tau-tui

The Textual terminal interface for tau: replies stream in as markdown, tool calls are cards with their tier, approval and result, agent events show in a live pane, tier-2 calls ask in a modal dialog, and the brand's τ mark opens every new session. One row of chrome at the top, at most four at the bottom, so 80x24 over mosh on a phone stays usable.

It is its own distribution (tau-tui, import tau_tui), uses only the public API of tau-core and plugs into the tau command through the tau.commands entry point (tui):

pip install tau-tui          # or, in this repo: uv sync --all-packages --group dev
tau tui                      # new session
tau tui --resume             # continue the most recent earlier session
tau tui --resume ID          # continue a given session
tau --home /path/to/tau --lang tr tui --profile travel

Configuration comes from tau.toml and .env like every other channel (TAU_HOME, TAU_DATA_DIR, TAU_PROFILE, TAU_LANG). A broken config or an unknown session end the command with a plain English message and exit code 2 before the interface opens. With no tau.toml anywhere a setup screen creates one first (home directory, provider, UI language; the same code as tau init); with a config whose role has no credentials the interface opens on its settings screen and the chat starts once a key is saved. /settings or F10 opens that screen at any time (see Settings). The chrome (hints, labels, notices, the approval dialog) follows [ui] language (en or tr); /lang switches it for one run.

Layout

Four bands by width (-phone < 60 cols, -narrow 60–99, -wide 100–139, -ultra ≥ 140) and three by height (-short < 20 rows, -mid, -tall ≥ 30). The header is one row, the composer a rounded box of one to five text rows, the hint bar one row.

80x24 (-narrow), idle after a two-tool turn:

 τ tau  home · 20260929-120000-abcd                                  brain · en
 › Salon lambası açık mı? Açıksa kapat.
 ● get_light_state(room="salon")                                  tier 0 · 38ms
   ⎿  {"on": true, "brightness": 80}
 ● set_light(room="salon", on=false)                  tier 2 · approved · 412ms
   ⎿  Light turned off
 τ Salon lambası kapatıldı.
   • Önceki durum: açık, %80 parlaklık
 ✓ Worked for 4s · 2 tool calls · ctrl+o expand
╭──────────────────────────────────────────────────────────────────────────────╮
│ ❯ Type a message…                                                            │
╰──────────────────────────────────────────────────────────────────────────────╯
 ? shortcuts · / commands · ctrl+j newline                          ctrl+q quit

100x30 (-wide): the events pane docks on the right (34 cols; 40 at -ultra), ctrl+e hides it. 160x45 (-ultra): a read-only session sidebar (28 cols) on the left; clicking a row resumes.

 τ tau  home · 20260929-120000-abcd                          brain · 7 tools · en
 › Salon lambası açık mı? Açıksa kapat.          │ Events · 5
 ● get_light_state(room="salon")   tier 0 · 38ms │ 12:00:01 ● get_light_state
   ⎿  {"on": true, "brightness": 80}             │          tier 0 · ok · 38ms
 τ Salon lambası kapatıldı.                      │ 12:00:02 ◆ approval set_light
                                                 │          tier 2 → approved (tui)

80x30 (-narrow -tall): an 8-row events strip under the transcript. Below 100 cols ctrl+e / F4 swaps the transcript for the events log in the same body; the composer stays live and Esc returns to the chat.

50x24 (-phone, Termius over mosh): the header drops the session id (/session shows it), tool badges move to the result line, the composer has no border, hints say ^j, the welcome card shows the 16-column mark, replies stream as plain text and settle into markdown when the segment ends, and the approval dialog is full width with stacked buttons.

 τ tau · home                                  en
 › Salon lambası açık mı? Açıksa kapat.
 ○ set_light(room="salon", on=false)
   ⎿  tier 2 · running
 ✻ Running set_light…
 ❯
 esc cancel · ^j newline                     busy

-short (< 20 rows): no hint row, the welcome mark is hidden, the composer grows to two rows.

Vocabulary

Every glyph is one cell wide and none is an emoji-presentation code point, so phone terminals draw them at the right width: › you · τ tau · ● finished tool call · ○ running or waiting for approval · ⎿ result line · ✻ busy verb / reasoning · ✓ ◆ ✗ turn footer (finished / limit, cancelled, interrupted / error) · ◆ needs you · ▸ notice · ❯ prompt.

State colours (deliberate; no green and no red in the palette, shapes double the colours): ok / success = Flare, error = Signal, rejected = Ember, running / pending = Core at 55 %, attention = Flare, notice = Ember. Rich renderables (the events pane) cannot take a TCSS alpha, so they use the literal blends theme.TEXT_MUTED (Core over Void at 60 %, the live clock and the running dot) and theme.TEXT_DISABLED (38 %, replayed clocks). The header τ breathes while the welcome card is on screen, runs the comet while a turn is busy and is static Ember otherwise; nothing else in the UI moves. The composer border is Ember, Signal while busy and Flare while an approval waits.

Keys and commands

Typing / into an empty composer opens a fuzzy menu of the slash commands (↑↓ pick, Tab complete, Enter run, Esc close). An exact name ranks first, and a command typed in full runs as typed even when the menu has not caught up with a burst of keys (a paste, mosh), so /session never runs /sessions. The commands are the shared spec of tau_core.commands, with the same wording as tau chat:

The Turkish aliases (/devam, /çık, /özet, /temizle, /durum, /rol, /ayarlar, …) keep working but are never listed; the menu shows the thirteen canonical names and every row is on screen at 80x24 (in the -short band it scrolls). /new, /resume, /compact, /clear, /model <role> and /settings wait for the current turn (a toast says so); /status and /model without an argument run any time.

Settings

SettingsScreen (/settings, /ayarlar, F10) edits tau.toml and .env without leaving the TUI, through tau_core.settings. It is a keyboard-only modal in the family of the approval dialog: a list of sections on the left, the section's form on the right, a status line and a hint row; the header stays visible above it. ↑↓ picks a section, Enter or Tab moves into its fields, shift+tab comes back, ctrl+s or F2 (the mosh-safe mirror; the app's own F2 picker is switched off while the screen is up) saves, Esc closes without writing.

 τ tau  home · 20260929-120000-abcd                                  brain · en
 ╭────────────────────────────────────────────────────────────────────────────╮
 │ ◆ Settings                                                                 │
 │                                                                            │
 │ Model roles        brain                                                   │
 │ Credentials        provider     fake                                   ▼   │
 │ Interface          model id     fake-brain                                 │
 │ Agent              max tokens                                              │
 │ Doctor             fallback     (none)                                 ▼   │
 │ Files              fallback id                                             │
 │                    ✓ no credentials needed                                 │
 │ ↑↓ section · tab fields · ctrl+s / F2 save · esc close                     │
 ╰────────────────────────────────────────────────────────────────────────────╯

• Model roles: per role a provider Select over every installed provider (a planned one such as ollama is listed in a note and cannot be chosen; a role that already names it keeps it, marked (planned)), the model id, max tokens, an optional fallback provider and model id, and the role's credential status (✓ credentials ok, ✗ missing AWS_REGION, AWS_PROFILE, ◆ missing … · fallback anthropic in use, ✓ no credentials needed).
• Credentials: the variables of the configured providers first (needed by brain, fast), then every other provider's, then every name in .env.example and .env. Each row shows set · sk-…abcd or not set (plus (from the environment; .env does not override it) when the shell sets the variable), an Input for a new value — a password one for keys, tokens, secrets and passwords, a plain one for AWS_REGION-like settings; the typed value is stripped — and a clear button (press again to keep). The names a start without credentials was missing are marked ◆ needed and the first of them has the focus.
• Interface: the UI language ([ui] language, applied live on save) and the profile (TAU_PROFILE in .env).
• Agent: context, window_size, max_turns, max_total_tokens (empty = none).
• Doctor: the tau doctor checks, run in a thread when the section is first shown.
• Files: the paths of tau.toml, .env, .env.example, persona/persona.md and persona/user.md with a mark for the ones that exist (build_app(show_paths=False) prints <home>/… instead, which the snapshot test uses).

A save first validates (a bad value stays in the dialog with the reason in the status line and nothing is written), then writes tau.toml through tomlkit (comments kept) and .env through python-dotenv (mode 0600, the changed names also exported to the process), then validates the saved files and, when a role, an agent option, a credential or the profile changed, rebuilds the agent on the same session through the app's factory, like /model (Settings saved; model reloaded., else Settings saved.). A language-only change relabels the chrome and rebuilds nothing. A credential value lives only in its Input until the save: it never reaches a Static, a notice, the text oracle, a screenshot or the log.

Two start-up paths use the same family of screens. tau tui without any tau.toml runs SetupScreen (run_setup): the home directory (~/.tau, editable), the provider for the brain and fast roles (detected from the environment and <home>/.env like tau init, with fake for a credential-free try) and the UI language; Create (Enter, ctrl+s, F2) runs tau_core.builder.init_home and the command continues into the normal app on that home, Esc leaves with exit code 2. A config whose role has no credentials opens the app without an agent: the header shows — for the session, the composer reads Add credentials in /settings and takes only /settings, /help, /lang and /quit, the settings screen opens on Credentials with the missing names marked, and the first agent is built as soon as a save gives the provider its key. A role that runs on its fallback provider is announced with one notice (brain: using the fallback provider anthropic (bedrock has no credentials). Add them in /settings.) instead of the terminal warning tau chat prints.

Approvals

TuiApprover (name = "tui") is the channel's Approver. A tier-2 call opens ApprovalScreen: tool, effect, the arguments as pretty JSON and the buttons Approve / Reject.

• y or e approves, n, h or Esc rejects; focus starts on Reject, so Enter rejects, and Tab only moves between the two buttons (PageUp/PageDown scroll the arguments). The keys work in both languages whatever the UI setting.
• y/e and Approve only work 0.5 s after the dialog opens (the hint reads (arming…) meanwhile), so keys typed for the chat cannot approve by accident.
• Anything other than an explicit approval rejects: a dialog closed another way, an approver with no running app, or a cancelled wait (the dialog is taken down so it cannot approve later).
• Rejecting does not cancel the turn: the model reads the refusal and answers.

The dialog is translucent: the chat stays visible behind it. Behind it the tool card reads ○ set_light(…) + ⎿ ◆ waiting for your approval, the composer and the header τ turn Flare and the hint bar says what to press. Nothing in this package approves on its own.

Sessions and events

/sessions lists the ten most recent earlier sessions with messages (numbered for /resume n, same order as tau chat); /resume alone resumes the most recent one. Resuming rebuilds the agent on the stored session (the previous one is closed), renders the history from the stored messages (user rows, markdown replies, tool cards with the recorded duration, approval and result), replays the stored agent events into the events pane before live ones, and keeps the current agent when the rebuild fails. The last 40 turns stay live; older ones fold into one line and PageUp at the top pages them in. One EventBus serves the app for its lifetime; events stamped with another session id are ignored.

/compact [focus] asks the agent's own model for a continuation summary, opens a new session whose history starts with that summary, links the two sessions in their metadata (compacted_from / compacted_to) and keeps the old one untouched; the transcript is reset like /new, then shows the summary as an info card and the notice Compacted N messages into a summary. New session <id>; the previous session <old id> is kept., the header shows the new id and the events pane starts empty. The summary is written in a worker, so the app stays responsive: the chrome is busy (✻ Summarizing…, esc cancel in the hint bar), read-only commands such as /status still answer, turns and session switches are refused, and Esc cancels the summarizer (Compaction cancelled; the session continues.). A failure (nothing to summarize, a model error, missing credentials) is a notice and the current session goes on. The details are in tau-core's README under "Compaction".

/model fast rebuilds the agent on the same session with that role (the transcript stays, the header and the welcome card follow) and remembers the role in the session's metadata, so /resume and tau tui --resume come back on it; build_app(role=...) pins a role for the whole run instead. An unknown role, missing credentials or an unavailable provider leave the current agent working and say why.

Knobs

Live markdown streaming needs ≥ 60 cols; two consecutive slow appends (a long open code fence) switch the segment to plain text until it ends. Replies are coalesced to at most one update every 80 ms.

Logs

While the interface runs the terminal belongs to Textual: details go to <TAU_DATA_DIR>/logs/tau.log (for example ~/.local/share/tau/logs/tau.log), the stderr handler of the tau command is detached for the duration, and the chat shows a short notice instead of a traceback.

Python API

from tau_tui import TauApp, build_app, resolve_resume, SessionNotFound
from tau_tui import TuiApprover, ApprovalScreen, Transcript, EventLog, rasterise
from tau_tui import SettingsScreen, SetupScreen, SetupResult, run_setup

app = build_app(config, session_store=store, session_id=None)  # builds the agent (channel "tui")
app.run()

build_app creates one EventBus and one TuiApprover, builds the first agent with tau_core.build_agent(..., channel="tui") and gives TauApp a factory agent_factory(session_id, *, role=None, seed_messages=None, config=None) that builds every later agent (/new, /resume, /compact, /model, a settings save) on the same bus, store and approver; a config handed to it becomes the config of every later build. role=None means brain, or the role a resumed session_id remembers from /model; an explicit role is pinned for the run. When the first agent cannot be built for lack of credentials, build_app returns the app without an agent (app.agent is None, app.missing_credentials names the variables) instead of raising. Tests pass model=FakeModel([...]), tools=[...], a fixed clock, tz=UTC, motion=False, show_durations=False, show_paths=False and a pinned version. run_setup(default_home=..., language=...) runs the setup app and returns a SetupResult (home, provider, language, created) or None. resolve_resume(store, resume, language=...) turns --resume [ID] into a session id and an optional notice (SessionNotFound for an unknown id). rasterise(cols) is the τ mark as half-block cells (docs/brand/README.md, "Terminal").

Development

uv run pytest packages/tau-tui -q                      # pilot + snapshot tests, offline
uv run pytest packages/tau-tui --snapshot-update       # after an intended visual change

Snapshots live under tests/__snapshots__/test_snapshots/ as .raw files: pytest-textual-snapshot 1.0 with syrupy 6 writes SVG under that extension, and the extension is pinned in test_snapshots.py (exactly one file per snapshot test, no glob). There are fourteen: welcome at 80x24, 50x24 and in Turkish, chat (a markdown list, a one-line fence and a rejected tier-2 card), busy, approval, menu (all thirteen rows) and picker at 80x24, wide 100x30 (busy chrome), ultra 160x45 (a fence and the last turn's cards expanded), the events view at 50x24, short 80x18, the settings screen on Model roles at 80x24 (show_paths=False) and the setup screen at 80x24. Every snapshot uses the fixed session id 20260929-120000-abcd (checked in every file except the setup screen, which has no session), a fixed clock, UTC event times, no motion and no durations, and is taken after the streams stopped; review a regenerated file before committing it (it must never contain a local path, which the test also checks). The pilot tests read the widgets through text oracles (Transcript.text, EventLog.text, HintBar.text, HeaderBar.text, ToolCard.text) so behaviour changes do not churn the snapshots. The mark fixtures under tests/fixtures/ are the text previews of rasterise(24) and rasterise(16); regenerate them together with the pinned SHA-256 in test_brand.py when the geometry in docs/brand/render.py changes.

License: MIT.