Update Dataset (PR → snapshot → steps → grapher)

Use this skill to run a complete dataset update with Claude Code subagents, keep a live progress checklist, and pause for user approval only when something needs attention.

Paired skill — keep in sync. /review-data-pr is the reviewer-side counterpart of this skill: it verifies the outcomes of the author-side steps defined here. Whenever you add, remove, or change a workflow step in this file, check whether review-data-pr/SKILL.md needs a matching reviewer-side check (and add it in the same commit if so). The reverse also holds — see the mirror note there. The creation-side skills /create-dataset and /create-snapshot belong to the same family: they point into this file's canonical sections (§5b-bis sanity bounds, §5c harmonization audit, §6b metadata quality, §6c metadata checklist + link verification, §6d scheduled issues, and the snapshot-related Guardrails), so when one of those sections changes, check whether the create skills need a matching edit in the same commit too.

The two announcements are owned elsewhere. /draft-data-update-slack-post owns the internal Slack form (step 9) and /owid-staff:draft-data-update-post (from the owid-staff plugin, owid/skills-private) owns the public /latest post (step 9b). Both steps here are delegation stubs on purpose — put wording, format, and style guidance in those skills, never back into this file.

Inputs

• <namespace>/<old_version>/<name>
• Get <new_version> as today's date by running date -u +"%Y-%m-%d"
• A bare <short_name> (no namespace/version) is also valid — it's what owid-issues reminder bodies use. Resolve it to <namespace>/<old_version>/<short_name> via the DAG: rg "/<short_name>:?$" dag/ -g "*.yml" | grep -v "^dag/archive" — the :? matters because active entries are YAML keys ending in : (a $-anchored pattern without it only matches dependency lines), and archived entries must never be resolution targets. Take the latest active version; ask the user if the short name is ambiguous across namespaces. Several space-separated short names (/update-dataset <short_name1> <short_name2>) mean a grouped update of related datasets: run the full workflow for each, on one shared branch/PR.

Read the dataset's scheduled-issue body before running etl update (step 6d locates the update-*.yml workflow — peek at it now, not only at 6d): it often carries grouped-update instructions naming companion datasets that must bump together, e.g. a deflator ("US PPI New Office Construction also needs to be updated alongside to adjust for inflation"). Also check the target's garden DAG entry for same-vintage dependency datasets (a data://garden/... dep sharing the old version date is usually a companion). Seed etl update with every companion snapshot so the chains move in one run — updating only the main dataset leaves the derived series (e.g. inflation-adjusted) with a silent NaN tail for the new months.

Optional trailing args:

• branch: The working branch name (defaults to current branch)

Assumptions:

• All artifacts are written to workbench/<short_name>/.
• Persist reusable update facts to workbench/<short_name>/update-context.yml as they are discovered. This is the canonical context artifact for the PR description, review handoff, and draft-data-update-slack-post.

Progress checklist (maintain, tick live, and persist to progress.md)

Every item below is mandatory. When an item names a skill, run that skill; a partial hand check doesn't count. Skip an item only on the user's explicit request, and list the skip in the PR's open items.

• Parse inputs and resolve: channel, namespace, version, short_name, old_version, branch
• Clean workbench directory: delete workbench/<short_name> unless continuing existing update
• Pipeline refresher: run scripts/pipeline_overview.py, brief the user on the chain (what each step does, external inputs, consumers, who last updated it) and give a provisional structure verdict — re-confirm it after the snapshot + meadow diffs (see 0b)
• Run ETL update workflow via etl-update subagent (help → dry run → approval → real run)
• Add yourself to dataset.owners in the new garden .meta.yml (don't reorder; preserve existing names and markers)
• Catalog # NOTE: / # TODO: comments carried over from the old step files into notes_to_check.md
• Detect any sanity_checks functions and their log-control flags; append to notes_to_check.md
• Create or reuse draft PR and work branch
• Update snapshot and compare to previous version; capture summary
• Find the producer's latest documentation for the new release (dataset page, methodology, codebook, main report, what's new); ask the user for anything missing, insufficient or inaccessible (see 3)
• Meadow step: run + fix + diff + summarize
• Garden step: run + fix + diff + summarize
• Surface new indicators: check meadow + garden diffs (and snapshot columns when meadow hardcodes a column subset) for new-version columns absent from the old; warn user + PR body, flag new-but-dropped columns, route rename pairs to the step-7 mapping
• Review sanity_checks output (enable log flag, re-run, scan log, revert flag) — if none found and the garden step does non-trivial logic, recommend adding them; if present but missing value bounds (positive / [0,1] / [0,100] per indicator type), suggest those too (see 5b-bis)
• Country harmonization audit: validate .countries.json against canonical regions (flag provider regions not yet in the regions dataset → /add-provider-regions), audit .excluded_countries.json, scan garden log for missing/unused/unknown warnings
• Region-provider drift: if this dataset's aggregates are in regions.yml (defined_by: <provider>), check whether the new version changed the provider's region set or country membership; if so, update regions.yml and re-propagate via /add-provider-regions
• Grapher step: run + verify (skip diffs), or explicitly mark N/A
• Re-evaluate each catalogued # NOTE: / # TODO: against fresh data; delete resolved workarounds + comments together, or record status in PR body
• Run /check-metadata-typos and /check-metadata-style on the new steps (see 6b)
• Changelog: decide whether this release deserves an entry in dataset.changelog of the garden .meta.yml, dated with the new version, whether or not the dataset has a changelog yet. Entries are optional. They are release notes for the people who use the data, so capture only changes that matter to them, briefly. Show the user the draft before committing it.
• Verify indicator-metadata coverage, dataset.update_period_days, snapshot DVC date_published and citation_full year (etl update copies both verbatim), and that all URLs resolve (HEAD-check) and every #fragment matches a real anchor in the target page (anchor pass, see 6c)
• Run the adversarial review (/fact-check-dataset) as a background agent once the garden step is built, with effort scaled to the update (see 6c-bis)
• Scheduled-issue workflow check (owid-issues, found via OWID_ISSUES_DIR in .env): locate the dataset's update-*.yml (exact / fuzzy / group match), verify cron vs the observed release cadence + update_period_days, filename convention, and that the issue body says to run /update-dataset <short_name> with no generic checklist; auto-fix body/title, ask before cron changes or new workflows — commits go straight to owid-issues main (see 6d)
• Commit, push, and update PR description
• Pull admin edits of the dataset's ETL-authored single charts into ETL (/pull-chart) before pushing their viz://chart steps: pull admin drift into the YAML, list it in the PR description, and confirm in sync on staging afterwards (see step 7)
• Run indicator upgrade on staging and persist report
• Run check-hardcoded-years after all remaps — numeric time pins below the new data's latest time hide the update; propose "latest" fixes with user sign-off (see step 7)
• Run the empty-entity audit (check-empty-entities) as a background agent after the indicator upgrade: full automated pass, scaled follow-up; regressions vs production = fix, pre-existing = document (see step 7)
• Run the referencing-prose audit (standard) — read the prose of every surface the sweep found for quantitative claims this update invalidates, and hand the recomputed numbers to content — never edit published prose yourself (see step 7)
• Post @codex review as a separate PR comment once the last data or metadata change is pushed, then address its comments (fix valid ones + resolve all threads). Request it again after every later push that changes the PR's own content (step code, metadata, snapshots, DAG entries), including fixes from the QA review, so the merged and announced version is the reviewed one; merging master or deleting the old version's files doesn't need one
• Run downstream-dependency check (rg "<namespace>/<old_version>/<short_name>" dag/ -g "*.yml" | grep -v "^dag/archive"); for each consumer outside the dataset's own chain, decide with the user whether to bump in this PR or document under "Downstream dependencies" for a follow-up PR (see "Downstream dependency check" section below for details)
• Run the silent-breakage check whenever downstream consumers were repointed in this PR: confirm the buildkite/etl-automated-staging-environment PR check is green, then triage the data-diff report's data-point losses and 🔴-tier datasets and run the full-report audit probes (see "Silent-breakage check" section)
• Ask the user whether to remove the old version; if yes, remove+archive its DAG entries now and relocate the new entries into the old slot, but KEEP the old step files until review sign-off — the consecutive-version review diffs them from disk; deleting the files is the final commit before merge (see "Removing the old version & reordering the DAG") — don't forget this step
• Hand off the QA links to the user (Anomalist + Chart Diff + Metadata Diff on the staging branch, plus the data-diff report) together with the staging links to the 5 most-viewed data pages, and wait for the user's sign-off — they read those pages themselves, and the announcements below wait for it (see "Final QA hand-off")
• Update update-context.yml with published chart count and 1–3 chart views for the public announcement
• Render Slack announcement via draft-data-update-slack-post (only after the QA sign-off), save to workbench, and notify user to post it to #data-updates-comms
• Draft the public-facing "Data update" post for OWID /latest via /owid-staff:draft-data-update-post (Mode A) — two versions, user picks, then the Google Doc in /Data updates, and hand the user the link (not added to the PR). The skill declines when we posted about this data less than six months ago; a declined post is a completed item — record the eligible date, don't override
• Right before merging, check the producer hasn't published a newer release since the snapshot (see "Final QA hand-off")
• After the merge, show both announcements again in the session: the Slack text in full, and the /latest Doc link (or, when the post was declined, its eligible date), with any number the merged data changed corrected (see step 8)

Persistence:

• After ticking each item, update workbench/<short_name>/progress.md with the current checklist state and a timestamp.
• Log lessons in workbench/<short_name>/lessons.md as you hit them (what you expected, what happened, the rule). Fold them into skills only when the user asks, and only lessons that would change what a future run does (apply this bar yourself, then list the changes per skill for the user); a quirk of one dataset goes in a # NOTE: in that dataset's step files instead. For each lesson, re-read the target skill on master first. If existing text already covers the point, sharpen that text in place. Add new text only when nothing covers it: one or two sentences stating the rule and its condition (CLAUDE.md, "Write skills and this file as rules, not history"), with anything longer (a procedure, a query) in a reference file next to the skill.

Checkpoints — when to pause

Default: keep going. Run through the full workflow without stopping unless one of the conditions below is met.

Stop and ask the user when:

• A step fails and the fix is ambiguous (multiple reasonable approaches, or you're unsure of the correct one)
• Data structure changed significantly (columns removed/renamed, large row count drops, schema changes that may affect charts)
• Country harmonization has new unmatched countries that need manual decisions
• The snapshot requires a manual download or credentials you don't have
• Indicator upgrade had imperfect matches (< 100% similarity) that need human review
• The producer's documentation for the new release is missing, insufficient or inaccessible (.claude/rules/producer-documentation.md)
• Anything that could silently break charts or lose data

Don't stop for:

• Routine assertion count updates (just update them and note in the summary)
• Clean step runs with only row increases
• Expected warnings (SettingWithCopyWarning, known unmapped territories)
• Straightforward filename/version reference updates

When you do stop, present a concise summary of the issue and what options exist.

When the update isn't a drop-in version bump

Some updates carry structural changes that make the standard rename-only flow the wrong tool. Recognise them up front and adjust the workflow.

Triggers — any of these means you're in restructure territory, not a version bump:

• short_name changes (producer rebranded the dataset).
• File format/schema changes (wide → long, different file extension with a different column set, new dimensions).
• Policy/indicator set changes substantially (splits, dropped composites, newly added areas).
• Score semantics change (e.g. binary → continuous with subnational coverage).

Workflow adjustments:

1. Skip etl update. The rename-only flow copies the old step files into a new folder — useless when the schema is different. Spot this fork early, at the snapshot/codebook stage, before running etl update. Scaffold the new chain via the create-etl-steps skill (wraps the wizard's cookiecutter templates) or launch the wizard UI with etlwiz and use its "ETL Steps" page — both produce a consistent meadow/garden/grapher skeleton to fill in. Then read the v1 scripts as a reference, not as a starting copy: port the source-specific logic that still applies to the new schema (column-rename maps, status/category normalisations, country harmonisation map, sanity checks, codebook-driven structural assertions).
2. Add the new chain to the DAG before removing the old. Leave both chains active while you build and validate v2; remove the v1 entries only once v2 is on staging and the chart remap is queued or done.
3. Decide on naming convention upfront. Ask the user whether to preserve v1 short_names where they map cleanly, or to adopt the source's fresh naming scheme. Fresh naming is cleaner but means the auto-Indicator-Upgrader can't help.
4. Hand-curate the v1 → v2 indicator mapping. When short_names change entirely, the auto-upgrader has nothing to match on, but the Indicator Upgrader also matches on title — so if v2 titles are descriptive (full sentences rather than the bare short_name), you can hand the user a table of v1 title → v2 title pairs and they can drive the chart remap from there. Generate this table from the v1 meta.yml + the v2 grapher catalog.
5. Before remapping a chart onto a successor indicator, diff the chart's config against the new indicator's shape. Three things silently break on a v1 → v2 remap and none of them fails a build: (a) the chart's pinned selectedEntityNames may not exist in the successor (v1 steps often computed regional aggregates the producer doesn't publish — check the retired step's aggregation method via git show <archive-marker-sha>:<path> and rebuild the aggregates in garden the same way, e.g. population-weighted means); (b) pinned yAxis min/max sized to the old index's range can clip the new one; (c) the subtitle/FAUST may describe the old construction ("combines 18 policies") and become factually wrong. When the producer publishes several related successor indices, their codebook usually says which is the primary measure and which is a transparency baseline — cite that framing when picking the chart's indicator.
6. Defer the Slack and /latest announcements until charts have been remapped. Both posts depend on charts.published_count and charts.selected_views from the v2 chain. Drafting them before the remap gives the wrong count (zero) and no representative views. Tell the user to ping you when the chart remap is done, then run steps 8 / 9 / 9b.

For the long-format with dimensions sub-case specifically (e.g. one row per (country, year, <dim1>, <dim2>)), use the modern OWID pattern:

• Meadow + garden: tb.format(["country", "year", <dim1>, <dim2>, ...], sort_columns=True).
• Aggregations: paths.regions.add_aggregates(tb, index_columns=[...full key...], regions=REGIONS, aggregations={...}).
• Grapher: pass long tables through unchanged; the framework auto-expands them into per-cell variables.
• Metadata: variables are keyed by the long-column name, with <% if <dim> == "X" and <dim2> == "Y" %>...<% endif %> Jinja blocks inside title, description_short, display.name. Grep this repo for tb.format(["country", "year" with more than two index entries to find current reference examples.
• Jinja coverage: after building the grapher dataset, verify every active (dim1, dim2) cell renders a non-empty value — read every column from the built grapher dataset and assert metadata.title is non-empty. A dimension combination with no matching <% if %> branch ships an untitled indicator.

Workflow orchestration

0. Initial setup
   • Check if workbench/<short_name>/progress.md exists to determine if continuing existing update
   • If starting fresh: delete workbench/<short_name> directory if it exists
   • Create fresh workbench/<short_name> directory for artifacts

0b) Pipeline refresher — brief the user before changing anything

The user may not have seen this dataset in a while, or someone else ran the last update. Before etl update, explain what the pipeline looks like now and whether this update is expected to change it.

1. Map the chain mechanically:

   .venv/bin/python .claude/skills/update-dataset/scripts/pipeline_overview.py <namespace>/<old_version>/<short_name> \
       > workbench/<short_name>/pipeline_overview.md
   

   It lists every step of the chain (including older-version upstream steps and same-version siblings such as extra snapshots or helper gardens), each step's files, the inputs from outside the chain, the direct consumers (other data steps, viz:// charts/MDims/explorers, export://; readers of a shared sibling such as a common snapshot are listed apart, as they don't consume this dataset), the garden owners, the commit that created the garden step (= the last update), and the previous version (archived, or still active mid-update). It also warns when a newer version is already active, which means you have the wrong old_version.
2. Skim the step scripts: the docstrings, run(), and the helper names are usually enough. Don't read every line of a 1,000-line garden.
3. Tell the user in about 5–8 lines, no more:
   • the chain in one line (snapshot (xlsx) → meadow → garden (+ income_groups_aggregations) → grapher), plus the number of snapshots if there's more than one;
   • what each non-trivial step does, in plain words ("meadow reshapes wide → long; garden harmonizes countries, adds population-weighted regional aggregates and per-capita columns");
   • key external inputs (population, regions, income groups, other datasets) and who consumes the output (counted by type; name the viz:// ones). The script only sees DAG steps, so when it lists none, say that admin-built charts still read the grapher dataset directly — don't let "no consumers" read as "unused";
   • who did the last update and when (the creating commit, with its PR number). When the script says the step was first built (no earlier version, active or archived), say this is the dataset's first update and name who built it, not who "last updated" it.
4. Close with a provisional structure verdict: usually "Expected: no structural change — drop-in version bump." If there's already evidence of a restructure (the triggers in "When the update isn't a drop-in version bump": a renamed short_name, a new file format or schema, a changed indicator set, changed score semantics), name it. Sources: the scheduled-issue body, the producer's release notes, or the new file's column list if it's cheap to fetch.

Don't pause for approval here. Keep going unless a restructure is already visible, which is an existing stop condition. The verdict is provisional, because the real evidence (column sets, shape, row counts) only appears at steps 3–4. Re-confirm it there. Record the final verdict as update_summary.pipeline_structure in update-context.yml, and put one line in the PR Summary: Pipeline structure: unchanged or Pipeline structure: changed — <what and why>.

1. Run ETL update command (etl-update subagent)
   • Inputs: <namespace>/<old_version>/<short_name> plus any required flags
   • Pick the URI that matches what's actually changing:
     • If the source data is changing — new source files, modified extractor, anything that affects the snapshot output — run from the snapshot URI with --include-usages. This bumps the whole chain (snapshot → meadow → garden → grapher) to the new version together:

       etl update snapshot://<ns>/<old_v>/<short>.<ext> --include-usages
       

     • Foundational / widely-used datasets (e.g. wb/*/income_groups, regions, population): add --direct-only. Plain --include-usages follows usages transitively and would try to version-bump every downstream consumer (income_groups has ~85 across 15 dag files). --direct-only restricts the bump to steps sharing the dataset's own namespace/version/short_name, i.e. just its chain. Caveat: --direct-only excludes sibling steps with a different short_name that belong to the same chain (e.g. income_groups_aggregations, which the grapher step also depends on) — pass those as extra seed steps so the grapher doesn't end up mixing a new-version garden with an old-version sibling. Dry-run and confirm the proposed set is exactly the chain before executing:

       etl update snapshot://<ns>/<old_v>/<short>.<ext> data://garden/<ns>/<old_v>/<sibling> --include-usages --direct-only --dry-run
       

     • If only garden logic / metadata is changing and the source data is unchanged, run from the garden URI. This bumps garden and grapher only; snapshot and meadow stay on the old version.
   • Either way, run etl update once. Don't call it separately per channel — that leaves stale version references in the DAG (e.g., new garden pointing to old meadow).
   • Perform help check, dry run, approval, then real execution; capture summary for later PR notes
   • After running, always verify the dag file: grep for the old version and confirm all internal references between the new steps point to the new version (e.g., garden depends on new meadow, not old meadow) — a garden still on the old meadow or old snapshot silently loads stale data.
   • Check every dependency that shares the dataset's short_name. etl update points it at the step's own predecessor: right for a rolling accumulator (each release builds on the last), wrong for a fixed historical vintage (restore the version the old step used). Read the step code; ask if unclear. Also check by hand any nested-form steps it reports as "not found in dag file".
   • etl update writes the new entries in the flat DAG form — convert them to the nested (compact) form now, while you're in the file, rather than leaving it until archiving (otherwise the flat block tends to survive the whole update unnoticed). See the example and load_dag() parse-check under "Removing the old version & reordering the DAG" step 4.

1a-bis) Add yourself to dataset.owners in the new garden .meta.yml

You've just become a contributor to this dataset, so add your canonical OWID name to its owners: list. Don't reorder — keep the existing primary first; append yourself at the end. Skip if you're already there.

Your canonical name must match an entry in the schema enum (schemas/dataset-schema.json). Resolve it from git config user.name via etl.owners.resolve_owner; if that returns None, add a mapping in etl/owners.py and a row in the schema enum before continuing.

Edit the YAML in place, preserving comments and the existing # review / # backport / # fasttrack markers on other entries.

1b) Check for outdated practices (check-outdated-practices skill)

• After etl update creates new step files, run the /check-outdated-practices skill on every new step file — including helper modules that etl update doesn't generate but you copied by hand (e.g. *_omms.py), since those carry legacy patterns too
• The skill reads the extension as the source of truth for the full pattern set (the geo.add_* aggregation/population helpers are flagged, not just geo.harmonize_countries) — don't rely on a remembered subset
• Fix any findings before proceeding — this avoids propagating legacy patterns into new versions
• geo.harmonize_countries → paths.regions.harmonize_names is mechanical and safe. geo.add_regions_to_table → paths.regions.add_aggregates changes the aggregation core — prove equivalence with a controlled A/B test, not a diff against the old feather. Build the new garden both ways against the same current catalog (swap the call, rebuild, save output; revert, rebuild, save output) and diff the two. Do NOT conclude "the helper shifts aggregates across all years" from a new-vs-old-feather diff — that conflates the helper with upstream-dataset drift (see step 5). In practice the two helpers are equivalent bar tiny historical edge cases (e.g. one region-year's population residual); if so, modernize. add_aggregates also auto-resolves income groups from the DAG, so it's the right tool when you later need WB income-group aggregates.

1c) Catalog # NOTE: / # TODO: comments in the copied step files (don't resolve yet)

• Run rg -n "#\s*(NOTE|TODO|FIXME|HACK|XXX):" snapshots/<namespace>/<new_version>/ etl/steps/data/{meadow,garden,grapher}/<namespace>/<new_version>/.
• Run the same rg over the old twin of every .dvc that etl update wrote: same filename, in the old version's folder (for f in snapshots/<namespace>/<new_version>/*.dvc; do rg -Hn "#\s*(NOTE|TODO|FIXME|HACK|XXX):" "snapshots/<namespace>/<old_version>/$(basename "$f")"; done). Repeat that for each companion snapshot that sits in another namespace or version folder. etl update byte-copies a snapshot's .py but rewrites its .dvc through a YAML load and dump (apps/step_update/cli.py), which drops every comment, so .dvc notes exist only in the old version at this point. Copy every one (bar the boilerplate filtered below) into the new .dvc by hand, at the same spot, and catalog it with the rest; whether it still applies is decided in step 6a, like any other note.
• Filter out generic boilerplate (e.g. # NOTE: To learn more about the fields, hover over their names. at the top of .meta.yml).
• Save the remaining actionable items to workbench/<short_name>/notes_to_check.md — one entry per annotation, recording file path, line number, which step it lives in (meadow/garden/grapher), and what the workaround does.
• Don't act on them yet. Resolution requires fresh data and happens after each step's run — see step 6a.

1d) Detect sanity-check logic in the copied step files Sanity checks live in two different forms — detect both:

• Function form — def sanity_check… / sanity_check…( call sites. Often gated by a module-level boolean flag (DEBUG, SHOW_SANITY_CHECK_LOGS, LONG_FORMAT) that defaults to False to keep normal runs quiet. Examples: etl/steps/data/garden/wb/.../world_bank_pip.py (SHOW_SANITY_CHECK_LOGS), etl/steps/data/garden/wid/.../world_inequality_database.py (DEBUG + LONG_FORMAT), etl/steps/data/garden/lis/.../luxembourg_income_study.py (no flag; prints unconditionally via tabulate).
• Inline comment form — # Sanity check / # Sanity checks / # sanity check marking an inline assertion block that isn't wrapped in a dedicated function. Very common: etl/steps/data/garden/emdat/.../natural_disasters.py, etl/steps/data/garden/emissions/.../national_contributions.py, etl/steps/data/garden/irena/.../renewable_capacity_statistics.py. These usually have no log flag — the block simply runs on every step execution and either passes or raises.

Run a combined sweep:

rg -n -i "def sanity_check|sanity_check\(|#\s*sanity check" \
    snapshots/<namespace>/<new_version>/ \
    etl/steps/data/{meadow,garden,grapher}/<namespace>/<new_version>/

Append a "Sanity checks" section to workbench/<short_name>/notes_to_check.md listing each hit — for each, record: file path + line number, which form (function vs. inline comment), the name of any log-control flag (function form only), and a one-line description of what's being asserted (read the surrounding 5–10 lines).

Don't act yet — the review happens in step 5b once the garden step has been run on the new data.

2. Create PR and integrate update via subagent (etl-pr)

   • Inputs: <namespace>/<old_version>/<short_name>
   • Create or reuse draft PR, set up work branch, and incorporate the ETL update outputs

3. Snapshot run & compare (snapshot-runner subagent)

   • Inputs: <namespace>/<new_version>/<short_name> and <old_version>

   Producer documentation for the new release. While the snapshot runs, find the producer's latest documentation per .claude/rules/producer-documentation.md. The links carried over in the .dvc and .meta.yml point at the previous edition, so search for the current one. The what's-new notes tell you which changes to expect in the snapshot, meadow and garden diffs, and they feed the changelog. The methodology is what §6b contrasts the carried-over text against.

   Hand-maintained snapshots + editorial data edits. Some snapshots have no url_download — the .py prompts for --path-to-file and the docstring says the data was "provided by email" / curated by hand. When the update is a small editorial correction (the user gives you the facts directly, e.g. "country X did Y in year Z"), you can produce the new snapshot yourself: copy the previous version's data file (data/snapshots/<ns>/<old_version>/<file>), change only the specific cells, and assert in a quick script exactly which rows/cells changed (and that all others are byte-identical) before running etls ... --path-to-file <edited>. Then update the .py docstring to document the edit and bump the .dvc date_published / citation_full year (ask the user whether it's a new producer release or an OWID-applied edit — see step 6c). Verify the user's stated facts against the existing data first — some may already be encoded from a prior release. Tell the user what's already present rather than blindly re-adding it.

   A file whose size moves a lot while its coverage stays put has changed shape — compare column sets before running meadow. The size is visible from a ranged GET or the .dvc long before the full download; when it drops (or jumps) by a large share with the same entities and years, the producer has usually added or removed columns. Diff the old and new column sets on the snapshot itself, and for any column that disappeared, grep every consumer of the dataset — downstream steps (it may be an index key there, not just a value), exports, and repos outside ETL that read those exports — before deciding how to handle it.

4. Meadow step repair/verify (step-fixer subagent, channel=meadow)

   • Run, fix, re-run; produce diffs
   • Save diffs and summaries
   • Watch for meadow input checks keyed on absolute row/column positions. Producers quietly restructure their files (e.g. dropping the legend rows above the first country, which shifts the row count and moves every country up). Data extraction that keys off content (drop rows without an id, then melt) survives, but hardcoded tb.loc[N] / exact-row-count asserts break — update them to the new positions/counts and drop a # NOTE so the next maintainer re-checks. The break is the check doing its job; don't loosen it into uselessness.
   • Re-confirm the step-0b structure verdict. The snapshot comparison and the meadow diff are the first real evidence of shape: added or removed columns, a wide ↔ long change, new dimensions, different files. If the verdict changes, tell the user in one line what changed, update update_summary.pipeline_structure, and switch to "When the update isn't a drop-in version bump". If it holds, say nothing until the PR line.

5. Garden step repair/verify (step-fixer subagent, channel=garden)

   • Run, fix, re-run; produce diffs
   • Save diffs and summaries

   Diff against a freshly-rebuilt old version, not the stale feather on disk. The old version's data/garden/.../ feather was built whenever it last ran — possibly against an earlier snapshot of a shared upstream dataset (population, regions, income_groups). A fresh build of the new version uses the current upstream, so a naive new-vs-old-feather diff shows differences in every population-weighted cell across all years and regions — pure upstream drift that has nothing to do with your change. Before trusting any diff, rebuild the old version on the current catalog and diff against that:

   .venv/bin/etlr data://meadow/<ns>/<old_version>/<short> --force --only
   .venv/bin/etlr data://garden/<ns>/<old_version>/<short> --force --only
   

   The apples-to-apples diff should collapse to just your intended change. Mention the drift separately in the PR (Chart Diff on staging will show it, because the live data is also stale relative to current upstream). Don't skip it.

   When NaN can appear on one side, don't let .fillna(False) hide it. In a cell-by-cell diff, (a - b).abs() <= tol evaluates to NaN when exactly one side is NaN; a downstream .fillna(False) then silently drops that real one-sided change. Treat "one side NaN, other side not" as a difference explicitly.

   Spot-check derived diff columns before quoting them. A pct-change column built via an assign/sort_values chain on merged Tables can silently misalign against its own rows. Before a revision magnitude goes into the PR body or an announcement, recompute it in a fresh plain-pandas frame straight from the raw old/new files and eyeball one row (old value, new value, pct) for arithmetic consistency — a mismatch on a single row means the whole column is garbage.

   Surface new indicators. Collect columns present in the new version but absent from the old — read both the meadow diff (step 4) and the garden diff, because the pipeline may drop columns between the two. etl diff prints additions as green + Column lines, but the ranked HTML report scores them at severity 0 — they sort last and will not surface on their own; read the text/JSON output, or compare column sets directly (set(ds_new[t].columns) - set(ds_old[t].columns) per table). Classify each find:

   • New in meadow and in garden → genuinely new indicator: warn the user and list it in the PR body under a "New indicators" heading — it appears on no chart until someone curates it, needs the same metadata coverage as the rest (see 6c), and may deserve a mention in the update announcement (step 9).
   • New in meadow but missing from garden → the pipeline drops it (explicit column selection or .drop(columns=...) in garden): surface it as "the source now ships X — currently dropped" and let the user decide deliberately whether to keep it.
   • Paired + Column / − Column → usually a rename, not an addition — route it to the indicator-upgrade mapping (step 7) instead.

   Long-format tables hide additions from the column diff. When a table keys its series by a dimension column (indicator, metric, question, …) with a stable value column, a new series adds rows, not columns — the column-set diff stays empty. Also diff the distinct combinations of the identifying dimensions, not each dimension's values separately — a new series can be a new tuple of existing values (e.g. sex=male and metric=count both existed before, but the (male, count) series is new): set(tb_new.reset_index()[dims].itertuples(index=False, name=None)) - set(tb_old.reset_index()[dims].itertuples(index=False, name=None)) where dims is the list of identifying dimension columns — built tables keep them in the index after Table.format(...), hence the reset_index() (with a single dimension this reduces to a value-set diff). The grapher step expands dimension values into separate variables, so the step-7 shortName pairing surfaces these too — but catch them here, before staging.

   New companion files at the source are new-indicator candidates too. Every check above diffs within the file we snapshot — a producer adding a separate file to the release (a pre-built index, a summary panel, a codebook-derived extract) is invisible to all of them. When listing the source's files (release page, OSF/Zenodo file API), diff the file inventory against the previous cycle — the baseline is the previous snapshot's .dvc companion-files # NOTE: when the host has no file-history API (/create-snapshot step 6 writes it; refresh it in the new .dvc either way) and treat each new data file as a candidate: ingest it (usually a second snapshot + its own table in the same steps), or record the deliberate skip in the PR body.

   If the meadow step itself selects a hardcoded column subset (tb[[...]], usecols=, .drop(columns=...)), a new source column never reaches any diff — grep the meadow step for such selections and, when present, compare the raw snapshot column set old-vs-new (snap.read_*().columns) to catch additions dropped at the door.

5b) Review sanity-checks output (only if step 1d catalogued any) Handling depends on the form catalogued in step 1d.

Function form with a log-control flag (e.g. SHOW_SANITY_CHECK_LOGS, DEBUG):

1. Flip the flag to True at the top of the garden step file.
2. Re-run the garden step, capturing output:

   .venv/bin/etlr data://garden/<namespace>/<new_version>/<short_name> --force --only \
       > workbench/<short_name>/sanity_checks.log 2>&1
   

3. Review the log: scan for AssertionError, error, warning, dropped, outliers flagged by country/year, unexpected totals. Surface actionable findings in the PR description under a "Sanity-check findings" collapsed section.
4. Revert the flag to its original value (usually False) before committing. Verify with git diff that the garden file has no unintended changes.

Function form with no flag, or inline # Sanity check(s) comment blocks:

1. Read each catalogued block (pull 5–15 lines of context around the hit) to understand what invariant is being tested.
2. Important: a sanity check can enforce its finding either by raising (assert, raise) or by logging (paths.log.warning, .critical, even .fatal). Logging variants do NOT fail the step — so "step 5 passed" is not proof that every invariant held. If the block uses logging, re-run the step and scan stdout/stderr for the relevant keywords; don't trust the exit code alone.
3. For non-trivial invariants (monotonicity, totals, bounds), also spot-check qualitatively against the fresh garden output via a short .venv/bin/python snippet.
4. Record any anomalies under "Sanity-check findings" in the PR description. No log artifact to keep here since the step's own output is the evidence.

In either form: if sanity_checks raise AssertionError on the new data, stop and decide with the user whether the assertion needs a threshold bump, whether upstream data genuinely broke, or whether the invariant is obsolete. If the check only logs, treat a new/expanding set of warnings the same way — they're the signal the sanity check was written to produce.

Watch for silent-delete patterns. Some sanity_checks functions also mutate the table — e.g. world_bank_pip's sanity_checks drops rows that fail invariants and reports the count via the log-control flag. With the flag off the deletions still happen; the reviewer just never learns which rows disappeared. When reading a sanity_checks function, scan for drop, filter, tb = tb[...] — anything that removes rows — and list every deletion in the PR body, not just the warning counts. If the deletion seems newly applicable to upstream fixes (e.g. the row should no longer be anomalous in the new release), that's a candidate for removing the workaround entirely.

5b-bis) Recommend sanity checks when the garden step lacks them (or lacks value bounds) Runs after the garden step is built. Two triggers:

• Step 1d catalogued no sanity checks and the garden step does more than a straight load-and-format (harmonization, column drops/renames, aggregations, derivations) — recommend adding sanity_check_inputs / sanity_check_outputs. A Codex review will flag this anyway; better to handle it during the update.
• The step has checks but you can spot missing value-bound coverage of the kinds below — suggest the additions.

Ground every threshold in the built data before writing it. Run min/max/nunique/value_counts per indicator on the built garden table first — never write plausible-sounding bounds. The classic trap: "% of GDP" indicators look like percentages but legitimately exceed 100 (debt and expenditure can reach several hundred percent of GDP in crisis years). A blind 0–100 assert fails the build on day one.

Suggested checks by indicator type (propose the applicable ones, not all):

Indicator type	Check
Share / proportion	values in [0, 1]
Percentage of a whole	values in [0, 100]
"% of GDP" and similar ratios	non-negative for levels (revenue, spending, debt); can exceed 100 — verify against data before capping. Where a ≤100 bound mostly holds, enforce it with a documented exception set (e.g. PFMH expenditure ≤ 100 outside {Equatorial Guinea, Kuwait, Kiribati} — each with a comment explaining why it's legitimate) so a new country crossing the line fails for review
Balances, growth rates, interest rates	can legitimately be negative — do NOT impose ≥ 0
Mutually exclusive share categories	components sum to 100% (or 1) per (country, year), within a small rounding tolerance (e.g. abs(sum - 100) < 0.1) — first verify against the data that the source's categories are actually exhaustive (some breakdowns ship without an "other" remainder, or only for a subset of rows)
Categorical flags / codes	exact value set (e.g. <= {0, 1}) and non-null
Input schema	set-equality on expected columns — catches the next source rename (e.g. debt → d) with a clear message
Coverage	country-count floors (≥ previous version); a drop is usually a parsing/mapping regression — re-audit before bumping the constant
Composite sums of optionally-reported components	never a bare .sum(axis=1) — it fabricates zeros on rows where the source reports no component (and hands the whole total to any remainder derived by subtraction). But don't jump to min_count=len(components) either: ground the rule in per-year reporting counts first — when component coverage widens over time (a category only exists from year X), requiring all components can wipe most of the series. The robust pattern is min_count=1 plus a withdrawn-component rule: a component reported by someone last year and by nobody this year marks a partially-published year — null the aggregate there. That rule is release-level by design (it catches a producer publishing without a component table); if one entity's component can plausibly go missing while others still report it, also compare component presence per aggregation key, accepting the false positives genuinely ragged reporting will add. And a component that vanishes for good is a schema retirement — shrink the expected component set rather than treating the year as partially published.
Deflated ("real"/"constant") series	the deflator must cover every period of the nominal series — a how="left" merge silently yields a NaN tail that ends the real series early when the deflator lags; also assert the base-period row (prefer the producer's own annual-average row, e.g. BLS's M13) exists

Implementation conventions: constants at the top, run() first, check functions below run(); sanity_check_inputs(tb) right after loading meadow, sanity_check_outputs(tb) right before paths.create_dataset(...); plain assert with messages that name the offending values. Reference example: imf/2026-06-12/public_finances_modern_history.py.

sanity_check_outputs runs before tb.format(...), so column names are still as the code produced them — format() lowercases/underscores them afterward. If the step builds columns like f"{col}_{status}_pop" where status is "Legal"/"Illegal"/"missing", the pre-format columns are status_Legal_pop (mixed case), not status_legal_pop. Select such columns case-robustly (e.g. [c for c in tb.columns if c.endswith("_pop")] then filter on c.lower()), or you'll get a KeyError that only surfaces at runtime. Keep update-specific facts (e.g. "country X is Legal in year Y") out of the committed checks — verify those via the garden diff at update time so the checks stay valid across future releases.

Negative-test the checks: after the step passes on real data, simulate each failure mode (rename a column, corrupt a flag, push a value out of bounds) and confirm the matching assertion fires — a check that never trips is untested code.

5c) Country harmonization audit Run after the garden step completes (and after 5b if it ran). Verifies that the country entities reaching the garden output are canonical, and that the mappings/exclusions consumed by paths.regions.harmonize_names(...) are well-formed. Output: workbench/<short_name>/harmonization_audit.md.

Modern API. Garden steps should be calling paths.regions.harmonize_names(tb, country_col=..., countries_file=..., excluded_countries_file=...) — the wrapper in etl/data_helpers/geo.py:1874. If you find a step still using the deprecated geo.harmonize_countries(...) directly, step 1b's /check-outdated-practices should already have flagged it; treat that as a separate cleanup. The audit below is API-agnostic — both call sites end up emitting the same three warning strings.

Source of truth. Canonical names come from two datasets, both consulted by the harmonizer:

• data/garden/regions/2023-01-01/regions — countries, continents, and OWID-defined aggregates. The runtime authority is paths.regions.tb_regions["name"]. This is built from etl/steps/data/garden/regions/2023-01-01/regions.yml plus a merge with regions.codes.csv and field defaults — don't parse the YAML in isolation or you'll miss the legacy entries and produce false positives.
• data/garden/wb/<latest>/income_groups — the four World Bank income-group aggregates (High-income countries, Upper-middle-income countries, Lower-middle-income countries, Low-income countries). OWID treats the latest version of this dataset as the official one, so the audit must resolve the version dynamically (don't pin a date — it goes stale when WB publishes a refresh). The names live in the classification column of the income_groups_latest table.

The audit's "canonical" set is the union of these two. A .countries.json entry looks like "Source name": "Target name" — the audit checks that every target name (the value the source gets harmonized to) appears in either dataset. Anything else is flagged.

1. Capture a fresh garden log:

   .venv/bin/etlr data://garden/<namespace>/<new_version>/<short_name> --force --only \
       > workbench/<short_name>/harmonization.log 2>&1
   

2. Scan the log for the three harmonization warnings. These are emitted by etl/data_helpers/geo.py (excluded list) and lib/datautils/owid/datautils/dataframes.py (mapping warnings) — the wording is stable:

   rg -n "missing values in mapping\.|unused values in mapping\.|Unknown country names in excluded countries file:" \
       workbench/<short_name>/harmonization.log
   

   For each warning, the entity list follows on subsequent lines (because harmonize_countries() is called with show_full_warning=True by default). Capture them.

3. Validate .countries.json target names against canonical regions + income groups. Each entry maps a source name (key) to a target / harmonized name (value); this check looks at the values. For each garden step in this update:

   import json
   from pathlib import Path
   from owid.catalog import Dataset
   
   # Resolve the canonical regions dataset dynamically (latest built version).
   # Don't pin a date — when the regions step version advances, a hard-coded path
   # would validate against a stale catalog and flag valid targets as non-canonical.
   regions_dirs = sorted(Path("data/garden/regions").glob("*/regions"))
   if not regions_dirs:
       raise RuntimeError(
           "No data/garden/regions/<version>/regions built locally — the audit can't "
           "run without the canonical regions catalog. Build it first with "
           "`.venv/bin/etlr data://garden/regions/<latest>/regions`."
       )
   tb_regions = Dataset(str(regions_dirs[-1]))["regions"]
   canonical_regions = set(tb_regions["name"].dropna().astype(str))
   
   # Add OWID's official income-group aggregates to the canonical set, if available.
   # OWID treats the latest income_groups version as official. This artifact is
   # often not built locally during a non-income-groups dataset refresh — degrade
   # gracefully (warn and skip) rather than aborting the audit.
   ig_dirs = sorted(Path("data/garden/wb").glob("*/income_groups"))
   if ig_dirs:
       ds_ig = Dataset(str(ig_dirs[-1]))
       canonical_income = set(ds_ig["income_groups_latest"]["classification"].dropna().astype(str).unique())
   else:
       print(
           "[WARN] No data/garden/wb/<version>/income_groups built locally — "
           "skipping income-group enrichment. The four WB income-group aggregates "
           "(High/Upper-middle/Lower-middle/Low-income countries) may surface as "
           "'not in canonical' until you build that dataset."
       )
       canonical_income = set()
   
   canonical = canonical_regions | canonical_income
   
   mapping = json.loads(Path("etl/steps/data/garden/<namespace>/<new_version>/<short_name>.countries.json").read_text())
   not_in_canonical = sorted({v for v in mapping.values() if v and v not in canonical})
   print("Targets not in OWID's canonical regions or income groups:", not_in_canonical)
   
   # Provider regional groupings the source ships (e.g. "Europe (WB)", "Asia and the Pacific (ILO)")
   # surface here when OWID hasn't added that provider's regions yet. Flag the "(Provider)"-suffixed ones:
   import re
   provider_region_candidates = [v for v in not_in_canonical if re.search(r"\([\w .&-]+\)$", str(v))]
   print("Look like provider regions not yet in the regions dataset:", provider_region_candidates)
   

   A non-empty not_in_canonical list means the mapping points at entities that aren't registered in either the regions catalog or the income-groups dataset. This isn't automatically a bug — it's a heads-up. Stop and decide with the user before proceeding — same pattern as the global "Checkpoints — when to pause" section at the top of this skill. Common causes (in order from "fix" to "accept"): typo, retired alias used as canonical, casing/whitespace mismatch, or a custom aggregate the source defines that OWID has no equivalent for. For typos/casing — fix the JSON. Judge each target by name, never by its suffix: one provider can have several canonical suffixes for different levels (the UN's continents are canonical as (UN), its subregions only as (UN M49) / (UN SDG)), so a suffix that is wrong on one target can be canonical on another. Before proposing renames, run this check on the existing mapping, keep every target it doesn't list, and propose new names only for the targets it does list.

   If the unmatched targets are a provider's regional grouping that OWID doesn't define yet — the provider_region_candidates above ("(Provider)"-suffixed names like "Europe (WB)", "Sub-Saharan Africa (ILO)"), or the source ships a region/subregion column feeding these — don't just accept them as outside-the-system. The proper fix is to add that provider's regions to the regions dataset with the /add-provider-regions skill (separate PR): they then become canonical, merge with regions/population infrastructure, and get a {provider}_region map indicator, instead of living outside the system. Surface this to the user and offer to run /add-provider-regions. Only for genuinely one-off, non-geographic groupings the regions system shouldn't own (BRICS, G7, G20) — accept and note in the PR description that those live outside the canonical system. For a real new historical region — add an entry to regions.yml in a separate PR.

4. Audit .excluded_countries.json. The file is optional; skip if it doesn't exist, and delete it if it is empty ([]). Before deleting, check the garden step: if it passes excluded_countries_file=paths.excluded_countries_path explicitly (older geo.harmonize_countries calls), that path is loaded unconditionally and the step fails with FileNotFoundError, so drop the argument (or move to paths.regions.harmonize_names, which loads the file only when it exists) in the same change:

   excluded_path = Path("etl/steps/data/garden/<namespace>/<new_version>/<short_name>.excluded_countries.json")
   if excluded_path.exists():
       excluded = json.loads(excluded_path.read_text())
       suspicious_canonical = sorted(set(excluded) & canonical)
       # Also surface continents and aggregates separately for review
       aggregates = set(tb_regions[tb_regions["region_type"].isin(["continent", "aggregate"])]["name"].dropna().astype(str))
       suspicious_aggregates = sorted(set(excluded) & aggregates)
       print("Excluded entries that ARE canonical regions:", suspicious_canonical)
       print("Excluded entries that are continents/aggregates:", suspicious_aggregates)
       print("Full excluded list for review:", sorted(excluded))
   

   suspicious_canonical is the actionable signal: each entry is a known country/region that we are dropping. Sometimes this is intentional (e.g. dropping "World" rows because the source double-counts them) — surface, don't auto-fix. Pause and ask the user if the list is non-empty. The full list is dumped so the LLM can also eyeball it for entities that aren't in canonical but look like real countries (typos, alternative names) we should be mapping rather than dropping.

5. Audit garden output entities. Always run this check, regardless of whether .countries.json exists or is populated — JSON mappings describe inputs to the harmonizer, but the entities that actually reach Grapher are whatever sits in the country column/index of the built garden tables. Inline country assignments (e.g. hardcoded tb["country"] = "England and Wales") and post-harmonization mutations both bypass the JSON check entirely; this is the only step that catches them.

   from pathlib import Path
   
   from owid.catalog import Dataset
   
   garden_dir = Path("data/garden/<namespace>/<new_version>/<short_name>")
   ds_garden = Dataset(str(garden_dir))
   
   entities: set[str] = set()
   for tname in ds_garden.table_names:
       tb = ds_garden[tname]
       # `country` can live in the index (after .format()) or as a regular column.
       if "country" in tb.index.names:
           entities.update(tb.index.get_level_values("country").dropna().astype(str).unique())
       elif "country" in tb.columns:
           entities.update(tb["country"].dropna().astype(str).unique())
       # tables with no country column are silently skipped (e.g. reference tables)
   
   output_not_in_canonical = sorted(entities - canonical)
   print("Garden output entities not in OWID's canonical regions or income groups:",
         output_not_in_canonical)
   

   Same triage rules as the JSON-targets check (Python check #3): typo / casing / alias / legitimately custom aggregate. A non-empty list means at least one entity that ships to Grapher isn't registered in either the regions catalog or the income-groups dataset. Stop and decide with the user before proceeding. Common fixes: typo or casing → patch the inline assignment (or .countries.json, whichever is the source) so the value matches the canonical name; alias → switch to the canonical name; legitimate custom aggregate → accept and note in the PR description that the entity lives outside the canonical system.

6. Write findings to workbench/<short_name>/harmonization_audit.md with six sections, populated only when non-empty. Each section must list every flagged entity, not just a count — counts alone aren't actionable, the user (or you) needs to read the actual names to judge whether each is intentional. For long lists (>20 entries) group by pattern when the grouping is obvious (e.g. ILO's " (ILO)"-suffixed regions vs. international orgs vs. derived "World ..." aggregates) so the reviewer can scan categories instead of one flat list. Sections:

   • ## Missing in mapping — countries in source data not in .countries.json (from log warning #1) — list each missing source name
   • ## Unused mappings — .countries.json entries the data never used (warning #2) — list each unused source→target pair
   • ## Unknown excluded entries — .excluded_countries.json entries not present in source data (warning #3) — list each
   • ## Targets not in OWID's canonical regions or income groups — target names from .countries.json that aren't registered in either dataset (Python check #3) — list each target name and the source names that map to it
   • ## Excluded entries matching canonical regions — possible over-exclusion (Python check #4) — list each
   • ## Garden output entities not in OWID's canonical regions or income groups — distinct country values found in the built garden tables that aren't in canonical regions or income groups (Python check #5) — list each entity

7. Surface in PR. If any section was populated, add a collapsed "Harmonization audit" section to the PR description (after the per-step sections) with the same listings, not just a summary. Empty sections can be omitted.

When you report progress to the user during the workflow, never just give a count — always include the list (or grouped categories) so they can judge in one glance.

Checkpoint summary:

• "Targets not in OWID's canonical regions or income groups" or "Garden output entities not in OWID's canonical regions or income groups" or "Missing in mapping" non-empty ⇒ stop, decide with user.
• "Excluded entries matching canonical regions" non-empty ⇒ stop, ask whether each exclusion is intentional.
• "Unused mappings" or "Unknown excluded entries" non-empty ⇒ surface in PR description; not a blocker.

5d) Region-provider drift — if this dataset defines OWID regions, propagate the change The harmonization audit (5c-3) catches providers not yet in regions.yml. This check is the counterpart: a provider whose aggregates are already in regions.yml can change its regions on a version bump (new/removed/renamed regions, or shifted country membership), and regions.yml won't update itself.

• Does this dataset own regions? grep "defined_by: <provider>" etl/steps/data/garden/regions/2023-01-01/regions.yml (e.g. ilo_1/ilo_2, maddison, wid, wb, who). No matches ⇒ skip this step.
• Re-extract the provider's region→country mapping from the new version — from the same source the regions were originally derived from (a region/subregion column, region entities, or a table-of-contents tier; see /add-provider-regions Step 1) — and diff it against regions.yml (defined_by: <provider> members), set-equality per region, plus new/removed/renamed regions. If the new version no longer carries the composition, the provider's section in the world-region-map-definitions article links to the defining source — use it to re-derive the provider's existing regions (regions newly introduced by this update won't be documented there yet — get those from the source or ask the user).
• If it drifted: stop and decide with the user (composition changes move every region-aggregated value downstream), then update regions.yml and re-propagate via /add-provider-regions — rebuild garden regions, refresh the {provider}_region grapher indicators + metadata, regenerate owid-grapher regions.data.ts (runRegionsUpdater), and update the article section.
• If unchanged: note it and move on.

6. Grapher step run/verify (step-fixer subagent, channel=grapher, add --grapher)
   • Skip diff

6a) Re-evaluate # NOTE: / # TODO: items from step 1c against fresh data Now that meadow, garden, and grapher have run on the new data, go back to workbench/<short_name>/notes_to_check.md and decide each item's fate. For each entry:

• Identify what the workaround does (read the surrounding code).
• Load the affected step's output with owid.catalog.Dataset (or inspect the raw snapshot) and compare corrected vs. uncorrected values. Cross-check the producer's release notes / changelog if available.
• If the upstream issue is fixed → delete the workaround and its # NOTE: / # TODO: comments in the same commit, then re-run the affected step (use --force --only, add --grapher for grapher) so downstream artifacts pick up the change.
• If the workaround is still needed → leave it and add a one-line status under a PR-description section titled "Not covered in this PR" (e.g. "Sierra Leone ×1000 correction still required — raw value in the 2026 file is still ~1/1000 of plausible"). These are deliberately deferred items the next updater should re-check. Delete the whole section if its last item gets resolved mid-PR.
• If you're uncertain → keep it, flag it in the PR description, and ask the user.

Do this before step 6b (metadata checks) so any re-runs triggered by comment-removal happen before the metadata sweep, not after.

6b) Metadata quality checks (mandatory) — run after all ETL steps are built Run all four checks on the newly built garden and grapher datasets so every issue surfaces together. Each skill writes results to the terminal; fix what comes up before moving on. These four read the text as authored; seeing it as rendered, on every chart, MDim view and explorer view each edit lands on, is Metadata Diff in the final QA hand-off below.

• Typos — /check-metadata-typos scoped to the current step. Run on each of the new .meta.yml files (garden first, then grapher). Accept or skip each suggested fix.
• Jinja spacing and style guide — /check-metadata-style on the grapher step. A mechanical pass first catches template artifacts like doubled spaces or stray newlines that only appear after Jinja rendering; then it audits user-facing fields (title, subtitle, description_short, display.name, presentation.*) against OWID's Writing and Style Guide. Rules live in .claude/skills/check-metadata-style/STYLE_GUIDE.md, so no Notion access is usually needed — the skill checks the file's Last synced from Notion date and refreshes it from Notion (in a separate PR) when it is more than two months old.
• Clarity for a general audience — read every user-facing field with non-specialist eyes. The other two skills enforce structure and style; this one judges whether the text is understandable.
• Contrast the carried-over text with the new documentation — etl update copies every .meta.yml and .dvc field verbatim, so read each indicator's text against the documentation found in step 3 (methodology, what's new) and propose a fix, citing the passage, for each claim that is wrong, stale or missing a caveat (.claude/rules/producer-documentation.md, "Contrast what's already written").
• Dimension sweep (dimensional indicators only) — a sentence written for one breakdown renders on all the sibling views, where the view's own filtering can make it false: a caveat that the data doesn't control for X is wrong on the variant grouped by X, a scope word like "all employees" overclaims on a variant filtered to a subgroup, a sentence about a toggle is wrong on views that exist for only one choice of that dimension. Render the text per dimension value (the grapher channel already has it flattened) and read each output as a reader of that chart. Prefer qualifying the wording so it holds everywhere over adding a Jinja branch. Full version: check 5 of the canonical suite in /generate-metadata.

Clarity checklist (do manually, no skill yet)

OWID readers are not domain experts. Walk each indicator's user-facing fields and flag anything that requires inside knowledge to parse.

Field	Clarity check
title / presentation.title_public	A non-specialist should know what the indicator measures from the title alone. Expand acronyms unless universally known (skip GDP; expand GWIS, MFI, SDG, IHME). Don't cram units into the title.
description_short	One or two short sentences: what the metric is and what it covers. Plain words, per .claude/rules/plain-writing.md. The chart subtitle is short by design — no run-on or stacked clauses.
description_key	Free-form markdown prose (paragraphs; sub-lists only where they help). Each paragraph or bullet should land a distinct, useful fact. Skip filler ("this dataset is widely used"); prefer substantive caveats (coverage gaps, methodology limits, what counts/doesn't count).
display.name	Short legend label. Reads naturally on a chart axis/legend; doesn't restate the title.
presentation.grapher_config.note	Concise footnote, ≤1 sentence ideally.
.dvc origin.description	Shown in the Sources tab. etl update copies it verbatim from the previous version, so reread it on every update rather than trusting it. It should describe the data product in a few plain paragraphs (.claude/rules/plain-writing.md, and its guidelines in schemas/definitions.json). Producer text that is factual but badly written — long, citation-heavy, repeating the producer, carrying staff credits, legal disclaimers or stale counts — gets rewritten, not just trimmed.

Flag and rewrite when you find:

• Acronyms or technical terms that aren't expanded the first time they appear
• Sentences that only make sense if you already know the data source
• Quantitative claims with no unit context (e.g. "burned area" without "in hectares" surfacing somewhere in the user-facing text)
• Inconsistent terminology between indicators in the same dataset (e.g. "wildfires" in one, "vegetation fires" in another)
• Wording that fails the talk test in .claude/rules/plain-writing.md (e.g. "anthropogenic emissions" → "human-caused emissions")
• Methodology-attribution claims ("following guidance from <agency>…") — open the cited link and confirm it actually says that. Agencies revise methodology, so a claim that was defensible when written can be stale now, the cited page may never have contained the guidance at all, and the 6c link check only proves the URL resolves, not the claim
• Coverage claims that describe our filters, not the source: check the years and entities in the raw file before describing them, and question filters that drop source data
• Scope qualifiers present in the origin title but absent from user-facing text — if the source is "Private Construction" / adults-only / market-exchange-rate-only, description_short and description_key must say so; the citation line alone doesn't reach readers
• Text that repeats what the reader has already read — a description_key bullet restating another bullet, description_short, or the title at the same level of detail. Unpacking description_short in the first bullet (full definition, how it's measured, what's included) is exactly what the panel is for; saying it again in other words is padding. Propose merging it into the bullet that already covers the ground, or dropping it

When a phrasing is ambiguous, propose a concrete rewrite — don't just flag it. Rewording a description_key bullet in place is yours to do; deleting or merging one is a proposal. List each bullet you'd remove with the reason, in the reply and the PR body, and keep it until the user agrees.

If any skill rewrites a .meta.yml, re-run the affected step so the built catalog reflects the edits. Add --grapher when the affected step is on the grapher channel — without it the local catalog is updated but staging stays stale, so the step 7 indicator upgrade sees the old text.

# garden / meadow:
.venv/bin/etlr <channel>/<namespace>/<new_version>/<short_name> --force --only
# grapher:
.venv/bin/etlr grapher/<namespace>/<new_version>/<short_name> --grapher --force --only

Then re-run the relevant check to confirm zero remaining violations.

6c) Indicator metadata coverage, dataset block, and link verification The other quality checks catch content issues; this step catches missing fields and broken URLs before they reach review.

Snapshot DVC freshness. etl update clones the previous snapshot's .dvc content verbatim except for date_accessed. Always re-check date_published and the year in citation_full / attribution under snapshots/<ns>/<new_version>/*.dvc — they will otherwise silently ship the old version's values. Set date_published to the producer's real release date when discoverable; otherwise copy date_accessed. Bump the year in citation_full and attribution to match. The license is cloned too: open the producer's terms page and confirm license.name and license.url still match it, since a label carried over from an earlier version can name terms the producer never granted. A license-only edit doesn't reach a grapher DB that already holds the origin (Origin._upsert_select in etl/grapher/model.py matches every field except license), so staging keeps the old license until another origin field changes.

• Citation year vs date_published year. After setting both, check whether the year inside citation_full / attribution matches date_published's year. If they differ, confirm it's intentional before shipping: it's legitimate when the producer labels the release by edition rather than publish date (e.g. UN IGME's "2025 report" published 2026-03-17 → citation_full (2025), date_published 2026), but otherwise it's a stale citation. When the gap is deliberate, leave a one-line note for the reviewer so they don't re-flag it.

• Last-Modified header as date_published source. When the producer's page states no release date (common on fully JS-rendered sites like the IMF Datamapper), the download URL's HTTP Last-Modified header is a defensible source — it's the server's own timestamp for the file, not an inference. Corroborate it against a release-named filename (e.g. …Dec 2025.xlsx + Last-Modified: Fri, 12 Dec 2025) and note the provenance when reporting to the user.

• Stale producer description on a JS-rendered page. If the .dvc meta.origin.description is producer text that may have changed but the page is an SPA shell (static HTML empty, WebFetch 403s, Wayback archives only the shell), don't burn time probing API endpoints and don't rewrite the producer's text to match the data — the blurb can legitimately lag their own releases. Ask the user to paste the page text from their browser, then diff it against the existing .dvc text and apply only the substantive changes. Clipboard pastes flatten typography (curly quotes → straight, en-dashes → hyphens) — keep the existing typographic punctuation unless the words themselves changed.

Mandatory fields per indicator. For every indicator in the garden .meta.yml, confirm these are set (either on definitions.common or per-indicator):

Field	Notes
title	Per-indicator
unit	Common is fine
short_unit	Common is fine
description_short	Per-indicator
description_key	Non-empty (markdown prose, or legacy bullet list); usually common
processing_level	minor or major
presentation.topic_tags	At least one tag
display.numDecimalPlaces	Common is fine
display.tolerance	Common is fine — chart tolerance for missing years
display.name	Per-indicator legend label. The data page title is the first one set of title_public, grapher_config.title, display.name, title, so a short label (e.g. "Women") needs title_public. Check each DisplayNameWarning in the grapher build log; some datasets rely on it on purpose (IVS)
presentation.attribution_short	Set explicitly — does NOT inherit from the origin's attribution_short (MySQL variables.attributionShort stays NULL if it's only on the origin). Place under definitions.common.presentation for the common case.

Conditional: if processing_level: major, every indicator with that level MUST also have description_processing. Keep processing_level: minor as the common default and override to major only on indicators that have a description_processing field — don't blanket-set major on the common block and then leave country-level proportions without their own processing note.

Not mandatory (skip if you don't need them): presentation.title_public (unless display.name is a short label, see above), presentation.title_variant, presentation.attribution.

A per-indicator display: block replaces the common one instead of merging (only presentation and grapher_config merge key by key; see _merge_variable_metadata in lib/catalog/owid/catalog/core/yaml_metadata.py). Repeat the common keys you need, e.g. numDecimalPlaces when adding tolerance.

Dataset block. Garden .meta.yml MUST include update_period_days:

dataset:
  update_period_days: <N>

This controls the auto-update cadence. Even when the rest of the dataset: block is empty, never strip update_period_days — leave the block in place with just that field.

Link verification. Run a HEAD request on every URL in the new .dvc and .meta.yml files (all channels — meadow .meta.yml files matter when they exist). Anything non-2xx is a signal, not a guaranteed break — always double-check before acting:

for url in $(rg --no-filename -No "https?://[^\"' ]+" snapshots/<namespace>/<new_version>/ etl/steps/data/{meadow,garden,grapher}/<namespace>/<new_version>/ \
    | sed -E 's/[).,;:>]+$//' \
    | sort -u); do
    printf "%s  %s\n" "$(curl -sI -L -o /dev/null -w '%{http_code}' --max-time 15 -A 'Mozilla/5.0' "$url")" "$url"
done

The --no-filename flag prevents rg from prepending path: to each match (otherwise the for-loop tries to curl path:url and every check returns 000). -A 'Mozilla/5.0' sometimes coaxes a real response out of Cloudflare-fronted hosts, but it doesn't always work — see the next note.

curl non-2xx ≠ broken. Cloudflare-fronted sites (notably ourworldindata.org) can return 404 to curl on URLs that work fine in a browser, depending on edge-node routing, IP geolocation, and cached state. Before treating a 4xx as a real failure:

1. Re-check with WebFetch (the built-in tool). It uses a different code path and a Mozilla/5.0 UA that Cloudflare usually accepts. A 200 with a coherent page body is authoritative — trust it over curl.
2. If WebFetch also fails, sanity-check the Wayback Machine. web.archive.org is itself not fetchable by WebFetch — query the availability API with curl instead: curl -sG "https://archive.org/wayback/available" --data-urlencode "url=<url>" (the --data-urlencode matters: interpolating a raw URL truncates the lookup at the first & and can false-clear a broken link). A capture is historical evidence, not a live-link check — read the snapshot timestamp: a capture from the last few days supports the bot-blocking hypothesis; an older capture proves nothing about today (the producer may have retired the URL since).
3. No automated check is decisive — a suspected-broken link always ends at flag-and-ask. curl, WebFetch, and Wayback only accumulate suspicion: a missing Wayback capture is non-evidence (plenty of live URLs were simply never archived), and a double curl+WebFetch failure can be bot-blocking (hosts like BLS block both curl and WebFetch while serving browsers fine), so never clear a link on Wayback evidence alone. The only decisive test is a real browser — offer to open the URL in the user's Chrome through the Claude in Chrome extension, or ask them to open it — so report the evidence trail (statuses, capture date or absence) and ask before rewriting any external link in metadata. Replacing a working link with a "safer" alternative because of a false-positive is worse than leaving the original. Apply the same restraint here as the global "Checkpoints — when to pause" section.

Fix any genuinely-non-2xx hit on url_main, url_download, license.url, or URLs referenced from description / description_key before continuing. The sed strips trailing markdown/punctuation chars (), ., ,, ;, :, >) so URLs inside [text](url) aren't reported as broken because of a stray closing paren.

Anchor fragments — HTTP 200 can't validate them. Servers ignore everything after #, so https://ourworldindata.org/poverty#key-insightsfdd (broken anchor) returns the same 200 as the real #key-insights. For every checked URL that carries a fragment, run a second pass against the page body:

for url in $(rg --no-filename -No "https?://[^\"' ]+" snapshots/<namespace>/<new_version>/ etl/steps/data/{meadow,garden,grapher}/<namespace>/<new_version>/ \
    | sed -E 's/[).,;:>]+$//' | sort -u | rg '#'); do
    page="${url%%#*}"; frag="${url#*#}"
    case "$frag" in *=*|*/*|'!'*) continue;; esac  # non-DOM fragments — see below
    # grep -F: fragments may contain regex metacharacters (#section.2) — ERE would false-OK id="section-2"
    if curl -sL --max-time 20 -A 'Mozilla/5.0' "$page" | grep -qF -e "id=\"$frag\"" -e "id='$frag'" -e "name=\"$frag\"" -e "name='$frag'"; then
        printf "OK         %s\n" "$url"
    else
        printf "NO-ANCHOR  %s\n" "$url"
    fi
done

• Skip non-DOM fragments (the case list): anything containing = (scroll-to-text #:~:text=, Google Sheets #gid=…, PDF #page=…) or / (hash-route state, whether slash-prefixed #/dashboard or not — e.g. FAOSTAT's #data/FBS), plus #!… hashbangs — none correspond to an element id and all would false-alarm. Real anchors are heading slugs and never contain = or /.
• NO-ANCHOR is a signal, not proof — same epistemics as the curl false-404 note above. Two false-positive modes: the page renders its anchors client-side (the raw HTML lacks the id), or curl got a Cloudflare challenge page instead of the real body. Confirm before flagging: check the fetched body is the real page (grep for </html> and a plausible <title>), then ask WebFetch whether the page has a section/heading matching the de-slugged fragment (key-insights → "Key insights") — OWID heading anchors are slugged headings, so heading-text presence is the authoritative check.
• On a confirmed missing anchor the reader still lands on the right page, just at the top — fix the fragment (grep the body's id=" values for the nearest real anchor) or drop it. Apply the same restraint as above: flag and ask before rewriting — the section may exist under a different slug, or the page may not have rebuilt yet.

Verification. After editing, re-run the affected step (with --grapher if grapher) so the catalog reflects the changes. Then confirm presentation.attribution_short actually landed:

from owid.catalog import Dataset
ds = Dataset("data/grapher/<ns>/<v>/<short_name>")
tb = ds["<table>"]
print(tb["<col>"].metadata.presentation.attribution_short)  # must NOT be None

Or after the staging upload:

make query SQL="SELECT shortName, attributionShort FROM variables WHERE catalogPath LIKE '%<ns>/<v>/<short_name>%'"

6c-bis) Adversarial data & metadata review (mandatory — run it in the background) Run /fact-check-dataset on garden/<namespace>/<new_version>/<short_name>. Step 6c verified that the links resolve; this step actually reads the producer's documentation behind them and cross-checks the data against independent sources online — it's the only step that can catch a mistake made by the source itself (unit slips, wrong-year values, stale pre-revision numbers).

Run it as a background agent once the garden step is built, and fold its findings in before 6d, the commit and the PR description, so the fixes land in this PR. Go deeper on red flags: large unexplained churn, an in-place source revision, a new producer, editorial claims riding on specific values.

For sources that compile national-office figures (OECD, Eurostat, UN agencies), compare a few countries' new years with the national office. A series shifted by one year at the source passes every diff and bound check.

Scope for an update: focus the metadata claim review on new/changed text, and the value cross-checks (a few per indicator on a large dataset) on the newly added data (latest wave/year) plus that skill's standard anchors; deep-review the top-viewed indicators + anomaly-flagged ones per its prioritization step. Apply its routing table: metadata fixes → propose them to the user and apply only the ones they approve (its suggested rewrites are reader-facing text), then re-run the step; confirmed source errors → <short_name>.corrections.yml; unconfirmed suspicions → list under "Not covered in this PR" for the reviewer. Save the report path (ai/adversarial-review-<short_name>-<date>.md) in update-context.yml and summarize any 🔴/🟡 findings in the PR body.

6d) Scheduled-issue workflow check (owid-issues) Every recurring data update is driven by a scheduled GitHub Actions workflow in the owid/owid-issues repo (.github/workflows/update-*.yml) that periodically opens a "Data update" issue. The conventions: schedule anything updated at least once per year but less than daily; filename update-{namespace}-{short_name}.yml; a cron schedule: trigger + imjohnbo/issue-bot creating the issue. This step runs now because 6c just established the two cadence facts the cron must match — dataset.update_period_days and the producer's actual release rhythm (source.release_date / next_release in update-context.yml).

Locate the workflow. The filename convention is loosely followed in practice, so search in widening circles — a miss on the exact name proves nothing:

1. Find the user's local checkout through OWID_ISSUES_DIR in the etl repo's .env. A line in .env doesn't set a shell variable by itself, so read it into one first (quotes stripped): OWID_ISSUES_DIR=$(sed -n 's/^OWID_ISSUES_DIR=//p' .env | tr -d "\"'"). People keep their repos in different places, so never assume a path and never pick one yourself:

   • Set → git -C "$OWID_ISSUES_DIR" pull, then use it.
   • Not set → ask the user whether owid-issues is already cloned on this machine. If it is, ask for the path and append OWID_ISSUES_DIR=<path> to .env. If it isn't, offer to clone it, ask where (gh repo clone owid/owid-issues <path>), and append that path to .env once the clone is done.

   Don't fall back to a gh api …/contents filename listing — the checks below need file contents (content grep for group workflows, cron/body/assignees parsing), and the commit step needs a working tree anyway. In the steps below, $OWID_ISSUES_DIR is that checkout.

2. Exact conventional name update-<namespace>-<short_name>.yml → fuzzy filename match (hyphen/underscore swaps, dataset-title words — e.g. update-gallup-ai-indicator.yml covers gallup/ai_indicator) → content grep (rg -il "<short_name>|<namespace>|<title words>" "$OWID_ISSUES_DIR/.github/workflows/").

3. Group workflows count. One workflow may cover a family of related datasets (e.g. update-climate.yml → /update-climate-data; the quarterly update-war-ucdp-preview-q*.yml set; an "… + OMM" title covering a derived chain). If a group workflow covers this dataset, verify that workflow — don't create a per-dataset duplicate.

If found, verify three things:

1. Frequency + timing. Parse the cron: line. The implied period must be consistent with update_period_days and with the release cadence observed this update. Check the timing within the cycle too: the issue should fire shortly after the producer's expected publication window, never before (existing precedents: Gallup 0 8 15 */3 * — mid-month, just after the wave publishes; OECD health expenditure 0 0 8 7 * — right after the early-July release). If this update revealed that the cadence or window shifted, propose a new cron with a # comment in the YAML explaining the timing choice (matching the existing style) — cron changes need user sign-off before committing.

2. Naming. The filename should be update-{namespace}-{short_name}.yml. Deviations → flag in the report only; don't rename (churn, and behavior doesn't depend on the filename). Exception: a file missing its .yml/.yaml extension is genuinely broken — GitHub Actions silently ignores it — fix that without asking.

3. Actionable issue body. The body should name the dataset (no version — versions go stale) and tell the next updater to activate the Claude skill:

   title: "Data update: <dataset title>"
   body: |
     Update the <dataset title> dataset (`<namespace>/<short_name>`).
   
     To run the update with Claude Code, run:
   
     ```
     /update-dataset <short_name>
     ```
   

   For group workflows, keep it a single command listing every member dataset — /update-dataset <short_name1> <short_name2> — or point at the family skill (e.g. /update-climate-data). If the body lacks the /update-dataset pointer or references a renamed path, refresh it — body/title fixes are auto-applied and reported afterwards, no need to ask. Also check assignees: still points at the dataset's current owner; flag a mismatch, don't auto-change.

   No generic checklist in the body. The update steps live in this skill, not in the issue. A generic to-do list ("Import the latest data from the source…", "Announce the update in #data-updates-comms") duplicates the skill and goes stale apart from it, so remove one if you find it, as part of the body fix. A list written for this dataset is different: some workflows carry steps no skill covers (the wildfires and UCDP workflows, for example). Keep those as they are and add the /update-dataset pointer around them.

If not found: per the conventions above, any dataset with update_period_days roughly in [2, 366] should have a scheduled issue (err on the side of scheduling too much). Propose creating one: copy an existing workflow as template (imjohnbo/issue-bot@v3.3.6 shape; keep close-previous: false and its WARNING comment), cron shortly after the expected release window, assignees: = the GitHub handle of the human directing this update (team table in CLAUDE.md), filename per the convention, title/body per the template above. If this update touched several related datasets, propose one grouped workflow rather than several. Creating a new workflow needs user sign-off.

Committing. Commit in $OWID_ISSUES_DIR straight to main — the standing exception to the branch-first rule; no branch, no PR — with an emoji+🤖 message (e.g. 🔨🤖 Point <short_name> update issue at /update-dataset), and push. Record the outcome (workflow file, cron, verdict, changes made) in progress.md and set source.scheduled_issue_workflow in update-context.yml. Nothing about this lands in the etl PR body beyond the existing tracking-issue link.

7. Indicator upgrade (optional, staging only)

   • First upload the new grapher dataset to the staging DB (required before the upgrader can detect it):

     STAGING=<branch> .venv/bin/etlr data://grapher/<namespace>/<new_version>/<short_name> --grapher
     

     Then confirm the variables actually landed in MySQL — data://grapher/... --grapher sometimes only builds the feather without upserting, leaving 0 rows in variables. If the count is 0, run the separate grapher:// step, which does the MySQL upsert:

     # verify
     STAGING=<branch> .venv/bin/python -c "from etl.config import OWIDEnv; print(OWIDEnv.from_staging('<branch>').read_sql(\"SELECT COUNT(*) n FROM variables WHERE catalogPath LIKE %(p)s\", params={'p':'%<namespace>/<new_version>/<short_name>%'}).n[0])"
     # if 0, force the upsert:
     STAGING=<branch> .venv/bin/etlr grapher://grapher/<namespace>/<new_version>/<short_name> --grapher
     

   • Locally, this upload can fail with AccessDenied on PutObject (no write access to the staging R2 bucket). Push and let the staging rebuild upload it. The variable count doesn't prove the upload worked (rows are written before it fails): before the indicator upgrade, wait for a green buildkite/etl-automated-staging-environment on your latest push and check that <indicators_url>/<new_variable_id>.metadata.json is served (OWID_ENV.indicators_url with STAGING=<branch>).
   • Then run the automatic upgrader:

     STAGING=<branch> .venv/bin/etl indicator-upgrade auto
     

   • auto matches on the variable name, so editing an indicator title in the same PR defeats it. It reports "No perfect matches found (no variables with identical names)" and saves nothing, even when shortName is unchanged on both sides — a one-character title edit is enough, such as a straight-to-curly apostrophe fix from /check-metadata-style. That output looks like a matching problem but is a naming one, so don't read it as "nothing to remap": compare the two datasets' name and shortName columns to confirm, then use the explicit mapping below.
   • auto can detect nothing for a legitimate version bump ("No dataset migrations detected. Nothing to do."). Don't conclude there's nothing to remap — fall back to an explicit mapping: pair old/new variable ids by shortName across the two dataset ids, store with WizardDB.add_variable_mapping(...), preview with cli_upgrade_indicators(dry_run=True), then apply (mechanics under "Indicator Upgrader CLI for one-shot chart remaps" in Guardrails). Map all indicators, not just charted ones — the full mapping is also what gives Anomalist's upgrade detectors complete coverage (see Final QA). The new-version shortNames left unpaired by this mapping are exactly the new-indicator list from step 5 — cross-check the two; a mismatch usually means a missed rename.
   • The mapping only covers old_version → new_version; charts stranded on older versions stay stranded. Previous cycles leave charts behind, and the upgrader cannot see them. After the main upgrade, sweep every version of the dataset, not just the one you bumped from:

     SELECT d.catalogPath, COUNT(DISTINCT c.id) AS charts,
            COUNT(DISTINCT CASE WHEN c.publishedAt IS NOT NULL THEN c.id END) AS published
     FROM charts c
     JOIN chart_dimensions cd ON cd.chartId=c.id JOIN variables v ON cd.variableId=v.id
     JOIN datasets d ON d.id=v.datasetId
     WHERE d.catalogPath LIKE '<namespace>/%/<short_name>' GROUP BY 1
     

     Don't filter to published charts — drafts count too: the upgrader remaps them as well, a draft left on ghost variables can be published later with stale data, and its variable references block archiving the old version exactly like a published chart's. Scope the predicate to the namespace, and escape literal underscores in every substituted component — the namespace as well as the short name (LIKE 'worldbank\_wdi/%/fertility\_rate') — short names recur across namespaces (fertility_rate exists under both demography and gapminder), a suffix-only LIKE '%<short_name>' sweeps unrelated datasets into the results, and an unescaped _ is a single-character wildcard in MySQL LIKE, so it can quietly match a similarly named dataset or namespace too; since the follow-on mapping pairs variables by shortName alone, either slip would redirect another dataset's charts onto your new variables. Anything on an older version is remappable the same way (one WizardDB.add_variable_mapping per old dataset id, paired by shortName). This matters beyond tidiness: the flagship chart for the very indicator an update exists to fix can be sitting on a version from months earlier and receive none of the new data. Whatever remains after the sweep is charts whose indicators have no shortName twin in the new version — which means renamed or retired, not automatically retired: a rename in an earlier cycle breaks the twin too, so first check the unpaired new-version shortNames and the producer's release notes for a successor, and map any true rename by hand (same WizardDB.add_variable_mapping mechanics, paired semantically instead of by shortName). Only the genuinely retired ones need a replace-or-retire decision and are the reason those old versions can't be archived.
   • Set GRAPHER_USER_ID alongside STAGING. cli_upgrade_indicators(dry_run=False) asserts on it (AssertionError: GRAPHER_USER_ID is not set!) only after computing the whole mapping and listing the affected charts, so a missing value fails late and looks like a mapping bug. The value is the human's own grapher user id — the repo's main .env has it; SELECT id, fullName FROM users confirms which is theirs (other .env.* files may carry a colleague's).
   • CRITICAL: After the upgrader finishes, always verify it actually worked by querying staging:

     mysql -h "staging-site-<branch>" -u owid --port 3306 -D owid -e "SELECT COUNT(*) FROM chart_dimensions cd JOIN variables v ON cd.variableId = v.id WHERE v.catalogPath LIKE '%<namespace>/<new_version>%'"
     

     If the count is 0, the upgrade did not run — re-run it.
   • Pull admin edits of ETL-authored single charts into ETL first (mandatory when the dataset has any). A single chart in etl/steps/viz/chart/ (one with chart_config_id) can carry edits made in the admin, and those win over ETL. A dimensions override is the dangerous one: it keeps the chart on the old variable ids after its step is repointed, because the admin copy of the array wins whole. Before pushing the chart steps below, run the pull-chart skill:

     .venv/bin/etl chart-pull --dataset <namespace>/<short_name> --env <production env file> --apply
     

     Read production (--env with a production env file, or ENV_FILE_PROD): without it the command falls back to staging-site-master, which can lag production by hours. Fold whatever it pulls into this PR, and list it per chart in the PR description. Judge every field it reports as replacing an ETL value, as the skill describes. After the chart steps are pushed to staging, re-run it with --env <branch> (no --apply): every chart must report in sync. A stale variable pins finding means the admin layer still names the old variable ids. The YAML already points at the new ones, so clear the pinned fields in the branch's staging admin, as the skill describes.
   • The auto-upgrader only remaps grapher charts — NOT ETL-defined explorers or MDims. Explorers (viz://explorer/...) and multidims (viz://chart/...) reference indicators by catalog path and are rebuilt by running their viz steps, which the indicator-upgrader never touches. If the dataset has any (check the DAG: rg "viz://(explorer|chart)/.*/<short_name>" dag/ -g "*.yml"), they'll still point at the old variables on staging until you re-run them:

     STAGING=<branch> .venv/bin/etlr viz://explorer/<ns>/latest/<short> viz://chart/<ns>/latest/<short> ... --grapher
     

     Verify none still reference the old version (both queries should return empty):

     # explorers
     mysql -h "staging-site-<branch>" -u owid -P 3306 -D owid -e "SELECT DISTINCT ev.explorerSlug FROM explorer_variables ev JOIN variables v ON ev.variableId=v.id WHERE v.catalogPath LIKE '%<ns>/<old_version>%'"
     # mdims
     mysql -h "staging-site-<branch>" -u owid -P 3306 -D owid -e "SELECT DISTINCT mdp.slug FROM multi_dim_x_chart_configs mx JOIN variables v ON mx.variableId=v.id JOIN multi_dim_data_pages mdp ON mdp.id=mx.multiDimId WHERE v.catalogPath LIKE '%<ns>/<old_version>%'"
     

   • Also verify narrative charts. Narrative-chart configs can pin a variableId in their own patch (not inherited from the parent chart), and that id can date from a version older than the one this update started from — left stale by a previous cycle. The auto-upgrader only carries old_version → new_version mappings, so it can never remap those, and the chart_dimensions count above can't catch them either: narrative-chart variable ids live only inside chart_configs. The upgrader warns about this case ("was NOT remapped: it pins indicators from a version of the upgraded dataset that the mapping does not cover") — watch its output for that warning. But the upgrader only visits narrative charts whose parent chart was affected by the mapping; a stale narrative chart whose parent no longer uses any mapped indicator is never visited and stays silent. So always run this catch-all scan over all narrative-chart configs:

     # STAGING=<branch> .venv/bin/python — scan narrative chart configs for variables on ANY old version
     import json
     from etl.config import OWID_ENV
     old_vars = set(OWID_ENV.read_sql(
         "SELECT v.id FROM variables v JOIN datasets d ON d.id = v.datasetId "
         "WHERE d.catalogPath LIKE %(p)s AND d.catalogPath NOT LIKE %(new)s",
         params={"p": "%/<short_name>", "new": "%<new_version>%"})["id"])
     df = OWID_ENV.read_sql("SELECT nc.id, nc.name, nc.parentChartId, JSON_EXTRACT(cc.config, '$.dimensions') AS dims "
                            "FROM narrative_charts nc JOIN chart_configs cc ON cc.id = nc.chartConfigId")
     stale = [(r["id"], r["name"], d["variableId"]) for _, r in df.iterrows() if r["dims"]
              for d in json.loads(r["dims"]) if d.get("variableId") in old_vars]
     print(stale)  # must be empty
     

     If any are found, remap them with an explicit mapping via the upgrader's own CLI helpers (load by parent chart id; cli_upgrade_indicators won't reach them because it finds charts via chart_dimensions):

     from sqlalchemy.orm import Session
     import etl.grapher.model as gm
     from apps.indicator_upgrade.upgrade import push_new_narrative_charts_cli
     from etl.config import OWID_ENV
     with Session(OWID_ENV.engine) as session:
         ncs = gm.NarrativeChart.load_narrative_charts_by_parent_chart_ids(session, [<parent_chart_id>])
     ncs = [nc for nc in ncs if nc.id in {<stale_nc_ids>}]
     errors = push_new_narrative_charts_cli(ncs, {<old_var_id>: <new_var_id>})
     

   • Pick <new_var_id> by matching the PARENT chart, not by short_name. A stale narrative pin is often a legacy, pre-dimensional variable — an old PPP year, or a flat short_name like headcount_215 — with no short_name twin in the new dataset, so the version-bump mapping (and any shortName join) can't reach it. A narrative chart is just a framing of its parent, so the correct target is whatever indicator the parent chart uses now (after the regular-chart upgrade has run). Read the parent's current variableIds and map each stale pin to the parent's equivalent by indicator role:

     import json
     pc = OWID_ENV.read_sql("SELECT cc.config FROM charts c JOIN chart_configs cc ON c.configId=cc.id WHERE c.id=%(i)s", params={"i": <parent_chart_id>})["config"].iloc[0]
     parent_var_ids = [d["variableId"] for d in json.loads(pc)["dimensions"]]  # the targets to map onto
     

     Different narrative charts (and even one chart's multiple pins) can come from different stale versions, so build the mapping per parent, not with one global dict (e.g. a $2.15/2017-PPP legacy pin → the parent's current $3/2021-PPP variable).
   • push_new_narrative_charts_cli takes the mapping directly (no WizardDB.add_variable_mapping needed) — pass {stale_id: parent_var_id}.
   • The auto upgrade only visits narrative charts when ≥1 regular chart is upgraded in the same run. Re-running etl indicator-upgrade after the regular charts are already remapped is a silent no-op for narratives — call push_new_narrative_charts_cli directly instead.
   • Verify with a fresh API read after the update: AdminAPI(OWID_ENV).get_narrative_chart(id)["configFull"] re-reads the just-rematerialized full config (the PUT recomputes and stores it); re-using a rendered-config value fetched before the update (or from a still-open SQL session's snapshot) will mislead you into thinking nothing changed. Note that configFull is itself the stored materialized full — fresh right after a save, but it goes stale again whenever the parent is edited later (see below).
   • Watch for stale FAUST overrides. These charts often also pin an old subtitle/footnote (e.g. "$2.15 per day" / "2017 prices") that no longer matches the parent. push_new_narrative_charts_cli migrates the indicator but only warns about the text — the fix is to reset the flagged field to the parent's exact text to restore inheritance (identical values drop out of the patch). Always ask the user before changing FAUST: it's reader-facing editorial text (Footnote, Axis, Unit, Subtitle, Title), so confirm the reset/rewrite first — never fold a FAUST change silently into the indicator upgrade. (Use _find_stale_faust_overrides(patch, parent_config, mapping) to list exactly which fields are stale; set only those to the parent's value and PUT via AdminAPI.update_narrative_chart. Leave numeric display overrides like tolerance alone unless asked — they may be intentional.) Then re-run the scan and confirm it's empty.
   • The upgrader's stale-FAUST warnings only cover narrative charts visited in that run (those whose parent was in the mapping). After the upgrade, run _find_stale_faust_overrides across every narrative chart carrying the dataset's variables, and pair it with a plain grep of the patches for old unit/base-year strings — the similarity check misses notes whose wording diverged from the parent, and pinned notes can be several releases stale, not just one.
   • Editing a parent chart's FAUST does not immediately reach its narrative children. Saving a parent does not cascade: a child's rendered config is rematerialized only when the child is saved, so after a parent edit it keeps serving the old inherited text (stored config and render) even though the field isn't pinned in the child's patch. To force re-derivation, re-save the child's stored patch unchanged (AdminAPI.update_narrative_chart(id, patch)). This is safe because the endpoint treats any payload as a full config and recomputes both stored configs against the current parent — patch = diff(payload, parent), full = merge(parent, patch) (updateNarrativeChart in grapher's adminSiteServer/apiRoutes/narrativeCharts.ts) — and a stored patch always carries $schema + dimensions plus the always-persisted props (entity/time/tab/chart-type selection), so pinned fields stay pinned and everything else re-inherits from the parent. Do not "refresh" by fetching configFull and PUTting it back: in this scenario that fetches the stale materialized full, and diffing it against the new parent pins the old inherited text into the child's patch as an override — permanently freezing exactly the text you meant to refresh. (PUTting a full config is the right pattern only when you mean to change the resolved view — the indicator remaps in apps/indicator_upgrade/upgrade.py and apps/chart_sync/cli.py edit the merged config first, then PUT it.) The re-save is required, not optional: nothing rematerializes the child on its own — the site render reads the same stored full — so waiting and re-checking never fixes it. Equally, never conclude from reading the child's stored config or render that the parent edit didn't take; it did, the child just hasn't been re-derived yet.
   • All three step-7 audits below share one surface sweep. find-chart-references enumerates every surface carrying the dataset's indicators (charts, MDim views, explorer views, narrative charts, article links/embeds, data insights, static viz, key-chart slots) and classifies each reference as embed / render / link. Run it once with --dataset-id <new grapher dataset id> --transitive --json <out> --gaps-json <gaps> and let all three audits read that list, instead of each re-deriving its own joins. The second file is the sweep's own list of what it could not see (also printed as COVERAGE GAP: lines); an empty result on any of those surfaces means unknown, not nothing references it, so every audit's outcome names them. It is also the tool to reach for outside these audits, whenever the question is "what would this change touch".
   • Hardcoded-time-bounds audit (mandatory — run the skill after all remaps). A chart pinned to maxTime: <old year> keeps opening at the old year after the update ships — the new data is on the indicator but invisible by default, and its edit history shows a human bumping the pin every cycle. Run the check-hardcoded-years skill in dataset mode: it sweeps the same surfaces as the empty-entity audit (charts, map tabs, MDim views, explorer views, narrative charts, article time= embeds/links) for numeric minTime/maxTime/timelineMinTime/timelineMaxTime/map.time pins and grades each against the new indicators' latest time (metadata fetches share the empty-entity audit's cache when both run). It's cheap enough to run unconditionally. A pin below the new max = the update is invisible on that surface — propose setting it to "latest" before merge, with user sign-off (reader-facing default-view change): charts get fixed on staging so the edit rides Chart Diff, MDim/explorer pins get fixed in their ETL YAMLs + export re-run, article pins go to content follow-up. A pin equal to the new max goes stale next cycle — propose the same fix now. Deliberate pins (year in title/subtitle/slug, narrative charts, single-year comparisons) are listed, not fixed — they're coupled to FAUST text.
   • Referencing-prose audit (standard — run after all remaps). The other two audits check whether a surface still renders correctly. This one checks whether the words around it are still true. Our own metadata prose is covered separately (see the "Grep metadata prose for numbers carried from the old release" guardrail); this is about published content that cites the dataset — article paragraphs, data-insight bodies, and above all titles, which carry the most-read numbers and are the easiest to forget. Take the surface list from find_references.py --json (the sweep above), never from a hand-rolled posts_gdocs_links join. Two of the surfaces that matter most don't appear in that table at all: a data insight holds its chart in front matter (content->>'$."grapher-url"'), and a featured metric lives in featured_metrics. A join on posts_gdocs_links undercounts, and what it silently drops can be the data insight whose title carries the headline number. The script is the right source, not an exhaustive one: two of the coverage gaps it reports (see the sweep above) bite this audit in particular — a chart nested inside an article layout container may leave no posts_gdocs_links row, and a data insight that records its chart anywhere other than grapher-url (or its narrative chart anywhere other than narrative-chart) is invisible. Carry those gaps into the outcome: "no stale claims" is a verdict on the surfaces swept, never on the site, so record it as "none in the N surfaces swept; not swept: …". When a surface you know cites the dataset is missing from the list (an article the content team named, a post the chart's own page links to), read it by hand and say so. For each gdoc-backed row, fetch the doc by the row's surface_id (the Google Doc id) and read both posts_gdocs.markdown (the body) and posts_gdocs.content->>'$.title' — the title is stored in the content JSON, not in markdown, so a body-only read misses the headline, and a body can be correctly hedged while the title states the bare multiple — and the title is what readers see in feeds and social cards. Then re-derive every quantitative claim against the new data. The distinction that decides whether there is anything to do:
     • Time-bounded claims survive appended periods — but not revisions of the periods they name. "In early 2023 it was around $4 billion; by late 2025 it had grown to $62 billion" is untouched by a new quarter, because it names the quarters it compares. It is not untouched by a revision: if this update changed the value or the date of a named period (a restatement, a corrected label, a moved period date), recompute it like any other claim. The exemption is for append-only updates; check the diff for touched periods before granting it.
     • Unbounded or "latest-implied" claims do. "has grown 1,300-fold", "now accounts for over 90%", "the biggest single segment" all silently re-point at whatever the newest data is. Recompute each one. Narrative charts carry claims of their own, and they go stale on staging, not later. A narrative chart's pinned title/subtitle often restates the post's headline ("…has grown 5-fold since late 2022"), but unlike a static image it renders live data: with maxTime: "latest" it draws the new months under the old claim the moment the remap lands. Read its merged config (AdminAPI.get_narrative_chart(id)["configFull"]) and recompute any claim in it. The fix belongs in this PR, with sign-off since it's reader-facing: either retitle it to the new figure, or pin its maxTime to the period the claim describes so it stays a dated snapshot. Decide one or the other, never both. Route a retitle through /edit-faust-metadata (it owns narrative-chart text and its blast radius); a maxTime pin goes straight to staging. When the parent is a chart, either edit reaches production with it once the parent's Chart Diff is approved. Chart-sync only carries narrative children of synced charts, so a narrative chart parented to an MDim view (parentMultiDimXChartConfigId) never syncs: make the same edit on production in the admin after merge, and keep it in the PR's open items until it's done. The data insight that embeds it is still the author's call. A static image changes the remedy, not the finding — and the sweep's kind cannot tell you which you have. kind=embed means the surface holds the chart by slug (a data insight's front-matter grapher-url, an announcement's cta button); it says nothing about what the reader sees, and both of those can sit above a static picture. Read the doc body instead — posts_gdocs.content->'$.body[*].type': an image block is a static picture, a chart block is a live grapher that picks new data up on its own. Where the picture is static, prose and image must agree — editing one sentence desyncs them and is worse than leaving both consistent-but-dated — so the honest options are "refresh image and text together" (a separate static-viz job) or "leave it, and record why". Say which you are recommending. Featured metrics and key-chart slots carry no prose, so they can't go stale this way; list them only so the surface count is honest. Route findings to content follow-up with the recomputed number attached, and never edit published prose yourself — it is reader-facing copy owned by whoever wrote it. Record the outcome (fixed / deliberately left, with the reason) in the PR's open items, so the next updater doesn't re-flag it — and say whether the update was append-only or revised named periods, since that decides whether bounded claims were in scope. The PR body alone doesn't tell the user: list each unsettled finding in your reply as waiting on their decision, with the current and proposed text, and repeat it in every report until they settle it. Link every finding, so content can act on it without searching. For each claim give the live page, the Google Doc (https://docs.google.com/document/d/<surface_id>/edit), a scroll-to-text link that lands on the sentence itself (<live url>#:~:text=<url-encoded claim>), and the chart whose numbers it cites. A table of quotes and recomputed numbers without links sends the reader back to search for every row. The sweep's --markdown report already builds the doc and scroll-to-text links; carry them into your own table rather than dropping them. The same applies to every other reference list you hand to a person — empty entities, hardcoded years, Codex follow-ups: name the surface and link it.
   • Empty-entity audit (mandatory — run it in the background after the indicator upgrade). An upgrade can leave a view pinned to entities that have no data in the new indicators — it renders with those series missing (or empty) and no error anywhere. The check-empty-entities skill audits charts (entity selections + map columnSlug), MDim views, explorer views, narrative charts, and published-gdoc country= references, and grades every finding against production. Its automated pass always covers every surface; on widely charted datasets, scale only the follow-up (start with changed data, most-viewed charts, and views most likely to break) and report unchecked surfaces as coverage gaps. When it runs: a pinned entity that had data on production and lost it on staging is a regression from this update — fix it before merge (remap the view or restore the entities); a gap identical on production is pre-existing — it still needs fixing (chart-config edit or content follow-up on the gdoc), just not necessarily in this PR, so list it in the PR body with a planned fix.

8. Update context for public announcement

   • Steps 8, 9 and 9b run only after the user has signed off the final QA hand-off. An announcement describes the data, and the review can still change it.
   • After the merge, show both announcements again in the session, since the drafts were handed over many turns earlier: the Slack text in full in the reply (not only the file link), and the /latest Doc link with its open steps, or, when 9b declined the post, its eligible date instead. First recheck their numbers against the merged data, because changes after the drafts (review fixes, a decision on suspect source values, a late metadata edit) can move them; update the saved drafts (slack-announcement.md, data-update.md) where they moved. Correct the Doc in place with the Google Docs connector (update_doc, per /owid-staff:draft-data-update-post's references/gdoc-format.md) and say which numbers changed, old and new; without that connector, list them next to the link for the user to fix by hand.
   • In workbench/<short_name>/update-context.yml (kept up to date as facts are discovered, see Assumptions), record at minimum:

     dataset:
       namespace: <namespace>
       old_version: <old_version>
       new_version: <new_version>
       short_name: <short_name>
       title: <public dataset title, if known>
       producer: <producer, if known>
     source:
       release_date: <snapshot origin date_published, if known>
       next_release: <best-effort, or null>
       url_main: <source page, if known>
       citation_full: <citation, if known>
       scheduled_issue_workflow: <owid-issues update-*.yml filename, or null (from 6d)>
     coverage:
       year_min: <garden min year>
       year_max: <garden max year>
       countries: <distinct countries/entities>
       includes_regions: <true/false>
       sparse_recent_year_note: <note, or null>
     charts:
       published_count: <published chart count>
       size_qualifier: <handful|moderate|large|massive>
       explorers: <list of published explorer slugs using this data, or []>
       mdims: <list of MDim slugs using this data, with published flag, or []>
       selected_views:
         - title: <chart title>
           slug: <chart slug>
           rationale: <why this represents the dataset>
     announcement:
       prior_posts: <list of {slug, type, published} for OWID posts already covering this data, or []>
       last_public_post: <date of the most recent announcement or data-insight, or null>
       eligible_from: <last_public_post + 6 months, or null if never posted>
       decision: <drafted | declined (cooldown) | n/a>
     update_summary:
       pipeline_structure: <"unchanged" | "changed — <what and why>"> (from 0b, re-confirmed at step 4)
       snapshot_diff: <short summary or artifact path>
       meadow_diff: <short summary or artifact path>
       garden_diff: <short summary or artifact path>
       notable_changes: []
       sanity_check_findings: []
       resolved_workarounds: []
     editorial_context:
       why_it_matters_snippets: []
       caveat_snippets: []
       interesting_update_snippets: []
     

   • Query the staging DB for published charts using the new dataset (filter on c.publishedAt IS NOT NULL). Draft/unlisted charts must not be counted in the announcement:

     SELECT c.id, cc.slug, cc.config->>'$.title' as title, cc.config->>'$.type' as type, cc.config->>'$.hasMapTab' as hasMapTab
     FROM charts c
     JOIN chart_configs cc ON cc.id = c.configId
     JOIN chart_dimensions cd ON cd.chartId = c.id
     JOIN variables v ON cd.variableId = v.id
     WHERE v.catalogPath LIKE '%<namespace>/<new_version>%'
       AND c.publishedAt IS NOT NULL
     GROUP BY c.id
     

   • Charts are not the only surface — also count the explorers and MDims that use this data. Many datasets feed published OWID explorers and multi-dimensional collections, which the grapher-charts query above misses entirely. Run the viz steps first (step 7) so these point at the new variables, then query both:

     -- Explorers (note isPublished — only published ones count for the announcement)
     SELECT DISTINCT ev.explorerSlug, e.isPublished
     FROM explorer_variables ev
     JOIN variables v ON ev.variableId = v.id
     JOIN explorers e ON e.slug = ev.explorerSlug
     WHERE v.catalogPath LIKE '%<namespace>/<new_version>%';
     
     -- MDims (note published flag — drafts are published=0)
     SELECT DISTINCT mdp.slug, mdp.published
     FROM multi_dim_x_chart_configs mx
     JOIN variables v ON mx.variableId = v.id
     JOIN multi_dim_data_pages mdp ON mdp.id = mx.multiDimId
     WHERE v.catalogPath LIKE '%<namespace>/<new_version>%';
     

     Chart-based explorers can also attach via explorer_charts.chartId (join through chart_dimensions) rather than explorer_variables — check that table too if the variable-based query comes up empty but the DAG shows an explorer step. Record published explorers/MDims under charts.explorers / charts.mdims in update-context.yml, and fold them into the "How many charts did this update affect?" answer (e.g. "10 published charts, 3 explorers, plus 3 draft MDims"). Only count published surfaces (isPublished=1 / published=1) toward the public announcement; note unpublished ones for QA.
   • Map the published chart count to size_qualifier: 1–9 = handful, 10–49 = moderate, 50–199 = large, 200+ = massive.
   • Pick 1–3 selected_views using these criteria (in order of preference):
     • Map views — immediately visual, readers can find their own country
     • Charts with punchy, standalone headlines — titles that make a clear claim work best for social sharing
     • Global trend charts (StackedArea / World) — show the big picture over time
     • Skip: population-weighted variants (harder to read quickly), within-regime breakdowns (too niche), country-specific views
   • Add snippets for the editorial prompts from source metadata, garden/grapher metadata, resolved sanity-check/workaround notes, and non-routine PR changes. Keep these as snippets/facts, not polished Slack prose.

9. Slack announcement

   • Run the draft-data-update-slack-post skill with workbench/<short_name>/update-context.yml as input. draft-data-update-slack-post is the canonical owner of the Slack form wording, copy-paste format, editorial framing, search URL, and any standalone fallback gathering. Do not duplicate that rendering logic here.
   • Save the rendered draft to workbench/<short_name>/slack-announcement.md.
   • If draft-data-update-slack-post reports missing mechanical fields, gather them, update update-context.yml, and re-render rather than inventing values. Ask the user if a missing field requires judgment.
   • Do not put the announcement in the PR at all — no embed and no pointer. The draft stays as the workbench/<short_name>/slack-announcement.md file (the user copies from there); comms drafts are internal and are kept out of the public data-update PR.
   • At the end of the update, tell the user, with a markdown link to the saved file so they can click through to open it: "Slack announcement drafted at [workbench/<short_name>/slack-announcement.md](workbench/<short_name>/slack-announcement.md). Please review and post it to #data-updates-comms." Always render the path as a markdown link […](…), not as inline-code — the chat UI renders it as clickable that way. (Slack can't be auto-posted — the user posts it.) In a git worktree, link by absolute path (/…/.claude/worktrees/<name>/workbench/<short_name>/slack-announcement.md): a relative link resolves against the editor's workspace root, i.e. the main checkout, whose gitignored workbench/ can still hold the previous update's draft, so the user reads stale text. The same goes for every workbench or ai/ artifact you hand over, including 9b's draft.

9b) Data update post (for OWID /latest)

• Run the /owid-staff:draft-data-update-post skill in Mode A, with workbench/<short_name>/update-context.yml and workbench/<short_name>/slack-announcement.md as input. That skill is the canonical owner of the CMS format, the house style, the CTA link rules, the Google Doc creation and styling, and the handoff wording. Do not duplicate any of it here.
• This is separate from the Slack announcement (step 9) — that one is a 10-field form for the internal channel; this one is a mini-blog-post for OWID readers, published on https://ourworldindata.org/latest.
• The skill drafts two versions for the user to choose between, saves the chosen one to workbench/<short_name>/data-update.md, and only then creates the Google Doc. Don't create the Doc yourself, and don't shortcut the sign-off: the Drive API can't edit or delete a Doc once created. Later fixes to the text go in place through the Google Docs connector when it is available.
• If the skill reports a missing fact, gather it (snapshot .dvc, garden .meta.yml, url_main), persist it back to update-context.yml, and re-run rather than inventing a value.
• "No post" is a valid outcome, not a blocked task. The skill checks what OWID has already published about this data and stops when the last announcement or data insight is less than six months old — a routine refresh of a frequently-updated dataset often lands inside that window. When it declines, record the prior post(s) and the eligible date under announcement: in update-context.yml and in the PR's open items, and move on; don't override it to get a draft. The Slack post (step 9) is unaffected — that one runs every update.
• Do not put the post in the PR at all — no embed and no pointer. Like the Slack draft, it stays in workbench/.

10. Codex review: trigger, address comments and resolve threads

• Post @codex review as a separate PR comment (not in the PR description) once the last data or metadata change is pushed, before the final QA hand-off, so its findings are handled before the user reviews. Request it again after every later push that changes the PR's own content — step code, metadata, snapshots, DAG entries, fixes from the QA review included: the review must cover the content that gets merged and announced. Merging master into the branch or deleting the old version's files doesn't change that content, so it doesn't need a new review. Before pushing again or posting another request, wait for the current round to complete and handle its findings: a later request doesn't clear them, they stay as open threads. Use:

  gh pr comment <pr_number> --body "@codex review"
  

• Codex's delivery channel depends on the verdict — poll both. A clean pass arrives as an issue comment ("Didn't find any major issues") from chatgpt-codex-connector[bot], with zero inline comments and no formal review object. A review with findings arrives as a formal review ("💡 Codex Review") with inline comments, and no issue comment. A watcher polling only one channel waits forever on the other outcome — treat a hit on either as completion. A third clean-pass shape: the automatic review on opening or marking ready can end with only a 👍 reaction on the PR and no comment at all — it counts only when the reacting login starts with chatgpt-codex-connector and it is stamped after the trigger. When the codex-pull-request-review-summary issue comment is present, it is the most direct completion signal: it is edited in place, and its table row shows ✅ Completed with the reviewed commit, which must match your HEAD.
• Wait ~60 seconds after posting @codex review, then poll both channels:

  gh api repos/owid/etl/issues/<pr_number>/comments | python3 -m json.tool   # clean-pass summary lands here
  gh api repos/owid/etl/pulls/<pr_number>/comments | python3 -m json.tool    # findings land here as inline comments
  gh api --paginate repos/owid/etl/issues/<pr_number>/reactions \
    --jq '.[] | select(.content == "+1" and (.user.login | startswith("chatgpt-codex-connector"))) | .created_at'   # bare-👍 clean pass
  

• A findings review can also carry its finding only in the review body (no inline comments, no resolvable threads) — list gh api repos/owid/etl/pulls/<n>/reviews and read each new review's body; polling only the two comment endpoints misses it (there is no thread to resolve — reply via a normal PR comment instead):

  # clean-pass summaries land in the issue comments:
  gh api repos/owid/etl/issues/<pr_number>/comments \
    --jq '.[] | select(.user.login | test("codex";"i")) | .body'
  # review-body-only findings (the third shape) — no thread, no issue comment:
  gh api repos/owid/etl/pulls/<pr_number>/reviews \
    --jq '.[] | select(.user.login | test("codex";"i")) | .body'
  

• Fetch open review thread IDs via GraphQL:

  gh api graphql -f query='{ repository(owner:"owid", name:"etl") { pullRequest(number:<pr_number>) { reviewThreads(first:20) { nodes { id isResolved comments(first:1) { nodes { body } } } } } } }'
  

• For each unresolved Codex comment:
  • If valid: apply the fix, commit, push, then resolve the thread:

    gh api graphql -f query='mutation { resolveReviewThread(input:{threadId:"<thread_id>"}) { thread { id isResolved } } }'
    

  • If not valid / not applicable: reply explaining why, then resolve the thread:

    gh api repos/owid/etl/pulls/<pr_number>/comments/<comment_id>/replies -f body="<explanation>"
    gh api graphql -f query='mutation { resolveReviewThread(input:{threadId:"<thread_id>"}) { thread { id isResolved } } }'
    

• If neither the inline-comments endpoint nor the issue-comments endpoint shows a Codex post after 60 s, wait another 60 s and retry (up to ~5 min total). Codex can take 5–10 min.

Committing and pushing

Commit and push incrementally as you go — after each step that produces code changes. Don't wait until the end. Use descriptive commit messages with appropriate emojis (the one auto-prepended by etl pr for the chosen category + 🤖 for AI-written code).

Verify the branch immediately before every commit (git branch --show-current). The session shares its checkout with the user's IDE — a branch switch there silently moves your shell too, and a commit then lands on whatever branch the IDE left behind. The failure is quiet: git push -u origin <branch> pushes the (empty) local PR branch, reports "Everything up-to-date", and the PR stays empty. Recover by cherry-picking the stray commit onto the right branch (don't force-move the other branch's pointer).

Keep the PR-body draft under workbench/<short_name>/ (or re-fetch it with gh pr view <num> --json body --jq .body before each edit) — not in the session scratchpad. The body gets re-edited throughout the update, and scratchpad files don't survive session resumes.

At the end of the workflow, update the PR description with:

• A tracking-issue link as the first line of the Summary — e.g. Tracks: [owid/owid-issues#NNNN](https://github.com/owid/owid-issues/issues/NNNN). Most data updates have a corresponding owid-issues ticket; try to find it by searching the title or <short_name> first, and ask the user for the issue number if you can't locate one rather than skipping the link silently.
• A summary of key changes at the top, including the Pipeline structure: … line from step 0b
• Collapsed <details> sections only for the pipeline steps that changed in a non-obvious way. Skip any step that's just the boilerplate generated by etl update — don't add a placeholder like "unchanged from boilerplate". The Summary already explains the why; per-step sections are only for the how, when the how isn't obvious from the diff.

A long-lived branch silently regresses shared charts

A branch that sits open while another dataset update merges to master will serve that dataset's old version on its staging server, because staging builds from the branch. Any chart combining both datasets then differs from production on two axes: yours (intended) and theirs (a regression). Approving such a diff syncs the branch's stale config back to production and reverts the other team's upgrade on a published chart.

It doesn't announce itself — CI is green, the chart renders, and the diff looks like the change you expected. So whenever a dataset update lands on master while your PR is open, or a chart diff shows a change you can't attribute to your own work, compare every dimension of the affected charts against production, not just the one you touched:

# for each chart id: list (shortName, dataset version) per dimension on staging and on prod

The fix is to merge master in, rebuild, and remap only the affected charts' foreign dimensions. Two scoping notes: chart-diff only covers charts your branch's datasets touch, so charts using only the other dataset are stale on staging but never sync — leave them, they're not yours to fix and repointing them adds review load. And remap by editing those chart configs directly rather than via a variable-level mapping, which would sweep in other charts sharing the same variables and pull them into the review queue.

Downstream dependency check

After completing the update, check if any other datasets depend on the old version of the updated dataset:

rg "<namespace>/<old_version>/<short_name>" dag/ -g "*.yml" | grep -v "^dag/archive"

Filter out the old dataset's own DAG entries (snapshot → meadow → garden → grapher chain). Any remaining references are downstream dependents that still point to the old version.

If downstream dependents exist, decide with the user whether to bump them in this PR or defer to a follow-up:

• Tell the user which datasets depend on the old version.
• Follow-up PR (default for a big fan-out): add a "Downstream dependencies" section to the PR description (not collapsed) listing the dependents, to be repointed in a separate PR. This is the usual two-PR pattern for foundational datasets (e.g. income_groups: chain-update PR, then a "🐝 Update all datasets to latest …" bulk-bump PR).
• Version-bumping a derived step inside the follow-up PR is a mini dataset update — run it as one. Repointing alone doesn't call for a bump (CLAUDE.md's "Versioning derived/OMM steps"), but when the follow-up also changes a derived/OMM step's logic or the owner opts to bump it, use .venv/bin/etl update data://garden/<ns>/<old_version>/<short_name> --include-usages --direct-only — it works fine for snapshot-less derived steps — rather than git mv + a hand-edited DAG, and treat the bump's charts like any update's: new variable IDs mean an indicator upgrade on staging (see step 7 for when auto detects nothing), then the step-7 audits (check-hardcoded-years, check-empty-entities) scoped to those charts. A hand-rolled git mv can reach the same end state, but it silently skips the audits and leaves no workbench trail.
• Bump in this PR (if the user wants it self-contained): repoint every downstream ref and remove/archive the old chain in the same PR. Mechanics to get right:
  • Bulk-replace with a negative-lookahead so a prefix match doesn't corrupt sibling short_names — e.g. re.sub(r"garden/wb/<old_v>/income_groups(?!_)", "garden/wb/<new_v>/income_groups", text) leaves income_groups_aggregations alone.
  • Remove the old own-chain block from dag/main.yml before the bulk sweep, or the sweep turns the old definition into a duplicate of the new key. Relocate the new block into the old slot (nested form) as part of the same edit.
  • Downstream datasets keep their own version and variable IDs — only their dependency on the updated dataset changes — so no chart remapping is needed for them; their aggregates just recompute against the new data (visible in Chart Diff). The indicator upgrade (step 7) still only concerns charts that use the updated dataset's own variables.
  • This is the only case where "Removing the old version" happens in the same PR — otherwise the old chain must stay until the follow-up repoints its consumers.

Silent-breakage check (downstream builds + value diff)

A foundational-dataset update can leave a downstream step building cleanly while quietly dropping data — a region whose aggregate can no longer be computed goes NaN, a reclassified country disappears, a join stops matching. Nothing raises; the feather is written; the gap only surfaces on a chart weeks later. Two existing commands cover it — no bespoke tool:

1. Do all downstream consumers still build?

.venv/bin/etlr --modified --continue-on-failure  # add --dry-run to list scope first

--modified detects the steps changed vs origin/master and expands to their full transitive downstream via the branch DAG (same machinery as chart-diff), runs them in dependency order, skips dependents of failed steps, and ends with a failure summary + non-zero exit. Staging runs the same check on every push: its bake (ops/templates/owid-site-staging/etl-build.sh) is etl run garden grapher explorers --modified --grapher --continue-on-failure with PREFER_DOWNLOAD=1, and --continue-on-failure re-raises the first failure at the end — so any consumer crash turns the buildkite/etl-automated-staging-environment PR check red. That makes the local run optional fast feedback, not the primary net. The caveat that matters: while that check is red, the data-diff report under-reports — dependents of the failed step are skipped, stay stale in the catalog, and diff as unchanged. Always confirm the check is green before trusting the report. Use --workers N to parallelize a big local fan-out.

Size the fan-out first (--dry-run lists the scope). Small (≲50 steps): the local run gives you the crash check in minutes, before burning a staging cycle. Foundational-dataset scale (hundreds of steps — an income-groups bump is ~560): skip the local build and rely on the staging check — locally it costs ~35 min and ~7 GB of data/ to duplicate what the bake does anyway (if you do run it, delete the builds afterwards; data/ is regenerable cache). Skip the local etl diff at that scale too — another ~25 min and a JSON in the hundreds of MB; owidbot's hosted report on the PR is the same comparison for free.

2. How much did their outputs change? The data-diff report answers this directly — it ranks everything by anomaly score (BARD, the metric Anomalist uses) and makes data loss unmissable, so you read its verdicts instead of scanning the diff yourself:

• On the PR (zero effort): open owidbot's data-diff HTML report (easiest via the full report link in owidbot's PR comment; direct URL under "Final QA hand-off") — it compares the staging build (new dependency) against production (old dependency). For a dependency bump the consumers' code is unchanged, so this is a clean old-dep-vs-new-dep comparison.
• Locally, after step 1: .venv/bin/etl diff REMOTE data/ --changed --include garden --output-html data-diff.html.

How to read it, in order:

1. "Top changes — what to watch": red "− lost N data point(s): labels…" entries lead both the Datasets and Indicators lists — each one is a coverage loss (the silent-breakage signature) and must be triaged: legitimate churn, or a silent drop? Below the losses come the biggest value changes with their median anomaly scores.
2. Tier strip + coverage chips: 🔴 datasets (score ≥ 15% or any coverage loss) need review; 🟡 a skim; 🟢 is rounding-level noise. A red − N row(s) removed: … chip on a dataset row always means data points disappeared, regardless of how small the share.
3. Filters: the tier dropdowns isolate 🔴 datasets/indicators; 📝 metadata-only separates pure metadata edits from value changes.

Full-report audit — deciding whether code changes are needed. For a dependency bump, "churn or bug?" is answerable mechanically, because only the dependency changed and consumer code didn't. Load the report JSON (DiffReport.from_json) and run these probes over all changed datasets, not just the Top-changes lists — each should come back empty or confined to dep-derived entities; anything else is a code-change candidate:

1. Structural: new/removed tables or columns anywhere → a dep bump must never do this.
2. "World" rows changed → some step computes World by summing the dep's groups (anti-pattern: World should come from the source or the full country set).
3. Raw-country rows changed → consumer code entangled with the dep beyond aggregation. (In bilateral datasets, verify every changed pair has an aggregate side; the country names that show up in samples are often just other dimension values — disaster types, element codes, age groups — so check dim columns, not string looks.)
4. Any indicator with >30% of rows changed → more than the aggregate rows moved.
5. Wipe vs edge, for every coverage loss: does the lossy entity still hold rows in the new build? Losing a few % at sparse edges (an aggregate-year-item combo emptying because members moved out, or dropping below a coverage threshold) is churn; an entity losing all its rows or going all-NaN is the bug signature — see the must-have trap below.

The samples cap at ~100 rows/diff, so pair the probes with the mechanism argument (only the dep changed) rather than treating them as a row-by-row proof.

The must-have trap. Some steps pin specific member countries as required for a group's aggregate. When the producer reclassifies a pinned country out of its group, the entire aggregate nulls for all years — a total-entity wipe, not edge churn. Two nets catch it: the geo must-have guard raises ValueError at build time (so the build check above fails loudly), and probe 5 is the backstop. The fix is updating the pinned list to the new membership — so whenever bumping income groups, check the repointed consumers' pinned member lists up front. Beware the rationalization risk: a wiped income-group aggregate looks like the expected churn the rest of the report is full of, and on a foundational dataset (population) it also announces itself as a dataset-count explosion — hundreds of consumers suddenly changed. Run probe 5 on every loss, however plausible the churn story.

Check which old version the data-diff compared against. After the old version leaves the DAG, the report can pair the new dataset with another active version of the same name and show tables or columns as removed. If the compared path isn't the real predecessor, compare with the published predecessor directly and note it in the PR body (not with etl diff --include, which keeps only DAG versions when several match):

import pandas as pd
from owid.catalog import Dataset
old = pd.read_feather("https://catalog.ourworldindata.org/garden/<ns>/<old_version>/<short_name>/<table>.feather")
new = Dataset("data/garden/<ns>/<new_version>/<short_name>")["<table>"].reset_index()

Then compare columns, row counts, and values on the shared keys.

Don't diff against stale local builds — a consumer built at an unknown earlier time (or downloaded from the catalog) conflates code drift with the dependency change and produces false positives; REMOTE/production is the trustworthy baseline because CI built it from master.

Complementary, not a replacement: data-diff sees every dataset (including ones with no charts), while Chart Diff shows how the same changes land on actual published charts and Anomalist flags per-country anomalies in the new data — the final QA hand-off (the last step before the announcements) covers all three. Use data-diff to find which datasets/indicators to worry about, then Chart Diff to judge what readers would actually see.

• When you deferred consumers to a follow-up PR: the checks above belong to that follow-up PR; here just confirm the "Downstream dependencies" list is complete (.venv/bin/etlr --modified --dry-run shows the affected set).

Removing the old version & reordering the DAG

After the ETL update, etl update appends the new version entries to the bottom of the main DAG file while the old version's entries stay in their original slot. Always ask the user whether to remove the old version — but never skip this checklist item, and when the user agrees, always do the reorder too.

Timing: dag entries now, files only after the PR review is done. Reviewers compare consecutive versions with the compare-previous-version VS Code extension, which is purely filesystem-based: it diffs the open step file against the same-named file in the nearest lower YYYY-MM-DD sibling folder, and never reads the dag. So split the removal in two:

• Dag entries: remove and archive right away (sequence in step 1 below). etl archive-dag reconstructs from the committed dag history and doesn't care whether the step files exist, and the version tracker only warns about archived steps missing code — archived steps that still have code are fine (that's why the periodic ":bomb: Delete archived step/snapshot code" cleanup commits exist).
• Old step files (.py, .meta.yml, .dvc): keep them until the reviewer signs off, then delete them as the final commit before merge. Don't modify them in the meantime — the extension diffs them verbatim, and any edit pollutes the consecutive-version comparison. Before that final commit, grep the tests for the old version's paths (rg "<ns>/<old_version>" tests/ apps/). A test that reads a real step file — to pick up its dataset.owners, say — fails as soon as the file is gone, and the unit-test build is the only thing that notices. Point such tests at the latest version found by globbing rather than at a new hardcoded one, so the next update doesn't break them again, and run make unittest before pushing the deletion. Paths that only appear as string fixtures need nothing.

The Final QA hand-off (Anomalist / Chart Diff / data-diff) is unaffected by either half (Anomalist's upgrade detectors need the old grapher dataset in the local catalog, not the step files or dag entries).

Workflow when the user agrees:

1. Delete the old version's dag entries. Remove its entries (snapshot → meadow → garden → grapher) from the main DAG file (e.g., dag/poverty_inequality.yml). Its files (etl/steps/..., snapshots/...) stay on disk until review sign-off (see "Timing" above). The archive dag (dag/archive/*.yml) is not edited by hand — etl archive-dag reconstructs it from git history, recording each removed step with the commit where it was last active (for recovery via git checkout).

   • Commit the removal BEFORE running etl archive-dag. The tool reconstructs from committed git history — an uncommitted working-tree removal is invisible to it, and it will instead pick up whatever earlier removals are already committed. Sequence: edit dag → commit → etl archive-dag → scope → commit the archive.
   • Branch runs are squash-safe. When run on a feature branch, the recovery marker points at the merge-base with origin/master (a master commit that survives the squash-merge), branch-only transient steps (created and reverted within the PR) are skipped automatically (archive_dag.skip_branch_only_step in the log), and re-runs don't churn markers that are already valid — only markers whose SHA is unreachable from master get replaced. No post-merge fixup pass is needed; committing more to the PR after archiving is fine.
   • A conflict in dag/archive/*.yml when merging master means both sides appended entries. Regenerate it, don't merge the blocks by hand. Resolve the file to master's version, finish the merge commit, then rerun etl archive-dag: your removal is now in committed history, so it re-appends your dataset's block with a fresh marker. Scope the result as in the next bullet and commit.
   • etl archive-dag reconciles the entire archive, not just your dataset. If the archive was stale, one run can append hundreds of unrelated lines across many archive files, plus marker comments — noise that doesn't belong in a data-update PR and that Codex will question. Keep the commit scoped: after running it, git checkout -- dag/archive/ to drop the unrelated files, then re-add only your dataset's block to dag/archive/main.yml (copy the exact entry archive-dag generated, including its # archived; last active in <sha> on <date> marker). The block you keep is genuine tool output, so it stays consistent with future full regenerations.

2. Move the new entries into the old slot so the dataset stays grouped with its neighbours and section comment. The new entries should not remain at the bottom of the main DAG.

3. Preserve the original section comment (same indentation as the old block) above the new entries.

4. Prefer the nested (compact) DAG format. etl update emits the flat form (each step a separate top-level key with a flat dep list); the loader (etl/dag_helpers.py:_parse_dag_yaml) also accepts the nested form, where the chain is declared inline and flattens to the same graph. The nested form is the team's preferred style and is usually what the archived old block already used:

   data://grapher/<ns>/<v>/<short>:
     - data://garden/<ns>/<v>/<short>:
       - data://garden/regions/2023-01-01/regions
       - data://meadow/<ns>/<v>/<short>:
         - snapshot://<ns>/<v>/<short>.csv
   

   Convert the relocated new entries to nested while reordering, so the active and archived blocks match. Verify it parses with .venv/bin/python -c "from etl.dag_helpers import load_dag; load_dag()" (a malformed nesting raises).

   When only some snapshots bumped, re-declare the unchanged inputs in the new block, including another dataset's whole chain (meadow, garden and snapshot) that the old block declared inline. If they were declared only inside the old nested block, deleting it leaves them undeclared, and the parse check above doesn't notice. Check every step reachable from the new version, not only its direct dependencies, so a re-declared chain with a bare leaf (say, the other dataset's meadow) is caught too:

   .venv/bin/python -c "
   from etl.dag_helpers import load_dag
   dag = load_dag()
   todo, seen = [s for s in dag if '/<namespace>/<new_version>/' in s], set()
   while todo:
       s = todo.pop()
       if s not in seen:
           seen.add(s)
           todo += dag.get(s, ())
   print(sorted(s for s in seen if s.startswith(('data://', 'data-private://')) and s not in dag))  # must be []
   "
   

5. Verify: rg "<namespace>/<old_version>/<short_name>" dag/ -g "*.yml" | grep -v "^dag/archive" returns nothing, and rg "<namespace>/<new_version>/<short_name>" dag/ -g "*.yml" shows the entries only in the main file (under the section comment), not at the bottom.

6. Run make check and commit with 🔨🤖 Remove old <name> entries and reorder DAG.

Expect a Codex false-positive on the archive edit. Because this step touches dag/archive/*.yml, Codex often flags it ("avoid updating archived DAG entries" — the AGENTS.md rule against editing archived files). This is expected: archiving is the explicitly-requested workflow step, and the rule's own "unless explicitly asked" exception applies. Reply citing that and resolve the thread — don't revert the archive. A second recurring flag on this step: Codex warns that removing the old chain will "leave the published chart on archived variables" / "remap charts before versioning the grapher step". If the indicator upgrade already ran on staging (step 7) and the old-variable scan came back empty, reply with that verification (the remapped configs sync to production on merge) and resolve.

Final QA hand-off — Anomalist, Chart Diff, Metadata Diff and data-diff

This is the last step of the pre-review work, and it comes before the announcements (steps 8, 9 and 9b): wait for the user's sign-off here before drafting either one. The old step files are still on disk at this point (deleting them is the final commit after review sign-off; see "Removing the old version & reordering the DAG"). Don't auto-run these — they're human-judgment tools. Hand off the four links so the user can review and click through, together with the 5 most-viewed data pages (see "The 5 most-viewed data pages" below):

• Anomalist — flags variables whose new values diverge from the old version beyond statistical thresholds. Catches accidental scale changes, base-year rebases that propagated the wrong way, and silent drops.

  http://staging-site-<container_branch>/etl/wizard/anomalist
  

  Check the upgrade detectors' coverage before handing off. Anomalist's upgrade_missing / upgrade_change detectors only compare old→new variable pairs from the wizard's variable-mapping table — and the indicator upgrader persists mappings only for charted indicators. If only some of the dataset's indicators are used in charts (the common case), the upgrade detectors silently skip the rest, and a partial mapping suppresses the shortName-inference fallback that would otherwise cover everything. Verify with WizardDB.get_variable_mapping_raw(): if it has fewer pairs than the dataset has indicators, rebuild the full mapping by shortName (old vs. new variables rows by datasetId) and re-run:

  STAGING=<branch> .venv/bin/etl anomalist --anomaly-types upgrade_missing --anomaly-types upgrade_change \
      --dataset-ids <new_dataset_id> --variable-mapping '<full json mapping>' --force
  

  Then spot-check the stored anomalies.dfReduced rows include indicators beyond the charted ones.

  Read upgrade_missing flags before calling them data loss. The detector pairs old and new indicators by shortName, so when the producer re-splits a dimension that is part of the shortName — comparability spells renumbered, a category recoded — the points move from one indicator to its sibling and every move is flagged as missing. Check whether the flagged points reappear under a sibling indicator with identical values before reporting a loss; if they do, it is a relabel to describe in the PR body (and, when the producer's release notes don't mention it, a question for the producer).

  The upgrade detectors also need the old grapher dataset in the local catalog (data/grapher/<ns>/<old_version>/<short>) — the FileNotFoundError names the new dataset id, but the missing files are usually the old version's. If the old chain has already been removed from the DAG (so etlr can't rebuild it), fetch its files straight from the public catalog:

  mkdir -p data/grapher/<ns>/<old_v>/<short> && cd data/grapher/<ns>/<old_v>/<short> && \
    for f in index.json <short>.feather <short>.meta.json; do \
      curl -sL -O "https://catalog.ourworldindata.org/grapher/<ns>/<old_v>/<short>/$f"; done
  

• Chart Diff — shows side-by-side before/after thumbnails for every chart that uses an upgraded indicator. Catches visual regressions the schema-level checks miss (axis ranges, color steps, legend changes).

  http://staging-site-<container_branch>/etl/wizard/chart-diff
  

  An approval is tied to the chart's updatedAt. A later upload to staging of an indicator that carries a presentation.grapher_config, a metadata-only one included, re-renders every chart inheriting from it through the admin API, which bumps that timestamp: those charts drop back to pending and owidbot's chart-diff check fails. Batch such edits before asking for approval, and after one tell the user Chart Diff needs approving again. Uploads that don't touch an indicator grapher_config keep the approvals; review their text changes in Metadata Diff.

• Metadata Diff — the text counterpart to Chart Diff. Compares the rendered reader-facing texts (chart title/subtitle/footnote, description_short, WYSK/description_key) against the same baseline, across charts, MDim views and explorer views. Catches what a config diff structurally cannot: text the chart inherits, so a reworded description_key in a garden step changes dozens of charts while Chart Diff shows nothing.

  http://staging-site-<container_branch>/etl/wizard/metadata-diff
  

  Sections are deep-linkable — ?diff-type=blast for the cross-surface blast radius, ?diff-type=charts / mdims / explorers for one surface at a time. Rejections export as metadata-rejections.md, naming the edit, the garden .meta.yml it was authored in and the dataset owner; nothing here gates the merge, so an unactioned rejection is an open item. Full description in /generate-metadata ("Reviewing the change on staging (Metadata Diff)"). owidbot's PR comment links it too, once the metadata-diff service is enabled in the deploy config.

• data-diff report — dataset/indicator-level value comparison of the staging build against production, ranked by anomaly score (see "Silent-breakage check" for how to read it). Covers every dataset the update touched, including ones with no charts — the perspective Anomalist and Chart Diff don't have. Easiest access: the full report link in owidbot's data-diff PR comment. The direct URL keeps the branch name's dots and underscores and replaces only characters outside [A-Za-z0-9._-] (e.g. /) with - — unlike the staging subdomain, which does replace ./_ (and it is not the truncated container name either):

  https://catalog.ourworldindata.org/diffs/<sanitized_branch>/data-diff.html
  

etl approve needs both sides pointed at the right databases: STAGING for the branch, ENV_FILE_PROD for production — and the tunnel the latter expects is usually down. The command reads pending diffs from and writes approvals to the global environment, which only targets the branch staging DB when STAGING is set — run it without STAGING=<branch> (or STAGING=1 from the branch checkout) and it quietly operates on your local dev DB instead, finding no pending diffs or approving in the wrong place. The production side comes from ENV_FILE_PROD (ENV_FILE_PROD=.env.live STAGING=<branch> .venv/bin/etl approve …) — without it the command asserts immediately. .env.live typically points at 127.0.0.1:3310, an SSH tunnel that is often not running; the reliable path is Tailscale, where the production grapher DB is reachable directly as host prod-db on port 3306 with the same live_grapher credentials (the pattern add-ivs-indicators/scripts/indicator_admin_table.py already uses). Being on Tailscale redirects nothing by itself — OWIDEnv.from_env_file() reads DB_HOST/DB_PORT straight from whatever file ENV_FILE_PROD names — so make the redirect real before running the commands below: either set DB_HOST=prod-db / DB_PORT=3306 in .env.live once (worth making permanent), or derive a one-off copy (gitignored by the .env.* pattern) and pass ENV_FILE_PROD=.env.prod-db instead:

sed -e 's/^DB_HOST=.*/DB_HOST=prod-db/' -e 's/^DB_PORT=.*/DB_PORT=3306/' .env.live > .env.prod-db

Note the approval writes to the staging chart-diff table and only reads production, so the prod connection is read-only.

Bulk-approve the easy chart diffs with etl approve before handing the rest to the human. On a dataset with many charts (e.g. WDI has 400+), most pending diffs exist only because the update changed no values a human needs to eyeball — either the underlying data is byte-identical (a version bump minted new variable IDs but the values didn't change) or it changed by a negligible source-revision amount. Reviewing those by hand in Chart Diff is wasted effort; let etl approve clear them first:

ENV_FILE_PROD=.env.live STAGING=<branch> .venv/bin/etl approve --dry-run                       # exact data match only — safe default, see counts first
ENV_FILE_PROD=.env.live STAGING=<branch> .venv/bin/etl approve --dry-run --allow-small-changes  # also count tiny source revisions (see below)
ENV_FILE_PROD=.env.live STAGING=<branch> .venv/bin/etl approve --allow-small-changes            # apply for real once the dry-run counts look right

• Plain etl approve only approves a chart when every dimension's underlying data is byte-identical between staging and prod (it hashes each dimension's actual data, not the raw variable ID — so a version bump that changed no values still gets approved).
• --allow-small-changes additionally approves charts where the only remaining difference is a handful of small-magnitude value changes (typical source revisions) — tune with --tolerance-pct (default 1% relative change per point), --tolerance-abs-floor (default 1e-6, guards near-zero values), --max-changed-points (default 5 per dimension, above which it's sent to manual review regardless of magnitude), and --max-new-points (default 1000 per dimension — new-coverage points, e.g. a fresh year, are given a generous allowance since they're expected from a routine update, but an unexpectedly large coverage jump still gets sent to manual review). It still requires every other part of the chart's config (title, subtitle, everything but the dimension values) to be identical.
• --show-data-diff prints the actual before/after values for skipped charts (per dimension: y/x/size/color) instead of just a hash mismatch — useful to see why a specific chart didn't qualify, or to sanity-check whether raising --tolerance-pct would be safe. Combine with --chart-id <id> to inspect one chart.
• Whatever's left after etl approve is what's actually worth a human's attention in Chart Diff — genuine content changes, added/removed country-year coverage, or other config differences.
• After an indicator upgrade, etl approve approves nothing — not even charts whose config is untouched. The upgrader saves each upgraded chart's config without the fields that equal a default: grapher's schema defaults, and every dimension display field the new indicator already sets. Production still carries them, so the strict equality fails on every upgraded chart, and --allow-small-changes fails the same way at its config pre-check. Before concluding a chart's config changed, compare with both layers filled back in, in this order: each dimension's display over its indicator's own display (variables.display plus unit/shortUnit), then the schema defaults. Filling only the schema defaults turns a missing tolerance into the default 0 and reports a change wherever the indicator sets 5.
• On a release that adds new observations, expect zero data-identical charts. New survey-years or a new year touch every country-level indicator, so "approve where config and data are unchanged" is legitimately empty. The useful split is then data changed, config unchanged (quick to review in bulk) versus config changed (look at each); hand the user the second list with what changed in each.
• Check for production edits made after the staging server was created — approving reverts them. Staging copied production when the branch's server was built; a colleague's edit to the same chart on production after that point is absent from the staging copy, and the upgrade's save carries the old config forward. Approving the diff then syncs the stale config back and silently undoes their edit. For every upgraded chart, compare production's charts.updatedAt against the last pre-upgrade revision on staging (chart_revisions); for any chart edited on production since, restore the edit on staging before approving.

Important: derive <container_branch> correctly. The staging hostname is not simply staging-site-<branch>. The container name is produced by get_container_name(branch) in etl/config.py:

1. Replace /, ., _ with - in the branch name.
2. Strip a leading staging-site- if present.
3. Truncate to the first 28 characters (Cloudflare DNS limit).
4. Strip any trailing -.

Branches over 28 chars therefore get clipped. Example: data-military-expenditure-2026 (30 chars) → container data-military-expenditure-20 → hostname staging-site-data-military-expenditure-20. The simplest way to get the correct value is to call the helper:

.venv/bin/python -c "from etl.config import get_container_name; print(get_container_name('<branch>'))"

Include owidbot's data-diff summary in the hand-off so the user knows the scale before clicking: pull the <summary><b>data-diff</b>: …</summary> line from owidbot's PR comment (gh pr view <num> --json comments) — e.g. ❌ 21 changed · 2 new · 4 identical · 16 skipped. If it reports removed datasets or errors, call those out explicitly.

Tell the user something like: "Final QA: please review Anomalist, Chart Diff and Metadata Diff in the Wizard, and the data-diff report for the dataset-level view — owidbot's summary: ❌ 21 changed · 2 new · 4 identical · 16 skipped. If anything looks off, let me know and I'll investigate."

These pages need a fresh staging build, so they're only meaningful after the PR's grapher upload to staging has completed and the staging server has rebuilt. Metadata Diff is the one that tells you when the server is behind — a 🚧 banner naming each stale dataset and the etlr grapher://grapher/<dataset> --grapher that fixes it. Read that banner before trusting any count: a stale dataset reports its diffs backwards, showing its older text as this branch's change.

The 5 most-viewed data pages — a human read, not a check you run. Most readers meet this dataset through a handful of pages, so those few deserve a person's full attention before the update ships. Find the 5 published charts using the new indicators with the most views over the past year. Run it against the branch's staging DB after the indicator upgrade, so the charts already point at the new version. Escape literal underscores in the path as \_, as in step 7. "Published" means the live isPublished flag in the config: charts.publishedAt is the first-publish date and stays set after a chart is unpublished. Use Python rather than make query here, because Make would swallow the $ in '$.isPublished':

STAGING=<branch> .venv/bin/python - <<'EOF'
from etl.config import OWIDEnv

df = OWIDEnv.from_staging("<branch>").read_sql(
    """
    SELECT DISTINCT cc.slug, COALESCE(ap.views_365d, 0) AS views_365d
    FROM charts c
    JOIN chart_configs cc ON cc.id = c.configId
    JOIN chart_dimensions cd ON cd.chartId = c.id
    JOIN variables v ON v.id = cd.variableId
    LEFT JOIN analytics_pageviews ap
      ON ap.url = CONCAT('https://ourworldindata.org/grapher/', cc.slug)
      AND ap.day = (SELECT MAX(day) FROM analytics_pageviews)
    WHERE v.catalogPath LIKE %(path)s
      AND cc.config->>'$.isPublished' = 'true'
    ORDER BY views_365d DESC
    LIMIT 5
    """,
    params={"path": r"grapher/<namespace>/<new_version>/<short_name>/%"},
)
print(df.to_string(index=False))
EOF

Give the user the 5 staging links (http://<container_name>/grapher/<slug>), most-viewed first with their yearly views, and ask them to open each page and read it the way a reader would: is this the best version of this page we can produce? Can anything be fixed, improved, rewritten or made clearer, so it introduces our readers to this data as well as it can? Don't pre-review these pages or offer your own verdict on them — the point is a person looking with fresh eyes, and an AI summary next to the links invites a skim. Then wait until the user says they've read all 5, make the changes they ask for, and only then close the hand-off. If the dataset has fewer than 5 published charts, hand over the ones it has.

In the same message, remind the user that when the dataset is important and the update isn't urgent, it helps to ask a colleague, ideally one who isn't an expert in the topic, to read the WYSK ("What you should know") of these 5 pages and say what was unclear, what they had to read twice, and what they expected to find but didn't. The user decides whether that applies; it's a suggestion, so the hand-off doesn't wait for the colleague's feedback.

Close every hand-off with an explicit list of what's still open — in the PR body, not just chat (chat scrolls away; the PR is what the reviewer and future-you actually read). This is one of the two cases where the list is worth writing out formally rather than as a sentence; see .claude/docs/open-items.md for what tends to get dropped. An update generates far more loose ends than it closes; this workflow's usual danglers: content/gdoc edits, producer error reports and any Metadata Diff rejection still unactioned (handed off — metadata-rejections.md is addressed to the dataset owner and nothing in the merge enforces it), reader-facing config changes held for sign-off (awaiting a decision), audits offered but not run and anything the staging build didn't cover (nobody checked it) — plus a fourth category, deferred to a follow-up PR, where the two-PR pattern is in play (downstream consumers to repoint, old-version archiving, charts on retired indicators): that list is the follow-up PR's scope, so losing it means redoing the analysis.

Before saying the update is ready to merge, confirm every checklist item that names a skill ran that skill, including the two background audits, with their findings handled. Run anything missing now, and wait for any audit still running.

Right before merging, check that the producer hasn't published a newer release since the snapshot. Updates take days, and some producers release monthly. Read the current release date where the snapshot got date_published (the API's publication-date attribute, the dataset page's stamp) and spot-check a few values against the snapshot. A newer release goes back to the user before the merge.

Guardrails and tips

• When you fix or edit step code, keep its steps labeled: a short comment per stage of run(), plus the why where the code doesn't show it, a line or two each (CLAUDE.md, "Comments in steps"). Add labels to an unlabeled step you have to read closely anyway, and shorten long comment blocks rather than adding to them.
• When you fix or edit step code, use pr.* / snap.* over pd.* wherever they have an equivalent (CLAUDE.md, "Preserving metadata/origins"). In particular, in a step that loads the snapshot (usually meadow), build tables from parsed rows with snap.read_from_records(rows), not Table(pd.DataFrame(...)) followed by .copy_metadata(...), and replace that pattern when you find it there. A garden step has no snap: derive its tables from the input Table (pr.merge, pr.concat, groupby) so the input origins carry over, and use pr.read_from_records(rows, columns=[...]) only for helper tables with no origin of their own (a mapping, a lookup). Don't swap a garden .copy_metadata(tb) for an originless pr.read_from_records: that drops the origins it was carrying.
• Don't push to the branch while a long-running snapshot extraction runs on its staging server. Every push redeploys that server, and the deploy runs make docs.build, whose first command is rm -rf site/ .cache/ — and .cache/ is also ETL's CACHE_DIR, where extraction scripts keep intermediate files. The extraction then fails with a FileNotFoundError on a file it wrote seconds earlier. Hold every push (commit on the server if you must) until the run finishes, or run it from a separate checkout that does not share .cache/ (an editable install pins BASE_DIR, so a second checkout reusing the main venv still writes into the main .cache/). Say so in the snapshot script's docstring when it is meant to run on staging.
• END_YEAR / "as of" framing for status/event datasets. When a dataset records events (and derives a status time series) and its latest event year lags the release date, you face a choice: forward-fill the latest status to the release year, or stop the series at the last event year and note the "as of" date in metadata. Prefer the latter — forward-filling invents data points for years with no source information (and shifts an END_YEAR-style constant ripples through the whole series). Keep the series at the last real year and add the currency note to description_processing and a description_key bullet (e.g. "The legal status shown for each country reflects the situation as of <Month Year>."). Confirm the choice with the user; they may change their mind.
• Converting a period label to a date: use the dates the producer states, don't re-derive their calendar. Non-calendar periods (fiscal quarters, ISO weeks, crop or school years, survey waves) come with real start/end dates in the producer's own releases. A rule inferred from a few examples — "the last Sunday of the month", "the first Monday" — will fit most periods and quietly miss the rest, and this error survives every schema, bound and sum check, because the values are all correct and only the dates they sit on move. So pin the derivation: keep a handful of period-end dates quoted from the producer and assert the code reproduces them, the way a codebook's worked examples serve as test vectors.
• OECD SDMX dataflow versions bump on new releases — a pinned URL goes 404/NoRecordsFound. The Data Explorer's "Developer API" links pin df[vs]/the REST path to a dataflow version (e.g. DSD_SHA@DF_SHA,1.0); when the producer publishes a new edition they may mint 1.1 and empty the old version, so last cycle's known-good URL returns NoRecordsFound. On that error, list versions with GET /public/rest/dataflow/<agency>/<id>/all?detail=allstubs. For reader-facing links (url_main, /latest posts) prefer the version-less explorer deep link (data-explorer.oecd.org/vis?df[ds]=DisseminateFinalDMZ&df[id]=<id>&df[ag]=<agency>), which always resolves to the latest release; in the snapshot's url_download, pin the version (deterministic) and expect to bump it each cycle.
  • Check every update, not only on errors — and don't trust "newest" blindly. The old version can stay up (no error at all) while a new one exists, and a new version can be listed while the OECD is still loading it, with a partial latest year (only a few donors) or whole runs of years missing. So before moving to a newer version, compare it with the pinned one: the same years and donors/recipients (no gaps), the latest year reasonably complete, and no unexplained breaks. Totals for overlapping years may legitimately differ, because a new release can revise history — surface those revisions and check that they are plausible (spread across the series rather than whole years or entities dropping out); a missing year or entity is a partial load, a shifted total is not on its own. A data URL with an empty version (...,DSD_X@DF_X,/key) silently takes the latest, partial or not — pin instead, and when a script queries the API, have it log the pinned and latest versions and warn when they differ.
  • DAC dataflows also link to versioned hierarchies of donors and recipients (HCL_DACDONORS, HCL_DACRECIPIENTS, named in the dataflow's HIER_CONTEXT annotation). A new hierarchy version carries the new World Bank income classification, and the OECD's own group totals follow whichever version the dataflow links to, which may be the older one. If you use the groups, decide which classification you want, and check group totals built from their members against the OECD's.
  • Release date: the explorer's "Last updated" stamp is in the API. It equals the validFrom of the dataflow's ContentConstraint type="Actual" (GET .../rest/dataflow/<agency>/<id>/<version>?references=contentconstraint, in UTC; the page shows local time). Use it for date_published. Versioned bulk-file names (...-v20260710) can be a day off.
• Re-test "manual upload" snapshots — the blocking may be inverse-UA. When a snapshot's docstring says the file is uploaded manually because "the website blocks the download request", verify that claim before carrying it into the new version. Some hosts (e.g. the IMF Datamapper) reject browser-like User-Agents with 403 while letting plain, honestly-identified clients through — the inverse of the usual bot-blocking — and the ETL downloader's default UA (DEFAULT_USER_AGENT in etl/download_helpers.py) is browser-like, so the original author may have misdiagnosed an automatable source. Test both directions (plain requests vs. browser UA) against the direct file URL. If the plain UA works: set url_download in the .dvc and pass user_agent="owid-etl/1.0 (https://ourworldindata.org)" (or similar plain UA) to snap.create_snapshot(...). Keep the snapshot .py script in that case — the script-less .dvc-only path (run_snapshot_dvc_only) calls create_snapshot() without a user_agent and would 403 — and say so in the docstring so nobody deletes it as "redundant".
• Switching a snapshot's source format or access method (a portal file to an API, wide to long): check the derived indicators and the units, not only the raw values. APIs often omit empty observations that a wide file kept as blank cells, so anything that counts rows or treats a missing row as zero (shares of partnerships, counts of reporters, completeness rates) shifts while every raw value still matches. Rebuild the full panel in garden if the logic depends on it, and compare those indicators against the previous version. APIs also often report base units where the file used thousands or millions, with the scale given only as a display attribute (SDMX SCALE, UNIT_MULT), so drop the old conversion and assert a known magnitude (a large country's total within a plausible range).
• Large monetary totals: save with repack=False when float32 would round them. paths.create_dataset repacks Float64 columns to float32 when the values agree within a relative tolerance of 1e-5 (series_eq in lib/repack), and float32 keeps about 7 significant digits, so totals in the trillions of dollars move by hundreds of thousands. When the data diff shows such rounding against the previous version, pass repack=False in each step that saves those columns.
• Manual-upload snapshots: also re-check for a stable download endpoint. Distinct from the inverse-UA case above: producers add direct links over time, so a snapshot that genuinely required a manual download last cycle may be automatable now. Check the producer's download page or API for a stable (ideally version-less) URL before carrying the manual flow forward; if one exists, convert the snapshot to a script-less url_download .dvc. Multi-file archives stay script-less too: snapshot the archive itself and read the member file in meadow via snap.extracted(). If the bundle includes a codebook or series-metadata file, consider passing it through meadow as an extra table and attaching per-indicator description_from_producer in garden — inventory its fill rates first, and skip fields owned elsewhere (units, license, dataset-level boilerplate).
• "Constant YYYY" labels must match the actual base period — whose convention that is depends on who deflates. For a real series OWID computes, verify the rebasing base is the YYYY annual average — prefer the producer's own annual-average row (e.g. BLS's M13 period) — not a single month that happens to fall in YYYY: a January base labeled as the year is ~5% off in a fast-inflation year, passes every schema check, and ships a mislabeled level. State the rebasing explicitly in description_processing. For a constant-price series the producer supplies, do the opposite: confirm and document their stated base-period methodology (some legitimately define "constant YYYY" against a specific month, quarter, or chain-linked reference period while naming the year) — never rebase their values to an annual average to make the label fit.
• Metadata rewrites don't reach charts that pin their own FAUST. When an update changes methodology or scope wording in indicator metadata (grapher_config.note, description_short, deflator/source claims), query the authored layer (the chart_configs row named by patchConfigId) for every affected chart (published and draft) to detect which FAUST fields are pinned: any chart whose patch overrides title/subtitle/note/sourceDesc keeps the stale text no matter what the indicator now says (sourceDesc is the chart's own source line; it overrides the indicator's attribution, so an old year range in it survives every update). Propose the minimal text edit (usually one sentence), ask before changing (reader-facing FAUST), then apply via the staging admin API with the full config as the payload — fetch the rendered config (gm.Chart.config, or chart_configs.config via charts.configId), edit the field, and pass that to apps.chart_sync.admin_api.AdminAPI.update_chart; the server re-derives the stored patch by diffing against the inherited layer, so inheritance survives. Do not PUT the bare patch: for charts inheriting from an indicator grapher_config, fields present only in the rendered config are silently dropped (in-repo callers like apps/indicator_upgrade submit chart.config, the resolved full config, for exactly this reason). Verify on the staging SVG render, and make sure the chart's diff gets approved in Chart Diff before merge: chart-sync copies a staging chart to production only when its diff is approved (apps/chart_sync/cli.py checks diff.is_approved; pending diffs are skipped), so an unapproved FAUST edit silently stays behind on staging while the PR merges. After the production bake, spot-check the live chart SVG for the new wording. When many charts pin the same sentence with an embedded year (deflator base years especially), swap by literal substring (constant 2023 US$ → constant 2024 US$), never by year-regex — neighbouring years in the same note (a GDP PPP year, a methodology-break year) must not move. And if the pinned sentence will restale at every release, propose the durable fix alongside: move it into indicator metadata via a definitions variable (e.g. constant {definitions.inflation_year} US$) and clear the chart-level overrides in a follow-up. Clearing a pinned FAUST field to "restore inheritance" is only safe if the field is actually in the inherited layer — check first. Chart FAUST inherits from presentation.grapher_config only; presentation.title_public does not enter the chart-config inherited layer, so a chart whose sole override is title has nothing to fall back to, and clearing it leaves the chart titleless. Read the layer before touching the chart — variables.patchConfigIdETL → chart_configs.config shows exactly which keys it carries. When the field is missing, the fix is two steps in order: add it to presentation.grapher_config in garden, rebuild and upload (grapher:// for the MySQL upsert), confirm it now appears in the inherited layer, and only then clear the chart-level override by PUTting the full config with that field set to the inherited value (identical values drop out of the patch). Verify on the rendered SVG, not just the stored config.
• Partial-year preliminary releases: sweep per-series max years and guard single-time views. When a producer's preliminary release adds a new year to only some series, check each charted indicator's latest year rather than the dataset's (metadata.json → dimensions.years). Stacked time-series charts self-protect — grapher drops a time point whose stack is incomplete — but a single-time discrete view (minTime: "latest" + StackedDiscreteBar) happily renders the partial year, and display.tolerance backfills the missing series with prior-year values, so the bar looks complete while silently mixing vintages (for shares, mixing denominators). Never conclude "the chart has all the data" from the render — check the series' own max years. Guard such views with timelineMaxTime: <last complete year>, not maxTime (a maxTime pin only moves the default; the slider still reaches the bad year), and remove the clamp at the detailed release.
• Audit blanket transformation rules per indicator before trusting them. Any garden rule that applies one transformation to a pattern-matched group of indicators — unit scaling by title keywords, sign conventions, currency or magnitude conversions — assumes the source stores the whole group in a single convention. Sources mix conventions, especially when different indicators come from different upstream producers. Scan each matched indicator's raw range to confirm it fits the assumed convention, split the rule where it doesn't, and guard both sides with sanity checks (inputs within the assumed convention; outputs within a data-grounded bound). "Matches the previous version" is not evidence of correctness — magnitude bugs are inherited from the old step; judge absolute plausibility against real-world values, not just old-vs-new equality. (For example, a ×100 rule shared by fraction-stored shares and ratios the source already stores in percent inflates the latter 100×.)
• Programmatic metadata: still curate the charted indicator(s). When a dataset's indicator metadata is generated in code (titles/units inferred from source names), the indicators actually used in charts deserve an explicit per-indicator block in the garden .meta.yml — description_short, description_key, display — layered on top (YAML merges per-field, so code-set fields like description_from_producer survive). Ground the bullets in the producer's codebook/methodology PDF rather than inferring from the data: the codebook yields the precise scope, the exact numerator/denominator, and caveats you won't guess (e.g. WWBI approximates the EEA 2004–2018 public sector from industry classifications). Leave a # NOTE: in the YAML explaining the programmatic/curated split for the next maintainer.
• Unguessable producer filenames: find the document index instead of brute-forcing URLs. When a producer's download filenames follow no stable convention, stop guessing after one or two 404s and look for the machine-readable index behind their JavaScript-rendered page — the page HTML itself is usually an empty shell, so curl/WebFetch on it finds nothing, but the index the page calls returns JSON with one entry per document (Q4 Inc investor sites, common for US-listed companies, expose it under /feed/). Record the endpoint as a # NOTE: in the snapshot so the next updater looks it up rather than repeating the search, and note that the same index answers "was any older file moved or renamed" — comparing its entries against the hardcoded URLs is a cheap moved-file check. It is not an in-place-revision check: a file replaced under the same URL looks identical in it, so for that use the index's per-document dates where it carries them, otherwise the file dates/hashes in the in-place-revision guardrail below.
• Scraped chart embeds: the page's own data tables are the canonical source — embed CDNs lag. When a snapshot's data lives in a chart embedded on the producer's page (Datawrapper and similar), don't fetch the chart platform's CDN endpoint (datawrapper.dwcdn.net/<id>/<version>/dataset.csv): the latest published chart version can trail the page by a full release, and the snapshot diff won't flag it. Producer pages server-render each embed's data as an HTML fallback table (<noscript><table>), so parse that instead: whole-page pd.read_html(io.StringIO(resp.text)), select the table whose columns exactly match the expected header, assert exactly one match. Related trap: the producer's visible "Updated" stamp and prose can lag their own data tables — trust the data, and when the newest rows have no stamped release date, date_published falls back to date_accessed (document why in a .dvc comment).
• Mixed-type object columns at meadow: when pd.read_csv produces an object column that mixes strings and NaN (common for sparse text columns like sources/comments/punishments), the feather repacker rejects it. Cast those columns to pandas "string" dtype before tb.format(...).
• paths.regions auto-resolves DAG dependencies: paths.regions.add_population(tb) and paths.regions.add_aggregates(tb, regions=[...]) pick up the population and income_groups datasets directly from the DAG. Don't paths.load_dataset("population") and pass it through unless the helper specifically asks for the dataset — the parameter is unused.
• WB income-group aggregates: add the four classification names (High-income countries, Upper-middle-income countries, Lower-middle-income countries, Low-income countries) to your REGIONS list and add data://garden/wb/<latest>/income_groups to the DAG. paths.regions.add_aggregates(...) auto-resolves the classification.
• Detect structural placeholders dynamically: when a source ships "balanced panel" rows that are zero everywhere by design (status combos that exist only for completeness), detect them at runtime (groupby(...).max() == 0) and assert the count matches the codebook. A coding change in the source then surfaces as a test failure instead of silently shipping noise.
• In-place source revisions: compare file dates/hashes, not version labels. Some producers replace the published file (and codebook) without bumping their stated version and without a changelog. When checking whether a source has a new release, look at the hosting platform's file-modification dates or hashes (e.g. the OSF API's date_modified per file) against the previous snapshot's date_accessed/md5 — an unchanged version label proves nothing. If you ingest such a revision, set date_published to the replacement date and leave a .dvc NOTE naming the behavior so the next updater re-checks. If the replacement appeared after etl update created the new version, move the steps to a version dated on or after the new date_published, so the folder isn't dated before its data.
• Category-state churn in combined categorical indicators. A source revision can add or remove the states a categorical indicator takes (a recode wave eliminating an enforcement state, a new cross-combination appearing). The net that catches it is metadata validation failing with "extra variables in YAML file / in table" on per-category regional variables. Fix all four surfaces together — the category map in step code, the sort: lists, the coding-description sentences, and the per-category regional _count/_pop YAML blocks — then re-run the phantom-category audit. Never keep a label the new data cannot produce.
• Guard silent catch-all buckets with an assert. A lookup that routes labeled source values into an existing fallback category ("requirement unknown", "other") degrades silently when the source adds a new label — the row lands in the fallback and nothing fires (unlike unmapped values that leak as new categories, which metadata validation catches). Assert observed labels ⊆ map keys, letting only genuinely-blank values reach the fallback.
• Grep metadata prose for numbers carried from the old release. Validated fields are covered by checks; prose numbers in description_key/descriptions (country counts, category counts, year ranges) are not. After a revision that changes panel composition, rg the new .meta.yml for the previous release's counts. The mirror of this check for published content that cites the dataset is the referencing-prose audit in step 7.
• corrections.yml overrides on categorical columns must use the source's current vocabulary. Assigning a retired label raises TypeError: Cannot setitem on a Categorical with a new category — pick the current-vocabulary value that maps to the same published output, and note the substitution in the entry.
• Codebook-vs-data inconsistencies: when the codebook documents one thing but the actual CSV shows another (placeholder claimed but non-zero rows present, etc.), preserve the data as-shipped and flag it in the PR description for the producer to confirm. Don't silently force the data to match the codebook.
• Grapher .meta.yml only when it adds something: the grapher step inherits everything via default_metadata=ds_garden.metadata, so drop the grapher .meta.yml if it only duplicates the garden values. Keep it only for genuine grapher-side overrides.
• Per-indicator description_processing reads better than a generic shared note: when an indicator is derived (combined-categorical buckets, regional aggregates, computed counts), spell out that indicator's derivation. Reusing named definitions for shared boilerplate is fine; just compose them into per-indicator sentences rather than dropping a single generic note across all indicators.
• description_key in definitions.common propagates only to indicators without their own list: if you want a bullet to appear on every indicator, either keep it on common.description_key and don't define per-indicator lists (it inherits), or prepend it explicitly to each per-indicator list (treats it as a "first bullet" pattern).
• Phantom-category audit on categorical indicators: after building categorical indicators, sweep every indicator and compare YAML sort: labels against the unique values that actually appear in the data. Phantom labels (declared in sort: or in a category map but never produced) clutter chart legends with empty buckets. Either drop them from sort: and description_key, or remove them from the map if they can never occur given the data shape. Re-run the audit on every data refresh — phantoms can reappear when a category is dropped upstream.
• NOTE: comments for the next maintainer when behaviour is data-conditional: when something in the code holds only because of the current data shape (e.g. "only 4 indicators have an EoE=0 row", "only Brazil 2025 is a transition-year artefact"), leave a # NOTE: comment near the relevant block asking the next data update to re-audit. Helps future maintainers spot which assumptions might decay before they bite.
• Indicator Upgrader CLI for one-shot chart remaps: when v1 → v2 short_names change so much that the auto-upgrader can't match them, drive the remap manually. Write a small script that calls WizardDB.add_variable_mapping(mapping={old_id: new_id, ...}, dataset_id_old=..., dataset_id_new=..., comments="...") with the explicit pairs, then run from apps.indicator_upgrade.upgrade import cli_upgrade_indicators; cli_upgrade_indicators(dry_run=True) to preview affected charts, and (dry_run=False) to apply. Mappings stay in the wizard DB until WizardDB.delete_variable_mapping() is called, so a slug-collision failure can be recovered by fixing the slug and rerunning the upgrade — only un-upgraded charts get reattempted. The active staging DB is inferred from the current git branch.

When the update is review-heavy and you need iterative back-and-forth with a topic owner over staging, see the report-indicator-changes skill for drafting the message.

Artifacts (expected)

• workbench/<short_name>/snapshot-runner.md
• workbench/<short_name>/progress.md
• workbench/<short_name>/notes_to_check.md (one entry per carried-over # NOTE: / # TODO:, plus detected sanity_checks functions and their log-control flags)
• workbench/<short_name>/sanity_checks.log (only if step 5b ran)
• workbench/<short_name>/meadow_diff_raw.txt and meadow_diff.md
• workbench/<short_name>/garden_diff_raw.txt and garden_diff.md
• workbench/<short_name>/harmonization.log and harmonization_audit.md (from step 5c)
• ai/adversarial-review-<short_name>-<date>.md (from step 6c-bis)
• workbench/<short_name>/indicator_upgrade.json (if indicator-upgrader was used)
• workbench/<short_name>/update-context.yml (canonical facts gathered during the update; consumed by draft-data-update-slack-post)
• workbench/<short_name>/slack-announcement.md
• workbench/<short_name>/data-update.md (public-facing post draft for OWID /latest, from step 9b)

Example usage

• Minimal catalog URI with explicit old version:
  • update-dataset data://snapshot/irena/2024-11-15/renewable_power_generation_costs 2023-11-15 update-irena-costs

---

Common issues when data structure changes

• SILENT FAILURES WARNING: Never return empty tables or comment code as workarounds — fix the parsing/transformations instead!
• Column name changes: If columns are renamed/split (e.g., single cost → local currency + PPP), update:
  • Python code references in the garden step
  • Garden metadata YAML (e.g., food_prices_for_nutrition.meta.yml)
  • Grapher metadata YAML (if it exists)
• Index issues: Check for unwanted index columns from reset_index() — ensure proper indexing with tb.format(["country", "year"]).
• Metadata validation: Use error messages as a guide — they show exactly which variables to add/remove from YAML files.