QuantGPT provides 8 tools with complete input schemas and descriptions, but quality is inconsistent. Tool names follow verb_noun convention (list_, validate_, run_, score_, diagnose_). Descriptions are present but vary in clarity, some are terse (list_operators: 16 chars), others adequate (run_backtest: 47 chars). Critical gap: NO output schemas documented for any tool. Parameters have types and enums where appropriate (e.g., run_backtest universe enum), but descriptions lack actionable constraints (e.g., date format YYYY-MM-DD stated but no validation guidance). Error handling is absent, no recovery guidance, no categorization of retryable vs fatal errors. Tools are READ_ONLY, reducing risk, but composition is weak: tools like score_factor, diagnose_factor, run_anti_overfit, run_rolling_validation accept identical parameter sets, suggesting potential consolidation. No tool annotations (readOnlyHint, idempotentHint) present despite all being read-only.

• Document output schemas for all tools. Specify fields, types, and structure. E.g., run_backtest returns {report_url: string, metrics: {sharpe_ratio: number, max_drawdown: number, ...}, status: 'success'|'error'}.
• Expand descriptions for list_operators and list_universes to 50-100 chars. Explain what structure they reveal and when to call them. E.g., 'Returns all supported operators and their syntax. Call this first to understand available factor expression building blocks.'
• Consolidate score_factor, diagnose_factor, run_anti_overfit, run_rolling_validation into a single analyze_factor tool with an 'analysis_type' enum parameter (score|diagnose|anti_overfit|rolling_validation). Reduces cognitive load and clarifies intent.
• Add error handling to all tools. Return structured errors: {error: string, code: 'invalid_expression'|'data_unavailable'|'timeout', recovery: 'Try validate_expression first' | 'Retry with smaller date range'}.
• Add tool annotations to all tools. Mark all as readOnlyHint=true since they perform no mutations. This signals to agents that calls are safe to retry.

Inferred effective spec: 2026-07-28+.

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