菜单
Manage references (BibTeX merge, cite-as-you-write resolution) and convert documents between Markdown / LaTeX / docx via pandoc. Use when the user says "build my bib file", "resolve [@cite] placeholders", "convert this draft to LaTeX / Word".
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Build a single self-contained HTML dashboard of all captured ideas and every paper's pipeline progress — per-section completion, conference name, deadline, and days-left, with click-to-sort columns. Use when the user says "show my dashboard", "overview of my ideas and papers", "which paper is due first", or triggers /dashboard.
Design and run experiments bound to a GitHub repo. Tracks the bound repo's URL + SHA, records versioned execution attempts (semver), and mirrors structured results locally so /paper can consume them when writing. Use when the user wants to design, log, sync, or analyze experiments. Triggered by /experiment.
Generate research figures — structural SVG (architecture / pipeline / concept) or data matplotlib plots — plus curate a reference-figure library. Each figure goes through a 6-step interactive flow (intent → kind → refs → size → palette → render) and lands in the correct paper or experiment scope. Triggered by /figure.
Socratic research-direction tool — walk the user through Socratic discussion, scout the last 3 years of arXiv + WebSearch, score value/feasibility, recommend venues, emit a brain-library index, hand off to /paper. Each idea is persisted as an on-disk manifest under outputs/idea-checks/<slug>/ and mirrored to AgentDB ideas/<slug>. Also retains the legacy horizontal-matrix and vertical-lineage modes. Use when the user says "help me find a research direction", "is this idea novel", "compare these N papers", "what should I learn for this idea", "where does this idea come from", or triggers `/idea-check`.
Summarize research papers (PDF or arXiv) into structured markdown (problem / method / results / contribution / limitations) and index each summary into the AgentDB `papers/` namespace for later semantic search. Use when the user drops PDFs into `inputs/papers/` or pastes an arXiv URL.
Bundle per-user state into a zip for cross-machine migration and restore it on the new machine without overwriting existing files. Excludes externally-reproducible artifacts (HuggingFace base models, etc.) registered via per-experiment external-artifacts.md. Triggered by /migrate.
| name | ref-manager |
| description | Manage references (BibTeX merge, cite-as-you-write resolution) and convert documents between Markdown / LaTeX / docx via pandoc. Use when the user says "build my bib file", "resolve [@cite] placeholders", "convert this draft to LaTeX / Word". |
STATUS: active.
research_assistant.refs.merge_bibtex/convert_document/scan_tex_cite_keys/render_latexall implemented.
outputs/drafts/ contains [@cite:slug] placeholders to resolve.Target: outputs/papers/<venue>/<direction>/refs.bib.
main.tex + sections/*.tex for \cite{<slug>} keys via
research_assistant.refs.scan_tex_cite_keys.papers/<slug> in AgentDB. If its
metadata.bibtex field is populated (lit-summarize stores it there when a
DOI was known), use that canonical entry directly — no network. Otherwise
fall through to step 3.research_assistant.lit.publisher_bibtex.fetch_bibtex_from_publisher(doi)
to fetch the canonical BibTeX (tries CrossRef → doi.org content
negotiation → CloakBrowser stealth fallback). Then memory_store it under
papers/<slug> so the next /cite run skips the network entirely.<direction>/refs.bib via
research_assistant.refs.merge_bibtex (dedupes by DOI → title+author →
entry key).\cite{slug} keys already match refs.bib
entries.Citation-debt check (evidence-first). A \cite{slug} in the prose with no
matching refs.bib entry is a citation debt — paper-architect surfaces it
on the progress board. Use
research_assistant.refs.unresolved_cite_keys(direction_dir) to list those keys
(it diffs scan_tex_cite_keys against bib_entry_keys(refs.bib)). After a
/cite run, the right outcome is an empty list; any remainder means the slug
has no papers/<slug> bibinfo and no known DOI — leave it as a [REF_NEEDED]
gap rather than inventing an entry.
research_assistant.refs.convert_document (shells out to pandoc).outputs/papers/<venue>/<direction>/refs.bib — per-direction BibTeX.*.tex, *.docx).papers/<slug> — read (bibinfo lookup); written when step 3 fetches a new
entry via publisher_bibtex.firstauthorYYYYkeyword — e.g.
vaswani2017attention. merge_bibtex preserves whatever key the source
entry carries; the convention only governs keys we mint ourselves when a
publisher feed omits one. Override per-venue by editing the entry in
<direction>/refs.bib after the merge.ieee for tech venues (default), acm for ACM venues
(SIGCOMM / OSDI / SOSP / NSDI / SIGMOD / VLDB), neurips for ML venues.
The venue's own _template/ .bst always wins when one is present; this
default only applies when no template ships.(Pandoc-on-PATH check shipped — convert_document raises a friendly
RuntimeError with the install hint when pandoc is missing.)