Pdf

English | 简体中文

Research Copilot is a Trellis-style, research-native, multi-platform CLI (rc) for running academic research as a controlled, stateful workflow. It models each piece of work as a task with a generic lifecycle (planning -> in_progress -> verify -> completed) crossed with a research activity kind (literature, ideation, experiment, writing, polish, review, rebuttal). Steering is injection-driven: each turn a coding-agent hook runs rc context, injecting the current workflow state plus a deterministic next-step recommendation computed from the task graph — you decide what to do next, nothing is auto-created.

This rebuild supersedes the previous Claude Code plugin architecture (320+ skills / MCP servers shipped as a plugin). See the redesign spec for the locked decisions and rationale.

• Author: ldm2060
• Repo: https://github.com/ldm2060/research_copilot

Status

Phase 0 — shipped:

• packages/core — the pure engine: task model, lifecycle FSM, research-state recommender, verify checks, workflow parser, context builder.
• packages/cli — the rc command-line tool.
• packages/adapters — the platform registry plus the Claude Code configurator.
• research-kit/ — workflow.md, 10 neutral rc-* agent templates, config defaults, spec templates.
• Claude Code is the one fully shipped platform (class-1 push injection via a UserPromptSubmit hook).

Other platforms are designed in the adapter registry; their adapters land in later phases (see the matrix below).

Installation

Quick Start (npx - no installation)

npx @research-copilot/cli init --user your-name --claude

This automatically installs both the CLI and the plugin package.

NPM (Recommended)

npm install -g @research-copilot/cli

Alternative Methods

• pnpm: pnpm add -g @research-copilot/cli
• Yarn: yarn global add @research-copilot/cli
• From Source: Clone and build (see INSTALLATION.md)
• GitHub Releases: Download pre-built archives (see Releases)

See INSTALLATION.md for detailed installation instructions, platform-specific notes, and troubleshooting.

Note: The rc init command automatically installs @research-copilot/plugin, which provides 6 research workflow skills and 10 agents for Claude Code and other supported platforms.

Quick Start Guide

New to Research Copilot? Start here:

• Getting Started Guide — Step-by-step tutorial from installation to your first task
• 快速入门指南（中文） — 中文版快速入门教程

The rc commands

kind is one of: literature, ideation, experiment, writing, polish, review, rebuttal. Full flag reference and examples: docs/usage/commands.md.

Skills

Research Copilot provides 6 high-level skills for orchestrating research workflows:

See docs/skills.md for detailed usage.

Platform support matrix

Class-1 platforms support per-turn push injection; class-2 platforms cannot, so the agent re-echoes an Active task: breadcrumb each turn and re-resolves state. See docs/dev/architecture.md and docs/dev/adding-a-platform.md.

Relationship to the old plugin

Research Copilot was previously a Claude Code plugin (a marketplace bundle of 320+ skills, 10 agents, 6 Python MCP servers, and a SessionStart guard hook). That architecture is superseded by this TypeScript-full-stack, multi-platform rebuild. The decision log (D1–D8) and migration path are recorded in the redesign spec and docs/dev/adr/0001-trellis-emulation.md.

Documentation

• Usage (researchers): docs/usage/ — command reference, Claude Code setup, a full workflow walkthrough.
• Development (contributors): docs/dev/architecture.md, docs/dev/core-api.md, docs/dev/adding-a-platform.md, docs/dev/testing.md, docs/dev/adr/0001-trellis-emulation.md.