Two tools with clear action verbs and reasonable descriptions (194 - 200 chars, within baseline). Both have input schemas with types and descriptions. However, output schemas are undocumented (only string returns visible), parameter descriptions lack format/constraint details (e.g., no regex for code validation, no enum for env_vars structure), and error handling is minimal, responses are free-text strings rather than structured objects with recovery guidance. Tool annotations present but incomplete (missing readOnlyHint/destructiveHint details for pip_install). No pagination, no per-item error reporting for batch operations.

• Define structured output schemas for both tools. python_run should return {status: 'success'|'failed'|'timeout', elapsed_s: number, stdout: string, stderr: string}. pip_install should return {success: boolean, packages: string[], elapsed_s: number, error?: string}.
• Add format/constraint details to parameter descriptions. For 'code': 'Python 3.9+ source code (max 50KB, no shell escapes or file I/O to /etc)'. For 'packages': 'Array of pip specifiers (1 - 20 items, e.g., ["numpy", "pandas>=2"])'. For 'env_vars': 'Object with string keys and string values (max 10 vars, no PATH/LD_LIBRARY_PATH overrides)'.
• Implement error categorization. Return structured errors: {error: string, category: 'retryable'|'user_fixable'|'fatal', suggestion?: string}. E.g., timeout → 'retryable: increase timeout param', invalid syntax → 'user_fixable: check code for syntax errors'.
• Fix tool annotations. Set python_run destructiveHint=true (code execution is destructive). Set pip_install destructiveHint=true (modifies environment). Add readOnlyHint=false to both (both modify state).
• Add optional dry_run parameter to python_run (default false) to preview execution without side effects. For pip_install, add a check_only parameter to validate packages without installing.