CHAPPiE Architecture

Active request path

API / CLI / Research
  -> web_infrastructure.chappie_runtime.CHAPPiERuntime
  -> TurnPipeline and TurnContext
  -> Brain, Memory, Life, Global Workspace
  -> GenerationGateway -> vLLM and Steering
  -> formatting and persistence

web_infrastructure/backend_wrapper.py is only a compatibility layer. Keep create_chappie_backend, init_chappie, CHAPPiERuntime and CHAPPiEBackend compatible.

Sync and streaming must build the same TurnContext, share preparation and finalization, and differ only at the output adapter.

Boundaries

• Runtime orchestrates; it does not duplicate Memory or Life semantics.
• brain/global_workspace.py selects and broadcasts active signals.
• brain/action_response.py creates prompt and action context.
• brain/steering_manager.py owns VAD, alpha, composite modes and layer profiles.
• brain/steering_backend.py performs activation injection.
• memory/ owns STM, LTM, retrieval, context files, sleep and forgetting.
• life/ owns homeostasis, goals, habits, attachment, time and development.

There are ten emotions in config/emotions.py. The Qwen 3.5 4B profile steers layers 10 through 26.

Historical v1

The multi-agent BrainPipeline is not production. Exact sources are under Legacy-Code/brain-pipeline-v1/; brain/brain_pipeline.py is a lazy compatibility loader. Historical agents must not be imported by the active runtime. Do not move or archive active steering with those agents.

Compatibility rules

• Preserve external API, SSE, CLI and persisted-data contracts.
• Do not reorder Intent, Memory, Life, Workspace, generation or finalization without a regression test.
• Treat provider priority, emotion mapping, sleep behavior and storage changes as separate migrations.
• Keep service and payload versions independent from the product version.

Required checks

python3 tests/test_runtime_architecture.py
python3 tests/test_local_first_runtime.py
python3 tests/test_life_simulation.py
python3 tests/test_reasoning_layering.py