The run_mcts tool has a critical structural flaw: its input schema declares 'possible_actions_fn' and 'evaluate_state_fn' as type 'callable', which is not a valid JSON Schema type. JSON Schema does not support callable/function types, these cannot be serialized, validated, or passed over JSON-RPC. The tool description is minimal (67 chars) and lacks guidance on when to use it or what the output structure contains. Parameters lack descriptions entirely. The schema is fundamentally broken for MCP transport.

• Replace 'callable' type with a proper JSON Schema definition. Since functions cannot be serialized, redesign the tool to accept state/action/reward data as structured inputs (e.g., a list of possible actions and their known rewards) rather than callback functions. Alternatively, split into separate tools: one to list possible actions, one to evaluate a state.
• Expand tool description to 100-200 characters. Explain: 'Executes Monte Carlo Tree Search to find the optimal next action in a multi-step planning problem. Use when you need to explore action sequences and balance exploration vs exploitation. Returns the best action, its estimated reward (q_value), and visit statistics.'
• Add descriptions to all parameters: initial_state (string format, e.g., 'problem description or state representation'), iterations (integer, range 10-1000, default 50), possible_actions_fn (redesign as structured input), evaluate_state_fn (redesign as structured input).
• Document the output schema explicitly: 'Returns {best_action: string|null, q_value: float (0-1), visits: integer, total_candidates: integer}. q_value is the estimated reward; visits is the number of times this action was explored.'
• Rename tool to 'plan_with_mcts' or 'search_optimal_action_mcts' to clarify intent and follow verb_noun convention.