deer-flow/backend/docs/TUI.md

DeerFlow Terminal Workbench (TUI)

deerflow is a terminal-native workbench for the DeerFlow harness. It runs embedded over DeerFlowClient — no Gateway, frontend, nginx, or Docker services required — while honoring the same config.yaml, checkpointer, skills, memory, MCP, and sandbox settings as the rest of DeerFlow.

Install & run

The TUI ships as an optional extra so the core harness install stays lean:

uv pip install 'deerflow-harness[tui]'    # or: pip install textual

Launch modes:

Command	Behavior
deerflow	Launch the TUI when stdin/stdout are TTYs
deerflow --tui	Force the TUI (clear diagnostic if textual is missing)
deerflow --tui-transparent	Use the terminal's default background when launching the TUI
deerflow --cli	Force headless/classic mode for one invocation
deerflow chat	Same TUI conversation surface
deerflow --continue	Resume the most recent thread
deerflow --resume THREAD	Resume a thread by id
deerflow --print "question"	Headless one-shot answer to stdout; on failure a concise stderr line and exit 1
deerflow --json "question"	Headless newline-delimited StreamEvents; on failure one terminal {"type": "error"} record and exit 1
deerflow --recursion-limit 250 --print "question"	Set the headless agent-loop super-step limit
echo "q" | deerflow --print	Read the message from stdin
DEER_FLOW_TUI=1 deerflow	Force the TUI via environment
DEER_FLOW_TUI_TRANSPARENT=1 deerflow	Persist terminal-background rendering via environment

--print and --json keep every word of an unquoted message, and words before a bare flag (deerflow summarize this repo --print) are the message, as with --cli. A message given on the command line wins over piped stdin. Words on both sides of the flag are rejected with exit 2, since their order cannot be recovered. --print and --json cannot be combined; pick one output mode.

If no TTY is available and no headless flag is given, deerflow prints guidance instead of hanging.

Provider failures (for example an expired credential) usually do not raise: the LLM error middleware turns them into a final AI message flagged deerflow_error_fallback. Headless runs treat that as a failure too — --print still writes the fallback text to stdout, then prints Error: LLM request failed (error_type=…, error_reason=…) to stderr; --json appends the same terminal {"type": "error"} record. Both exit 1.

Transparent rendering is opt-in; the solid DeerFlow palette remains the default. The transparent mode uses Textual's ansi_default background for the main screen, header, transcript, status, palette, composer, and modal surfaces while keeping truecolor foregrounds and selection highlights. Combine --tui-transparent with --tui when the UI also needs to be forced without a detected TTY.

Headless runs use a recursion limit of 100 by default. Pass a positive --recursion-limit when a longer agent loop is expected. This is a LangGraph super-step budget, so it can include model and tool-execution steps rather than mapping one-to-one to conversational turns. The max_recursion_limit setting is a Gateway safety ceiling for client-supplied values; it is not the default for trusted embedded CLI runs.

Surface

• Header — model, thread, project root, skill/tool counts.
• Transcript — user prompts, assistant answers, and compact tool cards (⚙ Read path ✓) with dimmed result previews. Finalized assistant messages render as Markdown (headings, bold, lists, code, links); the actively-streaming message stays plain text to avoid reflow jumpiness and snaps to Markdown when it completes. Transcript re-renders are coalesced (~16 fps) so streaming stays smooth on long threads.
• Status line — run state + animated spinner, model, thread title, token usage, and an esc interrupt hint while a run is active.
• Composer — rounded multiline input box. / opens the command palette. Pasted code, stack traces, and multi-paragraph prompts retain their line breaks and indentation; Enter sends the complete document.

Keys

Key	Action
Enter	Send message / accept palette selection
/	Open the slash-command palette
↑ / ↓	Palette navigation; otherwise move inside multiline input, then use history at the first/last row (wrapped rows included)
PageUp / PageDown	Scroll the transcript without moving focus from the composer
Tab	Complete the highlighted command (adds a trailing space)
Esc	Close the palette / overlay
Ctrl+C	Interrupt the active run, or quit when idle (again to skip waiting for an interrupted step)
Ctrl+L	Redraw · Ctrl+U clear composer

At the last input row, ↓ preserves the current draft unless you are browsing history. After ↑ recalls history, ↓ moves forward and restores the saved draft when it passes the newest entry.

At the first input row, ↑ leaves the draft, cursor, and undo history untouched when no input history is available. Recalling an entry or saved draft identical to the current input also preserves the cursor and undo history.

Transcript updates follow new output while the view is at the bottom. After you scroll upward, streaming refreshes preserve the reading position until you return to the bottom with PageDown.

Slash commands

/help /new /clear /goal /resume /threads (/switch) /model /skills /tools /mcp /memory /uploads /usage /config /quit, plus /<skill-name> task to activate any enabled skill for the current turn (same semantics as elsewhere in DeerFlow). /model and /threads open modal pickers, which show thread titles and model names exactly as written, square brackets included.

/clear removes the current transcript rows from the terminal display only; it keeps the active thread and persisted conversation intact. During an active run, /new and /clear ask you to wait for the run to finish instead of resetting in-flight display state.

/resume <id-or-title> switches conversations, while /resume without an argument opens the thread picker. During an active run, /resume, /threads, and /switch ask you to wait instead of switching away from in-flight output. An invalid /resume reference displays an error and leaves the current conversation intact. After an interrupt, late stream actions from the previous run are discarded when they reach the UI, even if the next prompt uses the same conversation. Its final usage, title, and completion event cannot replace those of the new run. An interrupted run cannot be stopped mid-step, so a tool call already in progress (a long shell command, for example) finishes first. Until it does, sending to the same conversation shows a notice instead of starting a second run there — two runs on one thread would overwrite each other's history. Send again once it finishes, or use /new or /resume to continue in another conversation meanwhile. Quitting (Ctrl+C or /quit) also waits for that step, or for a finished run still saving its title, with a notice; the TUI closes as soon as it is done. Press Ctrl+C again to quit at once; the unfinished step is then not saved. If a run cannot start, the TUI reports an error and returns to idle so you can retry.

Use /goal <condition> to set the active thread goal, /goal to show it, and /goal clear to clear it. /goal from-chat is an IM entry point for now; the TUI reports that it is not supported and points to /goal <condition>.

Architecture

The TUI is a UI shell over the existing embedded harness — it does not fork agent behavior.

cli.py          launch-mode planning (pure) + headless print/json + entry point
session.py      builds DeerFlowClient (+ checkpointer) and the persistence writer
runtime.py      StreamEvent  ->  reducer actions  (pure translate + threaded driver)
view_state.py   ViewState + reduce(state, action)  (pure, the testable heart)
message_format  compact tool summaries / truncation (pure)
command_registry slash-command registry + resolve (pure)
input_history   bounded ↑/↓ history (pure)
render.py       Rich renderers for header / transcript / status / palette (pure)
theme.py        palette + symbols
app.py          Textual App: composes widgets, drives runs on a worker thread,
                marshals actions back to the UI thread, renders ViewState
widgets/        Textual widgets (the multiline composer)
persistence.py  writes threads_meta so sessions appear in the Web UI (below)

DeerFlowClient.stream() is a synchronous generator, so the app runs it on a Textual worker thread and marshals each yielded action back to the UI thread via call_from_thread. The pure layers (everything except app.py and widgets/) have no Textual dependency and are unit-tested directly with synthetic StreamEvents.

Web UI visibility (shared persistence)

The Web UI lists conversations from the threads_meta SQL table (filtered by user_id), not from the checkpointer. An embedded run only writes the checkpointer, so a TUI thread would otherwise be invisible in the sidebar.

persistence.py closes that gap: on the first turn of a thread it writes a threads_meta row — owned by the local default user ("default") — into the same database the Gateway reads, and syncs the generated title afterward. This requires only the shared threads_meta store (built via deerflow.persistence.engine.init_engine_from_config), not the Gateway process. When the database backend is memory (no SQL store) the writer degrades to a silent no-op and the TUI still works.

All DB work runs on one long-lived background event loop, because a SQLAlchemy async engine is bound to the loop that created it.

Tests

Pure layers are TDD'd in backend/tests/test_tui_*.py; the Textual app, slash palette, and modal overlays are exercised through Textual's pilot harness with a fake in-process session (no live model). test_tui_persistence.py proves the threads_meta write/read round-trip.

cd backend && PYTHONPATH=. uv run pytest tests/ -k tui -q