ARC

August 21, 2026 ยท View on GitHub

Agent Research Copilot (ARC) is the research layer for theoretical-physics papers. It provides academic-paper access and source-aware research-domain, idea, planning, calculation, and checking workflows.

ARC v2 depends on AC Foundation for durable jobs, model calls, neutral documents, and proposer-reviewer orchestration. OCR proofreading, translation, Companion readers, and HTML publication belong to ALC.

Packages

  • arc-paper: arXiv, INSPIRE, DOI, paper caching, structural reads, citation traversal, and summaries.
  • arc-domain: evidence-bearing research-domain construction and exports.

The public arc-paper API is unchanged by the v2 repository split. Foundation commands remain available through arc-runtime; the plugin exposes only the arc-runtime, arc-paper, and arc-domain wrappers.

Install

ARC is distributed as an agent plugin with a lazy, SHA-locked private runtime. The runtime installs exact Git revisions of AC Foundation and ARC; no PyPI publication is assumed.

For Codex:

codex plugin marketplace add tririver/arc --ref stable
codex plugin add arc@arc

For Claude Code:

/plugin marketplace add tririver/arc@stable
/plugin install arc

For DeepSeek Harness:

dsh plugin --profile arc add github:tririver/arc

Check or prewarm the locked runtime:

plugins/arc/bin/arc-runtime doctor
plugins/arc/bin/arc-runtime setup

The default paper cache is .arc/cache/arc-paper below the launch directory; override it with ARC_PAPER_CACHE. Foundation runtime and neutral document state use AC_HOME, AC_RUNTIME_HOME, and AC_DOCUMENT_CACHE.

Use

Ask an installed agent directly for ordinary paper outcomes:

Summarize arXiv:0911.3380.
Find its references and citing papers.
Show the context around equation 2.30.

Name ARC for managed research workflows:

Use ARC to build a domain from arXiv:0911.3380 with papers since 2024.
Use ARC to develop and review ideas from that domain.
Use ARC to check this calculation.

Citation

If ARC contributes to your research, please cite:

Yanjiao Ma, Yi Wang, and Xingkai Zhang. ARC: An LLM-Native Agent Workflow for Theoretical Physics Research. ChinaXiv:202606.00234, 2026. https://chinaxiv.org/abs/202606.00234

Development

Python 3.11 or newer is required. Read AGENTS.md, keep generated work below ignored local/, and run:

AC_FOUNDATION_REPO_ROOT=../ac-foundation python scripts/check-generated-foundation.py
PYTHONPATH="$(find ../ac-foundation/packages packages -mindepth 2 -maxdepth 2 -type d -name src -print | paste -sd: -)" \
  python -m pytest --import-mode=importlib packages/*/tests tests
scripts/build-packages.sh

Prepare an approved release from a clean checkout with scripts/release-arc.sh <version>. The script updates both packages and plugin manifests, validates them, and pins the release source commit; it does not tag or publish.