opencode-timings

A tiny OpenCode TUI sidebar plugin that shows per-session timing — how much wall-clock the session has taken and how much of that was actually spent waiting on the model.

It renders into the same right-hand sidebar as the Quota / MCP / LSP / Todo / Files panels, reading the session's messages directly from the TUI's reactive state. Nothing is ever injected into the message stream, so there is zero context-window pollution.

Timing
api/wall ██████░░ 78%
api 31s · wall 40s
turns 4 · avg 8s
slowest 19s
tok/s avg 43 · last 51
tok in 1.2M · out 45k
per-turn ▄█▂▂
speed ▂▄█▆

Every row names itself, so there are no unlabeled numbers or glyphs to decode.

Metrics

| Row | Meaning | |------------|---------| | api/wall | How much of wall-clock was actual model inference, as a bar gauge and percent. | | api | Total assistant inference time — the sum of time.completed − time.created over every completed assistant message. | | wall | Span from the first to the last message timestamp. Includes the time you spend reading/typing between turns, so api is always a fraction of it. | | turns | Number of completed assistant messages, plus the average per-turn duration. | | slowest | The single slowest assistant message. | | tok/s | Generation speed: tokens the model generated (output + reasoning) over api time — avg across the session, last for the most recent turn. | | tok in / out | Total tokens the model read (input + cache read/write) vs generated (output + reasoning), across completed turns. | | per-turn | Sparkline of each recent turn's duration. | | speed | Sparkline of each recent turn's tok/s. |

The panel is always shown; before the first turn its values read zero.

A note on honesty: no metric reads the clock — everything derives from message timestamps, so idling between turns (or leaving a session open) moves only wall and the api/wall ratio, which exist precisely to show that contrast. The one thing timestamps can't separate is time inside a turn: a turn's span (completed − created) includes tool execution and any wait on you (permission prompts, questions), so on such turns api reads high and tok/s reads low — it's the effective speed you experienced, not the model's raw decode speed.

Install

Add it to the plugin array of the TUI config that OpenCode loads (~/.config/opencode/tui.json or tui.jsonc) — this is a TUI plugin, so it belongs in tui.json, not opencode.json:

{
  "$schema": "https://opencode.ai/tui.json",
  "plugin": ["@foae/opencode-timings@latest"]
}

OpenCode installs the plugin and its dependencies with Bun at startup. Restart OpenCode and open the session sidebar to see the Timing panel.

You can also pin a version, e.g. @foae/[email protected].

Configuration

Pass options using the tuple form ([spec, options]) in tui.json:

{
  "$schema": "https://opencode.ai/tui.json",
  "plugin": [
    ["@foae/opencode-timings@latest", {
      "mode": "fancy",
      "fields": {
        "ratio": true, "api": true, "wall": true, "turns": true, "avg": true, "slow": true,
        "tps": true, "lastTps": true, "tokIn": true, "tokOut": true,
        "sparkline": true, "speedline": true
      }
    }]
  ]
}

| Option | Values | Default | Meaning | |----------|--------|---------|---------| | mode | "fancy" | "simple" | "fancy" | fancy draws the gauge bar on the api/wall row and adds the two sparklines; simple is the same rows without the bar or sparklines. | | fields | object of booleans | all true | Each toggles exactly one value: ratio (the api/wall gauge + percent), api, wall, turns, avg, slow, tps (session-average tok/s), lastTps (latest turn's tok/s), tokIn, tokOut, sparkline (per-turn durations), speedline (per-turn tok/s) — the last two are fancy-only. Values that share a line drop out individually. |

With no options (a plain "@foae/opencode-timings@latest" string), it defaults to fancy mode with all fields shown.

Requirements

• OpenCode 1.15.x or newer (uses the TUI slot plugin API).

Development

Built and run with Bun.

• src/timing.ts — pure timing math, formatting, and config parsing (no JSX), unit-tested.
• src/tui.tsx — the SolidJS sidebar component and the slot registration.

bun install
bun run typecheck   # tsc --noEmit
bun test            # unit tests for the pure logic in src/timing.ts
bun run build       # compile src/ to dist/ with babel-preset-solid

The package ships dist/tui.js, pre-compiled with babel-preset-solid (universal renderer, moduleName: "@opentui/solid"). Shipping raw .tsx no longer works: OpenCode installs plugins under a node_modules path, and since @opentui/solid 0.4.x the host's Solid JSX transform skips node_modules, so raw JSX gets compiled by Bun's plain (non-reactive) JSX runtime and the panel renders once and never updates. The compiled output imports bare @opentui/solid / solid-js specifiers, which the host maps onto its own shared runtime instances.

opentui, solid-js, and the OpenCode plugin/SDK are peer dependencies — at runtime they come from the OpenCode host so the plugin shares its renderer; the devDependencies mirror them for local typecheck and tests.

License

MIT — see LICENSE.