fix(openai_agents): run base Agent __post_init__ in DeepEvalAgent - #3323

Summary

DeepEvalAgent now runs the base agents.Agent.__post_init__ normalization when constructed.

Why

DeepEvalAgent.__post_init__ (deepeval/openai_agents/agent.py:35) overrode the SDK's Agent.__post_init__ but only called patch_default_agent_runner_get_model(), never super().__post_init__(). Since DeepEvalAgent is a dataclass, its __post_init__ fully replaces the base one, so the base normalization never ran:

• model_settings kept the dataclass default. The SDK's default model is a GPT-5 model whose default settings carry reasoning=Reasoning(effort="none"), so every DeepEvalAgent used GPT-5 reasoning/verbosity defaults even for non-GPT-5 models such as gpt-4o, for which the base class swaps in generic ModelSettings().
• A dict passed as model_settings was never coerced by the SDK's _coerce_model_settings, leaving a raw dict on the field.
• Base field validation was skipped, so a non-list tools was silently accepted instead of raising TypeError.

Changes

• deepeval/openai_agents/agent.py: call super().__post_init__() before the runner patch.
• tests/test_integrations/test_openai_agents/test_deepeval_agent.py: deterministic regression tests (no API key, no network): model_settings parity with a base Agent for gpt-5 and gpt-4o, dict model_settings coercion, and TypeError on a non-list tools.

Testing

Verified with openai-agents 0.22.3 (the openai-agents CI job installs the latest SDK via pip install -U openai-agents, .github/workflows/test_integrations.yml:219).

With the fix:

pytest tests/test_integrations/test_openai_agents/test_deepeval_agent.py -q
4 passed in 0.03s

Without the one-line fix, the same file fails:

FAILED ...::test_model_settings_match_base_agent[gpt-5]
FAILED ...::test_model_settings_match_base_agent[gpt-4o]
FAILED ...::test_dict_model_settings_are_coerced  # AttributeError: 'dict' object has no attribute 'temperature'
FAILED ...::test_invalid_tools_are_rejected       # Failed: DID NOT RAISE TypeError

Full module suite (pytest tests/test_integrations/test_openai_agents/ -q): 15 failed, 4 passed, 1 skipped; the 15 failures are pre-existing missing-credential failures (identical 15 failed, 1 skipped on the unpatched tree) and CI runs this job only when OPENAI_API_KEY is set.