fix(decompose): address bot follow-up review (3 minor + 1 stale-diff false alarm)

Follow-up review (commit e1afcdf, 2026-05-01T21:19) flagged 5 items.
Triage:

- [Major false alarm] Scope creep: bot claimed extensive .github/* changes
  in this PR. Verified `git diff --name-only upstream/main..HEAD` — zero
  .github/* files in the actual diff. Bot was looking at a stale snapshot
  before the merge with upstream/main propagated. Will note this in the
  reply on the PR.

- [Major] PR body still said `Closes #225 (Phase 1 only)` even though
  the changeset was updated to `Refs #225`. Per PULL_REQUEST_TEMPLATE.md,
  `Closes` only fully resolves; this PR is Phase 1 of a multi-phase
  feature. Edited PR body via `gh pr edit` — `Closes #225` → `Refs #225`
  in the linked-issue line, and "This PR closes the Phase 1 part of #225"
  → "This PR addresses Phase 1 of #225".

- [Minor] Three changesets where convention is one. Consolidated
  `feat-decompose-to-ui-kit.md` + `feat-verify-ui-kit-visual-parity.md`
  into the existing `decompose-to-ui-kit.md` (which already had the
  correct `Refs #225` wording from the previous fix pass). Now one
  changeset, comprehensive description, all 3 packages bumped consistently.

- [Minor] README image URLs pointed at the fork's feature branch
  (`raw.githubusercontent.com/HomenShum/open-codesign/feat/decompose-to-ui-kit/...`)
  which would 404 after merge. Repointed both EN + ZH-CN README images
  at `OpenCoworkAI/open-codesign/main/...` so they keep working post-merge.

- [Minor] `dataUrlToBase64` used `dataUrl.indexOf('base64,')` — case-
  sensitive. RFC 2397 specifies `;base64` token is case-insensitive in
  practice (browsers accept both). Switched to a case-insensitive regex
  match (`/;base64,/i`). Throws same error message on miss; logic
  identical for canonical lowercase inputs.

Verification:
  - apps/desktop: pnpm typecheck (node + web tsconfig) clean
  - packages/core: pnpm typecheck clean

Author: HomenShum

.changeset/decompose-to-ui-kit.md

@@ -4,6 +4,14 @@
 "@open-codesign/i18n": patch
 ---
 
-Add **Decompose to UI Kit** — one-click in the chat sidebar emits a `ui_kits/<slug>/` folder shaped for coding-agent handoff (`index.html` + `components/*.tsx` + `tokens.css` + `manifest.json` + `README.md`). Built-in deterministic + vision verifiers self-check parity using a 12-question boolean rubric (`parityScore = passCount / totalChecks`, no LLM-fabricated floats) and re-iterate on gaps. Per-decompose cost surfaces inline as a toast.
+Add **Decompose to UI Kit** — opt-in sidebar action that emits a `ui_kits/<slug>/{index.html, components/*.tsx, tokens.css, manifest.json, README.md}` bundle shaped for downstream coding-agent handoff (Claude Code, Cursor). Decomposition is prompt-driven (no AST/parser deps); the orchestrator persists the structured plan to the virtual fs in a single atomic call. Output carries `schemaVersion: 1` so downstream consumers can evolve safely.
+
+Three new agent tools in `packages/core/src/tools/`:
+
+- `decompose_to_ui_kit` — orchestrator. Emits the full bundle from a source image + design brief.
+- `verify_ui_kit_parity` — deterministic verifier (no LLM, no cost): element-count parity, visible-text coverage, token coverage. Returns `passCount/totalChecks` derived score (no fabricated floats).
+- `verify_ui_kit_visual_parity` — vision-LLM judge wrapper. 12-check boolean rubric across 5 dimensions (layout / color / typography / content / components), anchor-calibrated reasoning-then-score chain-of-thought (WebDevJudge / Prometheus-Vision / Trust-but-Verify ICCV 2025). Host injects `renderUiKit` (headless screenshot) and `judgeVisualParity` (multimodal call) via the same deps interface as `generate_image_asset`. Without injections the tool returns `status: "unavailable"` and the agent proceeds with the deterministic verifier alone.
+
+`decomposePrompt.ts` (EN + ZH) walks the agent through decompose → verify (both) → reconcile gaps → iterate (max 2) → done with HONEST cost summary. Per-decompose cost surfaces inline as a toast.
 
 Refs #225 (Phase 1 of the requested image → componentization → prototype workflow). Phase 2 (cross-page flows, state machines, prototype orchestration) is tracked separately.

.changeset/feat-decompose-to-ui-kit.md

@@ -232,11 +232,11 @@ Add a `SKILL.md` to any project to teach the model your own taste.
 - **Comment mode** — click any element in the preview to drop a pin, leave a note, and let the model rewrite only that region
 - **Decompose to UI Kit** — one click in the chat sidebar emits a `ui_kits/<slug>/` folder (`index.html` + `components/*.tsx` + `tokens.css` + `manifest.json` + `README.md`) shaped for coding-agent handoff. Built-in deterministic + vision verifiers self-check parity using a 12-question boolean rubric (no floating-point arbitrary scores) and re-iterate on gaps. Per-decompose cost surfaces inline as a toast. See [BENCHMARKS.md](./BENCHMARKS.md).
 
-  ![Decompose to UI Kit — source image vs agent-emitted ui_kit, side-by-side parity check](https://raw.githubusercontent.com/HomenShum/open-codesign/feat/decompose-to-ui-kit/website/public/screenshots/decompose-to-ui-kit.png)
+  ![Decompose to UI Kit — source image vs agent-emitted ui_kit, side-by-side parity check](https://raw.githubusercontent.com/OpenCoworkAI/open-codesign/main/website/public/screenshots/decompose-to-ui-kit.png)
   <sub>Source image (gpt-image input) on the left, agent-emitted <code>ui_kit</code> rendered headlessly on the right. Parity score and status are derived deterministically — <code>parityScore = passCount / totalChecks</code> — from the 12-check boolean rubric. Numbers are from a real <code>e2e-opus-final</code> run, not a mock.</sub>
 
-  ![Iter-0 → iter-1 reconcile loop with honest score drift](https://raw.githubusercontent.com/HomenShum/open-codesign/feat/decompose-to-ui-kit/website/public/demos/decompose-iter-reel.gif)
-  <sub>4-frame reel from the <code>e2e-nodebench-iter</code> run: source → iter-0 (parityScore 0.82, 6 gaps) → iter-1 (parityScore 0.78, 5 gaps) → honest verdict. The agent fixed some gaps and introduced new layout drift; the boolean rubric exposes the regression instead of hiding it. <a href="https://raw.githubusercontent.com/HomenShum/open-codesign/feat/decompose-to-ui-kit/website/public/demos/decompose-iter-reel.mp4">MP4 version</a>.</sub>
+  ![Iter-0 → iter-1 reconcile loop with honest score drift](https://raw.githubusercontent.com/OpenCoworkAI/open-codesign/main/website/public/demos/decompose-iter-reel.gif)
+  <sub>4-frame reel from the <code>e2e-nodebench-iter</code> run: source → iter-0 (parityScore 0.82, 6 gaps) → iter-1 (parityScore 0.78, 5 gaps) → honest verdict. The agent fixed some gaps and introduced new layout drift; the boolean rubric exposes the regression instead of hiding it. <a href="https://raw.githubusercontent.com/OpenCoworkAI/open-codesign/main/website/public/demos/decompose-iter-reel.mp4">MP4 version</a>.</sub>
 - **Generation cancellation** — stop mid-stream without losing prior turns
 
 ### Preview and workflow

.changeset/feat-verify-ui-kit-visual-parity.md

@@ -230,11 +230,11 @@ brew install --cask opencoworkai/tap/open-codesign
 - **Comment mode**：点击预览中的任意元素，留下批注，模型只重写对应局部
 - **拆解为 UI Kit**：聊天侧边栏一键, 把当前 artifact 拆成 `ui_kits/<slug>/` 目录（`index.html` + `components/*.tsx` + `tokens.css` + `manifest.json` + `README.md`），形态对齐 coding agent 接入。内置确定性 + 视觉双 verifier 用 12 项 boolean check 自检（不是浮点分数），不达标自动迭代。每次成本以 toast 实时显示。详见 [BENCHMARKS.md](./BENCHMARKS.md)。
 
-  ![拆解为 UI Kit — source 与 agent 生成 ui_kit 并排对比](https://raw.githubusercontent.com/HomenShum/open-codesign/feat/decompose-to-ui-kit/website/public/screenshots/decompose-to-ui-kit.png)
+  ![拆解为 UI Kit — source 与 agent 生成 ui_kit 并排对比](https://raw.githubusercontent.com/OpenCoworkAI/open-codesign/main/website/public/screenshots/decompose-to-ui-kit.png)
   <sub>左边是 gpt-image 生成的 source 图，右边是 agent 输出的 <code>ui_kit</code> headless 渲染结果。parity score 与 status 完全由 12 条 boolean check 推导：<code>parityScore = passCount / totalChecks</code>，不是 LLM 自己打的浮点分。图中的数字是 <code>e2e-opus-final</code> 真实跑出来的，不是 mock。</sub>
 
-  ![iter-0 → iter-1 reconcile loop, 真实 score drift](https://raw.githubusercontent.com/HomenShum/open-codesign/feat/decompose-to-ui-kit/website/public/demos/decompose-iter-reel.gif)
-  <sub>来自 <code>e2e-nodebench-iter</code> 的 4 帧 reel: source → iter-0（parityScore 0.82, 6 个 gap）→ iter-1（parityScore 0.78, 5 个 gap）→ honest verdict。Agent 修了一些 gap 但又 introduced 新的 layout drift，boolean rubric 把 regression 直接 surface 出来不藏。<a href="https://raw.githubusercontent.com/HomenShum/open-codesign/feat/decompose-to-ui-kit/website/public/demos/decompose-iter-reel.mp4">MP4 版本</a>。</sub>
+  ![iter-0 → iter-1 reconcile loop, 真实 score drift](https://raw.githubusercontent.com/OpenCoworkAI/open-codesign/main/website/public/demos/decompose-iter-reel.gif)
+  <sub>来自 <code>e2e-nodebench-iter</code> 的 4 帧 reel: source → iter-0（parityScore 0.82, 6 个 gap）→ iter-1（parityScore 0.78, 5 个 gap）→ honest verdict。Agent 修了一些 gap 但又 introduced 新的 layout drift，boolean rubric 把 regression 直接 surface 出来不藏。<a href="https://raw.githubusercontent.com/OpenCoworkAI/open-codesign/main/website/public/demos/decompose-iter-reel.mp4">MP4 版本</a>。</sub>
 - **支持中途取消生成**：停止后也不会丢失之前的上下文和结果
 
 ### 预览与工作流

README.md

@@ -151,9 +151,14 @@ function extractFirstJsonObject(text: string): string {
 }
 
 function dataUrlToBase64(dataUrl: string): string {
-  const idx = dataUrl.indexOf('base64,');
-  if (idx === -1) throw new Error('Image must be a base64 data URL');
-  return dataUrl.slice(idx + 'base64,'.length);
+  // RFC 2397 says the `;base64` token is case-insensitive in practice (browsers
+  // accept both casings), so match either. Picks up the matched segment to
+  // compute the slice offset correctly.
+  const match = dataUrl.match(/;base64,/i);
+  if (!match || match.index === undefined) {
+    throw new Error('Image must be a base64 data URL');
+  }
+  return dataUrl.slice(match.index + match[0].length);
 }
 
 export interface VisionPromptInput {