Arbiter TUI

Arbiter's interactive terminal interface. One process, one shell window — but inside that window: a multi-pane layout where each pane is an independent conversation with its own agent, history, and streaming output. The TUI is what you get when you run arbiter with no arguments.

A pane is to a conversation what a tab is to a browser. A new pane is a new conversation against an agent of your choosing; multiple panes run side-by-side or stacked, each independently typing, streaming, and waiting for its agent. Background loops (long-running agent processes) live alongside; foreground panes can spawn child panes (/pane <agent> <msg>) whose results land back in the spawner when done.

Start with arbiter. The default layout is a single pane covering the full terminal; Ctrl-w v / Ctrl-w s splits it.

Screen anatomy

Every pane has the same chrome layout. From top to bottom:

┌──────────────────────────────────────────────────────────┐
│ agent · session title                       status/stats │  row 1: identity + status
│ ───────────────────────────────────────────────────────  │  row 2: header separator
│                                                          │
│   streamed model output, tool-call summaries, /cmd       │  scroll region
│   results, system messages …                             │  (scrolls; bounded ring)
│                                                          │
│ ─── [⠋ 3 tool calls…] ──────────────────────────────────  │  mid separator
│ ❯ user input here, multi-line if it wraps                │  input area (1..5 rows)
│ ─────────────────────────────────────────────────────── │  hint separator
│ esc interrupt · pgup/dn scroll · /agents · /help         │  hint row
└──────────────────────────────────────────────────────────┘

What lives where:

• Identity row. Left side: focused agent name and the session title. Right side: live status when the agent is working ("thinking…"), or aggregate token / cost stats when idle.
• Header separator. Plain dashed line; gets an accent colour on the focused pane in multi-pane layouts.
• Scroll region. Where everything the agent emits goes — streamed prose, tool-call summary lines, /cmd output, system notices. Scrollable with PgUp/PgDn (visual-row aware: wrapped lines count as multiple rows for navigation).
• Mid separator. Dashed line above the input. Doubles as the tool-call indicator: while a turn is firing tool calls, this row shows ⠋ N tool calls… instead of plain dashes.
• Input area. 1 row by default, grows up to 5 as readline wraps. Standard editing controls (arrows, history, tab-complete on slash commands).
• Hint row. Static legend of the most-used keys and commands. Hidden in multi-pane layouts (becomes clutter on every pane); the rows are still reserved as blank padding so input row position doesn't shift between modes.

Where to next

• Slash commands — the full /cmd catalogue, grouped by category.
• Keybindings — every key, chord, and modifier the editor recognizes.
• Panes — multi-pane layouts: split, focus, close, /pane spawn semantics.
• Streaming — what you see during a turn: thinking spinner, tool-call indicator, verbose mode, cancellation.
• Sessions — history persistence, session restore on relaunch, per-cwd scoping.

Configuration

Per-user state lives under ~/.arbiter/:

Everything else (current pane layout, scrollback, in-flight turns) is in-memory only and gone when the process exits — except for the session snapshot, which is restored automatically the next time you arbiter from the same directory.