@nerisma/pi-input-revamp

Replaces pi's input editor with a full rounded frame, a colored π prompt character, and a session metrics bar built into the border.

╭─ agent · anthropic/claude-sonnet-4-5 · high ──── 0.015$ · 15.2K/200K · 12.3% ─╮
│ π hello world                                                                 │
╰──────────────────────────────────────── T5 (3) · 12s · 0.004$ · OUT 8.3K ─────╯

The border and the π use the active theme's accent color. Each of the four corners of the border is a quadrant you fill with the info elements you want — agent, model, cost, context usage, per-turn metrics, and more. By default the top bar shows session-wide info and the bottom-right shows the last turn, but the whole layout is configurable (see Configuration).

Installation

pi install npm:@nerisma/pi-input-revamp

Or via settings.json:

{
  "packages": ["npm:@nerisma/pi-input-revamp"]
}

Configuration

The layout and animations are driven by ~/.pi/pi-input-revamp.json. The file is created with the defaults on first run, so you can just edit it. Missing fields fall back to the defaults, and changes are picked up on the next pi restart.

{
  "layout": {
    "topLeft": ["agent", "model", "thinking-level", "cwd", "duration", "tools", "tok"],
    "topRight": ["session-label", "ctx-percent", "ctx-tokens-full", "session-cost", "session-out", "session-hit", "session-miss"],
    "bottomLeft": [],
    "bottomRight": ["turn", "turn-duration", "turn-cost", "turn-out", "turn-hit", "turn-miss"]
  },
  "animations": {
    "typingPulse": true,
    "submitFlash": true,
    "metricPulse": true,
    "tokPulse": true
  }
}

Layout

Each quadrant — topLeft, topRight, bottomLeft, bottomRight — is an ordered list of element IDs joined with ·. Elements that have no data (e.g. a model that isn't set, or a turn that hasn't completed) are skipped silently, so empty quadrants simply collapse into the border.

Element	Shows
agent	Active agent name (PI_ACTIVE_AGENT)
model	provider/id of the current model
thinking-level	Thinking level (hidden when off)
cwd	Working directory ($HOME collapsed to ~)
duration	Elapsed session time (5m 12s)
tools	Number of tools actually sent on the last request (12 tools)
tok	Live token estimate of the text you're typing (~42 tok)
session-label	The literal SESSION tag
ctx-percent	Context window usage as a percentage (12.3%)
ctx-tokens	Context tokens in use (15.2K)
ctx-tokens-max	Context window size (200K)
ctx-tokens-full	Used / max (15.2K/200K)
session-cost	Whole-session cost in $
session-out	Whole-session output tokens (OUT 8.3K)
session-hit	Whole-session cache-read tokens (HIT 2.1K)
session-miss	Whole-session input / cache-miss tokens (MISS 1.2K)
turn-cost	Last completed turn cost in $
turn-out	Last completed turn output tokens (OUT 8.3K)
turn-hit	Last completed turn cache-read tokens (HIT 2.1K)
turn-miss	Last completed turn input / cache-miss tokens (MISS 1.2K)
turn	Turn number with metric-update count (T5 (3))
turn-duration	Duration of the last completed turn (12s)

Every element is self-contained: the session-* variants always report whole-session totals and the turn-* variants always report the last completed turn, regardless of which quadrant you place them in.

Showing another extension in the bar

Extensions can publish a status line through pi's ctx.ui.setStatus(key, text) API. Normally that text is rendered by pi's footer — which this extension hides — so those statuses would be invisible. To surface one, add a slot of the form ext:<statusKey> to any quadrant:

{
  "layout": {
    "bottomRight": ["turn", "ext:pi-quotas-usage", "turn-cost"]
  }
}

• <statusKey> is the exact key the other extension passes to setStatus() — it must match character for character. Check that extension's docs to find it.
• Placement and ordering work like any other element: the quadrant decides where it shows, and it's joined with the same · separator.
• The published text is shown as-is (extensions usually colour their own status), with no prefix added.
• When that extension currently has no status set, the slot shows the key name in the warning colour as a placeholder, so you can tell the slot is wired up and waiting for data rather than silently empty.

You can wire up as many as you like, in any corner — nothing needs to be declared beyond the ext:<statusKey> entry itself.

Animations

Each effect can be toggled independently under animations:

Key	Effect
typingPulse	Border and π lerp toward white the faster you type
submitFlash	Brief white border pulse when you submit (text goes non-empty → empty)
metricPulse	Metric text pulses toward white when its value changes
tokPulse	The ~N tok counter pulses each time the estimate updates

The thinking equalizer (the ▁▂▃…█ bar shown while the model works) is always on.

How it works

The extension registers a custom editor on session_start (and hides the default footer), subclassing pi's CustomEditor and overriding render(width).

From-scratch rendering. Instead of calling super.render() and post-processing, it builds every line itself. It reserves columns for the │ borders and the π prefix, then calls the inherited layoutText() to word-wrap the input (which keeps paste markers and grapheme segmentation intact). The cursor is drawn by inverting the grapheme under it (\x1b[7m…), and each line is padded to the inner width and wrapped in border characters. The top and bottom borders are produced by fitRoundedBorder, which fits a left and a right text into one line, truncating them when space runs short.

Session metrics. On each render it reads the session entries from ctx.sessionManager.getEntries() and aggregates token usage and cost (whole session, and the current/last turn separately). Context usage comes from ctx.getContextUsage().

Tool count. The extension subscribes to before_provider_request and keeps a reference to the tools array packed into the request payload. The count is read lazily at render time, after the whole hook chain has run, so it reflects any in-place filtering other extensions apply (for example MCP-bridged tools that inflate the active set but never reach the wire) and reports exactly what was sent on the last request. Before the first request it falls back to pi.getActiveTools().

Color animations. Several effects share a brightness engine that parses an ANSI foreground color to RGB and re-emits it in the same terminal mode (truecolor or 256-color), so the animation is never a no-op on 256-color terminals:

• Typing whitening — characters added are sampled over a sliding window to estimate WPM; the border and π lerp toward pure white the faster you type, with a fast attack and an exponential release when you stop.
• Submit flash — a non-empty → empty text transition triggers a brief white border pulse.
• Metrics pulse — when the session metrics change, the metric text pulses toward white and decays back.
• Thinking equalizer — while the model works, a VU-meter bar (▁▂▃…█) and a status word animate on their own line, driven by an independent sinusoid so the border stays at the fixed accent.

Animations run on short setInterval timers that request a re-render and stop themselves once the effect has fully decayed; all timers are cleared in dispose().

Compatibility

• pi >= 0.78

License

MIT © Sébastien SERVOUZE