| name | lumin-repo-lens-lab-codex |
| description | Codex-native lumin-repo-lens-lab for TS/JS repo structural review: find dead code, cycles, duplicate helpers/types, oversized modules, barrel fan-out, plan refactors, or answer 'does X exist anywhere?' with machine evidence. |
Lumin Repo Lens Codex
Purpose
Use this Codex-native wrapper for TS/JS repository structure review,
audit artifact interpretation, and refactor planning. It is a thin
front door over the shared lumin-repo-lens-lab engine; it does not duplicate
runtime code.
Prefer this wrapper in Codex when the user asks things like:
- "audit this repo"
- "what should I clean up next?"
- "is this helper/type/file already present?"
- "run lumin-repo-lens-lab full"
- "do a structural review"
Shared Engine
The shared engine must be installed as the sibling skill
lumin-repo-lens-lab. Set <audit-repo> to this path resolved relative to this
SKILL.md:
../lumin-repo-lens-lab/scripts/audit-repo.mjs
If that sibling path is missing, tell the user to link the generated
skills/lumin-repo-lens-lab directory into Codex as well. Do not copy the
engine into this wrapper.
Core Contract
NO STRUCTURAL CLAIM WITHOUT MACHINE EVIDENCE
NO ABSENCE CLAIM WITHOUT STATED SCAN RANGE
NO STRUCTURAL REVIEW WITHOUT A CHECKLIST GATE
Run the engine, read the artifacts, then make the claim. The short chat
answer is only output density; it never permits skipping checklist
triage.
Codex Workflow
For the first pass on a repo, stale artifacts, explicit audit/review,
large refactor planning, due diligence, or post-refactor review, run:
node <audit-repo> --root . --output .audit --profile full
For a small follow-up with a fresh full baseline, run:
node <audit-repo> --root . --output .audit --profile quick
Then read manifest.json, checklist-facts.json, fix-plan.json, and
the relevant raw artifacts under .audit/. Treat
audit-summary.latest.md as an artifact map, not the final answer.
Before any structural review answer, pass the checklist gate: triage
C/D/E/A/B/F using checklist-facts.json plus relevant raw artifacts.
For full, deep, exhaustive, due-diligence, CI, or formal review, also
open the sibling engine template at:
../lumin-repo-lens-lab/templates/REVIEW_CHECKLIST.md
For normal chat, keep the answer short and humane: what is stable, at
most three things worth smoothing next, what to keep as-is, and the scan
range/confidence. If a full profile ran but the user did not ask for a
full report, end with one required feature-discovery tail. This is not
decoration: it tells the user that the same evidence can be expanded
into a full checklist walk, formal report, or due-diligence handoff.
Give the user copyable phrases, for example: "full checklist로
펼쳐줘", "formal report로 써줘", or "due-diligence handoff로 정리해줘".
Do not omit it after full-profile short answers.
For a saved formal report, the final author re-reads the report before
final answer or handoff and manually checks headline counts, same-site
classifications, broad conclusions, and chat-persona leakage against the
cited artifacts or source. Do not replace this with a string heuristic.
Code-Change Requests
For add/edit/move/rename work, use the sibling lumin-repo-lens-lab-write-gate
skill when available. It owns pre-write reuse screening and post-write
delta checks. If the sibling skill is not installed, run the same engine
with --pre-write or --post-write and explain the missing sibling
surface.
Canon Work
For canonical draft or drift checks, use the sibling lumin-repo-lens-lab-canon
skill when available. This wrapper stays focused on Codex audit and
review flow.