Loading ToolBench
Scoring was not performed
All tool descriptions are in Chinese, preventing LLM comprehension in English-dominant multi-model environments. Descriptions like '获取所有项目列表' (Get all projects list) are short (20 - 30 chars) and lack actionable context for tool selection.
No output schemas documented. All tools return Dict[str, Any] with no hints about structure, fields, or types. LLMs cannot plan downstream operations or extract required IDs. For example, create_project should document: returns {id: string, name: string, color: string, ...}.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 25 | <=2025-11-25 | v2 |
| 2026-03-09 | F | 43 | - | v1 |
Parameter naming violates clarity convention. 'project_id_or_name' and 'tag_id_or_name' are overloaded and ambiguous. Should be separate optional params: 'project_id' and 'project_name' (or use a union type). LLMs struggle with dual-purpose params.
No tool annotations. DESTRUCTIVE operations (delete_project, delete_tag) lack destructiveHint=true. READ_ONLY tools (get_projects, get_tags) lack readOnlyHint=true. This prevents agents from reasoning about safety and undo/recovery.
No pagination or result limits. get_projects and get_tags return entire lists with no offset/limit params and no documented max count. Large projects/tags lists will exhaust context windows.
Error handling lacks recovery guidance. Functions catch exceptions and return wrapped dicts {success: False, info: error_string} but do not suggest next steps. E.g., 'Project not found' should suggest 'Call get_projects() to see available projects.'
Parameter descriptions are minimal or missing English translation. E.g., 'color' param describes format as '#FF0000' but does not explain if hex is required, if named colors are accepted, or default behavior.
merge_tags tool lacks clarity on merge semantics. Does 'merge_tags(source_name, target_name)' delete the source? Move items? Fail if source does not exist? Description does not answer these questions, forcing agent guessing.