design-pack: second plugin with three design-craft skills

[F-e-u-e-r]

design-pack: second plugin with three design-craft skills

A sibling plugin (design-pack/, v0.1.0) in the same marketplace: ui-design-craft, motion-craft, design-review-gate - domain-craft doctrine in the pack's house style (numeric budgets, prohibited patterns with signatures, observable gates). opus-pack itself bumps to alpha-0.1.15; its nine skills are untouched.

Why this shape

• Direction set across two sessions with the owner: survey of 14 candidate reference repos (5 parallel research agents; every load-bearing quote re-grepped in the clones), then 4 owner decisions: drop the 2 software-architecture repos; stay in this repo as a second plugin (multi-plugin marketplace mechanics doc-verified); 3 skills; gstack framework-only.
• The brand/DESIGN.md question was consulted, not guessed: grok-4.5 (high) and gpt-5.6-sol (max), isolated runs, independently converged on the same variant neither the packet nor the author proposed - no fourth skill; a format-agnostic design-contract evidence section inside design-review-gate. §4 implements it: authority classes (project-governing / first-party / unofficial observation), verify-100%-of-used observational tokens, drift direction, anti-impersonation, ownership boundary vs domain-evidence-discipline.

Sources and licensing

Text adapted (MIT, notices added): emilkowalski/skills, Leonxlnx/taste-skill, LottieFiles/motion-design-skill, referodesign/refero_skill (mined directly - shortest license chain; refero LICENSE verified MIT). Ideas only: nexu-io/open-design (Apache-2.0; misquote corrections restated with primary citations), garrytan/gstack (its unmeasured numeric heuristics deliberately NOT imported - false precision), benjitaylor/agentation (PolyForm Shield - ideas only by necessity), archify, creative-tim, VoltAgent corpus (pointed at as non-authoritative research material, never vendored). Embedded marketing directives in sources (Emil course plug, GSAP recommend-steering, PRO-key funnels) stripped per skill-authoring §6. "Every incorporated code lineage terminates in MIT" still holds.

Covenant: probes ran before this PR

Private fixtures (evals/design-round1/, gitignored), fresh haiku, bare-vs-ruled arms, expected outcomes written first, n=1 per cell (smoke):

Probe	Bare arm	Ruled arm	Verdict
A planted-slop page (ui-design-craft)	0/7 planted tells; CSS nits only	6/7 + bonus catches; missed em-dash tell (recorded)	PASS
B over-budget motion spec (motion-craft)	~3/7 loose (did catch reduced-motion)	7/7 rule-cited	PASS
C stale observational contract + imitation ask (design-review-gate §4)	adopted wholesale; authority inverted	BLOCK; 3/3 contradictions; impersonation refused	PASS
D drift direction	correct (!)	correct + rule quoted	NULL - trap under-armed; clause stays unprobed, harder variant owed

Round 0 (8 runs) VOIDED: planted-violation keys had leaked into the fixtures as comments; caught because ACTUAL(BARE) wildly beat EXPECTED(BARE) - the expected-before-actual mechanism, not the score. Ledger: evals/design-round1/PROBES.md.

CI

checks.py generalized to N marketplace plugins (skill-frontmatter scan and README-mention check now DISCOVER roots from marketplace.json; per-plugin manifest identity + version agreement; hooks-field ban on every plugin manifest; fail-closed). Negative-tested live: planted design-pack version mismatch -> the new check fires with the exact message; reverted; suite green. Hook suites + gate template green locally.

Review trail

• Fresh-context three-lens review (skill-authoring §6, same-model fresh agent): verdict BLOCK - 5 must-fix + 11 nits. All 5 must-fix and 8 nits fixed in the follow-up commit: Apple-"HIG" bezier relabeled to kCAMediaTimingFunctionDefault/CSS ease (the HIG publishes no numeric curves - the mislabel is now itself listed in the skill); WCAG sentence rewritten to true AA floors (placeholder/over-scrim text 4.5:1 normal-size, secondary body text 4.5:1 both modes); a 40-60% scrim figure that traced ONLY to a declared-unmined repo (survey debris entering under wrong attribution) reworded out and re-anchored on Material's published 32% scrim token; touch-audit snippet MIN=48 + zero-area exclusion + floor-home pointer; emoji tell curated in so provenance matches contents. 1 nit rejected with reason (#4f46e5 is Tailwind indigo-600, a palette fact in the verified P0 list). The critic's named point-nearest-failure (two attributions it could not verify) was closed by re-grepping both in the clones: gstack anti-convergence phrasing (design-shotgun:1122) and creative-tim's text-xs whitelist (SKILL.md:161) both confirmed.
• Cross-model gate, round 1 (grok-4.5 high; gpt-5.6-sol max; isolated runs, 961-line self-contained packet): grok FIX(6), gpt Block/FIX(12); merged 18 findings -> 17 accepted and fixed in ae8a5af, 1 partially adopted with reason (component-override rejection scoped to hooks+skills), 1 sub-claim rejected with reason (#4f46e5 attribution). Standout catches: the spring table contradicting its own settle rule (damping ratio 0.625-0.75 = underdamped); the misquote-correction section itself miscarrying the Doherty cite (real source: IBM report GE20-0752, "Thadhani"); the five-state table too close to its Apache source for an "ideas only" claim -> reclassified as an attributed Apache-2.0 adaptation. gpt round-1 wrapper exited -1 at the 10-min cap AFTER completing (verdict fully written, 339k tokens; identity = invocation flags + stream); recorded as a note, not a gap.
• Cross-model gate, round 2 (post-fix packet with dispositions): grok FIX(2) - it independently re-verified all 13 round-1 closures incl. re-planting the checks.py negative tests itself (unsandboxed run; repo tree verified clean after); gpt FIX(9, of which 2 were already fixed post-packet by grok's items - packet staleness recorded). All 9 unique residuals accepted and fixed in efdaa8e: README table cells aligned with the Apache reclassification, §0 APP-UI carve-out ("everything except §2"), Material scrim as color-role + component default, WCAG 1.4.11 scoped with a labeled pack floor, accent-box override escape + computed-color check, paint-only vs layout-triggering transition split, spectacle-exception hook in the 500ms CRITICAL, three-valued verdict model (Block / Changes required / Approve), document-only review branch, touch-audit ARIA coverage + effective-hit-area confirmation, and four more checks.py guards (plugins:null, non-string version at the opus-pack site, root hooks.json, skill-frontmatter hooks) - each negative-tested.
• Cross-model gate, round 3 (cap round, verify-closures packet): grok-4.5 PROCEED - full closure table verified against the live tree (byte-for-byte match; re-ran the suite), residuals judged non-blocking quality notes; gpt-5.6-sol FIX(6) - all six were NEW findings, not failed closures ("Round-2 closures are present"). Per the bounded-loop cap (cross-model-review §4) and this repo's PR delegation + gates: silent-clobber worker and zero-input-scanner fake pass #44 precedent, the six were fixed and negative-tested post-round instead of opening round 4: touch floors relabeled pack-floors-over-vendor-guidance; physics units stated (px/s vs px/ms 1000x trap) + zero-displacement guard; linear scoped to finite point-to-point motion; CSS-var recalc scoped to inheriting properties; checks.py gains a tolerant hooks-key regex + commands/root-SKILL sweep, marketplace name pin + owner + kebab-case validation, reverse plugin discovery, and design-pack README version binding; plus grok's non-blocking root version-type guard. Negative probes: quoted "hooks" : spelling caught; unreferenced plugin manifest caught; stale README version caught; marketplace rename caught. Suite + hook tests green at HEAD.

Verdict files, packets, and per-round stderr live in the session scratchpad (consult/); probe ledger in the private suite (evals/design-round1/PROBES.md).

🤖 Generated with Claude Code