cern-docs

name: cern-docs description: Use when the user asks how a CERN service or ATLAS / FTS framework operates and the answer should come from canonical operator documentation — what the official batch docs say about HTCondor / lxbatch, SWAN session settings and HTCondor pool, CERN Cloud (OpenStack) flavors and quotas, ML@CERN training and serving, ATLAS Athena / ASG / Tier-0 / databases, ATLAS computing and grid production, FTS3 (File Transfer Service) configuration / installation / REST API / messaging / monitoring. Backed by the cerndocs MCP (search_docs for BM25 search, fetch_doc for one page) over 8 indexed sources: atlas-sft, atlas-computing, atlas-databases, batch, cloud, ml, swan, fts. Does NOT cover HTCondor operational use — submitting, inspecting, killing jobs, or the Python bindings (use htcondor for that; cern-docs covers documentation queries about HTCondor), the fts-rest-* CLI itself (use fts-rest), Open Data dataset / recid / DOI lookup (use cern-opendata or atlas-opendata), live job-state inspection (use reana or shell tools), or multi-service infra recommendations (use infra-advisor). Disambiguator phrase: CERN docs BM25 multi-source. data_scope: both experiment: all

cern-docs — canonical CERN service & ATLAS-software documentation lookup

Upstream MCP: https://cern-mkdocs-mcp.app.cern.ch/mcp

This skill is a thin wrapper around the cerndocs MCP server. The server indexes eight CERN docs sites (BM25 cached for 24 h) and pulls page bodies live from the upstream git repos on demand. Two site shapes are covered transparently: MkDocs (search payload at /search/search_index.json) and legacy GitBook v2/CLI (no search payload — the server walks SUMMARY.md and indexes every linked markdown page).

When to load this skill

The user wants to know how a CERN service or ATLAS / FTS framework operates as documented. Examples:

• "How do I submit an HTCondor job from lxplus?"
• "What does the SWAN session timeout default to?"
• "Where are the GPU flavors documented in CERN Cloud?"
• "How does ATLAS Athena versioning work?"
• "What endpoints does ml.cern.ch's serving expose?"
• "How is the FTS3 server configured?"
• "What does the FTS3 docs say about OAuth2 token support / S3 / multihop?"

Do NOT load this skill for:

• Driving the FTS REST CLI (fts-rest-transfer-submit, fts-rest-whoami, …) → use fts-rest.
• Open Data dataset / recid lookup → use cern-opendata (portal records) or atlas-opendata (DSIDs, MC metadata).
• Live job-state inspection (a running REANA workflow, a current condor job's status) → use reana or the shell tools.
• Composite multi-service infra recommendations ("what stack should I use to do X end-to-end") → use infra-advisor.

Source IDs

The MCP exposes one source parameter to route queries to one of eight indexed sites. Pick the source that matches the user's domain; when ambiguous, search more than one — each source is its own BM25 index and a wrong-source query can miss a hit silently.

The MCP also exposes a docs://sources resource that lists the registered sources at request time — call it if you suspect the source list has grown beyond the eight above.

Tool usage

Step 1 — always: search_docs

search_docs(query, source, limit=10) — BM25 search. Returns title + URL + 200-char snippet only. Token-cheap. Always start here regardless of source.

Step 2 — source-class dependent

The eight sources split into two classes based on how their content is hosted:

Class A — Markdown-backed (atlas-sft, atlas-computing, atlas-databases, fts) The MCP can resolve a search hit's URL back to a .md source file in the upstream repo and fetch it. fetch_doc returns full page bodies:

• fetch_doc(url, mode="outline") — headings only. Confirm the right page first.
• fetch_doc(url, mode="sections:<heading>") — one section; use when the question is scoped to a known heading.
• fetch_doc(url, mode="markdown") — full page body. Most expensive; use only when the question genuinely needs the whole page.

For fts, the resolver follows GitBook conventions: …/docs/overview.html resolves to docs/overview.md; directory URLs and …/index.html resolve to the matching README.md.

Class B — search-only (batch, cloud, ml, swan) These sources are hosted outside the GitLab path resolver. fetch_doc returns "No matching source file" for all URLs from these sources.

• Use search_docs with a higher limit (limit=20) to cast a wider net.
• The 200-char snippet per result is the only retrieval mechanism available.
• When the question is narrow (e.g. a specific config option), rephrase and re-run rather than increasing the limit further.
• Cite the public URL from the search hit even when you can't fetch the body.

Output rules — what makes it into the user reply

• Cite the public docs URL that the search hit returned (e.g. https://swan.docs.cern.ch/sessions/ or https://fts3-docs.web.cern.ch/fts3-docs/docs/overview.html). The user can open it.
• Quote the relevant prose from the fetched body inline; do not link the user to the MCP URL or to any internal page identifier.
• If the MCP returns a Recovery Guide (auth-gated source with no token, unknown source ID), surface the recovery instructions to the user verbatim — they are operator-grade help text.

Example flows

"How does Athena versioning work?" (atlas-sft — Class A)

1. search_docs(query="Athena release versioning", source="atlas-sft")
2. Pick the top hit, fetch_doc(url=<hit URL>, mode="outline") to confirm.
3. fetch_doc(url=<hit URL>, mode="markdown") for the full body.
4. Quote the relevant section, cite the URL.

"How do I configure FTS3 for OAuth2 tokens?" (fts — Class A, GitBook)

1. search_docs(query="OAuth2 token configuration", source="fts")
2. fetch_doc(url=<hit URL>, mode="outline") to scout the page.
3. fetch_doc(url=<hit URL>, mode="sections:Configuration") for the relevant section.
4. Quote the section, cite the rendered .html URL.

"How do I submit a condor job from lxplus?" (batch — Class B, search-only)

1. search_docs(query="submit job from lxplus", source="batch", limit=20)
2. Extract the answer from snippets; cite the returned URL.
3. If snippets are thin, re-query with a tighter phrase: search_docs(query="condor_submit .sub file lxplus", source="batch", limit=20).

"What's the default SWAN session timeout?" (swan — Class B, search-only)

1. search_docs(query="session timeout default", source="swan", limit=20)
2. Extract from snippets; cite the returned URL.
3. (fetch_doc is not available for swan — snippets are the ceiling.)

Cross-source question ("compare HTCondor on batch vs SWAN's HTCondor pool"): run search_docs separately against batch and swan, present both. Don't let the BM25 ranker silently hide one source.

Limitation: FTS REST CLI docs

The FTS3 docs index covers the markdown under gitlab.cern.ch/fts/documentation/docs/. The build pipeline of the upstream site also pulls REST CLI / Python-bindings docs from a separate repo (cern-fts/fts-rest) into fts-rest/docs/ at publish time — those pages are not in this MCP's index (cross-repo inclusion isn't followed). For REST CLI command syntax, prefer the fts-rest skill (which has the command catalogue and --help output), or fall back to the upstream rendered URL the user can open in a browser.