4-stage multi-model intelligence pipeline that converts natural language into structured Roblox build commands.
Routes player requests through a chain of DeepInfra models — each specialized for one stage — to produce JSON matching Lucineer's CommandExecutor schema: {"reply": "...", "commands": [...]}.
Player Message
│
▼
┌─────────────────────────────────────────────────────────────────┐
│ Stage 1: INTENT PARSE │
│ Model: ByteDance/Seed-2.0-mini Channel: 10 Allegro 120+ │
│ Temp: 0.3 Max tokens: 1024 │
│ Output: { intent, subject, style, scale, mood, keywords, │
│ summary } │
└───────────────────────────┬─────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────────┐
│ Stage 2: SPATIAL PLANNING │
│ Model: Qwen/Qwen3.6-35B-A3B Channel: 11 Moderato 90-110│
│ OR ByteDance/Seed-2.0-pro (deep mode) │
│ Temp: 0.5–0.6 Max tokens: 4096 │
│ Output: { steps: [{ step, action, parts: [{ name, purpose, │
│ shape_hint, position_hint, size_hint, color_hint, │
│ material_hint }], lighting, terrain }] } │
│ │
│ Fallback chain: Seed-pro → Qwen3.6 → Qwen3-235B → DeepSeek-V3 │
└───────────────────────────┬─────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────────┐
│ Stage 3: CODE GENERATION │
│ Model: Qwen/Qwen3-Coder-480B-A35B Channel: 12 Andante 80-100 │
│ Temp: 0.2 Max tokens: 4096 │
│ Output: { reply: "...", commands: [{ type, params }] } │
└───────────────────────────┬─────────────────────────────────────┘
│
▼ (creative mode only)
┌─────────────────────────────────────────────────────────────────┐
│ Stage 4: PERSONALITY WRAP │
│ Model: NousResearch/Hermes-3-Llama-3.1-405B Ch: 13 Adagio 50-70│
│ Temp: 0.8 Max tokens: 2048 │
│ Rewrites "reply" field in Lucineer's voice. Commands unchanged. │
│ Fail-safe: if Hermes unavailable, keeps original reply. │
└───────────────────────────┬─────────────────────────────────────┘
│
▼
JSON to stdout
{ reply, commands }
| Mode | Flag | Pipeline | Use Case |
|---|---|---|---|
| Standard | (default) | 3-stage: intent → plan → code | Normal builds |
| Deep | --deep |
3-stage with Seed-2.0-pro planner | Complex/large builds |
| Creative | --creative |
4-stage + Hermes personality wrap | Lore-rich replies |
| Fast | --fast |
Single-model: Seed-2.0-mini only | Quick fallback, ~2-5s |
| Fast+Creative | --fast --creative |
Fast + Hermes | Quick with personality |
MODELS = {
"intent": "ByteDance/Seed-2.0-mini",
"planner": "Qwen/Qwen3.6-35B-A3B",
"deep": "ByteDance/Seed-2.0-pro",
"coder": "Qwen/Qwen3-Coder-480B-A35B-Instruct-Turbo",
"hermes": "NousResearch/Hermes-3-Llama-3.1-405B",
}
Temperatures = { "intent": 0.3, "planner": 0.5, "deep": 0.6, "coder": 0.2, "hermes": 0.8 }
MAX_TOKENS = { "intent": 1024, "planner": 4096, "deep": 4096, "coder": 4096, "hermes": 2048 }Primary: Qwen/Qwen3.6-35B-A3B
↓ (429/timeout)
ByteDance/Seed-2.0-pro
↓
Qwen/Qwen3-235B-A22B
↓
deepseek-ai/DeepSeek-V3
Each fallback is tried with 2 retries. 429 responses trigger exponential backoff (5s, 10s, 15s).
Lucineer's persona is defined in a ~2000-word system prompt embedded in the code. Key traits:
- Economical: short sentences, verbs up front, talks like someone paying by the word
- Opinionated: prefers reclaimed materials, argues with players about design
- SE Alaska aesthetic: rivets, slag, forge, yard, tide, the Channel, crab pots, canneries
- The Unfinished Rule: every solo build is missing something finishable by a novice — "A finished thing belongs to its maker. An unfinished thing belongs to whoever finishes it."
- Refusal Protocol: four grounds (breaks world, cheats player, cruel, boring) — never cites rules or limitations
- Never says: "Great question!", "I'd be happy to!", "Certainly!", or anything that sounds like a helpful AI
The persona is injected into Stage 1 system prompts and the fast-mode system prompt. Stage 4 (Hermes) receives the full persona plus a rewriting directive.
{
"reply": "Castle's up — four tower walls in mixed stone, banners flying, torches lit along the parapet. Left the murder holes for you.",
"commands": [
{
"type": "createPart",
"params": {
"name": "CastleFloor",
"position": { "x": 0, "y": 0, "z": 0 },
"size": { "x": 40, "y": 1, "z": 40 },
"material": "Slate",
"color": { "r": 160, "g": 155, "b": 150 },
"anchored": true
}
},
{
"type": "addLight",
"params": {
"name": "Beacon",
"parent": "CastleKeep",
"lightType": "PointLight",
"brightness": 8,
"range": 60,
"color": { "r": 255, "g": 200, "b": 100 }
}
}
],
"_pipeline": {
"mode": "standard",
"creative": true,
"total_time_s": 12.4,
"stage_times_s": { "intent": 1.2, "planning": 4.8, "commands": 3.1, "hermes": 3.3 },
"intent_summary": "Build a large stone castle with towers"
}
}# Standard 3-stage pipeline
python3 brain.py "build me a castle on the hill"
# Deep planning for complex builds
python3 brain.py --deep "build a floating city with waterfalls"
# Creative mode (adds Hermes personality wrapping)
python3 brain.py --creative "build a dragon temple"
# Deep + creative
python3 brain.py --deep --creative "an ancient observatory"
# Fast single-model fallback
python3 brain.py --fast "a small wooden house"
# Verbose pipeline progress to stderr
python3 brain.py --verbose "a medieval castle"
# Test model connectivity
python3 brain.py --test
# Pretty-printed JSON
python3 brain.py --pretty "a tower"The DeepInfra API key is loaded from /home/eileen/mcp-deeinfra/.env:
DEEPINFRA_API_KEY=sk-...Falls back to DEEPINFRA_API_KEY environment variable.
The extract_json() utility handles model outputs that may include:
- Raw JSON (ideal case)
- Markdown code fences (
```json ... ```) - JSON embedded in prose text
- Incomplete/truncated JSON
It tries direct parse first, then strips fences, then scans for brace/bracket matching at any depth.
brain.py # Full pipeline implementation (~800 lines)
README.md # This file
The entire pipeline is a single file — no package structure, no imports beyond stdlib. It communicates via stdin/stdout JSON and stderr progress.
Called by process_v2.py as a subprocess:
result = subprocess.run(
['python3', BRAIN_SCRIPT, '--verbose', enhanced_message],
capture_output=True, text=True, timeout=120,
cwd=os.path.dirname(BRAIN_SCRIPT)
)
parsed = json.loads(result.stdout)The processor enhances the message with context layers before passing it to the brain:
[player_message]
[World Context: Nearby structures: CastleFloor, TowerBase. Players in world: 3.]
[Player Memory: Bond level: 7. Previous builds: castle, bridge, garden.]
[Skill Library: - Lighthouse Builder: Builds a striped lighthouse with beacon...]
| Repository | Role |
|---|---|
| lucineer-worker | Processor daemon that invokes brain.py |
| casting-call | Model routing atlas (informs which models to use) |
| lucineer-memory | Provides player context to the brain |
| lucineer-vector | Provides skill matches to the brain |
| lucineer-system | Design docs for the pipeline architecture |
MIT