Description
When using the OpenAI agents library (v0.2.9) with realtime sessions, the tool_start events do not include the arguments/parameters being passed to the tool. This makes it impossible to display tool inputs in the UI, limiting visibility into what operations are being performed.
Current Behavior
tool_start events only contain: agent, tool, info, and type attributes
- The
info attribute contains only context with usage metrics
- Tool arguments are passed to the MCP server (visible in debug logs from overridden
call_tool method) but not exposed in the event object
- No apparent way to access the input parameters from the
RealtimeToolStart event
Expected Behavior
tool_start events should include an arguments or input field containing the parameters being passed to the tool
- This would allow UIs to display what inputs are being processed, similar to how
tool_end events expose the output/results
Example of desired behavior:
# Current tool_end event (working):
{
"type": "tool_end",
"tool": "AddNumbers",
"output": '{"result": 42}'
}
# Desired tool_start event (not currently available):
{
"type": "tool_start",
"tool": "AddNumbers",
"arguments": {"a": 15, "b": 27} # <-- This is missing
}
Environment
- openai-agents version: 0.2.9
- Python version: 3.13.2
- Use case: Realtime agent cohort with MCP server integration
Description
When using the OpenAI agents library (v0.2.9) with realtime sessions, the
tool_startevents do not include the arguments/parameters being passed to the tool. This makes it impossible to display tool inputs in the UI, limiting visibility into what operations are being performed.Current Behavior
tool_startevents only contain:agent,tool,info, andtypeattributesinfoattribute contains onlycontextwith usage metricscall_toolmethod) but not exposed in the event objectRealtimeToolStarteventExpected Behavior
tool_startevents should include anargumentsorinputfield containing the parameters being passed to the tooltool_endevents expose the output/resultsExample of desired behavior:
Environment