ChatDatabricks: _convert_lc_messages_to_responses_api incorrectly handling system message with multiple blocks

[ryanfortinOC]

Description

_convert_lc_messages_to_responses_api in databricks_langchain does not properly convert user, system, or developer messages when they contain content blocks. These messages are passed through unchanged with type: "text" (Chat Completions API format) instead of being converted to type: "input_text" (Responses API format).

This causes a 400 Bad Request error:

Invalid value: 'text'. Supported values are: 'input_text', 'input_image', 'output_text', 'refusal', 'input_file', 'computer_screenshot', and 'summary_text'.

Root Cause

The conversion function handles assistant messages correctly but simply passes through user/system/developer messages without converting content block types:

elif role in ("user", "system", "developer"):
    input_items.append(cc_msg)  # ← No conversion of content blocks!

This is problematic because _convert_message_to_dict() returns Chat Completions format (type: "text"), but the Responses API requires type: "input_text".

Why This Matters

1. LangChain middleware commonly builds system prompts as multiple content blocks

For example, deepagents middleware constructs system messages by combining multiple text blocks:

# From deepagents/middleware/_utils.py
system_message = SystemMessage(content=[
    {"type": "text", "text": "Part 1 of system prompt"},
    {"type": "text", "text": "Part 2 of system prompt"},
    # ... additional blocks
])

This pattern is common across LangChain tooling for composing prompts modularly.

2. Some models only support the Responses API

I'm using databricks-gpt-5-1-codex-mini via ChatDatabricks, which exclusively uses the Responses API — the Chat Completions endpoint is not available. This means there is no workaround; the model is completely unusable with any middleware that produces multi-block system messages.

Reproduction

from databricks_langchain import ChatDatabricks
from langchain_core.messages import SystemMessage, HumanMessage

llm = ChatDatabricks(endpoint="databricks-gpt-5-1-codex-mini", use_responses_api=True)

# Multi-block system message (common pattern from middleware)
messages = [
    SystemMessage(content=[
        {"type": "text", "text": "You are a helpful assistant."},
        {"type": "text", "text": "Additional instructions here."},
    ]),
    HumanMessage(content="Hello"),
]

llm.invoke(messages)  # Raises 400 Bad Request

Expected Behavior

The user, system, and developer message branches should convert content blocks the same way assistant messages do:

• type: "text" → type: "input_text"
• type: "image_url" → type: "input_image"

Environment

• langchain-openai: 1.1.7
• databricks-langchain: 0.16.1
• deepagents: 0.4.4
• Model: databricks-gpt-5-1-codex-mini (Responses API only)

---