[BUG] OpenAIModel tool message content sent as array instead of string breaks OpenAI-compatible endpoints (e.g., Kimi K2.5)

[har1101]

Checks

• I have updated to the lastest minor and patch version of Strands
• I have checked the documentation and this is not expected behavior
• I have searched ./issues and there are no duplicates of my issue

Strands Version

1.26.0

Python Version

3.13

Operating System

macOS 15.5

Installation Method

pip

Steps to Reproduce

1. Create an agent using OpenAIModel with an OpenAI-compatible endpoint
   (e.g., Kimi K2.5 via Bedrock Mantle or Moonshot API directly)
2. Register a simple tool (e.g., current_time from strands-agents-tools)
3. Ask the agent a question that triggers tool use: "What time is it right now? Please use your tools to answer."
4. Observe that the tool executes correctly and returns the correct timestamp
5. Observe that the model ignores the tool result and hallucinates a different datetime

Expected Behavior

The model should use the tool result (In this case, the exact current time. e.g., 2026-02-15T09:22:35) and return the correct datetime in its response.

Actual Behavior

The model returns a completely different, hallucinated datetime(e.g., 2025-04-26T18:25:31) despite the tool result containing the correct value.

Additional Context

Root Cause Analysis

OpenAIModel.format_request_tool_message() sends the tool result content
as an array of content blocks:

# Current behavior (openai.py L207-211)
return {
    "role": "tool",
    "tool_call_id": tool_result["toolUseId"],
    "content": [cls.format_request_message_content(content) for content in contents],  # array
}

While OpenAI's API accepts both array and string formats, many
OpenAI-compatible endpoints (including Kimi K2.5) only correctly process
the string format:

# Expected by Kimi K2.5 (per official docs)
{
    "role": "tool",
    "tool_call_id": tool_call.id,
    "name": tool_call_name,
    "content": json.dumps(tool_result),  # string
}

Kimi K2.5's official tool calling example
(https://huggingface.co/moonshotai/Kimi-K2-Instruct#tool-calling)
uses json.dumps(tool_result) (string format) for the content field.

When the model receives the array format, it fails to parse the tool result
and falls back to hallucinating the answer.

Verified Fix

Subclassing OpenAIModel and overriding format_request_tool_message() to
return content as a string resolves the issue:

class KimiCompatibleOpenAIModel(OpenAIModel):
    @classmethod
    def format_request_tool_message(cls, tool_result, **kwargs):
        text_parts = []
        for content in tool_result["content"]:
            if "json" in content:
                text_parts.append(json.dumps(content["json"]))
            elif "text" in content:
                text_parts.append(content["text"])
        return {
            "role": "tool",
            "tool_call_id": tool_result["toolUseId"],
            "content": "\n".join(text_parts),
        }

Debug Logs

Before fix (array format → hallucinated response):

formatted request=<{... 'content': [{'text': '2026-02-15T09:22:35', 'type': 'text'}] ...}>
Output: The current time is 18:25:31 UTC on April 26, 2025.  ← WRONG

After fix (string format → correct response):

formatted request=<{... 'content': '2026-02-15T09:30:54' ...}>
Output: The current time is 9:30:54 UTC on February 15, 2026.  ← CORRECT

Additional Context

• BedrockModel is not affected because it uses the Bedrock Converse API
  natively, and AWS Bedrock handles format translation to the model.
• This likely affects other OpenAI-compatible providers beyond Kimi K2.5
  (e.g., vLLM, SGLang, Ollama, etc.) that implement the OpenAI API spec
  with strict string-only content parsing for tool messages.

Possible Solution

Option A: Change format_request_tool_message() to always return content
as a string (most compatible with OpenAI-compatible endpoints).

Option B: Add a configuration option to OpenAIModel to switch between
array and string format for tool message content.

Option C: Add both formats support — use string format by default
(broader compatibility) with an option to use array format.

Related Issues

No response