oh-my-claudecode-research

Ingest a research paper into the project's two-folder reading library. Use when the user hands you a paper — a PDF path, a DOI, or a URL — and wants it read, summarized, and indexed. The skill writes a project-agnostic summary (with a cropped main figure) to the library's bibliographic-management/ folder, upserts a row in that folder's index.csv, judges whether the paper is relevant to the project's narrative (from CLAUDE.md

Write a structured Markdown record of the current session — the requests, the analyses run, the files touched, decisions made, and open next steps — to the project's session-logs dir (one dated file per invocation), then propagate any durable result/decision into the wiki knowledge vault. This is a working record for the user, not a scientific artifact and not a status/sync report. Reconstructs content faithfully from the actual conversation; never invents. Use at the end of (or partway through) a session, or when the user types /save-session-log.

Session-orientation workflow — read the project's Markdown corpus (CLAUDE.md, the outline, project MEMORY, the wiki landing page), build a genuine understanding of the study, then report a concise project summary and an honest current-status snapshot. Read-only with zero side effects — no exports, no memory edits, no report files, no analysis runs. Runs in light mode (cheap, core files only — the default for auto-trigger) or full mode (full corpus sweep). Use at the start of a session to reproduce the effect of "understand this project", or when the user types /session-start.

When a project artifact whose filename carries a version — the canonical outline or the figure deck — is bumped (e.g. v4→v5, ver7→ver8), propagate the new filename into every file that references the old one (CLAUDE.md, command/skill files, agent memories, manuscript anchors, export scripts) so no downstream tool points at a stale file, then offer to delete genuinely-obsolete archive files (never auto-delete). Reads the previously-current values from CLAUDE.md

Three-step pipeline from a Keynote (.key) or PowerPoint (.pptx) deck to manuscript + outline. Func 1 exports vector PDFs per slide; func 2 crops them (vector preserved) and renders an outline-grade PNG from the same cropped artifact; func 3 copies PDFs into the LaTeX manuscript figures dir and PNGs into the outline.md figures dir, then inserts/updates image links after each result heading.

Drive one figure from design idea to manuscript-ready vector PDF + outline-ready PNG. Loops `@figure-descriptor` → `@analysis-implementer` → `@reviewer` against a single fig-id in `figures.json`, with the `cropfig` skill auto-invoked at the end of each implement phase to keep manuscript + outline artifacts in lockstep. The third Phase 2 engine — a 3-agent loop, more complex than `/iterate-revision`'s 2-agent loop, and the only Phase 2 engine that executes real code per iteration. Safe to re-run; safe to resume after BLOCKED or HALT.

Revise one manuscript section against the reviewer team until DONE, BLOCKED, or HALT. Loops `@paper-writer` ↔ `@reviewer` with a venue-specific reviewer brief, recording every iteration's issues + verdict to `reviews.json` and updating `paper.json.sections[name]` status / iter. The first OMCR engine — a worked example of how to compose `skills/orchestrate/phases/*` primitives into a domain-specific loop. Safe to re-run; safe to resume after BLOCKED or HALT.

Find, summarize, and verify N papers on a topic, then drop verified entries into `references.bib` and the literature summary CSV. Dispatches `@literature-curator` (one or many instances) over a CrossRef/OpenAlex candidate list, runs every survivor through the `verify-citation` skill, and writes the run summary to `citations.json.last_sweep`. Sequential by default; opt-in `--parallel N` (1 ≤ N ≤ 4) fan-out for speed. Safe to re-run; idempotent against duplicate DOIs (existing BibTeX entries are skipped, not double-added).

Scaffold a LaTeX manuscript directory for a research project — copy the bundled skeleton (main.tex + sections/* + figures/ + references.bib + .gitignore + README), optionally apply a journal-specific documentclass from templates/journal-registry.json, optionally clone an Overleaf project and cache the Git credential helper (token never persisted to tracked files), commit on the default branch, and ask before pushing. Invoked by /start-research phase 6, but also standalone-callable when adding a manuscript dir to an existing project later.

Install OMCR infrastructure in a project — scaffold empty CLAUDE.md marker blocks (`## Project context` / `## Research stack` / `## Language preference`), create `.claude/agent-memory/` for the 6 core agents with canonical-template MEMORY.md, initialize empty `references.bib` + `references.csv`, and install a curated permission allowlist in `.claude/settings.json`. Does NOT ask about the user's research project (no working title / hypothesis / venue / preset interview). For that, run `/start-research` after this completes. Safe to re-run — never overwrites existing user content.

Internal primitive skill — engine skills compose these four phases. Not invoked directly by the user. Provides state-read, dispatch, evaluate, and loop primitives over `.claude/omcr-state/` so that engines (`/iterate-revision`, `/literature-sweep`, `/respond-reviewer`, `/figure-bake`, `/outline-expand`, `/supervisor-drive`) can be expressed as a small composition declaration instead of bespoke code.

Map-reduce engine — given an outline file and a `paper.json`, draft N sections in parallel by dispatching `@paper-writer` once per section in a single Agent-tool batch, then assemble the prose into each section's `paper.json.sections[name].path`. Passes a shared `nomenclature.md` payload to every dispatch (Phase 2 decision §4) so parallel writers share terminology. After merge, emits a non-blocking `terminology-drift.md` lint artifact listing terms that disagree across sections. First drafts only — does **not** call `/iterate-revision`. Safe to re-run; safe to scope with `--sections`.

Read a reviewer letter, classify each comment by type, dispatch per-comment responses to the right specialist agent (`@paper-writer`, `@analysis-implementer`, `@literature-curator`), and assemble a complete rebuttal letter. Structural comments are surfaced to user attention rather than auto-dispatched. Output defaults to LaTeX; accepts markdown or LaTeX input (auto-detected by extension). Safe to re-run; safe to resume after BLOCKED. The Phase 2 worked example of the classify-and-dispatch orchestration shape.

Interview-driven first-research-project initialization — asks about working title, field, hypothesis, target venue, datasets, narrative spine; fills the CLAUDE.md placeholders that /omcr-setup scaffolded; applies an optional domain preset to agent memory (when the existing MEMORY.md is still canonical-template); seeds @reviewer's persistent memory with target-venue aims/scope/editorial-priorities (registry-first, WebFetch fallback); scaffolds the LaTeX manuscript directory via the manuscript-scaffold skill. Requires /omcr-setup to have run first — will offer to run it if not. Safe to re-run; never overwrites filled-in answers, modified MEMORY.md, or existing manuscript content.

Autonomous OMCR orchestrator. Surveys state across the 5 OMCR state files (paper, reviews, citations, figures, rebuttals), picks the highest-priority engine from a hardcoded bottleneck-ranker, runs it through one of the 5 OMCR engines, then re-evaluates state from scratch and loops. Three modes — interactive (default), auto, plan-only. Six safety gates (HypothesisChange, NewCitation, NewExperiment, StructuralRewrite, BudgetExceeded, CriticalIssue) fire even in --auto. Single-target only currently. Halt-on-exception, no retry. Resume after halt is strict — requires explicit --resume <run-id> or --fresh.

Reconcile current state (captured figure deck) with final goal (outline document), refresh agent memories with factual drifts, and produce a dated status snapshot. Reads `## Research stack` config from the user's CLAUDE.md (Deck file / Outline file / Figure count / Report language / Sync report dir). Status-only output — never produces a TODO (that is the `todofig` skill's job). Called by the `/sync` slash command but also standalone-invocable.

Compare a captured-figure deck against an outline document and produce a prioritized TODO of gaps (P0/P1/P2). Reads `## Research stack` config from the user's CLAUDE.md (Deck file / Outline file / Figure count / Result pattern / Report language / Report output dir). Accepts an optional figure identifier in `$ARGUMENTS` (e.g. "Fig4") to restrict the analysis to a single figure. Called by the `/todofig` slash command but also standalone-invocable.

Verify that academic citations exist and match the metadata claimed for them, and optionally write verification results into a project summary table. Looks up DOIs against CrossRef, fetches canonical authors/title/year/journal and the abstract via OpenAlex, and reports any mismatch against a BibTeX entry or a free-form DOI. Use to gate every citation `literature-curator` adds, to audit an existing BibTeX file for fabricated or wrong entries, or to retrieve an abstract so the calling agent can judge whether the paper actually supports a given claim.