This server has 15 tools with significant definition gaps. Most tools lack proper input schemas (only 6 of 15 have visible schema definitions with typed parameters). Descriptions are present but often generic and under 100 chars, failing to explain WHEN to use each tool or what it returns. No output schemas are documented. Parameter descriptions are sparse or missing entirely. Error handling guidance is absent. The server reads like an API wrapper rather than an LLM-optimized tool suite. No tool annotations (readOnlyHint/destructiveHint) despite clear risk classifications in the metadata.

• Add output schemas for all tools. Document the structure of responses (e.g., list_eval_cases should return {cases: [{id, name, created_at}], total: number}).
• Rename 'audit.log' to 'post_audit_log' or 'create_audit_log' to follow verb_noun convention.
• Expand parameter descriptions to include valid values, constraints, and examples. E.g., 'mode' in get_file_variants: 'Mode of operation: diff (show changes), full (complete file), or summary (overview).'
• Add error recovery guidance to stateful tools. E.g., run_skill_pipeline: 'If the pipeline fails, call get_skill_pipeline_status(run_id) to check logs, then retry or contact support.'
• Annotate tools with readOnlyHint and destructiveHint. Mark list_*, get_*, benchmark_* as readOnly=true; mark create_*, ignore_*, run_skill_pipeline as destructive=true.