TailorAgent exposes 13 tools (with one duplicate 'skill' entry) across file operations, shell execution, search, and knowledge management. Most tools have descriptions in Chinese, but quality is inconsistent. Schemas are present for most tools with typed parameters, but descriptions lack depth and actionable guidance. Critical gaps: no error handling guidance, no output schemas documented, no tool annotations (readOnlyHint/destructiveHint), and destructive tools (bash, edit_file, write_file) lack confirmation patterns. File tools have reasonable parameter descriptions, but shell tools are under-documented. Knowledge and memory tools lack clarity on return structures.

• Add tool annotations: mark read_file, glob, grep, knowledge_search, get_current_time as readOnlyHint=true; mark bash, bash_output, edit_file, write_file, kill_shell as destructiveHint=true; mark get_current_time, read_file as idempotentHint=true.
• Document output schemas for all tools. E.g., bash returns {stdout: string, stderr: string, exitCode: number}; knowledge_search returns {results: [{id, content, score}], total: number}; memory returns {success: boolean, value?: string}.
• Expand bash/bash_output descriptions to 100+ chars: explain timeout behavior, when to use each variant, and that bash is destructive. Add error guidance: 'If command fails, check exit code and stderr. Retry transient failures; for permission errors, verify user context.'
• Add confirmation pattern for destructive file operations: edit_file and write_file should support a dry_run parameter that returns the diff without applying changes.
• Remove duplicate 'skill' tool entry and consolidate into single definition with clear parameter validation.

Inferred effective spec: <=2025-11-25.

• Relies on HTTP+SSE-only transport (deprecated) - migrate to Streamable HTTP