5 min readfeature
The footer quiet round: what a terminal owes you at rest
The area below the input showed four rows of chrome before you typed a word. This round made it two — and made every chip earn its row.
- #tui
- #ux
- #verification
The bug report was one sentence: "The area below the user keyboard input in the TUI has too many elements showing." We measured before touching anything. At rest — before you type a word — the footer rendered four stacked rows: a location line, a stats line, a line for every extension chip, and the hints. With the mission panel above the editor, that was eight rows of chrome on a 36-row terminal: 22% of the screen spent before any work happened.
What was actually wrong
The baseline capture (a real session over the live bridge, 128×36) made it concrete. After one trivial turn, the chip row read:
conf ✓ memory: up · stale (v0.0.51) turn 2.7s · ttft 2.7s · ctx 5% · in 26.8k out 2
Four defects, all visible in that one line:
- Every chip painted, always — sixteen
setStatuskeys existed, and the row admitted each one alphabetically, ambient or not. A memory version, a rules count and a dream recency meter answered questions nobody had asked yet. - It truncated mid-chip —
out 2was the cut tail ofout 2.1k (cached 26k). At width, the dump ate its own facts. - The same facts rendered twice — the context % lived on the stats
meter and the turn chip; in/out tokens on the stats chips and the
turn chip;
conf ✓on the mission panel and the footer; the active goal on the panel and its chip. - Nothing could turn it off — the footer's density had no kill switch, which our own UX card says means a feature isn't shipped.
What the field does
We reverse-engineered the below-the-input design of six terminal coding agents from their docs, source and changelogs (every claim sourced). The convergence is striking:
- Claude Code treats the status line as a trade — configuring a custom statusline makes it stop showing most keyboard hints — and budgets hints ("three showings" per session, 10-second auto-hides).
- Codex CLI hides its status row while streaming with the comment "to avoid duplicate [status indicators]" — duplicates are a named defect, not a style choice — and reserves a hint row for messages that expire.
- Gemini CLI made the footer a declarative, ordered item list: items
drop whole with
…at width, the error count is high-priority, and healthy items are simply "not shown when unavailable". - Crush ships the healthiest rule of all: the permission-mode badge is "an empty string in the default coding mode" — healthy states are silent. Transient messages get a 5-second TTL.
- OpenCode renders cost only when
cost > 0, subagent chips only when nonzero, hints only at ≥44 columns — conditional everything. - aider has no status bar at all: state prints when it changes and scrolls away.
NN/g's progressive-disclosure rule frames it: "the very fact that something appears on the initial display tells users that it's important." Painting ambient state by default tells users the wrong thing.
The contract we shipped
Four user journeys (first-time, daily, power, ops) turned that research into a manifest that dispositions every element below the input — persistent, situational, or demoted to a drill-in:
- Two rows at rest. One merged status line —
~/repo (main) · ctx ▕████░░░░░░▏ 42% (auto) · +3 -1 · ↑45k … agnes-3.0-flash— plus the contextual hint line. Location folded into the status row; the meter and the model are the survivors at every width; segments drop whole. - The chip row became an attention row. Failures and blockers paint
first (
✗ 2 to fix · /recovery,memory: down,mcp: 1 ✗,lsp: restarting), then active work (plan 2/5,fleet: 3 running, the turn meter, unseen✦learning events). Priority-ordered, whole chips drop with a dim+Ncounter — never a mid-chip cut again. A healthy engine, an all-connected MCP set, a ready language server: silent./engine,/mcp,/lsp,/statusand the dashboard answer on demand. - One fact, one home. The turn chip is
turn 3.4s · ttft 0.8s · ~$0.0084— the per-turn dynamics plus the session cost (which used to be the stats chip that degraded first at narrow widths). Context % and tokens stay on the status row. Confidence and the active goal live in the mission panel; kill the panel and their chips return. - A real kill switch.
UNOWN_FOOTER_CLASSIC=1restores the exact pre-round four-row footer, the ambient chip lines and the turn meter's full legacy format — one variable, both layers, read live.
Verified like a user, not like a test
The round's verification lane boots
the real TUI against the live stack and asserts on the pane itself — 9/9:
the merged row, the two-row rest, ambient silence with a healthy engine,
the busy hint swap, the turn-chip dedup, the engine-down attention row
(measured live: memory: starting… at ~1s, memory: down at ~8s — the
spawn-kick re-probe cadence), and the classic restore. The unit suites pin
both layouts: the vendored footer 43/43, the self-learning layer 5426/0.
The stills in this post are PTY-lane captures (the repo's
scripts/tui-capture/ pyte lane — real process, real keystrokes, the pane
text shipped beside every image); the tmux/ttyd lane that produces the
landing page's truecolor stills needs host binaries this machine lacks.
Nothing was retouched — re-run, not repaired, is the lane's rule.
The bottom area now answers the only three questions a status bar owes you at rest — where am I, what model, is context fine — in two rows, and spends a third only when something actually needs you.
Spotted something wrong or worth adding? The record is the repo — open an issue.