Critical gaps across all dimensions. Tool descriptions are minimal (10-50 chars, well below the 194-char baseline). Input schemas present but severely underdocumented: parameters lack type definitions and descriptions. No output schemas documented. Parameters like '__runtimeTarget' expose internal implementation details. No error handling guidance, no recovery paths, no validation rules. Tools combine multiple concerns (e.g., present_step handles node reveal, highlighting, zoom, mode, locking, and text in one call). No idempotency guarantees. Naming is inconsistent: some tools use verb_noun (render_graph, apply_graph_patch), others use imperative (highlight_nodes, clear_highlight). No pagination, no batch operations. This server prioritizes internal graph manipulation over user-facing clarity.

• Expand all tool descriptions to 100-200 characters. Include: what the tool does, when to call it (vs similar tools), and what it returns. E.g., 'Render a full runtime graph payload' → 'Render the complete graph visualization with all nodes, edges, and metadata. Call this after loading a new graph or applying major changes. Returns success status and rendered graph ID.'
• Document every parameter with type, constraints, and purpose. E.g., 'node' in add_node should be: 'Node object with required fields: id (string, unique), label (string, 1-100 chars), type (enum: input|process|output), position (object with x, y numbers). Example: {id: "node-1", label: "Process", type: "process", position: {x: 0, y: 0}}'
• Remove '__runtimeTarget' from all tool parameters. Inject it server-side from the MCP session context or environment. This is an internal implementation detail that should never be exposed to the LLM.
• Document output schemas for all tools. E.g., render_graph returns: {success: boolean, graphId: string, nodeCount: number, edgeCount: number, errors?: string[]}. This enables LLMs to chain calls and extract required IDs.
• Split present_step into focused tools: reveal_node (show a node), highlight_nodes (apply color), set_viewport (zoom/pan), set_presentation_mode (lock input, set title), update_explanation (change text). Each tool does one thing, making composition clearer and reducing parameter confusion.