Quiver: Math Research Agent

A VS Code extension that unifies the mathematical research workflow -- from idea exploration to proof development to paper writing -- in a single tool.

Features

@math Chat Participant

Type @math in Copilot Chat for quick math queries with LaTeX rendering.

• @math what is a derived category? -- definitions with auto-fetched context
• @math /search homological stability -- search arXiv, Wikipedia, nLab
• @math /arxiv spectral sequences -- arXiv-specific search with BibTeX
• @math /studio -- open the Research Studio panel

Research Studio (Branching Dialogue)

A webview panel for deep, non-linear exploration:

• Branching conversations -- fork from any message to explore alternative ideas. Active branch is highlighted; sibling branches appear dimmed.
• Math personas -- select from algebraist, analyst, geometer, topologist, number-theorist, or logician. Each shapes the LLM's proof strategies and vocabulary.
• Multi-agent mode -- send a question to multiple personas in parallel, get individual perspectives plus a synthesized answer.
• RAG auto-retrieval -- math entities are detected and looked up across arXiv, nLab, and Wikipedia automatically (toggleable).
• Attached papers -- drag-and-drop PDFs, .tex files, or paste arXiv IDs to inject paper context into the conversation.
• Lean4 verification -- optionally verify proof steps against a local Lean4 installation.
• Session management -- create, rename, switch between, and delete research sessions.

Write Tab (Paper Assistant)

Draft paper sections from your research and inject them into .tex files:

• Select a .tex file and see its section structure
• "Draft from branch" generates LaTeX from a research conversation
• Insert content at any section position using VS Code's WorkspaceEdit API
• Citations auto-formatted as \cite{} with BibTeX appended to .bib files

LLM Providers

Supports four providers via the VS Code Language Model API or direct API keys:

API keys are stored securely in VS Code's SecretStorage. A guided onboarding wizard runs on first activation.

Getting Started

1. Install the extension from the VS Code Marketplace (or build from source)
2. On first launch, the onboarding wizard guides you through provider setup
3. Type @math in Copilot Chat, or open the Research Studio via the command palette (Math Research Agent: Open Panel)

Building from Source

git clone https://github.com/andyzengmath/Quiver-Math-Research-Agent.git
cd Quiver-Math-Research-Agent
npm install
npm run compile
npm run build:webview

Press F5 in VS Code to launch the Extension Development Host.

Running Tests

npm run test:unit      # Mocha unit tests (243 tests)
npm run test:webview   # Vitest + React Testing Library

Architecture

src/
  extension.ts           # Activation, command registration
  services.ts            # Service factory (createServices)
  onboarding.ts          # First-run setup wizard
  llm/                   # Multi-provider LLM abstraction
    types.ts             # LlmProvider interface, error types
    service.ts           # LlmService registry
    providers/           # OpenAI, Anthropic, Google, VS Code LM
  dialogue/              # Branching conversation data model
    types.ts             # DialogueNode, DialogueTree, AttachedPaper
    tree.ts              # TreeManager (CRUD, fork, branch switch)
    storage.ts           # JSON persistence to .math-agent/
    context.ts           # ContextBuilder (path assembly, trimming)
  knowledge/             # RAG knowledge retrieval
    arxiv.ts             # arXiv API client
    wikipedia.ts         # Wikipedia MediaWiki API client
    nlab.ts              # nLab HTML scraper
    cache.ts             # Query result cache (24h TTL)
    entity-detector.ts   # Math entity detection (keyword + LLM)
    rag-orchestrator.ts  # Parallel source querying
  persona/               # Math persona system
    manager.ts           # 6 built-in + custom personas
  lean4/                 # Optional Lean4 integration
    service.ts           # Binary detection, verification
  papers/                # Paper attachment + TeX parsing
    manager.ts           # PDF/TeX/arXiv text extraction
    tex-parser.ts        # Section parsing, .bib detection
  chat/                  # Copilot Chat integration
    participant.ts       # @math chat participant
    multi-agent.ts       # Parallel persona orchestration
  webview/               # Research Studio panel
    panel.ts             # WebviewPanel lifecycle
    protocol.ts          # Typed message protocol
    message-handler.ts   # Handler registry pattern
    handlers/            # Feature-specific handlers
      send-handler.ts    # Message send + streaming
      branch-handler.ts  # Fork, switch, delete branches
      persona-handler.ts # Persona selection
      model-handler.ts   # Provider/model switching
      paper-handler.ts   # Paper attachment management
      lean4-handler.ts   # Lean4 verification UI
      write-handler.ts   # Paper drafting + file injection
      tree-handler.ts    # Session CRUD
webview-ui/              # React app for the webview
  src/
    App.tsx              # Tab navigation (Research / Write)
    components/          # UI components
    hooks/               # Custom React hooks
    utils/               # Markdown + KaTeX rendering

Settings

Data Storage

Research sessions are stored as JSON in .math-agent/trees/ within your workspace. This directory is auto-added to .gitignore. Context trimming memory files are saved to .math-agent/trees/{treeId}/memory/.

License

MIT