Server has 20 tools with explicit schemas and descriptions. Naming follows verb_noun pattern (project_get_active, code_read_current, ui_create). Descriptions are present (avg ~120 chars) but many lack actionable context for LLM selection. Schemas use Zod with type constraints (enums, min/max), but output schemas are undocumented. Tool annotations (readOnlyHint, destructiveHint) are correctly applied. Major gaps: no error recovery guidance, missing parameter descriptions in complex tools (code_batch, ui_apply_operations), and no pagination/result limits despite tools returning potentially large datasets.

• Document output schemas for all tools. For code_read_current, explicitly define: { cells: Array<{row, column, kind, text}>, revision: string, cellCount: number }.
• Add error recovery guidance to code_apply_current, code_insert_unit, and code_batch: 'If revision conflict occurs, call code_read_current to fetch the latest revision, then retry.'
• Replace discriminatedUnion in code_batch and ui_apply_operations with explicit operation type definitions. Document each operation variant (e.g., 'insert', 'update', 'delete') with required fields.
• Add pagination to code_read_current and code_read_range: accept 'offset' and 'limit' parameters, return 'totalRows' and 'hasMore' in response.
• Document the control schema for ui_upsert_control by referencing controlSchema from schema.ts in the description, or inline key properties: 'control: {id, type, x, y, width, height, ...}'