# PluginCompass Provider Management & Fallback System

## Overview

This document describes the architecture and functionality of the PluginCompass AI provider management system, including the admin panel configuration, provider limits, usage tracking, and fallback mechanisms.

---

## 1. Admin Panel Structure

### 1.1 Admin Panel Sections

The PluginCompass admin panel is accessible at `/admin` and provides the following management areas:

**Main Admin Areas:**
- **Build Models** (`/admin/build`) - Configure AI models available to users
- **Plan Models** (`/admin/plan`) - Configure planning provider chain
- **Plans** (`/admin/plans`) - Manage subscription plans and pricing
- **Accounts** (`/admin/accounts`) - User account management
- **Affiliates** (`/admin/affiliates`) - Affiliate program management
- **Withdrawals** (`/admin/withdrawals`) - Affiliate payout management
- **Tracking** (`/admin/tracking`) - Analytics and usage statistics
- **Resources** (`/admin/resources`) - System resource monitoring
- **External Testing** (`/admin/external-testing`) - WordPress testing configuration
- **Contact Messages** (`/admin/contact-messages`) - Customer inquiries

### 1.2 Admin Authentication

The admin panel uses session-based authentication with the following security measures:

- **Credentials**: Configured via environment variables `ADMIN_USER` and `ADMIN_PASSWORD`
- **Session Duration**: 24 hours (configurable via `ADMIN_SESSION_TTL_MS`)
- **Rate Limiting**: Maximum 5 login attempts per minute per IP
- **Account Lockout**: 15 minutes after failed attempts

**API Authentication:**
All admin API endpoints require authentication via session cookies. The endpoints include:
- Login: `POST /api/admin/login`
- Logout: `POST /api/admin/logout`
- Session check: `GET /api/admin/me`

---

## 2. Provider Management

### 2.1 Supported Providers

PluginCompass supports multiple AI providers for both planning and building:

**Build Providers:**
- OpenRouter (primary aggregator)
- Mistral
- Google (Gemini)
- Groq
- NVIDIA (NIM)
- Chutes AI
- OpenCode (internal/self-hosted)
- Ollama (self-hosted planning)

**Planning Providers:**
- OpenRouter
- Mistral
- Google (Gemini)
- Groq
- NVIDIA (NIM)
- Ollama (self-hosted)

### 2.2 Provider Configuration

Each provider requires API credentials configured via environment variables:

| Provider | Environment Variable | Default API URL |
|---------|---------------------|----------------|
| OpenRouter | `OPENROUTER_API_KEY` | `https://openrouter.ai/api/v1` |
| Mistral | `MISTRAL_API_KEY` | `https://api.mistral.ai/v1` |
| Google | `GOOGLE_API_KEY` | `https://generativelanguage.googleapis.com/v1beta2` |
| Groq | `GROQ_API_KEY` | `https://api.groq.com/openai/v1/chat/completions` |
| NVIDIA | `NVIDIA_API_KEY` | `https://api.nvidia.com/v1` |
| Chutes AI | `CHUTES_API_KEY` or `PLUGIN_COMPASS_CHUTES_API_KEY` | `https://api.chutes.ai/v1` |
| Ollama | `OLLAMA_API_URL` | Configurable self-hosted URL |

### 2.3 Model Discovery

The system automatically discovers available models from each provider:

1. **CLI-based Discovery**: Queries OpenCode CLI for available models
2. **Provider API Discovery**: Fetches model lists directly from provider APIs
3. **Manual Configuration**: Admin can manually add model configurations

**Model Configuration per Provider:**
Each configured model includes:
- Model identifier (provider/model format)
- Display label for users
- Tier classification (free, plus, pro)
- Icon association
- Provider priority order
- Media support flag (image uploads)

---

## 3. Provider Limits & Usage Tracking

### 3.1 Rate Limiting System

PluginCompass implements a flexible rate limiting system with the following components:

**Limit Types:**
- **Tokens per Minute (TPM)**: Token consumption rate limit
- **Tokens per Day (TPD)**: Daily token consumption limit
- **Requests per Minute (RPM)**: API call rate limit
- **Requests per Day (RPD)**: Daily API call limit

**Scope Levels:**
- **Per Provider**: Limits apply to all models from that provider
- **Per Model**: Limits apply to specific models only

**Default Behavior:**
- All limits default to 0 (unlimited)
- Limits are configurable per provider or per model
- Usage is tracked independently for each provider

### 3.2 Usage Tracking

The system tracks usage in real-time with the following characteristics:

**Tracked Metrics:**
- Tokens consumed per request
- Number of API requests
- Timestamps for rate window calculation
- Per-model breakdown when scoped

**Data Retention:**
- Usage data retained for 48 hours for rate limiting
- Aggregated statistics persisted for reporting
- Separate tracking for planning vs building

**State Files:**
- `provider-limits.json`: Saved limit configurations
- `provider-usage.json`: Recent usage data
- `token-usage.json`: User token consumption

### 3.3 Limit Enforcement

When a request is made, the system:

1. Identifies the provider and model
2. Checks applicable limits (provider-level or model-level)
3. Compares current usage against limits
4. Returns rate limit error if exceeded
5. Records usage after successful requests

---

## 4. Fallback System

### 4.1 Fallback Architecture

PluginCompass implements a multi-level fallback system for reliability:

**Fallback Levels:**
1. **Model-level Fallback**: Alternative models within same provider
2. **Provider-level Fallback**: Switch to alternative providers
3. **Ultimate Backup**: Final fallback model when all providers fail

### 4.2 Model Fallback Chain

**For Each Provider:**
Each provider has a configured fallback chain:

**OpenRouter:**
```
Primary Model → Backup 1 → Backup 2 → Backup 3 → Environment Fallbacks → Static Fallbacks
```

**Mistral:**
```
Primary Model → Backup 1 → Backup 2 → Backup 3 → Default (mistral-large-latest)
```

**Groq:**
```
llama-3.3-70b-versatile → mixtral-8x7b-32768 → llama-3.1-70b-versatile
```

**Google (Gemini):**
```
gemini-1.5-flash → gemini-1.5-pro → gemini-pro
```

**NVIDIA (NIM):**
```
meta/llama-3.1-70b-instruct → meta/llama-3.1-8b-instruct
```

### 4.3 Planning Chain Fallback

The planning system follows a configured priority chain:

1. Attempts first provider in chain
2. If rate limited or error occurs, moves to next provider
3. Continues through all configured providers
4. Returns error if all providers fail

**Planning Chain Configuration:**
- Configurable via admin panel
- Each entry: `{ provider, model }`
- Priority determines fallback order
- Supports provider prefix in model names (e.g., "groq/compound-mini")

### 4.4 Build Fallback Chain

For building operations, the system follows this sequence:

1. **Primary Model**: User-selected or auto-assigned model
2. **Provider Fallback**: Alternative providers with same model
3. **OpenCode Fallback**: Internal OpenCode processing
4. **Ultimate Backup**: Configured backup model (last resort)

### 4.5 Error Classification & Fallback Decision

Errors are classified to determine fallback behavior:

| Error Type | Example | Fallback Action |
|-----------|---------|-----------------|
| Rate Limit (429) | "Too many requests" | Wait 30s, switch provider |
| Server Error (5xx) | "Internal error" | Wait 30s, switch provider |
| Auth Error (401) | "Invalid API key" | Switch immediately |
| Billing Error (402) | "Insufficient credits" | Switch immediately |
| Model Not Found (404) | "Unknown model" | Wait 30s, switch model |
| User Error (400) | "Invalid request" | Return error, no fallback |
| Token Limit | "Context length exceeded" | Return error, no fallback |

**Continue Mechanism:**
- For early terminations, system sends "continue" message
- Retries up to 3 times with same model
- After 3 failures, switches to fallback model

---

## 5. Admin Configuration Interface

### 5.1 Model Management

The admin panel allows configuration of:

- **Add/Update Models**: Select from discovered models or add custom
- **Provider Priority**: Drag-and-drop reordering of provider fallback order
- **Tier Assignment**: free (1x), plus (2x), pro (3x) multipliers
- **Icon Selection**: Associate icons with models
- **Media Support**: Enable/disable image upload capability

### 5.2 Provider Limits Configuration

The limits interface provides:

- **Provider Selection**: Dropdown for each configured provider
- **Scope Selection**: Per-provider or per-model limits
- **Limit Inputs**: Numeric fields for TPM, TPD, RPM, RPD
- **Live Usage Display**: Current usage statistics per provider
- **Save/Reset**: Persist or revert limit changes

### 5.3 Ultimate Backup Configuration

The admin can configure a final fallback model that will be used when all other providers fail. This ensures system availability even during widespread outages.

---

## 6. Configuration Files

### 6.1 Environment Variables

Key configuration is done via environment variables:

```bash
# Provider API Keys
OPENROUTER_API_KEY=
MISTRAL_API_KEY=
GOOGLE_API_KEY=
GROQ_API_KEY=
NVIDIA_API_KEY=
CHUTES_API_KEY=

# Admin Authentication
ADMIN_USER=
ADMIN_PASSWORD=

# Rate Limiting
ADMIN_LOGIN_RATE_LIMIT=5
USER_LOGIN_RATE_LIMIT=10
API_RATE_LIMIT=100
```

### 6.2 Runtime State

The system maintains runtime state in:

- `.data/.opencode-chat/provider-limits.json` - Persisted limits
- `.data/.opencode-chat/provider-usage.json` - Recent usage
- In-memory state for active sessions and rate tracking

---

## 7. Security Considerations

### 7.1 Rate Limiting

- **Login Protection**: 5 attempts/minute, 15-minute lockout
- **API Protection**: 100 requests/minute per user
- **Provider Protection**: Configurable limits prevent abuse

### 7.2 Authentication

- Session-based auth with secure cookies
- OAuth support for Google and GitHub
- Rate-limited login attempts
- Session timeout enforcement

### 7.3 Data Protection

- Provider API keys stored securely
- Usage data retained only as needed
- No sensitive data in logs
- Encrypted session storage

---

## 8. Monitoring & Analytics

### 8.1 Tracking Metrics

The system tracks:

- **User Analytics**: Session duration, feature usage, model preferences
- **Business Metrics**: MRR, LTV, churn rate, CAC
- **Technical Metrics**: AI response times, error rates, queue wait times
- **Provider Metrics**: Per-provider usage and error rates

### 8.2 Admin Dashboard

The tracking page provides:

- Daily/weekly/monthly active users
- Revenue analytics
- Conversion funnels
- Error rate monitoring
- Resource utilization

---

## 9. Summary

PluginCompass provides a robust, multi-provider AI infrastructure with:

1. **Flexible Provider Management**: Support for 6+ AI providers with automatic model discovery
2. **Granular Rate Limiting**: Per-provider and per-model limits with configurable thresholds
3. **Intelligent Fallback**: Multi-level fallback chains ensure high availability
4. **Comprehensive Admin Control**: Full configuration through web-based admin panel
5. **Usage Tracking**: Real-time monitoring of token consumption and API usage
6. **Security Measures**: Rate limiting, authentication, and session management

This architecture ensures reliable AI-powered development while maintaining control over costs and system availability.