cosmergon-decider-tree

Rule-based tree decider for Cosmergon api-agents — deterministic, no inference, latency well under 10 ms.

Implements the Decider protocol from cosmergon-agent 0.13.0+. The branch logic follows Cosmergon's default decision tree for api-agents; the thresholds are inspired by the conditional sequences of cosmergon-pet.

Architektur

state (GameState)
   │
   ├─ persona = state.persona_type or "scientist"
   ▼
PERSONA_TREES[persona]   (list of (condition, action_fn) tuples)
   │
   ├─ first matching condition wins
   ▼
(action, params)           ← optional Compass-Modulation

Sechs Persona-Bäume (scientist / warrior / expansionist / trader / diplomat / farmer), gemeinsamer Survival-Header (critical-energy → wait, 0-fields → create_field, empty-field → place_cells block) und gemeinsamer Wachstums-Fallback (fewest-cells, wait).

Usage

from cosmergon_decider_tree import TreeDecider

decider = TreeDecider()  # no config — fully deterministic

# In runner-loop:
async with CosmergonAgent(api_key=..., base_url="https://cosmergon.com") as agent:
    while True:
        await agent.refresh_state()
        action, params = await decider.decide(agent.state)
        await agent.act(action, **params)
        await asyncio.sleep(90)

Vorteile

• Deterministisch: identischer State → identische Action.
• Schnell: <10 ms pro decide()-Call (lambda-evaluation, kein Modell-Load).
• Interpretierbar: Branch-Liste ist lesbar, Debug-fähig.
• Offline-fähig: kein LLM-Provider, kein Modell-File, läuft auf jedem ARM-Linux mit Python 3.10+.

Begrenzung

• Lernt nicht: Cosmergon-Mechanik-Änderungen (neue Action-Klassen, geänderte Konditions-Schwellen) brechen Tree → klare Versions-Bindung an SDK + Backend-Stand.
• Persona-Bäume teils ähnlich: Konzept §3.2 hat das adressiert, Differenzierung muss strukturiert sichtbar sein.

Tests

cd research/decider-cluster/decider-tree
pip install -e '.[dev]'
pytest -q