ULW-Research — Maximum-Saturation Research Orchestrator

What it does

ulw-research is a saturation research protocol, not a quick lookup:

1. Proposes the deliverable format first — PDF + DOCX by default, plus domain-appropriate alternatives (slides, standalone HTML, Markdown) and a template proposal (section skeleton, citation style, length, language). It then stops and waits for your answer.
2. Decomposes the question (Phase 0) — at least three orthogonal axes, plus a timestamped journal directory at .omo/ulw-research/<timestamp>/.
3. Launches the whole first wave at once (Phase 1) — explore workers for the codebase, librarian workers for web/docs/GitHub, browsing workers for WAF-blocked pages, repo-dive workers for SHA-pinned source reading. With team mode available, it fills the maximum roster, one member per axis.
4. Recursive EXPAND loop (Phase 2) — every worker must end with LEAD: markers; the orchestrator immediately spawns expansions until leads run dry (or 3 barren waves / a 5-wave ceiling), with bounded "excursions" that must be explicitly closed.
5. Empirical verification (Phase 3) — contested, undocumented, or performance-shaped claims are proven or refuted by running code.
6. Cited synthesis behind a visual-QA gate (Phases 4–5) — every claim carries a source or verification artifact, and rendered pages with charts/Mermaid must pass QA before delivery.

A distinctive feature is the orchestrator-owned evidence ledgers (claim-graph.md, observation-manifest.md, intent-diff.md, verification-economics.md) that make the path from observation → claim → verdict auditable.

Who it's for

• Anyone doing research where a wrong claim is expensive: legal, medical, financial, procurement, or public-facing work.
• Engineers who need an exhaustive map of a feature across a large codebase with file:line evidence.
• Power users already running parallel subagents (OMO/Codex multi_agent, OpenCode team_mode).
• Not for quick one-off questions — the run is deliberately token- and time-expensive, and the skill refuses to self-activate for ordinary Q&A or debugging.

Examples

• "/ulw-research map our entire auth flow and audit it against the OAuth spec" → codebase explore workers plus spec-reading librarians, then a cited PDF+DOCX audit report.
• "ulw-research: deep comparison of 5 vector databases on performance, licensing, and operating cost" → one axis per product, cross-checked against official docs, benchmarks, and GitHub issues, with contested numbers reproduced by an actual benchmark script.
• "ultradebate this architecture decision record" → at least two red-team members attack every contested claim, then a fresh refinement team sharpens the synthesis before the document is written.

Install the skill I found on Claude Skill Mart.
Copy the packages/shared-skills/skills/ulw-research folder from the GitHub repo code-yeongyu/oh-my-openagent into my ~/.claude/skills/code-yeongyu-ulw-research/.
When it's done, tell me in one line what this skill can do.