Fix MCP tool timeouts halting agent flow

[habema]

Fixes an issue where MCP tool timeouts (and other errors) would halt the entire agent flow instead of being handled gracefully.

Problem:
When an MCP tool timed out or errored, it would raise an AgentsException that propagated up and stopped the agent. This was inconsistent with regular function tools, which handle errors gracefully via failure_error_function.

Solution:
Wrapped the MCP tool invocation in to_function_tool with error handling that catches exceptions and converts them to error messages using default_tool_error_function, matching the behavior of regular function tools.

Changes:

• Added error handling wrapper in MCPUtil.to_function_tool()
• Errors now return error messages instead of raising exceptions
• New tests added

Repro Script:

import asyncio

from agents import Agent, Runner
from agents.mcp import MCPServerStreamableHttp
from agents.model_settings import ModelSettings


async def main():
    async with MCPServerStreamableHttp(
        name="Streamable HTTP Python Server",
        client_session_timeout_seconds=1, # Very low value for testing purposes
        params={
            "url": "https://docs.mcp.cloudflare.com/mcp",
        },
    ) as server:
        agent = Agent(
            name="Assistant",
            instructions="Use the tools to answer the questions.",
            mcp_servers=[server],
            model_settings=ModelSettings(tool_choice="required"),
        )

        message = "Can you give me some information on how to use the Workers AI Search binding?"
        result = await Runner.run(starting_agent=agent, input=message)
        print(result.final_output)


if __name__ == "__main__":
    asyncio.run(main())

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[seratch]

I will check later but overall it looks safe and good to me.

[seratch]

@codex review again

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: a64f752bb1

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Fix expected error message mismatch in MCP test

The test builds expected_error_msg from AgentsException("Error invoking MCP tool crashing_tool: Crash!"), but invoke_mcp_tool actually wraps failures as "Error invoking MCP tool <name> on server '<server>'", so the result returned by function_tool.on_invoke_tool will include the server name (e.g., fake_mcp_server) and this exact equality assertion will fail. To keep the test reliable, construct the expected message using the same format as invoke_mcp_tool or assert on a stable substring.

Useful? React with 👍 / 👎.