No gate reads inside a code fence in `skills/objectui/` — 123 TypeScript fences and 63 JSON fences are checked by nothing

[os-litant]

Filed out-of-scope from the #7251 skills-optimization flight (its DEFER row OUI-E-01). Unassigned.

Measured

At the flight's audit ref, skills/objectui/ carried 123 TypeScript fences and 63 JSON/JSONC fences. grep -rc "os:check" skills/objectui/ returns zero.

Every objectui gate that could read them excludes the surface, by construction rather than by accident:

Gate	Why it does not reach a fence in skills/
check-skills-paths.mjs	Deliberately excludes fences — "A fence is a worked example", and a tutorial fence may name a file the reader is about to create. It reads inline code spans in prose only.
check-doc-links.mjs	SCAN_ROOTS has no skills row.
check-doc-component-types.mjs	DOCS_ROOT = 'content/docs'; its header says so verbatim — "This gate walks content/docs and nothing else: not skills/**".
check-doc-snippet-types.mjs	content/docs + package READMEs.
check-doc-fence-languages.mjs	content/docs + package READMEs. It judges the fence language tag, never the contents.

Net: this repo enforces path existence in skill prose, and nothing else. A fence can import a symbol that does not exist, spell a key the renderer never reads, or restate a shape the spec changed, and stay green forever.

What already exists to copy

objectstack has packages/spec/scripts/check-skill-examples.ts, with the convention that an HTML comment containing os:check, placed on the line above a fence, opts that fence into a real type-check. A port would give objectui the same opt-in marker plus a runner over skills/**.

Why this is the highest-leverage structural addition here

It is the mechanism that stops the drift class from recurring. Two rounds of it have already been cleaned by hand — #3658 / #7094's truth sweep, and #7251's own correctness items, which included an eval asserting a mobile hook API (useSwipe / useLongPress) that has zero occurrences repo-wide. Every one of those was found by a human reading, which is precisely the economics check-skills-paths.mjs was created to end for paths.

Shape of the work

1. Port the marker convention and the runner; opt-in, so it starts at zero red.
2. Measure how many of the 123 fences compile as-is, and land the marker on the ones that already pass — that number is the honest starting ratchet.
3. Decide separately whether the ratchet should be shrink-only, the way check-doc-fence-languages.mjs treats its declared-file population.

Related: #7358 (the same package's other unguarded surface — the paths gate's scan root).

[os-litant]

Design input from implementing #7360 (PR #7408), not a request to change scope here.

The oracle this check picks is load-bearing, and the two candidates disagree on a real row today.

#7360's audit measured each must_contain token against guides/BASENAME.md — the guide matching the eval file's own basename. That per-guide oracle is what found its rows. But the "teach the thing" half of that card is closed by adding one bullet to guides/plugin-development.md naming PluginScope.registerComponent, and three assertions require that token: evals/plugin-development.json evals 1 and 2, and evals/page-builder.json eval 1.

After PR #7408 the counts are:

• registerComponent in guides/plugin-development.md: 0 to 1
• registerComponent in guides/page-builder.md: 0, unchanged

So evals/page-builder.json eval 1 is honest under a bundle-wide oracle (the answering agent loads the whole skill, router and all guides) and still untaught under a per-guide one. Whichever this check adopts, that row is either fine or red — worth deciding deliberately rather than discovering when the gate first runs.

Second input: whole-token comparison changes what the check can see. #7360's body already raises this; two measured examples, both in guides/testing.md, both counting 1 rather than 0 so a substring check reports them clean:

• create — matches create: vi.fn().mockResolvedValue({ id: '1', name: 'New' }), a mocked dataSource method, where the eval means the form mode
• view — matches the comment // vite preview

A third, FieldWidgetProps, matched only the head of FieldWidgetPropsSchema and is corrected in PR #7408. These three are the argument for comparing whole tokens; they are filed as rows in #7405 so they are not lost if this check lands with substring semantics.

Third, smaller input: must_not_contain deserves the opposite treatment. #7370 deliberately moved a token there as "props": { rather than the bare word, because a bare props fails any answer that names the rule while following it. A check that validates must_not_contain entries against the guides the way it validates must_contain would get that backwards.

---

Generated by Claude Code

[os-litant]

Claim: skills lane seat — graded and dispatched in one write (priority:p2 + pm:dispatched; domain:skills already held). Triage is vacant in objectui and this is the lane's own programme follow-up (filed from #7251's DEFER row OUI-E-01), so the seat grades it. Value gate (#13597): it is the mechanism that ends a drift class two hand-sweeps have already paid for — passes.
Session: session_01LraLgQVGq8egUwfYZpbYt1
Branch: claude/issue-7359-skill-fence-check
Worktree: objectui-issue-7359
Clause-②: no — a gate, no contract claim. Governed: the marker landing touches skills/objectui/** ⇒ the PR stays DRAFT (ui#6866 批 #7); review requests are the seat's step.

Scope, exactly the card's three steps with step 3 recorded not decided:

1. Port the <!-- os:check --> opt-in convention from objectstack's packages/spec/scripts/check-skill-examples.ts (marker on the line directly above a fence) as an objectui runner over skills/objectui/**/*.md — TypeScript/TSX fences type-checked, JSON/JSONC fences parsed — with a --self-test, a scripts/__tests__/ spec, a package.json script, and a workflow modelled on skills-paths.yml (no paths / paths-ignore, for the reason that file states; its test pins that).
2. Measure how many fences compile as-is at the branch point and land the marker on exactly those — the honest starting count, stated in the PR body. Opt-in ⇒ zero red on day one.
3. ⛔ Whether the marked population becomes shrink-only (the way check-doc-fence-languages.mjs treats its declared files) is NOT decided in this flight — record the number, add no ratchet.

Re-measured at objectui origin/main bf244f4, 2026-09-03T01:22Z: 112 ts/tsx/typescript fences and 56 json/jsonc fences across skills/objectui/**/*.md (the card's 123 / 63 were at its audit ref); os:check occurs 0 times. ⛔ The eval-token oracle discussed in comment 5514262423 (must_contain tokens vs the guides, whole-token vs substring) is a different check and is not this card — the seat records it separately. No guide prose is edited by this flight beyond the marker lines.

---

Generated by Claude Code

[os-litant]

Claim: dev seat (os-dev subagent of the dispatching PM session — same session id, distinct seat)
Session: session_01LraLgQVGq8egUwfYZpbYt1
Branch: claude/issue-7359-skill-fence-check
Worktree: /home/user/objectui-issue-7359
Branch point: 75e7436 (objectui origin/main at worktree creation — main moved on from the bf244f4 the seat re-measured at, so the starting population is re-measured at this sha).

Scope taken is exactly the seat's three steps in comment 5518884588: port the opt-in marker convention + a runner over skills/objectui/**/*.md, wire CI, and land the marker on exactly the fences that pass as-is — number recorded, no ratchet. The eval-token oracle in comment 5514262423 is a different check and is not built here.

---

Generated by Claude Code