LLM Configuration

LLM Configuration

Multi-Model Routing (v2.2.0+)

Software Factory automatically routes agents to the right model based on their role:

Category	Heavy Model	Light Model	Agent Tags
Reasoning	gpt-5.2	gpt-5-mini	reasoner, architect, strategist, planner
Production / Code	gpt-5.1-codex	gpt-5-mini	developer, tester, security, refactoring
Tasks	gpt-5-mini	gpt-5-mini	Generic agents
Redaction	gpt-5.1-codex	gpt-5-mini	doc_writer, tech_writer

Configure live in Settings → LLM (no restart required).

Darwin LLM Thompson Sampling

Same team (agent + pattern) A/B tests across LLM models automatically:

Beta(wins+1, losses+1) per (agent_id, pattern_id, technology, phase_type, llm_model)

• Warmup: random exploration for first 5 runs per context
• After warmup: Thompson Sampling picks the model with best Beta distribution
• View results in Teams → LLM A/B tab

Providers (multi-provider with fallback)

Provider	Models	Use Case
MiniMax	MiniMax-M2.5	Default (local dev)
Azure OpenAI	gpt-5-mini	Production lightweight
Azure AI Foundry	gpt-5.2, gpt-5.1-codex, gpt-5.1-mini	Production advanced
Demo	Mock	Testing (no API key)

Fallback Chain

minimax → azure-openai → azure-ai
Cooldown: 90s on HTTP 429 (rate limit)

Environment Variables

PLATFORM_LLM_PROVIDER=minimax          # or azure-openai, azure-ai, demo
PLATFORM_LLM_MODEL=MiniMax-M2.5       # or gpt-5-mini
LLM_RATE_LIMIT_RPM=50                 # requests per minute

# Azure AI Foundry (for gpt-5.2 / gpt-5.1-codex)
AZURE_AI_API_KEY=...
AZURE_DEPLOY=your-deployment-endpoint

# Azure OpenAI (for gpt-5-mini)
AZURE_OPENAI_API_KEY=...
AZURE_OPENAI_ENDPOINT=...

API Keys

Keys stored in ~/.config/factory/*.key (chmod 600).
⚠️ NEVER set *_API_KEY=dummy — use PLATFORM_LLM_PROVIDER=demo instead.

Routing Config API

# Get current routing matrix
GET /api/llm/routing

# Update routing matrix (flushes 60s cache instantly)
POST /api/llm/routing
Content-Type: application/json
{
  "routing": {
    "reasoning_heavy": {"provider": "azure-ai", "model": "gpt-5.2"},
    "reasoning_light": {"provider": "azure-openai", "model": "gpt-5-mini"},
    "production_heavy": {"provider": "azure-ai", "model": "gpt-5.1-codex"},
    "production_light": {"provider": "azure-openai", "model": "gpt-5-mini"}
  }
}

LLM A/B APIs

# Model fitness leaderboard
GET /api/teams/llm-leaderboard?technology=generic&phase_type=generic

# Recent A/B test results
GET /api/teams/llm-ab-tests?limit=20

Provider Notes

Provider	Notes
MiniMax	<think> tags consume tokens (min 16K context)
gpt-5-mini	NO temperature param; max_completion_tokens ≥ 8K
gpt-5.1-codex	Optimized for code; use max_completion_tokens
Demo	Returns mock responses, no external calls

Observability

Each LLM call is traced with: provider, model, tokens (in/out), cost, latency.
View: /api/llm/stats, /api/llm/traces, and Monitoring → MCP Tool Calls in the dashboard.

🇫🇷 Configuration LLM (FR) · 🇪🇸 ES