test-tps · v1.0.0 · 2026-09-19 · sha256 29212731f6808aa8

test-tps v1.0.0A

Immutable. This exact content never changes and is served at /api/v1/blob/29212731f6808aa8.

No tool to install first
claude plugin marketplace add https://markdownregistry.com/mp/a/art_u4t2kkbwolidm5ay/29212731f6808aa8dc2c70d7d89ea794efd2b199bfcecf9b73da0e0a65c11973.json

Then /plugin install test-tps@markdownregistry-u4t2kkbwolidm5ay in Claude Code. It installs this exact version, the one graded A here, pinned to the commit it was graded at, and leaves it there.

---
name: test-tps
version: 1.0.0
description: User-invoked benchmark for selected Pi provider/model. Measures TTFT, decode TPS, average TPS, latency, token usage, and aggregate percentiles. Use when user runs /skill:test-tps.
disable-model-invocation: true
argument-hint: "[--runs N] [--warmup N] [--output-tokens N] [--provider ID] [--model ID] [--json]"
---

# Test TPS

Benchmark selected Pi model. Invocation spends provider quota and may incur cost.

## Run

Resolve this skill directory, then run:

```bash
node <skill-dir>/scripts/test-tps.mjs <arguments>
```

Defaults: `PI_PROVIDER`, `PI_MODEL`, current reasoning level, 3 measured runs, 1 warmup, 256 requested output tokens.

Report command output verbatim. Do not reinterpret `N/A` metrics. Prefill TPS is `N/A` unless provider telemetry exposes prompt-processing duration. TTFT is latency, not prefill duration. Requested output tokens are a prompt target because Pi's CLI exposes no generic per-request max-output flag.

## Rules

- Use current selected provider/model unless arguments override them.
- Preserve failed-run messages.
- Never print credentials or auth files.
- For stable comparisons, keep provider, model, reasoning, prompt, run count, and output target identical.