fix(openai): handle assistant messages without content key - #3440

Summary

extract_input_parameters_from_completion raised KeyError: 'content' when a Chat Completions payload contained an assistant message without a content key. Because safe_extract_input_parameters swallows all exceptions, this degraded the whole payload to InputParameters(model="NA"), so the LLM span lost its input, messages, tools and tool descriptions.

Why

An assistant history message that only carries tool_calls may legally omit content: in the openai SDK, ChatCompletionAssistantMessageParam.__required_keys__ is empty and only role is marked Required (deepeval pins openai = "*", so any current SDK version applies).

extract_input_parameters_from_completion read the key unconditionally with message["content"] (deepeval/openai/extractors.py:49). The KeyError propagated to the bare except in safe_extract_input_parameters (deepeval/openai/extractors.py:25-26), which returned the placeholder InputParameters(model="NA"). The sibling render_messages (deepeval/openai/utils.py:118-119) reads the same messages with message.get(...), and stringify_multimodal_content (deepeval/openai/utils.py:61-62) already returns "" for None content, so a .get() here changes nothing downstream.

Changes

• deepeval/openai/extractors.py: read content with message.get("content") instead of message["content"] in extract_input_parameters_from_completion, mirroring render_messages.
• tests/test_core/test_models/test_openai_extractors.py: two regression tests — one against extract_input_parameters_from_completion (asserts model, input and rendered messages are preserved), one against safe_extract_input_parameters (asserts the payload no longer degrades to model="NA").

Testing

Both new tests fail on main:

• test_extract_input_parameters_with_assistant_message_without_content → KeyError: 'content' at deepeval/openai/extractors.py:49
• test_safe_extract_input_parameters_with_assistant_message_without_content → AssertionError: assert 'NA' == 'gpt-4o'

With the fix:

pytest tests/test_core/test_models/test_openai_extractors.py -q
→ 3 passed

pytest tests/test_core/test_models/test_openai_extractors.py tests/test_core/test_models/test_openai_model.py tests/test_core/test_imports.py -q
→ 62 passed

black --check deepeval/openai/extractors.py tests/test_core/test_models/test_openai_extractors.py   # black 25.12.0, as pinned in .github/workflows/black.yml
→ 2 files would be left unchanged.