Datasets:
| name: cli-bench | |
| version: "1.0" | |
| description: | | |
| CLI-Bench evaluates AI agents on their ability to use command-line tools | |
| to accomplish real-world tasks across developer platforms and services. | |
| Tasks span single-tool operations, multi-tool orchestration, and complex | |
| workflows involving fictional CLIs. | |
| task_count: 40 | |
| difficulty_distribution: | |
| easy: 20 | |
| medium: 10 | |
| hard: 10 | |
| categories: | |
| - devops | |
| - project_mgmt | |
| - communication | |
| - data_ops | |
| - custom_cli | |
| - composite | |
| tools: | |
| real_world: | |
| - gh # GitHub CLI | |
| - slack # Slack CLI | |
| - linear # Linear CLI | |
| - notion # Notion CLI | |
| - google # Google Workspace (Gmail, Calendar, Drive) | |
| - jira # Jira CLI | |
| - microsoft # Microsoft 365 | |
| fictional: | |
| - kforge # Artifact/deployment management | |
| - flowctl # Workflow/approval engine | |
| - meshctl # Service mesh control | |
| - datapipe # ETL pipeline builder | |
| - alertmgr # Alert/incident management | |
| scoring: | |
| dimensions: | |
| outcome: "Did the agent achieve the desired end state?" | |
| efficiency: "Did the agent use a reasonable number of commands?" | |
| recovery: "Did the agent handle errors or unexpected states gracefully?" | |
| default_weights: | |
| outcome: 0.6 | |
| efficiency: 0.2 | |
| recovery: 0.2 | |
| created: "2026-03-12" | |