Reported issues for cern-opendata-mcp-server

Pod holds 15 of 15 GitHub reports that passed its relevance review. This can include external user reports, maintainer-confirmed bugs, and concrete feature gaps. Treat them as evidence to inspect, not a count of distinct defects.

Back to cern-opendata-mcp-server.

Most discussed

feat(cern_opendata_search_records): category, keyword and LHCb magnet-polarity/stripping filters

Use case

The portal classifies 66,345 datasets, nearly all simulated (CMS, DELPHI, ATLAS), by physics process in categories (Higgs Physics › Standard Model, Exotica › Dark Matter, ...) and tags LHCb datasets with magnet polarity and stripping stream and version. cern_opendata_search_records exposes none of them: query reaches them only through unlisted field forms (categories.secondary:"Top physics"), and the tool drops their facets, so a caller can neither discover the…

Read the thread · 2026-10-01 · closed · 3 comments

bug(cern_opendata_get_records): citation drops the authors when no collaboration is set

Server version

0.1.0

mcp-ts-core version

0.13.10

Runtime

Node.js

Runtime version

Node 26.5.0

Transport

HTTP (Streamable HTTP)

Description

citation.text takes its creator from collaboration.name only, so a record that names authors but no collaboration gets a citation with no creator, opening with a bare (2014).. The portal's own "Cite as" for the same record lists the authors. Affects cern_opendata_get_records and the cern-opendata://record/{recid}…

Read the thread · 2026-10-01 · closed · 2 comments

bug(cern_opendata_list_files): on-demand records report "This record has no files"

Server version

0.1.0

mcp-ts-core version

0.13.10

Runtime

Node.js

Runtime version

Node 26.5.0

Transport

HTTP (Streamable HTTP)

Description

For records whose availability is ondemand, the portal's record API returns no _files, files or _file_indices, though distribution.number_files and distribution.size are set. cern_opendata_list_files then answers "This record has no files.", which reads as an empty dataset rather than one whose files sit on…

Read the thread · 2026-10-01 · closed · 2 comments

bug(cern_opendata_list_files): indexed files without a key make the whole record unreadable

Server version

0.1.1

mcp-ts-core version

0.13.10

Runtime

Node.js

Runtime version

Node 26.5.0

Transport

stdio

Description

cern_opendata_list_files rejects a whole record when one indexed file has no key. Every file in atlas-160006's eight file indexes (309 files, the JetSet2 release) carries only availability, checksum, filename, size and uri, and the index entries state no number_files and an empty availability. toCompactFile requires…

Read the thread · 2026-10-02 · closed · 1 comment

bug(cern_opendata_get_validated_runs): a CMS collision dataset with no list is told lists exist for CMS collision data only

Server version

0.1.1

mcp-ts-core version

0.13.10

Runtime

Node.js

Runtime version

Node 26.5.0

Transport

stdio

Description

cern_opendata_get_validated_runs explains every dataset that links no good-run list the same way: lists "exist for CMS collision data only, so simulated, non-CMS and non-collision records have none". For a CMS collision dataset that is false. Record 93950 (/ZeroBias/Run2017E-v1/RAW) is CMS collision data and has no list because no…

Read the thread · 2026-10-01 · closed · 1 comment

bug(cern_opendata_search_trigger_paths): paths without the HLT_ prefix (AlCa_, DST_, output modules) are unreachable

Server version

0.1.1

mcp-ts-core version

0.13.10

Runtime

Node.js

Runtime version

Node 26.5.0

Transport

stdio

Description

cern_opendata_search_trigger_paths prepends HLT_ to every path and its input pattern requires the prefix, but 365 of the portal's 4,135 CMS trigger path records (as of 2026-10-01) are named otherwise: 79 AlCa_ and 52 DST_ paths, 4 DQM_, 214 output modules (…Output), and 16 others such as HLTriggerFinalPath. None of them can…

Read the thread · 2026-10-01 · closed · 1 comment

bug(htmlToText): a bare < in portal HTML drops text up to the next >

Server version

0.1.1

mcp-ts-core version

0.13.10

Runtime

Node.js

Runtime version

Node 26.5.0

Transport

stdio

Description

htmlToText in src/services/cern-opendata/text.ts treats every < as the start of a tag, so a bare < in portal HTML drops all text up to the next >. Selection cuts written as |eta| < 2.4 are common in CMS dataset methodology. Record 5202 loses most of a sentence in content[], while structuredContent, which keeps the HTML as…

Read the thread · 2026-10-01 · closed · 1 comment

feat(cern_opendata_get_records): return variable dictionaries, physics categories, pile-up and LHCb run conditions

Use case

cern_opendata_get_records and cern-opendata://record/{recid} drop record metadata that answers common questions directly: what the variables in a derived ML sample mean, which physics process a simulated sample belongs to, which pile-up sample it was mixed with, and which magnet polarity and stripping an LHCb dataset uses.

Related: #11, #12

Proposed behavior

Add these Record fields from the search hits the lookup already reads, as received and absent when the record…

Read the thread · 2026-10-01 · closed · 1 comment

Most recent

feat(cern_opendata_get_records): cache lookup hits so deferred and body_offset walks don't re-read the portal

Use case

Related: #11

With the 64,000-byte response budget, a cern_opendata_get_records call returns what fits and lists the rest under deferred, and a long documentation body is read in 30,000-character slices with body_offset. Every follow-up call looks its ids up again, so the walk re-downloads hits the previous call already read:

Read the thread · 2026-10-02 · open · 0 comments

bug(cern_opendata_get_records): no response budget across records, and doc bodies past 30,000 characters are unreachable

Server version

0.1.1

mcp-ts-core version

0.13.10

Runtime

Node.js

Runtime version

Node 26.5.0

Transport

HTTP (Streamable HTTP)

Description

cern_opendata_get_records cuts each documentation body at 30,000 characters but never bounds the response: 20 long LHCb stripping pages return 628,420 bytes of structuredContent and 622,551 of content[] text, and ordinary records add up too (a CMS collision dataset costs a median 8 KB). The rest of a cut body is…

Read the thread · 2026-10-01 · closed · 1 comment

bug(cern_opendata_list_files): records with the largest manifests cannot be listed within the call budget

Server version

0.1.1

mcp-ts-core version

0.13.10

Runtime

Node.js

Runtime version

Node 26.5.0

Transport

HTTP (Streamable HTTP)

Description

cern_opendata_list_files reads the whole record (GET /api/records/{recid}) even when index names one file index, and cuts each attempt at 30 s within the 50 s call budget. Record 24464 is 16.3 MB (70 indexes, 32,618 files); once the portal needs over 30 s to send it, the retry restarts the download with about 19 s left…

Read the thread · 2026-10-01 · closed · 1 comment

bug(cern_opendata_search_records): malformed query answered with portal 500 surfaces as a generic upstream error

Server version

0.1.1

mcp-ts-core version

0.13.10

Runtime

Node.js

Runtime version

Node 26.5.0

Transport

HTTP (Streamable HTTP)

Description

The portal answers some malformed query strings with HTTP 500 instead of its 400 The syntax of the search query is invalid., varying per request: (foo drew 500 on 6 of 12 identical requests, title:( on 5 of 8. The service retries a 500, so when all three attempts get one the caller receives a generic…

Read the thread · 2026-10-01 · closed · 1 comment

bug(cern_opendata_search_trigger_paths): titles naming several datasets leave the suffix in path; notices ignore year

Server version

0.1.1

mcp-ts-core version

0.13.10

Runtime

Node.js

Runtime version

Node 26.5.0

Transport

HTTP (Streamable HTTP)

Description

218 CMS trigger records from 2011–2013 name two or three primary datasets in their title, such as High-Level Trigger path information HLT_Mu17_Mu8 (DoubleMu, DoubleMuParked datasets). The title parser knows only the singular ({Primary} dataset) suffix, so path keeps the whole suffix and dataset is absent. The…

Read the thread · 2026-10-01 · closed · 1 comment

bug(cern_opendata_search_records): sort mostrecent returns oldest first, title_desc returns A-Z

Server version

0.1.1

mcp-ts-core version

0.13.10

Runtime

Node.js

Runtime version

Node 26.5.0

Transport

HTTP (Streamable HTTP)

Description

cern_opendata_search_records describes sort: "mostrecent" as newest first, but it returns the oldest records first; title_desc (Z-A) returns A-Z, the same order as title. The portal's REST API takes a sort's direction only from a - prefix and ignores the default_order its sort options declare, so mostrecent…

Read the thread · 2026-10-01 · closed · 1 comment

bug(recid): records with prefixed recids (atlas-160006, cms-93956) are unreachable by recid

Server version

0.1.1

mcp-ts-core version

0.13.10

Runtime

Node.js

Runtime version

Node 26.5.0

Transport

HTTP (Streamable HTTP)

Description

The portal now mints recids as {experiment}-{number} (its release validator requires the form for new records); seven records carry one as of 2026-10-01, including atlas-160006 (JetSet2) and cms-93956. cern_opendata_search_records returns that recid, but every recid input accepts only 1–12 digits:…

Read the thread · 2026-10-01 · closed · 1 comment

The remaining reports are on the project's issue tracker.