Vestibule defines 2 tools with complete input schemas and detailed descriptions (194 - 250 chars each). Both tools have proper enum constraints, timeout bounds, and path validation. Naming is clear (run_code, read_workspace). However, output schemas are not documented, LLMs cannot predict return structure. Error handling is present but generic ('Internal error'). No tool annotations (readOnlyHint, destructiveHint). Security is strong (no secrets in params, path traversal prevention), but composition is minimal (only 2 tools, no batch variants). Descriptions are LLM-optimized and include prerequisites and constraints, meeting the 50 - 200 char baseline.

• Document output schema for both tools. E.g., run_code returns {type: 'text', text: string, exit_code?: int, stderr?: string}. read_workspace returns {type: 'text', text: string} with directory listing as newline-delimited or JSON array.
• Add tool annotations: run_code with destructiveHint=true (modifies /workspace), read_workspace with readOnlyHint=true. This signals to agents which tools have side effects.
• Enhance error messages with recovery guidance. E.g., 'Execution timeout (10s). Try: reduce code complexity, increase timeout_seconds (max 60), or split into smaller runs.'
• For read_workspace, document directory listing limits. E.g., 'Lists up to 1000 entries; use pagination or filtering for large directories.' Implement a limit parameter if needed.
• Clarify run_code output format in description: 'Returns stdout/stderr combined as text. Exit code 0 = success; non-zero = error. Exceptions are caught and returned as text.'

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

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