fix(components): flex declares the containment it renders - #6762

Fixes #6740

flex has always rendered schema.children, but its registration omitted isContainer while grid, card, container and stack — same directory, same ui namespace — all declare it. The render path never reads the flag, so nothing was broken at runtime; the flag's consumers are elsewhere, and the gap made them contradict the renderer.

Reproduced first, with the card's own control

Base b76ca6764, the card's re-check verbatim:

git grep -c isContainer -- packages/components/src/renderers/layout/flex.tsx   ->  no output, exit 1  (zero matches)
git grep -c isContainer -- packages/components/src/renderers/layout/grid.tsx   ->  grid.tsx:1, exit 0

That zero is a reading only because the positive control returns non-zero in the same corpus and under the same command shape.

The defect measured through the mechanism, inside this repo

@object-ui/sdui-parser is already a dependency of @object-ui/components, so the downstream path is constructible here — no NOT MEASURED fallback was needed. The manifest was built exactly the way the app builds it (getKnownTypes() + getMeta() fed to manifestFromConfigs, mirroring renderers/layout/page.tsx:462) and a node carrying children put through validateTree.

On base b76ca6764:

With this change, flex returns [] and the three controls are unchanged. This is the same warning objectstack measured downstream: its three shipped examples/app-showcase html pages drew 232 diagnostics, and every one of the 32 warnings was not-a-container on flex, with flex their only source. The 200 errors are a separate matter (objectstack#12924 owns it) and are untouched here.

The form is copied from the siblings, not invented

grid, card, container and stack all place isContainer: true in the registration's third argument, after defaultProps. This change matches that placement and value.

Where they disagree, and the call made. All four ui-namespace siblings pair isContainer with resizable: true and a resizeConstraints block. The page:* containers in containers.tsx declare isContainer: true on its own, with neither. So the flag is independent of designer resize affordances, and only isContainer is added here — minting a resize affordance for flex would be a behaviour change this card did not measure.

Reviewer attention: a second consumer, beyond the diagnostic

The card asked that anything altering behaviour past not-a-container be surfaced rather than absorbed. One thing does, and it is measured rather than reasoned about.

renderers/layout/react-page.tsx:77 builds the JSX scope of every kind:'react' page with if (!tag || cfg.isContainer) continue;, reading the predicate off getPublicConfigs(). While flex omitted the flag it was the one layout primitive of the five still injected there.

Measured by rendering real kind:'react' page source through SchemaRenderer:

The leaf-block row is the direction control: without it, "containers are dropped" would be indistinguishable from "everything is dropped".

Two things make this the correcting direction rather than a regression, but a reviewer should still rule on it:

1. content/docs/guide/react-pages.md has documented this behaviour all along and names this component first — "Layout containers are deliberately not injected. The scope builder skips every container ... so flex, grid, card and friends have no injected wrapper." The prose predates the 2026-08-21 rewrite that added the code citation; only flex did not obey it.
2. The behaviour being removed is itself silently broken. The injected wrapper is ({ children: _children, ...props }) — it discards children. Measured on base: a react page writing this component around a child rendered div data-obj-type="flex" class="flex flex-row ..." with the child gone. After the change the author gets the page-level error panel naming the identifier, and that doc page's stated remedy is real HTML with an inline style object.

So the change moves a react page from silently swallowing content to failing loudly with a documented fix. It is still a user-visible change on a published package, which is why it is called out here rather than left in the diff.

The pin covers the family, not just the one that was missing

The defect's shape is "three declare it and one does not". A pin covering only flex would let the next registration rot the same way and stay green while it did, so packages/components/src/__tests__/layout-containers-declare-containment.test.tsx asserts over all four, through validateTree rather than against the literal property — the property is only interesting because a mechanism reads it.

It carries three controls, each closing a way the pin could pass while measuring nothing:

• reachability — no unknown-component in the result, and the probe node really has children (the containment branch is guarded by node.children?.length);
• liveness — badge, a genuine leaf, must still draw not-a-container, so deleting the check or defaulting the flag on turns the file red;
• direction — the four leaf blocks must stay non-containers on the public tier, so the react-page assertion is about containers and not about everything.

A fourth test renders flex through the real SchemaRenderer and asserts its children still arrive, which is the scope guard for "nothing else changes".

Ablation

Against the committed implementation e44f941ba, with the restore trapped on EXIT INT TERM using absolute paths.

• Mutation confirmed on disk before anything was run: anchor count 1 -> 0, and blob 9cd1720a0... -> f5994b334... (a differing hash, not an exit code).
• Pin on the mutated tree: 2 failed, 9 passed, exit 1. The two reds are exactly the two consumers — "flex with children draws no not-a-container" (expected [ { severity: 'warning', ... } ] to deeply equal []) and "flex reports as a container on the public tier" (expected undefined to be true). The three sibling cases stayed green, since they legitimately declare the flag.
• Restore proven, not assumed: git checkout HEAD -- against the absolute path (naming HEAD, so the mutated index cannot be the source), then git diff HEAD empty and restored blob 9cd1720a0... equal to that path's HEAD blob hash.
• Pin on the restored tree: 11 passed, exit 0.

No rebuild step is involved and none is owed: vitest.config.mts aliases every @object-ui/* specifier to packages/*/src, and the pin reaches flex.tsx by relative import, so the mutated source is what ran. Nothing resolved through dist.

Gates

All run on e44f941ba with a clean tree, every exit code captured by redirect-then-capture and each result quoted from the tool's own verdict line.

The type-check number is qualified rather than quoted bare: packages/components/tsconfig.json excludes src/__tests__, so tsc --noEmit alone says nothing about the new pin. The chained tsc -p tsconfig.test.json is what covers it, and --listFiles was used to confirm the file is actually in that program rather than assumed to be.

The repo-wide pnpm lint is turbo run lint, which fans out to each package's own lint. Only this package's run was taken, and that is a measurement rather than a skip, on three pieces of evidence: the population is eslint's own (422 files, counted from --format json, not from a guess about what counts); both changed files are inside that population and were linted (the test file clean, flex.tsx carrying one pre-existing no-explicit-any warning at line 15 column 94 — the original registration signature, far from the line 146 edit); and eslint.config.js configures no projectService and no parserOptions.project, so type-aware linting is off and each file's verdict is a function of its own source alone — this diff cannot move a verdict in a package it does not touch.

CI owns the full farm and runs it regardless.

Generated by Claude Code