⚠️ Disclaimer. This repo grows on the fly out of my own thoughts and needs around AI assistance in computational research. It is a working draft, in need of human revision and selective improvement. Helpers and testers welcome!

🤖 Computational Research

A Wolfram-centric Codex and ChatGPT plugin for AI-assisted computational research. Original: Pavel Hajek’s Computational Research plugin.

• ⚙️ Autosetup installs the MCP servers and tools, and checks the environment.
• 📁 Autoorganization sorts a project into folders and keeps it tidy.
• 📚 Autoknowledge grows and maintains a wiki of what the project knows.
• 🔍 Autoresources gathers resources and keeps track of how to recover them.
• 🐺 Converts Wolfram notebooks to and from Markdown.
• 📓 Writes expository notebooks, and mathematics papers as notebooks.
• 📦 Turns code into a documented paclet and publishes it.
• 📝 Supports human-owned papers and paclet designs, with concise research companions and edits on request.
• 📐 Formalizes proofs in Lean.
• 🧮 Paper verification builds code and tests beside a mathematics paper and checks it statement by statement.
• 📔 Keeps a scientific journal of what was learned, if you want one.
• 🧬 Records the prompt behind everything it generates, if you want that.
• 🗃️ Autolab manages a backlog of work and works through it while you are away.
• 🤝 Shows you every deliverable, and revises it from the notes you write into it, keeping every version and remembering your corrections.
• 🧭 Gives a guided tour of the project.

🔄 Workflows

1. Organization of materials and resources

Organize your folder into a structure maintained by the LLM, or scaffold a new project from a research question.

Project/
  Code/          code items
  <PacletName>/  optional paclet folder
  Research/      research items
  Wiki/          knowledge base maintained by the LLM
  Work/          backlog managed by the LLM
  Resources/     downloaded material

The project is version-controlled with Git. Resources are downloaded and stored locally, with recovery instructions committed to Git instead of the downloaded files. Within each folder, Artifacts/ holds notes and notebooks on individual topics, while Archive/ holds superseded material. Documents intended as outputs are developed iteratively in the folder itself, as described below.

2. Assisted writing of output documents

The author develops the main document directly, in numbered iterations such as document1.tex, document2.tex, and so on. For each iteration, the LLM maintains a matching notes file, such as document1_notes.tex, following the document’s structure while exploring related questions and proposing additions.

The author decides what to incorporate and when to begin the next iteration. The LLM edits the main document only on explicit request.

See the document ownership rules.

3. Assisted paclet development

Paclet development follows the same workflow, with a design notebook as the main document. For a paclet named Infrageometry, keep Infrageometry1.nb and Infrageometry1_notes.nb in the repository root beside the paclet folder. The author develops the design in the notebook; the LLM proposes additions in the notes and implements the accepted design.

See the paclet design rules.

4. Backlog management and execution

Define and refine work items, have agents execute them in parallel in separate worktrees, and review the results. Work items follow the Kanban lifecycle Backlog → Ready → Active → UnderReview → Done, with each stage represented by a folder. The plugin can deploy a web control board for managing the backlog.

See the work-item format and execution details.

📥 Installation

This package uses the OpenAI plugin format and includes a repository marketplace. Add it as a source:

codex plugin marketplace add p135246/CodexPluginComputationalResearch

Then open the Plugins Directory in the ChatGPT desktop app, select Pavel Hajek — Research, install Computational Research, and test in a new chat. Availability of local marketplaces varies by surface. This repository is not a submission to the public OpenAI Plugins Directory.

For local development, clone this repository and add its absolute directory with codex plugin marketplace add /absolute/path/to/CodexPluginComputationalResearch. Invoke skills by name ($project-create, $document-propose) or describe the task.

Wolfram work requires an existing Wolfram Engine/Mathematica and a connected Wolfram AgentTools MCP server. Connect it through the host's MCP settings using the official server's instructions for that host; this package does not configure or launch another kernel. Lean and arXiv integrations are optional and discovered when needed. Run $project-check-env to check capabilities and license headroom.

Local scaffolding, builds, the board and CLI workers require filesystem and shell access. A web-only ChatGPT session needs those capabilities supplied by its environment. The board is a standalone local snapshot with a pending-change export; changes are reviewed and reconciled before applying to project files. Autolab uses native subagents when available, and the scheduled workflow uses bounded codex exec sessions. See runtime guidance for host differences and verification limits.

🧩 Functionality

⚙️ Autosetup

📁 Autoorganization

📚 Autoknowledge

🔍 Autoresources

🐺 Notebook and Markdown conversion

📓 Expository notebooks and mathematics papers

📦 Paclet development and publishing

📝 Human-owned papers and paclet designs

📐 Proof formalization in Lean

🧮 Paper verification

📔 Scientific journal

🧬 Prompt provenance

🗃️ Autolab

🤝 Document review and revision

🧭 Guided project tours

📄 License

Code: MIT. Ideas: CC BY 4.0.