Set up CoAutoResearch with a coding agent

These instructions are for the user’s coding agent. The requested outcome is a usable research workspace, ready for the user to define the research and choose when to start. Finish with a running, verified dashboard and its URL, plus the paper-export environment described in paper generation. Reuse the user’s existing installation and projects when available.

1. Identify the installation

If this repository is already open, use it. Otherwise clone https://github.com/YihongT/CoAutoResearch.git into an unused directory and read its README. Preserve existing checkouts and local changes; do not reset them or replace a global CLI as part of setup.

Read package.json to report the source version. When using npm instead, compare npm view co-auto-research version first. Do not claim the published package contains unreleased repository features.

The source CLI has no application dependencies or build step. Run it with node /absolute/path/to/CoAutoResearch/bin/auto-research.js; no npm link, global installation, Python virtual environment, or pip install is needed to start the dashboard. Individual research projects can need their own scientific dependencies later.

2. Check prerequisites and the selected backend

Verify Node.js 20+, Python 3.10+, and Git. Try python3, python, or Windows py -3; set COAUTO_PYTHON to an available supported executable when autodetection does not find it. Install missing prerequisites using the platform’s normal method, preserving existing versions and shell settings.

Reuse the user’s installed, authenticated Codex or Claude Code. If both are ready and the user has no preference, keep CoAutoResearch’s Codex default. Only one backend is needed.

Backend	Install	Login	Verify
Codex	`npm install -g @openai/codex@latest`	`codex login`	`codex --version` and `codex login status`
Claude Code, macOS/Linux/WSL	`curl -fsSL https://claude.ai/install.sh \| bash`	`claude auth login`	`claude --version` and `claude auth status`
Claude Code, Windows PowerShell	`irm https://claude.ai/install.ps1 \| iex`	`claude auth login`	`claude --version` and `claude auth status`

Ask the user to complete any interactive provider login. Do not read or copy credential files, log out an existing account, or ask for secrets in chat. Existing CLI authentication is sufficient. If the user chooses API billing, they can enter the appropriate key in Settings → Agent.

Run login on the host where CoAutoResearch runs. For headless Codex, the official alternative is codex login --device-auth. CLI executables outside PATH can be selected with COAUTO_CODEX or COAUTO_CLAUDE.

Verify current provider instructions when installation or login differs: Codex setup, Codex login, Claude setup, Claude login.

3. Start and verify the dashboard

From the source checkout, run:

node bin/auto-research.js doctor
node bin/auto-research.js ui

Keep the UI process running. If the agent’s command runner closes background processes when a task ends, start the command in a persistent terminal and tell the user which terminal must stay open. Use --port 8766 if the default port is already occupied. Only enable remote access if the user needs it; see remote setup.

Verify all of the following before reporting success:

The server remains alive and its printed local URL opens the dashboard.
The setup dialog identifies the chosen CLI and authentication correctly. The other backend may remain uninstalled. Use Refresh login and models after logging in or updating a CLI.
Create a disposable project in the UI, select the chosen backend, and confirm it opens. If the user already supplied a real project, open that instead. Do not start autoresearch merely to verify installation.
Settings → Agent shows models and compatible reasoning options. Save the chosen backend. If model discovery fails, resolve that error; do not substitute a guessed model ID or report the agent ready to run.
Return to the main research chat and verify its input is usable. Close and reopen the page to confirm the project persists.

Doctor checks tool installation; it does not prove provider access or model inference. Login and the browser checks above are separate acceptance steps. When a browser tool is unavailable, report precisely which check the user still needs to perform.

4. Prepare paper generation

Follow the Setup contract for the coding agent in paper-generation.md. Run the product paper-tools installer to install its four pinned skills, prepare the isolated Python environment, and verify a LaTeX-to-PDF build and page rendering. Reuse existing LaTeX and Poppler installations. Do not generate a research paper or start a trial as part of installation.

Keep the dashboard running while preparing these tools. Report dashboard readiness and paper-export readiness separately if a download, permission, or dependency prevents completing the latter. Installing skills alone is not a successful PDF setup. The dashboard itself still needs only the prerequisites above.

5. Leave the workspace ready

Report the installed/source version, dashboard URL, project folder, selected backend/model, the four skill locations and pinned revision, the paper-environment Python path, the PDF/render smoke-check result, and the exact command to reopen it. Tell the user to open Manuscript → Generate paper when they want a paper; the product handles generation through its internal agent. The default project folder is co-autoresearch-projects/ under the launch directory; respect an existing folder or explicit --projects-dir choice. Explain that the terminal owns the server and must remain open. Leave the dashboard ready for the user’s research brief and their decision to start autoresearch.