Tps Meter

English | 中文

A TUI plugin for opencode that displays real-time LLM output speed metrics next to the session prompt.

Display

Idle:

TPS 42.5 | AVG 38.2 | TTFT 0.8s

While the model is reasoning (a flowing particle stream appears to the right of TTFT):

TPS 42.5 | AVG 38.2 | TTFT 0.8s ⠒⠑⠂

• TPS — current tokens per second (live, color-coded)
• AVG — session-wide cumulative average from completed messages (color-coded)
• TTFT — time to first token, shown immediately on first token, persists until next generation (color-coded)
• Thinking stream — while reasoning streams, a FIFO particle stream (newest dot on the left, fading right, color shifts over time) appears to the right of TTFT. Only rendered while thinking — no space is reserved when idle.

Color Tiers

TPS & AVG (higher is better):

Tier	Range	Color
Slow	< 20 TPS	Red
Normal	20–50 TPS	Yellow
Fast	50–100 TPS	Green
Faster	> 100 TPS	Cyan

TTFT (lower is better):

Tier	Range	Color
Fast	< 10s	Green
OK	10–20s	Yellow
Slow	> 20s	Red

Installation

Important: This plugin must be configured in tui.json, not opencode.json.

One command (recommended)

opencode plugin @jack-yang/opencode-tps-meter -g

Installs the plugin and updates your config. Restart opencode — done. (Applies default options; to customize, use the tuple form in Configuration.)

Manual: npm

Add @jack-yang/opencode-tps-meter to the plugin array in ~/.config/opencode/tui.json:

{
  "plugin": [
    "@jack-yang/opencode-tps-meter"
  ]
}

GitHub

{
  "plugin": [
    "git+https://github.com/yangwudong/opencode-tps-meter.git"
  ]
}

Local file

1. Download tui.tsx to your opencode config directory:

curl -o ~/.config/opencode/tps-meter.tsx https://raw.githubusercontent.com/yangwudong/opencode-tps-meter/main/tui.tsx

1. Add to ~/.config/opencode/tui.json:

{
  "plugin": [
    "./tps-meter.tsx"
  ]
}

1. Restart opencode.

Configuration

All options are optional — defaults are sane. Pass them via the tuple form in tui.json:

{
  "plugin": [
    ["@jack-yang/opencode-tps-meter", {
      "spinner": {
        "theme": "tech",
        "colors": ["#00e5ff", "#2979ff", "#651fff"],
        "intervalMs": 100,
        "cells": 6
      },
      "tiers": {
        "tps":  { "slow": 20, "normal": 50, "fast": 100 },
        "ttft": { "fast": 10000, "ok": 20000 }
      }
    }]
  ]
}

Option	Default	Description
spinner.theme	"tech"	Color preset: "tech" (cyan→blue→purple) or "red" (KITT-style red). "none" disables the thinking stream.
spinner.colors	from theme	Custom gradient (hex strings); overrides theme.
spinner.intervalMs	100	Particle stream frame interval (ms).
spinner.cells	6	Number of cells (width) of the thinking stream.
tiers.tps	{20,50,100}	TPS/AVG thresholds (slow/normal/fast, in TPS); above fast is "faster".
tiers.ttft	{10000,20000}	TTFT thresholds in ms (fast/ok); above ok is "slow".

How It Works

Live TPS

Tracks text/reasoning delta events in a rolling 5-second window. Uses wall-clock duration (now - oldest sample) instead of inter-arrival time sums to avoid inflation from network buffering.

Token estimation uses ceil(byteLength / 4) with a calibration factor: when a message completes, the estimated token count is compared against actual tokens.output + tokens.reasoning from the message metadata, and a running median ratio corrects future estimates.

AVG

Session-wide cumulative average from all completed messages: sum(actual tokens) / sum(generation durations). Updates when a message completes. Generation duration = last delta - first delta per message (excludes tool execution time and TTFT).

TTFT

Time from message creation (info.time.created) to the first text/reasoning delta. Displayed immediately when the first token arrives, and persists until the next generation starts.

Development

git clone https://github.com/yangwudong/opencode-tps-meter.git
cd opencode-tps-meter
npm install
npm test

Pure measurement functions (measure.ts) have 29 unit tests. The TUI plugin (tui.tsx) is a standalone single file (all functions inline) for direct deployment.

Requirements

• opencode >= 1.4.3 (tested on 1.17.20)

License

MIT