# CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

## Repository overview

This repository contains the ZeaVis Edu application: a corn leaf disease classifier with a machine-learning pipeline (EfficientNetV2B0 training and export), a Rust/Axum/ONNX Runtime inference service, and a fullstack TypeScript application (React frontend, Elysia backend, PostgreSQL).

The ML pipeline lives under `Machine_Learning/`. The inference service lives under `apps/ml-service/`. Most ML commands should be run from the `Machine_Learning/` directory unless noted otherwise.

## Common commands

```bash
cd Machine_Learning
```

Activate the Python environment (already exists at repo root):

```bash
source ../.venv/bin/activate
pip install -r requirements.txt
```

Run local dataset preprocessing after placing `dataset_1.zip`, `dataset_2.zip`, and `dataset_3.zip` beside `preprocessing.py`:

```bash
python preprocessing.py
```

Export a trained Keras model to SavedModel and TFLite after placing the Colab-trained model at `best_model/best_model.keras`:

```bash
python save_model.py
```

Convert the SavedModel export to ONNX for the Rust ML service:

```bash
python convert_onnx.py
```

Convert the SavedModel export to TensorFlow.js via CLI:

```bash
export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
tensorflowjs_converter \
  --input_format=tf_saved_model \
  --output_format=tfjs_graph_model \
  --signature_name=serving_default \
  --saved_model_tags=serve \
  model/saved_model \
  model/tfjs_model
```

Open the training notebook locally if needed:

```bash
jupyter notebook notebook.ipynb
```

There is no project test suite, lint command, or build system configured in the ML pipeline.

## Fullstack app commands

The TypeScript application scaffold lives at the repository root and uses Bun workspaces with Moon tasks.

Install dependencies:

```bash
bun install
```

Run all development tasks through Moon:

```bash
bun run dev
```

Run type checks:

```bash
bun run typecheck
```

Run production builds:

```bash
bun run build
```

Run the API directly:

```bash
cd apps/api && bun run start
```

Run the web app directly:

```bash
cd apps/web && bun run dev
```

Run the ML service directly:

```bash
cd apps/ml-service && cargo run
```

Run the Telemetry stack:

```bash
# Start all telemetry services (Prometheus, Ingester, Vector, ClickHouse, Query Proxy, Telemetry UI)
make telemetry-up

# Local dev mode (port bindings exposed)
make telemetry-up-local

# Check health of all telemetry services
make telemetry-status

# View telemetry logs
make telemetry-logs [s=<service>]

# Build telemetry components
make telemetry-build

# Send a test metric
make telemetry-test-metric

# Stop telemetry
make telemetry-down
```

## High-level architecture

- `Machine_Learning/preprocessing.py` prepares the training dataset locally. It extracts three source ZIP files, merges selected class folders into `dataset/`, maps selected Mandarin labels from Dataset 3 via `desc.json`, removes known problematic image files, then creates `dataset.zip` for upload to Google Drive/Colab.
- `Machine_Learning/notebook.ipynb` is the training workflow intended for Google Colab with GPU enabled. It trains an EfficientNetV2B0-based classifier and saves the best model to Google Drive as `best_model.keras`.
- `Machine_Learning/save_model.py` is the production export step. It loads `best_model/best_model.keras`, rebuilds a clean EfficientNetV2B0 architecture without training-time augmentation layers, copies weights into that model, exports `model/saved_model/`, and writes `model/model.tflite`.
- `Machine_Learning/convert_onnx.py` converts the SavedModel to ONNX format (`model/model.onnx`) for use by the Rust inference service.
- TensorFlow.js export is intentionally done with the `tensorflowjs_converter` CLI rather than from Python to avoid protobuf/runtime conflicts documented in the README.
- `apps/ml-service/` is a Rust/Axum service that loads the ONNX model and serves HTTP endpoints for health checks, metadata, and image classification predictions. It uses ONNX Runtime for cross-platform inference performance.

## Telemetry architecture

The telemetry stack lives as a git submodule at `telemetry/` (repo `MythEclipse/Telemetry`). Architecture:

| Layer | Service | Role |
|-------|---------|------|
| Collector & Storage | **Prometheus** | Metric scraping & TSDB storage |
| System metrics | **Node Exporter** | CPU, memory, disk per host |
| Query | **Query Proxy** | REST API over Prometheus HTTP API |
| Visualization | **Grafana** | OSS dashboard & PromQL |
| Entry point | **Telemetry UI** | nginx + Vue 3 SPA |

Data flow: Node Exporter → Prometheus scrape (every 15s) → Grafana (PromQL) / Query Proxy (/api/metrics).

Each ZeaVis Edu service exposes a `GET /metrics` endpoint:

- **Web app** (`apps/web`): In dev mode, a Vite plugin serves client-side session metrics (page views, Web Vitals). In production, nginx proxies `/metrics` to the API service. Source: `apps/web/src/lib/telemetry.ts`, `apps/web/vite-plugin-metrics.ts`.
- **API** (`apps/api`): Uses `prom-client` for Node.js default metrics plus custom HTTP, auth, classification, and diagnosis counters/histograms. Source: `apps/api/src/lib/telemetry.ts`, exposed via `apps/api/src/routes/metrics.ts`.
- **ML service** (`apps/ml-service`): Uses the `prometheus` Rust crate for HTTP metrics, prediction counts, and model load status. Source: `apps/ml-service/src/telemetry.rs`.

All three share the `zeavis_` metric prefix and are scraped by Prometheus via `file_sd_configs` (see `telemetry/prometheus/targets/zeavis-edu.json`).

**IMPORTANT — Production architecture:** ZeaVis Edu apps and the Telemetry stack run on **separate VPS instances** connected via **Tailscale** (mesh VPN). Prometheus scrapes the API and ML service through their **Tailscale IPs** (e.g. `100.x.x.a:3000`), not via Docker hostnames. The target file has `__CHANGE_ME__` placeholders — replace with actual Tailscale IPs before deploying.

The telemetry stack is managed from the project root via `make telemetry-*` targets (see `Makefile`). Docker Compose defines 5 services (Prometheus, Node Exporter, Query Proxy, Grafana, Telemetry UI).

For **local single-host dev**, Prometheus can reach app services via a shared Docker network (`app-shared-net`). Use `make telemetry-up-local` for this mode.

## Fullstack application architecture

The root TypeScript workspace is a Bun + Moon monorepo:

- `apps/web/` contains the React + Vite + TypeScript frontend with React Router, TanStack Query, Zustand, Tailwind, and shadcn/ui-style components.
- `apps/api/` contains the Elysia backend with health/status routes and Drizzle/PostgreSQL configuration.
- `apps/ml-service/` contains the Rust/Axum inference service with ONNX Runtime for model predictions.
- `packages/shared/` contains shared TypeScript types and utilities consumed by both apps.

The backend reads `DATABASE_URL` for Drizzle/PostgreSQL, but the initial health/status endpoints do not require a live database connection. The ML service reads `MODEL_PATH` (default `../../Machine_Learning/model/model.onnx`) and `MODEL_INPUT_SIZE` (default `224`).

## Model labels and dataset mapping

The classifier targets four Indonesian labels:

- `Bercak Daun` — Gray Leaf Spot
- `Hawar Daun` — Northern/Southern Leaf Blight
- `Karat Daun` — Common Rust
- `Daun Sehat` — healthy corn leaf

Dataset handling is part of the model logic:

- Dataset 1 contributes `Bercak Daun`, `Hawar Daun`, and `Daun Sehat`; its `Karat Daun` folder is intentionally ignored because the README states it is not representative.
- Dataset 2 contributes `Common_Rust` mapped to `Karat Daun` and `Healthy` mapped to `Daun Sehat`.
- Dataset 3 is routed through Mandarin label mappings in `PEMETAAN_KATEGORI` inside `preprocessing.py`.

## Important generated/local artifacts

The following files/directories are generated or externally supplied during the ML workflow and may not exist in a fresh clone:

- `Machine_Learning/dataset_1.zip`, `dataset_2.zip`, `dataset_3.zip` — manually downloaded source datasets.
- `Machine_Learning/dataset/` and `Machine_Learning/dataset.zip` — generated by `preprocessing.py`.
- `Machine_Learning/best_model/best_model.keras` — trained model downloaded from Colab/Google Drive.
- `Machine_Learning/model/saved_model/`, `model/model.tflite`, `model/model.onnx`, and `model/tfjs_model/` — production exports.

## Android Google OAuth (Tauri) — known issues & fixes

The Tauri Android app uses Chrome's `intent://` protocol to bounce back from Google's OAuth browser page. Three bugs were found and fixed in commit `c75cba2`:

### 1. API base URL falls back to `http://tauri.localhost`

**Symptom:** Google login button navigates to `http://tauri.localhost/api/v1/auth/google` → 404.
**Root cause:** `auth-form.tsx` used `import.meta.env.VITE_API_BASE_URL || window.location.origin`. In Android WebView the origin is `http://tauri.localhost` (Vite dev server), not the API server.
**Fix:** Import shared `apiBaseUrl` from `api-client.ts` which already has the correct fallback: `import.meta.env.VITE_API_BASE_URL ?? 'https://zeavisedu.asepharyana.my.id'`.

### 2. `deep-link:get_current` IPC promise orphaned on SPA navigation

**Symptom:** `Cannot read properties of undefined (reading 'runCallback')` floods log; OAuth never completes.
**Root cause:** `plugin:deep-link|get_current` returns a JS promise that stays pending. When React Router's `navigate()` changes the URL (SPA, no page reload), the Tauri IPC bridge invalidates the pending callback reference — but the promise never resolves or rejects cleanly, so `.runCallback` is undefined.
**Fix (cold start):** `get_current` resolves via `window.location.href = target` (full reload). At boot there is no SPA state to lose, so a hard redirect is safe.
**Fix (warm start / `deep-link://new-url` event):** Store target in `sessionStorage` + dispatch a custom DOM event. A `<DeepLinkRouterHandler>` root layout route listens for the event and calls React Router's `navigate()`, keeping SPA state alive.

### 3. LoginPage `?token=` effect does not re-run on SPA navigation

**Symptom:** App navigates to `/login?token=xxx` but stays on the login form.
**Root cause:** The `useEffect` that reads `?token` and exchanges it for a session only listed `[setUser, queryClient, navigate]` as deps. React Router SPA navigation changes `location.search` but does not remount the component — so the effect never re-runs.
**Fix:** Added `location.search` to the effect's dependency array. Also added `visibilitychange` and `focus` event listeners as a backup — when the user returns from the Google OAuth browser tab, the app picks up the token from the URL even if the deep-link plugin's event was missed.

### Design rule for Tauri deep-link handlers

- **Cold start** (app was not running) → safe to use `window.location.href` (full reload). The React app has just booted, no state to lose.
- **Warm start** (app was running, user returns from system browser) → use React Router `navigate()` via custom events / sessionStorage. Do NOT use `window.location.href` — it triggers a full page unload which orphan Tauri IPC promises.

## Notes for future changes

- Keep README command examples and this file in sync when changing the ML pipeline.
- Preserve the current class label names unless the training notebook, preprocessing mappings, and downstream app/API expectations are updated together.
- `save_model.py` assumes the clean architecture matches the trained model weights exactly; changes to the notebook model architecture usually require corresponding changes in `build_clean_model()`.
- The Rust ML service expects the ONNX model at the path specified by `MODEL_PATH`. Ensure `convert_onnx.py` is run after `save_model.py` to generate the ONNX artifact before deploying the service.