pinchbench
от olearycrew
Coding Agents & IDEsКратко
Run PinchBench benchmarks to evaluate OpenClaw agent performance across real-world tasks.
Описание
Что делает этот навык
PinchBench measures LLM model performance as the brain of an OpenClaw agent across 23 standardized tasks. Tasks span calendar, email, research, coding, file management, and multi-step workflows. Results are graded automatically and can be submitted to a public leaderboard at pinchbench.com.
Provides 23 reproducible, auto-graded tasks with a public leaderboard, making model comparisons objective rather than anecdotal.
Когда его использовать
- Comparing two models on email triage and calendar tasks before picking one
- Running the automated-only suite for a quick performance check before a release
- Submitting benchmark scores to the PinchBench public leaderboard
- Testing a specific subset of tasks like market research or PDF summarization
- Verifying a new OpenClaw setup handles multi-step API workflows correctly
Пример сценария
User asks: How well does claude-sonnet-4 handle email and calendar tasks?
- 1
Register for a PinchBench API token using the --register flag
- 2
Run the benchmark targeting specific tasks: --suite task_01_calendar,task_16_email_triage,task_17_email_search
- 3
Wait for tasks to complete and results to save as JSON in the results/ directory
- 4
Inspect scores with jq to extract per-task grading means
- 5
Results upload automatically to the leaderboard using the registered token
Per-task scores for calendar and email tasks, plus an overall average, visible on the pinchbench.com leaderboard
Требования
- PinchBench API token for leaderboard submission (register once via --register flag)
Установка
clawhub install olearycrew/pinchbenchВыполните команду CLI или вставьте промпт в Claude Code, Cursor, Codex — любой кодинг-агент, которым пользуетесь.