feat: sync auto-next obligation gate hardening

README.md

@@ -2,78 +2,102 @@
 
 ## 中文說明
 
-這個 repo 是從較大的 OpenClaw workspace 中抽出的焦點工作流成果，主題是：
+這個 repo 目前聚焦兩條與 continuity 直接相關的成果：
 
 - **approved plan continuity hard-gate**
-- **dispatch receipt binding**
-- **anti-blackhole / completion-delivery watchdog groundwork**
+- **auto-next obligation gate**
 
-目標是避免兩類問題持續發生：
+目標是避免以下兩種 failure：
 
 1. **continuity failure / auto-next break**
-2. **subagent anti-blackhole / fake timeout**
+   - 任務已完成
+   - 下一步已知
+   - 但沒有真的 dispatch 下一顆 task
+   - 流程卻還是被當成正常收尾
+
+2. **task-boundary stop / 口頭續跑**
+   - 同一份 approved plan 內其實應該 auto-next
+   - 但主代理停在 task boundary
+   - 用 checkpoint / 口頭回報 / session metadata 取代真正 dispatch
 
 ## 目前已完成
 
 ### A. Continuity hard-gate
 - continuity evaluator
-- dispatch receipt binding groundwork
-- `derivedAction` continuity binding
+- receipt validator 最小欄位驗證
+- `derivedAction` / `nextDerivedAction` 納入 continuity 判定
 - `dry_run_dispatch` 不得冒充真 receipt
-- fake receipt authority 最小收緊
-- hook integration 已接入
+- fake receipt 不得放行
+- hook integration 已接入 `hooks/force-recall/handler.ts`
 
-### B. Anti-blackhole watchdog recovery
-- watchdog status recompute
-- 最小 recovery decision 閉環：
-  - `fetch_history`
-  - `respawn`
-  - `blocked`
-- owner-visible reporting payload
-- scenario matrix tests
+### B. Auto-next obligation gate
+- 新 failure reason：`missing_auto_next_dispatch`
+- 同一份 approved plan 中，若：
+  - 當前 task 已完成
+  - 下一顆 task 已知
+  - `sameApprovedPlan=true`
+  - `taskBoundaryStop=true`
+  - 非 `waiting_user` / `blocked` / `pending_verification`
+  - 非 `highRiskStop`
+  - 且沒有真實 next dispatch receipt
+  - ⇒ 直接 fail，不得停在 boundary 等主人再說「繼續」
+- receipt linkage hardening：receipt 現在要對到要求的 next-task handoff，而不是只要存在就算過
+- 新增最小 linkage 欄位：`nextTaskId`
+- checkpoint / session metadata / stale receipt / dry-run planner intent 不得冒充 auto-next dispatch proof
+
+## 驗證狀態
+- `node scripts/test_approved_plan_continuity_gate.mjs` → `17 passed / 0 failed`
+- `node scripts/test_force_recall_long_task_preflight.mjs` → PASS
+- `node --check hooks/force-recall/handler.ts` → PASS
+- `node --check scripts/approved_plan_continuity_gate.mjs` → PASS
+- `node --check scripts/approved_plan_dispatch_binding.mjs` → PASS
 
 ## 目前限制
-- continuity 仍偏 prompt-level hard-gate integration
-- watchdog recovery 目前驗收的是 decision / reporting / test slice，不是 live integration
+- 目前仍主要鎖在 continuity / force-recall 路徑，不是所有 entry points。
+- `sameApprovedPlan` 的上游證據仍可再更硬。
+- continuity plugin MVP 仍在後續產品化中，尚未整理成可直接讓其他 OpenClaw 安裝的插件包。
 
-## 下一步建議
-1. continuity runtime enforcement hardening
-2. watchdog live recovery integration
-3. escalation / receipt contract hardening
+## 下一步
+1. continuity 收尾覆核
+2. 回到 continuity plugin MVP
+3. 把目前 continuity 內核抽成可安裝、可設定、可測試、可依雙語 README 套用的插件 MVP
 
 ---
 
 ## English Description
 
-This repository is a focused export from a larger OpenClaw workspace covering:
+This repository currently focuses on two continuity-related hardening slices:
 
 - **approved plan continuity hard-gate**
-- **anti-blackhole / completion-delivery watchdog recovery**
+- **auto-next obligation gate**
+
+It prevents two core failure classes:
+
+1. **continuity failure / auto-next break**
+2. **task-boundary stop disguised as progress**
 
 ## Current State
 
 ### A. Continuity hard-gate
 - continuity evaluator
-- dispatch receipt binding groundwork
-- `derivedAction` continuity binding
-- `dry_run_dispatch` no longer accepted as a real receipt
-- fake receipt authority tightened
-- hook integration present
+- minimum receipt validation
+- `derivedAction` / `nextDerivedAction` continuity handling
+- `dry_run_dispatch` rejected as real receipt
+- fake receipt rejected
+- hook integration in `hooks/force-recall/handler.ts`
 
-### B. Anti-blackhole watchdog recovery
-- watchdog status recompute
-- minimal recovery-decision loop:
-  - `fetch_history`
-  - `respawn`
-  - `blocked`
-- owner-visible reporting payload
-- scenario matrix tests
+### B. Auto-next obligation gate
+- explicit failure reason: `missing_auto_next_dispatch`
+- task-boundary stop is now treated as continuity failure when same-plan auto-next is obligatory
+- receipt linkage hardening via `nextTaskId`
+- checkpoint / session metadata / stale receipt / dry-run intent can no longer stand in for real auto-next dispatch proof
+
+## Validation
+- continuity gate tests passing
+- force-recall preflight passing
+- syntax checks passing
 
 ## Current Limitations
-- continuity remains prompt-level rather than engine-level
-- watchdog recovery is validated as a decision/reporting/test slice, not live execution integration
-
-## Suggested Next Steps
-1. continuity runtime enforcement hardening
-2. watchdog live recovery integration
-3. escalation / receipt contract hardening
+- scoped mainly to the continuity / force-recall path
+- upstream `sameApprovedPlan` evidence can still be hardened further
+- plugin packaging is still pending