Summary
OpenAIModel.__init__ declares observer: MessageObserver = MessageObserver — the default value is the class object, not an instance. Any caller that omits observer crashes as soon as the streaming path fires, because the unbound method is called with only one argument.
Reproduction
from nexent.core.models import OpenAIModel
m = OpenAIModel(model_id="gpt-4o-mini", api_key="sk-...") # observer omitted
m([{"role": "user", "content": [{"type": "text", "text": "Say hello in three words."}]}]) self.observer.add_model_new_token(new_token)
TypeError: MessageObserver.add_model_new_token() missing 1 required positional argument: 'new_token'
The HTTP request to the provider succeeds first — the failure happens while handling the response, so the call is billed and then thrown away, which makes it a slightly expensive way to find out.
Passing an instance works:
from nexent.core.utils.observer import MessageObserver
m = OpenAIModel(observer=MessageObserver(), model_id="gpt-4o-mini", api_key="sk-...")
# nexent replied: the stub answered
Where
nexent/core/models/openai_llm.py:18:
def __init__(self, observer: MessageObserver = MessageObserver, temperature=0.2, ...):
The docstring right below says "observer: MessageObserver instance for tracking model output", so the intent looks clear and this reads like a missing ().
Suggested fix
Either default to None and build one lazily — the usual way to avoid a shared mutable default:
def __init__(self, observer: MessageObserver | None = None, ...):
self.observer = observer if observer is not None else MessageObserver()or make it required. = MessageObserver() in the signature would also work but gives every instance the same observer, which is probably not what you want.
Same shape appears in openai_vlm.py and openai_long_context_model.py if they share the signature — I did not check those.
Environment
nexent 1.8.1, Python 3.12, Windows. Found while pointing a model at a local OpenAI-compatible endpoint.
Unrelated but possibly useful
While I was here: omitting api_base leaves OpenAIServerModel to construct the client without a base_url, so the OpenAI SDK falls back to OPENAI_BASE_URL. That makes it easy to point nexent at a local endpoint for testing without touching config — I confirmed the resolved value is the local address and that the run reproduces byte-for-byte offline afterwards. Mentioning it only because "how do I aim this at my own endpoint" is a common question; happy to leave it out if it is noise here.
Summary
OpenAIModel.__init__declaresobserver: MessageObserver = MessageObserver— the default value is the class object, not an instance. Any caller that omitsobservercrashes as soon as the streaming path fires, because the unbound method is called with only one argument.Reproduction
The HTTP request to the provider succeeds first — the failure happens while handling the response, so the call is billed and then thrown away, which makes it a slightly expensive way to find out.
Passing an instance works:
Where
nexent/core/models/openai_llm.py:18:The docstring right below says "observer: MessageObserver instance for tracking model output", so the intent looks clear and this reads like a missing
().Suggested fix
Either default to
Noneand build one lazily — the usual way to avoid a shared mutable default:or make it required.
= MessageObserver()in the signature would also work but gives every instance the same observer, which is probably not what you want.Same shape appears in
openai_vlm.pyandopenai_long_context_model.pyif they share the signature — I did not check those.Environment
nexent1.8.1, Python 3.12, Windows. Found while pointing a model at a local OpenAI-compatible endpoint.Unrelated but possibly useful
While I was here: omitting
api_baseleavesOpenAIServerModelto construct the client without abase_url, so the OpenAI SDK falls back toOPENAI_BASE_URL. That makes it easy to point nexent at a local endpoint for testing without touching config — I confirmed the resolved value is the local address and that the run reproduces byte-for-byte offline afterwards. Mentioning it only because "how do I aim this at my own endpoint" is a common question; happy to leave it out if it is noise here.