{
  "version": 3,
  "created_at": "2026-07-08T14:39:05.284Z",
  "task_summary": "Exhaustive multi-database EoE literature review: disease, immune mechanisms, pMHC, and targeted arms",
  "agents": [],
  "phases": [
    {
      "name": "Plan",
      "delegations": [
        {
          "steps": [
            {
              "title": "Scope, taxonomy & multi-database query matrix",
              "description": "Fix the review taxonomy for four parts — (I) exhaustive EoE (epidemiology, natural history, clinical presentation, diagnosis/histology, endoscopy, therapeutics, guidelines, outcomes/QoL, EoE beyond esophagus/EGIDs); (II) immune cell types & mechanisms (eosinophils, mast cells, CD4/Th2, ILC2, iNKT, basophils, dendritic/antigen-presenting cells, B cells & IgG4, Tregs, epithelium & barrier, fibroblasts/remodeling, cytokine axes IL-4/5/13/TSLP/IL-33, IFN-γ); (III) pMHC diagnostics & therapeutics (antigen-specific T cells, TCR discovery, tetramers/multimers, AIM assays, tolerance/anergy, pMHC therapeutics across immunology); (IV) targeted arms (EoE genetics/HLA/GWAS, atopic march & prevention, teplizumab/T1D template, diagnostic status quo & unmet need, approved+pipeline therapeutics). Build a query matrix (6-12 queries per subtopic) spanning PubMed, OpenAlex, bioRxiv/medRxiv, and ClinicalTrials.gov. Write the taxonomy + query matrix to review_scope.json. Deliverable: the scope file (no artifact yet).",
              "title_short": "scope"
            },
            {
              "title": "Multi-database systematic retrieval",
              "description": "Execute the full query matrix across PubMed (host.mcp mcp-pubmed — abstracts, PMIDs), OpenAlex (search_openalex kernel helper), and bioRxiv/medRxiv (host.mcp mcp-biorxiv) via the repl tool, writing per-query hits to handoff JSON incrementally (DOI, PMID, title, authors, year, venue, cited_by, abstract where returned, source_db, query, theme). Pull the EoE therapeutic pipeline from ClinicalTrials.gov (mcp-clinical-trials) separately for Part IV. Expect thousands of raw hits before dedup. Save raw pulls to handoff/retrieval_raw/*.json. Print per-database, per-theme raw counts.",
              "title_short": "retrieve"
            },
            {
              "title": "Deduplicate & build master reference library",
              "description": "Merge all raw pulls; deduplicate by normalized DOI (fallback PMID/title-match) across databases and themes; keep the union with a multi-theme tag list per paper and a source-database list. Verify a sample of DOIs resolve (verify_dois helper) and flag any Crossref retraction (update-to). Save eoe_master_reference_library.csv (target several hundred to 800+ unique DOI-verified references) with columns: doi, pmid, title, authors, year, venue, cited_by, themes, source_dbs, tier(blank), retraction_flag. Print total unique, per-theme counts, year distribution.",
              "title_short": "dedup+library"
            },
            {
              "title": "Citation-graph expansion on landmarks",
              "description": "Identify the ~25-40 landmark papers (highest cited_by per theme + known anchors: Alexander 2014, Kottyan genetics, Blanchard 2006, Dilollo/Hill 2025, Herold 2019, Du Toit 2015, Wen 2019, Rothenberg atlas). For each, walk one step backward (references) and one forward (cited-by) via expand_citations; fold new on-topic hits into the master library (re-dedup). This surfaces the seminal roots and the newest extensions a keyword sweep misses. Update eoe_master_reference_library.csv (version 2). Print how many new references the expansion added per theme.",
              "title_short": "citation-expand"
            },
            {
              "title": "Relevance screen & evidence tiering",
              "description": "Screen every library entry with an abstract for relevance + assign Tier-1 (landmark/primary), Tier-2 (supporting primary), Tier-3 (review/context/guideline) using parallel host.llm over abstracts (batched, platform kernel; triage only, not prose) with a rubric; hand-confirm the Tier-1 set. Drop off-topic keyword false-positives (flag, don't delete). Update the tier + relevance columns (library version 3). Print the Tier-1 roster per part with one-line relevance notes and the tier histogram.",
              "title_short": "screen+tier"
            },
            {
              "title": "Part I — exhaustive EoE review",
              "description": "Write Part I as referee-grade prose (paragraphs opening on synthetic claims, inline DOI links, organized by theme not paper): epidemiology & rising incidence, natural history & fibrostenotic progression, clinical presentation across ages, diagnosis & histologic criteria (15 eos/hpf, EoEHSS), endoscopic features (EREFS), the EGID family & EoE beyond the esophagus, treatment overview (the 3 D's: drugs/diet/dilation), guidelines evolution (AGREE/consensus), and outcomes/quality-of-life burden. Tie to this project's prior-phase omics where relevant. Save eoe_review_part1_disease.md; run style_pass once before saving.",
              "title_short": "part I"
            },
            {
              "title": "Part II — immune cell types & mechanisms",
              "description": "Write Part II — the mechanistic deep dive — as prose per cell type/axis: eosinophils (recruitment via CCL26/CCR3, effector functions, EPX), mast cells (SIGLEC6+ disease state, tryptase, mast-epithelial crosstalk), CD4/Th2 & pathogenic effector Th2, ILC2s, iNKT cells, basophils, dendritic/antigen-presenting cells & epithelial MHC-II, B cells & local IgG4, Tregs & tolerance, epithelial barrier (DSG1, FLG, SPINK7) & TSLP/IL-33 alarmins, fibroblasts & remodeling, and the IL-4/IL-13/IL-5 and IFN-γ cytokine axes. Anchor each to primary literature and this project's single-cell/meta-signature findings. Save eoe_review_part2_immune_mechanisms.md; style_pass once.",
              "title_short": "part II"
            },
            {
              "title": "Part III — pMHC diagnostics & therapeutics",
              "description": "Write Part III covering the emerging pMHC field (primary + reviews, since new): antigen-specific T-cell detection (tetramers/multimers, AIM assays), the Hill-Spergel EoE antigen-specific-T-cell/TCR platform (eoeTCR-4, prior-art/patent framing), food-specific T cells across allergy, pMHC-based diagnostics, and pMHC/antigen-specific therapeutics for tolerance/anergy/deletion across immunology (tolerogenic approaches, multimer-based, T1D/celiac/MS precedents). Honest boundary: what is EoE-specific vs borrowed from adjacent fields. Save eoe_review_part3_pmhc.md; style_pass once.",
              "title_short": "part III"
            },
            {
              "title": "Part IV — targeted arms",
              "description": "Write Part IV as four linked sections: (a) EoE genetics/HLA/GWAS (TSLP, CAPN14, susceptibility-vs-effector split, HLA-II presentation); (b) atopic march & prevention with the teplizumab/T1D template (reconcile with atopic_march_prevention_review.md, no duplication); (c) diagnostic status quo & unmet need (elimination diet + serial endoscopy burden, biomarkers); (d) approved + pipeline therapeutics (PPI, swallowed topical steroids/budesonide, dupilumab, cendakimab, benralizumab, others — with the ClinicalTrials.gov pipeline pull). Save eoe_review_part4_targeted.md; style_pass once.",
              "title_short": "part IV"
            },
            {
              "title": "Evidence-map / gap-analysis figure",
              "description": "Render eoe_evidence_map.png (publication-grade, apply_figure_style): a theme × evidence-strength matrix showing per-subtopic reference counts by tier, the temporal depth (year span) of each theme, and explicit flags where evidence is thin or absent (EoE-specific prevention trials, DQ epitope benchmarking, pMHC-therapeutic clinical data in EoE). A companion panel maps the preprint's core claims to their strongest supporting tier. Save the figure.",
              "title_short": "evidence-map"
            },
            {
              "title": "Consolidate master review + gap analysis",
              "description": "Assemble a master document eoe_exhaustive_literature_review.md: executive scope statement (and the honest 'comprehensive-from-open-APIs' ceiling), a navigation index linking Parts I-IV, a cross-cutting synthesis of what is established vs contested vs open, a dedicated gap-analysis section anchoring each gap to what establishes it, and the consolidated DOI-verified reference list with tiers. Cross-reference the evidence-map figure. Carry standing disclaimers (computational predictions; Hill-Spergel prior-art/patent; no clinical validation).",
              "title_short": "consolidate"
            },
            {
              "title": "Bundle the exhaustive-review package",
              "description": "Tar the master document, Parts I-IV, the master reference library CSV, and the evidence-map figure into eoe_exhaustive_litreview_bundle.tar.gz; save all as artifacts (separate save_artifacts per language). Close with the final reference count, per-part and per-tier breakdown, database provenance, and the key honest gaps the preprint must foreground.",
              "title_short": "bundle"
            }
          ]
        }
      ],
      "id": "phase-0"
    }
  ],
  "desired_outputs": [
    "Deduplicated DOI-verified master reference library (CSV, theme-tagged, tiered)",
    "Part I — exhaustive EoE review (markdown)",
    "Part II — immune cell types & mechanisms review (markdown)",
    "Part III — pMHC diagnostics & therapeutics review (markdown)",
    "Part IV — targeted arms: genetics/HLA, prevention/atopic march, diagnostic gap, therapeutic pipeline (markdown)",
    "Evidence-map / gap-analysis figure",
    "Bundled review package (tar.gz)"
  ],
  "feasibility": {
    "rationale": "Four literature databases confirmed available (PubMed, OpenAlex, bioRxiv/medRxiv, ClinicalTrials.gov) plus Crossref full-text fetch and citation-graph expansion — all free, no paid credits or GPU needed. Parallel host.llm relevance-screening of abstracts is the one step that could draw inference budget; it runs on the platform kernel and is used only for triage, not prose. Scope is genuinely large (target 400-800+ unique DOIs across all arms); main limits are paywalled full text (abstract-level synthesis where needed) and API rate limits (paced, incremental saves). 'Most exhaustive of all time' is the aim; honest ceiling is 'the most comprehensive synthesis achievable from open bibliographic APIs in this session,' which is stated in the review.",
    "confidence": "high"
  }
}