The idea: for the same ~3 h of UI work, the toolkit arm made 1,507 calls to the MCP arm's 515. The toolkit arm's result was closer to the wireframes. The extra calls came from two things:
- 7 drafting agents each ran check → exec → mx check → fidelity on every script, then the lead re-applied it.
- Round 1 introduced four regressions (N1 shell detached, N2
.d-none precedence, N3 32 px autoFit columns, N4 focus ring).
Proposals:
- UI agents apply directly, with one mx check per batch.
- Cross-cutting fixes first, done once as SCSS.
- Re-review only the pages that changed.
- Add N1–N4 as a pre-apply lint in
check-page-shell.sh / learned-css-that-never-applied.md.
Field evidence: contrib/inbox/2026-09-29-ui-loop-efficiency.md (#165).
Done = the next UI round needs about half the calls, with no fidelity regression.
The idea: for the same ~3 h of UI work, the toolkit arm made 1,507 calls to the MCP arm's 515. The toolkit arm's result was closer to the wireframes. The extra calls came from two things:
.d-noneprecedence, N3 32 px autoFit columns, N4 focus ring).Proposals:
check-page-shell.sh/learned-css-that-never-applied.md.Field evidence:
contrib/inbox/2026-09-29-ui-loop-efficiency.md(#165).Done = the next UI round needs about half the calls, with no fidelity regression.