govuk-orgchart

name	govuk-orgchart
description	Crawl, extract and query the GOV.UK government org chart -- ministerial roles, the people who hold them, the organisations they serve, and past office-holders. Use whenever a question is about who holds which UK ministerial role, which department they sit in, past holders of an office, or cross-joining with the Parliament Members API or Wikidata. Loads its query capability via a bundled script (no MCP server, no boot-time tokens).
allowed-tools	Bash(python3 ${CLAUDE_SKILL_DIR}/scripts/query.py *) Bash(./scripts/govuk_sparql_serve.sh *) Bash(python3 scripts/govuk_*.py *)
license	Open Government Licence v3.0 (upstream gov.uk content); MIT (this skill's crawler / extractors / query layer)
metadata	{"provenance":{"tier":3,"operator":"forgetmenot (crawl + extraction)","service":"gov.uk (Government Digital Service / Cabinet Office)","upstream-data":"GOV.UK /government/ pages (HTML + parallel /api/content/ JSON), Crown copyright under OGL v3.0","citation-short":"GOV.UK (via forgetmenot govuk-orgchart extraction)","citation-formal":"GOV.UK org-chart pages (Government Digital Service / Cabinet Office); crawl + heuristic extraction by forgetmenot, retrieved {date}","confidence":"derived","confidence-notes":"Triples from upstream JSON-LD / RDFa / microdata = derived (confidence ≈ gov.uk's own structured data); factoids = heuristic (BeautifulSoup templates over gov.uk markup; QA harness in third_party/govuk/html/orgcharts/extractors/factoids/qa.json reports per-template accuracy)."}}

GOV.UK org-chart skill

GOV.UK publishes a public org chart across pages under /government/: ministerial roles, the people who hold them, organisations (departments, agencies, NDPBs), and historical figures. Every public page is double- served as HTML and as structured JSON at the parallel /api/content/ path -- both are crawled and stored locally so extraction can be re-run offline. The corpus lives at third_party/govuk/html/orgcharts/.

Querying (the on-demand part)

To answer an org-chart question, run the bundled query script. It uses the local rdflib-endpoint at 127.0.0.1:8765 if it's up, otherwise parses the N-Quads in-process (slower first time, no setup needed):

python3 ${CLAUDE_SKILL_DIR}/scripts/query.py 'SELECT ?holder WHERE {
  GRAPH ?g {
    <https://www.gov.uk/government/ministers/chancellor-of-the-exchequer>
      <https://forgetmenot.local/govuk#roleHolder> ?holder .
  }
} LIMIT 5'

Common prefixes are auto-prepended (govuk:, schema:, parl:, owl:, dcterms:, xsd:). For a quick dump of everything we know about a URI:

python3 ${CLAUDE_SKILL_DIR}/scripts/query.py --describe https://www.gov.uk/government/people/rachel-reeves

--tsv switches output to tab-separated for easier reading in chat.

This pattern (bundled script invoked from the skill body) is the shipped way to give a session a capability without paying the always-on MCP cost: nothing about SPARQL or rdflib enters context until the skill is invoked.

Pipeline (when extending or refreshing the corpus)

# 1. Crawl. Priority-frontier crawler: roles -> people -> orgs -> announcements.
python3 scripts/govuk_crawl.py --resume --max 500 --workers 5
python3 scripts/govuk_crawl.py --resume \
    --wikidata-seeds third_party/data/wikidata/data/people-bridge.jsonl
        # supplementary seeds for politicians the natural link graph misses

# 2. Fetch the parallel /api/content/ JSON for each crawled page.
python3 scripts/govuk_fetch_api.py --workers 5

# 3. Extract. Templates + API parser; per-page Turtle, rolls up to N-Quads.
python3 scripts/govuk_extract_factoids.py --refresh --workers 8
python3 scripts/govuk_extract_triples.py     # generic RDFa/JSON-LD/microdata

# 4. (Optional) Long-lived SPARQL endpoint for faster repeated queries.
./scripts/govuk_sparql_serve.sh              # 127.0.0.1:8765/

# 5. QA: anchor cases + cross-checks. Exits non-zero on regression.
python3 scripts/govuk_qa.py

# 6. Visual report (PDF).
python3 scripts/govuk_report.py

What's in the RDF

Prefix	URI	Role
fm:	https://forgetmenot.local/vocab#	This project's vocabulary -- everything we model ourselves
govuk:	https://www.gov.uk/vocab/meta#	Reserved for <meta name="govuk:*"> attributes gov.uk publishes directly. We do NOT invent under this namespace -- see docs/vocab.md.
schema:	http://schema.org/	Standard
dcterms:	http://purl.org/dc/terms/	Standard

Per-page Turtle holds:

• schema:Person, fm:Organisation / schema:GovernmentOrganization, fm:MinisterialRole, fm:RoleTenure
• fm:CurrentOfficeHolder / fm:FormerOfficeHolder (driven by the Content API's explicit current=true|false flag)
• govuk:contentId (stable UUID), govuk:schemaName, govuk:publishingApp, dcterms:modified -- structural identifiers literally lifted from <meta name="govuk:*"> tags
• Provenance qualifiers: fm:apiSourced true (from /api/content/ JSON) vs fm:proseExtracted true (from biography prose regexes). Two independent extractors run and their disagreements surface in QA.
• Cross-page relations: fm:hasMinister, fm:holdsRole, fm:roleHolder, fm:previouslyHeldBy, fm:previouslyHeldRole, fm:partOf, fm:tenureStart, fm:tenureEnd

The rollup at third_party/govuk/html/orgcharts/extractors/factoids/all.nq puts every triple in a named graph equal to the gov.uk page URL it came from, so the file itself answers "which gov.uk page is this fact from?". If raw source text gets attached for LLM-verification, it lives in a separate named graph (<page-url>#raw convention) — never inline with the extracted triples. See docs/vocab.md.

Cross-corpus joins

Direct URL-string or name joins are brittle (departments rename, people add suffixes). The repo carries two bridge layers:

• Wikidata bridge at third_party/data/wikidata/. The people-bridge.ttl is owl:sameAs between gov.uk person URIs and Wikidata QIDs, with parl:memberId for the Members API. Refresh with third_party/data/wikidata/scripts/refresh.py.
• gov.uk content-id -- the extractor lifts the publishing-app UUID onto every page subject so URL renames don't break joins within the gov.uk corpus.

Worked example -- voting record of the current Chancellor in one SPARQL query, gov.uk → Wikidata → parliament.uk, no name matching:

PREFIX govuk: <https://forgetmenot.local/govuk#>
PREFIX owl:   <http://www.w3.org/2002/07/owl#>
PREFIX parl:  <https://id.parliament.uk/schema/>

SELECT ?holder ?wikidata ?parliamentId WHERE {
  GRAPH ?g {
    <https://www.gov.uk/government/ministers/chancellor-of-the-exchequer>
      govuk:roleHolder ?holder .
  }
  ?holder owl:sameAs    ?wikidata ;
          parl:memberId ?parliamentId .
}
# -> Rachel Reeves, Q574896, 4031

Data-quality discipline

This skill operates under the data-quality discipline: anchor cases, cross-extractor checks, provenance qualifiers on every triple, upstream issues for source bugs (danbri/forgetmenot#20), and a hard rule against cobbling hybrid query workarounds. Run scripts/govuk_qa.py before declaring an extraction change "done".

Coverage caveats

• gov.uk's past-PMs index lists Sunak/Truss/Johnson as cards without dedicated /history/past-prime-ministers/<slug> URLs (they appear in details.appointments_without_historical_accounts in the Content API). _extract_pms_index_from_api() reads them.
• 13% of the Wikidata-bridge politician set is not reachable from the current org chart's natural link graph; pass --wikidata-seeds to govuk_crawl.py to close.
• Welsh-language .cy URLs are intentionally skipped (translations, not primary records).

On MCP

scripts/govuk_mcp_sparql.py is registered in the project .mcp.json as the govuk-sparql server -- intentionally eager-loaded because SPARQL over this corpus is a core capability for any org-chart question. Two tools:

• sparql_query(query) -- run any SPARQL SELECT or ASK against the corpus, prefixes auto-prepended.
• sparql_describe(uri) -- shortcut for every triple mentioning a URI.

Tool Search defers individual tool schemas until invoked, so the boot-time cost is the server registration, not the full tool surface.

The bundled scripts/query.py here exists as a fallback for sessions without MCP available (e.g. running the pipeline outside Claude Code, or in a subagent context where MCP servers aren't proxied).

Before either path: the rdflib HTTP endpoint must be running --

./scripts/govuk_sparql_serve.sh   # 127.0.0.1:8765/

Provenance to cite

Tier 3 — third-party. GOV.UK is operated by the Government Digital Service (Cabinet Office). The crawl, extraction and query layer are ours. Upstream content is under OGL v3.0; our heuristics can be wrong.

• Inline cite: "(GOV.UK, via govuk-orgchart extraction)" — once per paragraph in user-facing answers.
• When stating a factoid (role-holder, office, start/end date), prefer hedged phrasing ("per GOV.UK as of {date}"). QA results for the extractor live at third_party/govuk/html/orgcharts/extractors/factoids/qa.json.
• When the data is a triple lifted from gov.uk's own JSON-LD / RDFa / microdata, cite it as such — that's gov.uk's structured data, not our parsing. Triples / factoids are tagged differently in the extractor output.
• Cross-corpus joins (gov.uk content_id ↔ parliament.uk member id ↔ Wikidata QID) flow through the bridge files at third_party/data/wikidata/data/people-bridge.{jsonl,ttl}. Cite both endpoints when joining: e.g. "Yvette Cooper (parliament.uk member 4514; gov.uk /government/people/yvette-cooper; Wikidata Q333062), via the forgetmenot bridge".
• Never present heuristic factoids as authoritative.
• See ../../docs/provenance.md for the cross-skill rules.