Distinguish agent final answers, refusals, and pending tool results

Goal

Have the application present only a true final result as the completed answer and handle other states explicitly.

Steps

1. Enumerate run-result and item types, marking final answers, refusals, and pending work. [openai-results]
2. Render a final answer only when result.final_output is not None and the run is complete; approval pauses leave it None. [openai-results]
3. Use result.new_items to distinguish MessageOutputItem, ToolCallItem, ToolCallOutputItem, HandoffCallItem, and approval items. [openai-results]
4. Retain result-type information for diagnosis without exposing internal prompts or tool arguments. [openai-results]
5. Verify parsing and rendering with fixtures for every result type. [openai-results]

How to verify

Cover final answer, refusal, tool call, and error results; only final-answer text should reach the user response.

Troubleshooting

If the UI shows an empty answer, inspect the result type and pending tool items instead of treating every run item as text.