End-to-end cookbooks

The repository's notebooks progress from credential-free execution to live model generation. Launch them from the repository root:

uv sync --frozen --all-groups
uv run jupyter notebook cookbook/

Learning path

1. Credential-free quickstart validates, inspects, and executes a complete program with FakeToolProvider.
2. Custom provider implements ToolProvider, runs the reusable provider contract, and executes a program through the extension.
3. Calculator connects to a local MCP server and asks a configured OpenAI model to generate arithmetic, higher-order, and conditional programs.
4. Memory adds conversational memory to model-backed generation.
5. Game statistics composes local MCP tools into a data-processing workflow.

The first two tutorials need no credentials or network access and execute fully in normal CI. The remaining notebooks require OPENAI_API_KEY; CI validates their notebook structure, cleanliness, Python syntax, server paths, and deterministic MCP tools without making model calls.

CI execution contract

A notebook opts into credential-free execution with:

{
  "metadata": {
    "treelang": {
      "ci_execute": true
    }
  }
}

make cookbooks rejects committed cell outputs or execution counts, compiles every code cell, and executes each opted-in notebook in memory from the repository root. The executed notebook is never written back to the working tree.

When adding a tutorial:

1. State its learning goal and prerequisites in the first Markdown cell.
2. Keep every code cell independently readable and use assertions for important outcomes.
3. Prefer treelang.testing fakes for deterministic examples.
4. Mark it ci_execute only when it needs no credentials, network, or external service.
5. Clear all outputs and execution counts before committing.
6. Add it to this learning path and run make cookbooks.