itchyshin

Run Rose's narrow pre-publish consistency gate for gllvmTMB README, vignettes, pkgdown navigation, NEWS, roxygen, or Rd changes. Cross-checks claims against the validation-debt register, the README stable-core feature matrix, the AGENTS.md scope-boundary rule, and (per Phase 0C closeout 2026-05-16) the export ↔ pkgdown reference parity, the removed-article cross-reference cascade, the Preview-banner citation discipline, the REWRITE-PREP contract, and the ROADMAP slice "Done when" deliverable rule.

Audit a completed gllvmTMB task or phase before closing it. Checks implementation, equations, examples, tests, docs, pkgdown, roadmap, NEWS, known limitations, stale wording, after-task reporting, AGENTS.md Rule

Triage gllvmTMB pkgdown articles into tiers (1 = public worked example, 2 = technical reference, 3 = retire) and judge each against the 10 reader-first style rules. Use when adding, removing, or rewriting any vignette under `vignettes/` or `vignettes/articles/`.

Review and improve gllvmTMB prose in README files, vignettes, pkgdown articles, after-task reports, release notes, design docs, and manuscript-style text for clarity, concrete claims, stable terminology, citations, and reader fit.

Hard review gate for scientific R/ggplot figures before merge or release. Use to judge whether screenshots, generated plots, vignettes, or plot helper outputs are immediately publication-ready; fails default-looking ggplot output, illegible labels, misleading intervals, cramped facets, poor scales, and weak visual hierarchy.

R/ggplot2 implementation specialist for publication-quality plot helpers in R packages. Use when editing or creating plot_* functions, package themes, color scales, confidence interval geoms, facets, labels, vdiffr visual tests, export helpers, or when replacing poor default ggplot output with manuscript-ready graphics.

R package engineer for plot helper APIs, roxygen documentation, tests, vignettes, CRAN-safe examples, namespace imports, and package integration. Use when creating or refactoring R package plotting functions, especially ggplot helpers that need publication-quality defaults and reliable behavior.

Scientific illustrator and statistical graphics editor for R package ggplot figures. Use when figures look like poor default ggplot output, need publication quality, need critique/redesign, or must visualize model results, correlations, predictions, intervals, or parameter surfaces for interpretation.

Close a hsquared or HSquared.jl task with repo-visible evidence. Use after meaningful docs, code, tests, CI, coordination, GitHub issue, or public-readiness work.

Review the future bridge from hsquared R calls to HSquared.jl. Use for marshalling, JuliaCall or other bridge choices, result-shape parity, engine controls, and R placeholder/error wording.

Review the HSquared.jl engine contract for input payloads, result shape, storage policy, diagnostics, controls, and numerical assumptions. Use when changing docs or code at the R-Julia boundary or Julia engine API.

Review hsquared formula, covariance, and relationship grammar. Use when adding or changing R syntax, Julia formula equivalents, unsupported-syntax errors, examples, docs, or the R-Julia model specification.

Rehydrate the hsquared or HSquared.jl project before substantial work. Use when resuming a thread, starting a new slice, recovering after interruption, checking coordination state, or preparing edits that depend on repo, CI, roadmap, or twin-lane status.

Dispatch hsquared work across coordinator, R, and Julia lanes. Use before starting a substantial slice, spawning agents, steering the Julia twin thread, touching shared docs, or deciding which named review lenses are active.

Review pedigree validation, ID recoding, sorting, founder handling, unknown parents, and sparse Ainv construction for hsquared and HSquared.jl. Use before claiming pedigree or animal-model support.

Review hsquared README, design docs, vignettes, issue text, and release prose for reader-first, concrete, honest wording. Use when public or long-form prose changes.

Audit the freqTLS-versus-bayesTLS benchmark for a fair configuration, sound cache provenance, and the correct R-SHRIMP data rebuild.

Audit and improve freqTLS figures, galleries, pkgdown articles, and ggplot recipes, enforcing the Confidence-Eye uncertainty contract, before Florence, Rose, Pat, Fisher, and Grace call a figure done.

Review freqTLS before a public, CRAN, GitHub, or internal release.

Add simulation-based parameter-recovery tests for freqTLS models.

Audit a completed freqTLS task or phase before closing it, checking implementation, equations, examples, tests, docs, pkgdown, roadmap, NEWS, the capability matrix, known limitations, stale wording, and after-task reporting.

Review freqTLS profile-likelihood confidence intervals for correctness, equivariance, chi-square calibration, and honest open/boundary/multimodal handling.

Review and improve freqTLS prose in README files, vignettes, pkgdown articles, after-task reports, release notes, design docs, and manuscript-style text for clarity, concrete claims, stable terminology, citations, and reader fit.

Design a compact parameter-recovery simulation plan for the freqTLS 4PL thermal-load-sensitivity model.

Review and improve drmTMB prose in README files, vignettes, pkgdown articles, after-task reports, release notes, design docs, and manuscript-style text for clarity, concrete claims, stable terminology, citations, and reader fit.

Audit a completed drmTMB task or phase before closing it, checking implementation, equations, examples, tests, docs, pkgdown, roadmap, NEWS, known limitations, stale wording, and after-task reporting.

Audit and improve drmTMB figures, figure galleries, pkgdown articles, simulation reports, and ggplot recipes when plots look poor, inconsistent, misleading, too sparse, missing raw or replicate data, or need Florence, Rose, Pat, Fisher, and Grace visual QA before being called done.

Audit a completed task or phase before closing it, checking implementation, equations or algorithms, examples, tests, docs, roadmap, release notes, stale wording, and after-task reporting.

Review statistical, machine-learning, simulation, or data-transformation implementations before merging.

Review an R package before a public, CRAN, GitHub, or internal release.

Add a new drmTMB distribution family with likelihood, simulation, tests, and documentation.

Add simulation-based parameter-recovery tests for drmTMB models.

Add a new non-Gaussian response family to GLLVM.jl with Laplace-approximated likelihood, link functions, Julia multiple-dispatch wiring, ADEMP recovery test, and documentation. Use when extending Phase 3 (non-Gaussian families) beyond the Gaussian-only v0.1.0 pilot.

Add ADEMP-style simulation-based parameter-recovery tests for GLLVM.jl models using Test.jl and StableRNGs.

Mandatory closure routine for every completed GLLVM.jl task or phase. Writes a compact report to docs/dev-log/after-task/YYYY-MM-DD-<task>.md covering goal, files changed, tests added, benchmark numbers, R-parity verdict (vs gllvmTMB), JET/Allocs/Aqua verdicts, Rose audit verdict, remaining risks, and the next command. Enforces the drmTMB-grade Definition of Done the maintainer asked for: implementation + tests + docs + examples + check-log + after-task audit + Rose audit verdict.

Audit and improve GLLVM.jl figures, Documenter.jl pages, simulation reports, and CairoMakie recipes when plots look poor, inconsistent, misleading, too sparse, missing raw or replicate data, or need Florence, Rose, Pat, Fisher, and Grace visual QA before being called done.

Review Julia GLLVM likelihoods and parameterizations before merging.

Use when working on Julia package hygiene — Project.toml, exports, Aqua.jl, JET.jl, Allocs.jl, version bumps, Manifest stability.

Review and improve GLLVM.jl prose in README files, tutorials under docs/src/, Documenter pages, after-task reports, release notes, design docs, and manuscript-style text for clarity, concrete claims, stable terminology, citations, and reader fit.

Use when porting/translating any gllvmTMB R-side API to GLLVM.jl, building parity tests under GLLVM_PARITY_TESTS, or auditing API mirror coverage.

Keep drmSEM CRAN-clean — R CMD check 0 errors/0 warnings, examples that need drmTMB wrapped in \dontrun, Remotes for the GitHub engine, no compiled code in drmSEM itself, DESCRIPTION hygiene, ASCII, reasonable file sizes, reproducible RNG. Use before release or when touching DESCRIPTION/examples/CI.

Reader-facing docs for drmSEM — README, the vignette, and pkgdown reference sync. Pair equations + R syntax + interpretation, show component-labelled output and the distribution-mediated decomposition, and tell users what to try when a path/family is unsupported. Use when writing user-facing prose.

THE core skill. Enforces component-labelled paths — never describe or compute a non-mu path as a mean effect. Use for ANY prose, docs, message, or code that states what a drmSEM path MEANS.

How drmSEM talks to drmTMB — ONLY through R/extractors.R. The drmTMB return-shape facts, the dpar:term coefficient convention, and the rule to confirm parameterizations against a live fit before trusting samplers.

Build basis sets and run any-component d-separation / Fisher's C correctly — adjacency = direct parent on ANY component, causal ordering, conditioning on parents, LRT df bookkeeping, honest framing as drmSEM's definition.

R package craft for drmSEM — roxygen2 on every export, S3 conventions, the document/test/check dev loop, NAMESPACE sync, cli messaging, tests shipped with implementation, small focused commits.

The simulation-based effect engine — do()-style topological propagation, mean vs distribution mediation, direct/indirect/total decomposition, Monte-Carlo coefficient uncertainty. Use whenever computing or describing an effect estimand. Coefficient-product mediation is banned.

Testing discipline for drmSEM — pure-logic kernels run without drmTMB, integration tests gated by skip_if_not_installed, every effect/estimand/d-sep rule needs a known-DGP recovery test, evidence tiers recorded in the validation ledger. Use when adding tests or any new estimand/rule.

Audit a completed task or phase before closing it, checking implementation, equations or algorithms, examples, tests, docs, roadmap, release notes, stale wording, and after-task reporting.

Review and improve project prose in README files, vignettes, pkgdown or Quarto pages, after-task reports, release notes, design docs, and manuscript-style text.

Design compact simulation, replay, or calibration tests for statistical models, machine-learning workflows, agent-based models, and stochastic data pipelines.