| name | dag-deepen |
| tier | core |
| runtime | on-demand |
| trigger | mention |
| description | Architecture-deepening scan as a DAG: deterministic git-hotspot discovery picks the highest-friction modules, one agent leaf per hotspot concurrently hunts shallow modules (deletion test, deep-interface sketches), a synthesis leaf dedups cross-hotspot repetition and ranks by leverage, HTML report of candidates. Trigger: improve architecture / 找重构机会 / deepen modules / architecture report / dag-deepen. Skip: what to delete (/dag-slim) / diff review (/dag-review) / design vocabulary only (/codebase-design). |
| metadata | {"source":"oh-my-dag","version":"1.0.0","methodology":"deterministic hotspot discovery + per-hotspot agent fanout (prebuilt plan, zero conductor LLM) + cross-hotspot synthesis + leverage ranking"} |
| disable-model-invocation | true |
/dag-deepen — architecture-deepening candidate scan
Engine-shaped port of Matt Pocock's improve-codebase-architecture: instead of one agent
wandering the repo, a deterministic hotspot pass (git touch frequency over recent
commits, zero LLM) picks the modules where friction actually lives, then a fleet of
agent leaves scans them concurrently — each hunting shallow modules with the deletion
test — and a synthesis leaf merges cross-hotspot duplication (the same logic hand-rolled
in N places = one shared-util candidate, the global win a single scan can't see) and
ranks by leverage. Output: an HTML report of before/after interface sketches.
Usage
bun run scripts/dag-deepen.ts [<scope-path>] \
[--commits N]
[--hotspots K]
[--model M]
[--out path]
Name a direction to skip repo-wide ranking: bun run scripts/dag-deepen.ts "src/harness/plan".
Report lands in $TMPDIR/omd-deepen-<ts>.html (Tailwind + Mermaid cards, one per candidate:
files / friction / deletion-test verdict / before-after / strength).
Discipline
- Hotspot-first is YAGNI for architecture: only recently-touched modules get scanned —
friction you aren't paying is friction you don't refactor. Widen with
--commits, don't
tour the whole repo.
- Candidates are inputs to
/grill, not conclusions: every run ends by telling you to
grill the chosen candidate before touching code. dag-deepen never edits files, never
opens a PR.
- Vocabulary lives in
/codebase-design (shallow module / deletion test / deep
interface / leverage / locality): the scan prompts inject that single source — read it
there, don't re-derive it here.
