[Fix] Error from OpenAI Responses API with web_search

[dawidkulpa]

Title

[Fix] Error from OpenAI Responses API with web_search

Relevant issues

When OpenAI Responses API is called with web_search tool, and the response contains those web_search calls, LiteLLM throws an error with 503 status:

litellm.ServiceUnavailableError: litellm.MidStreamFallbackError: litellm.APIConnectionError: APIConnectionError: OpenAIException - Error parsing chunk: Chat provider: Invalid output_item  {'id': 'ws_0e5386c91dd50f830068ebe883849081919a0164a2c5f5a7f5', 'type': 'web_search_call', 'status': 'in_progress'},
Received chunk: type=<ResponsesAPIStreamEvents.OUTPUT_ITEM_ADDED: 'response.output_item.added'> output_index=1 item={'id': 'ws_0e5386c91dd50f830068ebe883849081919a0164a2c5f5a7f5', 'type': 'web_search_call', 'status': 'in_progress'} sequence_number=563.

Pre-Submission checklist

Please complete all items before asking a LiteLLM maintainer to review your PR

• I have Added testing in the tests/litellm/ directory, Adding at least 1 test is a hard requirement - see details
• I have added a screenshot of my new test passing locally
• My PR passes all unit tests on make test-unit
• My PR's scope is as isolated as possible, it only solves 1 specific problem

Type

🐛 Bug Fix

Changes

• when transforming Responses API response to Completion API format, unsupported web_search calls are skipped now.

Tests

Before change:

After change:

[vercel]

@dawidkulpa is attempting to deploy a commit to the CLERKIEAI Team on Vercel.

A member of the Team first needs to authorize it.

[krrishdholakia]

hey @dawidkulpa, shouldn't we be trying to return it to the user, instead of ignoring it?

[dawidkulpa]

Hey! Good question - I was debating it with myself as well. The issue is that would be a real hack, as Completions API just don't support giving statuses from these kind of calls. We could map it to something simple, but that would definitely be against what API provides.
I honestly would like that, but skipped it just because it's a hack and didn't want to bother you with explanations and potentially more difficult maintainability.
Just let me know what's your opinion, I can implement it :)