rorchestra 0.1.0

Roblox/Luau AI orchestration system for codebase understanding and scoped edit generation

Roblox/Luau AI Orchestration System

A Python orchestration layer that maintains hierarchical memory of a Rojo/Luau codebase, assembles compact context packets, and dispatches Gemini CLI workers for scoped code edits.

Quick Start

cd orchestrator
pip install -e ".[dev]"

# Ingest a Rojo project
python -m app.main ingest C:\path\to\your\rojo\project

# Generate memory summaries
python -m app.main summarize --repo-id 1

# Request an edit
python -m app.main edit "Add error handling to the data save module" --repo-id 1 --scope DataManager --side server

# Validate a patch
python -m app.main validate --task-id 1

# Check live Studio state (uncertainty-triggered)
python -m app.main check ui_existence StarterGui.ScreenGui.MainHUD

# View system status
python -m app.main status

# Enter the interactive REPL (Rorchestra)
python -m app.rochester
# Inside the REPL, you can use:
# /edit "description"  -> Propose an edit
# /apply <id>          -> Apply a patch and rebuild memory cascade
# /mcp                 -> Check connected MCP servers and statuses

Architecture

Layer	Purpose
Adapters	Rojo, luau-lsp, Gemini CLI, MCP dispatcher
Services	Ingest, graph, memory, summarization, packets, workers, validation, MCP
Policies	Safety gates, MCP trigger policy
Telemetry	JSONL event log
Storage	SQLite (ORM), file-based artifact store

MCP Integration

Uses a canonical capability dispatcher that routes through:

• Primary: Official Roblox Studio MCP
• Fallback: Community robloxstudio-mcp (filtered toolset)

Raw MCP output is stored out-of-band in artifacts/mcp_raw/ — never injected into planner context.

Memory Model

Memory records are invalidation-driven, not time-based. A record is only stale when its source files change:

• Accepted patches trigger invalidate_by_file()
• Stale scopes are re-summarised on demand via summarize

Requirements

• Python 3.11+
• Rojo (on PATH)
• luau-lsp (optional, for static validation)
• Gemini CLI (for worker invocations and summarisation)