Track 3 — Build the LaTeX demo from the starter

Goal: run the LaTeX starter in the browser workspace, see a rendered image, then complete the trainee exercises. Full staged path is in the roadmap.

1. Open the workspace + clone the project

Open browser VS Code → https://aiworkspace.konomondo.org (same container as Track 2 — Node 22 + TeX Live + Claude Code). Open a terminal (Terminal → New Terminal) and fetch the projects, then enter the LaTeX one:

clone-projects                 # clones latex-starter + hermes into ~/projects
cd projects/latex-starter

2. Prove rendering works (no LLM)

npm install
npm run smoke

Expect runtime/outputs/smoke/latex.svg. This confirms latex + dvisvgm work before any model is involved.

3. Run the agent + UI

npm run dev

The workspace already has your OLLAMA_API_KEY in its environment, so the agent just works. The dev server listens on port 3583 inside the container (published on the server as 3584) — open it from VS Code’s Ports panel, which forwards it for you. Keep or edit the prompt, click Run agent: the agent calls render_latex and the image appears in a few seconds. Watch the terminal — Flue streams the tool calls there.

If the agent replies in prose instead of calling the tool, set FLUE_MODEL (in the project’s .env) to a stronger tool-capable cloud model (gpt-oss / qwen3 / glm).

4. Build toward the architecture

The goal is to grow this seed into the latex-mcp MCP architecture — follow the staged roadmap (expose the render over MCP → drive it from an orchestrator → design for scale → test + review).

The // TODO(trainee): markers in the code are good warm-ups before the MCP work: add an explain_steps tool (src/shared/render-tools.ts), a Regenerate button (public/index.html), and extend the promptfoo tests (npx promptfoo@latest eval).

Do it all spec-first with Claude Code: brainstorm → plan → execute → test.

Done when: you’ve rendered an image from a prompt, and you can state the next roadmap stage (expose render_latex as an MCP tool) and how you’d start it.