AI assist fails for any openai o1 model

[awooler]

Describe the bug

Hello, very happy to see that the openai o1 model family is now available through AI assist. However when I try to use any of the o1 models, I see this error:

openai.BadRequestError: Error code: 400 - {'error': {'message': "Unsupported value: 'messages[0].role' does not support 'system' with this model.", 'type': 'invalid_request_error', 'param': 'messages[0].role', 'code': 'unsupported_value'}}

I can confirm that I don't have any custom instructions input to my marimo settings, and I believe o1 models don't support system instruction, so assuming that marimo is trying to pass system instructions by default which triggers the issue.

Environment

{
  "marimo": "0.10.17",
  "OS": "Darwin",
  "OS Version": "24.1.0",
  "Processor": "i386",
  "Python Version": "3.9.0",
  "Binaries": {
    "Browser": "132.0.6834.112",
    "Node": "--"
  },
  "Dependencies": {
    "click": "8.1.7",
    "docutils": "0.17.1",
    "itsdangerous": "2.2.0",
    "jedi": "0.19.2",
    "markdown": "3.7",
    "narwhals": "1.18.4",
    "packaging": "24.2",
    "psutil": "5.9.8",
    "pygments": "2.18.0",
    "pymdown-extensions": "10.12",
    "pyyaml": "6.0.2",
    "ruff": "0.8.3",
    "starlette": "0.27.0",
    "tomlkit": "0.13.2",
    "typing-extensions": "4.12.2",
    "uvicorn": "0.34.0",
    "websockets": "14.1"
  },
  "Optional Dependencies": {
    "altair": "5.5.0",
    "pandas": "2.2.3",
    "pyarrow": "18.1.0"
  },
  "Experimental Flags": {
    "chat_sidebar": true,
    "tracing": false
  }
}

Code to reproduce

No response

[mscolnick]

hey @awooler, we can likely check if the model starts with o1 and remove ignore the system prompt. would you want to contribute this fix? i don't use the o1 models myself yet.

[awooler]

hey, I was mistaken that the o1 models don't support a system prompt, that used to be the case but seems like they added a "developer" prompt instead.

however see this thread: https://community.openai.com/t/how-is-developer-message-better-than-system-prompt/1062784

they say that the "developer" message should be forwards/backwards compatible with the "system" message, which others are reporting is not the case for them as well, so probably will stick to sonnet for now until fixed on their side

[mscolnick]

Sounds good, we can wait until openai new models' API becomes more stable