shira-hermes/api/services/agent_runner.py

"""Agent runner — wraps OpenAI client with tool-calling loop via ai-gateway."""

from __future__ import annotations

import json
import logging
from typing import Any

from openai import AsyncOpenAI

from mcp_server.espocrm_client import EspoCrmClient
from mcp_server.tools.crm_tools import register_crm_tools
from mcp_server.tools.document_tools import register_document_tools
from mcp_server.tools.legal_tools import register_legal_tools
from api.services.skills import register_skill_tools
from api.services.memory import register_memory_tools
from api.services.delegate import register_delegate_tool

logger = logging.getLogger("shira.agent")


class AgentRunner:
    """
    Runs an agentic conversation loop using the OpenAI SDK pointed at ai-gateway.

    The loop:
    1. Send messages + tools to the model
    2. If the model returns tool_calls, execute them and feed results back
    3. Repeat until the model returns a text response (no tool_calls)
    4. Return the final text
    """

    def __init__(
        self,
        ai_gateway_url: str,
        ai_gateway_api_key: str,
        model: str = "sonnet",
        max_tokens: int = 4096,
        max_iterations: int = 10,
    ):
        self.client = AsyncOpenAI(
            base_url=f"{ai_gateway_url.rstrip('/')}/v1",
            api_key=ai_gateway_api_key,
        )
        self.model = model
        self.max_tokens = max_tokens
        self.max_iterations = max_iterations

    async def run(
        self,
        system_prompt: str,
        messages: list[dict[str, str]],
        context: dict,
        espocrm_url: str,
        espocrm_api_key: str,
        depth: int = 0,
        blocked_tools: set[str] | None = None,
        allowed_toolsets: list[str] | None = None,
    ) -> str:
        """Run a conversation with tool-calling loop. Returns the final text response.

        Args:
            depth: Current delegation depth. 0 = main conversation, 1 = child agent.
            blocked_tools: Tool names to exclude (used by delegation to prevent recursion).
            allowed_toolsets: If set, only register tools from these categories (crm, documents, legal).
        """

        # Build tools from context
        crm = EspoCrmClient(espocrm_url, espocrm_api_key)
        case_id = (context.get("case") or {}).get("id") or (context.get("drillDown", {}).get("case", {}).get("id"))
        user_id = (context.get("currentUser") or {}).get("id")

        tools_registry: dict[str, dict] = {}

        # Core CRM/document/legal tools
        should_register_all = allowed_toolsets is None
        if should_register_all or "crm" in (allowed_toolsets or []):
            register_crm_tools(tools_registry, crm, case_id, user_id, context)
        if should_register_all or "documents" in (allowed_toolsets or []):
            register_document_tools(tools_registry, crm, case_id, context)
        if should_register_all or "legal" in (allowed_toolsets or []):
            register_legal_tools(tools_registry, crm, case_id, context)

        # Phase 3 tools (only for main conversation, not child agents)
        if depth == 0:
            register_skill_tools(tools_registry)
            register_memory_tools(tools_registry, case_id, user_id)
            register_delegate_tool(tools_registry, context, espocrm_url, espocrm_api_key)

        # Remove blocked tools (for delegation safety)
        if blocked_tools:
            for name in blocked_tools:
                tools_registry.pop(name, None)

        # Build OpenAI tools format
        openai_tools = []
        for name, tool in tools_registry.items():
            openai_tools.append({
                "type": "function",
                "function": {
                    "name": name,
                    "description": tool["description"],
                    "parameters": tool["parameters"],
                },
            })

        # Build messages with system prompt
        openai_messages = [{"role": "system", "content": system_prompt}]
        openai_messages.extend(messages)
        self._last_messages = openai_messages

        # Agentic loop
        for iteration in range(self.max_iterations):
            logger.info("[agent] depth=%d iteration %d, messages=%d", depth, iteration + 1, len(openai_messages))

            try:
                response = await self.client.chat.completions.create(
                    model=self.model,
                    messages=openai_messages,
                    tools=openai_tools if openai_tools else None,
                    max_tokens=self.max_tokens,
                )
            except Exception as e:
                logger.error("[agent] API error: %s", e)
                return f"שגיאה בתקשורת עם שירות ה-AI: {e}"

            choice = response.choices[0]
            message = choice.message

            # If no tool calls, return the text
            if not message.tool_calls:
                return message.content or "לא הצלחתי להבין את הבקשה. נסי שוב."

            # Process tool calls
            openai_messages.append(message.model_dump())

            for tool_call in message.tool_calls:
                fn_name = tool_call.function.name
                try:
                    fn_args = json.loads(tool_call.function.arguments)
                except json.JSONDecodeError:
                    fn_args = {}

                logger.info("[agent] tool call: %s(%s)", fn_name, json.dumps(fn_args, ensure_ascii=False)[:100])

                tool_def = tools_registry.get(fn_name)
                if not tool_def:
                    result = f"❌ כלי לא ידוע: {fn_name}"
                else:
                    try:
                        handler = tool_def["handler"]
                        result = await handler(**fn_args)
                    except Exception as e:
                        logger.error("[agent] tool error %s: %s", fn_name, e)
                        result = f"❌ שגיאה בהפעלת {fn_name}: {e}"

                openai_messages.append({
                    "role": "tool",
                    "tool_call_id": tool_call.id,
                    "content": str(result),
                })

        # Exhausted iterations
        logger.warning("[agent] max iterations (%d) reached", self.max_iterations)
        return "הגעתי למספר המקסימלי של פעולות. נסי לשאול שוב בצורה ממוקדת יותר."

    def get_messages(self) -> list[dict]:
        """Return the conversation messages from the last run (for post-conversation learning)."""
        return getattr(self, "_last_messages", [])