managarten/CLAUDE.md
Till JS 7901a5df2f docs(local-llm): document browser-local LLM stack and CSP requirements
Add packages/local-llm/CLAUDE.md as the package-level reference for
browser-local LLM inference. The package went through a non-trivial
engine swap from WebLLM/Qwen to transformers.js/Gemma 4 E2B on
2026-04-08, and the bring-up surfaced enough sharp edges that the
next person (or AI agent) touching this code will save real time
having them written down in one place rather than re-discovering
them error by error.

Captured topics:

- What the package is, what library/model is currently used, and
  the deliberate engine-agnostic API surface that lets future swaps
  stay contained to this package.

- Why we chose transformers.js + Gemma 4 over staying on WebLLM
  (MLC compilation lag for new model architectures) and what the
  return path looks like once MLC ships Gemma 4 builds.

- The seven CSP directives that browser-local inference needs and
  WHY each one is required:
    * script-src: 'wasm-unsafe-eval', cdn.jsdelivr.net, blob:
    * connect-src: huggingface.co + *.huggingface.co + cdn-lfs-*,
                   *.hf.co + cas-bridge.xethub.hf.co (XET CDN),
                   cdn.jsdelivr.net (for the WASM preload fetch)
  Including the subtle "jsDelivr is needed in BOTH script-src and
  connect-src" trap that produces identical-looking error messages
  for two distinct underlying causes.

- The Vite SSR module-cache gotcha: CSP additions made in
  packages/shared-utils/security-headers.ts do NOT hot-reload across
  the workspace package boundary, while additions made directly in
  apps/mana/apps/web/src/hooks.server.ts do. Includes the diagnostic
  pattern (compare which additions show up in the next CSP error
  vs which don't) and the workaround (move them into hooks.server.ts
  via setSecurityHeaders options).

- The two-step tokenization pattern that's mandatory for
  Gemma4Processor: apply_chat_template(tokenize:false) → string, then
  processor.tokenizer(text, return_tensors:'pt'). The collapsed
  apply_chat_template(return_dict:true) path looks shorter but
  produces a malformed input shape and crashes model.generate() deep
  inside the forward pass with "Cannot read properties of null
  (reading 'dims')" — opaque from the call site.

- The transformers.js v4 quirk that model.generate() returns null
  (not a tensor) when a TextStreamer is attached. The streamer is
  the only stable text channel; the engine always attaches one and
  uses the streamer's collected text as the canonical output, with
  a chars/4 fallback for token counts.

- API surface (Svelte 5 example), how to add a new model to the
  registry, deploy notes (no base image rebuild needed for local-llm
  changes alone, but IS needed if shared-utils CSP defaults change),
  browser cache semantics, and hard browser support requirements
  (WebGPU, ~1.5–2 GB VRAM for E2B q4f16, no CPU/WASM fallback).

Also link to the new doc from the root CLAUDE.md Shared Packages
table so people land on it from the standard discovery path.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-04-08 23:27:50 +02:00

178 lines
9.7 KiB
Markdown

# CLAUDE.md
Guidance for Claude Code when working in this repo.
## Monorepo Overview
pnpm workspace monorepo with two consolidated tops:
- **`apps/mana/apps/web`** — unified SvelteKit frontend serving 27+ product modules under `mana.how`. One build, one IndexedDB, one auth session, one deployment.
- **`apps/api`** (`@mana/api`) — unified Hono/Bun backend API server. Consolidates per-module compute servers; routes registered under `/api/v1/{module}/*`.
Per-product directories under `apps/{product}/` still exist for landing pages, mobile apps, and product-specific packages, but the active web frontend and API both live in the two consolidated apps above.
- **Package Manager:** pnpm 9.15.0
- **Build System:** Turborepo
- **Node:** 20+
- **Primary doc:** [`apps/mana/CLAUDE.md`](apps/mana/CLAUDE.md) — module structure, data layer, encryption, routing.
### Repo layout
```
apps/
├── mana/ # Unified frontend (SvelteKit web + Expo mobile + Astro landing)
├── api/ # Unified backend API (Hono/Bun) — @mana/api
├── {product}/ # Per-product landing pages, mobile apps, packages
│ # Standalone (own container, not unified): manavoxel
games/ # arcade, voxelava, whopixels, worldream
services/ # Backend services (Hono/Bun, Go, Python) — see list below
packages/ # Shared workspace packages (@mana/*)
docs/ # Long-form docs (deployment, hardware, postmortems, etc.)
.claude/guidelines/ # Coding conventions — read before changing code
```
### Active services (`services/`)
`mana-auth` (3001), `mana-sync` (3050), `mana-credits`, `mana-user`, `mana-subscriptions`, `mana-analytics`, `mana-search` (3021), `mana-crawler`, `mana-api-gateway`, `mana-notify`, `mana-media`, `mana-llm`, `mana-image-gen`, `mana-video-gen`, `mana-stt`, `mana-tts`, `mana-voice-bot`, `mana-events`, `mana-landing-builder`. Each non-trivial service has its own `CLAUDE.md`.
## Coding Guidelines
Always consult before changing code:
| Document | Purpose |
|----------|---------|
| [`.claude/GUIDELINES.md`](.claude/GUIDELINES.md) | Overview |
| [`.claude/guidelines/code-style.md`](.claude/guidelines/code-style.md) | Formatting, naming, linting |
| [`.claude/guidelines/sveltekit-web.md`](.claude/guidelines/sveltekit-web.md) | Svelte 5 runes, stores |
| [`.claude/guidelines/expo-mobile.md`](.claude/guidelines/expo-mobile.md) | React Native, NativeWind |
| [`.claude/guidelines/hono-server.md`](.claude/guidelines/hono-server.md) | Hono/Bun servers |
| [`.claude/guidelines/database.md`](.claude/guidelines/database.md) | Drizzle ORM, pgSchema |
| [`.claude/guidelines/authentication.md`](.claude/guidelines/authentication.md) | Mana Auth integration |
| [`.claude/guidelines/error-handling.md`](.claude/guidelines/error-handling.md) | Result types, error codes |
| [`.claude/guidelines/testing.md`](.claude/guidelines/testing.md) | Vitest, mock factories |
| [`.claude/guidelines/design-ux.md`](.claude/guidelines/design-ux.md) | UI patterns, a11y |
## Development Quick Start
See [`docs/LOCAL_DEVELOPMENT.md`](docs/LOCAL_DEVELOPMENT.md) for the full setup.
```bash
pnpm docker:up # PostgreSQL, Redis, MinIO
pnpm setup:env # Generate per-app .env files from .env.development
pnpm setup:db # Create databases + push schemas
# Start the unified Mana app (most common)
pnpm run mana:dev
# Project-specific full stack (auth + backend + web with auto DB setup)
pnpm dev:chat:full
pnpm dev:todo:full
pnpm dev:picture:full
# … one per project
# Service-only
pnpm dev:auth # mana-auth (3001)
pnpm dev:sync # mana-sync Go server (3050)
```
Quality:
```bash
pnpm run build
pnpm run type-check
pnpm run format
```
## Key Architecture Notes
These are the patterns that span the repo. Service-/app-specific details live in their own CLAUDE.md.
### Local-first data layer
The unified Mana app uses **one IndexedDB** (`mana`) with all 120+ collections. Module stores write directly to Dexie tables; hooks in `database.ts` track changes into `_pendingChanges` tagged by `appId`. The unified sync engine (`sync.ts`) groups by `appId` and pushes to `mana-sync` (Go, port 3050), which persists field-level LWW into PostgreSQL with RLS.
Full architecture, sprint history, threat model:
- [`apps/mana/apps/web/src/lib/data/DATA_LAYER_AUDIT.md`](apps/mana/apps/web/src/lib/data/DATA_LAYER_AUDIT.md)
- [`apps/mana/CLAUDE.md`](apps/mana/CLAUDE.md)
### At-rest encryption
Sensitive user content in 27 tables is AES-GCM-256 encrypted before hitting IndexedDB. Master key lives in `mana-auth`, KEK-wrapped (`MANA_AUTH_KEK` env, must be set in prod). Optional zero-knowledge mode via Settings → Sicherheit.
When touching sensitive fields:
1. Add the table to `apps/mana/apps/web/src/lib/data/crypto/registry.ts` with the field allowlist
2. `await encryptRecord(tableName, record)` before writes
3. `await decryptRecords(tableName, visible)` after Dexie reads, before the type converter
Default new user-typed fields to **encrypt**; default new IDs/timestamps/sort-keys to **plaintext**.
### Authentication
All servers use `@mana/shared-hono` with `authMiddleware()`. Tokens are EdDSA JWTs issued by `mana-auth` with claims `{sub, email, role, sid, tier, exp, iss, aud}`. Cross-app SSO works across `*.mana.how`. See [`.claude/guidelines/authentication.md`](.claude/guidelines/authentication.md) and `services/mana-auth/`.
**Adding an app to SSO** requires updating *all three*:
1. `trustedOrigins` in `services/mana-auth/src/auth/better-auth.config.ts`
2. `CORS_ORIGINS` for mana-auth in `docker-compose.macmini.yml`
3. Run `pnpm test -- src/auth/sso-config.spec.ts` from `services/mana-auth/`
### Access tiers
`guest < public < beta < alpha < founder`. Apps gate themselves via `requiredTier` in `packages/shared-branding/src/mana-apps.ts`; the JWT carries a `tier` claim; `AuthGate` enforces it client-side. Admin API at `PUT /api/v1/admin/users/:id/tier`.
### Database (PostgreSQL)
Two databases: **`mana_platform`** (all services + app server-side data, schema-isolated via `pgSchema`) and **`mana_sync`** (sync engine, write-heavy). Always use `pgSchema('name').table(...)`, never plain `pgTable()`. Adding a new schema: see [`.claude/guidelines/database.md`](.claude/guidelines/database.md).
### Object storage
MinIO (Docker, S3-compatible) in both local and prod. Console: http://localhost:9001 (`minioadmin`/`minioadmin`). Use `@mana/shared-storage` helpers. Pre-configured per-project buckets (`picture-storage`, `chat-storage`, `cards-storage`, …).
### Turborepo: avoid recursive turbo calls
**CRITICAL**: Parent workspace packages (e.g. `apps/chat/package.json`) must NEVER define `type-check`, `build`, or `lint` scripts that call `turbo run <task>`. Root turbo already orchestrates those — defining them in children causes infinite recursion (10+ minute hangs, thousands of duplicate tasks). Only `dev` is OK to delegate to turbo from a parent package, since it's persistent and typically scoped.
## Shared Packages (`packages/`)
| Package | Purpose |
|---------|---------|
| `@mana/shared-auth` | Client-side auth for web/mobile |
| `@mana/shared-hono` | Hono middleware (auth, health, errors) |
| `@mana/shared-storage` | S3/MinIO helpers |
| `@mana/shared-branding` | App registry, tiers, branding |
| `@mana/shared-types` | Common TS types |
| `@mana/shared-utils` | Utility functions |
| `@mana/shared-ui` | React Native UI components |
| `@mana/shared-theme` | Theme config |
| `@mana/shared-i18n` | i18n |
| `@mana/local-store` | Local-first store primitives — used by unified Mana, manavoxel, arcade, and shared-uload/-stores/-links |
| `@mana/local-llm` | Browser-local LLM inference (transformers.js + Gemma 4 E2B, WebGPU). Powers `/llm-test` and the playground module. See [`packages/local-llm/CLAUDE.md`](packages/local-llm/CLAUDE.md) for the CSP requirements and the transformers.js v4 gotchas. |
## Adding Dependencies
```bash
pnpm add -D <pkg> -w # Workspace root (dev tools)
pnpm add <pkg> --filter @mana/web # A specific app
pnpm add <pkg> --filter @mana/shared-utils # A shared package
```
## Environment Variables
Single source of truth: **`.env.development`** (committed). After editing, run `pnpm setup:env` to regenerate per-app `.env` files with the right prefixes (`EXPO_PUBLIC_*` for mobile, `PUBLIC_*` for SvelteKit, no prefix for Hono/Bun servers). Mapping logic in `scripts/generate-env.mjs`. Full guide: [`docs/ENVIRONMENT_VARIABLES.md`](docs/ENVIRONMENT_VARIABLES.md).
## Server Access
- **Production (Mac Mini):** `ssh mana-server` (Cloudflare Tunnel). See [`docs/MAC_MINI_SERVER.md`](docs/MAC_MINI_SERVER.md). Useful: `./scripts/mac-mini/status.sh`, `./scripts/mac-mini/deploy.sh`, `./scripts/mac-mini/build-app.sh <app>`.
- **GPU server (Windows, RTX 3090):** `ssh mana-gpu` (192.168.178.11, LAN only). Hosts STT/TTS/image-gen/video-gen/Ollama. See [`docs/WINDOWS_GPU_SERVER_SETUP.md`](docs/WINDOWS_GPU_SERVER_SETUP.md).
## Reference Docs
| Path | When you need it |
|------|------------------|
| [`apps/mana/CLAUDE.md`](apps/mana/CLAUDE.md) | **Default** — module pattern, routing, encryption usage |
| [`apps/mana/apps/web/src/lib/data/DATA_LAYER_AUDIT.md`](apps/mana/apps/web/src/lib/data/DATA_LAYER_AUDIT.md) | Sync engine deep-dive, encryption rollout, threat model |
| [`docs/LOCAL_DEVELOPMENT.md`](docs/LOCAL_DEVELOPMENT.md) | First-time setup, `dev:*:full` commands |
| [`docs/ENVIRONMENT_VARIABLES.md`](docs/ENVIRONMENT_VARIABLES.md) | All env vars |
| [`docs/DATABASE_MIGRATIONS.md`](docs/DATABASE_MIGRATIONS.md) | Migration workflow + rollback |
| [`docs/DEPLOYMENT.md`](docs/DEPLOYMENT.md) | Production deployment |
| [`docs/PORT_SCHEMA.md`](docs/PORT_SCHEMA.md) | Which service runs on which port |
| Service-specific `CLAUDE.md` files | Service internals |