@erlin-ai/pi-token-summary

Token/cost inline stats for the pi coding agent: dim transcript line with this turn's output, cumulative tokens, context %, and cost. Modes via /token-summary verbose|brief|off.

Packages

Package details

extension

Install @erlin-ai/pi-token-summary from npm and Pi will load the resources declared by the package manifest.

npm repohome report
$ pi install npm:@erlin-ai/pi-token-summary
Package
@erlin-ai/pi-token-summary
Version
0.5.2
Published
Oct 5, 2026
Downloads
1,277/mo · 27/wk
Author
erlinerd
License
MIT
Types
extension
Size
27.9 KB
Dependencies
0 dependencies · 1 peer
Pi manifest JSON
{
  "extensions": [
    "./src/index.ts"
  ]
}

Security note

Pi packages can execute code and influence agent behavior. Review the source before installing third-party packages.

README

pi-token-summary

CI License: MIT

English | 中文

Inline token/cost stats for the pi coding agent. After each agent reply, a dim stats line is appended to the conversation:

↓1.2k ⟳ · Σ20.5k · ctx 63% · ⚡130.8 tok/s (4.9s · TTFT 6282ms) · $0.11 · glm-5.3-flash
  • ↓this turn: output tokens of this reply
  • Σ: cumulative session output
  • ctx: context window usage
  • $: cumulative cost
  • model name

The stats line is a custom session entry (appendEntry + registerEntryRenderer): it renders in the TUI only, never enters LLM context, and historical lines re-render when a session is resumed.

The latest turn's TPS is kept resident in pi's footer as a status entry (ctx.ui.setStatus) — the default footer is kept, this only adds to it. While a stream is in flight it shows the running TPS (⚡12.3 tok/s (3.2s)), then the exact value once usage lands. Every streamed chunk repaints it — no throttle, no easing — so the number moves at the pace the provider streams. The ⚡ icon and number are colored by speed, dim red when slow and bright emboldened green when fast: #A31010 under 10, #E3170D under 30, #EF7C00 under 60, #B8D430 under 90, bold #7CFC00 under 120, bold #39FF8E at 120+. The footer is independent of the inline mode: off hides the transcript line only.

Display modes

/token-summary verbose   # after EVERY assistant message (incl. tool-call turns)
/token-summary brief     # once at the end of each turn (default)
/token-summary off       # no inline lines; totals still tracked, report works
/token-summary           # full session report + current mode

The mode persists in ~/.pi/agent/pi-token-summary.json across restarts. Default: brief.

Note: pi's turn_end event fires after every assistant message; a turn with tool calls produces multiple messages. verbose shows each one, brief shows once when stopReason is not toolUse (i.e. the turn truly ended).

How it works

pi's extension API provides the message usage on turn_end; pi.appendEntry() persists the rendered line to the session file and pi.registerEntryRenderer() renders it dim in the transcript. Cumulative values are seeded from the session file at startup, so resumed sessions continue accurately. Purely local — no network requests.

Install

pi install github:erlinerd/pi-token-summary   # GitHub
pi install npm:@erlin-ai/pi-token-summary               # once published to npm

Takes effect in new sessions.

Uninstall

pi uninstall @erlin-ai/pi-token-summary

Development

npm install
npm test        # tsx --test
npx tsc --noEmit

License

MIT