| name | pandadata-warehouse |
| description | Build, repair, query, and maintain a local Pandadata market-data warehouse backed by DuckDB and Parquet for repeat quant research, backtests, factor studies, and batch reports. Use when the user asks to cache Pandadata or panda_data results locally, create or update a local market-data database layer, perform incremental data refreshes, query downloaded data with DuckDB SQL, design table partitions and watermarks, validate local data against Pandadata APIs, or avoid repeated Pandadata API calls. |
| license | GPL-3.0-only |
| metadata | {"organization":"QuantSkills","organization_url":"https://github.com/quantskills","repository":"skill-pandadata-warehouse","repository_url":"https://github.com/quantskills/skill-pandadata-warehouse","project_type":"skill","collection":"pandadata-warehouse","creator":"abgyjaguo","creator_url":"https://github.com/abgyjaguo","maintainer":"abgyjaguo","maintainer_url":"https://github.com/abgyjaguo"} |
| quantSkills | {"project_type":"skill","category":"data-api","tags":["pandadata","warehouse","duckdb","parquet","data-engineering"],"platforms":["claude-code","codex","openclaw","cursor"],"status":"stable","requires":["skill-pandadata-api"],"validation_level":"runnable","maintainer_type":"community","summary_zh":"Pandadata 本地数据仓库：用 DuckDB 与 Parquet 缓存、增量刷新、查询和校验行情数据，减少重复 API 调用。","summary_en":"Pandadata warehouse skill for caching, refreshing, querying, and validating local DuckDB and Parquet market-data stores."} |
{
"version": 1,
"task": {
"placeholder": "补充数据表、接口、分区、SQL 查询、校验或修复要求"
},
"fields": [
{
"key": "operation",
"label": "仓库操作",
"type": "select",
"default": "refresh",
"options": [
{ "value": "initialize", "label": "初始化仓库" },
{ "value": "refresh", "label": "增量刷新" },
{ "value": "query", "label": "SQL 查询" },
{ "value": "validate", "label": "数据校验" },
{ "value": "repair", "label": "修复仓库" }
]
},
{
"key": "asset_class",
"label": "资产类别",
"type": "select",
"default": "stock",
"options": [
{ "value": "stock", "label": "A 股" },
{ "value": "index", "label": "指数" },
{ "value": "future", "label": "期货" },
{ "value": "option", "label": "期权" },
{ "value": "hk_us", "label": "港美股" },
{ "value": "factor", "label": "研究因子" }
]
},
{
"key": "universe",
"label": "标的范围",
"type": "textarea",
"placeholder": "代码列表、指数成分或明确的下载范围；分钟数据必须限定范围"
},
{
"key": "start_date",
"label": "开始日期",
"type": "date"
},
{
"key": "end_date",
"label": "结束日期",
"type": "date"
},
{
"key": "warehouse_path",
"label": "仓库路径",
"type": "text",
"placeholder": "留空则使用 ~/.pandadata/warehouse"
}
],
"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}}请对 Pandadata 本地仓库执行 {{operation}}，资产类别为 {{asset_class}}。{{#universe}}标的范围：{{universe}}。{{/universe}}{{#start_date}}开始日期：{{start_date}}。{{/start_date}}{{#end_date}}结束日期：{{end_date}}。{{/end_date}}{{#warehouse_path}}仓库路径：{{warehouse_path}}。{{/warehouse_path}}确认真实 API 契约，以分区 Parquet 保存原始字段并通过 DuckDB 暴露视图；优先增量、保持可恢复，校验覆盖范围、重复键、行数与 API 抽样，任何破坏性修复先列出影响并征得确认，输出中文报告。"
}
Pandadata Warehouse
Use this skill to design and operate a local Pandadata warehouse for heavy data reuse. Keep the warehouse faithful to Pandadata source fields, store raw data in partitioned Parquet, expose query views through DuckDB, and refresh only the missing trading periods unless a table requires a full partition rewrite.
Workflow
- Clarify the user's target scope: asset class, table family, symbol universe, date range, frequency, warehouse path, and whether the request is initialize, refresh, query, validate, or repair.
- Load
references/warehouse-playbook.md before designing tables, writing refresh code, rebuilding partitions, or deciding validation rules.
- Use the
pandadata-api skill or local Pandadata documentation to confirm exact panda_data.get_* method names, parameters, response fields, credentials, and date formats before writing or running API calls.
- Keep raw warehouse data close to the API contract. Preserve source column names and types when practical; place derived calculations in downstream analysis views or reports.
- Store Parquet under a stable warehouse root, defaulting to
~/.pandadata/warehouse, and keep metadata such as source method, partition keys, date coverage, row counts, refresh time, and validation status in _meta.json or an equivalent manifest.
- Create or refresh DuckDB views that read the Parquet partitions with
read_parquet, then run downstream analysis through SQL against the local view layer.
- Validate each material refresh by checking row counts, date ranges, duplicate keys, partition coverage, and a small sample against fresh Pandadata API results.
- Report any missing credentials, unavailable SDK, failed partitions, stale datasets, or destructive repair actions clearly before continuing.
Supported Warehouse Families
Use this skill for these local table families unless the user narrows the scope:
| Family | Typical source methods | Default partition |
|---|
| Trading calendar | get_trade_cal, get_last_trade_date | exchange |
| A-share daily bars | get_stock_daily and adjusted daily variants | year |
| A-share minute bars | get_stock_min | symbol and month |
| Index bars | get_index_daily, get_index_min | year or month |
| Futures bars | get_future_daily, adjusted futures daily variants, get_future_min | year or month |
| Options bars | get_option_daily | year |
| HK/US daily bars | get_hk_daily, get_us_daily | year |
| Adjustment factors | get_adj_factor | year |
| Research factors | get_factor | factor name and year |
Treat the method list as a routing map, not an API contract. Confirm the exact method details with pandadata-api before coding.
Core Rules
- Prefer incremental refreshes keyed by the latest completed trading date. Use full partition rewrites only for datasets whose history can change, such as forward-adjusted prices.
- Do not run broad minute-bar or all-market downloads unless the user explicitly supplies a symbol universe or confirms the expected size.
- Never delete, overwrite, or rebuild existing partitions silently. First list the affected files, explain why the rebuild is needed, and get explicit confirmation unless the user already authorized that exact destructive action.
- Keep failed downloads resumable. Record failed partitions and leave successful partitions intact.
- Make freshness visible in answers and generated code: include source method, last local date, latest API/trading date checked, and any stale-data warning.
Maintainer, License, and Limits
- Created and maintained by
abgyjaguo for the QuantSkills community in the quantskills/skill-pandadata-warehouse repository.
- Licensed under GNU GPL v3.0 only (
GPL-3.0-only); keep the root LICENSE file with the full GPLv3 text when publishing or packaging.
- Treat outputs as data engineering and research support, not investment advice. Do not promise returns, profitability, production readiness, certification, or official endorsement.
- Keep credentials, API tokens, private datasets, and confidential material out of examples, metadata, logs, warehouse files, and generated reports.
Runtime Compatibility
- Codex: use this
SKILL.md directly and agents/openai.yaml for UI metadata.
- Cursor: use
agents/cursor-rule.mdc as the project rule/loader.
- Claude Code, Hermes, and OpenClaw: use
agents/portable-loader.md to point the runtime at this skill root and require it to read SKILL.md first.
- Keep runtime-specific files as thin adapters. Update this file first when behavior changes, then sync the adapters so they do not contradict the canonical workflow.
