Gemma Chat - (For Windows, Linux, Raspberry Pi): Ollama backend + installers

[revdarkness]

Summary

Adds first-class Windows support by replacing the MLX-LM backend with a thin Ollama HTTP client. Mac/MLX path is left untouched in spirit (and could be reintroduced behind a platform check), but this PR is scoped as a Windows-only release line tagged v0.1.0-win.1.

• Delete src/main/mlx.ts (Python venv + mlx_lm.server subprocess manager)
• Add src/main/ollama.ts — locate/pull/warm + chat-stream client. Uses Ollama's OpenAI-compatible /v1/chat/completions so the SSE parser is identical to the original. Honors OLLAMA_HOST env var.
• src/shared/types.ts — AVAILABLE_MODELS now uses gemma3:{1b,4b,12b,27b} Ollama tags; DEFAULT_MODEL = gemma3:4b
• src/main/workspace.ts — wsRunBash runs PowerShell on Windows (GEMMA_SHELL=bash opt-out for Git Bash users); deny list extended for Windows destructive patterns
• src/main/index.ts — mac-only BrowserWindow options (titleBarStyle, vibrancy, trafficLightPosition) now guarded by process.platform === 'darwin'; icon picks .ico on win32
• electron-builder.yml — added win + nsis x64 target with desktop/start-menu shortcuts
• build/icon.ico generated from existing icon.png via png-to-ico
• package.json — dist:win / dist:mac convenience scripts; description updated
• README rewritten for Windows with Ollama install, model table, env vars, and a v0.2 llama.cpp roadmap note

The window-frame and shell choices are platform-conditional, so reintroducing the MLX backend on darwin in a future PR would be additive — none of the mac-only chrome was removed unconditionally.

Why Ollama instead of MLX

MLX is Apple-only. Ollama runs natively on Windows, ships its own GPU acceleration (CUDA / ROCm preview), and exposes the same OpenAI-compatible chat surface the original MLX server already targeted, so the chat-streaming code carried over with zero changes. It also removes the Python venv bootstrap entirely, which is the biggest source of first-run friction on Windows.

A v0.2 will add an optional bundled llama.cpp backend so users don't need to install Ollama separately.

Test plan

• npm run typecheck (node + web) passes on Windows 11
• npm run build (electron-vite) passes
• npm run dist:win produces dist/gemma-chat-0.1.0-win.1-setup.exe (NSIS, x64) — local smoke
• First-run flow: Ollama not installed → user-facing error with download link
• First-run flow: Ollama running, model not pulled → streaming pull progress in setup UI
• Chat mode: tokens stream, tool calls execute, web search + bash tool work under PowerShell
• Build mode: file writes land in workspace dir, live preview iframe updates
• Model switching via the header switcher pulls + warms a new tag
• Verified on macOS that the mac code path still compiles (BrowserWindow opts are spread conditionally)

Supply-chain note

Ran npm audit + lockfile-integrity scan + install-hook review on the fork: lockfile is clean (663/663 entries with sha512- SRI hashes, all from registry.npmjs.org), six lifecycle scripts present and all attributable to known maintainers (electron, esbuild, fsevents, onnxruntime-node, protobufjs warning-only, sharp). The 15 audit advisories are routine electron 34 / electron-builder 25 transitive noise, addressable via npm audit fix --force to bump electron → 42 / electron-builder → 26.

Credits

All original design, UX, agent loop, XML tool protocol, live-streaming write logic, and overall product taste are @ammaarreshi's. This PR is a mechanical port for the Windows side and a credit/link added to the README. Happy to split into smaller commits, drop the README changes, or reshape this as a Linux+Windows PR with a runtime backend switch if you'd rather — your call.

🤖 Generated with Claude Code

[revdarkness]

Smoke-tested locally on Windows 11 + Node 20.19. Two follow-up commits pushed to revdarkness:windows:

1. Node 20.19 ESM/CJS interop crash fix. With "type":"module" in package.json and the main bundle at out/main/index.js, Node treats the main process as ESM. Electron 34's embedded Node 20.19 then crashes in cjsPreparseModuleExports when our ESM bundle imports the CJS electron module — TypeError: Cannot read properties of undefined (reading 'exports'). Fix: emit main + preload as CommonJS with a .cjs extension via rollupOptions.output.format = 'cjs'. Renderer stays ESM/React/Vite as before. Same fix the wider electron-vite community has converged on for this Node version.

2. gemma4:latest added as the recommended default. Ollama now publishes Gemma 4 (8B Q4_K_M, 9.6 GB) as gemma4:latest, which lines up with the upstream README's original intent. The gemma3:* tags remain as smaller alternatives.

Verified end-to-end against gemma4:latest: streaming data: {...}\n\n SSE chunks ending with data: [DONE] parse cleanly through the existing chatStream() reader. App launches, Welcome → Setup → chat flow loads.

Happy to split the CJS fix into its own PR if you'd prefer it land independently of the Windows port — it benefits any Node 20.19+ user, including the macOS path.