Unified format to define thinking/resoning

[rong-xyz]

Checked

• I searched existing ideas and did not find a similar one
• I added a very descriptive title
• I've clearly described the feature request and motivation for it

Feature request

I am a User that uses models and langchain across multiple providers and users. Unlike Human, AI, Tool Messages that are well defined and is shared amoing different chat models. It seems to me that the thinking in different models are DIVERGING and this is not good.
Eg: In google
https://github.com/langchain-ai/langchain-google/pull/937/files
And in openai it looks like this

This sort of arbitrary definition is not good for a feature that is increasingly common

Motivation

We need a well defined universal definition and schema on how to deal with thought tokens and process them in the sdk, Eg the google genai can't even take the AI msg with thought token back to itself

Proposal (If applicable)

No response

[ccurme]

Hi @rong-xyz, thanks for raising this. We definitely need an update!

We are planning development for a new feature that will allow users to opt-in to a standard AIMessage format that catches up with a few recent innovations from major providers (reasoning output, citations, and potentially other items like web search results and code execution blocks).

Under this format, content would in all cases be a list. For example:

[
    {
        "type": "reasoning",
        "reasoning": "...",
        "id": "...",
    },
    {
        "type": "text",
        "text": "...",
    },
]

This is not a small undertaking unfortunately because it is a breaking change. So:

1. We'd like to enable users to opt-in to the feature;
2. We'd like to standardize a swath of types at once (instead of just reasoning), to minimize breaking changes.

[ccurme]

By the way, yesterday we released a step toward this for ChatOpenAI: https://python.langchain.com/docs/integrations/chat/openai/#responses-api

Setting output_version="responses/v1" when instantiating ChatOpenAI opts-into a new format in which reasoning blocks are structured into message content, similar to Anthropic (although still in OpenAI-native format with nested reasoning summaries).