A leave management system for handling employee leave applications with approval/rejection workflows
LeaveManagementSystem has basic tool definitions with schemas and descriptions present, but falls short of production quality. All 4 tools have input schemas and descriptions, but descriptions are minimal (10-50 chars, well below the 194-char baseline). Parameter descriptions are present but generic. No output schemas documented. Error handling is minimal (bare strings like 'Invalid index'). No tool annotations (readOnlyHint, destructiveHint). The in-memory database and lack of pagination/limits are architectural concerns. Tool naming is acceptable (verb_noun pattern), but descriptions lack context about when to use each tool and what they return.
Submit a leave application with employee ID, date range, and reason
Approve a leave application by index
Reject a leave application by index
Retrieve all leave applications from the database
Descriptions are too short (10-50 chars vs 194-char baseline). 'Retrieve all leave applications from the database' lacks context about pagination, limits, or return structure. LLMs cannot determine when to call view_leaves vs filtering via apply_leave results.
No output schemas documented. view_leaves returns List[Dict] but LLMs don't know the structure (fields: employee_id, from_date, to_date, reason, status). apply_leave returns a bare string. Agents cannot plan downstream calls or extract data reliably.
Error handling is non-actionable. 'Invalid index' tells the LLM nothing, no guidance on valid range, no suggestion to call view_leaves first. approve_leave and reject_leave should return structured errors with recovery hints.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 44 | 2026-07-28+ | v2 |
No tool annotations. approve_leave and reject_leave are destructive (modify state) but lack destructiveHint. view_leaves is read-only but lacks readOnlyHint. Agents cannot distinguish safe from risky operations.
view_leaves has no pagination or limit. In production, returning all leaves (potentially thousands) blows context windows and wastes tokens. Should accept limit and offset parameters and return a total count.