[LennartPurucker]

Why

Claims reached the other curators only after a separate claim PR was merged to main. That needed an ask, CI and a merge, so with five curators and their agents nobody knew in time who was working on what.

What

Claims now live on the branch claims and are pushed straight to it, with no PR. The branch holds one YAML file per claimed dataset, plus a README table that serves as the team's board.

• Atomic. A claim is a commit on the tip that was just fetched. git rejects a push that isn't a fast-forward, so when two people claim the same dataset at once, only one push lands; the other command fetches again and reports the holder.
• Access. Whoever can push to the repository can claim. There is nothing new to host or sign up for, and any git remote works ($DATA_FOUNDRY_CLAIMS_REMOTE, $DATA_FOUNDRY_CLAIMS_BRANCH).
• Commands. data-foundry-curation claims list|claim|release|renew|check|sync|import-records|init (src/data_foundry/curation/claims.py). It uses git plumbing only and never touches your working tree or branch. Claims expire after 7 days unless renewed, and an expired one can be taken over with --steal (the branch log records it).
• Records. claimed_by and the WIP value become a mirror of the board, written by claims claim/release/sync. The claim PR to main goes away.
• Warn only. Nothing blocks a write. Warnings come from:
  • save_record, when it writes a record another curator holds (ClaimWarning);
  • a Claude Code hook (.claude/settings.json, .claude/hooks/claims_hook.py), which lists the board at session start and warns before an Edit or Write on another curator's dataset;
  • the dashboard.
• Dashboard. The Claimed by column shows the board (fetched every 20 s). Picking your own name claims the dataset; emptying the cell releases it.
• Pages. The Pages workflow fetches the board before building. A workflow that lives only on the claims branch dispatches the Pages build on every claim.
• Docs.
  • New: .claude/skills/knowledge-data-foundry/references/claims.md.
  • The screen, add and review skills now claim before they work.
  • Updated: README, CLAUDE.md, GETTING_STARTED.md (section 7), and a CHANGELOG entry.

Already done on the remote

• The claims branch was created and the 39 claims from the records were imported (all Lennart's).
• A ruleset on claims forbids force-push and deletion, so the log can't be rewritten.

For each curator, once

git config data-foundry.curator <Name>   # a claimed_by value of curation/vocabularies.yaml
.venv/bin/python -m data_foundry.curation.cli claims list

Claude Code may ask you to approve the new project hooks the first time.

Tests

The new tests/test_claims.py uses a local bare repo as the remote and two clones as two curators. It covers:

• a race on the same dataset (one winner) and on different datasets (both land);
• release without force refused; expired claims taken only with steal; renewal;
• the cache; the records mirror and the import;
• the save_record warning, the hook (warns, never decides, fails open), the CLI and the dashboard endpoint.

Full suite: 1325 passed, 175 skipped.

🤖 Generated with Claude Code

https://claude.ai/code/session_01YHxFEkgCy8bE5DhoykkptE