Simplified-ReasoningOfficial·29 Agent Skills Included

Pi-Bench

Benchmark for proactive assistants in long-horizon workflows

Evaluates proactive personal assistants across 100 multi-turn tasks in five professional personas. Measures Proactivity and Completeness scores to reveal how well models uncover hidden user intents over long sessions. Provides ready-made model configs, Docker runtime, and a public leaderboard for reproducible comparisons.

58|1

↗

npx skills add Simplified-Reasoning/Pi-Bench --all -g -y

All Skills in This Repository (29)

Pure Emerald Level Indicators

📦 In Repo

Simplified-Reasoning

memory

Separate persistent facts from event history and search HISTORY.md with grep.

summarize

Condense long-form text from URLs, local files, and YouTube links into summaries.

clawhub

Search the ClawHub registry by natural-language queries and install skills into the nanobot workspace.

skill-creator

Create and package modular AgentSkills with SKILL.md frontmatter metadata.

github

Retrieve GitHub issues, pull requests, and CI run details via gh CLI.

tmux

Start isolated tmux sessions, send keystrokes, and capture pane output.

weather

Retrieve current weather and forecasts for locations via wttr.in with Open-Meteo JSON fallback.

cron

Schedule reminders and recurring tasks with cron expressions and intervals.

jargon-translator

Translate workplace jargon between plain Chinese and professional phrasing.

law-exam-trainer

Convert law exam videos and documents into structured practice questions.

Data Analysis

Convert raw data into validated insights with uncertainty and caveats.

china-tax-law

Cite PRC tax laws and provisions for compliance and dispute guidance.

Official

Intermediate

Frequently Asked Questions

FAQPage Schema

How to install Pi-Bench?▼

Run `npx skills add Simplified-Reasoning/Pi-Bench --all -g -y` in your terminal to install all skills in this suite globally.

What does Pi-Bench measure?▼

It measures Proactivity (whether an assistant discovers hidden user intents early) and Completeness (whether final outputs satisfy checklist requirements) across long-horizon, multi-session tasks.

Which models can Pi-Bench evaluate?▼

It ships with ready-made configs for Claude, DeepSeek, Gemini, GLM, MiniMax, and Doubao models, and you can add any custom model via a YAML config file.

How do I run a Pi-Bench evaluation?▼

After setup, run `pibench --model-id deepseek-v3.2 --run 3` from the repository root to execute three leaderboard-style repeats of the benchmark.

What personas and tasks does Pi-Bench include?▼

It includes 100 multi-turn tasks across five personas: researcher, marketer, pharmacist, law trainee, and financier, each with persistent workspaces and hidden intents that emerge over time.