🐛 Bugfix: default OpenAIModel observer to a fresh instance, not the class

[XixianWasTaken]

Summary

Fixes #3921

OpenAIModel.__init__ declared observer: MessageObserver = MessageObserver — the default was the class object, not an instance. Any caller that omitted observer crashed in the streaming path with:

TypeError: MessageObserver.add_model_new_token() missing 1 required positional argument: 'new_token'

…and only after the provider request had already succeeded (and been billed).

Changes

• sdk/nexent/core/models/openai_llm.py
  • observer now defaults to None; __init__ builds a MessageObserver() lazily when it is omitted, so each model gets its own observer rather than a signature-level instance shared across every model. Explicit observers are stored unchanged.
  • Docstring updated to describe the default.
• test/sdk/core/models/test_openai_llm_observer_default.py (new)
  • Default is a MessageObserver instance, not the class.
  • Two models constructed without an observer get distinct instances.
  • Explicit instance is preserved; explicit None behaves like omitting it.
  • Streaming call with observer omitted succeeds and the default observer captures the streamed tokens (client.chat.completions.create mocked, no network).

The subclasses mentioned in the issue (openai_vlm.py, openai_long_context_model.py) already declare observer as a required positional parameter, so they were not affected and are unchanged.

Verification

Backend venv per AGENTS.md (Python 3.11.9, uv sync --extra data-process --extra test, uv pip install -e "../sdk[dev]"), each file run in its own pytest process as test/run_all_test.py does:

File	Result
test/sdk/core/models/test_openai_llm_observer_default.py	5 passed
test/sdk/core/models/test_openai_llm.py	94 passed
test/sdk/core/models/test_openai_long_context_model.py	22 passed
test/sdk/core/models/test_openai_vlm.py	18 passed
test/sdk/core/models/test_vlm_adapter.py	39 passed
test/sdk/core/models/test_dashscope_vlm_adapter.py	4 passed
test/sdk/core/agents/test_nexent_agent_context_runtime_factory.py	5 passed
test/backend/agents/test_context_budget_v2_pipeline.py	2 passed

With the fix temporarily reverted, the new test file fails 4/5 with the exact TypeError from the issue, confirming it reproduces the bug.

Ruff was not run locally (not available offline in this environment); added lines are within the SDK's 119-character limit and both files compile.

🤖 Generated with Claude Code