Ask a human before your MCP tool does something irreversible — elicitation on both protocol revisions, with a two-call confirmation token for clients that cannot be asked
Two tools with complete schemas and descriptions, but descriptions are minimal (under 50 chars) and lack context about the approval/confirmation flow. Tool names are clear action verbs (delete_, archive_). Parameters are well-typed with descriptions. However, descriptions do not explain WHEN to use these tools, what the confirmation token mechanism is, or how the two-call fallback works, critical context for LLM selection. Output schemas are not documented. Error handling is present but not described in tool definitions.
Archives a thing after asking.
Deletes things after asking.
Tool descriptions are too brief (under 50 chars) and lack context. 'Deletes things after asking' does not explain the approval flow, when to use this tool vs direct deletion, or how the confirm_token parameter works. LLMs cannot infer the two-call confirmation pattern from the description alone.
Output schemas are not documented. The tool definitions show input schemas but do not specify what delete_things or archive_thing return. LLMs cannot plan downstream actions or extract confirmation tokens without knowing the response structure.
Parameter descriptions lack actionable detail. 'Optional confirmation token from a previous call' does not explain how to obtain the token, when it expires, or what happens if it is invalid. The resourceKey parameter is not documented in the tool definitions, only in the ApprovalRequest interface.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 54 | 2026-07-28+ | v2 |
No error handling guidance in tool descriptions. The approval flow can return 'declined', 'rejected', or 'pending' outcomes, but tool definitions do not explain these states or what the LLM should do next. Error recovery is opaque.
Tool definitions do not declare destructive/reversible risk levels. The rubric notes delete_things as DESTRUCTIVE and archive_thing as REVERSIBLE, but the tool schemas lack toolAnnotations (destructiveHint/idempotentHint) to signal this to clients.