reasoning-agent
Public
Advanced AI reasoning plugin that enables LM Studio to invoke Google Gemini models with automatic API key rotation, intelligent error handling, and quota management.
gemini-3-flash-preview)| Setting | Type | Default | Description |
|---|---|---|---|
| API Keys | string | (required) | Comma-separated Google AI API keys |
| Model Name | string | gemini-3-flash-preview | Gemini model to use |
| Max Tokens | number | 2048 | Maximum response length (100-10000) |
| Temperature | number | 0.7 | Response creativity (0-2) |
| Retry Attempts | number | 3 | Attempts per key before rotation |
| Request Timeout | number | 30000 ms | Timeout per request (1000-120000 ms) |
| Detailed Logging | boolean | false | Enable debug logging |
The plugin distinguishes between three error categories:
| Error Type | Trigger | Action |
|---|---|---|
| Quota Error | 429, rate_limit, quota exceeded | Immediately rotate to next key |
| Auth Error | 401, 403, invalid API key | Immediately rotate to next key |
| Transient Error | Network timeout, service unavailable | Retry (up to retryAttempts) |
Enable Detailed Logging in config to see:
Check LM Studio's console/logs for these messages while running in dev mode.
Before deploying to production, verify:
curl -X POST https://generativelanguage.googleapis.com/v1beta/models/gemini-3-flash-preview:generateContent?key=YOUR_KEYnpm install in plugin directorynpm run buildreasoning_invokeParameters:
Returns:
MIT License - Free for personal and commercial use
Found a bug? Want to improve the plugin?
For issues related to:
Future enhancements:
Version: 1.0.0
Last Updated: April 2026
LM Studio Compatibility: v0.2.28+
cd reasoning-agent
npm install
lms dev # For development with hot reload
# OR
lms push # To deploy to your LM Studio installation
key1,key2,key3
User: "Use the reasoning tool to analyze the implications of quantum computing on cybersecurity"
LM Studio: Calls reasoning_invoke with the prompt
Plugin:
1. Sends to Gemini API (key 1)
2. Gets response back
3. Formats as Markdown
AI: Returns formatted reasoning to user
User: Asks AI a complex question
AI: Calls reasoning_invoke → Plugin tries key 1
Key 1: Gets 429 quota error
Plugin: Automatically rotates to key 2
Key 2: Successfully returns response
AI: Uses the reasoning in its answer
reasoning_invoke(
prompt="Analyze this...",
useThinking=true,
temperature=0.3,
maxTokens=4096
)
┌─────────────────────────────────┐
│ AI Calls reasoning_invoke │
└────────────┬────────────────────┘
▼
┌─────────────────┐
│ Key 1: Attempt │
│ Retries: 1-3 │
└────────┬────────┘
│
┌──────▼──────┐
│ Success? ✓ │ YES ──→ Return Response
│ Quota? ✗ │
│ Auth? ✗ │
└──────┬──────┘ NO
▼
┌─────────────────┐
│ Key 2: Attempt │
│ Retries: 1-3 │
└────────┬────────┘
│
┌──────▼──────┐
│ Success? ✓ │ YES ──→ Return Response
└──────┬──────┘ NO
▼
┌─────────────────┐
│ Key 3: Attempt │
│ Retries: 1-3 │
└────────┬────────┘
│
┌──────▼──────┐
│ Success? ✓ │ YES ──→ Return Response
└──────┬──────┘ NO
▼
Return Error Message
(All keys exhausted)
[ReasoningAgent] Attempting request with key 1/3 (abc123***)
[ReasoningAgent] Using model: gemini-3-flash-preview, temperature: 0.7, maxTokens: 2048
[ReasoningAgent] Attempt 1/3 passed: 429 quota error (Will rotate key)
[ReasoningAgent] Attempting request with key 2/3 (def456***)
[ReasoningAgent] Successfully generated response (1245 chars)
reasoning-agent/
├── manifest.json # Plugin metadata
├── package.json # Dependencies & scripts
├── tsconfig.json # TypeScript compiler config
├── .gitignore # Git ignore patterns
├── README.md # This file
└── src/
├── index.ts # Entry point (registers config + tools)
├── config.ts # Configuration schema for LM Studio UI
└── toolsProvider.ts # Core implementation (API calls + key rotation)
npm run build
lms dev
lms push
{
prompt: string // Required: The prompt for Gemini
systemPrompt?: string // Optional: Custom system prompt
useThinking?: boolean // Optional: Enable thinking (default: true)
temperature?: number // Optional: Override temperature
maxTokens?: number // Optional: Override max tokens
}
## Reasoning Agent Response
[Gemini's response]
---
**Model:** gemini-3-flash-preview | **Temperature:** 0.7 | **Max Tokens:** 2048