CHAPPiE Backend

Entrypoints

• App API: python3 app.py, port 8010
• Steering API: python3 -m brain.steering_api_server, port 8000
• Training: python3 -m Chappies_Trainingspartner.training_daemon
• Runtime factory: web_infrastructure.chappie_runtime.create_chappie_backend
• Compatibility factory: web_infrastructure.backend_wrapper.create_chappie_backend

Runtime modules

File	Responsibility
chappie_runtime.py	public facade, construction and lifecycle
turn_pipeline.py	shared sync/stream turn order
turn_context.py	input normalization and context gates
contracts.py	typed internal contracts
generation.py	current brain lookup and generation
formatting.py	parsing, sanitizing and API message formatting
persistence.py	chat, STM and finalization writes
backend_wrapper.py	lightweight legacy imports only

Routers use public runtime methods. Do not add new private-wrapper coupling.

Providers

• vLLM is the fixed production web-chat route.
• Ollama and Groq remain supported for factory, CLI, training and research.
• There is no active Cerebras provider.
• Preserve provider fallback order unless a separate behavior change is approved.
• GenerationGateway must resolve the current brain lazily after settings reload.

API and streaming

Preserve request schemas, response fields, status codes, SSE event names, ordering and completion behavior. Sync and stream use the same TurnContext and finalization intent.

Tests

python3 tests/test_runtime_architecture.py
python3 tests/test_api_contract.py
python3 tests/test_vllm_response_handling.py
python3 tests/test_ollama_response_handling.py
python3 tests/test_training_daemon_lifecycle.py

Use requirements/ci.txt for deterministic backend tests. Do not require GPU weights in offline CI.