When adding/changing tests, optimize for (1) clear intent, (2) stability against refactors, and (3) determinism.

Practical rules:

• Assert observable behavior, not intermediate implementation details. If you need to validate a request payload, mock the SDK client and inspect what was actually sent (not private helpers/step-by-step logic).
• Keep assertions intent-focused and resilient: use targeted field assertions rather than full object/dump equality unless the exact structure is truly contract-stable. If a full expected dict is stable/readable, compare to that dict directly.
• Ensure deterministic behavior in test inputs/expected outputs (avoid unordered set/dict conversions without sorting).
• Use fixtures (e.g., autouse) for repeated setup/teardown patterns like forcing a fresh module import.
• Avoid duplicating existing coverage for the same behavior; extend the existing test where the behavior is already asserted.

Example: readable, stable expectation

result = convert_to_openai_function(MyModel)
assert result == {
    "name": "MyModel",
    "parameters": {"type": "object", "properties": {}, "required": []},
    "strict": True,
}

Example: avoid brittle intermediate-step tests (assert the actual SDK call)

client = MagicMock()
# arrange llm to use mocked client
llm.invoke("hello", prompt_cache_key="k")
called_kwargs = client.create.call_args.kwargs
assert called_kwargs["prompt_cache_key"] == "k"

Example: fix nondeterministic ordering

items = list(set(items))          # flaky
items = sorted(set(items))        # deterministic