pi-token-meter

English | 简体中文

What it is

A session token meter for pi. It shows the numbers growing in real time, and prints a settlement line when a round ends (that line never enters the model context).

Live line (next to the spinner):

── ⠋ 12s · ↓1.2k 45tok/s ──────────────────────────

Result line:

3m32s · ↑286k ↓1.2k R13.4M W1.2M 57tok/s · claude-opus-5 (high) $20.16

[!TIP] ↑ input input (uncached input)
↓ output output (thinking tokens included)
R/W cache read / cache write (shown when non-zero)
model name the model ID (the same one pi's model list and status bar show)
cost computed from the model price list (per-million-token prices, tiered pricing included, 1h cache writes doubled); same source and same accounting as pi's status-bar cost
tok/s output tokens ÷ pure generation time. Generation time accumulates "first output token → end" per model call, so it excludes first-token wait and tool execution
round duration wall-clock time of a whole round, from the moment the message is sent until the round settles
estimates a usage report the provider sends while streaming is taken as-is. Before it reports, the input side is split into input / cache read by "context size − last cached prefix"

Features

• Live counters: duration / ↑ input / ↓ output / R/W cache / speed / $ cost, growing in real time
• Arrow animation presets: Steady (the changing numbers are highlighted), Blink (the changing numbers blink per frame), Merge (only the changing side is shown)
• Wait timer: after a message is sent, the round's timer is appended after pi's working line (Working 1s), and is replaced by the full live line as soon as the model starts answering
• Round settlement: one round = from sending the message to fully settling (automatic retries / compaction continuations / queued messages included). Once it ends, one line is written at the end of the transcript
• Complete accounting: every model call in the round is summed; nested model calls inside tools (subagents and the like) and the summary generation of automatic compaction count as well; on interrupt or error, whatever was spent is still settled
• Visual configuration: /token-meter-settings opens a settings panel in the same style as pi's own /settings

[!NOTE] Numbers during streaming are estimates or incremental reports; the round settlement is what counts (the settled value agrees with the provider bill exactly)
Idle-time cache warming happens outside a round and is not counted in it; it does count toward pi's own session totals
When a model has no prices configured (cost is 0) the cost shows $0; token counting is unaffected

Install

pi install npm:@hayston/pi-token-meter
pi install -e npm:@hayston/pi-token-meter     # one-shot trial

Configuration file

[!TIP] Use /token-meter-settings to open the visualization settings panel

At ~/.pi/agent/token-meter.json

Global

Live line fields (submenu)

Result line fields (submenu)

[!WARNING] Missing fields, wrong types and broken JSON all fall back to the defaults, and the refresh interval snaps to a preset, so a bad file can never drag a session down
language is re-resolved every time the panel opens. Changing Language in the panel reopens it immediately in the new language

Development

node test/run.mjs           # all
node test/run.mjs unit      # copy, config, metering, formatting, panel, event wiring
node test/run.mjs lifecycle # real AgentSession: extension discovery, registration result, /reload