Both tools have minimal descriptions (under 100 chars) and lack parameter-level documentation. Input schemas are present but sparse: 'weights' and 'sample_gradients' have only type and description at the top level, with no constraints, ranges, or validation guidance. No output schema is documented. The tool names lack action verbs (register_task, compute_penalty are weak; should be 'freeze_task_weights' or 'calculate_ewc_penalty'). No error handling, no recovery guidance, no idempotency markers. This is a proof-of-concept implementation, not production-grade.

• Rename tools with clear action verbs: 'freeze_task_weights' (instead of 'register_task') and 'calculate_ewc_penalty' (instead of 'compute_penalty'). This helps LLMs infer intent from the name alone.
• Expand tool descriptions to 150 - 250 characters. Include WHEN to call (e.g., 'Call after training a task to completion'), WHAT it does (e.g., 'Freezes weights and estimates Fisher importance'), and WHAT it returns (e.g., 'Returns penalty value and per-weight gradients'). See pattern:tool-description.
• Add parameter constraints: 'weights: dict of float values in range [-1e6, 1e6], at least 1 key' and 'sample_gradients: list of 1 - 10000 gradient dicts, each with same keys as weights'. Specify what happens if constraints are violated.
• Document output schemas explicitly. For 'freeze_task_weights', return {status: string, total_tasks: int, fisher_diagonal: dict}. For 'calculate_ewc_penalty', return {ewc_penalty: float, penalty_gradients: dict, num_consolidated_tasks: int}.
• Add error handling: return {error: string, recovery_hint: string} when inputs are invalid. E.g., 'sample_gradients is empty, provide at least 1 gradient dict per sample' or 'weights keys do not match prior task, ensure all prior task keys are present'.