opencode-tps

A plugin for OpenCode that displays a live tokens-per-second (tok/s) meter in the terminal UI, plus a persistent avg / max / min summary that stays on screen after each response completes.

Fork of williamcr01/opencode-tps with a persistent post-response summary, scale-consistent numbers, and a tuned token estimator.

What it does

While the model is streaming, the meter shows a live tok/s reading (5-second rolling window) in the bottom-right of the prompt.

When the response completes, the meter freezes a summary for that message:

tok/s 38.2 avg · ↑51.0 ↓22.4

• avg — uses the real output-token count from the model's usage report whenever the API exposes it (info.tokens.output), falling back to an estimate otherwise.
• max / min — peak and floor of the per-second live readings during the response. The model API doesn't expose per-second real-token counts, so these are estimate-derived and then rescaled by the message's real/estimate ratio, keeping all three numbers on the same scale.

The summary stays on screen until the next assistant response starts streaming, then it's replaced by a fresh live reading.

Installation

Via OpenCode CLI

opencode plugin @mesaleh/opencode-tps

Via npm

1. Add the plugin to your opencode.json:

{
  "$schema": "https://opencode.ai/config.json",
  "plugin": ["@mesaleh/opencode-tps"]
}

2. Install the package:

cd ~/.opencode
npm install @mesaleh/opencode-tps

Requirements

• OpenCode >= 1.3.14
• OpenCode TUI (Web UI does not support this plugin)

How it works

The plugin subscribes to message.part.delta, message.updated, and message.part.updated events from OpenCode:

• During streaming: estimates tokens per delta (~5.5 bytes per token) and renders a live tok/s value over a 5-second rolling window. The slot shows tok/s - when no tokens are being generated.
• Per-message accumulator: in parallel with the rolling window, the plugin tracks first-delta timestamp, cumulative estimated tokens, and the observed max/min of the live reading (after a 3-second warm-up so a single early sample can't pin an artificial floor).
• On completion: computes avg = tokens / duration, preferring the real output-token count from the message info when available. Rescales max and min by the same real/estimate ratio so all three displayed numbers share a scale.
• Across tool calls: the live rolling window clears on tool transitions, but the per-message accumulator keeps going, so the final avg reflects the whole response.

License

MIT — see LICENSE. Original work © williamcr01.