opencode-throughput

Real-time LLM performance monitoring plugin for OpenCode. Tracks latency, throughput, token usage, and cost per model with toast notifications and JSONL logging.

Features

• TTFT (Time To First Token) — measures how fast the model starts generating
• TPS (Tokens Per Second) — generation throughput during streaming
• Total Latency — end-to-end request time
• Token Usage — input, output, reasoning, cache read/write
• Cost Tracking — per-request and cumulative cost
• Toast Notifications — real-time performance summary after each LLM response
• JSONL Logging — persistent logs at ~/.opencode/throughput.jsonl
• Benchmark Tool — query historical stats on demand from the AI agent

Installation

需要在两个配置文件中分别声明插件：

1. opencode.json — 服务端插件（事件监听、数据采集）

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

2. .opencode/tui.json — TUI 插件（侧边栏显示）

{
  "$schema": "https://opencode.ai/tui.json",
  "plugin": ["opencode-throughput"]
}

OpenCode 会自动从 npm 安装插件。

Usage

Once installed, the plugin runs automatically:

1. After each LLM response, a toast notification appears with metrics:

   claude-sonnet-4 | 2.1s TTFT | 58.3 tok/s | 4.2s | ↑1.2k ↓892

2. Logs are written to ~/.opencode/throughput.jsonl (one JSON object per line)

3. Query benchmarks — ask the AI agent to use the benchmark tool:

   "Show me the benchmark stats for claude" "Compare model performance" "Show the last 10 requests"

Benchmark Tool Args

| Arg | Type | Description | |-----|------|-------------| | model | string (optional) | Filter by model name (partial match) | | last | number (optional) | Only show the last N entries |

Log Format

Each line in ~/.opencode/throughput.jsonl:

{
  "ts": "2026-03-21T10:00:00.000Z",
  "model": "anthropic/claude-sonnet-4",
  "providerID": "anthropic",
  "modelID": "claude-sonnet-4",
  "sessionID": "abc123",
  "messageID": "def456",
  "ttft_ms": 2100,
  "tps": 58.3,
  "latency_ms": 4200,
  "inputTokens": 1200,
  "outputTokens": 892,
  "reasoningTokens": 0,
  "cacheReadTokens": 500,
  "cacheWriteTokens": 200,
  "cost": 0.0234,
  "finish": "stop"
}

How It Works

The plugin hooks into OpenCode's event system:

• message.updated — captures AssistantMessage.time.completed, tokens, and cost
• message.part.updated — captures the first TextPart.time.start for TTFT calculation

Development

Local development

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

Build

bun install
bun run build

Publish

npm publish --access public

License

MIT