Add unit tests for OpenAIChatAgent

[vaibhavhariram]

Summary

• 51 tests across 9 test classes in hud/agents/tests/test_openai_chat.py
• Covers init (api_key, base_url, openai_client, completion_kwargs, checkpoint injection), schema sanitization (_sanitize_schema_for_openai: anyOf nullables, prefixItems→array, nested recursion, key allowlist), tool schema formatting to OpenAI function format, _oai_to_mcp conversion, system message formatting, format_blocks (text/image/mixed), format_tool_results (text/image/empty/structured content), error handling (API errors, JSON truncation), and get_response/_invoke_chat_completion params passthrough

Test plan

• uv run pytest hud/agents/tests/test_openai_chat.py -q → 51 passed, 0 failed
• All async tests use pytest-asyncio (auto mode via asyncio_mode = "auto" in pyproject.toml)
• No real network calls — OpenAI client fully mocked

🤖 Generated with Claude Code

---

Note

Low Risk
Low risk: adds a new test module only, with no production code changes. Main risk is tightening behavior expectations that could require follow-up fixes if current behavior differs.

Overview
Adds hud/agents/tests/test_openai_chat.py, a comprehensive unit test suite for OpenAIChatAgent.

The tests cover initialization paths (HUD gateway vs. explicit api_key/base_url/openai_client, completion_kwargs, and checkpoint injection), schema sanitization and tool schema conversion to OpenAI function format, plus message/result formatting (get_system_messages, format_blocks, format_tool_results) including image handling.

Also adds coverage for get_response / _invoke_chat_completion behavior: tool call parsing, passthrough of allowed completion kwargs (while protecting model/messages/tools), finish-reason handling, and graceful error responses for API failures/truncated JSON.

^{Reviewed by Cursor Bugbot for commit 4ec040c. Bugbot is set up for automated code reviews on this repo. Configure here.}

[Ruthwik-Data]

Solid coverage — 51 tests across 9 classes with fully mocked OpenAI client is the right approach for a unit test suite at this layer.

A couple of observations:

reasoning_content handling not tested. The mocks set reasoning_content = None throughout, but the actual OpenAIChatAgent likely has a code path that handles non-None reasoning_content (e.g., o1/o3 models). Worth adding at least one test where reasoning_content is a non-empty string to ensure it doesn't get silently dropped or cause a formatting error downstream.

format_tool_results with mixed content types. The tests cover text/image/empty/structured content individually, but no test exercises a ToolResult with a mixed list (e.g., [TextContent, ImageContent] in the same result). In eval pipelines that return evidence alongside screenshots, this combination is common and worth a dedicated case.

Minor: The make_tool_call / make_tool_result factories are clean and reusable — if this project grows the test suite further, it may be worth moving them to a shared conftest.py so other agent tests can use them too.

Overall the test plan is well-structured and the async handling via pytest-asyncio in auto mode is the right call. Happy to see this merged once the above gaps are addressed or consciously scoped out.