OpenCode TPS Counter Plugin

This plugin posts a compact generation speed line after assistant responses in OpenCode:

▣ Lat.: <value>s | E2E TPS: <value>

What the metrics mean

• E2E TPS: Effective local throughput in this OpenCode session.
  • Formula: (output tokens + reasoning tokens) / (effective end - turn start - merged tool execution time)
  • effective end prefers the last streamed token time when available, and falls back to message completion time.
  • This is measured from OpenCode message/event timestamps, not provider-native telemetry.
• Lat.: Time to first token for the assistant turn.
  • Formula: first streamed token time - turn start
  • If streaming timing is unavailable, it shows n/a.

How the plugin measures a turn

• The plugin reports only when an assistant message is finalized with finish: "stop".
• It groups all assistant messages in that turn by parentID (tool-call steps plus the final assistant stop).
• It sums output + reasoning tokens across the grouped assistant messages.
• It tracks stream windows from both stored parts and live message.part.updated events so latency still works when part timing is incomplete.
• It merges overlapping tool runtimes and subtracts that merged tool time from the end-to-end duration.

Important limitation

These numbers are based on OpenCode-local timing and token accounting. They are useful for relative comparisons in the same environment, but they will not always match provider dashboards (for example OpenRouter throughput) because provider-side queueing/transport internals are not exposed in plugin events.

Reliability is highest when a provider streams chunks continuously. If a provider buffers output and flushes large chunks (or nearly the full response) at once, the measured generation window becomes artificially short and the reported rate can spike.

Example: some Z.ai responses can arrive in bursts where most assistant text appears nearly at once. In that case, the displayed value is less representative of true model-side throughput. This behavior comes from provider streaming characteristics and is not fully fixable from plugin-side event timing.

Trigger behavior

The plugin reports when an assistant message reaches finish: "stop" on message.updated. This is more reliable than waiting for session.idle.

Install from npm

Add the plugin package to your OpenCode config. OpenCode installs npm plugins automatically.

Example opencode.json:

{
  "plugin": ["@howaboua/opencode-tps-counter@^0.1.0"]
}

Local development

Install dev dependencies:

bun install

Type-check the plugin:

bun x tsc --noEmit

Use local plugin wiring while developing:

{
  "plugin": ["file:///absolute/path/to/opencode-tps-counter/index.ts"]
}