Two tools with basic schemas but significant quality gaps. powershell_command has proper parameter types and descriptions, but returns unstructured text strings instead of structured objects, violating pattern:response-shaper. get_powershell_info lacks parameter descriptions (empty input schema) and returns raw JSON wrapped in emoji strings. Both tools lack error categorization and recovery guidance. Descriptions are present but generic. No tool annotations (readOnlyHint/destructiveHint). powershell_command is high-risk (WRITE) but has no confirmation or dry-run pattern.

• Refactor responses to return structured JSON: {success: bool, stdout: string, stderr: string, return_code: int, command: string} instead of emoji-prefixed strings. LLMs parse structured data reliably.
• Add error classification: return {success: false, error: {type: 'retryable'|'user_fixable'|'fatal', message: '...', suggestion: '...'}} so agents know whether to retry, ask user, or abort.
• Add dry-run support to powershell_command: new parameter 'dry_run: bool' that uses PowerShell -WhatIf flag. Prevents accidental destructive execution.
• Translate get_powershell_info description to English and expand: 'Retrieve PowerShell version, OS info, and runtime details. Use this to verify environment before executing commands.' Add readOnlyHint=true annotation.
• Add timeout and encoding parameters to get_powershell_info for consistency, or document why they differ from powershell_command. Currently inconsistent interface.