wataboocolosus/triadees

Purpose and Entry Points

Tríade Ω is a governed local AI-agent prototype: it runs auditable cognitive cycles, retains scoped memory, performs gated learning, and can operate background workers and experimental federated nodes. Primary CLI and server entry point: triade_digimon.py. The API serves the official React SPA through apps/single_port_app.py; route composition is under apps/routes/. The browser client starts at frontend/src/main.tsx.

Core Cognitive Run

The orchestrated request path is implemented around the central runner and typed contracts: triade/core/runner.py, triade/core/contracts.py, and triade/core/central.py. A run moves from input and signals through memory/context, Crystal regulation, planning, safety, response, persisted artifacts, and verification. triade/core/expression_cortex.py is the presentation boundary: it converts governed internal results into human-facing output rather than exposing raw internal state.

Triadic State and Memory Boundary

Central cognition is modulated by the hypothalamus and Bodega memory/context layers: triade/core/hypothalamus.py, triade/core/bodega_global_context.py, and triade/core/crystal.py. SQLite is the operational persistence layer, with schema evolution in triade/memory/migrations/ and semantic retrieval/storage in triade/memory/semantic_store.py and triade/memory/semantic_search.py. Memory confidence, scope, provenance, identity continuity, and semantic-governance rules form a hard boundary: candidate/experimental content is not stable truth by default.

Governance, Safety, and Evidence

Actions and outputs pass policy, permission, coherence, safety, and verification controls: triade/core/safety.py, triade/core/permission_governor.py, and triade/core/verification.py. Runtime execution adds durable task status, leases, receipts, and recovery through triade/runtime/governed_task_executor.py, triade/runtime/task_leases.py, and triade/runtime/runtime_recovery.py. Sensitive operations are intended to be blocked, sandboxed, or human-approved rather than executed directly.

Learning and Neuron Lifecycle

Learning is a gated pipeline rather than direct memory mutation: candidate discovery/evidence progresses through validation and production injection in triade/learning/pipeline.py, triade/learning/evidence_bridge.py, and triade/learning/validation.py. Specialized “neurons” are designed, evaluated, trained, certified, and lifecycle-managed by triade/neuron_factory/lifecycle.py, triade/neuron_factory/evaluation.py, and triade/core/neuron_registry.py. Regression and rollback protection are separate release gates in triade/regression/gate.py and triade/regression/rollback.py.

Always-On Runtime and Model Layer

Background operation uses supervised workers, queues, scheduling, concurrency, and watchdogs: triade/workers/worker_loop.py, triade/workers/worker_supervisor.py, and triade/runtime/watchdog.py. Resource policies can degrade execution mode rather than stop safe observation; resource control resides in triade/core/resource_governor.py. Ollama is the recommended local inference and embedding service, routed by triade/models/model_router.py and accessed via triade/models/ollama_client.py; safe fallback exists but has reduced cognitive/learning claims.

External Surfaces and Deployment

The public/API boundary includes route modules such as apps/routes/api.py, apps/routes/knowledge.py, and apps/routes/health.py; dashboard endpoints are read-only by design. Federation is an authorization- and evidence-gated private-node subsystem, not established distributed production: triade/federation/dispatch.py, triade/federation/evidence_gate.py, and triade/federation/registry.py. An Android edge node can preprocess locally and relay work via android/triade-node/app/src/main/java/local/triade/node/TriadeNodeService.java and android/triade-node/app/src/main/java/local/triade/node/RelayClient.java. Deployment targets include Docker/Compose, systemd units, Caddy, and cloud manifests in compose.cloud.yml, deploy/systemd/, and render.yaml.