ANcpLua / repo fleet / spawn deck · 2026-05-15

Fan out four scanners over 349 repos

One pass over gh repo list ANcpLua --limit 1000 classified everything into 4 buckets (45 delete / 59 review / 64 archive / 181 keep). Reviewing 200+ candidates manually is the failure mode this whole skill exists to avoid. Spawn four specialized teammates instead — each takes one detection axis, returns evidence-backed verdicts, the lead aggregates into a triage HTML.

github_dark palette pattern: 21-agent-spawn-deck runs in parallel no destructive ops — verdicts only

01Teammate prompts — copy one per terminal tab

02 · Synthesis step (after all four return)

Each teammate writes a JSON verdict file under ~/repo-audit/scans/<teammate>.json with shape {repo, verdict, confidence, evidence[]}. When the four files exist, the lead re-runs the classifier with the new evidence and writes the final triage as a fresh repo-fleet-cleanup-triage.html (pattern 18-editor-triage-board) into the user's project area. The drag-drop in that triage overrides any teammate's verdict — they're advisory.

Why fan-out instead of one big audit? Each scan axis (empty / fork-status / stub-name / activity) has a different evidence model. One agent doing all four ends up paging in 4× the context per repo and confusing axes. Four specialists, each with a tight scope, finish in ~1/4 the wall-clock with sharper verdicts. Pattern lifted from html-effectiveness's spawn-deck style (originally the shadow-ai-inventory by another agent).

What this deck cannot do for you