Server provides 6 tools with varying quality. All tools have descriptions (10-100 chars range), and most have parameter schemas visible. However, several critical gaps undermine production readiness: (1) Tool names lack clear action verbs, 'backlog-read', 'backlog-write', 'backlog-done' are vague and don't follow verb_noun convention ('get_backlog_items', 'create_backlog_item', 'complete_backlog_item' would be clearer). (2) Parameter descriptions are minimal or missing context. For example, 'backlog-write' action enum lists 7 operations but doesn't explain when to use each or their prerequisites. (3) Output schemas are not documented, tool handlers return JSON strings or plain text without declared structure. (4) Error handling is ad-hoc; most handlers throw generic errors like 'Unknown action' without recovery guidance. (5) No indication of idempotency, retry safety, or consequences for destructive operations like prune/clear. (6) Dependency management (e.g., 'backlog-ticket-done' validates dependencies but doesn't explain what dependencies are or how they work). These gaps are typical of community tools but fall short of production-grade agent tooling.
Mark backlog items as complete with optional summary - done operation and list
Prune completed/done backlog items - remove old archived items from COMPLETED_Backlog
Read-only access to backlog items - list and view backlog work items
Mark backlog todos as complete with dependency validation
Read-only access to backlog todos - list and filter todos for a backlog item
Manage backlog items - create, amend, submit, approve, reopen, and mark as wontfix
Tool names do not follow verb_noun convention. Names like 'backlog-read', 'backlog-write', 'backlog-done' are ambiguous. 'backlog-write' performs 7 different actions (create, list, amend, submit, approve, reopen, wontfix), this violates the single-responsibility principle and forces LLMs to guess which action to invoke.
'backlog-write' combines multiple operations (create, amend, submit, approve, reopen, wontfix) into one tool with an action enum. This violates single responsibility and increases cognitive load. Should split into separate tools: create_backlog_item, update_backlog_item, submit_backlog_item, approve_backlog_item, reopen_backlog_item, mark_backlog_wontfix.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 46 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 33 | - | v1 |
No output schemas documented. Tool handlers return JSON strings or plain text without declared field structures. LLMs cannot know what fields to expect or how to chain results to downstream tools. E.g., 'backlog-read' returns a JSON object with fields like 'topic', 'priority', 'status', 'version', 'created', 'agent', 'session', 'description', 'filepath', 'age', 'isStale', but this structure is not declared in the tool definition.
Parameter descriptions lack actionable detail. E.g., 'status' param in backlog-read simply says 'Filter backlog items by status' but does not explain what status values mean or when to use each. 'action' enum in backlog-write lists 7 values but does not explain prerequisites (e.g., can you approve without first submitting?).
Destructive operations (backlog-prune with clear action) lack confirmation or dry-run guidance. 'dryRun' parameter exists but tool description does not highlight that 'clear' is irreversible and dangerous. No recovery guidance if user accidentally invokes it.
Error messages are generic and not actionable. E.g., 'Backlog item not found: <topic>' does not suggest alternatives (search_backlog?) or next steps. 'Unknown action: <action>' in backlog-write does not list valid actions. No guidance on retryable vs fatal errors.
No documentation of idempotency or retry safety. backlog-write 'create' action will fail if item exists, but backlog-write 'amend' archives and recreates, this is not idempotent. LLMs cannot safely retry without risking duplicates or data loss.
Tool composition and chaining unclear. E.g., backlog-ticket-done 'validates dependencies' but the input schema does not include a dependencies field, and the description does not explain what dependencies are or how to satisfy them. This breaks tool-chaining patterns.