Harlock

A pure-Elixir TUI framework for Unix terminals. TEA-style model / update / view loop on top of OTP, with first-class focus traversal, layout constraints, mouse support, ANSI cell-diff rendering, and a small termios NIF for direct /dev/tty control.

defmodule Counter do
  use Harlock.App     # imports the view DSL (box/1, text/2, vbox/1, …)

  def init(_), do: %{n: 0}

  def update({:key, {:char, ?+}, []}, m), do: %{m | n: m.n + 1}
  def update({:key, {:char, ?-}, []}, m), do: %{m | n: max(0, m.n - 1)}
  def update({:key, {:char, ?q}, []}, _), do: :quit
  def update(_, m), do: m

  def view(m) do
    box(
      title: "Counter",
      border: :rounded,
      child: text("count: #{m.n}")
    )
  end
end

Harlock.run(Counter)

Run it with mix run. Not from an IEx prompt: IEx's own terminal driver reads the same tty, and the app would not receive keystrokes. Log output is held back while the app runs and printed after it exits, so a Logger call does not draw over the screen.

A more realistic app wires focus traversal, a selectable table, a scrollable viewport, and a side-effect via Cmd — all together. Tab moves focus between the two boxes; the focused widget owns its keys. Full source: examples/overview.exs.

defmodule Overview do
  use Harlock.App
  alias Harlock.Cmd

  def init(_) do
    %{
      tasks: [
        %{id: 1, name: "compile", state: "done"},
        %{id: 2, name: "test", state: "running"},
        %{id: 3, name: "dialyzer", state: "queued"},
        %{id: 4, name: "credo", state: "queued"},
        %{id: 5, name: "publish", state: "blocked"}
      ],
      selected: 1,
      log: for(i <- 1..40, do: "[#{i}] event line #{i}"),
      log_offset: 0
    }
  end

  def update({:key, {:char, ?q}, []}, _), do: :quit

  def update({:key, {:char, ?r}, []}, m) do
    cmd =
      Cmd.from(fn -> Enum.map(1..3, &"[refresh] new line #{&1}") end)
      |> Cmd.map(fn lines -> {:refreshed, lines} end)

    {m, cmd}
  end

  def update({:refreshed, lines}, m), do: %{m | log: lines ++ m.log}

  # The runtime auto-routes scroll keys to the focused viewport and
  # delivers this message; the app just writes where the offset lives.
  def update({:harlock_scroll, :log, new_offset}, m), do: %{m | log_offset: new_offset}

  # A focused table routes row movement too, so there is no key dispatch here.
  def update({:harlock_select, :tasks, id}, m), do: %{m | selected: id}

  def update(_, m), do: m

  def view(m) do
    vbox(
      constraints: [fill: 1, length: 1],
      children: [
        hbox(
          constraints: [percentage: 40, fill: 1],
          children: [
            box(
              title: "Tasks",
              border: :rounded,
              border_style: [dim: true],
              focus_style: [fg: :cyan, bold: true],
              # :tasks lives on the table so row movement routes to it; the box
              # mirrors its focus for the border.
              focus_proxy: :tasks,
              child:
                table(
                  focusable: :tasks,
                  columns: [
                    column(title: "#", width: {:length, 3}, render: &Integer.to_string(&1.id)),
                    column(title: "name", width: {:fill, 1}, render: & &1.name),
                    column(title: "state", width: {:length, 8}, render: & &1.state)
                  ],
                  rows: m.tasks,
                  row_id: & &1.id,
                  focused_row: m.selected,
                  selection: {:single, m.selected}
                )
            ),
            box(
              title: "Log",
              border: :rounded,
              border_style: [dim: true],
              focus_style: [fg: :cyan, bold: true],
              # :log lives on the viewport, because that is what receives the
              # scroll keys. focus_proxy: lights the box up with it.
              focus_proxy: :log,
              child:
                viewport(
                  focusable: :log,
                  offset: m.log_offset,
                  content_height: length(m.log),
                  child:
                    vbox(
                      constraints: List.duplicate({:length, 1}, length(m.log)),
                      children: Enum.map(m.log, &text/1)
                    )
                )
            )
          ]
        ),
        text("Tab focus  arrows/PgUp/PgDn or wheel scroll  click select  r refresh  q quit",
          style: [dim: true]
        )
      ]
    )
  end
end

Harlock.run(Overview, nil, mouse: true)

Installation

def deps do
  [{:harlock, "~> 0.9"}]
end

Harlock builds a small termios NIF (c_src/termios.c) and an exec helper (c_src/exec_helper.c, installed as priv/harlock_exec, which Cmd.exec uses to run programs with the terminal) — elixir_make handles both automatically. Requires a C compiler and make available at install time. macOS, Linux, and *BSD are supported; Windows native is not (WSL works).

Why Harlock

If you've written a Phoenix LiveView app you already know how to use Harlock — init / update / view, message-passing for events, side-effects as Cmd values. The runtime is a single OTP supervision tree: terminal owner → IO → cmd executor → TEA loop. Any crash shuts the tree down and the terminal owner restores the tty — checked in a real pty for crashes, killed supervisors, and programs started with Cmd.exec.

Compared to alternatives:

• Owl is a styled-output library ("println but pretty"). Harlock is a full interactive runtime — focus, layout, dirty-flag rendering, async cmds, resize handling.
• Ratatouille wraps termbox via a C port. Solid, but the C dep is bigger and the runtime model is its own thing. Harlock is pure Elixir for rendering, with a small in-process NIF only for terminal control — closer to "Elixir all the way down" if that matters to you.
• ratatui-via-port approaches (Rust binary speaking a wire protocol to BEAM) ship as two artifacts: your Elixir release plus a separately-compiled Rust binary that has to be on PATH at runtime. Harlock ships as one mix dependency — its only native pieces build from source with it, and there is no version-skew between BEAM and renderer. The element tree is also ordinary Elixir data, which makes testing and composition easier than a wire-protocol boundary.

Status

Harlock is v0.9. The API is intentionally narrow and stable for the primitives it ships; widgets and ergonomics are still landing. Anything @moduledoc false is internal and free to change.

See ROADMAP.md for the full plan through v1.0.

Examples

./scripts/run.sh counter    # simplest possible app — count up/down, by key or button
./scripts/run.sh sysmon     # live BEAM process monitor: sort, pause, dialogs with buttons, mouse
./scripts/run.sh contacts   # contact manager: search, list, dialog with buttons, mouse
./scripts/run.sh showcase   # tabs, viewport, widgets, form inputs, key and mouse events
./scripts/run.sh notes      # multi-line textarea: wrap toggle, readline editing, undo, click to place
./scripts/run.sh explorer   # tree + select + menu: async-loaded nodes, focus changes, mouse
./scripts/run.sh dashboard  # telemetry + logger subscriptions: sparkline, scrollable log, controls
./scripts/run.sh nodes      # BEAM node explorer: windowed table, lazy supervision tree, mouse
./scripts/run.sh overview   # the README's second snippet: table, viewport, Cmd, mouse
./scripts/run.sh editor     # file browser: open files in $VISUAL/$EDITOR, Ctrl-Z, mouse

The scripts/run.sh wrapper is in the GitHub repo — clone the repo to run the examples. The hex package itself is the library; apps depend on :harlock and build their own runtime entry point (see the Counter snippet above).

Every example runs with the mouse on and has tests under test/examples, which start each one with the options its --run uses and fail if any example compiles with a warning.

contacts exercises most of the core primitives: tab focus traversal, text_input fields, buttons and a checkbox, an overlay with focus_trap, async save via Cmd.from, styled text, a custom theme, a status bar with a current-focus indicator, and the mouse.

showcase is a five-tab tour of the widgets — a clickable tab bar, a 200-row log viewer with viewport + scrollbar and styled levels, a long form that uses scroll-into-view to keep the focused field visible, a widget gallery with animated progress/spinner/statusbar/keybar and a button and checkbox driving them, an event inspector for modified arrows (Ctrl-Up, Alt-Left, etc.) and the raw mouse events an app receives, and an order form with radio groups, a check list, a list box and a textarea.

nodes is a BEAM node explorer — observer for people on SSH — and the largest example: a process list, supervision trees, and memory over time. It shows the two patterns that matter for real data. The process table uses a window function, because enumerating pids is one cheap list while Process.info/2 on all of them is not, so only the rows about to be drawn get hydrated. The supervision tree loads children through a Cmd on expansion, because which_children/1 is a call into another process and the window function runs during rendering. With the mouse, the tabs switch on a click, the wheel scrolls the process list, a click on a process shows its status, current function, links and queue underneath, and a click on a marker expands a supervisor.

dashboard wires two push subscriptions into one screen: Sub.telemetry feeds job durations to a sparkline, Sub.logger turns log calls into update/2 messages, and Sub.interval drives the workload. It emits its own telemetry because a standalone example has nothing else to listen to — but the work runs inside a Cmd, so the handler fires in a different process from the UI exactly as it would for a real query. Point the same subscription at [:ecto, :repo, :query] and only the transform changes, to read Ecto's :total_time — in native time units — instead of :duration. Pausing removes only the interval from subs/1, so the runtime visibly stops one subscription and leaves the others running. A radio group sets the workload's size, Pause and Clear are buttons as well as keys, and the log keeps a scrollable history, newest first, whose view holds still while new lines arrive above it.

editor is a file browser that hands the terminal to your editor. Enter, or the "Open in editor" button, runs $VISUAL, then $EDITOR, then vi on the selected file through Cmd.exec; the app redraws and reloads the preview when the editor exits. Ctrl-Z suspends to the shell through Cmd.suspend, and the mouse selects files, presses the button and scrolls the preview. Without a directory argument it works on a scratch directory of sample files.

explorer puts tree, select and menu in one app. Its deps node starts with nothing loaded: expanding it marks the node in flight, returns a Cmd, and the fetched children arrive as an ordinary message — the pattern any tree over a filesystem or a remote node needs. The filter dropdown opens over the actions and closes when focus leaves it, by Tab or by a click elsewhere, through {:harlock_focus, from, to}. The filtered node list is rebuilt from the model in view/1 rather than handed to the widget, because the model owns what is displayed. The mouse works throughout: node markers, rows, filter choices, actions, and pane borders.

sysmon is a live process table refreshed on a timer. A radio group picks the sort order, unticking Live drops the interval from subs/1 so the rows hold still for reading, and help and quit are overlays with buttons that trap focus while y, n and Esc keep working.

notes is one textarea and one update/2 clause for every edit: typing, word motions and kills, vertical motion that keeps its goal column, and a click that places the cursor all arrive as the same {:harlock_edit, …} message. Undo and redo come from Harlock.UndoStack, held in the model.

Testing your app

Harlock.Test boots an app under a headless backend — no /dev/tty required — and exposes synchronous helpers:

test "Tab cycles focus through the form" do
  h = Harlock.Test.start_app(MyApp, init_arg)

  Harlock.Test.send_key(h, :tab)
  assert Harlock.Test.focused(h) == :email

  Harlock.Test.send_key(h, :tab)
  assert Harlock.Test.focused(h) == :submit

  Harlock.Test.stop(h)
end

Same code path as the real runtime — only the bytes-in / bytes-out boundary is mocked.

Smoke tests

The scripts in priv/*_smoke.exs exercise the real runtime and termios NIF in a pty via script(1): resize, crash restoration, Cmd.exec with typed input and Ctrl-Z inside the program it runs, suspend under a job-control shell, mouse reporting, log output held back while an app runs, and two examples. They run in CI.

./scripts/smoke.sh

Picks the right flag syntax for BSD vs util-linux script automatically.

Contributing

Issues and PRs welcome at https://github.com/thatsme/harlock. The codebase is about 11k lines of Elixir and 1k lines of C. Start with lib/harlock/app/runtime.ex — everything else is reachable from there.

License

MIT. See LICENSE.