wtf is deep-learning-auto-research?

deep_learning_auto_research is an opt-in workflow tool for Codex, an AI coding assistant. It takes a markdown system plan and breaks it into traceable implementation blocks. Each block gets relevant research papers attached, block-specific evidence extracted, and an implementation spec created. Coding is gated behind approved context, meaning the AI cannot start writing code until you explicitly approve the research and specs for each block. This is not a command-line tool you run directly. Instead, you start an MCP server (a background process that exposes tools to Codex), then ask Codex to use it with prompts like "Use deep_learning_auto_research" followed by a specific action. The server stores all workflow state and markdown files on your machine. You need Node.js and npm installed, and you build the server with npm install and npm run build before adding it to your Codex configuration. The workflow has 16 steps. You create a project, feed it your markdown plan, and have Codex propose semantic implementation blocks derived from your actual plan rather than generic phases like "setup" or "deployment." You review and approve those blocks, set a target language and framework (such as Python and PyTorch), then prepare research context for each block. Codex can search online for papers or you can attach your own. It extracts only research relevant to that specific block, you approve it, then a spec is generated from the block, papers, and extracted research. Only after you approve the spec does Codex implement that single block. Each project creates a structured folder with a .planner directory containing state, graph, and audit logs, plus a blocks directory where each block gets its own folder with markdown files for the block definition, papers, extracted research, spec, and implementation. The tool enforces dependencies, so a block cannot be implemented until the blocks it depends on are implemented or verified. A final synthesis step combines all verified blocks. The README provides prompt templates for every stage, with placeholders for project paths, block IDs, languages, and frameworks.