17 tools with complete schemas and descriptions, but significant gaps in error guidance, parameter constraints, and composition clarity. All tools have descriptions (10-150 chars, mostly adequate) and input schemas with typed parameters. However, many descriptions lack actionable context for LLM selection. Tier-gated architecture is well-designed but not reflected in tool annotations. No enum constraints on status/type parameters. Error handling returns codes but minimal recovery guidance. Output schemas are implicit rather than documented.

• Add enum constraints to all status/type parameters. E.g. 'status' in get_ticket_status should declare valid values: ['open', 'in_progress', 'resolved', 'closed']. This prevents LLM hallucination and makes the tool self-documenting.
• Enhance error messages with recovery guidance. Instead of returning {code: 'UNKNOWN_SESSION'}, return {code: 'UNKNOWN_SESSION', message: 'This session is not signed in. Call open_session(session_id, employee_id) first.'}. This enables LLMs to self-correct.
• Add tool annotations (readOnlyHint, destructiveHint, idempotentHint) to expose the tier-gated architecture. Mark Tier 0 tools as readOnlyHint=true, Tier 1 write tools as destructiveHint=true, and Tier 2 tools as requiring human handoff via description.
• Document output schemas explicitly. For each tool, add a 'Returns' section in the description or a separate schema field. E.g. lookup_policy returns {question: string, passages: [{document: string, section: string, text: string, score: float}]}.
• Clarify backend-only tool usage. For open_session, confirm_proposal, and reset_demo, explain the control_key authentication and when an agent should call them (e.g. 'Called by the backend to bind a session to an employee; agents should not invoke this directly').