Single tool 'get_company_snapshot' has a reasonable name and description, but critical gaps undermine quality. The input schema is present with proper type and description for the 'symbol' parameter. However, the output schema is completely undocumented, the function returns a complex nested dict with 'meta', 'financials', 'valuation', 'market_data', and 'coverage' fields, but no schema is provided to the MCP server definition. Error handling is minimal: exceptions are caught but responses lack actionable recovery guidance. The tool accepts both formats ('600519' and 'sh.600519') but this flexibility is not documented in the parameter description. No pagination, no rate limiting, no per-field error reporting. The implementation calls external Baostock API but exposes no timeout, retry, or partial-failure handling to the agent.

• Add explicit output schema to MCP tool definition documenting all returned fields: meta (symbol, as_of, data_source), financials (dict), valuation (dict), market_data (dict with daily_price, 52w_high, 52w_low, latest_price), coverage (dict with missing_fields array).
• Enhance parameter description to document both accepted formats: 'Stock symbol in format sh.XXXXXX (Shanghai) or sz.XXXXXX (Shenzhen). If only 6-digit code provided (e.g. 600519), exchange is auto-detected: codes starting 6 or 9 → Shanghai, 0 or 3 → Shenzhen.'
• Add error classification to responses: return structured error objects with category (retryable|user_fixable|fatal), message, and recovery hint. E.g. 'Baostock login failed (retryable): service temporarily unavailable. Retry in 30 seconds.'
• Implement timeout (e.g. 30s) for Baostock API calls and return clear timeout error with retry guidance instead of hanging.
• Return per-field success/failure in coverage.missing_fields with reasons (e.g. 'price_series: no data for date range', 'valuation: no profit data available for recent quarters'). This lets the LLM understand data quality and decide whether to proceed.