper_wiki.py /absolute/path/to/paper-wiki
Use --config <path/to/paper-wiki.toml> only for a temporary validation profile or another repo layout.
If your environment provides a skill validator, run it against this skill folder after editing the skill itself.
After editing validate_paper_wiki.py or its body contracts, run:
python -m unittest discover -s scripts -p 'test_*.py' -v
Visualization
paper-wiki/viz.html is required. Use viz.html as the canonical filename; treat vis.html as a typo unless the user explicitly asks for a separate alias.
scripts/generate_viz.py renders a self-contained Cytoscape graph + detail-pane viewer by injecting scripts/templates/viz.html, scripts/static/viz.css, scripts/static/viz.js, and the vendored libraries in scripts/vendor/. Keep all of these alongside the script; the generated viz.html inlines the vendored libraries so it works fully offline (no CDN). Papers, topics, and each concept type get distinct node colors. The viewer UI and text are English throughout; Chinese appears only inside a paper note's own body. The generator excludes papers_zh/ from graph nodes and edges, but embeds a matching Chinese title and body under its canonical paper ID. For papers with a mirror, show a 中文 button in the detail pane to switch that note to Chinese and an English button to switch back; the choice sticks across papers until it is switched back.
The viewer has two tabs over a shared detail pane: Graph and Timeline. The
tab is reflected in the URL fragment (#graph, #timeline) so either view can be
linked or bookmarked directly; #graph is the default. The search box is shared by
both tabs; every other control belongs to one of them. The split between the left
view and the detail pane starts at half the window and is resized by dragging the
divider between them; each side keeps a 260px minimum, and the graph re-measures as
the divider moves.
The Graph tab is a paper-wiki knowledge map, not a flat force graph:
- Collapsed by default. It opens showing only topics and concepts on a deterministic grid (same layout every time), each labelled with its paper count. Papers stay hidden until a topic is opened, so the first view is a readable index rather than a hairball.
- Expand on demand. Double-clicking a topic (or the "Expand N papers" button in the detail pane) fans that topic's papers around it without moving the other topics. "Expand all" runs the fcose force layout to lay out the whole clustered map; "Collapse all" and "Reset view" return to the grid.
- Focus. Clicking any node highlights it and its direct neighbors and dims the rest; the highlighted topics and concepts and the clicked node itself keep readable labels however far the map is zoomed out, and the clicked node keeps a minimum on-screen size; the type legend (top-left) filters to one type; search reveals matching hidden papers.
The Timeline tab is the chronological reading surface:
- Papers and sources only. Topics and concepts are rewritten repeatedly, so their
timestampmeans "last touched" and would read as noise in a reading feed. - Two orderings. "Added" sorts by
timestamp(when the note entered the wiki), "Published" bysubmittedfor papers andpublishedfor other sources; the arrow flips newest-first and oldest-first. Entries carrying neither publication date group under "No date" at the end. Same-day entries fall back to publication date and then concept ID, so the order never drifts between reloads. - Compact rows. Month headers with entry counts, then one row per entry: day, a
statusglyph, and the English title, with the description on hover. Clicking a row renders that note in the shared detail pane. - Status row. The detail pane shows a paper's or source's
status. When os-ui's dev server is behind the page (itsGET /api/paper-wiki/statusanswerswritable), the row becomes unread / skimmed / read buttons; a click runsset_status.py, regeneratesviz.html, and repaints the timeline glyph. Opened from disk or a static host, the row stays read-only. - Show in graph. While the Timeline tab is up, the note header carries a "Show in graph" button. It switches to the Graph tab, reveals a paper together with the papers it links with, highlights that neighborhood, and frames it with the zoom capped at the whole-map fit, so the note's position among its topics and concepts stays readable against the dimmed map. The button is hidden on the Graph tab, where the note is already on screen.
generate_viz.py copies timestamp, submitted, published, and status from frontmatter onto graph nodes so the timeline can order and label entries; the validator already requires timestamp on every note and submitted on every paper.
These are viewer behaviors baked into the bundled assets — regenerating viz.html is all that is needed to pick up viewer changes; no per-paper edits are involved.
Comparison Tasks
When comparing papers, write the comparison as a separate note only if the user asks for a durable artifact. Otherwise answer in chat and reference the paper files. Compare along concrete axes such as problem framing, method, evidence, assumptions, failure modes, reusable artifacts, and relevance to the user's research direction.
04 同仓库更多 Skills
workflow-skill-creator
Distills a completed user workflow or interaction into a reusable agent skill. Use when the user asks to turn their workflow, interaction, or multi-step process into a skill, or when they say "make this a skill", "create a skill from what we just did", "package this workflow" or similar. Do not use for creating skills from scratch without an existing workflow (use a generic skill-creator for that).
uv
Checks whether the uv Python package manager is installed and installs it if missing. Ensures uv is on PATH. Use when another skill requires uv as a prerequisite.
uniprot-database
Access protein metadata, function, taxonomy, and sequences across UniProtKB, UniParc, and UniRef. Use when searching for proteins, mapping identifiers, or retrieving functional annotations and publications. Don't use for sequence alignment, protein folding, or sequence similarity search (use specialized skills for those tasks).
unibind-database
Queries the UniBind database for experimentally validated transcription factor (TF) binding sites. Use when retrieving direct TF-DNA interaction datasets, downloading binding site coordinates (BED/FASTA) for local analysis, or listing available datasets by species, cell line, or TF name. Don't use to query specific intervals, locations, genes, motif models or expression data.
ucsc-conservation-and-tfbs
Fetch Evolutionary Conservation scores (phyloP, phastCons) and Transcription Factor Binding Sites (TFBS) from the UCSC Genome Browser. Use when analyzing whether genomic variants or regions are evolutionarily conserved, functionally important, or bounded by TF regulators across major projects (ENCODE, JASPAR, ReMap).
string-database
Query the STRING database for protein-protein interactions (PPIs), functional enrichment, and homology. Use when the user asks about interactions between specific proteins, interaction evidence, confidence scores, protein interaction partners, or pathway enrichments.