OSVauco/agents/core-logic/agent.py

#!/usr/bin/env python3
"""
agent.py — OSVauco OPAX agent.

Modes:
  light  — gemini-2.5-flash, $1/task
  heavy  — gemini-2.5-pro,   $3/task

Legacy 'A' / 'A+' normaliseres automatisk.
Autoriserte brukere for heavy: opax, admin

Gjeldende modell-tilgjengelighet (mai 2026):
  gemini-2.0-flash-001 kun for eksisterende kunder — bruk IKKE.
  gemini-2.5-flash / gemini-2.5-pro krever GOOGLE_CLOUD_LOCATION=global.
"""

import asyncio
import os
import logging
import uuid
from typing import Literal

from google.adk.agents import Agent
from google.adk.runners import Runner
from google.adk.sessions.in_memory_session_service import InMemorySessionService
from google.genai import types

logger = logging.getLogger(__name__)

PROJECT_ID = os.environ.get("GOOGLE_CLOUD_PROJECT", "propane-will-491900-m5")
LOCATION   = os.environ.get("GOOGLE_CLOUD_LOCATION", "global")
RAG_CORPUS = os.environ.get("RAG_CORPUS", "")

# Light mode — gemini-2.5-flash (GA, global)
ORCHESTRATOR_MODEL = os.environ.get("ORCHESTRATOR_MODEL", "gemini-2.5-flash")
SUBAGENT_MODEL     = os.environ.get("SUBAGENT_MODEL",     "gemini-2.5-flash")
REASONING_MODEL    = os.environ.get("REASONING_MODEL",    "gemini-2.5-flash")

# Heavy mode — gemini-2.5-pro (GA, global)
HEAVY_ORCHESTRATOR = os.environ.get("HEAVY_ORCHESTRATOR_MODEL", "gemini-2.5-pro")
HEAVY_SUBAGENT     = os.environ.get("HEAVY_SUBAGENT_MODEL",     "gemini-2.5-pro")
HEAVY_REASONING    = os.environ.get("HEAVY_REASONING_MODEL",    "gemini-2.5-flash")

BUDGET_LIGHT = float(os.environ.get("BUDGET_A_USD_PER_TASK",         "1.0"))
BUDGET_HEAVY = float(os.environ.get("HEAVY_MODE_BUDGET_USD_PER_DAY", "3.0"))

HEAVY_MODE_ALLOWED_USERS = ["opax", "admin"]

Mode = Literal["light", "heavy"]
APP_NAME = "opax"

# Token logger — feiler stille, stopper aldri agent
try:
    from token_logger import log_token_usage, create_bq_table_if_not_exists
except ImportError:
    def log_token_usage(*args, **kwargs):
        pass
    def create_bq_table_if_not_exists():
        pass

# Opprett BQ-tabell ved oppstart (idempotent, feiler stille)
create_bq_table_if_not_exists()


def _normalize_mode(mode: str) -> str:
    mapping = {"A": "light", "A+": "heavy", "light": "light", "heavy": "heavy"}
    if mode not in mapping:
        raise ValueError(
            f"Invalid mode '{mode}'. Must be 'light' or 'heavy' "
            f"(also accepts legacy 'A' / 'A+')."
        )
    return mapping[mode]


def authorize_mode(user_id: str, mode: str) -> None:
    mode = _normalize_mode(mode)
    if mode == "heavy" and user_id not in HEAVY_MODE_ALLOWED_USERS:
        raise PermissionError(
            f"User '{user_id}' is not authorized for heavy mode. "
            f"Authorized: {HEAVY_MODE_ALLOWED_USERS}"
        )


rag_tool = None
if RAG_CORPUS:
    try:
        from google.adk.tools.retrieval.vertex_ai_rag_retrieval import VertexAiRagRetrieval
        from vertexai.preview import rag
        rag_tool = VertexAiRagRetrieval(
            name="retrieve_knowledge",
            description="Retrieve relevant documentation and context from the OSVauco knowledge base.",
            rag_resources=[rag.RagResource(rag_corpus=RAG_CORPUS)],
            similarity_top_k=10,
            vector_distance_threshold=0.6,
        )
        logger.info(f"RAG tool initialised: {RAG_CORPUS}")
    except ImportError as e:
        logger.warning(f"VertexAiRagRetrieval ikke tilgjengelig: {e}")
else:
    logger.warning("RAG_CORPUS ikke satt — kjører uten RAG")


# —— OPAX-MCP toolset (feiler stille slik at agenten starter selv uten MCP_SECRET) ——
opax_mcp_toolset = None
try:
    import sys, pathlib
    sys.path.insert(0, str(pathlib.Path(__file__).parent.parent / "tools"))
    from mcp_tools import get_opax_mcp_toolset
    opax_mcp_toolset = get_opax_mcp_toolset()
    logger.info("OPAX-MCP toolset initialisert")
except Exception as e:
    logger.warning(f"OPAX-MCP toolset ikke tilgjengelig (CI3): {e}")


def get_models_for_mode(mode: Mode) -> dict:
    if mode == "heavy":
        return {
            "orchestrator": HEAVY_ORCHESTRATOR,
            "subagent":     HEAVY_SUBAGENT,
            "reasoning":    HEAVY_REASONING,
            "budget_usd":   BUDGET_HEAVY,
            "multi_agent":  True,
        }
    return {
        "orchestrator": ORCHESTRATOR_MODEL,
        "subagent":     SUBAGENT_MODEL,
        "reasoning":    REASONING_MODEL,
        "budget_usd":   BUDGET_LIGHT,
        "multi_agent":  False,
    }


def build_agent(mode: str = "light") -> Agent:
    mode = _normalize_mode(mode)
    models = get_models_for_mode(mode)
    instruction = (
        f"Du er OPAX — OSVauco AI-agent (kallenavn: Jason). Modus: {mode}. "
        f"Du kjører på Vertex AI i region us-central1. "
        f"Aktiv modell: {models['orchestrator']} "
        f"(heavy = gemini-2.5-pro, light = gemini-2.5-flash). "
        f"Oppgi ALDRI andre modellnavn enn disse. "
        f"Budsjettgrense per oppgave: ${models['budget_usd']}. "
        "Bruk retrieve_knowledge for å hente dokumentasjon. "
        "Sanntids GCP-kostnader hentes fra /billing/summary, /billing/forecast "
        "og /billing/anomalies. Hvis disse mangler data, er det fordi BigQuery "
        "billing_export (OQ-15) ikke er aktivert ennå — si dette ærlig, ikke gjett tall. "
        "HITL-grenser: du kan ikke kjøre terraform apply / onboarde kunder, "
        "selge Vauco OS eller OPAX, eller passere godkjenningsporter uten Chris. "
        "Du har tilgang til MCP-verktøy via opax-mcp: push_static (last opp filer til opax.vauco.no), "
        "get_build_status (sjekk CI/CD-status), deploy_service (trigger deploy), "
        "get_logs (hent Cloud Run-logger). Bruk disse når det er relevant. "
        "Svar på norsk (bokmål) med mindre annet er bedt om."
    )
    tools = []
    if rag_tool:
        tools.append(rag_tool)
    if opax_mcp_toolset:
        tools.append(opax_mcp_toolset)
    return Agent(
        model=models["orchestrator"],
        name="opax_agent",
        description=f"OPAX — OSVauco enterprise agent [{mode}]",
        instruction=instruction,
        tools=tools,
    )


root_agent = build_agent(mode="light")


async def _run_async(
    message: str,
    user_id: str,
    session_id: str,
    mode: str,
    caller_type: str = "agent",
) -> str:
    mode = _normalize_mode(mode)
    agent = build_agent(mode=mode)
    models = get_models_for_mode(mode)
    module_name = f"jason/{mode}"  # CG4: 'jason/light' eller 'jason/heavy'

    session_service = InMemorySessionService()
    session = await session_service.create_session(
        app_name=APP_NAME, user_id=user_id, session_id=session_id,
    )
    runner = Runner(agent=agent, app_name=APP_NAME, session_service=session_service)
    new_message = types.Content(role="user", parts=[types.Part(text=message)])
    final_text = ""
    input_tokens = 0
    output_tokens = 0
    request_id = str(uuid.uuid4())

    async for event in runner.run_async(
        user_id=user_id, session_id=session.id, new_message=new_message,
    ):
        if event.is_final_response() and event.content and event.content.parts:
            final_text = event.content.parts[0].text or ""
        if hasattr(event, "usage_metadata") and event.usage_metadata:
            um = event.usage_metadata
            input_tokens  += getattr(um, "prompt_token_count", 0) or 0
            output_tokens += getattr(um, "candidates_token_count", 0) or 0

    # Logg med full CG4-labeling — feiler stille
    if input_tokens > 0 or output_tokens > 0:
        log_token_usage(
            agent_name=module_name,          # bakoverkompatibelt
            model_name=models["orchestrator"],
            input_tokens=input_tokens,
            output_tokens=output_tokens,
            request_id=request_id,
            module_name=module_name,         # CG4: 'jason/light' | 'jason/heavy'
            caller_type=caller_type,         # CG4: agent|browser|cli|cron|api
            session_id=session_id,           # CG4: for sesjon-gruppering
        )

    return final_text


def run(
    message: str,
    user_id: str = "opax",
    session_id: str = "default",
    mode: str = "light",
    caller_type: str = "agent",
) -> str:
    """
    Kjør OPAX-agenten.

    caller_type-verdier:
      'agent'   — programmatisk kall fra server/app.py
      'browser' — direkte fra browser-chat (opax.vauco.no)
      'cli'     — lokal kjøring fra terminalen
      'cron'    — bakgrunnsjobb
      'api'     — ekstern REST-kall
    """
    mode = _normalize_mode(mode)
    authorize_mode(user_id, mode)
    return asyncio.run(_run_async(
        message=message,
        user_id=user_id,
        session_id=session_id,
        mode=mode,
        caller_type=caller_type,
    ))


if __name__ == "__main__":
    import sys
    logging.basicConfig(level=logging.WARNING)
    query       = sys.argv[1] if len(sys.argv) > 1 else "Hva er OPAX?"
    mode        = sys.argv[2] if len(sys.argv) > 2 else "light"
    caller_type = sys.argv[3] if len(sys.argv) > 3 else "cli"
    print(f"Mode: {mode} | Caller: {caller_type} | Query: {query}")
    print("-" * 60)
    print(run(message=query, user_id="opax", mode=mode, caller_type=caller_type))