CLI-Bench: Benchmarking AI Agents on CLI Tool Orchestration

CLI-Bench is a benchmark for evaluating the ability of AI agents (e.g., LLM-based coding assistants) to use command-line interface tools to accomplish real-world developer tasks. Unlike existing benchmarks that focus on code generation or isolated API calls, CLI-Bench tests whether agents can orchestrate multiple CLI tools end-to-end across realistic workflows spanning project management, DevOps, communication, and data operations.

Overview

Property	Value
Tasks	40
Categories	6 (devops, project_mgmt, communication, data_ops, custom_cli, composite)
Tool Adapters	12 (7 real-world + 5 fictional)
Difficulty	20 easy, 10 medium, 10 hard
Format	YAML task definitions with declarative initial/expected state

Task Categories

• devops: Infrastructure and deployment operations (CI/CD, monitoring, alerts)
• project_mgmt: Issue tracking, sprint management, task coordination across platforms
• communication: Messaging, notifications, channel management via Slack and email
• data_ops: Data pipeline construction, ETL operations, report generation
• custom_cli: Tasks using fictional CLIs that cannot be memorized from training data
• composite: Multi-tool workflows requiring coordination across 2-3 tools in sequence

Tool Adapters

Real-World Tools (7)

Tool	Domain
gh	GitHub CLI (issues, PRs, repos, actions)
slack	Slack CLI (messages, channels, users)
linear	Linear CLI (issues, projects, cycles)
notion	Notion CLI (pages, databases, blocks)
google	Google Workspace (Gmail, Calendar, Drive)
jira	Jira CLI (issues, sprints, boards)
microsoft	Microsoft 365 (Teams, Outlook, OneDrive)

Fictional Tools (5) — Memorization-Proof

Tool	Domain
kforge	Artifact registry and deployment management
flowctl	Workflow engine with approval gates
meshctl	Service mesh topology and traffic control
datapipe	Declarative ETL pipeline builder
alertmgr	Alert routing, escalation, and incident management

Fictional tools are designed so that agents cannot rely on memorized CLI syntax from pre-training. Agents must read the provided tool adapter specifications and reason about correct usage from first principles.

Task Format

Each task is a YAML file containing:

id: cb-001
title: "List open issues in a GitHub repo"
difficulty: easy
category: project_mgmt
description: |
  Natural language description of the task objective.
tools_provided:
  - gh
initial_state:
  gh:
    repos:
      acme-corp/web-platform:
        issues:
          - number: 42
            title: "Fix login redirect loop"
            state: open
            assignee: alice
expected_state:
  gh:
    command_history:
      - pattern: "gh issue list.*--repo acme-corp/web-platform.*--state open"
    output_contains:
      - "42"
scoring:
  outcome: 0.6
  efficiency: 0.2
  recovery: 0.2

• initial_state: The simulated environment state before the agent acts.
• expected_state: Declarative assertions on command patterns, state mutations, and expected outputs.
• scoring: Per-task weight overrides for the three evaluation dimensions.

Evaluation Metrics

CLI-Bench scores agents along three dimensions:

Metric	Weight (default)	Description
Outcome	0.6	Did the agent achieve the desired end state? Verified via declarative state assertions.
Efficiency	0.2	Did the agent use a reasonable number of commands? Penalizes excessive retries or unnecessary exploration.
Recovery	0.2	Did the agent handle errors or unexpected states gracefully? Tests resilience to failed commands and ambiguous outputs.

The aggregate score per task is a weighted sum. The benchmark also reports pass^k (the fraction of tasks solved within k attempts), providing a measure of reliability across repeated runs.

Difficulty Levels

• Easy (20 tasks): Single-tool, single-command operations with straightforward state assertions.
• Medium (10 tasks): Single-tool multi-step workflows or tasks requiring conditional logic.
• Hard (10 tasks): Multi-tool composite workflows requiring sequential orchestration, error recovery, and cross-tool state propagation.

Usage

With the datasets library

from datasets import load_dataset

dataset = load_dataset("ChengyiX/CLI-Bench")

Loading YAMLs directly

import yaml
from pathlib import Path

tasks = []
for task_file in sorted(Path("data/tasks").glob("cb-*.yaml")):
    with open(task_file) as f:
        tasks.append(yaml.safe_load(f))

print(f"Loaded {len(tasks)} tasks")
print(f"Categories: {set(t['category'] for t in tasks)}")

Loading tool adapter specifications

import yaml
from pathlib import Path

adapters = {}
for adapter_file in Path("tool_adapters").glob("*.yaml"):
    with open(adapter_file) as f:
        adapter = yaml.safe_load(f)
        adapters[adapter_file.stem] = adapter

print(f"Loaded {len(adapters)} tool adapters")

Repository Structure

data/
  metadata.yaml          # Benchmark metadata and configuration
  tasks/
    cb-001.yaml          # Individual task definitions
    cb-002.yaml
    ...
    cb-040.yaml
tool_adapters/
    gh.yaml              # GitHub CLI adapter spec
    slack.yaml           # Slack CLI adapter spec
    ...
    kforge.yaml          # Fictional: artifact management
    flowctl.yaml         # Fictional: workflow engine
    meshctl.yaml         # Fictional: service mesh
    datapipe.yaml        # Fictional: ETL pipelines
    alertmgr.yaml        # Fictional: alert management

Citation

@misc{cli-bench-2026,
  title={CLI-Bench: Benchmarking AI Agents on Command-Line Tool Orchestration},
  author={Chengyi Xu},
  year={2026},
  url={https://github.com/minervacap2022/CLI-Bench}
}

Links

• GitHub: https://github.com/minervacap2022/CLI-Bench
• License: Apache 2.0