@qijenchen/design-system 0.1.0-beta.13 → 0.1.0-beta.131

package/ds-canonical/skills/deep-audit-cross-codex/references/phase-b-codex-brief.md

@@ -1,165 +1,51 @@
-# Phase B — Codex parallel audit + 比稿辯論
+# Phase B — independent peer audit and reconciliation
 
-## B.0 Transport discovery(per codex-collab/SKILL.md Step 0.4)
+## 1. Build the peer packet
 
-3-test 順序固定:
-```bash
-ls -la node_modules/.bin/codex && node_modules/.bin/codex --version   # 1 Local CLI(canonical)
-which codex 2>/dev/null && codex --version 2>/dev/null                 # 2 Global(罕見)
-ls -la ~/.codex/auth.json                                              # 3 Auth(sanity)
-```
-
-**Decision**:1 ✅ → local CLI / 1 ❌ + 2 ✅ → global / 1+2 ❌ + 3 ✅ → `npm install` / 全 ❌ → 報 user。**禁 Explore agent 替身**(M31)。
-
-## B.1 Brief template
-
-⚠️ **Faithful relay invariant**(M31 Step 0.05,user-verbatim,不可只送 paraphrase):
-
-```
-## User 原話(verbatim)
-「<user 在本 session 對 deep audit 的原話 quote,中英 / 符號 / 圖文 ref 全保>」
-
-## Claude Phase A 結果摘要
-
-### 全盤閱讀
-- 讀完:CLAUDE.md / 31 active M-rules / spec.md ×<N> / tokens / patterns / memory active
-
-### 全 dim NO-SAMPLE deep audit findings
-- P0: <列具體 file:line + 違反 spec / rule>
-- P1: <同>
-- P2: <同>
-
-### SSOT-UI/UX 已 ASK user
-- <列 N 項>
-
-### Autonomous landed
-- <列 N 項 file:line + commit hash>
+Use the frozen Phase A inventory, rubric, user wording, repository identity, and primary evidence without including Phase A conclusions. The packet must include digests so the response can prove input equality.
 
-### Phase A 結論(不 verify 但抓到)
-- <列 grep 看到 / suspect drift / 未跑 visual>
+Require the peer to:
 
-## 你的任務(Phase B,獨立)
+- identify provider, model, context, and transport;
+- run in a separate read-only context;
+- cover the full inventory and every applicable rubric dimension without sampling;
+- verify claims against source and actual runtime artifacts;
+- report complete per-dimension coverage, including zero-finding evidence;
+- return findings, counterexamples, uncertainty, and literal verification results.
 
-1. **全盤閱讀**(per A.0 file list,不 sample):
-   - CLAUDE.md / `.claude/rules/*` / `.claude/references/*`
-   - `packages/design-system/src/**/*.spec.md` 全部(~60 file)
-   - `packages/design-system/src/tokens/**/*.spec.md` + `packages/design-system/src/patterns/**/*.spec.md`
-   - `~/.claude/.../memory/MEMORY.md` index + active project memory
+Use the transport contract in the cross-provider collaboration workflow. An unavailable or unverified transport produces `REVIEW-BLOCKED`.
 
-2. **全 dim deep audit NO-SAMPLE**(對齊 `.claude/skills/design-system-audit/SKILL.md` Group A-P 全 dim per SKILL.md):
-   - 每 dim 全 DS-wide 掃,禁 sample top N
-   - 每 finding 必 cite: <file:line> + <quote> + <違反 spec / rule>
+## 2. Validate the response
 
-3. **整理完整報告**(P0 / P1 / P2 分類):
-   ```
-   ### Codex Phase B audit
-   - P0: <list>
-   - P1: <list>
-   - P2: <list>
+Before considering conclusions, verify:
 
-   ### 跟 Claude Phase A 對照
-   - Claude 抓 + 你不同意:<list + cite>
-   - 你抓 + Claude 漏:<list + cite>
-   - 兩邊都漏的盲區:<list + cite>
+1. author, primary, and peer provider identities are explicit; the author equals the bound primary and the peer differs from both;
+2. inventory and rubric digests match;
+3. context isolation and read-only enforcement are evidenced;
+4. worktree fingerprints are unchanged;
+5. every required dimension and inventory partition is accounted for;
+6. every material claim has primary evidence.
 
-   ### 真理由 disagreement(必 cite battle)
-   - <題 1>: Claude 立場 / 你立場 / spec.md path:line 引文 / world-class ≥3 家對照
-   ```
+Reject partial, mutated, same-provider, framed, or identity-free responses.
 
-4. **禁**:
-   - frame 答案進 Claude 思路(不可 paraphrase Claude 結論)
-   - sample / heavy agent skip(NO-SAMPLE STRICT)
-   - pass-through 共識（必獨立 own report）
-   - 直接給 user verdict（共識由 Claude Phase B.5 synthesize）
+## 3. Reconcile independently
 
-## 請獨立解讀 user 原話(per M31 Step 0.05)
+For each primary and peer finding:
 
-User 原話可能跟我 paraphrase 不同。請以「你直接看 user 原話該怎麼解讀」為主,
-不要被 Claude 結論 frame。如果 user 想要的方向你判斷跟我不同,直接寫出來。
-
-## Cite invariant(M22)
-
-每 claim 必附 inline cite:URL / GitHub source path + line / screenshot ref。
-無 cite = unverified rumor,自動降 P2 「pending verification」。
-```
-
-## B.2 Step 4 self-check 必過(M22/M23/M27→M23(c)/M8)
-
-對 codex reply 跑 4 題:
-
-| # | Check | 失敗 → |
-|---|---|---|
-| 1 | M22 cite check — codex 引的 benchmark 有 inline URL / GitHub line? | reply 要求補 cite |
-| 2 | M23 DS-first — codex 建議是否覆蓋 DS 既有 canonical(spec.md / token / variant)? | 必 grep verify DS 既有,有命中 → 拒絕 codex / 修正 |
-| 3 | M23(c) prop namespace — codex prop name 是否撞 DS 既有? | wrap-and-rename |
-| 4 | M8 ≥3 source — codex 只引 1 家? | reply 要求補到 3 家 |
-
-## B.3 Step 4.5 Verify each claim(anti-pass-through)
-
-逐條 codex finding 過:
-
-```
-Finding N: <codex claim 一句話>
+1. locate the governing canonical owner;
+2. reproduce the claimed source/runtime fact;
+3. scan counterexamples across the full relevant inventory;
+4. classify it as verified, false, partial, or evidence-blocked;
+5. preserve disagreements rather than resolving them by vote.
 
-Verify:
-1. Grep DS-internal: rg "<keyword>" src/ → <hit / miss + 行號>
-2. WebFetch external(若 cite world-class): URL → <真 quote 對照>
-3. Run invariant script(若 codex 給數字): bash <script> → <實 value vs codex value>
-4. Counter-example scan: grep <pattern> DS-wide → <反例 / 無>
+Use this comparison record:
 
-Verdict: ✅ verified / ❌ FALSE / ⚠️ partial
-Reasoning: <一句話>
-```
-
-## B.4 Cite battle template
-
-```
-Disagreement N: <topic>
-
-| Side | Position | Cite | Quote |
-|---|---|---|---|
-| Claude Phase A | <stance> | <file:line / URL> | 「<quote>」 |
-| Codex Phase B | <stance> | <file:line / URL> | 「<quote>」 |
-
-WebFetch external:
-- Polaris: <URL> → 「<quote>」
-- Material: <URL> → 「<quote>」
-- Atlassian: <URL> → 「<quote>」
-
-Verdict: <Claude / Codex / Both / STOP-ask-user>
-Reasoning: <cite + world-class consensus + DS internal canonical alignment>
-```
-
-## B.5 Step 5 比稿 matrix(per finding)
-
-```
-| Finding | Claude v1 | Codex v1 | Verdict | Final synthesized |
+```markdown
+| Decision axis | Primary claim + evidence | Peer claim + evidence | Independent verification | Final state |
 |---|---|---|---|---|
-| <N> | <stance> | <stance> | ✅ accept C / ✅ accept C+ / ✅ accept Claude / 修正 = synthesize / 重啟 | <final 方案> |
+| <axis> | <claim> | <claim> | <artifact> | accept-primary / accept-peer / synthesize / blocked |
 ```
 
-**禁** pass-through 直接 paste codex / 「兩邊都對」打太極 / vote without cite。
-
-## Phase B complete output
-
-```markdown
-# Phase B 完成
-
-## Codex independent findings(NO-SAMPLE)
-- P0: <N> / P1: <M> / P2: <K>
-
-## 比稿結果
-- Claude 抓 + Codex 漏: <list>
-- Codex 抓 + Claude 漏: <list>(這些是 Phase A 漏掉的盲區!)
-- Cite battle resolved: <K 題,各題 verdict>
-- Cite battle deferred(evidence 對等): <N 題,STOP 等 user>
-
-## 共識 SSOT-UI/UX(中文人話 propose)
-<決策 1-N per triage-rubric.md format>
+## 4. Complete the audit
 
-## 共識 autonomous landed(commit <hash>)
-- <N> 項
-- file:line + diff link
-
-→ 進 Phase C
-```
+Return author, primary, and peer provider identities, matching input digests, both coverage ledgers, verified overlaps, findings unique to each pass, rejected claims, unresolved disagreements, authorization decisions, final verification artifacts, and residual risk. Never describe a single-provider, author/peer collision, or incomplete run as independent review.

package/ds-canonical/skills/deep-audit-cross-codex/references/skill-relationships-antipatterns-benchmarks.md

@@ -0,0 +1,28 @@
+# Relationships, anti-patterns, and review anchors
+
+## Workflow composition
+
+- Use the design-system audit skill as the sole owner of audit dimensions and detailed rubrics.
+- Use the cross-provider collaboration workflow as the sole owner of independent-peer transport and brief equality.
+- Use the proposal workflow for genuine product／UI／UX SSOT decisions retained by the user.
+- Use the canonical-placement workflow for approved propagation and drift prevention.
+- Use the governance-pruning workflow when the audit discovers duplicate, stale, or over-concrete rules.
+
+This workflow composes those owners; it does not copy their mutable rules or counts.
+
+## Permanent anti-patterns
+
+- skipping canonical inputs because they appear unchanged;
+- sampling, representative subsets, or unrecorded omissions;
+- showing primary conclusions to the peer before its independent pass;
+- treating a same-provider worker or renamed adapter as an independent peer;
+- accepting a response without matching inventory/rubric digests;
+- allowing reviewer mutation or failing to compare worktree fingerprints;
+- accepting claims without reproducing their evidence;
+- resolving disagreement by vote, authority label, or vague consensus;
+- running repeated generative audits after verified material findings reach zero;
+- claiming completion while a required gate is unavailable or unproven.
+
+## External review anchors
+
+Use current primary sources when comparing this protocol with mature review systems. Relevant patterns include independent academic peer review, maintainers plus public patch review, adversarial design review, separation of author and approver, immutable build provenance, and reproducible verification. These are comparison anchors, not substitute evidence for repository policy.

package/ds-canonical/skills/deep-audit-cross-codex/references/triage-rubric.md

@@ -1,91 +1,26 @@
-# Triage rubric — SSOT-UI/UX vs Non-SSOT scope classifier
+# Triage and convergence rubric
 
-## Scope classifier(critical 第 1 步)
+## Human-owned product decisions
 
-### SSOT-UI/UX substantive(必 ASK user)
+Stop and request a decision only when a proposed change creates a genuine unresolved product／UI／UX SSOT choice, including user-visible behavior, information architecture, interaction model, UI pattern, component or copy semantics, visual language, or another user-visible trade-off. Explain the evidence, viable options, trade-offs, recommendation, and downstream propagation in plain language.
 
-任一命中 → **STOP propose 等 user A/B**:
+## Standing-delegated engineering decisions
 
-- 動 `packages/design-system/src/components/<X>/<X>.tsx` 的 cva variant / size / state(視覺 surface)
-- 動 spec.md 的「視覺結構 / canonical / SSOT / 何時用 / 何時不用 / 邊界案例 / Layout Family」段落 substantive meaning
-- 加 / 刪 / 改 token(`tokens/**/*.css` / `*.spec.md`)
-- 加新 prop / 重命名 prop / 改 prop default value(API contract)
-- 跨元件 design language 改動(同類 row primitive 重新訂 padding / dismiss canonical 跨元件改)
-- 新 family 宣告 / 改 family 歸屬
-- 新 pattern 加入 `patterns/` 或既有 pattern scope 改變
+Architecture, canonical governance policy implementation, ownership wiring, release trust, security boundaries, provider capabilities, adapters, CI/CD, migration, release, rollout, rollback, and failure remediation are engineering decisions. Resolve them autonomously under the canonical Standing Authorization using evidence, independent review, hard gates, least privilege, readback, and recovery. Examples also include implementation drift back to an existing contract, broken pointers, deterministic generated-view regeneration, tests that enforce an unchanged rule, dead code, typos, or behavior-preserving refactors. Verification remains mandatory.
 
-### Non-SSOT(AUTO 整批做完,M33 anti-defer)
+## Finding materiality
 
-- Bug fix(spec 既有 canonical + code 跑掉 → 對齊 spec)
-- Code clean(unused import / dead export / typo / 排版)
-- Refactor 不動行為(extract helper / rename internal symbol / consolidate condition)
-- 命名一致對齊既有 canonical(per `# 命名與語言一致性`)
-- Test / audit / verify(playwright / invariant script / hook regex 加廣)
-- Audit dim / hook 加廣(對齊 spec wording 廣度,M34)
-- Pointer / cross-link 補完整(reciprocal Dim 3 / SSOT pointer)
-- Spec typo / markdown layout / 標點
-- Memory file rotate(已 absorb → demote)
-- Governance / hook / skill 內部 refactor(無 BLOCKER 邏輯改)
+Adversarially reproduce every raw finding before action:
 
-## SSOT-UI/UX propose 中文人話 format(per `feedback_propose_in_plain_chinese.md`)
-
-```
-### 決策 N:<一句話標題,zero jargon>
-
-**現況**:
-- <目前 spec / code 行為,具體一句話人話>
-- 證據:<file:line>
-
-**影響**:
-- 不改:<具體後果>
-- 改了:<具體後果>
-
-**選項**:
-- A. <做法 1>
-  - 後果:<好處 / 代價>
-  - 涉及檔案:<count + file 列表>
-- B. <做法 2>
-  - 後果:<好處 / 代價>
-  - 涉及檔案:<count + file 列表>
-- C. 不動
-  - 後果:<漂移持續累積 / 或維持現況的成本>
-
-**我推**:<A / B / C> 因 <理由>
-```
-
-### 4-Q gate(propose 前必 inline 跑,失敗 → 不列 option)
-
-| Q | Check | 失敗 → |
+| Class | Meaning | Action |
 |---|---|---|
-| Q1 M22 cite | 3-column owner table(spec path:line / canonical sentence / conflicting code)| 撤回 propose,先補 anchor |
-| Q2 M17 SSOT | 既有 token / primitive / pattern 列消費清單 | 撤回,先 consume 既有 |
-| Q3 Rule-of-3 | 同概念 ≥ 3 處 → 選 SSOT 其他 pointer | 撤回,選 SSOT home |
-| Q4 M10 下游 | 修上游 ≥ 3 處下游 redundant 可清 | 提案加上「下游 N 處同步 retire」 |
-
-### 禁用 jargon list(propose 內 zero tolerance)
-
-❌ `L1-L7` / `canonical` / `primitive` / `SSOT` / `consume` / `traits` / `M-rule` / `cva` / `tier` / `tokens` / `wrapper`
-✅ 翻成:`主檔` / `設計原則` / `共用零件` / `通用基底` / `共用設計來源` / `沿用` / `行為類型` / `規則 N` / `樣式拼裝表` / `層級` / `設計變數` / `外殼`
-
-## Non-SSOT autonomous 7 軸 simultaneous optimize
-
-每次動 code 前 mental check:
-
-1. **言簡意賅** — comment / spec 文字短而精;禁長 docstring;禁不必要 comment(per CLAUDE.md `# Tone and style`)
-2. **效率 + 效能** — 避 unnecessary re-render / memo gap / O(N²) algorithm / context thrashing
-3. **SSOT 鐵律** — 動 visual decision 前 grep DS 既有 token / variant / pattern(M23)/ spec.md owner(M29)/ wrapper extends primitive(M30)/ token 而非 hardcode(M17)
-4. **易懂 + 維護 + 擴充** — file size budget(governance ≤ 200/300/500/800)/ function ≤ 80 / naming 一致 / public API 穩定
-5. **世界級 + 一致設計語言** — mindset #1(Polaris / Material / Atlassian / Ant / Carbon / Apple HIG ≥ 3 cite)+ M22 inline cite + M26 propose 前 WebFetch ≥ 3 source
-6. **完整 self-verify** — M20 best-practice score ≥ 80 / M31 5-step(Layer A own + Layer C codex 比稿)/ M32 pixel-quantified audit
-7. **自動 self-improve** — M14 5-layer pipeline(對話結論 → spec / hook / SKILL / CLAUDE.md / memory 該動的全動)+ M20 stop hook auto-score
-
-任一軸明顯 degrade → 撤回方案重來,**禁** 「先這樣 / 之後優化」defer keyword(M33 BLOCKER)。
-
-## 共識決策(Phase B.5)同 format
+| material | user, contract, accessibility, security, release, or verified regression impact | route genuine product／UI／UX choices to the user; remediate engineering findings under Standing Authorization |
+| marginal | preference or wording with no material effect | record without churn |
+| false positive | source or an applicable exception disproves the claim | reject with evidence |
+| evidence-blocked | required source/runtime evidence is unavailable | retain blocker; never call PASS |
 
-Phase B 共識(Claude + Codex 雙 verify PASS)走相同 propose / autonomous 分流:
+## Convergence
 
-- SSOT-UI/UX 共識 → 中文人話 propose user 拍板
-- Non-SSOT 共識 → autonomous batch(7 軸 optimize)
+Stop iterative generative review when a complete pass, followed by adversarial verification, yields zero material findings and zero regressions. Do not chase stylistic zero and do not stop while inventory, rubric, identity, isolation, or deterministic verification is incomplete.
 
-**禁** Phase B 共識降回 Phase A 已 ASK 的同 decision 重 ASK(浪費 user 時間)。已 ASK 的 → 等 user 拍板;新發現(Phase B 抓的 Phase A 漏)→ 新 propose。
+The final report must include input digests, per-dimension coverage for both providers, verified findings, rejected findings, unresolved decisions, corrections made, exact verification artifacts, and residual risk.

package/ds-canonical/skills/deep-audit-cross-codex/references/upstream-directives-r-mapping.md

@@ -0,0 +1,34 @@
+# Upstream directives and SSOT mapping
+
+## Integrity contract
+
+The design-system audit skill owns the dimension inventory and detailed rubric. This workflow always consumes the complete current inventory and must not duplicate a mutable dimension count. A change to the audit rubric is made at its owner and inherited here through a path and digest.
+
+The shared governance instruction, provider registry, skill-semantics registry, canonical rules, and release/fleet contracts remain their own owners. Generated provider views are projections, never alternate policy sources.
+
+## Permanent objectives
+
+The workflow must preserve these outcomes across local, cloud, source-repository, template, and consumer environments:
+
+- complete rather than sampled review;
+- equal evidence and judgment criteria for both providers;
+- a genuinely independent second pass;
+- explicit SSOT ownership and deterministic propagation;
+- plain-language user decisions only for genuine product/UI/UX SSOT choices or tradeoffs;
+- autonomous completion of engineering/governance policy, architecture, remediation, release,
+  rollout, and other standing-delegated work whether substantive or not;
+- reproducible verification, clean maintainable code, and no fabricated PASS;
+- provider-specific mechanisms confined to registered adapter bindings;
+- fail-closed onboarding for any future provider until compatibility is evidenced.
+
+## Evidence mapping
+
+For every objective, the final report identifies:
+
+1. canonical source path and digest;
+2. runtime/provider binding used;
+3. coverage or verification artifact;
+4. downstream generated views and consumers checked;
+5. unresolved blocker or explicit incompatibility reason.
+
+This mapping replaces historical provider-specific instructions and prevents the workflow from forking when a tool, model, transport, or repository surface changes.

package/ds-canonical/skills/delivery-handoff/SKILL.md

@@ -60,6 +60,11 @@ Purpose: 產品 / feature 通過 prototype → audit → stakeholder 決策後
 
 ### Phase 1 — Inventory 生成
 
+**Deterministic 源優先(2026-07-07 治理進化收尾:本 skill 原「幾乎零機械依賴」列冊補強)**:
+- Component / story 母集 = `node scripts/gen-ds-story-manifest.mjs` 產出的 manifest(禁憑記憶列)
+- Token 使用 = `rg -o -- '--[a-z][a-z0-9-]*' <feature files> | sort | uniq -c`(機械計數,非印象;字元類含 0-9 — DS token 名 46% 含數字如 `--chart-1`/`--black-a02`,缺數字會截斷合併成假 key)
+- 消費統計交叉驗 `scripts/audit-orphan-tokens.mjs` 的消費通道邏輯(同一套 5 通道定義)
+
 scan 目標 feature 的 code,自動 inventory:
 
 1. **Component 清單**(哪些 DS 元件被此 feature 消費):
@@ -108,7 +113,9 @@ flowchart TD
   Confirm -->|取消| Review
 ```
 
-AI 根據 code(route / onClick / navigation)auto-generate,user review 後微調。
+AI 根據 code(route / onClick / navigation)auto-generate，並以 source/readback 驗證。
+若 flow 可由既有 requirement 與實作唯一重建，直接修正 evidence drift；只有仍存在會改變
+產品／UI／UX SSOT 的真實 flow 取捨才 P2H。
 
 ### Phase 3 — Per-screen spec sheet
 
@@ -166,9 +173,9 @@ Storybook title: `Features/{FeatureName}/Handoff`
 
 此 page 就是「Figma-like 的 inspectable 交付」— stakeholder 打開就看到全景。
 
-### ⚠️ Checkpoint — User review handoff package
+### Checkpoint — Handoff evidence/readback receipt
 
-Phase 4 產出後 pause,讓 user review:
+Phase 4 產出後記錄並驗證:
 
 ```
 📦 Handoff Package Ready
@@ -186,13 +193,17 @@ Export formats:
 - Storybook URL: {link}
 - Markdown(可 export PDF): {path}
 
-請 review:
-(a) 通過,準備 share 給 {audience}
-(b) 需要修改: ...
-(c) 還缺 ... 補齊
-(d) Audience 換人,調整語氣 / 深度
+Readback:
+- requirement／source coverage: {complete / findings}
+- links／screens／states／a11y evidence: {verified / findings}
+- audience binding: {source / exact requirement}
 ```
 
+可由 requirement、source、tests 與既有 audience 唯一判定的 finding 由 agent 自主修正並
+繼續。只有 handoff 揭露新的產品／UI／UX outcomes 取捨，或對外分享會構成無法從既有
+requirement 推導的 business commitment，才停在 shared governance 定義的 human-only
+boundary；handoff milestone 本身不是 user engineering-approval gate。
+
 ---
 
 ## 世界級 handoff 標準(reference)