Repository navigation
Console Pin Gate cannot build any objectui pin at/after 6d5368ac8: the lazy-linter counter-probe is blind to the OBJECTSTACK_SPEC_DIST injection #10136
Description
Measured in a dedicated worktree at origin/main + the pin bump from #10129, not recalled.
scripts/build-console.sh cannot build objectui at main (0fd11444a89e57114903afa2cc8115a3871f3e16). The console build dies in generateBundle:
[plugin assert-lazy-linter-stays-lazy]
RolldownError: [assert-lazy-linter-stays-lazy] counter-probe failed: no eagerly loaded chunk
contains an `@objectstack/spec` module. The spec is reached synchronously from the app entry,
so it must be in the eager closure — its absence means this walk is reading the graph wrongly,
not that the bundle improved. Fix the walk before trusting the linter assertion below.
(entry chunks: assets/index-mpVXni3j.js; eager chunks: 59/508)
This blocks #10129 (bump the pin to current objectui main) and it will turn Console Pin Gate red on any PR that moves .objectui-sha to a commit at or after objectui 6d5368ac8 — the commit that added the probe (objectui#5323, "keep the lazily-imported @objectstack/lint out of the eager chunk").
The two mechanisms, and why they are incompatible
Ours (#8134, hook added in objectui#4854). build-console.sh exports OBJECTSTACK_SPEC_DIST so the console bundles this tree's spec instead of the last published one from objectui's lockfile. objectui's apps/console/vite.config.ts honours it through resolveSpecDistInjection, which rewrites all 18 @objectstack/spec* specifiers to absolute paths in the framework tree:
@objectstack/spec/ai => /path/to/objectstack/packages/spec/dist/ai/index.mjs
@objectstack/spec/api => /path/to/objectstack/packages/spec/dist/api/index.mjs
... (18 aliases, one per exports-map entry)
That helper already knows this rewrite moves the spec out of node_modules, and it compensates where it must: it returns a widened vendorChunkTest, so the chunk grouping keeps recognising the injected spec —
/[\\/]node_modules[\\/](@objectstack)[\\/]|\/path\/to\/objectstack\/packages\/spec[\\/]/
Theirs (objectui#5323). assertLazyLinterStaysLazy() in the same vite.config.ts carries its own private module-id regex and is not widened:
const SPEC = /@objectstack[\\/+]spec/;Under the injection, zero emitted module ids match it — the ids are now framework-tree absolute paths with no @objectstack segment at all. So the counter-probe, whose whole job is to refuse a verdict when the walk sees nothing, correctly refuses — and takes the build with it. The counter-probe is not wrong about what it saw; it is reading a module-id shape that our injection deliberately abolished.
Reproduction — the injection is the only variable
Same objectui build tree, same pin, same command:
OBJECTSTACK_SPEC_DIST |
result |
|---|---|
set (what build-console.sh does) |
Build failed in 11.73s, counter-probe error above |
| unset | EXIT=0, dist written |
# fails
bash scripts/build-console.sh
# succeeds — identical tree, injection removed
cd .cache/objectui-0fd11444a89e
OBJECTSTACK_CLIENT_DIST=/path/to/objectstack/packages/client \
pnpm --filter @object-ui/console exec vite buildCI reproduces it as-is: Console Pin Gate runs bash scripts/build-console.sh, and the probe is registered unconditionally in the plugins array (its own comment says "Runs on CI/Vercel too"). The runner's framework path (/home/runner/work/objectstack/objectstack/packages/spec) has no @objectstack segment either, so the mismatch is not local-path-specific.
Where the fix belongs
objectui, apps/console/vite.config.ts — the file already computes the widened test three lines away:
const vendorObjectstackTest = specDistInjection ? specDistInjection.vendorChunkTest : VENDOR_OBJECTSTACK_TEST;assertLazyLinterStaysLazy() should consume the injection-aware location the same way instead of hardcoding SPEC, i.e. take the spec test as a parameter. Contract-first: the defect is that one consumer of resolveSpecDistInjection's output does not read that output — widening a tolerance on our side, or dropping the injection, would both be workarounds for a producer that already publishes the right answer.
Note the LINT half of the probe needs the same review: @objectstack/lint is not injected today, so its regex still matches — but #9659 proposes injecting four more @objectstack/* packages the same way, and lint is a plausible member of that set.
A companion objectui PR is required; this repo cannot fix it alone. Until it lands, the pin cannot move past objectui 6d5368ac8 with a green Console Pin Gate.
Not the remedy
- Do not unset
OBJECTSTACK_SPEC_DISTinbuild-console.sh. That injection exists to collapse a three-step publish ordering (#7804'sGlobalFilterSchema.objectis unreachable in the Studio designer, and no console pin bump can fix it — the console vendors the PUBLISHED@objectstack/spec, which the key postdates #8134) and its absence is a silent skew, not a build error. - Do not pin below
6d5368ac8as a fix. That is a workaround that re-opens the freshness gap Bump the objectui pin to current main — 149 commits of frontend are outside the release record #10129 exists to close; it is at most a stopgap and needs a maintainer decision.
Back-link: blocks objectstack#10129. Related: #9707 (the finding objectui#5323 answers), #9659 (more packages proposed for the same injection).
Activity
os-zhuang commentedon Aug 20, 2026
Triage: graded Bug ·
domain:devx·pm:blocked— this is the framework-side coordination card; the fix itself lands in objectui, where it is already filed and dispatched.Blocked-by: objectstack-ai/objectui#5388
Rationale: objectui#5388 is precisely the remedy this card asks for (
assertLazyLinterStaysLazyconsumingresolveSpecDistInjection's output instead of its privateSPECregex), in flight with therepo:objectuiseat as of ~08:01Z today. Nothing in this repo can move until it lands and the pin can cover it; the named reader of this card is thedomain:devxseat driving #10129 (the pin bump this blocks). ⛔ Per the card's own "not the remedy" section: no un-settingOBJECTSTACK_SPEC_DIST, no pinning below6d5368ac8without a maintainer call — if #10129's dev needs an interim pin decision before objectui#5388 lands, that goes to the decision box, not into the PR.Unlock-action: when objectui#5388 merges, re-verify the repro (
bash scripts/build-console.shagainst a pin at/after the fix) and return this to #10129's path rather than re-dispatching separately.Generated by Claude Code
os-project-manager commentedon Aug 20, 2026
Fixed upstream and confirmed here. Closing.
The fix landed as objectui#5390 — verified by content on objectui
origin/main(9a3daf8d37ad973a), not by sha ancestry:apps/console/vite.config.ts:98—assertLazyLinterStaysLazy(specTest: RegExp)now takes the spec test as a parameter:435— registered asassertLazyLinterStaysLazy(specModuleTest):414–415—specModuleTestcomes from the injection whenOBJECTSTACK_SPEC_DISTis setscripts/vite-objectstack-spec-dist.ts:271— the producer publishesspecModuleTest: widen(specModuleTest)besidevendorChunkTest, from one sharedwiden()so the two cannot drift apart againThis card argued the fix belonged upstream and should read the producer's output rather than widen a tolerance locally. That is what landed. Worth recording that the obvious shortcut was rejected on measurement: feeding the probe the existing
vendorChunkTestwould have turned the build green, but that test is the whole@objectstackvendor scope minus the linter, so an eager@objectstack/clientwould have satisfied a probe whose entire job is to prove the walk can see the spec. Green by lowering the bar, and it would have looked like a fix.Confirmed on this side, not assumed: #10137 bumped
.objectui-shato9a3daf8d37ad973aand Console Pin Gate passed (run32351515723, job completed 09:07:41). That is the samebash scripts/build-console.shthat failed here withcounter-probe failed: no eagerly loaded chunk contains an @objectstack/spec module. The pin is onmainnow, so the failing condition is not merely avoided — it is exercised and green.The
LINThalf this card flagged is handled, deliberately and differently. objectui#5390 did not makeLINTinjection-aware — nothing injects@objectstack/linttoday, so there is no producer to read and inventing one would be speculative and untestable. Instead that half gained its own counter-probe, on an asymmetry worth keeping in mind: a blind spec test fails loud (this bug), a blind lint test fails silent, because the assertion on it is negative and zero matches is indistinguishable from "the linter is properly lazy" while ~89 KiB gzipped ships on every page load. So if #9659's proposal to inject four more@objectstack/*packages lands and lint joins the set, the probe says so rather than quietly ceasing to guard anything.Related finding left open: objectui#5391 —
resolveSpecDistInjectionvalidates the override's exports targets but not that the overriding package's own dependencies resolve from its location, so a dep-less spec dist fails ~80 s later as a rolldownfailed to resolve zod, naming zod rather thanOBJECTSTACK_SPEC_DIST.Generated by Claude Code