Knowledge-Graph-Hub

Knowledge modeling review of KG-Microbe transforms and merged KG for alignment with METPO, Biolink Model, and KGX specification. Use when auditing transform output quality, validating categories/predicates, checking CURIE prefix registration, or preparing a release.

Add a new data source transform to KG-Microbe end-to-end — from "should we even ingest this?" through deep research, cross-reference analysis, semantic design, code scaffold, integration, verification, and shipping the PR. Bakes in the CLAUDE.md 7-step checklist plus the research + analysis + verify phases that keep the KG coherent. Use whenever the ask is "ingest X", "add source Y", or "integrate a new database into kg-microbe".

Ship a messy topic branch as a set of clean, focused PRs. Walks through triage of committed commits + working-tree modifications + untracked files, extracts misfiled commits to their own branches, splits orthogonal changes into separate PRs, gitignores build noise, opens follow-up issues for deferred items, and merges in the correct dependency order. Use when a working branch has accumulated mixed commits, uncommitted work, and dozens of untracked scratch files and needs to reach master.

Emit a Codex-ready review prompt for the KG-Microbe repository focused on code logic, consistency, robustness, bugs, bottlenecks, and scalability. Use before delegating a deep review pass to Codex (via the codex:rescue subagent) or another external code-review agent so the target is precisely scoped and the review dimensions are enforced.

Determine whether local KG-Microbe transform outputs (data/transformed/<source>/) and merged KG (data/merged/) are current relative to origin/master. Compares latest commit times on origin/master touching each transform's code directory against local output mtimes; also checks merge stage against merge_utils/, merge.yaml, and every transform output. Use before cutting a release, before running kg-release, or when triaging "why did my merged KG change".

Work with KG-Microbe's unified chemical mapping system (`mappings/kgmicrobe_unified_entity_mappings.sssom.tsv.gz` and `kg_microbe/utils/chemical_mapping_utils.py`). Use when adding a new mapping source, regenerating the unified file, debugging a missing ChEBI lookup, validating mappings against OLS, or reasoning about which source wins when sources disagree.

Render a GTDB phylogenetic diagram from a KG-Microbe merged release with each clade sized by the count of non-taxonomy edges (phenotypes, growth media, chemicals, etc.) incident on it. Folds NCBITaxon and kgmicrobe.strain edges onto their GTDB equivalent via in-graph close_match, GTDB metadata, and the published NCBI2GTDB tables. Persists the resolved mapping and a gap report. Use when you need to see *where in the GTDB tree the metadata is concentrated* — which clades are well-characterized vs sparse.

Generate a structured Markdown report on every post-transform and post-merge operation needed to take KG-Microbe from per-source TSVs to the final shipped data products (merged KG, RDF copy, release tarballs). Reports each operation's purpose, command, inputs, outputs, severity, and current freshness against the on-disk repo state. Use when planning a release, onboarding, auditing what is stale, or producing a "what's left to ship" punch list.

Use when curating or expanding a KG-Registry resource page. Applies to stub resources and incomplete existing pages stored as Markdown with YAML front matter under resource/.

Use when asked everywhere a given resource is used or consumed downstream — which knowledge graphs, aggregators, or products draw on it, and any recorded real-world uses — based on what KG-Registry records. Reads the live hosted KG-Registry Parquet files over HTTP (no local clone needed), combining reverse provenance (other resources naming this one as a source) with the resource's own usages field.

Use when asked to find KG-Registry resources similar to a given one — alternatives, related graphs, or near-duplicates. Reads the live hosted KG-Registry Parquet files over HTTP (no local clone needed), combining structured signals (shared domains, taxa, category, node/edge scale, shared upstream sources) with softer heuristics (description similarity and overlapping publication authors) into a ranked shortlist.

Use when asked to find KG-Registry resources about a topic, subject area, or domain (e.g. "everything about RNA interactions", "what proteomics knowledge graphs are there", "find resources mentioning Reactome"). Queries the live hosted KG-Registry Parquet files over HTTP — no local clone needed — mapping a free-text topic to the registry's domain vocabulary plus keyword search.

Use when asked what upstream data sources a given resource (usually a knowledge graph) is built from — its inputs, dependencies, or provenance — based on what KG-Registry records. Reads the live hosted KG-Registry Parquet files over HTTP (no local clone needed), walks each product's original_source and secondary_source, resolves the referenced ids, and optionally recurses to build the full upstream tree.

Use when asked to curate the next N KG-Registry stub resources. Selects uncatalogued stub pages not already listed in reports/curation_problems.tsv and dispatches single-resource curation in parallel.

Use when updating the URL for a specific KG-Registry product, especially when the original URL is dead, redirected, versioned, or replaced by a newer canonical landing or download page.

Use when validating or quality-checking KG-Registry resource pages after curation, especially for schema compliance, provenance slot correctness, and stub-cleanup issues.