Test agent without message_params to ensure backward compatibility.

An exception mid-batch is caught and reported; results from earlier steps (including images) survive.

Test passing beta feature headers.

Test building tool result with base64 image.

Test handling of text extraction markers in tool results.

Test handling circular references in messages.

Dict-typed fields can't be set via env var; fail at load time rather than letting a raw str through to AttributeError later.

Test combining multiple parameter types.

Test complete state initialization flow.

Test complete flow from user input to message rendering.

Test handling concurrent async operations.

Test behavior when session state is modified during rendering.

Test concurrent calls to setup_state.

Test creating new event loop when existing is closed.

Test creating new event loop when none exists.

Test passing custom headers through message_params.

Test handling deeply nested content structures.

Test performance with deeply nested content structures.

Test message integrity with empty content list.

Test message integrity with missing content.

Test message integrity with missing role.

Test message integrity validation with valid messages.

If any step errors, the batch is is_error=true and the API rejects image blocks in errored tool_results, so prior-step images become text placehol

Test handling error during event loop creation.

Test that exceptions during rendering are propagated.

Test successful tool execution.

Test tool execution with error.

Test extracting text from assistant message.

Test extracting text when there's no text content.

Test that text extraction returns None for non-assistant messages.

Regression: after a force-prune the kept-count must climb again (min, min+1, ...) rather than getting pinned at min_images forever.

Test rendering a complete conversation with various message types.

Test get_or_create_event_loop when another loop is running.

Ctrl+C during hold_key's sleep must still release the keys.

Tool params are built before any screenshot, so __init__ must compute the post-resize size up front rather than defaulting to raw screen size.

Test MessageRenderer initialization.

Test recovery from initialization errors.

Test initialization with empty session state.

Test initialization with None session state.

Within one interval window, growing the history by one image must not change which earlier images were replaced; that is what keeps the prompt

Test that invalid metadata fields are properly rejected by the API.

Test handling invalid sender types.

Test rendering very large conversation history.