[Bug]: OpenAIResponses + FunctionAgent always returns an empty string response when NOT streaming

[jbrowning]

Bug Description

When the following conditions are met:

1. A FunctionAgent with tools
2. When OpenAIResponses is the llm
3. When the response is NOT streamed...

LlamaIndex will return an empty string response immediately after the tool call instead of fetching the LLM's final response.

Version

0.14.2

Steps to Reproduce

Example repo: https://github.com/jbrowning/llama-index-openairesponses-bug

Relevant Logs/Tracbacks

[dosubot]

This is a confirmed issue: when using FunctionAgent with tools and OpenAIResponses as the LLM in non-streaming mode, the agent returns an empty string after a tool call instead of the expected final LLM response. This happens because FunctionAgent expects the LLM to populate message.content in its ChatResponse after tool calls, but OpenAIResponses does not do this correctly in non-streaming mode, leading to an empty string being returned as the final response (source, source).

This appears to be a design oversight in the interaction between FunctionAgent's non-streaming logic and the OpenAIResponses wrapper. The codebase is heavily optimized for streaming, and the non-streaming fallback path does not always trigger a final LLM completion after tool execution.

Workarounds: switch to streaming mode if possible, or use a different LLM wrapper such as OpenAILike if you need non-streaming responses. There is no direct fix in the current release, and no open ticket tracks this exact bug, but related issues and PRs suggest the maintainers are aware of similar response handling problems. Keep an eye on future updates for a permanent fix.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}