nv5466/deltahyperion

Purpose and Entry Points

DeltaHyperion DH2 is a Windows-oriented, local research runtime for a persistent artificial subject (“Delta”), emphasizing continuity, bounded agency, local privacy, and controlled external actuation. Primary lifecycle entry point is delta.ps1; desktop interaction enters through delta_desktop_launcher.py, which uses the direct launcher rather than OpenClaw’s generic UI. Legacy stack scripts remain separate compatibility paths.

Paths: delta.ps1, delta_desktop_launcher.py, delta_launcher.py

Runtime Orchestration and Agency

The runtime is composed of supervised loops and decision subsystems rather than a single web service. Startup resolves project/state roots, reads local bridge/launcher configuration, and starts missing services. Agency combines goals, endogenous state, decision logic, autonomy supervision, and preservation constraints before an action is authorized.

Paths: delta_autonomy_supervisor.py, delta_agency_engine.py, delta_decision.py

State, Continuity, and Persistence Boundary

Mutable runtime state belongs outside the checkout under DELTA_STATE_ROOT; new code should resolve this through configuration and persist through the storage layer. The repository retains many tracked JSON/JSONL files as legacy/migration artifacts. DeltaDB/MuDB provides canonical continuity indexing, migrations, repository access, and workers; snapshots, backup, lifecycle, and retention support recovery and bounded history.

Paths: delta_config.py, delta_storage.py, deltadb/database.py

Context Assembly and Retrieval

Conversation and decision inputs are assembled from explicitly scoped sources: files, memories, conversations, command outputs, live state, and optionally DCode evidence. The retrieval package routes requests, gathers evidence, fuses/reranks it, emits prompt packets, and records receipts. Prompt-boundary and redaction layers constrain what reaches model calls.

Paths: delta_retrieval/router.py, delta_retrieval/fusion.py, delta_prompt_boundary.py

Models and Execution Interfaces

Local model execution is centered on Ollama, with profiles, scheduling, queues, workers, and topology support. Modelfiles and the training subtree support local fine-tuning/export workflows, but training is a separate pipeline from the operational runtime. The direct launcher bridges user messages to runtime/model execution; the desktop client exposes Chat, Runtime, Dashboard, Logs, and Settings.

Paths: delta_ollama_client.py, delta_model_scheduler.py, delta_launcher_bridge.py

Safety, Governance, and External Boundaries

The core boundary is that persistent identity, memory, private context, and continuity data remain local. Containment, runtime security, egress validation, secret redaction, contact/message safety, and self-modification gating regulate effects and disclosures. Policies and causal receipts make decisions and external results auditable.

Paths: delta_containment.py, delta_egress.py, delta_self_modification_gate.py

Integrations and Host Observation

DeltaScope is a typed, host-scoped inventory/telemetry subsystem supporting Windows, Linux, macOS, and Raspberry Pi through platform adapters and a CLI. OpenClaw is an optional bridge/plugin boundary, not the primary chat route. DCode is a separately configured executable integration: a single bridge invokes it without a shell, limits mappings/payloads, validates identity, redacts returned evidence, and preserves the existing path as authoritative until explicit cutover.

Paths: delta_scope/manager.py, openclaw-plugin-delta-core/index.ts, delta_dcode_bridge.py