ci: make gradients a UX review finding

[CrysisDeu]

Summary

Adds a flat-fill rule to the UX review lane (main + fork variants): a NEW decorative gradient fill (linear-, radial-, conic-gradient used as a background or surface color) on chrome, a dialog, or an exported image such as a share card is a finding. The expected fill is one solid theme token, or the brand purple #7c3aed on an outward-facing artifact.

The rule is scoped to the defect class it comes from. It explicitly does not flag functional gradients (mask-image scroll-edge fades, loading shimmer, streaming glow) or the shipped gradient mechanisms (appstore gradient art in components/appstore/gradient.ts, which reads as content rather than chrome, and the session 'gradient' color mode) — those are deliberate and test-pinned, and an "any gradient" rule would misfire on every PR that touches them.

The rule is stated once in website/docs/theming-contract.md (next to the existing "never a hardcoded #hex" rule, reachable from website/AGENTS.md), so an author sees it before review; both reviewer prompts cite that doc.

Why

The brand system is flat. A gradient background on a share card in #8040 went through the UX lane without comment because no lens named gradients; it was caught by a human instead. This puts the rule where the reviewer will apply it every time — lens 7 (Consistency & Habituation), next to the existing "theme tokens over hardcoded colors" bullet.

Prompt + doc change, no new lane/gate/script. Both ux-review.yml and fork-ux-review.yml get the identical bullet so fork PRs are held to the same bar.

Tested

• Both workflow files parse as YAML
• scripts/check_brand_name.py (added lines) and scripts/docs-lint.sh pass locally
• CI runs the workflow-shape and docs tests

no linked issue: prompt-only reviewer tightening surfaced during #8040 review.

[github-actions]

Opus 4.8 Review — ✅ no blocking findings

Reviewed 616d0afe82f1a1e74a6a31b4990bc53bfec474fe — this comment is updated in place on each push.

Review details

No findings.

[OPUS-REVIEWED] 616d0af

Verdict parsed from the review's SHA-scoped output markers for commit 616d0afe82f1a1e74a6a31b4990bc53bfec474fe.

False positive or not applicable? A repository writer can comment:
/ai-review override fable 616d0afe82f1a1e74a6a31b4990bc53bfec474fe: <one-sentence reason>

[github-actions]

Design Review (Fable 5) — ✅ PASS

Design-level review of 616d0afe82f1a1e74a6a31b4990bc53bfec474fe — updated in place on each push. A BLOCK verdict blocks PR readiness; PASS/CONCERNS are advisory.

Design-Verdict: PASS

A reviewer-lens gap with a named escaped defect (#8040), fixed at the exact layer that missed it — prompt plus author-facing doc, no new gate.

Watch

• The rule text (including the exclusion list and the #7c3aed literal) now lives verbatim in three files; when a new shipped gradient mechanism lands or the brand hex changes, all three must move together or the lane starts false-flagging — consider making the two prompt copies terser and letting theming-contract.md own the exclusion details.

[DESIGN-REVIEWED] 616d0af

[github-actions]

First Principles Review (Fable 5) — ✅ PASS

Premise-level review of 616d0afe82f1a1e74a6a31b4990bc53bfec474fe — why this exists and whether the shipped surface is the smallest honest version. Updated in place on each push. A BLOCK verdict blocks PR readiness; PASS/CONCERNS are advisory.

All checks are done. The rule duplicates nothing (grep for gradient in website/AUTOSDE.yaml, check-theme-colors.mjs, and all workflows: zero pre-existing hits), the fork/main duplication is that file pair's documented mirroring pattern, the carve-outs name real shipped mechanisms (components/appstore/gradient.ts exists; 'gradient' mode in sessionColors.ts; #7c3aed in 21 files including ShareCard.tsx), and the provenance is a reported defect (#8040). Final review:

First-Principles-Verdict: PASS

A gradient shipped past the UX lane because no lens named gradients; this puts the rule exactly where that lens failed, scoped to the defect class.

What this change ships

Intent: stop new decorative gradient fills from passing UX review unflagged, as one did in #8040. FIX (a review-lane gap with a reported escape), shipped as a prompt+doc addition.

1. UX review now flags a new decorative gradient fill on chrome/dialogs/share cards — justified (reported defect, feat: share assistant messages as branded social cards #8040)
2. Fork PRs are held to the identical rule — justified (fork-ux-review.yml is a declared mirror of ux-review.yml)
3. Theming contract tells authors "fills are flat," with named carve-outs — justified (author sees the rule before the reviewer does)

Verification that grounds the PASS: no existing mechanism covers this — gradient appears nowhere in website/AUTOSDE.yaml, website/scripts/check-theme-colors.mjs (which checks raw literals, not gradients built from tokens), or any workflow before this change (counts: 0, 0, 0). The neighbouring "theme tokens over hardcoded colors" bullet does not catch a token-built gradient, so this is not a second spelling. Every carve-out names a real, test-pinned mechanism in the tree, so the rule won't misfire on the shipped gradient art. The change sits at cause level: the lens that missed the defect is the thing amended, and the author-facing contract now states the rule the reviewer enforces. All three items are declared in the description.

[FIRST-PRINCIPLES-REVIEWED] 616d0af

[github-actions]

GPT 5.6 Review — ✅ no blocking findings

GPT 5.6 completed its review of 616d0afe82f1a1e74a6a31b4990bc53bfec474fe and found no blocking issues.

This comment is updated in place on each push.

Review details

No findings.
[GPT-REVIEWED] 616d0af

False positive or not applicable? A repository writer can comment:
/ai-review override gpt 616d0afe82f1a1e74a6a31b4990bc53bfec474fe: <one-sentence reason>

[CrysisDeu]

Design Review — Watch: "Any gradient is a finding" contradicts the shipped dashboard — fixed in d827b60

The bullet's premise — "the brand system is flat, so any CSS gradient … on a user-facing surface is a finding" — is falsified by the codebase it governs … Scope the rule to the actual #8040 failure class: decorative background/fill gradients, with functional gradients (fade masks, scroll cues, loading shimmer) and existing gradient features exempt.

Verified the cited sites exist (20 non-test files under website/src carry a gradient; index.css:2086 mask-image fades, appstore/gradient.ts header says the art "reads as content, not chrome", sessionColors.ts ships a 'gradient' mode). The bullet in both ux-review.yml and fork-ux-review.yml is now scoped to a new decorative gradient fill (background/surface color) and names the exemptions explicitly: mask-image scroll-edge fades, loading shimmer, streaming glow, the appstore gradient-art module, and the session 'gradient' color mode.

Suggestion: state the rule once in the theming contract and have both prompts cite it — fixed in the same commit. Added a "Fills are flat" paragraph to website/docs/theming-contract.md next to the existing "never a hardcoded #hex" rule (reachable from website/AGENTS.md); both prompt bullets now point at that doc.

[CrysisDeu]

First Principles — Subtraction: shrink the bullet to the harm #8040 demonstrated — fixed in d827b60

Shrink the bullet in both ux-review.yml and fork-ux-review.yml from "any CSS gradient … dashboard chrome, dialogs, and exported images … alike" to the harm #8040 demonstrated: decorative gradient fills (backgrounds) on new surfaces, explicitly not mask-image fades, shimmer skeletons, or the appstore gradient-art mechanism.

Done as stated. The bullet now reads "a NEW decorative gradient fill (linear-, radial-, or conic-gradient as a background or surface color) … is a finding" and carries an explicit NOT-a-finding list: mask-image scroll-edge fades, loading shimmer, streaming glow, components/appstore/gradient.ts, and the session 'gradient' color mode. The Watch item (59 occurrences / test-pinned *scrollEdges* masks would earn a finding under the old wording) is the same defect and is closed by the same edit; the PR body's "the brand system is flat" claim is now qualified to fills rather than all gradients, and the rule is recorded once in website/docs/theming-contract.md so the two YAML copies cite a single source.

[github-actions]

UX Review (Fable 5) — ✅ PASS

UX-level review of 616d0afe82f1a1e74a6a31b4990bc53bfec474fe — updated in place on each push. A BLOCK verdict blocks PR readiness; PASS/CONCERNS are advisory.

UX-Verdict: PASS

No product surface changes — this is a reviewer-prompt and contributor-doc edit; nothing a dashboard user sees or does is touched, and all cited exemptions (components/appstore/gradient.ts, session 'gradient' mode) exist.

[UX-REVIEWED] 616d0af

[CrysisDeu]

Bundle Size Gate red on d827b60 is main-inherited. The gate reports assets/t-*.js: 740.0 KB exceeds its 740.0 KB budget by 4 B; this PR touches no file under website/src. Three locale-adding merges on main (4e5398d, de19df8, 390ac8d) put the i18n runtime chunk over its ceiling, and sibling PRs on the same main window show the same red. Unblocker: #8411 moves the ceiling to 780 KB (re-measured). This PR will be rebased once #8411 lands.

[CrysisDeu]

Rebased onto main 434ff7c, which carries #8412's re-measured t ceiling (777 KB), so the inherited Bundle Size Gate red is cleared without any change to this PR's diff. #8411 was closed as redundant.

[chenmingwei23]

Tier 1 auto-approve: chore/CI tooling (3 files). Criteria: no conflict, no requested changes, security path denylist clean, design-doc gate clean, SAST annotations clean (Semgrep + CodeQL both success, zero PR-scoped alerts), security checklist all-NO, AI reviewers green. Category: adds a flat-fills gradient rule to the UX review workflow prompts plus its theming-contract doc; workflow prompt text with no runtime impact.

[iamwhatever]

Tier 1 auto-approve: chore / build tooling (3 files). Criteria: no conflict, no requested changes, security path denylist clean, design-doc gate clean, SAST annotations clean, security checklist all-NO, AI reviewers green. Category: adds a "flat fills" gradient criterion to the UX review prompt in .github/workflows/ux-review.yml and its fork variant, and documents the same rule in website/docs/theming-contract.md — prompt/doc prose only, no job gating logic and no runtime code path.