reasoning_agent.py

"""Reasoning agent — emits AG-UI REASONING_MESSAGE_* events.

Shared by reasoning-custom (custom amber ReasoningBlock) and
reasoning-default (CopilotKit's built-in reasoning slot).

Why a reasoning model + Responses API:
The OpenAI Responses API streams `response.reasoning_summary_text.delta`
items only for native reasoning models (gpt-5, o3, o4-mini, etc.).
CopilotKit's bridge translates those into AG-UI REASONING_MESSAGE_*
events with `role: "reasoning"`, which the frontend renders via the
`reasoningMessage` slot. gpt-4o / gpt-4o-mini do not emit reasoning
items, so a non-reasoning model would never light up the slot.
"""

from __future__ import annotations

import os

from deepagents import create_deep_agent
from langchain.chat_models import init_chat_model

from src.agents._header_forwarding_middleware import HeaderForwardingMiddleware

SYSTEM_PROMPT = (
    "You are a helpful assistant. For each user question, first think "
    "step-by-step about the approach, then give a concise answer."
)

REASONING_MODEL = os.environ.get("OPENAI_REASONING_MODEL", "gpt-5.4")

# No full CopilotKitMiddleware — this demo exercises only reasoning-token
# streaming through the OpenAI Responses API and doesn't consume frontend
# tools or app context. We still attach the minimal HeaderForwardingMiddleware
# so the inbound ``x-aimock-context`` (and other ``x-*``) headers reach the
# outgoing /v1/responses call; without it the LangGraph run swallows them
# inside ``configurable`` and aimock 404s with no fixture match. The minimal
# middleware does ONLY header propagation — no App-Context injection, no
# tool-merging, no state-surfacing.
graph = create_deep_agent(
    model=init_chat_model(
        f"openai:{REASONING_MODEL}",
        use_responses_api=True,
        reasoning={"effort": "medium", "summary": "detailed"},
    ),
    tools=[],
    system_prompt=SYSTEM_PROMPT,
    middleware=[HeaderForwardingMiddleware()],
)