Two tools with minimal documentation and incomplete schemas. register_task and compute_penalty have basic descriptions (60-80 chars) but lack parameter descriptions entirely. Input schemas present but parameters lack type information beyond object/array containers. No output schema documented. No error handling guidance. Tool names lack action verbs (register_task vs. freeze_task_weights; compute_penalty vs. calculate_ewc_penalty). No indication of idempotency, destructiveness, or recovery paths. Composition is reasonable (two distinct concerns) but parameter naming is generic (weights, sample_gradients, w, g).

• Add per-parameter descriptions and type constraints. E.g., 'weights: dict of parameter names (str) to float values. Example: {"w_layer1": 1.5, "w_layer2": -0.8}. All keys must be strings; all values must be floats.'
• Rename tools with clear action verbs: 'freeze_task_weights' (register_task), 'calculate_ewc_penalty' (compute_penalty). Verb-noun naming helps LLMs parse intent.
• Document output schemas formally. E.g., register_task returns {status: string, total_tasks: integer, fisher_diagonal: dict}. compute_penalty returns {ewc_penalty: float, penalty_gradients: dict}.
• Add error handling with recovery guidance. E.g., 'If sample_gradients is empty, provide at least one gradient dict per sample.' 'If weight keys don't match prior tasks, only penalize overlapping keys.'
• Add tool annotations in MCP schema: register_task with destructiveHint=true (modifies state), compute_penalty with readOnlyHint=true (no side effects).