mx-data

Use mx_data.py to query Eastmoney MX data through natural language and produce structured Excel/JSON outputs.

Trigger Boundaries

Use for exact data questions: price, close/open/high/low, volume, turnover, PE/PB, ROE, revenue, profit, balance sheet items, dividends, shareholders, executives, company profile, sector/index/fund quotes.
Prefer this skill whenever the answer depends on current or historical market data rather than model memory.
Do not use it for event interpretation, news causality, announcements, research reports, policy text, or market narrative; use mx-search first for those.
Do not use it to screen a universe by constraints such as "ROE > 15%" or "PE < 20"; use mx-xuangu.

Required: a clear natural-language query with entity, metric, and time range or date.
Ask a concise follow-up if the entity, metric, or period is ambiguous and the ambiguity changes the result.
Use narrower periods for daily-level data. Multi-year daily queries can create large Excel/JSON files and overload context.
Require MX_APIKEY in the environment. Optional: set MX_OUTPUT_DIR; otherwise output goes to ~/.codex/skills-output/mx_data/output.

Normalize the user request into one focused query. Include the entity code/name, metric, market if needed, and date range.
Run:

python /Users/lu/.codex/skills/mx-data/mx_data.py "贵州茅台近三年净利润 营业收入"

Inspect the terminal preview first. Open generated Excel/JSON only when deeper parsing is needed.
If the raw JSON shape matters, load references/result-fields.md.
Answer with the data source, query used, time range, key values, and generated file paths when files were created.

State that the data came from Eastmoney MX via mx-data.
Include units, dates, and whether values are real-time, daily, annual, or another granularity.
For tables, summarize the important rows/columns and point to the generated .xlsx and _raw.json.
Do not turn raw data into investment advice unless the user explicitly asks for analysis; even then, distinguish data from judgment.

Missing MX_APIKEY: ask the user to configure it; do not invent values.
Empty result: broaden entity/metric wording or reduce condition specificity.
Large output: rerun with a narrower time range or fewer metrics.
API limit/auth errors: report the code and stop.
Network/JSON errors: retry once only if the failure looks transient; otherwise report the failure and preserve any raw output path.

Use evals/evals.json after changing this skill or its script.
Positive cases should trigger mx-data; adjacent news, screening, watchlist, and trading prompts should not.