paper-toolkit

Deterministic Python tooling for the fixed steps of academic-paper development — workspace setup, evidence-DAG management, structured checks, LaTeX composition, compile + page introspection, figure rendering, literature search, and an analysis pipeline that feeds the manuscript.

The toolkit contains no LLM calls. All writing, review, and judgment lives in the companion Claude Code skills, which call into this CLI for everything that should be reproducible.

Why

Paper drafts drift between agents, sessions, and editors. The toolkit pins the mechanical parts — schemas, checks, file layout, LaTeX compose, evidence DAG — so the agent only owns prose. Two skills ship with the package:

• agentsociety-analysis — runs the 6-stage analysis pipeline

(frame → explore → claims → refine → produce → synthesis) against an experiment SQLite database, emitting bilingual reports + publication-grade figure specs.

• agentsociety-generate-paper — turns analysis outputs into a

Springer-Nature-style manuscript: intake, drafting, checks, skeptical review, revision loop, and compile.

Install

Requires Python ≥ 3.11.

# from a clone (recommended for now)
git clone https://github.com/Yokumii/paper-toolkit.git
cd paper-toolkit
uv sync

# or with pip
pip install .

LaTeX (TeX Live / MacTeX) is required for paper compile-once and

paper figure render. The toolkit shells out to latexmk / bibtex /

pdflatex and parses their logs deterministically.

Quick start

# 1. create a workspace
uv run paper init --title "Demo" --venue nature --language en --workspace ./demo

# 2. seed an evidence node + claim
uv run paper evidence add-claim --workspace ./demo \
    --node-id c_warming --label "Mean temperature rose 1.1 K" --strength primary

# 3. render a figure from a JSON spec
uv run paper figure render --spec ./demo/paper/figure_specs/f_trend.json --workspace ./demo

# 4. compose and compile
uv run paper compose pack-figures --workspace ./demo
uv run paper compose write-main --workspace ./demo
uv run paper compile-once --workspace ./demo

# 5. structured checks
uv run paper check style --workspace ./demo
uv run paper check claim-coverage --workspace ./demo
uv run paper check figures --workspace ./demo
uv run paper check figure-qa --workspace ./demo
uv run paper status --workspace ./demo

Run uv run paper <group> --help for the full verb list in each group, or see the dev branch for the long-form design and plan documents.

CLI surface

| Group | Purpose | |---|---| | paper init / status / scan | Workspace lifecycle and snapshotting | | paper evidence | Evidence DAG: nodes, edges, claim/evidence linking | | paper template | List and expand section templates | | paper compose | Pack figures, write main.tex, write refs.bib | | paper compile-once / page | LaTeX compile + page-metadata inspection | | paper check | style, figures, claim-coverage, figure-qa | | paper figure / table | Render figures and LaTeX tables from JSON specs | | paper lit | Direct-API search (arXiv / CrossRef / OpenAlex) → refs.bib | | paper refs | Inspect and dedup refs.bib (DOI + title-Jaccard fallback) | | paper analysis | 6-stage analysis pipeline + lift-to-evidence bridge |

Every command emits a JSON Envelope on stdout (success, payload, errors) so the skills can route on structured results rather than parsing prose.

Using the Claude Code skills

The skills live under skills/ and are picked up automatically when paper-toolkit is installed as a Claude Code plugin.

// in .claude-plugin/marketplace.json
{
  "plugins": [
    {
      "name": "paper-toolkit",
      "source": "https://github.com/Yokumii/paper-toolkit"
    }
  ]
}

Inside a Claude Code session:

/skill agentsociety-analysis        # run the analysis pipeline
/skill agentsociety-generate-paper  # draft, check, and compile the manuscript

Both skills read their SKILL.md router first, then dispatch to per-stage prompts and subagents (data-explorer, claim-extractor, figure-reviewer,

report-producer, skeptical-reviewer, …). All deterministic steps go through this CLI; the skills never invent file paths or schemas.

Repository layout

src/paper_toolkit/      # the CLI + library (no LLM calls)
  cli/                  # Typer groups (init, evidence, compose, check, figure, lit, refs, analysis, ...)
  models/               # Pydantic v2 schemas (paper.json, evidence graph, claims, ...)
  analysis/             # 6-stage pipeline: db, state, claims, synthesis, lift
  figures/              # JSON-spec → matplotlib renderer (Arial fallback, pdf.fonttype=42, 89/183mm)
  checkers/             # style, figures, claim-coverage, figure-qa
  refs/                 # bib parser + dedup engine
  lit/                  # arXiv / CrossRef / OpenAlex search
  typeset/              # main.tex composition + LaTeX log parsing
skills/
  agentsociety-analysis/
  agentsociety-generate-paper/
tests/                  # 293 unit + integration tests

Design principles

• No LLM in the toolkit. Every CLI verb is pure Python + stdlib + a few

scientific deps (matplotlib, pypdf, httpx). The agent decides what to write; the toolkit decides what's legal.

• Schemas at every boundary. Pydantic v2 with extra="forbid" on every

artifact: paper.json, evidence graph, claims, figure specs, check reports.

• Derived state, no phase machines. State files record facts; stages are

derived from artifact presence + schema validity, so the agent can't lie about progress.

• Idempotent bridges. paper figure register and

paper analysis lift-to-evidence are safe to re-run — they update existing rows in place.

Development

uv sync --extra dev
uv run pytest -q
uv run ruff check src tests

The full design history and per-phase implementation plans live on the dev branch (docs/design.md, docs/plans/).

License

MIT — see LICENSE.