Sandy provides 6 tools with mixed quality. Tool definitions are visible and mostly include schemas, but descriptions are inconsistent in depth and usefulness for LLM selection. Naming follows verb_noun convention (sandy_*, prime), which is good. However, schemas lack comprehensive parameter descriptions, and several tools have vague or missing guidance on error conditions and prerequisites. The server implements a specialized sandbox execution domain (TypeScript in Docker + AWS SDK access), which is niche and well-scoped, but tool composition and chaining could be clearer.

• Rename 'prime' to 'get_sandbox_capabilities' or 'get_skill_definition' to follow verb_noun convention and clarify its purpose for LLMs.
• Expand sandy_run description to explain: (1) How script paths are resolved relative to the session directory; (2) How 'args' are passed to the script (CLI args? environment?); (3) What the return value contains (stdout, stderr, exit code?); (4) Typical execution time and timeout behavior.
• Document sandy_check's 'baseline' vs. 'connect' actions: What does each check? When should the LLM call one vs. the other? What constitutes success/failure for each?
• Add comprehensive error handling documentation to all tools. For each tool, document: (1) Common failure modes (resource not found, timeout, permission denied); (2) Actionable recovery steps for the LLM (e.g., 'If session not found, call sandy_create_session'); (3) Categorize errors as retryable or user-fixable.
• Document the expected workflow and tool chaining: Add a 'Getting Started' section in descriptions explaining typical sequences (e.g., 'Typically: (1) sandy_create_session → (2) sandy_run → repeat (2) as needed').
• Add JSON schema documentation for tool responses. For sandy_create_session and sandy_resume_session, explicitly document the response structure (e.g., { "sessionId": string, "scriptsPath": string, "createdAt": ISO8601 }).
• Clarify parameter relationships. For sandy_check, document that 'imdsPort' is required when action='connect' and ignored when action='baseline'.
• For sandy_image, add expected resource requirements and timing in the description (e.g., 'Image creation typically takes 5-10 minutes; uses 2GB disk space').

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

• Relies on Logging (deprecated) - log to stderr or use OpenTelemetry