pi-gauge

Live LLM speed in pi's status bar — throughput (tokens/sec) and latency (time-to-first-token) at a glance, updating while the model streams.

v0.1.0 work-in-progress — status-bar format may change before 1.0.

Check out my other Pi extensions

• — Auto-syncs Pi theme with OS dark/light mode.
• — Real-time prompt cache hit rates and token metrics.
• — Session task checklist with dependencies and a TUI renderer.
• — Prunes verbose tool outputs from context while preserving history.
• — Dot-grid visualization of context window token usage.
• — Live tokens/sec and TTFT in the status bar.
• — Tracks subscription quotas and rate limits across AI providers.

What you get

Just ask something in pi and the status bar shows a little pill like this:

[⚡ 42.1 t/s ⏱ 412ms]

• ⚡ 42.1 t/s — how fast tokens are arriving
• ⏱ 412ms — how long you waited for the first token (TTFT)

It ticks live during streaming, then settles on the final numbers. Waiting for the first token? You'll see [⏱ —] until it arrives.

Install

pi install npm:pi-gauge

Requires a pi coding agent with extension support. Settings persist to ~/.pi/agent/pi-gauge.json.

How it looks with different settings

All examples below are the exact strings the extension renders (same renderText the status bar uses), for one identical reply — 84.2 tokens in 2 s with 412 ms TTFT — unless noted.

Default (pill + emoji)

$ /gauge
pi-gauge: tps=on ttft=on mode=e2e display=pill icons=emoji

status:  [⚡ 42.1 t/s ⏱ 412ms]

Hide one metric

/gauge tps off      →   [⏱ 412ms]        # latency only
/gauge ttft off     →   [⚡ 42.1 t/s]     # throughput only

Turn both off and the gauge hides itself (status cleared).

Compact mode (no brackets, no t/s)

/gauge display icon

status:  ⚡ 42.1 ⏱ 412ms

Back to default with /gauge display pill:

status:  [⚡ 42.1 t/s ⏱ 412ms]

Icon sets: emoji vs Nerd Font

/gauge icons emoji   (default — works everywhere)

status:  [⚡ 42.1 t/s ⏱ 412ms]

/gauge icons nerd    (needs a Nerd Font in your terminal)

status:  [ 42.1 t/s  412ms]

The nerd glyphs are tachometer U+F0E4 for t/s and clock U+F017 for TTFT. If you see boxes, your terminal isn't using a Nerd Font — switch back with /gauge icons emoji.

Combine them freely — compact + nerd is the minimal look:

/gauge display icon
/gauge icons nerd

status:    42.1  412ms

Fast vs slow models (same prompt)

fast model:   [⚡ 142 t/s ⏱ 210ms]
slow model:   [⚡ 6.2 t/s ⏱ 1.85s]    # durations ≥1s render as seconds

Number formatting: t/s shows one decimal below 100 (42.1), whole numbers at/above (142). Durations show ms under a second, s with two decimals above (1.85s).

Measurement modes: e2e vs stream

Same reply (100 tokens, 412 ms TTFT, 2 s total), two ways to count:

mode=e2e:     [⚡ 50.0 t/s ⏱ 412ms]   # tokens ÷ whole request (request → message_end)
mode=stream:  [⚡ 63.0 t/s ⏱ 412ms]   # tokens ÷ streaming only (first token → message_end)

e2e (default) answers "how long did I wait overall?" — comparable across providers. stream answers "how fast does it feel once it starts?" — TTFT is identical either way, only the t/s denominator changes.

Provider note: Anthropic/Google stream cumulative usage, so live t/s is exact. OpenAI usually sends usage only in the final chunk, so the live number is a chars÷4 estimate that snaps to the exact value at message_end.

The /gauge command

Bare /gauge in the TUI opens a bordered settings dialog with all five settings as toggle rows, a live demo preview (mocked reply: 84.2 tokens in 2 s, 412 ms TTFT) so you can see each change as you make it, and a ● unsaved flag in the title while edits are uncommitted. ↑↓ move, ←→/space/enter toggle, ctrl+s saves (persists to ~/.pi/agent/pi-gauge.json and re-renders the status bar), esc exits without saving. Outside the TUI (or with subcommands) the classic notify-based behaviour is unchanged. Subcommands and their values tab-complete in the TUI (/gauge m → mode, /gauge mode → e2e|stream).

| Command | What it does | |---|---| | /gauge | Open settings dialog (TUI) / show current settings (otherwise) | | /gauge tps on\|off | Show/hide throughput (throughput also works) | | /gauge ttft on\|off | Show/hide latency (latency also works) | | /gauge mode e2e\|stream | Switch the t/s denominator | | /gauge display pill\|icon | Bracketed pill vs bare compact | | /gauge icons emoji\|nerd | Emoji glyphs vs Nerd Font glyphs | | /gauge graph [model] | History chart: per-call TPS + TTFT bars for this session (TUI overlay; text summary elsewhere). ←→ cycles the model filter, ↑↓/PgUp/PgDn scroll, esc closes |

Every change saves immediately and re-renders the status bar if metrics exist.

History graph (/gauge graph)

Every assistant reply also appends one small record (pi-gauge-sample: model, tokens, TTFT, e2e/stream durations) to the session JSONL via appendEntry — invisible in the transcript, zero LLM context cost, and it survives session resume. /gauge graph renders it:

• TUI: bordered overlay with per-call TPS bars (denominator follows your mode e2e|stream setting) stacked over per-call TTFT bars, footer averages, ←→ cycles the model filter (all → model…), ↑↓/PgUp/PgDn scroll long sessions, esc closes. /gauge graph gpt-5.4 pre-filters.
• Outside the TUI (--print, rpc/json): prints a per-model text summary (calls, avg t/s, avg TTFT) instead of the overlay.

Sessions from before this feature simply show "no samples yet".

How it works

before_provider_request ──► first token ──► …streaming… ──► message_end
        │ TTFT = ──────────┘                                    │
        └────────────── e2e duration ───────────────────────────┘
                              └────── stream duration ──────────┘

• Request start is anchored on before_provider_request, TTFT ends at the first *_delta, totals snap to exact usage.output at message_end.
• Status writes are throttled (~4/sec) so fast streams don't spam the UI.
• State re-anchors on every LLM call (tool loops make several per turn); aborted/empty streams keep partial metrics instead of crashing.

Dev

npm install
npm run build   # src/*.ts → dist/ (dist/ is COMMITTED — pi installs with --omit=dev and no build step)
npm test

Regenerating the demo GIF

The GIF at the top is recorded with VHS (Charm's terminal-screen recorder — you write a .tape script, it types the commands in a headless terminal and exports a .gif).

• Script: vhs/demo.mjs — imports the real renderText from dist/, so the GIF can never drift from what the status bar shows.
• Tape: vhs/gauge.tape — font, theme, window size, and timing.
• Output: assets/demo.gif

brew install vhs ffmpeg   # once
npm run build             # demo imports from dist/
vhs vhs/gauge.tape        # → assets/demo.gif (~17 s)

License

MIT © 2026 Arnav Gupta