Loading ToolBench
Single tool 'trash' has a clear verb-noun name and solid schema with proper types and descriptions. Input parameters are well-defined with minItems constraint and default values. Tool description (106 chars) is concise and explains the reversible nature. Output is structured (status, error fields). However, no output schema is documented for the LLM, and error handling lacks recovery guidance. The tool does one thing well but misses patterns for actionable error messages and downstream chaining.
Move files or directories to system trash (recycle bin). Never permanently deletes. Set verbose=true to see per-file results.
Output schema not documented. LLM cannot infer response structure (status, error fields, per-item results). Agents cannot reliably extract or chain results.
Error messages lack recovery guidance. 'Access denied, outside allowed directories' tells the LLM nothing about what to do next. Should suggest checking path or requesting directory allowlist expansion.
No confirmation or dry-run pattern for destructive operation. Although trash is reversible, agents should be able to preview what will be trashed before committing.
Tool description does not document the allowed directories constraint or that paths must be absolute. LLM may pass relative paths or paths outside allowlist without understanding why they fail.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 61 | 2026-07-28+ | v2 |
No batch operation guidance. If agent needs to trash 100 files, it must call the tool once with all paths. Tool accepts array but description does not clarify this is the intended pattern for bulk operations.