Build your own frontend

Tau’s Textual app is one frontend, not the architecture. A custom UI plugs into the same primitives the built-in TUI uses:

CodingSession   — owns the coding-agent environment
AgentEvent      — describes assistant text, tool calls, results, errors
Frontend state  — belongs to your UI

The reusable tau_agent package stays independent of terminal frameworks, widgets, keybindings, config paths, and slash-command UX. Build against tau_coding.session.CodingSession, not Textual widgets.

CodingSession provides the environment (provider/model, tools, persistence, skills, prompt templates, project context, slash-command handling, compaction). Your frontend provides the interface (prompt input, transcript rendering, command entry, cancellation, pickers).

Minimal event loop

async for event in session.prompt(user_text):
    render_event(event)

The stream yields provider-neutral AgentEvent values from tau_agent.events (see the agent loop for the list). Render from these, never from provider-specific chunks. Treat AgentStartEvent/AgentEndEvent and non-recoverable ErrorEvent as the source of truth for “is the agent working?”.

Steering and follow-ups

If the user submits while a run is active, queue instead of starting a second run:

async for event in session.prompt(user_text, streaming_behavior="steer"):
    adapter.apply(event); redraw(state)

Use streaming_behavior="follow_up" for a prompt that waits until the run would otherwise stop. Overlapping session.prompt(...) calls without streaming_behavior are rejected so two runs can’t mutate one transcript. QueueUpdateEvent carries pending queued text for badges/status.

Slash commands

Slash commands belong to tau_coding. Before treating input as a prompt:

result = session.handle_command(text)

If result.handled, apply the requested effect (exit_requested, clear_requested, new_session_requested, compact_summary, message) and show reference/status output outside the durable conversation. If result.compact_summary is not None, call await session.compact(result.compact_summary) (an empty string means “use the built-in prompt as-is”).

/skill:<name> is intentionally not a command — pass it through to session.prompt(...), which expands it before the run.

Restoring and switching sessions

Initialize the visible transcript from session.messages (the built-in TuiState.load_messages() is a reference). ToolResultMessage preserves structured metadata (e.g. edit patches), so you can render restored tool results without reading JSONL directly.

For session switching, use tau_coding.session_manager.SessionManager — list_sessions(session.cwd), then await session.resume(session_id) (or load a fresh CodingSession with storage=jsonl_session_storage(record.path)), then rebuild the transcript from session.messages.

Cancellation, pickers, keybindings

Cancel with session.cancel() — keep consuming events until the stream ends.
Read picker data directly from the session: command_registry.list_commands(), skills, prompt_templates, available_model_choices, available_models, available_providers, thinking_level, available_thinking_levels, session_manager. For model changes from another provider, call set_provider(...) then set_model(...).
Keybindings and themes are frontend policy. The built-in app reads ~/.tau/tui.json via tau_coding.tui.load_tui_settings(), but your UI can ignore it.

What not to depend on

Avoid coupling to private CodingSession attributes, provider-specific response chunks, Textual internals, or the raw JSONL structure (use SessionManager / CodingSession). Stick to the event, message, tool, harness, and session primitives.