Tracking: build the owdata index metapackage

[larnsce]

What this is

Tracking issue for owdata: the index metapackage for the openwashdata R data packages. Decisions of record: dev/decisions-2026-09.md. Rewritten 2026-09-04.

Why

The org holds a few dozen uniform R data packages built with washr. The only catalog was a Google Sheet rendered on openwashdata.org, and since washr 1.1.0 nothing writes to it. There is no dataset- or variable-level index, and openwashdata.r-universe.dev indexes only washr.

owdata harvests the org weekly into three tables (packages, datasets, variables), ships them inside the package behind owd_packages(), owd_datasets(), owd_variables(), owd_search() and owd_install(), exports them as CSV for openwashdata.org, and produces the r-universe registry. A package is published when it carries a Zenodo DOI; there are no version tiers.

Work items, in order

• Write CLAUDE.md #16 CLAUDE.md
• Scaffold the owdata package with washr 1.1.0 #1 Scaffold with washr 1.1.0, workflows committed by hand
• Build hardened metadata parsers with fixtures #3 Parsers and fixtures
• Build the harvest orchestrator with local and GitHub backends #4 Harvest orchestrator with local and GitHub backends, value gates, version bump, registry draft
• Define and ship the catalog data model #5 Data model and CSV exports
• Implement the exported functions #6 Exported functions and the startup message
• Weekly harvest workflow that opens a pull request #9 Weekly harvest workflow that opens a pull request
• Register all packages on r-universe #8 r-universe registry, hand-committed
• Website switchover: openwashdata.org renders the harvested CSVs #11 Website switchover, the launch criterion; then close washr #69

Closed by the decisions: #2 (version tiers), #7 (catalog articles on pkgdown), #10 (migration campaign); PRs #13, #14, #15. The branch of #15 is the parts bin.

Conventions

• One maintainer. Branches and pull requests to main; no dev branch. The harvest bot opens a PR, never pushes to main.
• Code under R/ uses base R and the declared Imports (tibble, utils). Tidyverse packages are fine in data-raw/ and tests.
• The harvest is a script, never a prompt. No owdata skill.
• Style in all prose: no em dashes, no emojis, blank line after headings, en-GB.
• The repository goes public before Register all packages on r-universe #8.

Open question

The July count was about 70 packages; the GitHub harvest found 54. Check once whether the gap is private repositories, record the answer, and stop counting to 70.

[larnsce]

Premortem summary (advisory, nothing here blocks implementation)

A premortem was run on this plan on 2026-07-18 (frame: twelve months out, the plan has failed; eight independent failure investigations plus an adversarial counter-check). Full report and transcript: docs/premortem-report-2026-07-18.html and docs/premortem-transcript-2026-07-18.md on branch claude/implementation-premortem-jd5udn. Per-issue findings are recorded as comments on #2, #4, #5, #8, #9, #10, and #11.

Most likely failure: the migration campaign stalls (about 45 heavyweight reviews on one part-time maintainer, with the grandfather rule hiding the stall because legacy rows satisfy the launch metric with zero reviews done).

Most dangerous failure: no mandate. The plan publicly reclassifies other people's published, DOI-holding work and opens bot PRs into their repos, with no consent, announcement, or objection-handling step in any issue. The safe retreat under objection (tiers become informational) would gut #2, #6, #7, and #10.

Hidden assumption across all findings: mechanical signals substituting for human agreement and attention. Version plus DOI substitutes for review consent, green schema tests substitute for content correctness, an unowned follow-up substitutes for adoption, and publishing a policy substitutes for ratifying it.

Suggested (not blocking) checkpoints:

• Written tier-policy sign-off and a contributor announcement before the first public harvest.
• A reviews-per-month target for the migration campaign, with a fallback design if throughput stays under about 4 per month.
• The website switchover treated as a launch criterion with an owner and a date rather than an unowned follow-up.

Also worth naming: all twelve issues are authored and so far executed by one person who is concurrently rebuilding pkgreview and shipping washr. The coding phases fit that mode; phases 3 to 5 are coordination-heavy and are where the premortem expects the stall.

---

Generated by Claude Code