reasoning

Streamlit UI rendering and aggregation for the reasoning extension.

The Streamlit chat client receives AGENT_REPLY_REASONING_STEP and AGENT_REPLY_TASK_LIST_UPDATE payloads from the WebSocket as the predefined reasoning state runs. Rather than rendering each step as its own chat row (which floods the UI), this module folds them into a single synthetic REASONING_TRACE message in the chat history and renders it as one collapsible expander with a task panel — mirroring the React UI in the BESSER-Agentic-Framework-UI repo.

Two halves:

• append_reasoning_step() and apply_task_list_update() — called from baf.platforms.websocket.streamlit_ui.websocket_callbacks’s on_message. They mutate the most recent in-progress trace in HISTORY (or open a new one) and trigger a Streamlit rerun.

• write_reasoning_trace() and is_empty_trace() — called from baf.platforms.websocket.streamlit_ui.chat’s write_message to render the aggregated trace as a single expander.

The module is self-contained: no public API leaks back to the rest of the streamlit_ui package, so the existing chat/websocket code stays otherwise unchanged.

baf.platforms.websocket.streamlit_ui.reasoning._is_task_only_tool_call(step)

True if a llm_tool_calls event invokes only built-in task tools.

baf.platforms.websocket.streamlit_ui.reasoning._last_in_progress_trace(history)

Return the most recent in-progress MessageType.REASONING_TRACE message in history, or None if every trace is closed (or none exists yet). Walks from the end so the latest open trace is found first.

baf.platforms.websocket.streamlit_ui.reasoning._visible_steps(steps)

Filter out bracket markers, task events, and task-only tool calls.

baf.platforms.websocket.streamlit_ui.reasoning.append_reasoning_step(streamlit_session, step)

Fold a streamed reasoning step into the active trace in HISTORY.

• reasoning_started → append a fresh REASONING_TRACE message with in_progress=True.

• reasoning_finished → append the bracket marker, flip in_progress to False so the UI collapses the expander.

• Any other step → append to the currently-open trace.

Defensive fallback: if no open trace exists when a non-bracket step arrives (history replay, reconnect, etc.) a new trace is created so the step is never silently dropped.

baf.platforms.websocket.streamlit_ui.reasoning.apply_task_list_update(streamlit_session, tasks)

Replace the task snapshot of the active trace.

Each task_list_update payload carries the full current snapshot, so the UI just swaps the trace’s tasks list rather than diffing. Defensive: opens a fresh trace if none is currently in progress.

baf.platforms.websocket.streamlit_ui.reasoning.is_empty_trace(content)

True when a reasoning trace has no observable steps and no tasks.

Empty traces typically result from a pure LLM-only reply that did not invoke any tools or plan any tasks — rendering them would just produce an empty chat bubble.

baf.platforms.websocket.streamlit_ui.reasoning.write_reasoning_trace(content)

Render an aggregated reasoning trace as a single collapsible expander.

• While in_progress is True the expander is open by default so the user can watch the agent think.

• Once the loop finishes (in_progress flipped to False) the expander defaults to collapsed; the user can re-open it for inspection.

• Steps are grouped by their step number so each loop iteration shows a single STEP N header. Task-related events are omitted from the step list — the task panel below is the source of truth for task progress.

• The task panel renders the live snapshot with colored status boxes.