effect/ai: provider web_search tool-result is treated as local tool output and sent to OpenAI /responses API as `function_call_output`

[tensor2077]

What version of Effect is running?

effect/ai 0.33.0

What steps can reproduce the bug?

1. Enable the provider-defined web_search tool and send a streaming /responses request (gpt5.2).
2. Ask the model to perform a web search (e.g., “Use web search to define AI agent and summarize.”).
3. After the first turn completes, send a second message in the same chat session (any text).
4. Observe HTTP 400: No tool call found for function call output with call_id ws_....

What is the expected behavior?

success response

What do you see instead?

{
"error": {
"message": "No tool call found for function call output with call_id ws_029e7d51de056f790069536f1df6488196ab10f32c7a3d915e.",
"type": "invalid_request_error",
"param": "input",
"code": null
}
}

Additional information

After the first turn (a web search + answer), the persisted history looks like:

 {
    "content": [
      {
        "role": "system",
        "content": "You are an AI assistant.",
        "options": {}
      },
      {
        "role": "user",
        "content": "Use web search to look up the concept of an acquisition and summarize briefly.",
        "options": {}
      },
      {
        "role": "assistant",
        "content": [
          {
            "type": "tool-call",
            "id": "ws_029e7d51de056f790069536f1df6488196ab10f32c7a3d915e",
            "name": "web_search",
            "params": {
              "action": {
                "type": "search",
                "query": "acquisition definition",
                "sources": [{ "type": "url", "url": "https://en.wikipedia.org/wiki/Acquisition" }]
              }
            },
            "providerExecuted": true,
            "options": {}
          },
          {
            "type": "text",
            "text": "An acquisition is ...",
            "options": {}
          }
        ],
        "options": {}
      },
      {
        "role": "tool",
        "content": [
          {
            "type": "tool-result",
            "id": "ws_029e7d51de056f790069536f1df6488196ab10f32c7a3d915e",
            "name": "web_search",
            "isFailure": false,
            "result": { "status": "completed" },
            "options": {}
          }
        ],
        "options": {}
      }
    ]
  }

So far, everything looks correct.
However, sending any second user message fails with:

  {
    "error": {
      "message": "No tool call found for function call output with call_id ws_029e7d51de056f790069536f1df6488196ab10f32c7a3d915e.",
      "type": "invalid_request_error",
      "param": "input",
      "code": null
    }
  }

ERROR cause

The provider-executed tool result

 {
        "role": "tool",
        "content": [
          {
            "type": "tool-result",
            "id": "ws_029e7d51de056f790069536f1df6488196ab10f32c7a3d915e",
            "name": "web_search",
            "isFailure": false,
            "result": { "status": "completed" },
            "options": {}
          }
        ],
        "options": {}
      }

was mistakenly treated as a local tool call result and sent back to the OpenAI API as a function_call_output.

The request payload look like this

{
    "model": "gpt-5.2",
    "reasoning": {
        "summary": "auto"
    },
    "input": [
        {
            "role": "developer",
            "content": "You are an AI assistant."
        },
        {
            "role": "user",
            "content": [
                {
                    "type": "input_text",
                    "text": ""
                }
            ]
        },
        {
            "role": "user",
            "content": [
                {
                    "type": "input_text",
                    "text": "Use web search to look up the concept of an acquisition and summarize briefly."
                }
            ]
        },
        {
            "role": "assistant",
            "content": [
                {
                    "type": "output_text",
                    "text": "{AI_RESPONSE_TEXT}"
                }
            ]
        },
        {
            "type": "function_call_output",
            "call_id": "ws_029e7d51de056f790069536f1df6488196ab10f32c7a3d915e",
            "output": "{\"status\":\"completed\"}"
        },
        {
            "role": "user",
            "content": [
                {
                    "type": "input_text",
                    "text": "go on"
                }
            ]
        }
    ],
    "include": [
        "web_search_call.action.sources"
    ],
    "text": {
        "format": {
            "type": "text"
        }
    },
    "tools": [
        {
            "search_context_size": "medium",
            "type": "web_search"
        }
    ],
    "tool_choice": "auto",
    "stream": true
}

[IMax153]

@tensor2077 - thanks for reporting the issue. I'm happy to fix this, but unsure what the expected shape of the second request should be?

Looking at the Vercel AI SDK implementation, the logic seems the same.

[tensor2077]

@IMax153
I have made a request through OpenAI’s official Agents SDK (openai-agents-js).

Below is the request payload of my second message

{
  "model": "gpt-5.2",
  "instructions": "You are an AI assistant.",
  "input": [
    {
      "role": "user",
      "content": [
        {
          "type": "input_text",
          "text": "Use web search to look up the concept of an acquisition and summarize briefly."
        }
      ]
    },
    {  // <- the tool call item
      "id": "ws_03df6a7f46996cdb006955e82dfd688194b1a151580763df4f",
      "type": "web_search_call",
      "action": {
        "type": "search",
        "query": "business acquisition definition concept"
      },
      "status": "completed"
    },
    {
      "type": "message",
      "id": "msg_03df6a7f46996cdb006955e83017948194aff1c37dda1a5354",
      "role": "assistant",
      "content": [
        {
          "type": "output_text",
          "text": "An **acquisition** is a business transaction where one company (the **acquirer**) **buys enough of another company (the target)**—typically by purchasing its **shares/stock** or its **assets**—to obtain **control** over it. In many acquisitions, the acquired company becomes a **subsidiary** of the acquirer (they remain separate legal entities, but the parent controls key decisions). ([law.cornell.edu](https://www.law.cornell.edu/wex/acquisition?utm_source=openai))\n\nThis is different from a **merger**, which usually combines two companies into a more unified organization (often framed as closer to “equals”), whereas an acquisition is a **purchase/takeover**. ([britannica.com](https://www.britannica.com/money/mergers-and-acquisitions-strategy?utm_source=openai))",
          "annotations": []
        }
      ],
      "status": "completed"
    },
    {
      "role": "user",
      "content": [
        {
          "type": "input_text",
          "text": "go on"
        }
      ]
    }
  ],
  "include": [],
  "tools": [
    {
      "type": "web_search",
      "search_context_size": "medium"
    }
  ],
  "stream": false
}

reference

https://github.com/openai/openai-agents-js/blob/5d9b751b6ebe6d0fb94a74a4d93a3d6f31790f96/packages/agents-openai/src/openaiResponsesModel.ts#L1100-L1113

    if (item.type === 'hosted_tool_call') {
      if (
        item.providerData?.type === 'web_search_call' ||
        item.providerData?.type === 'web_search' // for backward compatibility
      ) {
        const entry: OpenAI.Responses.ResponseFunctionWebSearch = {
          ...camelOrSnakeToSnakeCase(item.providerData), // place here to prioritize the below fields
          type: 'web_search_call',
          id: item.id!,
          status: WebSearchStatus.parse(item.status ?? 'failed'),
        };

        return entry;
      }

https://github.com/openai/openai-node/blob/cb3987b4840aedd3ca8400ca8fa444fd927ada73/src/lib/responses/ResponseStream.ts#L203

#accumulateResponse(event: ResponseStreamEvent): Response {

---

The code snippet I use to send and log request:

https://gist.github.com/tensor2077/64f689156385b31d9867b843dfe2e48c