feat(governance): 新增 learning writeback approval package

docs/superpowers/specs/2026-04-15-MASTER-ai-autonomous-flywheel-v2.md

@@ -634,7 +634,7 @@ Alert / Sentry / SigNoz / Gitea / Market Watch / Operator
 | `docs/evaluations/ai_agent_communication_learning_contract_2026-06-11.json` | 2026-06-11 committed snapshot；完成度 `35%`，runtime worker / DB migration / Telegram direct send 全部 false |
 | `apps/api/src/services/ai_agent_communication_learning_contract.py` | 只讀 loader；強制驗證 runtime / migration / Telegram / SDK / route 權限都未開 |
 | `GET /api/v1/agents/agent-communication-learning-contract` | 治理 API；只回傳 committed contract，不啟動 worker、不碰 DB/Redis、不呼叫外部服務 |
-| `docs/evaluations/ai_agent_interaction_learning_proof_2026-06-11.json` + `GET /api/v1/agents/agent-interaction-learning-proof` | P2-403A / P2-403B / P2-403C 互動、接手、學習、成長、read model gate 與 Redis dry-run gate 證據面；目前 live session、message、handoff、learning write、Telegram receipt 全部 `0`，下一步 P2-403D |
+| `docs/evaluations/ai_agent_interaction_learning_proof_2026-06-11.json` + `GET /api/v1/agents/agent-interaction-learning-proof` | P2-403A / P2-403B / P2-403C / P2-403D 互動、接手、學習、成長、read model gate、Redis dry-run gate 與 learning writeback approval package 證據面；目前 live session、message、handoff、learning write、Telegram receipt 全部 `0`，下一步 P2-403E |
 | `docs/evaluations/ai_agent_live_read_model_gate_2026-06-11.json` + `GET /api/v1/agents/agent-live-read-model-gate` | P2-403B AgentSession / Redis Streams live read model gate；定義 safe fields、Redis envelope、worker gate、rollback plan 與 no-write smoke，不連 DB、不讀寫 Redis、不啟動 worker |
 
 #### 3.2.1c 2026-06-11 AI Agent 主動營運委派與版本生命週期契約
@@ -696,6 +696,9 @@ Repo / registry / release notes / K8s / host / observability / backup evidence
 | `docs/schemas/ai_agent_redis_dry_run_gate_v1.schema.json` | P2-403C Redis Streams consumer group dry-run gate schema；鎖定 Redis connection / consumer group / XADD / XREADGROUP / ACK / dead-letter / replay / Telegram / learning writeback 全部 false |
 | `docs/evaluations/ai_agent_redis_dry_run_gate_2026-06-11.json` | P2-403C committed snapshot：fixture-only consumer group dry-run、handoff envelope、ack / dead-letter / replay idempotency、frontend redaction；live counts 全部 `0` |
 | `GET /api/v1/agents/agent-redis-dry-run-gate` | 只讀 API；只回傳 Redis dry-run gate，不連 Redis、不建立 consumer group、不 XADD / XREADGROUP / ACK、不寫 dead-letter、不 replay、不發 Telegram |
+| `docs/schemas/ai_agent_learning_writeback_approval_package_v1.schema.json` | P2-403D learning writeback approval package schema；鎖定 KM write / PlayBook trust update / timeline learning / replay score / Telegram / runtime worker 全部 false |
+| `docs/evaluations/ai_agent_learning_writeback_approval_package_2026-06-11.json` | P2-403D committed snapshot：owner review、rollback、redaction、blocked write actions；live writes 全部 `0` |
+| `GET /api/v1/agents/agent-learning-writeback-approval-package` | 只讀 API；只回傳 learning writeback approval package，不寫 KM、不更新 PlayBook trust、不寫 timeline / replay score、不發 Telegram |
 | `/zh-TW/governance?tab=automation-inventory` | P2-402G governance UI：顯示可委派能力、版本生命週期、Host / K3s / stateful 只讀關卡與 Telegram / redaction gate；不提供執行按鈕 |
 
 **採用順序：**
@@ -709,7 +712,7 @@ Repo / registry / release notes / K8s / host / observability / backup evidence
 7. 先建立可被統帥看見的互動 / 接手 / 學習 / 成長證據面。✅ P2-403A 已完成；live count 仍為 `0`，不宣稱 runtime 已運作。
 8. 建立 AgentSession / Redis live read model gate。✅ P2-403B 已完成；safe fields、Redis envelope、worker gate、rollback plan、no-write smoke 已接入治理頁，但 live DB query / Redis consumer / worker 仍未授權。
 9. 建立 Redis Streams consumer group dry-run、handoff envelope、ack / dead-letter / replay gate。✅ P2-403C 已完成；fixture-only dry-run、idempotency、redacted evidence 與治理頁已接入，但 Redis runtime consumer / ACK / replay 仍未授權。
-10. 最後才進人工批准後的 learning writeback / smoke / canary / production rollout。下一步 P2-403D learning writeback approval package。
+10. 建立 learning writeback approval package，先固定 owner review、rollback、redaction 與 blocked write actions。✅ P2-403D 已完成；KM / PlayBook trust / timeline learning / replay score live write 仍未授權。下一步 P2-403E Telegram receipt approval package。
 
 #### 3.2.1d 2026-06-11 Agent 互動、學習與成長證據面
 
@@ -724,7 +727,7 @@ Repo / registry / release notes / K8s / host / observability / backup evidence
 | Live read model | Agent 心跳、最近狀態、資料時間 | P2-403B gate 已完成；live count 仍為 `0` |
 | 訊息收據 | 發話、收話、ack、dead letter、replay | P2-403C dry-run gate 已完成；live receipt count 仍為 `0` |
 | 接手紀錄 | OpenClaw / Hermes / NemoTron parent turn / handoff chain | P2-403C handoff envelope 已完成；runtime handoff count 仍為 `0` |
-| 學習回寫 | KM、Playbook trust、learning failure、replay score | P2-403D 待辦 |
+| 學習回寫 | KM、Playbook trust、learning failure、replay score | P2-403D approval package 已完成；live learning write count 仍為 `0` |
 | Telegram 收據 | action-required / approval / failure digest queued / delivered / acknowledged | P2-403E 待辦 |
 
 **本波已建立的可執行契約（只讀，不授權 runtime）：**
@@ -732,7 +735,7 @@ Repo / registry / release notes / K8s / host / observability / backup evidence
 | 檔案 / API | 用途 |
 |---|---|
 | `docs/schemas/ai_agent_interaction_learning_proof_v1.schema.json` | 互動、接手、學習、成長、Telegram receipt 與前端 redaction schema |
-| `docs/evaluations/ai_agent_interaction_learning_proof_2026-06-11.json` | P2-403A + P2-403B + P2-403C committed snapshot；完成度 `65%`，live truth counts 全部 `0` |
+| `docs/evaluations/ai_agent_interaction_learning_proof_2026-06-11.json` | P2-403A + P2-403B + P2-403C + P2-403D committed snapshot；完成度 `72%`，live truth counts 全部 `0` |
 | `apps/api/src/services/ai_agent_interaction_learning_proof.py` | 只讀 loader；強制 live flags / DB / Redis / Telegram / transcript / 私有推理全部關閉 |
 | `GET /api/v1/agents/agent-interaction-learning-proof` | 治理 API；只回傳證據面，不啟動 worker、不碰 live DB/Redis、不發 Telegram |
 | `docs/schemas/ai_agent_live_read_model_gate_v1.schema.json` | P2-403B live read model gate schema；強制 DB / Redis / worker / Telegram / learning writeback 仍需批准 |
@@ -743,6 +746,10 @@ Repo / registry / release notes / K8s / host / observability / backup evidence
 | `docs/evaluations/ai_agent_redis_dry_run_gate_2026-06-11.json` | P2-403C committed snapshot；fixture-only dry-run、handoff envelope、ack / dead-letter / replay idempotency、frontend redaction |
 | `apps/api/src/services/ai_agent_redis_dry_run_gate.py` | 只讀 loader；拒絕 Redis network call、consumer group creation、runtime replay、Telegram send 與 writeback |
 | `GET /api/v1/agents/agent-redis-dry-run-gate` | 治理 API；只回傳 Redis dry-run gate，不連 Redis、不建立 consumer group、不 XADD / XREADGROUP / ACK、不 replay、不發 Telegram |
+| `docs/schemas/ai_agent_learning_writeback_approval_package_v1.schema.json` | P2-403D learning writeback approval package schema；強制 KM / PlayBook trust / timeline learning / replay score / Telegram / runtime worker 全部維持未授權 |
+| `docs/evaluations/ai_agent_learning_writeback_approval_package_2026-06-11.json` | P2-403D committed snapshot；owner review、rollback、redaction、blocked write actions 與 live write total `0` |
+| `apps/api/src/services/ai_agent_learning_writeback_approval_package.py` | 只讀 loader；拒絕 KM write、PlayBook trust update、timeline learning、replay score write、Telegram send 與 runtime worker |
+| `GET /api/v1/agents/agent-learning-writeback-approval-package` | 治理 API；只回傳 learning writeback approval package，不寫 KM、不更新 PlayBook trust、不寫 timeline / replay score、不發 Telegram |
 | `/zh-TW/governance?tab=automation-inventory` | 顯示證據階梯、目前真相、三 Agent lane、可觀測訊號、runtime gates 與 redaction policy |
 
 **硬性紅線：**
@@ -1818,6 +1825,13 @@ Phase 6 完成後
 - 更新 `ai_agent_interaction_learning_proof_2026-06-11.json`：整體完成度 `65%`，current task `P2-403C`，next task `P2-403D`；live AgentSession / Redis events / handoff / learning write / Telegram digest receipt 全部仍為 `0`。
 - 本波仍不連 Redis、不建立 consumer group、不 XADD / XREADGROUP / XACK、不寫 dead-letter、不 runtime replay、不發 Telegram、不做 learning writeback、不讀取或輸出 secret；下一步 P2-403D 才做 learning writeback approval package。
 
+### 2026-06-11 23:59 (台北) — §3.2 / §5 — 完成 P2-403D Learning Writeback Approval Package — 先把學習回寫拆成人工批准包，不再用 no-action 包裝自動化
+
+- 新增 `ai_agent_learning_writeback_approval_package_v1` schema / committed snapshot / loader / API / 測試，定義 KM、PlayBook trust、timeline learning、replay score 的 proposed delta、owner review、rollback、redaction 與 blocked write actions。
+- `apps/web/src/app/[locale]/governance/tabs/automation-inventory-tab.tsx` 接入 `GET /api/v1/agents/agent-learning-writeback-approval-package`，顯示 learning writeback KPI、批准包欄位、禁止欄位、review gates、learning lanes、truth flags 與 live write count。
+- 更新 `ai_agent_interaction_learning_proof_2026-06-11.json`：整體完成度 `72%`，current task `P2-403D`，next task `P2-403E`；live AgentSession / Redis events / handoff / learning write / Telegram digest receipt 全部仍為 `0`。
+- 本波仍不寫 KM、不更新 PlayBook trust、不寫 timeline learning、不寫 replay score、不發 Telegram、不啟動 runtime worker、不讀取或輸出 secret；下一步 P2-403E 才做 Telegram receipt approval package。
+
 ### 2026-04-15 (台北) — 全檔 — 建立 v2 骨架，§0/§1 完成 — 統帥批准「單 MASTER + 4 道閘門」結構
 
 - 從 v1（plans/2026-04-15-MASTER-ai-autonomous-flywheel.md）繼承核心發現