| name | kr-bundle-pipeline |
| description | Use when building, validating, or debugging KR snapshot bundles and the daily batch flow.
Trigger on build_kr_snapshot.py, run_daily_batch.py, snapshot.json, manifest.json, summary.json,
GitHub Actions daily-market-data.yml, ci_safe profile behavior, or eit-research bundle loading.
|
KR Bundle Pipeline
Use this skill for the KR snapshot export path that feeds eit-research.
Workflow
- Choose the correct profile:
ci_safe for GitHub-hosted or headless environments- official profiles only for local or self-hosted flows
run_local_collection.py --dart-mode cache_only only after live OpenDART times out- do not run broad KOSPI200/all-ticker live DART without bounded cache-backfill controls
- Validate the exported bundle files and coverage fields.
- Check consumer compatibility against
/home/seok436/projects/eit-research.
- Keep benchmark and market-cap behavior aligned with the current
ci_safe contract.
- If the change affects timing semantics, also use
point-in-time-guardrails.
Timing Contract
For partial current-month local bundles, keep decision_date equal to the actual
as_of date. Do not expand it to the calendar month-end. execution_date should
be the next KRX trading day when pykrx can identify one from known business days;
otherwise fall back to the next weekday. Never emit the next-month first business
day as a placeholder for a partial-month bundle.
Coverage Reporting
Do not treat fundamental_tickers and filing_tickers as full DART coverage.
They count objects present in snapshot.json. For local cache-only DART runs, also
report:
quarters_nonemptyfiling_text_nonemptymarket_cap_nonnulllast_close_nonnull- actual DART cache coverage for the target month:
missing_fundamental,
missing_filing, latest quarter distribution, cache size/volume
DART Collection Safety
For KOSPI200 bundles, live OpenDART is enrichment, not the gate for KRX/pykrx market data.
If live DART times out, stop live calls and resume with --dart-mode cache_only.
Only use live DART backfill when it is universe-limited, delayed (5s+ per ticker), resumable,
skips cached ticker/month keys, and stops on transient network/rate-limit symptoms.
If strict DART returns empty fundamentals/filing during controlled backfill, stop live calls;
do not push through repeated 013 responses.
Do not use controlled backfill progress.json as completion evidence by itself. It is
only a resume checkpoint; diskcache eviction can make completed tickers disappear from
the actual cache when the size limit is too small.
Evidence Standard
When reporting a KR bundle run, cite the run root and validation reports from that run.
Always separate market coverage (price_tickers, market_cap_nonnull,
last_close_nonnull, sector_map, benchmark_bars) from DART enrichment coverage
(quarters_nonempty, filing_text_nonempty). Do not encode one date-specific run as
the durable baseline for future collections.
Expected shape for a healthy KOSPI200 local bundle:
- raw and bundle validation have
failed=0
- universe size is 200 for
kospi200
- price, market cap, last close, and sector map are complete or explicitly explained
- last price date is at or before
decision_date
- DART coverage is reported as actual non-empty quarters and filing text, not object counts
Read Next
- Contract and commands:
references/contract.md
- Batch artifact expectations:
references/batch.md
Commit Format
If you make a commit while using this skill, follow @rules/commit-messages.md:
<type>: <message> then a blank line then <body>.
