[BUG] OpenAI and Ollama adapters send an empty tool message when a tool returns a dict

Start with _extract_function_response_content in python/packages/kagent-adk/src/kagent/adk/models/_openai.py and _extract_response_content in python/packages/kagent-adk/src/kagent/adk/models/_ollama.py; both return an empty string for dict results without a content list or result key. Run the reproduction script from python/ with uv run python repro.py to see the empty tool messages. Done means both helpers JSON-encode other dicts (matching the Go extractFunctionResponseContent) and the repro shows the dict content reaching the model.

📋 Prerequisites

• I have searched the existing issues to avoid creating a duplicate
• By submitting this issue, you agree to follow our Code of Conduct
• I am using the latest version of the software
• I have tried to clear cache/cookies or used incognito mode (if ui-related)
• I can consistently reproduce this issue

🎯 Affected Service(s)

App Service

🚦 Impact/Severity

Blocker

🐛 Bug Description

When a tool returns a dict, the Python OpenAI and Ollama adapters send the model an empty tool message. The model never sees what the tool returned.

ADK passes a dict returned by a tool through unchanged (only non-dict results are wrapped as {"result": ...}), so this covers:

• any tool that returns a dict, like the {"status": "success", "report": ...} style used in the ADK docs
• ADK's own error responses, e.g. {"error": "Invoking get_weather() failed as the following mandatory input parameters are not present: city ..."} for a missing argument, an unknown tool name, or a rejected confirmation. The model is supposed to read these and retry, but it gets an empty message instead.
• kagent's own tools that return dicts, e.g. ask_user ({"status": "pending", ...}) and the remote A2A tool ({"status": "failed", "error": ...})

Both helpers only handle a string, a dict with a content list (MCP results), or a dict with a result key, and return "" for everything else:

• python/packages/kagent-adk/src/kagent/adk/models/_openai.py, _extract_function_response_content. Used for Chat Completions and the Responses API, so OpenAI, Azure OpenAI, Foundry and Mistral are all affected.
• python/packages/kagent-adk/src/kagent/adk/models/_ollama.py, _extract_response_content

The Go ADK (extractFunctionResponseContent in go/adk/pkg/models/base.go) JSON-encodes other dicts, and the Python Bedrock and SAP AI Core adapters also send the dict as text, so only these two Python helpers drop it.

🔄 Steps To Reproduce

This runs a real ADK agent with kagent's OpenAI model against a small local fake OpenAI server that records the tool messages it receives. The fake model calls get_weather first without its argument, then with it.

import asyncio, json, threading
from http.server import BaseHTTPRequestHandler, HTTPServer
from google.adk.agents import Agent
from google.adk.runners import InMemoryRunner
from google.genai import types
from kagent.adk.models._openai import OpenAI

calls = ["{}", '{"city": "Paris"}']  # first without the required argument, then with it
sent = []

class FakeOpenAI(BaseHTTPRequestHandler):
    def do_POST(self):
        body = json.loads(self.rfile.read(int(self.headers["Content-Length"])))
        sent[:] = [m for m in body["messages"] if m["role"] == "tool"]
        if calls:
            msg = {"role": "assistant", "content": None, "tool_calls": [{"id": f"call_{len(sent) + 1}", "type": "function",
                   "function": {"name": "get_weather", "arguments": calls.pop(0)}}]}
        else:
            msg = {"role": "assistant", "content": "Done."}
        data = json.dumps({"id": "x", "object": "chat.completion", "created": 0, "model": "gpt-4o",
                           "choices": [{"index": 0, "message": msg, "finish_reason": "stop"}]}).encode()
        self.send_response(200); self.send_header("Content-Length", str(len(data))); self.end_headers(); self.wfile.write(data)

    def log_message(self, *args): pass

def get_weather(city: str) -> dict:
    """Get the current weather for a city."""
    return {"status": "success", "report": f"Sunny, 25C in {city}"}

async def main():
    server = HTTPServer(("127.0.0.1", 0), FakeOpenAI)
    threading.Thread(target=server.serve_forever, daemon=True).start()
    model = OpenAI(type="openai", model="gpt-4o", api_key="test", base_url=f"http://127.0.0.1:{server.server_port}/v1")
    runner = InMemoryRunner(agent=Agent(name="weather", model=model, tools=[get_weather]), app_name="demo")
    session = await runner.session_service.create_session(app_name="demo", user_id="u")
    msg = types.Content(role="user", parts=[types.Part.from_text(text="Weather in Paris?")])
    async for event in runner.run_async(user_id="u", session_id=session.id, new_message=msg):
        for part in event.content.parts if event.content else []:
            if part.function_response:
                print(f"tool result in ADK : {part.function_response.response}")
    for m in sent:
        print(f"sent to the model  : tool message for {m['tool_call_id']} -> content={m['content']!r}")

asyncio.run(main())

Run it from python/ with uv run python repro.py.

🤔 Expected Behavior

The model receives the tool result, e.g. the dict serialized as JSON:

sent to the model  : tool message for call_1 -> content='{"error": "Invoking `get_weather()` failed ..."}'
sent to the model  : tool message for call_2 -> content='{"status": "success", "report": "Sunny, 25C in Paris"}'

📱 Actual Behavior

tool result in ADK : {'error': 'Invoking `get_weather()` failed as the following mandatory input parameters are not present:\ncity\nYou could retry calling this tool, but it is IMPORTANT for you to provide all the mandatory parameters.'}
tool result in ADK : {'status': 'success', 'report': 'Sunny, 25C in Paris'}
sent to the model  : tool message for call_1 -> content=''
sent to the model  : tool message for call_2 -> content=''

The tool ran and ADK has both results, but both tool messages sent to the model are empty.

💻 Environment

• OS and version: macOS 26.4 (arm64)
• Application version: main at 12aa4d46, Python 3.13.15, google-adk 2.12.0
• Kubernetes: not needed to reproduce (runs locally against a fake OpenAI server)

🔧 CLI Bug Report

No response

🔍 Additional Context

A possible fix is to fall back to json.dumps(response, default=str) in both helpers instead of returning "", which matches what the Go ADK does. I'd be happy to open a PR with that and unit tests for both adapters, if that sounds right.

📋 Logs

📷 Screenshots

🙋 Are you willing to contribute?

• I am willing to submit a PR to fix this issue