PaperWatch

Turn arXiv and conference paper floods into Codex-readable research packets.

无 API 成本的 Codex 论文情报流水线。

PaperWatch is a local-first literature intelligence workflow for researchers who want to keep up with fast-moving papers without building a paid model API pipeline.

It collects papers, builds reproducible indexes, prepares reading packets, downloads selected PDFs, extracts page-marked text, and then lets Codex do the actual reading inside your current session.

collect papers
  -> rank and index
  -> generate Codex packet
  -> download selected PDFs
  -> extract page-marked text
  -> Codex reads and writes the research report

Why PaperWatch Exists

Most paper automation tools blur an important boundary: they make a script generate fluent summaries and call that "reading."

PaperWatch keeps the boundary explicit.

Mechanical work	Research work
Fetch metadata from arXiv, OpenReview, CVF	Decide which papers matter
Store papers in SQLite	Compare directions and trends
Rank by keywords and heuristics	Read PDFs and judge contributions
Generate Markdown/CSV packets	Write deep summaries and learning plans
Download and extract PDFs	Explain limitations and next steps

Scripts prepare the evidence. Codex reads the evidence.

What You Can Do With It

• Build a daily arXiv paper briefing.
• Process full OpenReview venues such as ICLR, ICML, and NeurIPS.
• Collect CVF Open Access papers for CVPR, ICCV, and ECCV.
• Create full conference CSV indexes.
• Split papers into direction-specific reading lists.
• Download must-read PDFs into local folders.
• Extract PDF text with page markers for Codex.
• Use a bundled Codex skill to make the workflow reusable.

Install

git clone https://github.com/<your-name>/paperwatch.git
cd paperwatch

python -m venv .venv
.\.venv\Scripts\Activate.ps1
python -m pip install -e .

Optional extras:

# OpenReview Python client
python -m pip install -e ".[conference]"

# PDF extraction
python -m pip install -e ".[pdf]"

# Tests and skill validation
python -m pip install -e ".[dev]"

Quick Start

1. Make a Daily Paper Packet

Edit config.yaml, then run:

paperwatch daily --limit 12

PaperWatch writes:

reports/codex-pack-YYYY-MM-DD.md

Ask Codex:

Use the PaperWatch workflow.
Read reports/codex-pack-YYYY-MM-DD.md and write today's Chinese paper briefing.

2. Process an OpenReview Conference

For ICLR, ICML, NeurIPS, and similar venues:

paperwatch --config configs/iclr-2026-all.yaml conference-pack --limit all
paperwatch --config configs/iclr-2026-all.yaml conference-analyze

Typical outputs:

reports/conference-pack-YYYY-MM-DD.md
reports/iclr-2026-full-analysis-YYYY-MM-DD.md
reports/iclr-2026-all-papers-YYYY-MM-DD.csv

3. Process CVPR / ICCV / ECCV

CVF conferences use CVF Open Access pages instead of OpenReview:

python scripts/collect_cvf_conference.py `
  --conference CVPR `
  --year 2025 `
  --database data/paperwatch.sqlite

The script preserves raw HTML and parsed metadata so the collection can be audited.

4. Deep-Read Selected PDFs With Codex

Download selected PDFs:

python scripts/download_must_read_pdfs.py `
  --input-dir reports/iclr-2026-by-direction `
  --output-dir data/pdfs/iclr-2026-must-read

Extract page-marked text:

python scripts/extract_pdfs_for_codex_reading.py `
  --manifest data/pdfs/iclr-2026-must-read/manifest.csv `
  --output-dir data/extracted/iclr-2026-must-read-codex `
  --indices all `
  --max-pages all

Ask Codex:

Read the extracted files in data/extracted/iclr-2026-must-read-codex one by one.
For each paper, write a detailed Chinese summary based on the extracted text.
Do not use script-generated summaries as final content.

The Codex Skill

PaperWatch includes a reusable Codex skill:

skills/paperwatch-research/

Use it when asking Codex to:

• update a daily paper report;
• triage a top conference;
• download selected PDFs;
• extract PDF text;
• write detailed summaries from extracted papers;
• keep the no-API boundary clear.

The most important rule in the skill:

Script-generated summaries are not deep reading. Deep summaries require Codex to read a packet, PDF, or extracted PDF text.

Common Commands

paperwatch collect
paperwatch list --limit 10
paperwatch codex-pack --limit 12
paperwatch conference-pack --limit 30
paperwatch conference-analyze
paperwatch analyze --limit 20
paperwatch report

If the paperwatch command is not installed, use:

python -m paperwatch <command>

Repository Layout

src/paperwatch/              CLI, collectors, ranking, reports
scripts/                     conference and PDF helper scripts
configs/                     reusable source configs
skills/paperwatch-research/  Codex skill
docs/                        workflow and adapter documentation
examples/                    small safe examples
tests/                       unit tests

What Stays Local

PaperWatch is designed to keep generated and personal research artifacts out of git:

data/
reports/
*.sqlite
*.sqlite-journal

These folders may contain PDFs, extracted text, raw conference HTML, private notes, and generated reports.

Documentation

• Project definition
• Workflow guide
• Conference adapters
• Open-source checklist

Development

python -m pip install -e ".[dev]"
python -m pytest
python C:\Users\zb\.codex\skills\.system\skill-creator\scripts\quick_validate.py skills\paperwatch-research

Status

PaperWatch is early-stage research tooling. The workflow is useful today, but source adapters may need updates when arXiv, OpenReview, or CVF change their pages or APIs.

License

MIT. See LICENSE.