Rune Documentation¶

Rune is a local-first coding agent that uses LoRA weight space as episodic memory. It implements a 4-phase template-driven pipeline (decompose → plan → code → integrate), parallel swarm orchestration, a Doc-to-LoRA hypernetwork, TIES/DARE adapter merging, and a flat adapter registry with lineage tracking.

Core Subsystems¶

• Pipeline — 4-phase coding pipeline with Jinja2 templates, per-phase iteration, and early stopping. Entry: scripts/rune_runner.py
• Adapter Registry — SQLite + filesystem store for LoRA adapters with write-once enforcement, fitness queries, and lineage tracking. Entry: libs/adapter-registry/
• Hypernetwork — Perceiver-based Doc-to-LoRA hypernetwork generating rank-8 LoRA adapters in a single forward pass. Entry: libs/model-training/

Documentation¶

• Architecture: 4-Phase Pipeline — Pipeline phases, swarm execution, template system
• Architecture: Adapter Storage — Registry schema, write-once policy, querying
• Architecture: Monorepo Mapping — Component layout, integration points
• Architecture: GPU Strategy — Multi-GPU coordination, pipeline parallelism
• Swarm Architecture — Fat orchestrator, training pool, evolution worker
• Implementation Plan — Original phased build plan with current status annotations
• Components Overview — All services and libraries
• Risk Matrix — Primary research risks with mitigations
• Build Order — Component dependency chain