Shell tool in Responses API not available bug

[michprev]

Since today, the native local shell tool stopped working. The LLM complains that it has no shell available. Tested with GPT 5.1 and 5.2. Tested both with leaving environment unset and setting it to local.

[realglyph]

Need an update to this fix @OpenAI_Support

[davide.sanvito]

Same issue here @OpenAI_Support

[sven.den.hartog]

Same issue. It’s very annoying that features will break within a few days of implementing around it. Maybe there should be a set of features that are marked as “beta” so we won’'t build production ready code around it. @OpenAI_Support

[wendyjiao]

Hi, openai engineer here. Can I get some request id for the failed calls?

[realglyph]

@wendyjiao have lost them, but you can reproduce this in the playground select the built in tools: apply_patch and shell then add a custom function tool i.e weather tool.

if you try prompting it, the model says “i dont have access to the shell/apply_patch tools”

removing the custom function tool is the only fix for them to work properly

[davide.sanvito]

As of today, the bug seems fixed

[wendyjiao]

We’ve fixed the bug yesterday. Thanks everyone for reporting this!

[jwehrstein]

The issue appeared today again @wendyjiao @OpenAI_Support
Experienced it with gpt-5.4-2026-03-05 (with tool-search-tool).
Shell & Apply-Patch tools are not visible to the agent, only my custom function tools.

LLM Output states that it does not have Shell or ApplyPatch Tool. Not even seeing Tool-Search-Tool. Though response log (properties section) shows all tools are present (respID: resp_036dd8334816af9c0069b1716f53c0819abd03a41f0be5d38e )

[mjasny]

Also with older models like “gpt-5.3-codex” it is still not working reliably and all our agent workflows are broken! Please fix!! @wendyjiao @OpenAI_Support

[wendyjiao]

thanks for reporting, taking a look

[wendyjiao]

@jwehrstein could you share what your input looks like? couldn’t repro on my side.

[jwehrstein]

Thanks for looking into it. You can e.g. check this multi-turn trace (send via Agents SDK today Mar 11, 2026, 1:43 PM UTC) with response ID resp_036dd8334816af9c0069b1716f53c0819abd03a41f0be5d38e - i hope you have access to it. It is a 12k token input, containing several user/assistant interactions and a system prompt explaining all tools (including shell and apply-patch tool explanations similar to the github examples). Tools include two custom FunctionTools, and the OpenAI provided Shell-Tool, Apply-Patch Tool and ToolSearchTool.

Can I assist you with additional information?

[_j]

I checked hosted shell and shell function in the chat playground.

Both seem to be working. I returned some stdout of the AI’s failure (to make it think it had Powershell), in the site’s offered template for the multiple function call return, and AI tried again.

Hosted, transmitted:

{
  "model": "gpt-5.2",
  "input": [
    {
      "role": "developer",
      "content": [
        {
          "type": "input_text",
          "text": "..."
        }
      ]
    },
    {
      "role": "user",
      "content": [
        {
          "type": "input_text",
          "text": "..."
        }
      ]
    }
  ],
  "tools": [
    {
      "type": "shell",
      "environment": {
        "type": "container_auto"
      }
    }
  ],
  "text": {
    "format": {
      "type": "text"
    },
    "verbosity": "medium"
  },
  "reasoning": {
    "effort": "medium",
    "summary": "auto"
  },
  "stream": true,
  "store": false,
  "include": [
    "reasoning.encrypted_content",
    "web_search_call.action.sources"
  ]
}

Local

{
  "model": "gpt-5.2",
  "input": [
    {
      "role": "developer",
      "content": [
        {
          "type": "input_text",
          "text": "..."
        }
      ]
    },
    {
      "role": "user",
      "content": [
        {
          "type": "input_text",
          "text": "..."
        }
      ]
    }
  ],
  "tools": [
    {
      "type": "shell"
    }
  ],
  "text": {
    "format": {
      "type": "text"
    },
    "verbosity": "medium"
  },
  "reasoning": {
    "effort": "medium",
    "summary": "auto"
  },
  "stream": true,
  "store": false,
  "include": [
    "reasoning.encrypted_content",
    "web_search_call.action.sources"
  ]
}

The problem might be persistence of another forum topic’s issue: developer functions along with shell or apply_patch breaks the API.

I added a custom function with plausible plaintext, also gave a try with patch function, and the AI still gave its “bad shell code” a retry, no error.

---

Moar

GPT-5.2

GPT-5.4

All local == all functions