23 tools with consistent naming (verb_noun pattern) and complete input schemas loaded from JSON files. However, descriptions are uniformly terse (10-80 chars, well below the 194-char baseline), parameter descriptions are minimal or absent, and output schemas are not documented. Tools follow a repetitive pattern (workbookPath + workbookContent + outputPath + outputMode) but lack guidance on when to use each export/report variant. No error handling documentation, no recovery guidance, and no indication of which operations are idempotent or destructive beyond the Risk field.
Descriptions are uniformly terse (10-80 chars vs. 194-char baseline). LLMs cannot distinguish between 8 export/report tools with similar names and minimal guidance on when to use each variant.
Parameter descriptions are missing or minimal. 'outputMode' enum values ('path', 'content', 'base64') lack explanation of when to use each. 'workbookPath' vs 'workbookContent' dual parameters lack guidance on mutual exclusivity or precedence.
Expand tool descriptions to 100-200 chars. For export tools, explain the output format and use case: 'Exports a project overview projection (JSON) suitable for high-level status reporting. Use this to generate executive summaries.' Distinguish between project_overview_view, task_edit_view, and phase_detail_view with concrete examples.
Add parameter descriptions for all enum values. Document outputMode: 'path: write to file system; content: return as string; base64: return as base64-encoded binary (for XLSX/binary formats).' Clarify workbookPath vs workbookContent: 'Provide either workbookPath (file path) or workbookContent (JSON string); if both provided, workbookContent takes precedence.'
Document output schemas in tool descriptions or via a separate schema reference. Example: 'Returns {ok: boolean, operation: string, diagnostics: Diagnostic[], stdout?: string, stderr?: string, exitCode: number}.' This enables LLMs to extract and chain results.
Add error handling guidance to descriptions. Example: 'On invalid patch: returns {ok: false, diagnostics: [{message: "Patch validation failed: ..."}]}. Retry with corrected patch or call mikuproject_ai_validate_patch first to debug.'
Expose Risk metadata as MCP tool annotations. Add readOnlyHint: true for READ_ONLY tools; destructiveHint: true for WRITE tools. This enables agents to reason about retry safety and side effects without parsing descriptions.
Output schemas are not documented. Tools return structured results but LLMs cannot see what fields to expect, forcing them to guess at response structure and breaking downstream tool chaining.
No error handling documentation. Tools do not indicate which errors are retryable, what recovery steps are available, or how to interpret failure modes. No guidance on invalid inputs or constraint violations.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Risk field exists in metadata but is not exposed via MCP tool annotations, preventing agents from understanding side effects and retry safety.