cc-devflow 4.5.2 → 4.5.3

package/.claude/skills/cc-act/CHANGELOG.md

@@ -1,5 +1,18 @@
 # CC-Act Skill Changelog
 
+## v1.8.0 - 2026-04-28
+
+- add remote state consistency rules for issue, PR, tracker, `needs-info`, and `ready-for-agent` closeout handoffs
+- require dangerous git actions to have explicit ship semantics and user confirmation before reset, clean, branch deletion, or whole-tree restore
+- require tooling changes to prove the real hook, adapt, publish, lint, or verification entrypoint still works
+
+## v1.7.0 - 2026-04-28
+
+- carry cc-check QA feedback-loop and behavior evidence into PR briefs, handoffs, and release notes so ship materials preserve expected/actual/reproduction truth
+- add durable follow-up brief rules for current behavior, desired behavior, key interfaces, acceptance criteria, and out-of-scope boundaries
+- teach shared follow-up and evidence extraction to include report-card followup briefs and QA architecture follow-ups
+- extend readiness dashboard guidance with feedback-loop quality and behavior evidence checks
+
 ## v1.6.4 - 2026-04-28
 
 - add a readiness dashboard covering review freshness, review quality, specialist facets, QA coverage, browser QA, failure ownership, documentation release, and PR body accuracy

package/.claude/skills/cc-act/PLAYBOOK.md

@@ -26,6 +26,7 @@
 2. 确认 `review/report-card.json` 是 `pass`，且没有未解释的 gaps / reroute
 3. 确认 `planning/tasks.md` 不再有未完成项
 4. 确认 `review.freshness` 新鲜、`runtime.failureOwnership` 无未解释失败、`qa.coverageAudit` / `qa.browserEvidence` 有证据或明确 skip
+5. 确认 `qa.feedbackLoop` / `qa.behaviorEvidence` 能支撑行为结论；不可复现时必须写清缺什么 artifact / 权限 / 输入
 
 如果 gate 没闭合，直接回 `cc-check` 或 `cc-do`，不要在 `cc-act` 自我安慰。
 
@@ -82,7 +83,7 @@ Ship 必须属于这 4 种模式之一：
 3. 检查提交边界，按逻辑单元拆分，保证提交顺序不引用未来代码。
 4. 如果有 WIP commit，只能用非破坏性 rebase / fixup 处理，不允许盲目 soft reset。
 5. push 前比较 local / remote HEAD；PR 前检查是否已有打开 PR / MR。
-6. 生成 readiness dashboard：review freshness、review quality、QA coverage、browser QA、failure ownership、documentation release、PR body accuracy。
+6. 生成 readiness dashboard：review freshness、review quality、QA coverage、browser QA、feedback loop、behavior evidence、failure ownership、documentation release、PR body accuracy。
 
 ## Phase 3: Build Delivery Pack
 
@@ -114,8 +115,10 @@ Ship 必须属于这 4 种模式之一：
 - review packet path / summary
 - finding triage summary
 - QA / claim evidence summary
+- QA behavior evidence and feedback-loop quality
 - readiness dashboard
 - PR body accuracy check
+- Durable follow-up briefs: current behavior、desired behavior、key interfaces、acceptance criteria、out of scope
 
 缺这些字段时，可以生成 local handoff，但不能声称 PR body 已经可 review。
 
@@ -183,6 +186,8 @@ Ship 必须属于这 4 种模式之一：
 - 长期方向写 `devflow/roadmap/roadmap.md`
 - 下一轮待排队动作写 `devflow/roadmap/backlog.md`
 - 不要把噪音和碎念回写成系统真相
+- follow-up 必须是 durable brief：用领域语言写 current behavior、desired behavior、key interfaces、acceptance criteria、out of scope
+- 独立行为拆独立条目；有依赖关系时写明顺序，方便下一轮并行或排队
 
 ## Phase 7: Declare The Next Entry
 
@@ -201,6 +206,7 @@ Ship 必须属于这 4 种模式之一：
 4. `cc-simplify`、单测、e2e、commit/push 的结果是不是都能追溯？
 5. PR body / release note / handoff / changelog 说的是不是同一套现实？
 6. readiness dashboard 有没有 blocker 或 stale warning？
+7. follow-up 是不是行为契约，而不是“改某文件某行”的易腐烂 TODO？
 
 如果第 1 或第 3 题答案不是“能”，说明 `cc-act` 仍然太重或太糊。
 

package/.claude/skills/cc-act/SKILL.md

@@ -1,6 +1,6 @@
 ---
 name: cc-act
-version: 1.6.4
+version: 1.8.0
 description: 'Use when verified work must be shipped or handed off with a clear landing path: run simplify and required tests, create or update a PR, prepare a local handoff, close out merged work, sync docs, write release notes, and fold follow-ups back into backlog or roadmap.'
 triggers:
   - 准备提 PR
@@ -127,7 +127,8 @@ tool_budget:
 3. 运行 `scripts/verify-act-gate.sh --dir <requirement-dir>`，确认 gate 真的闭合。
 4. 运行 `scripts/detect-ship-target.sh`，识别当前分支、base branch、PR 状态与推荐 ship 路径。
 5. 检查 `review.freshness`、`runtime.failureOwnership`、`qa.coverageAudit`、`qa.browserEvidence`，确认 readiness dashboard 没有 blocker。
-6. 如果在 `cc-act` 期间因为 `cc-simplify`、单测、e2e、review 修复而改了代码，必须回 `cc-check`，不能带着旧证明继续 ship。
+6. 检查 `qa.feedbackLoop`、`qa.behaviorEvidence`、`qa.architectureFollowUps` 和 follow-up brief，确认交付材料继承的是行为证据，不是聊天记忆或易腐烂 TODO。
+7. 如果在 `cc-act` 期间因为 `cc-simplify`、单测、e2e、review 修复而改了代码，必须回 `cc-check`，不能带着旧证明继续 ship。
 
 ## Ship Modes
 
@@ -213,6 +214,9 @@ tool_budget:
 7. PR idempotency：已有打开的 PR / MR 只更新 body，不重复创建。
 8. Review range：PR brief / PR body 必须写清 `cc-check` 审过的 base/head SHA、review packet、finding triage 摘要。
 9. Post-integration verification：本地合并或 post-merge closeout 后，必须在 merged result 上跑必要 gate；不能只继承合并前绿色。
+10. Follow-up durability：PR brief / release note / backlog writeback 里的 follow-up 必须写成行为契约，包含 current behavior、desired behavior、key interfaces、acceptance criteria、out of scope；不要把当前文件路径或行号当成长期计划。
+11. Remote state consistency：如果本次 closeout 触碰 GitHub issue / PR / tracker，必须记录当前 state、目标 state、允许转换、已保留事实和下一位 owner；`needs-info` 必须保留已确认事实和具体问题，`ready-for-agent` 必须有可执行 brief。
+12. Tooling smoke：如果本次改动影响 hook、pre-commit、lint、publish、adapt 或验证脚本，必须跑真实入口或最接近真实入口的 smoke；只读配置文件不等于工具链可用。
 
 ## Readiness Dashboard
 
@@ -222,9 +226,11 @@ PR / handoff 之前必须把 readiness 压成一屏事实：
 2. Review quality：记录 `review.qualityScore`、specialist facet 覆盖、finding triage 摘要。
 3. QA coverage：记录 `qa.coverageAudit` 的 coverage、gaps、e2e/eval requirement。
 4. Browser QA：UI / 用户路径变更必须有 `qa.browserEvidence`，否则要有 skip reason。
-5. Failure ownership：`runtime.failureOwnership` 不能有未解释的 `in-branch` 或 `ambiguous` failure。
-6. Documentation release：README / CLAUDE / architecture / handoff / changelog 的同步状态必须可审计。
-7. PR body accuracy：PR body 必须从当前 `pr-brief.md`、当前 diff、当前 report-card 重建；已有 PR body 只能被刷新，不能被继承。
+5. Feedback loop：bugfix / 行为变更必须有 `qa.feedbackLoop`，否则要有不可复现或不适用原因。
+6. Behavior evidence：expected / actual / reproduction steps 必须能被 reviewer 独立理解。
+7. Failure ownership：`runtime.failureOwnership` 不能有未解释的 `in-branch` 或 `ambiguous` failure。
+8. Documentation release：README / CLAUDE / architecture / handoff / changelog 的同步状态必须可审计。
+9. PR body accuracy：PR body 必须从当前 `pr-brief.md`、当前 diff、当前 report-card 重建；已有 PR body 只能被刷新，不能被继承。
 
 readiness dashboard 有 blocker 时，不能创建或更新 PR，只能 reroute 到 `cc-check` / `cc-do` 或生成 local handoff。
 
@@ -236,6 +242,7 @@ readiness dashboard 有 blocker 时，不能创建或更新 PR，只能 reroute
 2. 丢弃未合并工作必须要求用户显式确认；没有确认时只能转 `local-handoff`。
 3. branch cleanup 只发生在 merge / PR / discard 语义已经清楚之后。
 4. `post-merge-closeout` 必须记录 merged-result verification：命令、exit status、关键观察、失败时 reroute。
+5. 危险 Git 动作必须有明确 ship 语义：`git push` 只在 `create-pr` / `update-pr` / release lane 执行，`git reset --hard`、`git clean -f`、`git branch -D`、整树 restore/checkout 必须先列出对象并取得用户确认；没有确认时写 handoff，不执行。
 
 ## Documentation Release
 
@@ -249,6 +256,16 @@ readiness dashboard 有 blocker 时，不能创建或更新 PR，只能 reroute
 6. Discoverability：新增文档必须能从 README、CLAUDE 或 handoff 入口找到。
 7. TODO/backlog cleanup：只把有 diff 证据完成的事项移到 completed；新 follow-up 写回 backlog/roadmap。
 
+## Issue And Follow-Up Handoff
+
+当本次交付产生或更新 issue / PR / backlog follow-up 时，文案必须耐 refactor：
+
+- 先写用户可观察行为：what happened / expected / reproduction 或 desired behavior。
+- 再写 agent 可执行 contract：acceptance criteria、allowed scope、blocked by、verification command。
+- 不把短期文件路径、行号、函数名写成长期事实；只有当前 reviewer 需要定位时才放进 PR brief 的实现备注。
+- 多个 follow-up 必须按独立可验证切片拆开，并标明 `AFK` / `HITL`、blocked-by 和 owner。
+- 已经问过或确认过的信息不能在 `needs-info` 里丢失；问题必须具体到 reporter 能直接补证据。
+
 ## Loop
 
 1. 先锁定 ship 事实：当前分支、base branch、PR 状态、requirement 状态。

package/.claude/skills/cc-act/assets/PR_BRIEF_TEMPLATE.md

@@ -39,6 +39,8 @@
 - Specialist review facets:
 - QA coverage:
 - Browser QA:
+- Feedback loop:
+- Behavior evidence:
 - Failure ownership:
 - Documentation release:
 - PR body accuracy:
@@ -57,6 +59,14 @@
 - Fresh evidence:
 - Merged-result verification:
 
+## QA Behavior Evidence
+
+- Feedback loop:
+- Expected behavior:
+- Actual behavior:
+- Reproduction steps:
+- Consistency:
+
 ## Documentation Sync
 
 - `CLAUDE.md`:
@@ -81,7 +91,11 @@
 
 ## Follow-Ups
 
-- 
+- Current behavior:
+- Desired behavior:
+- Key interfaces:
+- Acceptance criteria:
+- Out of scope:
 
 ## Risks
 

package/.claude/skills/cc-act/assets/RELEASE_NOTE_TEMPLATE.md

@@ -20,10 +20,19 @@
 
 - 
 
+## QA Behavior Evidence
+
+- Feedback loop:
+- User-visible behavior:
+- Reproduction / verification path:
+
 ## Ops Notes
 
 - 
 
 ## Follow-Ups
 
-- 
+- Current behavior:
+- Desired behavior:
+- Acceptance criteria:
+- Out of scope:

package/.claude/skills/cc-act/references/closure-contract.md

@@ -10,6 +10,7 @@
 6. 如果文件结构变了，就同步对应目录的 `CLAUDE.md`
 7. PR / handoff 必须记录 `cc-check` 审过的 base/head SHA、review packet、finding triage 摘要
 8. readiness dashboard 必须说明 review freshness、QA coverage、browser evidence、failure ownership、documentation release、PR body accuracy
+9. behavior handoff 必须带上 QA feedback loop、expected / actual / reproduction steps，以及 durable follow-up brief
 
 ## Ship Decision Contract
 
@@ -35,6 +36,7 @@
 9. 删除 branch、worktree、未合并提交、归档 requirement 前必须列出对象；丢弃未合并工作需要显式确认
 10. verification 每次进入 `cc-act` 都必须重新跑；只有 push、PR 更新、文档生成等动作可以因为幂等状态跳过
 11. PR body accuracy 必须对照当前 report-card、当前 diff、当前 commits；旧 body 不能作为证据源
+12. follow-up 回写必须用行为契约表达，包含 current behavior、desired behavior、key interfaces、acceptance criteria、out of scope；不能只写文件路径或聊天 TODO
 
 ## Memory Consolidation
 
@@ -52,6 +54,7 @@
 - maintainer 知道怎么验证
 - PR / handoff / release 材料反映同一套事实
 - readiness dashboard 没有 blocker，PR body accuracy 已检查或明确阻塞
+- QA behavior evidence 和 feedback loop 已进入 PR / handoff / release 材料
 - post-merge closeout 反映 merged result 的验证事实，而不是只反映合并前事实
 - 下一轮计划入口更清楚
 - 文档入口可发现，changelog 不丢历史，TODO / backlog 只记录有证据的事项

package/.claude/skills/cc-act/scripts/cc-act-common.sh

@@ -189,6 +189,30 @@ req_act_collect_followups() {
 
   if [[ -f "$report_card" ]]; then
     jq -r '(.gaps // [])[]?' "$report_card" 2>/dev/null | sed '/^$/d' > "$out_file" || true
+    jq -r '
+      (.followupBriefs // .review.followupBriefs // [])[]?
+      | [
+          ("summary: " + (.summary // .title // "follow-up")),
+          ("current: " + (.currentBehavior // "not recorded")),
+          ("desired: " + (.desiredBehavior // "not recorded")),
+          ("interfaces: " + ((.keyInterfaces // []) | join(", "))),
+          ("acceptance: " + ((.acceptanceCriteria // []) | join("; "))),
+          ("out-of-scope: " + ((.outOfScope // []) | join(", ")))
+        ]
+      | join(" | ")
+    ' "$report_card" 2>/dev/null | sed '/^$/d' >> "$out_file" || true
+    jq -r '
+      (.qa.architectureFollowUps // [])[]?
+      | [
+          ("summary: " + (.summary // .title // "architecture follow-up")),
+          ("current: " + (.currentBehavior // "not recorded")),
+          ("desired: " + (.desiredBehavior // "not recorded")),
+          ("interfaces: " + ((.keyInterfaces // []) | join(", "))),
+          ("acceptance: " + ((.acceptanceCriteria // []) | join("; "))),
+          ("out-of-scope: " + ((.outOfScope // []) | join(", ")))
+        ]
+      | join(" | ")
+    ' "$report_card" 2>/dev/null | sed '/^$/d' >> "$out_file" || true
   fi
 
   if [[ -f "$manifest" ]]; then
@@ -238,7 +262,9 @@ req_act_collect_evidence() {
     (.evidence // [])[]?,
     ((.claimEvidence // [])[]? | "- " + (.claim // "claim") + ": " + (.status // "unknown") + " via " + (.commandOrArtifact // "artifact") + " - " + (.keyObservation // "")),
     (if .qa.coverageAudit then "- qa.coverage: " + (.qa.coverageAudit.status // "unknown") + ", gaps=" + (((.qa.coverageAudit.gaps // []) | length) | tostring) else empty end),
-    (if .qa.browserEvidence then "- qa.browser: " + (.qa.browserEvidence.status // "unknown") + ", mode=" + (.qa.browserEvidence.mode // "unknown") else empty end)
+    (if .qa.browserEvidence then "- qa.browser: " + (.qa.browserEvidence.status // "unknown") + ", mode=" + (.qa.browserEvidence.mode // "unknown") else empty end),
+    (if .qa.feedbackLoop then "- qa.feedbackLoop: " + (.qa.feedbackLoop.status // "unknown") + ", mode=" + (.qa.feedbackLoop.mode // "unknown") + ", signal=" + (.qa.feedbackLoop.signalSharpness // "not recorded") else empty end),
+    (if .qa.behaviorEvidence then "- qa.behavior: " + (.qa.behaviorEvidence.status // "unknown") + ", expected=" + (.qa.behaviorEvidence.expectedBehavior // "not recorded") + ", actual=" + (.qa.behaviorEvidence.actualBehavior // "not recorded") else empty end)
   ' "$report_card" 2>/dev/null | sed '/^$/d' > "$out_file" || true
   req_act_dedup_file "$out_file"
 }

package/.claude/skills/cc-act/scripts/render-pr-brief.sh

@@ -157,6 +157,23 @@ browser_qa_summary="$(jq -r '
     (if (($browser.skipReason // "") != "") then ", skip=\($browser.skipReason)" else "" end)
   end
 ' "$report_card")"
+feedback_loop_summary="$(jq -r '
+  .qa.feedbackLoop as $loop |
+  if $loop == null then
+    "not recorded"
+  else
+    "status=\($loop.status // "unknown"), mode=\($loop.mode // "unknown"), determinism=\($loop.determinism // "not recorded"), reproductionRate=\($loop.reproductionRate // "not recorded")" +
+    (if (($loop.blockedReason // "") != "") then ", blocked=\($loop.blockedReason)" else "" end)
+  end
+' "$report_card")"
+behavior_evidence_summary="$(jq -r '
+  .qa.behaviorEvidence as $behavior |
+  if $behavior == null then
+    "not recorded"
+  else
+    "status=\($behavior.status // "unknown"), boundary=\($behavior.userFacingBoundary // "not recorded"), expected=\($behavior.expectedBehavior // "not recorded"), actual=\($behavior.actualBehavior // "not recorded"), steps=\((($behavior.reproductionSteps // []) | length))"
+  end
+' "$report_card")"
 failure_ownership_summary="$(jq -r '
   (.runtime.failureOwnership? // [])
   | if length == 0 then
@@ -210,6 +227,8 @@ pr_body_accuracy_summary="body must be regenerated from this pr-brief, current r
   echo "- Specialist review facets: $specialist_review_summary"
   echo "- QA coverage: $qa_coverage_summary"
   echo "- Browser QA: $browser_qa_summary"
+  echo "- Feedback loop: $feedback_loop_summary"
+  echo "- Behavior evidence: $behavior_evidence_summary"
   echo "- Failure ownership: $failure_ownership_summary"
   echo "- Documentation release: $documentation_release_summary"
   echo "- PR body accuracy: $pr_body_accuracy_summary"
@@ -252,6 +271,18 @@ pr_body_accuracy_summary="body must be regenerated from this pr-brief, current r
     echo "- No evidence lines captured yet."
   fi
   echo
+  echo "## QA Behavior Evidence"
+  echo
+  echo "- Feedback loop: $feedback_loop_summary"
+  echo "- Behavior evidence: $behavior_evidence_summary"
+  jq -r '
+    .qa.behaviorEvidence as $behavior |
+    if $behavior == null then empty else
+      (($behavior.reproductionSteps // [])[]? | "- Reproduction step: " + .),
+      (($behavior.domainLanguage // [])[]? | "- Domain language: " + .)
+    end
+  ' "$report_card" 2>/dev/null | sed '/^$/d' || true
+  echo
   echo "## Documentation Sync"
   echo
   echo "- \`CLAUDE.md\`: $claude_status"

package/.claude/skills/cc-act/scripts/verify-act-gate.sh

@@ -47,6 +47,8 @@ failure_ownership_open="$(jq -r '
 ' "$report_card" 2>/dev/null || echo 0)"
 coverage_status="$(jq -r '.qa.coverageAudit.status // "unknown"' "$report_card" 2>/dev/null || echo unknown)"
 browser_status="$(jq -r '.qa.browserEvidence.status // "unknown"' "$report_card" 2>/dev/null || echo unknown)"
+feedback_loop_status="$(jq -r '.qa.feedbackLoop.status // "skipped"' "$report_card" 2>/dev/null || echo skipped)"
+behavior_evidence_status="$(jq -r '.qa.behaviorEvidence.status // "skipped"' "$report_card" 2>/dev/null || echo skipped)"
 
 remaining_tasks="0"
 if [[ -f "$tasks_file" ]]; then
@@ -62,6 +64,8 @@ fi
 [[ "$failure_ownership_open" -eq 0 ]] || { echo "Gate open: failure_ownership_open=$failure_ownership_open" >&2; exit 1; }
 [[ "$coverage_status" != "blocked" && "$coverage_status" != "fail" && "$coverage_status" != "pending" ]] || { echo "Gate open: coverage_status=$coverage_status" >&2; exit 1; }
 [[ "$browser_status" != "blocked" && "$browser_status" != "fail" && "$browser_status" != "pending" ]] || { echo "Gate open: browser_status=$browser_status" >&2; exit 1; }
+[[ "$feedback_loop_status" != "blocked" && "$feedback_loop_status" != "fail" && "$feedback_loop_status" != "pending" ]] || { echo "Gate open: feedback_loop_status=$feedback_loop_status" >&2; exit 1; }
+[[ "$behavior_evidence_status" != "blocked" && "$behavior_evidence_status" != "fail" && "$behavior_evidence_status" != "pending" ]] || { echo "Gate open: behavior_evidence_status=$behavior_evidence_status" >&2; exit 1; }
 [[ "$remaining_tasks" -eq 0 ]] || { echo "Gate open: remaining_tasks=$remaining_tasks" >&2; exit 1; }
 
 cat <<EOF
@@ -75,5 +79,7 @@ REVIEW_FRESHNESS=$review_freshness
 FAILURE_OWNERSHIP_OPEN=$failure_ownership_open
 COVERAGE_STATUS=$coverage_status
 BROWSER_STATUS=$browser_status
+FEEDBACK_LOOP_STATUS=$feedback_loop_status
+BEHAVIOR_EVIDENCE_STATUS=$behavior_evidence_status
 REMAINING_TASKS=$remaining_tasks
 EOF

package/.claude/skills/cc-check/CHANGELOG.md

@@ -1,5 +1,17 @@
 # CC-Check Skill Changelog
 
+## v1.10.0 - 2026-04-28
+
+- add test fixture honesty review for partial fixtures, generated stubs, casts, and missing mock payload fields
+- require fixture shortcuts to preserve public seam behavior instead of hiding contract or input-shape problems
+
+## v1.9.0 - 2026-04-28
+
+- add QA feedback-loop evidence so bugfix and behavior checks record reproduction mode, determinism, signal quality, and blocked loop reasons before a pass verdict
+- add behavior evidence fields for expected behavior, actual behavior, reproduction steps, consistency, and domain language
+- tighten test-quality review around public-interface tests, system-boundary mocks, and architecture follow-ups when no correct test seam exists
+- update the report-card template, renderer, and gate validation for the new QA evidence shape while keeping older reports structurally compatible
+
 ## v1.8.4 - 2026-04-28
 
 - add review freshness, quality score, specialist review facets, and finding confidence fingerprints so stale or noisy review evidence cannot masquerade as current proof

package/.claude/skills/cc-check/PLAYBOOK.md

@@ -41,6 +41,7 @@ NO PASS WITHOUT FRESH EVIDENCE
    - task review proof
    - requirement diff review
    - claim evidence matrix
+   - QA feedback loop and behavior evidence
    - QA regression / test-quality proof
    - QA coverage and browser evidence
    - review freshness and finding confidence
@@ -69,11 +70,12 @@ NO PASS WITHOUT FRESH EVIDENCE
 2. Task review proof
 3. Requirement diff truth
 4. Claim evidence matrix
-5. QA regression and test quality
-6. QA coverage and browser evidence
-7. Review freshness and confidence calibration
-8. Failure ownership
-9. Spec alignment and sync readiness
+5. QA feedback loop and behavior evidence
+6. QA regression and test quality
+7. QA coverage and browser evidence
+8. Review freshness and confidence calibration
+9. Failure ownership
+10. Spec alignment and sync readiness
 
 ## Claim Evidence Matrix
 
@@ -121,13 +123,16 @@ review 还要带 `freshness`：`status`、`reviewedCommit`、`currentCommit`、`
 
 行为变化、bugfix、边界条件、用户可见流程必须补 `qa`：
 
+- `feedbackLoop`：用什么 loop 证明现实，速度、确定性、信号锋利度、复现率如何
+- `behaviorEvidence`：用户边界、expected / actual、复现步骤、稳定性、领域语言
 - `regressionProof`：red command、red failure reason、green command、是否恢复最终状态
-- `testQuality`：是否验证真实行为、mock 边界、是否存在 test-only production API
+- `testQuality`：是否经公共接口验证真实行为、mock 是否只停在系统边界、是否存在 test-only production API
+- `architectureFollowUps`：没有正确 test seam 时记录 seam / hidden coupling / shallow module 的后续改造
 - `tddException`：纯配置、生成文件、throwaway prototype 等例外和替代验证
 - `coverageAudit`：覆盖率、codepath / user-flow map、缺口、是否需要 e2e / eval、测试质量星级
 - `browserEvidence`：UI / 用户路径变更的 affected routes、截图、console、health score、issues，或明确 skip reason
 
-测试只绿过一次，不能证明 regression test 有效；断言 mock 本身，不能证明真实行为。
+测试只绿过一次，不能证明 regression test 有效；断言 mock 本身，不能证明真实行为。没有可信反馈环时，`pass` 不成立。
 
 ## Failure Ownership
 
@@ -171,6 +176,26 @@ review 还要带 `freshness`：`status`、`reviewedCommit`、`currentCommit`、`
   },
   "qa": {
     "status": "pass",
+    "feedbackLoop": {
+      "status": "pass",
+      "mode": "targeted-test",
+      "commandOrArtifact": "npm test -- src/feature/feature.test.ts",
+      "speed": "fast",
+      "determinism": "high",
+      "signalSharpness": "fails only when the target behavior is absent",
+      "reproductionRate": "1/1",
+      "attempts": [],
+      "blockedReason": ""
+    },
+    "behaviorEvidence": {
+      "status": "pass",
+      "userFacingBoundary": "feature action",
+      "expectedBehavior": "the user-visible behavior succeeds",
+      "actualBehavior": "verified by targeted test",
+      "reproductionSteps": [],
+      "consistency": "deterministic",
+      "domainLanguage": []
+    },
     "regressionProof": [],
     "testQuality": [],
     "coverageAudit": {
@@ -193,6 +218,7 @@ review 还要带 `freshness`：`status`、`reviewedCommit`、`currentCommit`、`
       "issues": [],
       "skipReason": "not a UI or user-path change"
     },
+    "architectureFollowUps": [],
     "tddException": null
   },
   "quickGates": [],
@@ -251,3 +277,4 @@ review 还要带 `freshness`：`status`、`reviewedCommit`、`currentCommit`、`
 4. `review.status` 是真实现实，还是我脑补的绿色？
 5. 如果把这份 `report-card.json` 给下一位接手者，他知道接下来去哪吗？
 6. diff review 是否同时覆盖了 plan completion、scope drift、critical pass、doc staleness？
+7. feedback loop 是否真的证明了用户描述的行为，而不是只证明附近代码能跑？

package/.claude/skills/cc-check/SKILL.md

@@ -1,6 +1,6 @@
 ---
 name: cc-check
-version: 1.8.4
+version: 1.10.0
 description: Use when a planned or investigated change needs fresh verification evidence, layered gate proof, review truth, and an honest pass fail blocked verdict before entering cc-act.
 triggers:
   - 验收这个需求
@@ -120,6 +120,7 @@ NO PASS WITHOUT FRESH EVIDENCE
    - task-level review proof
    - requirement-level diff review
    - claim evidence matrix
+   - QA feedback loop and behavior evidence
    - QA regression / test-quality proof
    - spec alignment / sync readiness
 4. **Freeze Verdict**
@@ -138,7 +139,7 @@ NO PASS WITHOUT FRESH EVIDENCE
 
 ## Verification Layers
 
-`cc-check` 不是只看“测试是不是绿的”，而是至少看 9 层：
+`cc-check` 不是只看“测试是不是绿的”，而是至少看 10 层：
 
 1. **Runtime Layer**
    - 测试、lint、typecheck、build、脚本 gate
@@ -154,6 +155,7 @@ NO PASS WITHOUT FRESH EVIDENCE
 6. **QA Test Layer**
    - 回归测试是否有 red/green 证据
    - 测试是否验证真实行为，而不是 mock 或 test-only production API
+   - 反馈环是否能稳定复现或证明用户描述的行为
 7. **Review Freshness Layer**
    - review 是否绑定当前 `headSha`
    - 从 review 到当前 HEAD 是否还有新增 commit
@@ -164,6 +166,9 @@ NO PASS WITHOUT FRESH EVIDENCE
 9. **Failure Ownership Layer**
    - 失败是本分支引入、基线已存在、环境阻塞，还是归属不明
    - 归属不明默认不能支撑 `pass`
+10. **Behavior Contract Layer**
+   - expected / actual / reproduction steps 是否用用户和领域语言写清
+   - follow-up 是否是行为契约，而不是易腐烂的文件行号 TODO
 
 任何一层失真，都不能写 `pass`。
 
@@ -187,13 +192,27 @@ NO PASS WITHOUT FRESH EVIDENCE
 
 `cc-check` 必须区分“有测试”和“测试证明了正确行为”：
 
-1. 回归测试必须记录 red/green 证据；red 要因为目标行为缺失而失败，不是语法、fixture 或 mock 写错。
-2. 测试应验证真实行为；如果依赖 mock，必须说明 mock 的边界和为什么不会测试 mock 本身。
-3. 生产代码里新增仅测试使用的 API，默认是坏味道，必须 blocking，除非有明确生产生命周期理由。
-4. 复杂 mock setup 超过测试主体时，优先要求 integration / contract test 解释。
+1. 先建立反馈环，再谈修复：failing test、curl / HTTP、CLI fixture、headless browser、trace replay、throwaway harness、bisect / differential loop 都可以，但必须说明速度、确定性、信号锋利度和复现率。
+2. 回归测试必须记录 red/green 证据；red 要因为目标行为缺失而失败，不是语法、fixture 或 mock 写错。
+3. 测试应从公共接口验证真实行为；不准为了方便直接测私有实现。
+4. mock 只允许站在系统边界：外部 API、数据库、时间、随机数、文件系统、网络。mock 自家模块、断言内部调用次数或顺序，默认是 review finding。
+5. 生产代码里新增仅测试使用的 API，默认是坏味道，必须 blocking，除非有明确生产生命周期理由。
+6. 复杂 mock setup 超过测试主体时，优先要求 integration / contract test 解释。
+7. test fixture 必须诚实表达 contract：partial fixture、generated stub、`as` / `any` / 双重 cast、缺字段 mock payload 都要说明真实字段与填充字段；如果这些技巧让测试绕过公共 seam 或隐藏错误输入，默认是 review finding。
+8. 如果没有正确测试 seam，不要硬造脆弱测试；记录 `qa.architectureFollowUps`，说明缺失 seam / hidden coupling / shallow module，并按严重度决定 reroute 或 follow-up。
 
 这些事实写入 `qa.regressionProof` 和 `qa.testQuality`。如果本需求没有行为测试空间，必须记录 `tddException` 或替代验证命令。
 
+## QA Behavior Evidence
+
+用户可见行为、bugfix、regression、工作流、CLI 行为和 API 行为都必须留下行为证据：
+
+1. `qa.feedbackLoop` 记录本轮用什么 loop 证明现实，包含 `status`、`mode`、`commandOrArtifact`、`speed`、`determinism`、`signalSharpness`、`reproductionRate`、`attempts`、`blockedReason`。
+2. `qa.behaviorEvidence` 记录 `userFacingBoundary`、`expectedBehavior`、`actualBehavior`、`reproductionSteps`、`consistency`、`domainLanguage`、`status`。
+3. bugfix 不能只写“代码改了”；必须证明用户描述的原始症状已经被同一条或更可信的反馈环覆盖。
+4. 不能复现时，verdict 默认 `blocked` 或回 `cc-investigate`，并写清尝试过哪些 loop、还缺什么 artifact / 权限 / 输入。
+5. QA issue / follow-up 必须用行为和验收条件表达，不写易失效的文件路径或行号，除非它是当前 review finding 的证据位置。
+
 ## QA Coverage And Browser Evidence
 
 测试不是数量游戏。`cc-check` 必须判断测试覆盖了哪条真实路径：

package/.claude/skills/cc-check/assets/REPORT_CARD_TEMPLATE.json

@@ -42,6 +42,32 @@
   ],
   "qa": {
     "status": "blocked",
+    "feedbackLoop": {
+      "status": "blocked",
+      "mode": "targeted-test",
+      "commandOrArtifact": "npm test -- src/feature/feature.test.ts",
+      "speed": "fast",
+      "determinism": "high",
+      "signalSharpness": "targeted failure would indicate the user-visible behavior regressed",
+      "reproductionRate": "not recorded",
+      "attempts": [
+        "targeted unit test"
+      ],
+      "blockedReason": "red/green reproduction is not recorded yet"
+    },
+    "behaviorEvidence": {
+      "status": "blocked",
+      "userFacingBoundary": "feature behavior",
+      "expectedBehavior": "target behavior succeeds for the user",
+      "actualBehavior": "not proven yet",
+      "reproductionSteps": [
+        "run the targeted verification command"
+      ],
+      "consistency": "not recorded",
+      "domainLanguage": [
+        "feature behavior"
+      ]
+    },
     "regressionProof": [
       {
         "behavior": "original symptom",
@@ -81,6 +107,23 @@
       "issues": [],
       "skipReason": "template example is not a UI browser QA scenario"
     },
+    "architectureFollowUps": [
+      {
+        "summary": "Add the missing public test seam before widening coverage",
+        "currentBehavior": "review proof depends on task metadata instead of a direct behavior seam",
+        "desiredBehavior": "tests verify the behavior through a stable public interface",
+        "keyInterfaces": [
+          "feature public API"
+        ],
+        "acceptanceCriteria": [
+          "target behavior can be reproduced without private implementation hooks"
+        ],
+        "outOfScope": [
+          "rewriting unrelated modules"
+        ],
+        "status": "deferred-minor"
+      }
+    ],
     "tddException": null
   },
   "quickGates": [

package/.claude/skills/cc-check/references/gate-contract.md

@@ -16,6 +16,17 @@
 
 所有通过结论都必须来自本次新鲜证据；旧输出只能当线索，不能直接继承 verdict。
 
+## QA Feedback Loop
+
+行为变更和 bugfix 的 evidence 还必须说明反馈环：
+
+- `mode`：failing test、curl / HTTP、CLI fixture、browser、trace replay、bisect、differential loop 等
+- `determinism`：反馈是否稳定，flaky 时复现率是多少
+- `signalSharpness`：失败是否指向目标行为，而不是语法、fixture 或 mock 问题
+- `blockedReason`：无法建立 loop 时缺少什么 artifact、权限、服务或输入
+
+没有可信 loop 的 bugfix 默认不能 `pass`。
+
 ## Reroute
 
 - `none`

package/.claude/skills/cc-check/references/review-contract.md

@@ -98,15 +98,29 @@ Review finding 不只是“发现过”，必须有处置结果：
 
 Review 必须判断测试是否证明行为：
 
+- 反馈环是否可信：速度、确定性、信号锋利度、复现率是否足够支撑结论
+- bugfix 是否复现并覆盖了用户描述的原始症状，而不是附近的另一个失败
+- expected / actual / reproduction steps 是否能让 reviewer 独立复现或判断缺件
 - 回归测试是否有 red/green 证据
 - red 是否因为目标行为缺失而失败
 - green 是否包含 targeted test 和必要的 broader gate
-- mock 是否必要，且没有断言 mock 本身
+- 测试是否通过公共接口覆盖行为
+- mock 是否只停在系统边界，且没有断言 mock 本身或内部调用顺序
 - 生产代码是否新增 test-only API
 - integration / contract test 是否比复杂 mock 更直接
+- 如果没有正确测试 seam，是否记录了架构 follow-up，而不是造易碎测试
 - coverage audit 是否映射真实 codepath / user flow / error state / edge case
 - UI 或用户路径变更是否有 browser evidence、截图、console 结果，或明确 skip reason
 
+## Durable Follow-Up Facts
+
+Review 产生的 QA issue 或 follow-up 必须可长期执行：
+
+- 用领域语言描述用户或系统行为，不把当前文件路径 / 行号当成唯一真相
+- 写清 current behavior、desired behavior、key interfaces、acceptance criteria、out of scope
+- 独立行为拆成独立条目；有依赖关系时写明顺序
+- `deferred-minor` 只能用于不阻塞本次交付的 minor 项，并且必须进入 `cc-act` follow-up writeback
+
 ## Failure Ownership
 
 失败归属必须结构化写入 `runtime.failureOwnership[]`：
@@ -125,6 +139,8 @@ Review 必须判断测试是否证明行为：
 - `important` / `critical` finding 未处理前，不算通过
 - `important` / `critical` finding 缺 triageStatus，不算通过
 - QA test quality 缺失且本次涉及行为变化，至少是 `blocked`
+- 行为变更缺 `qa.feedbackLoop` / `qa.behaviorEvidence` 且没有明确例外，至少是 `blocked`
+- bugfix 没有复现原始症状，也没有解释不可复现原因，不能通过
 - review freshness 缺失、过期或与当前 head 不一致，不能绿灯
 - UI / 用户路径变更缺 browser evidence 且无 skip reason，不能绿灯
 - `runtime.failureOwnership` 仍有 `in-branch` 或 `ambiguous` 未解释失败，不能绿灯