Reasoning Messages

"""Reasoning agent for LlamaIndex.Shared by `reasoning-custom` (custom amber ReasoningBlock slot) and`reasoning-default` (CopilotKit's built-in reasoning slot). The systemprompt asks the model to think step-by-step before answering, so the LLMproduces a reasoning channel that the chat UI can render.Why a reasoning model + the OpenAI Responses API------------------------------------------------Mirrors the langgraph-python parity gold standard(`init_chat_model("openai:<reasoning-model>", use_responses_api=True,reasoning={"effort": "medium", "summary": "detailed"})`). The OpenAIResponses API streams `response.reasoning_summary_text.delta` items only fornative reasoning models (gpt-5, o3, o4-mini, …); a non-reasoning model likegpt-4.1 on the chat-completions wire emits NO reasoning channel against realOpenAI, so the reasoning slot would only ever light up under aimock. Routingthrough `OpenAIResponses` with a reasoning model makes the chain of thoughtstream against a REAL provider; aimock renders the fixture's abstract`reasoning` field into the same Responses-API shape for deterministic tests.(LlamaIndex pins the default to `gpt-5` rather than langgraph's `gpt-5.4`because LlamaIndex 0.5.6 rejects model names absent from its context-sizetable at workflow construction — see REASONING_MODEL below.)Uses `get_reasoning_ag_ui_workflow_router` (a thin in-package extension of thestock `get_ag_ui_workflow_router`) so the model's reasoning summary deltassurface as AG-UI `REASONING_MESSAGE_*` events. The stock router reads onlyassistant text and silently drops the reasoning channel; see`_reasoning_router.py` for the three framework gaps it closes (and for how`_extract_reasoning_delta` reads the Responses-API summary delta off`resp.raw`, which LlamaIndex's own stream processing does not surface). Thefrontend `CopilotChatReasoningMessage` slot then composes with the flow."""from __future__ import annotationsimport osfrom llama_index.llms.openai import OpenAIResponsesfrom agents._reasoning_router import get_reasoning_ag_ui_workflow_routerSYSTEM_PROMPT = (    "You are a helpful assistant. For each user question, first think "    "step-by-step about the approach, then give a concise answer. Keep "    "responses brief -- 1 to 3 sentences max.")# Reasoning-capable model routed through the OpenAI Responses API.## Default is `gpt-5` (a native reasoning model), NOT the langgraph gold# standard's `gpt-5.4`. LlamaIndex 0.5.6's `OpenAIResponses.metadata` resolves# the context window through `openai_modelname_to_contextsize()`, which raises# `ValueError: Unknown model` for names outside its hard-coded table —# `AGUIChatWorkflow.__init__` reads `llm.metadata.is_function_calling_model`,# so an unrecognized name (like `gpt-5.4`) crashes workflow construction at# startup. `gpt-5` is in both that table AND the O1_MODELS reasoning list, so# it streams reasoning natively against real OpenAI. Deployments can override# via OPENAI_REASONING_MODEL (with any name LlamaIndex 0.5.6 recognizes).REASONING_MODEL = os.environ.get("OPENAI_REASONING_MODEL", "gpt-5")# `summary: detailed` requests the streamed reasoning summary; `effort:# medium` mirrors the gold config. We pass these through BOTH# `reasoning_options` (idiomatic; honored for O1_MODELS like gpt-5) AND# `additional_kwargs` (unconditionally merged into the /v1/responses body by# `OpenAIResponses._get_model_kwargs`), so the `reasoning` param still reaches# the wire if a deployment overrides to a reasoning model outside the# O1_MODELS allowlist._REASONING_PARAMS = {"effort": "medium", "summary": "detailed"}_openai_kwargs = {}if os.environ.get("OPENAI_BASE_URL"):    _openai_kwargs["api_base"] = os.environ["OPENAI_BASE_URL"]reasoning_router = get_reasoning_ag_ui_workflow_router(    llm=OpenAIResponses(        model=REASONING_MODEL,        reasoning_options=_REASONING_PARAMS,        additional_kwargs={"reasoning": _REASONING_PARAMS},        **_openai_kwargs,    ),    frontend_tools=[],    backend_tools=[],    system_prompt=SYSTEM_PROMPT,    initial_state={},)

"""Reasoning agent for LlamaIndex.Shared by `reasoning-custom` (custom amber ReasoningBlock slot) and`reasoning-default` (CopilotKit's built-in reasoning slot). The systemprompt asks the model to think step-by-step before answering, so the LLMproduces a reasoning channel that the chat UI can render.Why a reasoning model + the OpenAI Responses API------------------------------------------------Mirrors the langgraph-python parity gold standard(`init_chat_model("openai:<reasoning-model>", use_responses_api=True,reasoning={"effort": "medium", "summary": "detailed"})`). The OpenAIResponses API streams `response.reasoning_summary_text.delta` items only fornative reasoning models (gpt-5, o3, o4-mini, …); a non-reasoning model likegpt-4.1 on the chat-completions wire emits NO reasoning channel against realOpenAI, so the reasoning slot would only ever light up under aimock. Routingthrough `OpenAIResponses` with a reasoning model makes the chain of thoughtstream against a REAL provider; aimock renders the fixture's abstract`reasoning` field into the same Responses-API shape for deterministic tests.(LlamaIndex pins the default to `gpt-5` rather than langgraph's `gpt-5.4`because LlamaIndex 0.5.6 rejects model names absent from its context-sizetable at workflow construction — see REASONING_MODEL below.)Uses `get_reasoning_ag_ui_workflow_router` (a thin in-package extension of thestock `get_ag_ui_workflow_router`) so the model's reasoning summary deltassurface as AG-UI `REASONING_MESSAGE_*` events. The stock router reads onlyassistant text and silently drops the reasoning channel; see`_reasoning_router.py` for the three framework gaps it closes (and for how`_extract_reasoning_delta` reads the Responses-API summary delta off`resp.raw`, which LlamaIndex's own stream processing does not surface). Thefrontend `CopilotChatReasoningMessage` slot then composes with the flow."""from __future__ import annotationsimport osfrom llama_index.llms.openai import OpenAIResponsesfrom agents._reasoning_router import get_reasoning_ag_ui_workflow_routerSYSTEM_PROMPT = (    "You are a helpful assistant. For each user question, first think "    "step-by-step about the approach, then give a concise answer. Keep "    "responses brief -- 1 to 3 sentences max.")# Reasoning-capable model routed through the OpenAI Responses API.## Default is `gpt-5` (a native reasoning model), NOT the langgraph gold# standard's `gpt-5.4`. LlamaIndex 0.5.6's `OpenAIResponses.metadata` resolves# the context window through `openai_modelname_to_contextsize()`, which raises# `ValueError: Unknown model` for names outside its hard-coded table —# `AGUIChatWorkflow.__init__` reads `llm.metadata.is_function_calling_model`,# so an unrecognized name (like `gpt-5.4`) crashes workflow construction at# startup. `gpt-5` is in both that table AND the O1_MODELS reasoning list, so# it streams reasoning natively against real OpenAI. Deployments can override# via OPENAI_REASONING_MODEL (with any name LlamaIndex 0.5.6 recognizes).REASONING_MODEL = os.environ.get("OPENAI_REASONING_MODEL", "gpt-5")# `summary: detailed` requests the streamed reasoning summary; `effort:# medium` mirrors the gold config. We pass these through BOTH# `reasoning_options` (idiomatic; honored for O1_MODELS like gpt-5) AND# `additional_kwargs` (unconditionally merged into the /v1/responses body by# `OpenAIResponses._get_model_kwargs`), so the `reasoning` param still reaches# the wire if a deployment overrides to a reasoning model outside the# O1_MODELS allowlist._REASONING_PARAMS = {"effort": "medium", "summary": "detailed"}_openai_kwargs = {}if os.environ.get("OPENAI_BASE_URL"):    _openai_kwargs["api_base"] = os.environ["OPENAI_BASE_URL"]reasoning_router = get_reasoning_ag_ui_workflow_router(    llm=OpenAIResponses(        model=REASONING_MODEL,        reasoning_options=_REASONING_PARAMS,        additional_kwargs={"reasoning": _REASONING_PARAMS},        **_openai_kwargs,    ),    frontend_tools=[],    backend_tools=[],    system_prompt=SYSTEM_PROMPT,    initial_state={},)