HERMES AGENT HARNESS

# hermes_agent_harness.py - LangGraph + Pydantic + Evaluator Judge Pattern
from typing import List, Dict, Any, Literal, Optional
from pydantic import BaseModel, Field, ValidationError
from langgraph.graph import StateGraph, END

# ==========================================
# 1. PYDANTIC SCHEMAS (STRUCTURAL & JUDGE)
# ==========================================

class UserRequest(BaseModel):
    query: str = Field(..., min_length=2, max_length=1000)

class EvaluationResult(BaseModel):
    """Pydantic schema for the Evaluator Node (Judge LLM)."""
    is_satisfactory: bool = Field(description="True if candidate answer completely fulfills user goal.")
    score: float = Field(ge=0.0, le=1.0)
    critique_feedback: Optional[str] = Field(default=None, description="Actionable feedback if unsatisfactory.")

class HermesState(BaseModel):
    """LangGraph central state model enforced by Pydantic."""
    query: str
    messages: List[Dict[str, Any]] = []
    pending_tool_call: Optional[Dict[str, Any]] = None
    tool_results: List[Dict[str, Any]] = []
    draft_response: Optional[str] = None
    eval_result: Optional[EvaluationResult] = None
    retry_count: int = 0
    max_retries: int = 3

# ==========================================
# 2. LANGGRAPH NODES & EVALUATOR NODE
# ==========================================

def agent_reasoner_node(state: HermesState) -> Dict[str, Any]:
    # Worker LLM plans tool call or generates response
    pass

def evaluator_node(state: HermesState) -> Dict[str, Any]:
    """Dedicated Judge LLM evaluating draft response against original goal."""
    prompt = f"GOAL: {state.query}\nDRAFT: {state.draft_response}"
    
    # Call secondary LLM forcing EvaluationResult output schema
    result: EvaluationResult = judge_llm.with_structured_output(EvaluationResult).invoke(prompt)
    
    if not result.is_satisfactory:
        # Inject critique feedback back into context for re-try
        critique_msg = {"role": "user", "content": f"CRITIQUE: {result.critique_feedback}"}
        return {
            "eval_result": result,
            "retry_count": state.retry_count + 1,
            "messages": state.messages + [critique_msg]
        }
    return {"eval_result": result}

def judge_decision_router(state: HermesState) -> str:
    if state.eval_result and state.eval_result.is_satisfactory:
        return "end"
    if state.retry_count >= state.max_retries:
        return "end"  # Exit to avoid infinite loop
    return "retry_reasoner"

# ==========================================
# 3. BUILD GRAPH
# ==========================================

workflow = StateGraph(HermesState)
workflow.add_node("reasoner", agent_reasoner_node)
workflow.add_node("evaluator", evaluator_node)

workflow.add_conditional_edges(
    "evaluator",
    judge_decision_router,
    {"end": END, "retry_reasoner": "reasoner"}
)