`kind:'html'` page source loses the space between a text run and an adjacent inline element — `A <strong>x</strong> page` renders as `Axpage`

[os-sales]

Found while implementing #5470 (rewriting apps/console/src/sdui-tiers-preview.tsx's
page sources off Tailwind). Filing rather than fixing — outside that card's declared file
surface, and it is in packages/sdui-parser, not the console.

What happens

packages/sdui-parser/src/parse.ts:177, inside the children loop:

const text = this.readTextRun();
const trimmed = text.replace(/\s+/g, ' ').trim();
if (trimmed) children.push(trimmed);

The collapse (\s+ → one space) is correct and matches HTML's own whitespace model.
The .trim() is not: it deletes the leading and trailing space of every text run,
including the space that separates that run from an adjacent inline element. HTML
collapses a whitespace run to one space; it does not remove it.

Measured

Rendered through the real PageRenderer in Chromium against
apps/console/sdui-tiers-preview.html (vite dev), reading textContent:

source	rendered
A <strong …>kind:'html'</strong> page — …	Akind:'html'page — …
Full HTML tag set in the <em …>html</em> tier.	Full HTML tag set in thehtmltier.
A trusted <a …>react tier</a> behind a flag.	A trustedreact tierbehind a flag.

Present before and after #5470's change, byte-for-byte identical output — measured by
checking out apps/console/src/sdui-tiers-preview.tsx at 7e811687a (26 classNames in
its page sources, i.e. the pre-change file) and re-rendering the same page:

NEW "intro": "Akind:'html'page — native HTML tags and the blocks' structured props, …"
BASE "intro": "Akind:'html'page — plain native HTML tags, Tailwind classes, …"

So this is not a regression from that card — it is the tier's own behaviour, and it has
been visible in the repo's own worked example for as long as the example has existed.

Why it matters beyond the preview

kind:'html' is the tier the guide tells authors and AI agents to reach for by default
(content/docs/guide/react-pages.md §Choosing between them), and the tag set it advertises
is exactly the prose set — strong, em, a, code, b, i, abbr, q, cite
(packages/components/src/renderers/basic/html-elements.tsx). Any authored sentence with
emphasis or a link inside it renders with the words run together. It is silent: the page
renders, the structure is right, nothing reports anything.

What needs deciding

The fix is small but the exact rule is worth stating rather than guessing, because it is a
save-time parse that pins metadata:

• (a) Replace .trim() with HTML's rule — collapse the run to a single space, and drop
  a run only when it is entirely whitespace and sits between block-level boundaries.
  Highest fidelity; needs a decision about what counts as a boundary in a schema tree that
  has no block/inline distinction of its own.
• (b) Narrower: keep one leading/trailing space whenever the run is adjacent to a
  sibling element (rather than to the start/end of the parent). Mechanical, no boundary
  taxonomy, slightly over-generous inside block containers like <ul>.

Either way this changes parser output for existing sources, so it wants a pin in
packages/sdui-parser/src/__tests__ and a look at whether any snapshot depends on the
current trimming.

[huangyiirene]

Triage first-touch (bare-card sweep): queued pm:queue, Bug, domain:ui (lands in this repo's packages/sdui-parser). User-visible in the tier the guide recommends by default — any sentence with inline emphasis renders with words run together, silently.

Rule choice, decided at triage: option (b) — keep one leading/trailing space when the text run is adjacent to a sibling element (not at the parent's start/end). Mechanical, no block/inline taxonomy invented for a schema tree that has none, and "slightly over-generous inside <ul>" is a whitespace nit while (a)'s boundary taxonomy is a design project. Pin the three measured cases + a <ul>-shaped control showing the over-generosity is bounded; sweep existing parser snapshots for dependence on the current trimming (a snapshot that changes is part of the fix, reviewed not silently accepted). ⚠️ Twin check required in the PR body: objectstack also carries a packages/sdui-parser — establish whether its parse.ts shares the .trim(); if yes, file the upstream twin card (don't fix cross-repo in this PR). Size S.

---

Generated by Claude Code

[os-zhuang]

Claim: PM loop round 17
Session: session_01EuPCi56cnGyykygi3z9w4m
Branch: claude/issue-5661-sdui-parser-inline-whitespace
Worktree: objectui-issue-5661
Domain: domain:ui
File surface: packages/sdui-parser/src/parse.ts and packages/sdui-parser/src/__tests__/** (stop on breach; explain in the report)
Container & model: S, mode:subagent, model: opus
Clause-②: no
Serial constraints cleared: no in-flight card shares packages/sdui-parser. The one sibling still running this round is #5619 (packages/auth + packages/app-shell + apps/console) — disjoint. #5710/PR #5713 and #5637/PR #5714 both merged at 20:42Z / 20:43Z and released their holds.

Tier basis. No path-derived mandate applies (objectui carries no scripts/pm/; the objectstack copy of dispatch-gates.mjs derives against the wrong tree and is not used). Content judgement: this changes parser output, not its accept/reject behaviour, and restores HTML's own whitespace rule rather than inventing one — a declared = enforced repair. Clause ② does not trip. Size follows triage's own S; tier held at the default judgement档 rather than the floor because "which existing snapshot legitimately changes" is a judgement, not a gate verdict.

Premise re-verified on origin/main (now 190fbd01d, after this round's two merges). The defect stands: parse.ts:176-177 still reads

const text = this.readTextRun();
const trimmed = text.replace(/\s+/g, ' ').trim();

⚠️ Line numbers differ from the card's citation — the card says :177, the readTextRun() call is at :176 and the .trim() at :177. Re-derive from the tree; ⛔ do not cite the card's numbers in the PR.

⚠️ The twin check triage required is already answered — measured, not assumed. objectstack does carry packages/sdui-parser/src/parse.ts and it carries the identical construct at :168 (readTextRun() at :167). So the twin is real and the upstream card is owed. ⛔ Per triage, do not fix it cross-repo in this PR — file the upstream twin card with the measurement and link it. Note the line numbers differ between repos (objectui :177 / objectstack :168), so the two files have already drifted; say so in the twin card rather than implying they are a copy.

---

Generated by Claude Code

[os-zhuang]

os-dev-report

{
 "issue": 5661,
 "status": "done",
 "branch": "claude/issue-5661-sdui-parser-inline-whitespace",
 "pr": "https://github.com/objectstack-ai/objectui/pull/5722",
 "head_sha": "051568b5e",
 "premise_still_valid": true,
 "premise_notes": [
 "Issue body read in full and NOT truncated: it ends with the 'What needs deciding' section closing on the snapshot note ('...it wants a pin in packages/sdui-parser/src/__tests__ and a look at whether any snapshot depends on the current trimming.'). Both comments read (triage first-touch 2026-08-22T13:25Z; PM claim 2026-08-22T20:47Z).",
 "ZONE 2.1 confirmed: defect stands on origin/main 190fbd01d.",
 "ZONE 2.2 confirmed: the card's ':177' citation is off by one for the first line. Re-derived: readTextRun() at :176, .trim() at :177. Card line numbers not cited in the PR.",
 "ZONE 2.3 confirmed by re-measurement, not taken on trust: objectstack origin/main (acb4dbc) packages/sdui-parser/src/parse.ts carries the identical statement at :167/:168. Full diff of the two files: the ONLY difference is objectui's provenance stamp (import + 9-line block, objectui#4000), which is exactly the 9-line offset. Twin filed, not fixed cross-repo.",
 "ZONE 2.4 resolved as OUT OF SCOPE: :275 `const trimmed = raw.trim()` is inside interpretBrace(raw) — it normalises a braced ATTRIBUTE's source text before JSON.parse / before it becomes an { $expr } marker. Expression syntax, not the rendered content model; no separator space is at stake. Different defect class, untouched, no card filed."
 ],
 "ruling_followed": "option (b) — leading/trailing space kept when the run is adjacent to a sibling; parent start/end still trim. Option (a) not implemented.",
 "summary": "Replaced the unconditional .trim() in parseChildren with a position-aware decision: collapse the run to single spaces (unchanged), then keep one leading space when a sibling precedes the run and one trailing space when a sibling element follows it. Parent-edge trimming is preserved. Added a 10-test pin suite covering the three card-measured cases, the collapse, the parent-edge behaviour that must not move, a <ul> control bounding the rule's known over-generosity, and two cases asserted through the real SchemaRenderer via textContent. Patch changeset for @object-ui/sdui-parser included.",
 "files_changed": [
 "packages/sdui-parser/src/parse.ts (+23/-2)",
 "packages/sdui-parser/src/__tests__/inline-whitespace.test.tsx (new, 189 lines)",
 ".changeset/sdui-parser-inline-whitespace-5661.md (new, patch bump)"
 ],
 "fence_respected": true,
 "fence_notes": "No file outside packages/sdui-parser/src/** was written. packages/auth, packages/app-shell, apps/console, packages/components, packages/types, content/docs, apps/site: read-only reconnaissance only. objectstack: read-only (git show origin/main), zero edits.",
 "tests": {
 "new_suite": "pnpm exec vitest run packages/sdui-parser/src/__tests__/inline-whitespace.test.tsx => 'Test Files 1 passed (1) / Tests 10 passed (10)'; lock VERDICT command-exit 0.",
 "reverse_verification": "Fix COMMITTED first (bded4edf4), then `git checkout origin/main -- packages/sdui-parser/src/parse.ts` (revert proven on disk: removed line grep-count 1, new-code grep-count 0), suite re-run => 'Tests 9 failed | 1 passed (10)', lock VERDICT command-exit 1. Direction as predicted: RED. The single green is 'still drops the edge space at the parent's own start and end' — the half of .trim() that was already correct. Render case reproduced the card's exact string: expected \"A kind:'html' page — native HTML tags.\", received \"Akind:'html'page — native HTML tags.\". Restored via `git checkout <branch> -- parse.ts` (proven: removed line 0, new-code 3) and `git status --porcelain` empty.",
 "union_batch_1": "pnpm exec vitest run --maxWorkers=2 packages/sdui-parser/ packages/layout/ packages/types/ apps/console/ => 'Test Files 140 passed (140) / Tests 1486 passed (1486)'; lock VERDICT command-exit 0 (held 307s).",
 "union_batch_2a": "pnpm exec vitest run --maxWorkers=2 packages/components/src/__tests__/ => 'Test Files 104 passed (104) / Tests 946 passed (946)'; lock VERDICT command-exit 0 (held 221s).",
 "union_batch_2b": "pnpm exec vitest run --maxWorkers=2 packages/components/src/renderers/ packages/components/src/ui/ packages/components/src/notifications/ => 'Test Files 78 passed (78) / Tests 692 passed (692)'; lock VERDICT command-exit 0 (held 202s).",
 "union_total": "322 test files, 3124 tests, 0 failures.",
 "union_sha_note": "The three union batches ran at bded4edf4. `git diff --name-only bded4edf4 HEAD` = .changeset/sdui-parser-inline-whitespace-5661.md ONLY — no source or test file differs. At the final head 051568b5e the decisive gates were re-run: `pnpm exec vitest run packages/sdui-parser/` => 'Test Files 7 passed (7) / Tests 52 passed (52)', check:control-bytes OK, changeset:check OK.",
 "invocation_note": "All vitest runs from the worktree ROOT with no `--`, per scripts/vitest-invocation-guard.mjs (a package-cwd run is refused and would otherwise go green on another package's files).",
 "resource_note": "Every build/test went through /home/user/objectstack/scripts/pm/os-verify-lock.sh; verdicts read from its VERDICT line, never a bare $?. Longest wait 71s. Sibling holder (issue #5619) observed and waited out, never bypassed."
 },
 "snapshots": {
 "changed": [],
 "statement": "NO snapshot changed and none was updated (no `-u` run was performed anywhere). packages/sdui-parser contains no snapshot files at all. The repo holds exactly two .snap files — packages/components/src/__tests__/__snapshots__/snapshot.test.tsx.snap and .../snapshot-critical.test.tsx.snap. Both owning tests were executed in union_batch_2a and passed with the .snap files unmodified (`git status --porcelain` empty after the run; a mismatch fails rather than rewrites). `vitest list --filesOnly` confirmed all 104 real test files in that directory were collected, so neither snapshot test was silently skipped. Conclusion for triage's sweep: nothing in the repo depended on the old trimming."
 },
 "gates": [
 {"gate": "turbo run type-check lint --filter=@object-ui/sdui-parser", "verdict": "PASS — 'Tasks: 9 successful, 9 total'; tsc --noEmit && tsc -p tsconfig.test.json clean; eslint '0 errors, 7 warnings', all 7 pre-existing in render.test.tsx and verify.ts, none in a file this PR touches"},
 {"gate": "pnpm check:control-bytes", "verdict": "PASS — 'check-control-bytes: OK (scanned 4770 tracked text file(s); skipped 85 binary)'; plus an independent grep -naP over the three changed files, zero hits"},
 {"gate": "pnpm check:phantom-deps", "verdict": "PASS — 'Every in-scope import is declared by the package that publishes it.'"},
 {"gate": "pnpm check:self-import", "verdict": "PASS — 'No package names itself inside its own src/.'"},
 {"gate": "pnpm check:esm-specifiers", "verdict": "PASS — 'no un-ledgered package emits an extensionless relative specifier'"},
 {"gate": "pnpm changeset:check", "verdict": "PASS — fixed group OK, privatePackages declared, 'No changeset declares a major bump.'"},
 {"gate": "pnpm lint:coverage", "verdict": "PASS — '46/46 packages linted, 0 with outstanding errors (0 total).'"},
 {"gate": "pnpm type-check:coverage", "verdict": "PASS — '45/46 via type-check ... 1 not compiled' and '41/41 packages compile their tests'"},
 {"gate": "pnpm check:eager-closure", "verdict": "NOT RUN — exits 2 with 'No eager-closure report at apps/console/dist/eager-closure.json ... broken gauge, not a passing budget' unless the console is vite-built. Declared narrowing: this diff adds no import to any package entry, so it cannot move that budget"},
 {"gate": "pnpm lint (repo-wide)", "verdict": "NARROWED, declared — `lint` is `turbo run lint`, i.e. each package's own `eslint .` plus //#lint:root (which DID run here as a task dependency: 0 errors, 26 pre-existing warnings). Only one package has changed files and eslint.config.js declares no type-aware linting (no `project` / `projectService`), so no untouched package's verdict can move. CI runs the full farm regardless"}
 ],
 "gate_derivation": "Derived from objectui's own .github/workflows/ (ci.yml, lint.yml, control-bytes.yml, changeset-guard.yml, changeset-presence.yml, node-esm-load-gate.yml, performance-budget.yml) and root package.json. scripts/pm/dispatch-gates.mjs was NOT used — it lives in objectstack and derives against that repo's workflows.",
 "open_questions": [],
 "out_of_scope_findings": [
 "filed as objectstack-ai/objectstack#11148: objectstack's packages/sdui-parser/src/parse.ts carries the identical .trim() at :167/:168 on origin/main acb4dbc — the upstream twin triage required. Unassigned, no labels, cross-linked both ways. Its in-repo consumer (packages/lint/src/validate-jsx-pages.ts) is diagnostics-only, so nothing there is visibly broken today; the card states that rather than overclaiming, and records that the two files have already drifted so a line-number cherry-pick will not apply."
 ]
}

---

Generated by Claude Code

[os-zhuang]

Read-back check on the report above: the literal os-dev-report first line survived, so the report is machine-visible. One sanitizer casualty to correct, in tests.reverse_verification: the angle-bracketed placeholder in the restore command was eaten on the write path. It reads git checkout -- parse.ts; what was run is

git checkout claude/issue-5661-sdui-parser-inline-whitespace -- packages/sdui-parser/src/parse.ts

The restore proof either side of it is unaffected (reverted-line grep-count 0, new-code grep-count 3, git status --porcelain empty). Nothing else in the JSON lost bytes — angle-bracketed tokens inside backticks (<ul>, <li>) came through intact, this one was bare.

---

Generated by Claude Code