Add support for adding function call items to conversation history via `updateHistory`

[aiddun]

Please read this first

• Have you read the docs? Yes - Agents SDK docs
• Have you searched for related issues? Yes

Describe the feature

Feature: Add support for manually creating and adding function call items to conversation history via updateHistory.

Current Limitation:

Currently, when using updateHistory() to modify the Realtime API conversation history, function call items are ignored with a warning (see code here):

} else if (addition.type === 'function_call') {
  logger.warn(
    'Function calls cannot be manually added or updated at the moment. Ignoring.',
  );
}

This means that while history_updated events include tool calls in the returned history array (RealtimeItem[] which includes RealtimeToolCallItem), there's no way to manually add or restore tool calls when managing conversation history.

[ruskaruma]

Hello @seratch and @aiddun im happy to take this on if it’s unclaimed and not an urgent priority.

I’ll need some time to work through it, but I think it’s worth addressing properly. Let me know if you’re okay with that or have any constraints in mind.

[seratch]

I am not sure why the items are ignored in the initial implementation by @dkundel-openai but there might be a certain reason out there. If enabling function_call items does not bring any issues (perhaps not though), it would be fine to make changes.

[ruskaruma]

That makes sense. I wasn’t sure whether function call items were intentionally excluded due to assumptions in downstream handling or stability concerns, so I wanted to check before making changes. I can dig into the original constraints and evaluate whether enabling function_call items in updateHistory introduces any inconsistencies with the Realtime state machine or event flow.

[ruskaruma]

Before proceeding, it would be great to get some context from @dkundel-openai on the original design decision here. I’ll wait for your input before proposing an implementation. If it looks safe or can be guarded behind clear validation rules, I’m happy to follow up with a small, well scoped change and a narrow test case.

[ruskaruma]

Hello @aiddun,@seratch and @dkundel-openai, I have completed a full audit of the realtime history pipeline and confirmed the function_call restriction is localized to one single serialization branch.

The SDK here already receives and stores completed tool calls in history; updateHistory currently just refuses to restore them, creating a oneway asymmetry. The OpenAI Realtime API documentation describes support for adding function_call items when populating a conversation’s history.

My proposal here to resolve this issue is to allow restoring completed, read-only tool call records only (status=== "completed", output !==null), with strict valdation and no execution, approval, lifecycle hooks, or tracing.

This keeps semantics unambiguous while enabling session restoration and history replay.

Before implementing, I just want to confirm whether the original restriction was intended purely as a safety guard against execution, or if there are other constraints I should account for. Let me know your thoughts here.