finding(skills): five wrong shipped-API facts in the published objectui guides, all in fences no gate reads

[claude]

Filed by the skills seat from the --measure sweep run on #7675 (that card asked for os:check markers; this is what the sweep found instead).

What

Five sites in the published skills/objectui/guides/ bundle state a shipped-API fact that the built dist/*.d.ts contradicts. Each is a fact an AI reader copies verbatim, and none is reachable by any gate today: all five sit in UNMARKED fences, which check-skill-examples.mjs ignores by design.

Measured on a472b07 after turbo run build with the gate's own --build-filter:

guide · fence	diagnostic	the shipped fact
auth-permissions.md:22	TS2322 at :29 — authClient does not exist on AuthProviderProps	the AuthProvider element is handed an authClient prop the props type has no member for
auth-permissions.md:105	TS2559 at :110 — AuthClient has no properties in common with CreateAuthenticatedFetchOptions	createAuthenticatedFetch is called with the client itself, not with its options object
auth-permissions.md:237	TS2322 at :242 — Element is not assignable to "hide" | "disable" | "custom" | undefined	a closed union prop is handed a JSX element
i18n.md:94	TS2339 x4 at :98 — formatCurrency, formatDate, formatNumber, formatRelativeTime	all four ARE exported by @object-ui/i18n, as module-level functions; the guide destructures them off useObjectTranslation(), whose return type has none of them
project-setup.md:276	TS2322 at :280 — value does not exist on MetadataProviderProps	the MetadataProvider element is handed a value prop the props type has no member for

The i18n.md row is the cheapest and the most clearly wrong: the four functions exist, at packages/i18n/dist/index.d.ts:49, just not where the guide says.

Why it is a card and not a rider

Each fix is a skills judgement on a GOVERNED surface (two approvers), and the fences stay UNMARKED afterwards regardless — every one of them also references undefined placeholder identifiers, so none can be opted into the gate without a rewrite. Correcting the fact and opting the fence in are two different pieces of work; this card is the first.

Re-derive

node scripts/check-skill-examples.mjs --build-filter   # feed to turbo run build
node scripts/check-skill-examples.mjs --measure 2>&1 | grep '\[semantic\]' | grep guides/

---

Generated by Claude Code

[claude]

Claim: PM loop round 5 — self-triaged into the skills lane (p2) and executed on the branch the sweep already opened (draft PR #7677), by the same dev
Session: session_019RfFHiRCSs3JXLK4cwcfox
Branch: claude/issue-7675-skill-guide-fences-os-check (PR #7677 — its body becomes Fixes #7676)
Worktree: objectui-issue-7675
Domain: domain:skills
File surface: skills/objectui/guides/auth-permissions.md (three sites), skills/objectui/guides/i18n.md (one site, four helpers), skills/objectui/guides/project-setup.md (one site), plus the usePullToRefresh ref fix already on the branch in mobile.md — each a one-line correction of a shipped-API fact to what dist/*.d.ts declares, ⛔ no rewrite of any fence, ⛔ no markers, ⛔ no package source.
Container & model: S, mode:subagent, model: opus
Clause-②: no (documentation corrected to the shipped API; the API is untouched)
Serial constraints cleared: PR #7677 is the only open PR on skills/objectui/guides/; the fixes ride on it. Governed (skills/**) ⇒ draft PR, in-seat review, os-zhuang + hotlong requested by the seat, human merge.

Grading (22:2xZ): p2 — six copy-and-break shipped-API facts in the published corpus (an AI reader copies them verbatim), none reachable by any gate today; no runtime break, the packages are right, so not p1. The card's own boundary is kept: facts only, fences stay fragments.

---

Generated by Claude Code

[claude]

os-dev-report
{
"issue": 7676,
"status": "done",
"branch": "claude/issue-7675-skill-guide-fences-os-check",
"pr": "#7677",
"premise_still_valid": true,
"summary": "All five carded sites corrected, plus one repeat instance of the same fact, each a one-line correction to what the built dist/*.d.ts declares; no fence rewritten, no marker added, no package source touched. (1) mobile.md gesture fence: usePullToRefresh returns an object (ref, isRefreshing, pullDistance), not a bare ref, so the call site now destructures the ref member. (2) auth-permissions.md provider fence: the AuthProvider prop is client, not authClient - and correcting the name exposed that authUrl is REQUIRED on AuthProviderOptions, so the same line now supplies it; the nesting-order fence at the end of the guide stated the identical wrong prop and was corrected the same way, which is the sixth correction. (3) auth-permissions.md fetch fence: createAuthenticatedFetch takes CreateAuthenticatedFetchOptions (only sameOriginOnly is optional there), never the auth client; the bearer token comes from storage. (4) auth-permissions.md guard fence: PermissionGuard's fallback is the closed union 'hide' | 'disable' | 'custom' and custom content belongs in fallbackContent, so the JSX element moved to the right prop. (5) i18n.md formatting fence: the four format helpers are module-level exports of @object-ui/i18n (packages/i18n/dist/index.d.ts:49), not members of useObjectTranslation()'s return - and each takes an options object, not a positional format string, so the import line, the destructure line and the formatDate call were all corrected; correcting only the import would have introduced a fresh error on the positional format string one line down. (6) project-setup.md provider fence: app-shell's MetadataProvider takes adapter (plus optional ttlMs), not value. Per guide: auth-permissions.md 4 corrections, i18n.md 1 site (4 helpers, 2 lines), project-setup.md 1, mobile.md 1. All six fences stay UNMARKED per the mandate. PR #7677 body rewritten: first line is now the closing relation to this card, with the six-row correction table, per-row differential evidence, gate verdicts and the changeset verdict; it stays a draft, no ready flip, no reviewers requested, no auto-merge.",
"tests": "DIFFERENTIAL DIAGNOSTIC per site - the named TS code present in a --measure log taken BEFORE the edit and absent from one taken AFTER, with the fence's remaining diagnostics being the placeholder class. Row 1 mobile.md:48:35 TS2322 (pull-to-refresh result object not assignable to the element ref prop) -> ABSENT; fence keeps only TS2304 CardContent. Row 2 auth-permissions.md:29:19 TS2322 ('authClient' does not exist on the provider props) -> ABSENT; that fix exposed an interim TS2741 ('authUrl' is missing but required), which is ALSO ABSENT after the required prop was supplied - reported rather than hidden, since it means the first correction alone was incomplete; fence keeps only TS2304 AppContent. Row 3 auth-permissions.md:110:53 TS2559 (AuthClient has no properties in common with CreateAuthenticatedFetchOptions) -> ABSENT, and that fence now has NO diagnostics at all. Row 4 auth-permissions.md:242:56 TS2322 (Element not assignable to the closed union) -> ABSENT; fence keeps only TS2304 AccessDenied and BulkDeleteButton. Row 5 i18n.md:98:11/:27/:39/:53 TS2339 x4 (none of the four helpers exists on the hook's return) -> ALL FOUR ABSENT with no new diagnostic replacing them; fence keeps only TS7031 implicit-any on its own untyped example props - reported precisely, as TS7031 is not the TS2304 placeholder class. Row 6 project-setup.md:280:21 TS2322 ('value' does not exist on the provider props) -> ABSENT; fence keeps only TS2304 myAdapter. CORPUS LEVEL: semantic failures 81 -> 80 over 95 judged ts fences, and exactly ONE verdict flipped in the whole corpus - auth-permissions.md:105 went FAIL -> pass, making it the first unmarked fence in skills/objectui/guides/ that compiles. Left UNMARKED per the mandate. Every edit was proven on disk BEFORE any measurement by anchor grep counts in both directions (old text -> 0, new text -> 1) plus git diff --stat; the i18n import anchor appeared twice in the file, so that edit was scoped by line index to the target fence and the marked+passing fence at :39 was verified untouched. GATE UNION re-run ON THE FINAL COMMIT a7dfbcc, every exit code captured by redirect-then-capture (never through a pipe): check:skill-examples exit 0 - 'Marked: 13 ts fence(s) (floor 13), 70 json fence(s) (floor 70)', 'Semantic phase: 13 of 13 ts fence(s) judged, 0 failed', 'JSON phase: 70 fence(s) parsed, 0 failed', 'Bare any: 1 finding(s), 1 declared in KNOWN_BARE_ANY_EXAMPLES, 0 NOT declared, 0 declared row(s) no longer red', 'Every marked skill example holds up against the built types'. MARKED POPULATION AND BOTH FLOORS UNCHANGED at 13 ts / 70 json - no marker moved, so MARKED_FLOOR is not re-pinned. check:skills-paths exit 0 'OK (88/89 stated path(s) resolve across 20 guide file(s); 1 baselined)'; check:skill-eval-tokens exit 0 'Every must_contain token is taught by its own skill bundle.'; check:doc-fences exit 0 'every TypeScript block in 227 document(s) is fenced ts/tsx/typescript'; check:control-bytes exit 0 'OK (scanned 6258 tracked text file(s); skipped 85 binary)'; check:governed-queue-guard exit 0 '132 cases pass'. Control-byte self-scan (grep -naP over the C0 range plus DEL) over all three newly edited guides returned no match. CHANGESET VERDICT (objectui has no skip-changeset label, the script decides): node scripts/check-changeset-presence.mjs exit 0 - 'Compared the working tree with a472b07 (merge-base with origin/main): 4 file(s) changed, 0 of them published source of a package the release covers, 0 of them a manifest whose published contract moved, 0 under a package changesets ignores, 0 changeset(s) added.' / 'No source or published contract of a released package changed in this range, so no changeset is owed.' RATCHETS (whole-file lines, before -> after): auth-permissions.md 348 -> 348; i18n.md 187 -> 185; project-setup.md 345 -> 345; mobile.md 168 -> 168; published SKILL.md bundle sum 137 lines, untouched. Net -2 lines, so no content was purchased. PR body read back to its end after the update: stored text is the sent body verbatim plus the harness's appended footer, first line intact, all six table rows present, model-name grep on the stored body ZERO. No ablation was run - nothing here required one.",
"mcp_calls": "0 - every GitHub read and write this round went through repo-scoped REST with $GITHUB_TOKEN (claim-comment read, PR update, PR body read-back, comment create, comment read-back); 0 for the whole run",
"open_questions": [],
"out_of_scope_findings": [
"NOT filed, reported for the seat's call: correcting row 3 made skills/objectui/guides/auth-permissions.md:105 the first unmarked guide fence that compiles clean (semantic failures 81 -> 80, one verdict flip). Opting it in would raise MARKED_FLOOR ts 13 -> 14. No card was filed because ruling A settled the marker direction for this corpus (guides stay prose-first, no markers), so whether this single now-eligible fence is an exception is the seat's decision, not a defect to triage."
]
}

---

Generated by Claude Code

[claude]

ACCEPT — governed in-seat review of PR #7677, head a7dfbcca (skills seat, 2026-09-04T22:3xZ).

Implemented-by: claude/issue-7675-skill-guide-fences-os-check (os-dev subagent; the branch the sweep opened, kept for this card)
Reviewed-by: session_019RfFHiRCSs3JXLK4cwcfox (skills seat)

Verified in a detached compare worktree at the PR head, not from the report: four published guides, 9 insertions / 11 deletions, every hunk a one-line correction of a shipped-API fact to what the built dist/*.d.ts declares — AuthProvider takes client (and the required authUrl), at both sites; createAuthenticatedFetch() takes its options, not the client; PermissionGuard's fallback is the closed union and the element moves to fallbackContent; the four i18n format helpers are module exports of @object-ui/i18n taking options objects, so the import, the destructure and the formatDate call all change together; usePullToRefresh returns an object whose ref is destructured; MetadataProvider takes adapter. No fence rewritten, no marker added, no package source touched. The dev's per-site differential diagnostics (the named TS code present before and absent after, the residue being placeholder-only) are the verification the seat relies on — one corrected fence (auth-permissions.md's fetch example) now compiles clean, the first unmarked guide fence that does; corpus semantic failures 81 → 80. Gates on the tree: check-changeset-presence (no changeset owed — objectui's decider), check-control-bytes, check-skills-paths exit 0; the dev's report adds check:skill-examples unchanged at 13/70 (both floors kept), check:skill-eval-tokens, check:doc-fences, check:governed-queue-guard green; net −2 lines. PR checks at the head: 22 success, 3 skipped, 3 still running, none failed. PR body starts Fixes #7676, six-row 改前 → 改后 table with the differential evidence, no model identifier; dev report first on this card with the literal os-dev-report line; zero MCP calls.

Residue, accepted without a card: the now-clean fetch fence could be opted into the marker (13 → 14); ruling A on objectui#7675 keeps the guides prose-first, and a single marker is not worth a governed PR of its own — it rides on the next guide correction that touches that file.

Landing regime: governed (skills/**) ⇒ the PR stays a draft; reviewers os-zhuang + hotlong requested in this same act; human merge.

---

Generated by Claude Code