Block Trade Radar

WorkBuddy PandaData 数据源覆盖

本技能在 WorkBuddy 专家包中运行时，所有实时和历史金融数据必须来自已连接的 pandadata Connector。若下文、参考资料或脚本提到 Python panda_data SDK、 AkShare、Tushare、网页抓取、直连 HTTP 或本地凭证，以本节为准：不得使用这些 方式获取正式数据。

1. 数据型任务必须先完成 auth_status 和至少一次真实的 call_pandadata 业务调用；在收到 Connector 返回前，禁止输出分析、排名、数字结论或“无数据”。本 Skill 的流程不得绕过 主 Agent 的最低接口调用清单。
2. 先按本 Skill 的 references/pandadata-interface-contracts.md 选择已登记方法和参数。只要任务能映射到已登记接口， 就直接通过 call_pandadata 传入该业务方法和 params；常规调用前不得执行接口检索。
3. 仅当已登记方法被 Connector 明确报告为参数契约不兼容、字段契约变化或调用失败时， 才对该方法调用一次 get_method_doc，修正参数后最多重试一次。
4. 仅当本地接口表没有匹配项，或 Connector 明确报告方法不存在/不受支持时，才调用 search_methods 动态发现接口；不得靠猜测连续试用名称相近的 get_* 方法。
5. 只有 call_pandadata 实际返回 0 行时才允许写“无数据”。必须先完成一次复查调用：校验 最新交易日与代码格式，放宽日期窗口，移除非必填过滤条件，或使用登记的备用参数；仍为 0 行才可如实报告，并保留两次调用回执。0 行不得触发 search_methods。
6. 不向 call_pandadata 添加未登记参数或顶层行数限制；记录实际方法、参数、数据日期、 频率、复权口径、行数、空值和错误状态。
7. 包内脚本只可处理 Connector 已返回的数据或执行纯本地计算与校验，不得自行联网取数。

最终答案必须包含“数据调用回执”表：接口、实际参数、状态、行数、数据日期范围和关键字段。 缺少回执表示任务未完成，必须继续调用工具而不是结束回答。

权限不足、配额限制、空结果、延迟发布和字段缺失都必须明确披露，不得切换到其他数据源 或用模型推断补数。

{
  "version": 1,
  "task": {
    "placeholder": "补充扫描窗口、行业范围、折溢价阈值或重点席位；留空则按默认近期开窗扫描"
  },
  "fields": [
    {
      "key": "symbol",
      "label": "股票代码（可选）",
      "type": "text",
      "placeholder": "如：600519.SH；留空扫描全市场"
    },
    {
      "key": "date",
      "label": "截至日期",
      "type": "date",
      "help": "留空由前端使用今天，并自动按交易日处理"
    }
  ],
  "prompt_template": "{{#task}}任务与材料：\n{{task}}\n\n{{/task}}{{#attachments}}用户上传的材料（已放入工作区）：\n{{attachments}}\n\n{{/attachments}}请运行 A 股大宗交易雷达{{#symbol}}，聚焦 {{symbol}}{{/symbol}}{{#date}}，截至 {{date}}{{/date}}，计算相对同日收盘价的折溢价率，识别机构专用买卖方向、重复折价接盘与买卖营业部相同的疑似对倒，并按成交额和折溢价排名，输出中文报告。"
}

Block Trade Radar

Use this skill to scan and read A-share block-trade (大宗交易) activity: for the whole market over a window or a single name, join every block trade to the same-day close to compute a discount/premium rate (折溢价率), read whether the buyer or seller is 机构专用 (institutional seat), flag repeated discounted takeovers and same-branch (buyer==seller) prints, and rank names by block-trade amount and by premium/discount. Prefer Pandadata as the data source, keep every figure traceable to get_block_trade (plus the get_stock_daily close used for the discount math) and a trade date, and never invent prices, amounts, discounts, or directions.

Scope And Positioning (read first to avoid overlap)

This skill is the block-trade discount/premium view. It is deliberately distinct from its siblings:

• Unlike market-daily-review (daily whole-market review that merely lists the day's block trades as one line item): this skill is a block-trade-centric analysis — same-day discount/premium math, institutional-direction read, repeat-discount and wash-print flags, and a cross-sectional premium/amount ranking. If the user wants a full end-of-day market review, hand off to market-daily-review.
• Unlike smart-money-profiler (龙虎榜 / 北向 / 两融 "smart money" across 营业部 seat behaviour on the daily公开信息 board): block trades are a separate off-exchange channel with their own discount signal; this skill reads get_block_trade in depth rather than the 龙虎榜 board. Names surfaced here can be cross-checked there.
• Unlike event-risk-alert (per-name watchlist risk monitoring — unlocks, pledges, reductions): a block trade may implement a shareholder reduction but is not itself a filing-level risk event. If the user wants to watch their own holdings for reduction/unlock risk, hand off to event-risk-alert.
• Unlike a-share-stock-dossier (single-name deep due diligence): this skill is a block-trade scan, not a company dossier.

Block Trade Model (read before analysis)

get_block_trade returns one row per block-trade print (成交笔), keyed by symbol + date + sequence_id. A single name on a single day may have several prints; a large reduction is often split into many.

• Discount/premium (折溢价率) — not a returned field; you compute it. 折溢价率 = price / 同日收盘价 - 1, where the close comes from get_stock_daily for the same symbol and date. Negative = 折价 (below market, the common case — a buyer demands a discount to absorb a block); positive = 溢价 (above market, rarer, sometimes signals demand). Always state the close you used.
• Direction / institutional flag — buyer and seller are 营业部/席位 names. 机构专用 on the buyer side = an institution is absorbing the block (接盘); on the seller side = an institution is distributing (出货). Read direction from the seat text verbatim; do not infer identity beyond what the string says.
• Same-branch print (对倒式) — when buyer == seller (same 营业部), the print may be an internal transfer / 过券 rather than a genuine change of beneficial ownership. Flag it; do not count it as directional institutional flow.
• Scale — amount (成交额, 元) and volume (成交量, 股) size each print. price is the block price. Aggregate per name per day for a name-level view; keep print-level detail for the timeline.
• Dates — date is the trade date. A scan over a window is a snapshot of prints in that window.

Workflow

1. Resolve the target: whole-market scan over a window, or a single name's block-trade timeline. Confirm the date window (default a recent trailing window, e.g. last ~30 trading days, for a market scan).
2. Read references/block-trade-playbook.md before the first run in a session. Use it for the routing table, the discount/premium formula and edge cases, direction/wash-print rules, aggregation rules, the report skeleton, empty-data handling, and the QA checklist.
3. Use the registered get_block_trade and get_stock_daily contracts in references/pandadata-interface-contracts.md, then call them directly through call_pandadata. Use search_methods only when no registered method covers the task, and use get_method_doc only after an explicit contract error; do not invent parameters, fields, symbols, or credentials.
4. Collect evidence:
   • Block trades: get_block_trade for the window (empty symbol for whole market; a specific symbol for one name's history).
   • Same-day close for the discount math: get_stock_daily for the traded symbols over the window.
   • Identity & industry: get_stock_detail and get_stock_industry to name prints and roll them up by sector.
   • Calendar: get_last_trade_date / get_trade_cal to bound the window.
5. Compute per print: 折溢价率 against the same-day close; tag buyer/seller institutional direction; flag same-branch prints. Aggregate per name per day (总成交额, 笔数, 加权折溢价率), then rank: largest by amount, deepest 折价, notable 溢价, repeat-discount names, and net institutional 接盘/出货. Keep raw print counts long enough to cite source method, window, and missing-close status.
6. Generate the Markdown report following the skeleton in the playbook. Save to reports/block-trade/<scope>-<date>.md (e.g. reports/block-trade/market-20260706.md) unless the user gives another path.
7. Run scripts/validate_report.py <report-path> after writing. Fix missing sections, missing source notes, a missing discount-basis (close) note, missing wash-print/institutional caveats, missing window/date labels, or a missing disclaimer before presenting the result.

Interface Map

Routing aid only; the exact call contract must still come from pandadata.

Analysis Modes

• Whole-market scan: all prints in the window → 折溢价 distribution, net institutional direction, largest-by-amount and deepest-discount leaders, repeat-discount names, and industry distribution. Separate genuine directional prints from same-branch (buyer==seller) prints.
• Single-name timeline: one ticker's block-trade history — each print's date, discount/premium vs that day's close, buyer/seller direction, and whether prints cluster (e.g. a reduction being worked off in pieces).
• Discount read: a persistent, deepening 折价 with 机构专用 on the seller side may indicate distribution pressure; a shrinking discount or 溢价 with institutional buyers may indicate absorption. State these as relative observations, not signals to act.
• Wash-print screen: surface buyer==seller prints separately and exclude them from the institutional net; they are internal transfers, not a change in beneficial ownership.

Report Rules

• Write in Chinese unless the user requests another language.
• Always state the discount basis. 折溢价率 is derived; name the close (get_stock_daily, same symbol+date) it was computed against. Never present a 折溢价率 without its price basis.
• Never over-read a single print. One large 折价 block is a transaction, not a verdict; aggregate and look for repetition before calling anything a pattern.
• Read direction from buyer/seller text verbatim; 机构专用 is a seat label, not a named institution — do not invent who it is. Flag buyer==seller prints as possible 对倒/过券.
• Mark the scan window and as-of date. Prints accumulate; a scan is a snapshot — state the snapshot date and the trade-date range.
• Separate facts (raw price, amount, volume, seats), derived metrics (折溢价率, per-name aggregates, ranks, net institutional flow), and judgment. Label all derived calculations.
• Treat empty API results as evidence. State "无数据" with the method name and queried window instead of silently omitting a section. If the same-day close is missing for a print, mark 折溢价率 as 待补 rather than fabricating a basis.
• Keep the tone factual and structural. Use "折价接盘", "净接盘/净出货", "可能提示承接/派发意愿" rather than directional calls; never give trading instructions or personalized investment advice.

Automation (optional scheduling)

When the user asks for an automated block-trade radar, create a task that runs on trading days after market close (e.g. after 18:00 Asia/Shanghai) to catch that day's block-trade prints. Make it idempotent: if reports/block-trade/<scope>-<date>.md exists, regenerate and overwrite. Skip non-trading days.

Resource Guide

• references/block-trade-playbook.md: routing table, discount/premium formula and edge cases, direction/wash-print rules, aggregation rules, report skeleton, empty-data handling, and the QA checklist.
• scripts/validate_report.py: checks the report for required sections, source notes, the discount-basis (close) note, institutional/wash-print caveats, window/date labels, and the disclaimer.

Quality Bar

• Every material claim traces to get_block_trade, a trade date, the scan window, and (for 折溢价率) the get_stock_daily close used.
• 折溢价率 is always presented with its price basis; a missing close is marked 待补, never fabricated.
• Institutional direction is read from buyer/seller text; 机构专用 is never named as a specific institution.
• buyer==seller prints are flagged and excluded from the directional net.
• End every report with this disclaimer: 本报告基于公开数据与规则化分析生成，仅供研究参考，不构成任何投资建议。