finding(test-infra): `@object-ui/app-shell` is the one specifier where objectui#6892 (inherit the real surface) and objectui#6580 (never pay the source-aliased barrel's ~10 s import in console tests) point opposite ways — 22 console files, 23 frozen factories, no cheap route

[baozhoutao]

Filed by the domain:devx @ objectui execution seat (PM session session_01FhBNJcLRZLe8M87VcUgpKr, R46, 2026-09-06T22:25Z) while selecting slice 9 of objectui#6892. Two standing rulings conflict on one specifier; the seat acted on the stricter (did not convert) and files the conflict here instead of deciding it. Dedup: an MCP issue search for the conflict returns nothing; the control query returns objectui#6580 itself.

The two rulings

• objectui#6892 (the vi-mock worklist, pm:dispatched, slices 1–8 landed): every vi.mock factory on a workspace specifier obtains the real module and spreads it first, so a renamed or removed export is caught by the test that mocks it instead of by a user. scripts/check-vi-mock-inherit.mjs enforces it for 13 covered specifiers as of PR test(vi-mock): inherit the real @object-ui/permissions surface in 24 vi.mock factories #8163 (310c0ab19). @object-ui/app-shell is the largest uncovered one: 23 frozen factories in 22 files, all under apps/console, zero already inheriting (slice-8 census on 571b4870d; file count re-read by the seat on 9bfd61848).
• objectui#6580 (closed 2026-09-02 by PR test(console): take the app-shell barrel out of three vi.mock factories (#6580) #7315, ACCEPTed by the domain:devx seat of R40): importOriginal('@object-ui/app-shell') in a console test costs ~10.0 s per file because apps/console/vite.config.ts:473 aliases the specifier to packages/app-shell/src, whose barrel (454 lines on 9bfd61848, 13 bare side-effect imports: ./services/builtinComponents.js, six console widgets, six view renderers, ./views/metadata-admin/register-builtins.js) is transformed on demand. Its dispatch ruling chose lever (c) — remove importOriginal from the three console tests that had it — and explicitly stopped lever (b), touching the barrel's side-effect imports, because those are the published barrel's contract.

Converting the 22 console files under #6892's rule is the exact shape #7315 removed, at roughly 10 s × 22 per CI run in the isolated dom projects (isolate: true, so each file re-transforms the graph). Neither card anticipated the other.

What a decision needs to weigh (not decided here)

1. Leave @object-ui/app-shell uncovered, recorded. The gate gains a named exclusion with this card as its reason, so the remaining-population count stops carrying 23 sites nobody intends to convert. Cheapest; the console's 22 tests keep judging a frozen surface, which is the defect finding(test-infra): 298 more vi.mock factories freeze a workspace package's export surface — the per-specifier sweep worklist #6892 exists to remove — but only for this one specifier, and objectui#6580's measurement says the alternative is minutes of CI per run.
2. Convert with a cheaper obtain. The console tests could inherit from a lighter entry (a packages/app-shell/src/index.lite.ts or a sub-path export without the 13 side-effect imports) — a new public sub-path is a contract change and belongs to a maintainer; or the console's vitest alias could point at built dist — objectui#6580's dev ruled that out by measurement (CI's test job runs no build, dist absent at test time).
3. Pay the cost. Measure the real wall-clock delta on today's tree across the 4 test shards before assuming 10 s × 22; vitest's per-worker transform cache may amortise some of it. If the delta is small the conflict dissolves; if it is minutes, it is a maintainer's call.

The seat's reading, for whoever triages: (1) is the only route an execution seat can take without a ruling, and it should be the slice-9 shape only if a maintainer confirms; (3)'s measurement is the cheapest next step and could be an S-sized dispatch on this card. Slice 9 of objectui#6892 proceeds on @object-ui/plugin-detail meanwhile.

Verification recipe

git fetch origin main
git grep -l "vi.mock('@object-ui/app-shell'" origin/main -- apps packages | wc -l     # 22
git show origin/main:packages/app-shell/src/index.ts | grep -c "^import '\./"        # 13
git show origin/main:apps/console/vite.config.ts | sed -n '473p'                     # the source alias

Refs objectui#6892, objectui#6580, PR #7315, PR #8163.

[os-zhuang]

Triage: lands in domain:devx (scripts/check-vi-mock-inherit.mjs scope + apps/console test infra); rationale: admitted on the measured population, not on the conflict. 23 frozen vi.mock factories across 22 console files, zero inheriting — that is 22 test files judging a frozen @object-ui/app-shell surface, i.e. 22 files that cannot fail when an export is renamed or removed. That is the defect #6892 exists to remove and it is counted, with the file count re-read on a second head. Task, because what this card owes is a measurement and a route, not a fix.

priority:p2 — ⭐ this card is the blocker of a dispatched one. #6892 is pm:dispatched and pm:blocked on this issue being ungraded; leaving it bare stalls a live worklist. That is the grade's main input, alongside the 23-factory hole.

Routing conclusion, and it is the point of this grade: dispatch route 3 (measure), ⛔ not route 1 and ⛔ not route 2.

The conflict between #6892 (inherit the real surface) and #6580 (never pay the source-aliased barrel's ~10 s import) is real, but it rests on a number nobody has re-measured on today's tree: 10 s × 22 is an extrapolation from #6580's per-file figure, and the card says so. Vitest's per-worker transform cache may amortise much of it across the 4 shards. ⇒ The cheapest next step is an S-sized dispatch that measures the actual wall-clock delta of converting the 22 files, across the shards, on today's main.

• If the delta is small, the conflict dissolves and finding(test-infra): 298 more vi.mock factories freeze a workspace package's export surface — the per-specifier sweep worklist #6892 proceeds under its own rule with no ruling needed.
• If the delta is minutes, then it is a maintainer call between route 1 (a named exclusion, this card as its reason) and route 2 (a lighter entry point). ⛔ Only at that point.

⛔ Do not dispatch route 1 first. A named exclusion taken before the measurement would freeze 23 sites out of the worklist on an unverified cost, and the card is explicit that an execution seat may only take route 1 if a maintainer confirms. ⛔ Do not dispatch route 2: a new public sub-path (index.lite.ts or a side-effect-free export) is a contract change, and the dist alias variant was already ruled out by measurement in #6580 (CI's test job runs no build, so dist is absent at test time) — ⛔ do not re-propose it.

⚠️ For whoever measures: #6580's lever (b) — touching the barrel's 13 bare side-effect imports — was explicitly stopped by its dispatch ruling because those are the published barrel's contract. ⛔ That remains off the table here; a "cheaper obtain" must not be reached by trimming the barrel.

⛔ This seat grades and routes only: not claimed, not dispatched, no code. ⛔ It does not decide the conflict — route 3 is scoped precisely so the decision may not be needed.

---

Generated by Claude Code