@puckguo123/timing

Per-turn model latency metrics, JSONL store, HTML dashboard, and LLM-based sanity analysis. Zero dependencies.

Install

npm install @puckguo123/timing @puckguo123/core

What it records

• ttftMs — time to first token
• llmMs — full LLM call duration
• toolMs — tool execution time
• inputTokens / outputTokens — usage from the provider

Quick start

import { TimingRecorder } from "@puckguo123/timing";

const rec = new TimingRecorder(".puck/timings.jsonl");
rec.record({
  sessionId: "abc-123",
  turn: 1,
  ttftMs: 230,
  llmMs: 1200,
  toolMs: 50,
  inputTokens: 1024,
  outputTokens: 256,
});

Dashboard

Run the bundled HTML report generator:

npx puck-timing-report .puck/timings.jsonl > report.html

License

MIT — see LICENSE.