Azure Responses API: Bad Request when Codex sends tool output (shell_tool). function_call_output expects string, Codex appears to send object

What version of the IDE extension are you using?

v26.311.21342

What subscription do you have?

Azure OpenAI / no OpenAI ChatGPT plan

Which IDE are you using?

VS Code

What platform is your computer?

Darwin 25.3.0 arm64 arm

What issue are you seeing?

• When using Codex v0.114.0 or the VS Code Codex extension with Azure OpenAI (Responses API), normal chat works but any tool execution (shell_tool / exec_command) results in “Bad Request”. However, downgrading the VS Code Codex extension to v0.4.69 makes the issue disappear.
• Futhermore, disabling shell tools avoids the error. Azure accepts tool calls and tool outputs when the output is a string, but returns 400 if the output is an object.

What steps can reproduce the bug?

1. codex exec "List files in this repo."
2. Observe "Bad Request" after tool call attempt.

What is the expected behavior?

Tool call succeeds, and model returns output.

Additional information

Environment

• Codex version: 0.114.0
• VS Code Codex extension: 0.4.69 fixes it; newer versions (v26.311.21342) reproduce the issue
• OS: macOS
• Provider: Azure OpenAI (Responses API)
• Model/deployment: gpt-5.2-codex
• base_url: https://<resource_name>.cognitiveservices.azure.com/openai/v1
• wire_api: responses
• model_reasoning_effort: medium

Additional Evidence (Azure curl test)

Azure accepts function_call_output when output is string:

POST /responses
{
"model": "gpt-5.2-codex",
"previous_response_id": "<RESP_ID>",
"input": [{
"type": "function_call_output",
"call_id": "<CALL_ID>",
"output": "{"files":["a.txt","b.txt"]}"
}]
}
=> 200 OK

Azure rejects output object:

POST /responses
{
"model": "gpt-5.2-codex",
"previous_response_id": "<RESP_ID>",
"input": [{
"type": "function_call_output",
"call_id": "<CALL_ID>",
"output": { "stdout": "a\nb\n", "stderr": "", "exit_code": 0 }
}]
}
=> 400 invalid_request_error: Invalid type for input[0].output (expected string or array of objects)

Azure also rejects type "tool_output":

=> 400 invalid_request_error: Invalid value 'tool_output' (supported includes function_call_output)

Logs (Codex)

In codex-tui.log:

• Tool call logged: ToolCall: exec_command {"cmd":"ls"}
• Immediately after: Turn error: Bad Request

Hypothesis

Codex may be sending tool output in a structured object format that Azure Responses API rejects (expects a string or array of objects). This causes 400 on tool output submission.