Layer1 Fin Data

Fin Data

Use this skill for read-only structured financial data queries.

Do not use this skill for price forecasts, direct buy/sell recommendations, position sizing, strategy backtests, or raw SQL execution.

User-facing outputs must preserve 证据来源、时间窗口、数据覆盖限制和安全边界；do not turn structured market data into direct investment advice.

Runtime Binding

This skill defines the Fin Data logical source contract; it does not declare a separate user-visible MCP connection.

• In WorkBuddy, call Fin Data tools through the single tongzhou-fin-research Connector. Runtime tool names use fin_data__<tool>, for example fin_data__get_latest_snapshot.
• In Codex and retained legacy clients, the same canonical tool may appear as a bare name or a client-qualified server/tool name. Match the canonical tool suffix and keep the Fin Data source label.
• Do not bypass an available tongzhou-fin-research Connector by selecting a similarly named global or deferred financial tool.

Progressive References

Load only the reference needed for the current question:

• references/market.md: latest price, historical series, index snapshots, market daily reports
• references/screening.md: rankings, filters, counts, limit-up/limit-down, consecutive patterns
• references/entity.md: security lookup, profiles, baskets, constituents, entity links
• references/macro_financial.md: macro indicators, commodities, rates, financial indicators, sector valuation
• references/limitations.md: data coverage, realtime wording, unsupported claims, safety boundaries

Read this skill's references/<name>.md file before filling detailed tool parameters.

For combination queries, read multiple references. Example: "白酒行业今日涨幅前五" may need references/entity.md for basket discovery and references/screening.md for ranking.

Routing

Choose the narrowest tool family first:

The table uses canonical tool suffixes for readability. Under the WorkBuddy Connector, prepend fin_data__ when selecting the actual tool.

Required Rules

• Use MCP tools only. Never generate or pass raw SQL.
• 对每个返回字段分别判定状态：日期、数值、币种、单位或口径缺失时使用字段级 partial，展示已返回字段并明确缺口，不从模型记忆补值。
• 不支持的市场不得替代：工具明确不覆盖证券、市场、频率或指标时标记 unsupported，不得换成同名证券、相关指数、宽基指数或示例数字。
• 超时、服务失败、协议错误和参数校验失败是 error；错误不是空结果。只有调用成功且目标行集为空才是 empty。
• 对可修正的身份、日期或单一参数错误，一个数据任务最多两次恢复；每次只修正一个维度，不循环切换指标、市场或工具。
• Expose only virtual schema/tool names to users, not physical database tables or fields.
• Resolve relative time before calling tools: 今天/昨日/本周/本月/最近N天 must map to explicit dates or a supported relative window; 最近N个交易日 must use trading-day tools/limit=N, not calendar-day date ranges.
• Mention the returned trade_date, report_date, data_timestamp, or date range in the answer.
• Display returned market dates without adding a weekday. Do not infer a weekday, holiday, or non-trading reason from model memory. State trading-day status only when the tool returns explicit trading-calendar evidence.
• If an answer nevertheless pairs an ISO date with a weekday, verify it with a deterministic calendar calculation. If that cannot be verified, omit the weekday; never infer it from model memory. Do not infer a holiday or non-trading reason from the weekday alone.
• In user-visible output, call the source 公开行情数据. Do not expose Gateway/MCP/Connector names, server or tool names, physical tables, archive/cache implementation labels, or raw metadata keys.
• If the returned trading date is different from the user's calendar phrase, state the returned trading date explicitly instead of hiding the mismatch.
• If a metric/tool is not available, say the current atomic capability is not covered. Do not invent data.
• Event-window market reactions are descriptive statistics from close prices. Do not call them strategy backtests, predictions, or investment advice.
• For two or more comparable symbols in the same market, prefer the bounded batch tool instead of repeating the single-symbol tool.
• For common broad indices, use known standard codes directly: 上证指数 000001.SH, 深证成指 399001.SZ, 创业板指 399006.SZ, 沪深300 000300.SH.
• For uncommon securities, baskets, industries, concepts, themes, or macro indicators, search/resolve first instead of guessing IDs. For industry/theme rankings, prefer search_baskets before rank_securities/screen_stocks.
• Keep investment wording descriptive and evidence-based. Do not turn data into deterministic forecasts or direct recommendations.
• If a multi-part user question asks for market data plus another data domain, finish the market-data table/summary and then continue with the other domain before finalizing.
• Do not answer "will query / please wait" after tool use. Continue calling the needed tool or provide the completed answer.
• Treat a non-empty list/table in MCP output as retrieved evidence. Summarize those rows; do not call it an empty result.
• search_security, resolve_entities, and get_security_profile return identity evidence only. They never support a price, return, volume, or market-date claim. For a price request, call a supported numeric market-data tool after identity resolution; if that call does not succeed, output no market number.
• For a Hong Kong latest-available close, call query_data(ticker=<canonical code>, market="hk_stock", granularity="daily", metrics=["close"], limit=5) and select the latest returned trading row. Do not turn a validation or empty-result issue into an authentication/session diagnosis.
• Preserve returned units and currency exactly. If a field has no unit or currency, label that field 单位未返回 or 币种未返回; do not infer 股/手/CNY/HKD from the market and do not manually convert raw values to 万、亿、百分比 or another display unit.
• If metadata.extras.cache_status is stale, label it as a recent fallback and display each returned market date; never call it realtime/current data.
• ETF candidate rankings compare disclosed tracking/name and daily market fields only. They are not fund recommendations and do not contain AUM or bid/ask spread.
• For investment-sensitive wording, unsupported claims, or empty-result caveats, read references/limitations.md.

Output Shape

Prefer compact tables for lists and rankings. Include:

• name/code when applicable
• metric values and units when available
• date/time of the data
• any important limitation from references/limitations.md