Summary

• detect empty or whitespace-only OpenAI structured-output content before Pydantic JSON parsing
• raise a ModelProviderError with response diagnostics such as response_id and finish_reason
• add regression coverage for None, empty string, and whitespace-only content

Why

When OpenAI chat completions return an empty structured-output message, browser-use currently surfaces a low-level Pydantic EOF error. That makes agent failures much harder to diagnose after long runs. This keeps the error at the provider boundary and preserves useful completion metadata.

Fixes #4786.

Tests

• uv run pytest tests/ci/models/test_llm_openai_structured_output.py
• uv run ruff format --check browser_use/llm/openai/chat.py tests/ci/models/test_llm_openai_structured_output.py
• uv run ruff check browser_use/llm/openai/chat.py tests/ci/models/test_llm_openai_structured_output.py
• git diff --check

Summary by cubic

Detects and reports empty OpenAI structured-output before JSON parsing, returning a clear provider error with diagnostics instead of a pydantic EOF. Adds tests for None, empty, and whitespace-only messages.

• Bug Fixes
  • Detects None/empty/whitespace message.content and raises ModelProviderError (502) with response_id, finish_reason, and base_url when available.
  • Prevents low-level pydantic EOF errors from leaking past the provider boundary.
  • Adds regression tests covering all empty-content cases.

^{Written for commit f1f456e. Summary will update on new commits.}