Skip to content

2.3.3.md

Latest commit

 

History

History
18 lines (13 loc) · 3.68 KB

2.3.3.md

File metadata and controls

18 lines (13 loc) · 3.68 KB
Raw
summary Fix format() drop-offs across 4 tools; fall back to per-ID calls when batch rejects on a single bad NCT
breaking false

2.3.3 — 2026-04-21

Fixed

  • analyze_trial_landscape prompt — focusAreas argument unusable — The arg was declared as z.array(z.string()), but per the MCP spec, prompts/get sends arguments as Record<string, string> with no wire format for arrays. Sending an array failed protocol-layer validation; sending a comma-string failed the tool-layer schema. Changed focusAreas to a comma-separated z.string(); generate() splits on ,, trims, and filters. Verified via HTTP: "phases, sponsors" now produces Focus the analysis on: **phases, sponsors**. Closes #27.
  • clinicaltrials_get_study_record — format() still dropped hasResults, resultsSection, derivedSection, documentSection — Follow-up to #18. Added four render blocks: **Has Results:** with a pointer to clinicaltrials_get_study_results when true; **MeSH Conditions:** and **MeSH Interventions:** from derivedSection.*BrowseModule (prefers browseLeaves, falls back to meshes); a ## Results Summary stub with counts for outcome measures, adverse events, participant-flow periods, and baseline measures; a ## Documents (N) section with kind tags (Protocol/SAP/ICF) and upload dates from documentSection.largeDocumentModule.largeDocs. Extended RawStudyShape with documentSection and browseLeaves. Verified live on NCT03722472 (renders all four) and NCT02194738 (confirms MeSH Interventions path with 16 drug/procedure terms). Closes #28.
  • clinicaltrials_get_field_values — format() silently dropped up to 235 values — Hard .slice(0, 15) on topValues with no tail message. structuredContent carried up to 250 values, but content[] consumers lost everything beyond the 15-value cap. Added an explicit tail: … and N more values in structuredContent (of X unique; topValues capped at 250). Documented the 250 cap in the output schema's .describe(). Verified on LeadSponsorName (49,996 unique values → 15 rendered + tail message). Closes #30.
  • clinicaltrials_get_field_definitions — format() dropped field description — The field tree's description is the one place in the server's tool surface where authoritative data-model documentation lives, but format() never read the property, so LLM-only clients lost the context. Added a <description> line beneath each field when present, for both top-level and nested renders. Verified on protocolSection.identificationModule — nctId now renders its canonical description in content[]. Closes #31.
  • clinicaltrials_get_study_results — partial-invalid batch cascaded failure to valid NCTs — Follow-up to #23. The batch endpoint (filter.ids) rejects the whole request if any single ID is malformed or nonexistent, and the handler mapped that batch-wide error to every input ID — so a single bad ID dropped all valid results. Changed the catch block to fall back to sequential per-ID getStudy calls (honors the service's ~1 req/sec rate limit); only truly failing IDs now land in fetchErrors. Added erroredIds Set to prevent the downstream missing-ID loop from double-counting fallback failures. Verified on [NCT03722472, NCT05956821, NCT00000000]: valid IDs succeed, only NCT00000000 errors. Closes #32.

Changed

  • RawStudyShape — Extended derivedSection.conditionBrowseModule and interventionBrowseModule with browseLeaves: Array<{ id?, name?, relevance? }>; added top-level documentSection.largeDocumentModule.largeDocs: Array<{ filename?, hasIcf?, hasProtocol?, hasSap?, label?, typeAbbrev?, uploadDate? }>. All additive — no breaking changes.