Rob Ross palette OS (phase 1 MVP)

Local-first palette generation engine that learns a user's aesthetic from heterogeneous taste sources and compiles it into a versioned design genome.

Repository: github.com/ledoit/Rob-Ross

If your checkout folder is still named robross-palette-engine, you can rename it to rob-ross-palette-engine locally (close editors first if Windows reports “permission denied”).

Core principles

• 100% local execution
• No paid APIs
• No cloud dependencies at runtime
• Deterministic color math done in code (not by LLM)

Stack

• Python 3.11+
• Ollama (mistral or llama3)
• LlamaIndex
• ChromaDB
• sentence-transformers
• colormath
• Rich
• Typer
• Palette Studio (optional): FastAPI + Alpine.js — local UI on 127.0.0.1:8765, animated CSS mesh background, no separate frontend build step

Quick start

1. Create environment:
   • python -m venv .venv
   • Windows Bash: source .venv/Scripts/activate
2. Install dependencies:
   • pip install -r requirements.txt
3. Install Ollama and pull a local model:
   • Install from https://ollama.com/download
   • Run:
     • ollama pull mistral
     • or ollama pull llama3
4. Optional environment variables (prefer ROB_ROSS_*; legacy ROBROSS_* still works):
   • setx ROB_ROSS_OLLAMA_MODEL mistral (PowerShell: $env:ROB_ROSS_OLLAMA_MODEL="mistral" for session only)
   • setx ROB_ROSS_EMBED_MODEL sentence-transformers/all-MiniLM-L6-v2
5. Run CLI:
   • python cli.py --help
6. Palette Studio (infinite select → regen loop in the browser):
   • python -m studio
   • Open http://127.0.0.1:8765 (override host/port with ROB_ROSS_STUDIO_HOST / ROB_ROSS_STUDIO_PORT; set ROB_ROSS_STUDIO_RELOAD=true for dev autoreload)
   • Check variants to set the shortlist for the next run, edit the brief, Regenerate. Heart adds a palette to the export roster (same as roster add).

Commands

• python cli.py ingest <source>
• python cli.py build-genome
• python cli.py generate --task "make 11 ide palettes and 5 web palettes then build a superset of 25"
• python cli.py quick "black and yellow" — prompt-driven IDE batch (--variety, --adherence, --fresh to wipe ide_palette_*.json first)
• Taste moods are weighted-sampled per batch when genome/user_loop_state.json exists (auto-created on first quick). Hearts / roster add / roster shortlist add bump weights. Optional reproducibility: ROB_ROSS_SEED=12345.
• python cli.py preview — HTML mock-editor gallery
• python cli.py roster add ide_palette_02 — final pick (VS Code export list); alias roster export-add
• python cli.py roster shortlist add ide_palette_02 — shortlist (best-of-batch; biases the next quick only)
• python cli.py export-themes
• python cli.py feedback
• python cli.py superset --input-palettes outputs/palettes --count 25

Layout

• genome/ genome JSON, theme_roster.json (export + shortlist), user_loop_state.json (taste mood weights + event tail), history snapshots
• sources/ raw and processed extracted principles
• vector_store/ Chroma persistence (rob_ross_principles collection)
• outputs/palettes/ generated palettes JSON
• outputs/reports/ rationale markdown reports
• core/ implementation modules (quick_session.py is shared by CLI quick and Studio)
• studio/ FastAPI app + templates + static CSS
• tests/ unit tests
• scripts/export_vscode_themes.py — installable Cursor/VS Code themes (rob-ross-ide-palettes)
• vscode-themes/ local theme extension package