wizard-checkpoints/harness-interface-spec.md

Harness Interface Specification

Standard Interface (All Harnesses Must Implement)

class Harness(ABC):
    '''Abstract base for all agent harnesses'''
    
    @abstractmethod
    def execute(self, prompt: str, context: Context) -> Response:
        '''Execute a prompt and return response'''
        pass
    
    @abstractmethod
    def tool_call(self, tool: str, params: dict) -> Result:
        '''Execute a specific tool'''
        pass
    
    @abstractmethod
    def session_save(self, session: Session) -> Checkpoint:
        '''Save session state'''
        pass
    
    @abstractmethod
    def session_load(self, checkpoint: Checkpoint) -> Session:
        '''Load session state'''
        pass
    
    @property
    @abstractmethod
    def capabilities(self) -> List[str]:
        '''Return list of supported capabilities'''
        pass
    
    @property
    @abstractmethod
    def metrics(self) -> Metrics:
        '''Return performance metrics'''
        pass

class ClawHarness(Harness):
    '''Claw Code implementation'''
    
    def __init__(self, binary_path: str, model: str):
        self.binary = binary_path
        self.model = model
        self.bridge = OllamaBridge()  # or direct
    
    def execute(self, prompt, context):
        # Claw-specific: fast, local, tool-heavy
        return self.bridge.run(prompt, model=self.model)

class HermesHarness(Harness):
    '''Hermes Agent implementation'''
    
    def __init__(self, profile: str):
        self.profile = profile
        self.gateway = GatewayClient()
    
    def execute(self, prompt, context):
        # Hermes-specific: complex workflows, skills
        return self.gateway.chat(prompt, profile=self.profile)

class OllamaHarness(Harness):
    '''Ollama local implementation'''
    
    def __init__(self, model: str = "qwen2.5:1.5b"):
        self.model = model
        self.client = OllamaClient()
    
    def execute(self, prompt, context):
        # Ollama-specific: fully local, offline
        return self.client.generate(prompt, model=self.model)

Usage (Wizard Code)

# Config-driven harness selection
config = load_profile("allegro")
harness = HarnessFactory.create(config.harness.default)

# Same code, different runtime
response = harness.execute("Write Python hello world", context)

# Fallback chain
for harness_name in config.harness.fallbacks:
    try:
        harness = HarnessFactory.create(harness_name)
        return harness.execute(prompt, context)
    except HarnessError:
        continue

Configuration Schema

# ~/.hermes/profiles/allegro/config.yaml
harness:
  default: claw
  fallbacks:
    - name: hermes
      condition: "complex_workflow"
    - name: ollama
      condition: "offline"
  
  claw:
    binary: /opt/claw/claw-cli
    model: claude-3-7-sonnet
    bridge: ollama  # or direct
  
  hermes:
    profile: allegro
    gateway: http://localhost:8644
  
  ollama:
    model: qwen2.5:1.5b
    host: http://localhost:11434

tools:
  - file
  - terminal
  - web

selection_criteria:
  speed: claw > ollama > hermes
  capability: hermes > claw > ollama
  offline: ollama only

---

Design by Ezra
Implementation by Allegro