AirPlan V2 initial release — unified scheduler with 12 sub-modes

Merges 8 V1 plugins into 1 unified plugin, adds 4 new components (dep, tst, sec, rvr). 12 sub-modes: arc, eng, do, dbg, xdb, sdb, ndb, ctx, dep, tst, sec, rvr. L1 code-level guarantees: atomic writes, file locks, evidence gating, forced debug routing, 3-phase arc gate, evidence-first debug gate, Chinese language lock, scheduler boundary. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-06-10 16:24:26 +08:00
commit 2c4b3340bf
81 changed files with 6005 additions and 0 deletions
--- a/.claude-plugin/plugin.json
+++ b/.claude-plugin/plugin.json
@@ -0,0 +1,67 @@
+{
+  "name": "airplan-v2",
+  "version": "2.0.0",
+  "description": "AirPlan V2 — 统一制品驱动开发调度器。单一插件整合规划(arc)、执行(do)、调试(dbg)、GUI验证(xdb)、静态分析(sdb)、网络调试(ndb)、部署(dep)、测试(tst)、安全扫描(sec)、需求审查(rvr)。L1代码级保障。",
+  "author": {
+    "name": "AirPlan Team",
+    "email": "noreply@airlongdian.fun",
+    "url": "https://airlongdian.fun"
+  },
+  "homepage": "https://airlongdian.fun/plugins/airplan-v2",
+  "repository": "https://airlongdian.fun/plugins/airplan-v2",
+  "license": "MIT",
+  "keywords": [
+    "airplan",
+    "scheduler",
+    "orchestrator",
+    "debugger",
+    "static-analysis",
+    "deployment",
+    "testing",
+    "security",
+    "v2"
+  ],
+  "skills": "./skills/",
+  "commands": "./commands/",
+  "interface": {
+    "displayName": "AirPlan V2",
+    "shortDescription": "统一制品驱动开发调度器 (12子模式)",
+    "longDescription": "V2 将 V1 的 8 个独立插件合并为 1 个统一插件，新增 4 个组件(部署/测试/安全/需求审查)，共 12 个子模式：arc(架构)、eng(调度)、do(执行)、dbg(调试)、xdb(GUI)、sdb(静态)、ndb(网络)、ctx(上下文)、dep(部署)、tst(测试)、sec(安全)、rvr(需求)。L1代码级保障：不依赖LLM自觉。",
+    "developerName": "AirPlan Team",
+    "category": "Productivity",
+    "capabilities": [
+      "Interactive",
+      "Write",
+      "Bash"
+    ],
+    "websiteURL": "https://airlongdian.fun/plugins/airplan-v2",
+    "privacyPolicyURL": "https://openai.com/policies/row-privacy-policy/",
+    "termsOfServiceURL": "https://openai.com/policies/row-terms-of-use/",
+    "defaultPrompt": [
+      "使用 AirPlan V2 初始化项目规划上下文",
+      "使用 AirPlan V2 运行架构规划 (arc)、调度引擎 (eng)、任务执行 (do)",
+      "使用 AirPlan V2 进行调试 (dbg)、GUI验证 (xdb)、静态分析 (sdb)",
+      "使用 AirPlan V2 部署 (dep)、测试 (tst)、安全扫描 (sec)、需求审查 (rvr)"
+    ],
+    "brandColor": "#6366F1",
+    "screenshots": []
+  },
+  "modes": {
+    "arc": "架构规划器 — 分析依赖、写集冲突、产出执行计划",
+    "eng": "调度引擎 — 波次派发、监控、合并、修复编排",
+    "do": "任务执行器 — 单任务切片执行与结果生成",
+    "dbg": "调试器 — 7步调试工作流、根因分析、修复回滚",
+    "xdb": "GUI验证器 — 截图取证、GUI操作验证、DRM/KMS截图",
+    "sdb": "静态分析器 — 多语言静态分析、diff模式",
+    "ndb": "网络调试器 — 抓包分析、TLS解密、远程探测",
+    "ctx": "上下文管理器 — 压缩、Token估算、锁检测",
+    "dep": "部署器 — SSH远程构建、部署、systemd管理",
+    "tst": "测试运行器 — 统一测试执行与结果报告",
+    "sec": "安全扫描器 — 敏感数据检测与脱敏",
+    "rvr": "需求审查器 — 交付物与需求一致性验证"
+  },
+  "requirements": {
+    "python": ">=3.10",
+    "system": ["git", "fcntl"]
+  }
+}
--- a/.gitignore
+++ b/.gitignore
@@ -0,0 +1,7 @@
+__pycache__/
+*.pyc
+*.pyo
+.bak
+*.tmp
+*.lock
+.DS_Store
--- a/airplanV2审查后.md
+++ b/airplanV2审查后.md
@@ -0,0 +1,209 @@
+# AirPlan V2 设计文档 — OPUS 4.8 审查报告
+
+> **原文档**: airplanV2-Qwen3.7-Max设计.md (Draft 0.1, 2026-06-09)
+> **审查日期**: 2026-06-09
+> **审查结论**: 整体架构方向正确，设计深度足够，存在 12 项需修正问题和 8 项需补充内容
+
+---
+
+## 一、需修正问题
+
+### 1.1 编号错误：两个 3.7.4
+
+Section 3.7 中 AirRvr 和事件索引层均编号为 `3.7.4`。事件索引层应为 `3.7.5`。
+
+### 1.2 AirArc 工具限制配置格式不对
+
+3.2.6 中给出了 JSON 格式的工具白名单配置：
+
+```json
+{
+  "allowed_tools": ["Read", "Glob", "Grep"],
+  "denied_tools": ["Write", "Edit", "Bash", "NotebookEdit"],
+  "deny_plan_mode": true
+}
+```
+
+Claude Code 的 SKILL.md 使用 YAML frontmatter 定义工具限制，格式应为：
+
+```yaml
+---
+name: AirArc
+allowed-tools: [Read, Glob, Grep]
+---
+```
+
+不存在 `denied_tools` 和 `deny_plan_mode` 字段（这两个是虚构的 API）。实际的 plan mode 阻断依赖 SKILL.md 中的强制性文本指令 + `allowed-tools` 白名单双重保障。文档中的伪 JSON 配置块应修正为实际可用的 YAML frontmatter。
+
+### 1.3 P0-7 修复方案核心仍为 Prompt 指令，非代码级保障
+
+P0-7（AirEng 无子线程状态轮询）的修复方案（3.2.8）主体是一个 SKILL.md 中的伪代码循环指令，依赖 LLM 自觉执行。尽管末尾有「技术补充：engine monitor_engine() 增加 wall-clock 超时检测」，但：
+
+- wall-clock 超时只是兜底检测（2小时硬上限），不能替代 5 分钟级的状态轮询
+- 真正解决 P0-7 需要在 `engine.py` 中增加代码级定时轮询循环（如 `while` 循环 + `time.sleep`），而非依赖 prompt 指令
+- 建议将 3.2.8 的主体从 SKILL.md 文本改为 `engine.py` 中的 `monitor_engine()` 重构，SKILL.md 指令仅作为辅助提醒
+
+### 1.4 Phase 2 实现 PlanDelta 消费者，但生产者到 Phase 4 才实现
+
+3.2.11 的 `TaskGraph.apply_delta()` 在 Phase 2 (T-2.9) 实现，但其数据来源——AirArc 的增量重规划（产出 `PlanDelta`）——在 Phase 4 (T-4.6) 才实现。这意味着 Phase 2~4 期间，`apply_delta()` 没有真正的生产者。建议：
+
+- 将 AirArc 增量重规划从 Phase 4 提前到 Phase 2，与 T-2.9 同步交付；或
+- Phase 2 先实现全量 replace 模式（Arc 产出完整 DAG，Eng 整体替换），Phase 4 再升级为增量 delta 模式
+
+### 1.5 区域冲突检测未定义「区域」概念
+
+3.2.12 的 `RegionConflictDetector._check_region_overlap()` 是核心判断逻辑，但设计中没有定义什么是「区域」、如何从任务中提取区域信息、两个区域如何判重叠。这是 worktree 隔离并行的关键前提，不能留空。建议至少明确：
+
+- 区域定义来源：函数边界 / 类定义 / CSS 选择器块 / 标记注释分隔
+- 区域信息由谁提供：Arc 规划时静态分析 / Worker 执行前动态分析 / todo.md 中的区域标注
+- 重叠判据：行号区间重叠 / AST 节点冲突 / 文本 diff 冲突
+
+### 1.6 Token 估算比率命名容易误解
+
+3.3.2 中 `RATIOS` 字典的值实际含义是「每 token 对应的字符数」，但命名 `RATIOS` 和注释「中文字符 → token」暗示反向。建议重命名为 `CHARS_PER_TOKEN` 并在注释中明确。
+
+### 1.7 AirDbg 快照可能包含无关变更
+
+3.5.2 的 `pre_fix_snapshot` 使用 `git commit -am` 提交所有修改。在无 worktree 隔离的场景下，这会混入其他并行任务的文件变更，导致回滚时误伤。建议改为仅提交当前任务写集范围内的文件，或仅在 worktree 隔离模式下使用。
+
+### 1.8 AirTst 结果格式不支持嵌套测试结构
+
+3.7.2 的结果格式为扁平结构，但 GoogleTest 有 test suite → test case 层级，pytest 有 module → class → function 层级。建议扩展为：
+
+```json
+{
+  "framework": "googletest",
+  "totalTests": 128, "passed": 127, "failed": 0, "disabled": 1,
+  "duration": "4.2s",
+  "suites": [
+    {
+      "name": "NetworkTest",
+      "totalTests": 32, "passed": 32, "failed": 0,
+      "testCases": [
+        {"name": "ConnectTimeout", "status": "passed", "duration": "0.05s"}
+      ]
+    }
+  ],
+  "failures": []
+}
+```
+
+### 1.9 KMS 截图硬编码 sudo
+
+3.6.1 中 `KmsGrabCapture` 直接使用 `sudo ffmpeg`，这要求密码免密 sudo 配置，在生产环境中是安全隐患。建议增加权限检测和降级策略：
+
+- 检测当前用户是否有 `/dev/dri/card0` 读写权限
+- 无权限时提示用户配置 udev 规则或将用户加入 video 组，而非直接 sudo
+- sudo 仅作为显式 opt-in 的 fallback
+
+### 1.10 迁移策略未覆盖 todo.md → DAG 过渡
+
+Section 5 的迁移策略讨论了 state.json 的兼容性，但未涉及最核心的格式变迁——todo.md 被 TaskGraph 替代。现有项目的 todo.md 如何转换为 DAG？转换后人类如何查看/编辑任务状态（todo.md 的双重角色：机器可读 + 人类可读）？建议：
+
+- 保留 todo.md 作为人类可读视图，DAG 作为内部调度结构
+- 提供 `todo.md → DAG` 导入器和 `DAG → todo.md` 导出器
+- 或明确声明 V2 放弃人类直接编辑 todo 的便利性，改为通过 AirArc 间接操作
+
+### 1.11 度量指标缺少量化手段
+
+Section 7 中部分指标无法客观测量：如「AirEng 非中文输出」「AirEng 轮询遗忘」「AirDo 跳过 AirDbg」的目标值设为 0 次，但当前没有机制检测这些行为是否发生。建议为每个指标定义检测方法：
+
+- 非中文输出：事件日志中记录 AirEng 输出语言，正则检测非中文字符比例
+- 轮询遗忘：monitor_engine 事件日志的时间间隔分析，超过 10 分钟无轮询记录即为遗忘
+- 跳过 AirDbg：事件日志中 blocked→merge 路径无 debug.session 事件即为跳过
+
+### 1.12 worktree 合并冲突处理缺失
+
+3.2.12 描述了 SOFT 冲突时创建 worktree 并行执行，完成后 `git merge` 回主分支。但未定义合并冲突时的处理流程。两个 worktree 即使修改同一文件的不同区域，git merge 仍可能因相邻行冲突而失败。建议增加合并冲突升级策略：
+
+```
+MergeResult.CONFLICT → 自动升级到 AirDbg 解决 → 仍失败则降级为串行重执行
+```
+
+---
+
+## 二、需补充内容
+
+### 2.1 插件发现与加载机制
+
+文档新增了 AirDep、AirTst、AirSec、AirRvr 四个插件，但未说明引擎如何发现和加载它们。V1 的插件加载机制是什么？V2 是否保持一致？建议在 3.7 新增一节说明插件注册规范。
+
+### 2.2 AirRvr 的成本模型
+
+AirRvr 对每个任务运行独立的 LLM 审查（加载需求文档 + diff + 验证证据），Token 消耗可能远超任务执行本身。以 100 任务项目为例，AirRvr 全量审查的 API 成本需要估算。建议：
+
+- 定义审查触发策略：默认逐任务，可配置为波次审查或里程碑审查
+- 给出不同模式下的 Token 消耗估算
+- 考虑轻量审查模式（仅对比 Done When 文本，不加载完整 diff）
+
+### 2.3 环境修复持久化方案
+
+P3-4（环境特定修复不可持久化）在问题列表和痛点汇总中出现多次，但在 Phase 实施计划中无对应修复任务。建议在 Phase 2 或 3 增加 T-X.Y：「运维模式库：持久化环境修复脚本，支持跨重启自动应用」。
+
+### 2.4 AirEng 与 AirRvr 的集成协议
+
+AirRvr 的审查结果（pass / conditional-pass / fail）需要集成到 AirEng 的 merge 流程中，但文档未定义两者之间的接口协议。3.7.4 第 5 点仅概念性描述，缺少：
+
+- AirEng 如何调用 AirRvr（文件系统 / 函数调用）
+- AirRvr 审查的超时和失败处理（审查本身卡死怎么办）
+- 审查结果与修复预算的关系（fail 后重试几次）
+
+### 2.5 事件日志的保留策略
+
+3.7.5 的事件日志是 JSONL 格式无限追加，与 state.json 的 P2-2（列表无界增长）面临同样问题。建议定义事件日志的轮转/截断策略。
+
+### 2.6 安全扫描的误报处理
+
+AirSec（3.7.3）发现敏感数据时「阻止合并并通知用户」，但没有误报处理机制。静态扫描的正则匹配容易产生误报（如代码中的示例密钥、Base64 编码的二进制数据）。建议增加：
+
+- 白名单机制（已知安全的值、测试 fixture）
+- 人工确认流程（首次发现时通知用户判断，后续同模式自动放行）
+- AirSec 的判定为 advisory 而非 blocking（可配置）
+
+### 2.7 并发度可配置的具体机制
+
+T-2.7 提到「并发度可配置」，但未说明配置方式。是通过配置文件？环境变量？引擎命令行参数？`engine.py:527` 的硬编码常量替换为什么？建议明确。
+
+### 2.8 AirContext 压缩失败后的降级路径
+
+3.3.1 提到压缩验证失败后「重试一次，仍失败则放弃压缩并通知用户」。但放弃压缩意味着上下文持续增长直到超过模型窗口，最终导致任务失败。建议增加三级降级：
+
+1. 重试（换 prompt）
+2. 换模型压缩（如切换到更便宜的模型做摘要）
+3. 激进截断（保留最近 N 轮 + ADR 引用，丢弃中间轮次）
+
+---
+
+## 三、架构层面评论
+
+### 3.1 Prompt 指令 vs 代码保障的边界
+
+文档中有多处将 prompt/SKILL.md 文本指令作为 P0 级缺陷的主要修复手段（P0-5 AirArc 阻断、P0-6 中文锁定、P0-7 轮询循环、P0-8 AirDbg 强制路由）。这些 prompt 级修复在 LLM 遵循度足够高时有效，但不构成硬性保障。
+
+建议在文档中明确标注每项修复的保障层级：
+
+- **L1 代码级**：由引擎代码强制执行，不依赖 LLM 行为
+- **L2 指令级**：由 SKILL.md 指令约束，LLM 可能偏离
+- **L3 建议级**：最佳实践文档，无强制机制
+
+对于 P0 级缺陷，修复应至少达到 L1 或 L1+L2 双重保障。当前 P0-7 的修复仅为 L2，应升级。
+
+### 3.2 DAG 调度与人类可读性的平衡
+
+TaskGraph 替代静态 todo.md 表格是架构上正确的方向，但会牺牲 V1 的核心优势之一：任何人都可以用文本编辑器打开 todo.md 了解项目状态。建议在 DAG 之上保持 todo.md 作为导出视图（非调度数据源），确保人类可读性不丢失。
+
+### 3.3 Phase 优先级合理性
+
+Phase 1 集中修复 P0 缺陷，方向正确。但 T-1.10~T-1.13 均为 prompt 指令修改（0.5d 每项），实际工作量可能被低估——prompt 调优往往需要多轮测试验证，而非一次性编写。建议将 0.5d 调整为 1d 或合并为 2 个任务（AirArc/AirEng 指令重写 + AirDo/AirDbg 路由重写）。
+
+---
+
+## 四、审查总结
+
+| 类别 | 数量 | 严重程度 |
+|------|------|---------|
+| 需修正（事实性错误或设计缺陷） | 12 | 1.1/1.2 为事实错误，其余为设计需要完善 |
+| 需补充（缺失内容） | 8 | 影响完整性和可实施性 |
+| 架构评论 | 3 | 建议性，非阻塞 |
+
+**总体评价**：文档对 V1 问题的诊断全面准确，V2 的架构改进方向正确且设计深度足够。12 项修正是实施前应解决的前置条件，8 项补充可在实施过程中逐步完善。推荐在修正 1.1~1.12 后进入 Phase 1 实施。
--- a/commands/airplan.md
+++ b/commands/airplan.md
@@ -0,0 +1,46 @@
+---
+description: "AirPlan V2 - unified scheduler with 12 modes"
+argument-hint: "[arc|eng|do|dbg|xdb|sdb|ndb|ctx|dep|tst|sec|rvr] [sub-command]"
+allowed-tools: "[Read, Glob, Grep, Bash, Write, Edit]"
+---
+
+# /airplan
+
+AirPlan V2 - unified scheduler. V1 8 plugins merged into 1, adds 4 components, total 12 sub-modes.
+
+## Sub-modes
+
+| Mode | Role | Key Files |
+|------|------|-----------|
+| arc | Architecture planner | execution-plan.json, task-graph.json |
+| eng | Scheduler engine | state/aireng/state.json |
+| do | Task executor | state/airdo/tasks/{task_id}/result.json |
+| dbg | Debugger | state/airdbg/sessions/*.json |
+| xdb | GUI validator | state/airxdb/artifacts/ |
+| sdb | Static analyzer | state/airsdb/reports/ |
+| ndb | Network debugger | state/airndb/captures/ |
+| ctx | Context manager | state/aircontext/ |
+| dep | Deployer | state/airdep/sessions/ |
+| tst | Test runner | state/airtst/reports/ |
+| sec | Security scanner | state/airsec/ |
+| rvr | Requirements reviewer | state/airrvr/reviews/ |
+
+## Usage
+
+1. Parse $ARGUMENTS; default to "status" when empty.
+2. Run from project root:
+
+```bash
+python "$HOME/plugins/airplan/scripts/airplan.py" --mode <MODE> --project . [OPTIONS]
+```
+
+3. Common sub-commands:
+   - arc: --sub enter|status|parallel-review|incremental-replan --todo AirPlan/todo.md
+   - eng: --sub enter|status|plan|dispatch|monitor|merge|intervene
+   - do: --sub enter|status|finish --task-id <id> --result <path>
+   - dbg: --sub start|snapshot --task-id <id>
+   - dep: --sub deploy --task-id <id> --host <host> --binary <path>
+   - tst: --sub run --task-id <id> --framework <name>
+   - sec: --sub scan --task-id <id> --scan-path <path>
+
+4. Runtime auto-bootstraps missing AirPlan/ files.
--- a/commands/arc.md
+++ b/commands/arc.md
@@ -0,0 +1,49 @@
+---
+description: "AirPlan arc - architecture planner, READ-ONLY, never writes code"
+argument-hint: "[enter|status|parallel-review|incremental-replan]"
+allowed-tools: "[Read, Glob, Grep]"
+deny-plan-mode: true
+---
+
+# /arc
+
+AirArc is a pure architecture planner. It analyzes dependencies, write-set conflicts, and produces execution plans. It NEVER writes code, modifies source files, or enters plan mode.
+
+## Hard Rules (violated = bug)
+
+1. **READ-ONLY** — you only Read, Glob, Grep. Never Write, Edit, Bash.
+2. **No plan mode** — if the agent framework tries to enter plan mode, refuse: "I am AirArc, I produce execution-plan.json, not code changes."
+3. **Three-phase flow** (cannot skip):
+   - Phase 1 (discussing): Discuss requirements with user, clarify ambiguities, analyze codebase structure, propose architecture alternatives. **Forbidden to write execution-plan.json.**
+   - Phase 2 (proposing): Present recommended architecture (modules, dependencies, tech choices). Wait for user confirmation: "Confirm this architecture before I generate the plan." User objections return to Phase 1. **Forbidden to write execution-plan.json.**
+   - Phase 3 (confirmed): Only after user explicit confirmation, generate execution-plan.json and task-graph.json.
+
+## Sub-commands
+
+### enter
+
+```bash
+python "$HOME/plugins/airplan/scripts/airplan.py" --mode arc --project . --sub enter
+```
+
+### status
+
+```bash
+python "$HOME/plugins/airplan/scripts/airplan.py" --mode arc --project . --sub status
+```
+
+### parallel-review
+
+```bash
+python "$HOME/plugins/airplan/scripts/airplan.py" --mode arc --project . --sub parallel-review --todo AirPlan/todo.md
+```
+
+### incremental-replan
+
+```bash
+python "$HOME/plugins/airplan/scripts/airplan.py" --mode arc --project . --sub incremental-replan --todo AirPlan/todo.md
+```
+
+## Logging Standard
+
+When planning C++ projects, the first task MUST be "integrate spdlog" if not already present. All generated code must use spdlog, not std::cout/qDebug/printf.
--- a/commands/eng.md
+++ b/commands/eng.md
@@ -0,0 +1,87 @@
+---
+description: "AirPlan eng - scheduler engine, dispatches isolated workers, never codes directly"
+argument-hint: "[run|status|plan|dispatch|monitor|merge|intervene]"
+allowed-tools: "[Read, Glob, Grep, Bash, Write, Edit]"
+---
+
+# /eng
+
+AirEng 是调度引擎，不是执行器。它派发隔离 Worker、监控、合并结果。绝不直接实现代码。
+
+## 硬规则（违反=bug）
+
+1. **Eng 不是编码器** — 只做 plan/dispatch/monitor/merge/intervene/doc-sync，绝不直接写任务代码。
+2. **必须通过 /do 派发** — 每个任务必须 spawn 隔离的 /do 子代理（fork_context=false）。
+3. **保持父线程精简** — 父线程只做调度操作。
+4. **直接写代码唯一例外** — Worker 硬阻塞无法自恢复时的紧急干预，干预后立即回到调度模式。
+5. **必须使用中文** — 所有状态报告、进度通知、问题描述均使用中文。禁止英文输出。
+6. **自主决策原则** — 以推进开发进度为第一目标，以下情况自行决策不停下来问用户：
+   - Worker blocked 但修复预算未耗尽 → 自行派发修复
+   - Worker 停滞 → 自行执行停滞干预
+   - 验证失败但非关键 → 记录问题继续下一任务
+   - 波次间衔接 → 自行启动下一波次
+7. **仅以下情况才询问用户**：修复预算耗尽且任务仍 blocked；需求歧义无法继续；系统资源耗尽；用户显式暂停。
+
+## 子命令
+
+### run (默认) — 一次完整调度步骤
+
+```bash
+python "$HOME/plugins/airplan/scripts/airplan.py" --mode eng --project . --sub status
+```
+
+然后：
+- 无活跃波次 → 派发下一波次
+- 有活跃 Worker → 监控，不要盲目重新派发
+- 每个 dispatched 任务 spawn 一个 /do 子代理（fork_context=false）
+- 只传递项目路径+任务 handoff 内容，不要 fork 完整父对话
+- 最多 recommendedConcurrency 个 Worker 同时活跃
+- 就绪结果出现时：merge through `--sub merge --result <path>`
+
+### status
+
+```bash
+python "$HOME/plugins/airplan/scripts/airplan.py" --mode eng --project . --sub status
+```
+
+### plan
+
+```bash
+python "$HOME/plugins/airplan/scripts/airplan.py" --mode eng --project . --sub plan --todo AirPlan/todo.md
+```
+
+### dispatch
+
+```bash
+python "$HOME/plugins/airplan/scripts/airplan.py" --mode eng --project . --sub dispatch
+```
+
+读取 AirPlan/state/aireng/dispatch/ 中的派发清单，为每个任务 spawn 一个隔离的 /do Worker 子代理。
+
+### monitor
+
+```bash
+python "$HOME/plugins/airplan/scripts/airplan.py" --mode eng --project . --sub monitor
+```
+
+合并就绪结果，检测停滞 Worker，更新 nextAction。
+
+### merge
+
+```bash
+python "$HOME/plugins/airplan/scripts/airplan.py" --mode eng --project . --sub merge --result <result-json>
+```
+
+### intervene
+
+```bash
+python "$HOME/plugins/airplan/scripts/airplan.py" --mode eng --project . --sub intervene
+```
+
+仅用于无法通过常规监控或重新派发解决的硬阻塞。
+
+## 无人值守模式
+
+- 每 300 秒重新检查活跃 Worker
+- 当前波次收敛后自动派发下一波次
+- 仅在状态达到 completed 或用户决策阻塞时停止
--- a/lib/air_runtime/init.py
+++ b/lib/air_runtime/init.py
@@ -0,0 +1,20 @@
+# air_runtime — AirPlan V2 核心运行时库
+
+from air_runtime.io import atomic_json_write, safe_json_load
+from air_runtime.lock import FileLock
+from air_runtime.utils import ordered_unique, session_stamp, normalize_policy, sanitize_task_id, sanitize_marker
+from air_runtime.events import EventLog
+from air_runtime.task_graph import TaskGraph, PlanDelta
+from air_runtime.worktree import WorktreeIsolation, RegionConflictDetector, ConflictLevel
+from air_runtime.evidence_gate import EvidenceGatePolicy, EvidenceClass
+from air_runtime.contracts import now_iso, WorkerResult, TaskRecord, DeploymentRecord
+from air_runtime.paths import (
+    airplan_root, todo_path, plan_path,
+    state_root, engine_state_path, worker_state_path,
+    arc_state_path, dbg_state_path, xdb_state_path,
+    sdb_state_path, ndb_state_path, ctx_state_path,
+    dep_state_path, tst_state_path, sec_state_path, rvr_state_path,
+    required_project_artifacts,
+)
+
+__version__ = "2.0.0"
--- a/lib/air_runtime/pycache/init.cpython-313.pyc
+++ b/lib/air_runtime/pycache/init.cpython-313.pyc
--- a/lib/air_runtime/pycache/contracts.cpython-313.pyc
+++ b/lib/air_runtime/pycache/contracts.cpython-313.pyc
--- a/lib/air_runtime/pycache/deploy_runtime.cpython-313.pyc
+++ b/lib/air_runtime/pycache/deploy_runtime.cpython-313.pyc
--- a/lib/air_runtime/pycache/events.cpython-313.pyc
+++ b/lib/air_runtime/pycache/events.cpython-313.pyc
--- a/lib/air_runtime/pycache/evidence_gate.cpython-313.pyc
+++ b/lib/air_runtime/pycache/evidence_gate.cpython-313.pyc
--- a/lib/air_runtime/pycache/installer.cpython-313.pyc
+++ b/lib/air_runtime/pycache/installer.cpython-313.pyc
--- a/lib/air_runtime/pycache/io.cpython-313.pyc
+++ b/lib/air_runtime/pycache/io.cpython-313.pyc
--- a/lib/air_runtime/pycache/lock.cpython-313.pyc
+++ b/lib/air_runtime/pycache/lock.cpython-313.pyc
--- a/lib/air_runtime/pycache/paths.cpython-313.pyc
+++ b/lib/air_runtime/pycache/paths.cpython-313.pyc
--- a/lib/air_runtime/pycache/project_bootstrap.cpython-313.pyc
+++ b/lib/air_runtime/pycache/project_bootstrap.cpython-313.pyc
--- a/lib/air_runtime/pycache/review.cpython-313.pyc
+++ b/lib/air_runtime/pycache/review.cpython-313.pyc
--- a/lib/air_runtime/pycache/review_runtime.cpython-313.pyc
+++ b/lib/air_runtime/pycache/review_runtime.cpython-313.pyc
--- a/lib/air_runtime/pycache/sdb_backends.cpython-313.pyc
+++ b/lib/air_runtime/pycache/sdb_backends.cpython-313.pyc
--- a/lib/air_runtime/pycache/sec_runtime.cpython-313.pyc
+++ b/lib/air_runtime/pycache/sec_runtime.cpython-313.pyc
--- a/lib/air_runtime/pycache/task_graph.cpython-313.pyc
+++ b/lib/air_runtime/pycache/task_graph.cpython-313.pyc
--- a/lib/air_runtime/pycache/test_runtime.cpython-313.pyc
+++ b/lib/air_runtime/pycache/test_runtime.cpython-313.pyc
--- a/lib/air_runtime/pycache/todo_parser.cpython-313.pyc
+++ b/lib/air_runtime/pycache/todo_parser.cpython-313.pyc
--- a/lib/air_runtime/pycache/utils.cpython-313.pyc
+++ b/lib/air_runtime/pycache/utils.cpython-313.pyc
--- a/lib/air_runtime/pycache/worktree.cpython-313.pyc
+++ b/lib/air_runtime/pycache/worktree.cpython-313.pyc
--- a/lib/air_runtime/pycache/xdb_capture.cpython-313.pyc
+++ b/lib/air_runtime/pycache/xdb_capture.cpython-313.pyc
--- a/lib/air_runtime/contracts.py
+++ b/lib/air_runtime/contracts.py
@@ -0,0 +1,109 @@
+"""
+数据契约 — 保持 V1 契约完整性，新增 DeploymentRecord、AirRvr 审查报告等结构。
+"""
+
+from __future__ import annotations
+
+import re
+from dataclasses import dataclass, field
+from datetime import datetime, timezone
+from typing import Any
+
+
+def now_iso() -> str:
+    return datetime.now(timezone.utc).isoformat()
+
+
+@dataclass
+class WorkerResult:
+    task_id: str
+    status: str  # done | blocked | failed
+    summary: str = ""
+    files_changed: list[str] = field(default_factory=list)
+    validations: list[dict] = field(default_factory=list)
+    document_updates: list[dict] = field(default_factory=list)
+    evidence: list[dict] = field(default_factory=list)
+    risks: list[str] = field(default_factory=list)
+    blockers: list[str] = field(default_factory=list)
+    deploy_required: bool = False
+    deploy_info: dict | None = None
+
+    def validate_for_finalize(self, brief: dict | None = None) -> None:
+        if not re.fullmatch(r"[A-Za-z0-9_\-\.]+", self.task_id):
+            raise ValidationError(f"invalid task_id: {self.task_id!r}")
+
+        if self.status not in ("done", "blocked", "failed"):
+            raise ValidationError(f"invalid status: {self.status}")
+
+        if self.status == "done" and not self.validations and not self.files_changed:
+            raise ValidationError("done without validations or file changes")
+
+        if self.deploy_required:
+            deploy_validations = [
+                v for v in self.validations
+                if v.get("kind") in ("remote-deploy-verify", "remote-binary-md5")
+            ]
+            if not deploy_validations:
+                raise ValidationError("deploy_required but no deploy verification in validations")
+
+    def to_dict(self) -> dict[str, Any]:
+        return {
+            "taskId": self.task_id,
+            "status": self.status,
+            "summary": self.summary,
+            "filesChanged": self.files_changed,
+            "validations": self.validations,
+            "documentUpdates": self.document_updates,
+            "evidence": self.evidence,
+            "risks": self.risks,
+            "blockers": self.blockers,
+            "deployRequired": self.deploy_required,
+            "deployInfo": self.deploy_info,
+        }
+
+    @classmethod
+    def from_dict(cls, data: dict) -> WorkerResult:
+        return cls(
+            task_id=data.get("taskId", ""),
+            status=data.get("status", ""),
+            summary=data.get("summary", ""),
+            files_changed=data.get("filesChanged", []),
+            validations=data.get("validations", []),
+            document_updates=data.get("documentUpdates", []),
+            evidence=data.get("evidence", []),
+            risks=data.get("risks", []),
+            blockers=data.get("blockers", []),
+            deploy_required=data.get("deployRequired", False),
+            deploy_info=data.get("deployInfo"),
+        )
+
+
+@dataclass
+class TaskRecord:
+    task_id: str
+    task: str = ""
+    files_dirs: str = ""
+    done_when: str = ""
+    status: str = "TODO"
+    validations: str = ""
+    adr: str = ""
+
+    @property
+    def text_for_classification(self) -> str:
+        return f"{self.task} {self.files_dirs} {self.done_when}"
+
+
+@dataclass
+class DeploymentRecord:
+    task_id: str
+    host: str
+    binary_path: str
+    md5: str = ""
+    service_name: str = ""
+    service_status: str = ""
+    deploy_at: str = field(default_factory=now_iso)
+    smoke_test_passed: bool | None = None
+
+
+class ValidationError(Exception):
+    pass
--- a/lib/air_runtime/deploy_runtime.py
+++ b/lib/air_runtime/deploy_runtime.py
@@ -0,0 +1,133 @@
+"""
+AirDep 部署运行时 — V2 新增组件。
+SSH 远程构建 + 二进制传输 + systemd 生命周期管理 + 部署验证。
+"""
+
+from __future__ import annotations
+
+import hashlib
+import subprocess
+from dataclasses import dataclass
+from pathlib import Path
+
+from air_runtime.contracts import DeploymentRecord, now_iso
+from air_runtime.io import atomic_json_write, safe_json_load
+from air_runtime.paths import dep_state_path
+
+
+@dataclass
+class DeployTarget:
+    host: str
+    user: str = "root"
+    port: int = 22
+    build_dir: str = "/tmp/airdep-build"
+    deploy_dir: str = "/opt/app"
+
+
+@dataclass
+class DeployResult:
+    task_id: str
+    success: bool
+    binary_md5: str = ""
+    service_status: str = ""
+    error: str = ""
+    journal_excerpt: str = ""
+
+
+def _ssh(target: DeployTarget, cmd: str, timeout: int = 120) -> subprocess.CompletedProcess[str]:
+    return subprocess.run(
+        ["ssh", "-o", "StrictHostKeyChecking=accept-new", "-p", str(target.port),
+         f"{target.user}@{target.host}", cmd],
+        capture_output=True, text=True, timeout=timeout,
+    )
+
+
+def _scp(target: DeployTarget, local: Path, remote: str, timeout: int = 300) -> subprocess.CompletedProcess[str]:
+    return subprocess.run(
+        ["scp", "-o", "StrictHostKeyChecking=accept-new", "-P", str(target.port),
+         str(local), f"{target.user}@{target.host}:{remote}"],
+        capture_output=True, text=True, timeout=timeout,
+    )
+
+
+def _md5_file(path: Path) -> str:
+    h = hashlib.md5()
+    with open(path, "rb") as f:
+        while chunk := f.read(8192):
+            h.update(chunk)
+    return h.hexdigest()
+
+
+def deploy(task_id: str, project_root: Path, target: DeployTarget,
+           local_binary: Path, service_name: str,
+           build_cmd: str | None = None,
+           smoke_test_cmd: str | None = None) -> DeployResult:
+    """执行完整部署流程：构建 → 传输 → systemd → 验证。"""
+    state_file = dep_state_path(project_root)
+
+    try:
+        # Step 1: 远程构建（可选）
+        if build_cmd:
+            result = _ssh(target, f"cd {target.build_dir} && {build_cmd}", timeout=600)
+            if result.returncode != 0:
+                return DeployResult(task_id=task_id, success=False,
+                                   error=f"build failed: {result.stderr[-500:]}")
+
+        # Step 2: 二进制传输 + MD5 校验
+        remote_tmp = f"/tmp/{local_binary.name}"
+        scp_result = _scp(target, local_binary, remote_tmp)
+        if scp_result.returncode != 0:
+            return DeployResult(task_id=task_id, success=False,
+                               error=f"scp failed: {scp_result.stderr[-500:]}")
+
+        local_md5 = _md5_file(local_binary)
+        md5_result = _ssh(target, f"md5sum {remote_tmp} | cut -d' ' -f1")
+        remote_md5 = md5_result.stdout.strip()
+        if local_md5 != remote_md5:
+            return DeployResult(task_id=task_id, success=False,
+                               binary_md5=f"local={local_md5} remote={remote_md5}",
+                               error="md5 mismatch after transfer")
+
+        # Step 3: 部署二进制
+        _ssh(target, f"mv {remote_tmp} {target.deploy_dir}/{local_binary.name}")
+
+        # Step 4: systemd 生命周期
+        _ssh(target, f"systemctl daemon-reload")
+        _ssh(target, f"systemctl restart {service_name}")
+        status_result = _ssh(target, f"systemctl is-active {service_name}")
+        service_status = status_result.stdout.strip()
+
+        # Step 5: 冒烟验证（可选）
+        smoke_passed: bool | None = None
+        if smoke_test_cmd:
+            smoke_result = _ssh(target, smoke_test_cmd, timeout=60)
+            smoke_passed = (smoke_result.returncode == 0)
+
+        # Step 6: 记录部署产物
+        record = DeploymentRecord(
+            task_id=task_id,
+            host=target.host,
+            binary_path=str(target.deploy_dir / local_binary.name),
+            md5=local_md5,
+            service_name=service_name,
+            service_status=service_status,
+            smoke_test_passed=smoke_passed,
+        )
+
+        # 持久化
+        sessions = safe_json_load(state_file) or {"sessions": []}
+        if isinstance(sessions, dict):
+            sessions.setdefault("sessions", []).append(record.__dict__)
+        atomic_json_write(state_file, sessions)
+
+        return DeployResult(
+            task_id=task_id,
+            success=service_status == "active",
+            binary_md5=local_md5,
+            service_status=service_status,
+        )
+
+    except subprocess.TimeoutExpired as exc:
+        return DeployResult(task_id=task_id, success=False, error=f"timeout: {exc}")
+    except Exception as exc:
+        return DeployResult(task_id=task_id, success=False, error=str(exc))
--- a/lib/air_runtime/events.py
+++ b/lib/air_runtime/events.py
@@ -0,0 +1,102 @@
+"""
+事件日志模块 — V2 可观测性基础设施。
+JSONL 格式无限流式追加，与 state.json 互补：state.json 是当前快照，事件日志是完整时间线。
+"""
+
+from __future__ import annotations
+
+import contextlib
+import json
+import logging
+import os
+import tempfile
+from datetime import datetime, timezone
+from pathlib import Path
+
+logger = logging.getLogger(__name__)
+
+# 事件类型常量
+TASK_DISPATCHED = "task.dispatched"
+TASK_COMPLETED = "task.completed"
+TASK_BLOCKED = "task.blocked"
+MERGE_STARTED = "merge.started"
+MERGE_COMPLETED = "merge.completed"
+REPAIR_CREATED = "repair.created"
+REPAIR_RESOLVED = "repair.resolved"
+INTERVENTION_STALL = "intervention.stall"
+XDB_CAPTURED = "xdb.captured"
+DEBUG_SESSION = "debug.session"
+CONTEXT_COMPACTED = "context.compacted"
+DEPLOY_COMPLETED = "deploy.completed"
+TEST_RUN = "test.run"
+SEC_SCAN = "sec.scan"
+REVIEW_SESSION = "review.session"
+ENGINE_CYCLE = "engine.cycle"
+WORKER_TIMEOUT = "worker.timeout"
+LOCK_ACQUIRED = "lock.acquired"
+LOCK_RELEASED = "lock.released"
+STALE_LOCK_CLEANED = "stale_lock.cleaned"
+
+
+def now_iso() -> str:
+    return datetime.now(timezone.utc).isoformat()
+
+
+class EventLog:
+    """结构化事件日志（JSONL 格式），支持轮转截断。"""
+
+    MAX_LINES = 10000
+
+    def __init__(self, path: Path, max_lines: int = MAX_LINES):
+        self._path = path
+        self._max_lines = max_lines
+        self._pending_merge_complete = None
+
+    def emit(self, event_type: str, payload: dict | None = None) -> None:
+        entry = {
+            "ts": now_iso(),
+            "type": event_type,
+            **(payload or {}),
+        }
+        self._path.parent.mkdir(parents=True, exist_ok=True)
+        with open(self._path, "a") as f:
+            f.write(json.dumps(entry, ensure_ascii=False) + "\n")
+        self._maybe_rotate()
+
+    def _emit_with_completion(self, event_type: str, payload: dict) -> None:
+        """emit 并确保 MERGE_STARTED / MERGE_COMPLETED 成对。"""
+        self.emit(event_type, payload)
+
+        # 自动补全配对事件
+        if event_type == "merge.started":
+            self._pending_merge_complete = payload.get("taskId")
+        elif event_type == "merge.completed":
+            self._pending_merge_complete = None
+
+    def _maybe_rotate(self) -> None:
+        """原子轮转：先写临时文件，再 os.replace。"""
+        if not self._path.exists():
+            return
+        try:
+            with open(self._path, "r", encoding="utf-8", errors="replace") as f:
+                lines = f.readlines()
+            if len(lines) <= self._max_lines:
+                return
+
+            keep_count = self._max_lines // 2
+            kept_lines = lines[-keep_count:] if len(lines) > keep_count else lines
+
+            fd, tmp_path = tempfile.mkstemp(dir=self._path.parent, suffix=".tmp")
+            os.close(fd)
+
+            try:
+                with open(tmp_path, "w", encoding="utf-8") as f:
+                    f.writelines(kept_lines)
+                os.replace(tmp_path, self._path)
+                logger.info("rotated event log %s, kept %d/%d lines", self._path, len(kept_lines), len(lines))
+            except Exception:
+                with contextlib.suppress(Exception):
+                    os.unlink(tmp_path)
+                raise
+        except OSError:
+            pass
--- a/lib/air_runtime/evidence_gate.py
+++ b/lib/air_runtime/evidence_gate.py
@@ -0,0 +1,74 @@
+"""
+任务类型感知的证据门控 — V2 P0-1 修复。
+替代 V1 无差别触发 AirXDB 的逻辑，根据任务特征差异化要求证据类型。
+"""
+
+from __future__ import annotations
+
+import enum
+from dataclasses import dataclass
+
+
+class EvidenceClass(enum.Enum):
+    GUI_REQUIRED = "gui_required"          # 需要截图/GUI 操作验证
+    NETWORK_REQUIRED = "network_required"   # 需要抓包/流量验证
+    STATIC_ANALYSIS = "static_analysis"    # 需要静态分析
+    CODE_ONLY = "code_only"                # 代码级验证即可
+
+
+@dataclass
+class EvidenceGatePolicy:
+    """基于任务特征的差异化证据要求。
+
+    关键词匹配 → 确定性分类，零额外 API 调用，可测试可覆盖。
+    用户可在 todo.md 中用 [no-xdb] / [no-ndb] 标记显式跳过。
+    """
+
+    GUI_INDICATORS: tuple[str, ...] = (
+        "gui", "ui", "render", "layout", "dialog", "osd",
+        "overlay", "visual", "screenshot", "display",
+        "widget", "pane", "toolbar", "settings_dialog",
+        "canvas", "button", "window", "popup", "menu",
+        "drm", "kms", "opengl", "vulkan",
+    )
+
+    NETWORK_INDICATORS: tuple[str, ...] = (
+        "network", "rtsp", "http", "tcp", "udp", "tls",
+        "dns", "proxy", "socket", "stream", "port",
+        "packet", "pcap", "bandwidth", "latency",
+    )
+
+    STATIC_ANALYSIS_INDICATORS: tuple[str, ...] = (
+        "cppcheck", "clang-tidy", "mypy", "ruff", "lint",
+        "static analysis", "compile_commands",
+    )
+
+    SKIP_MARKERS: tuple[str, ...] = ("[no-xdb]", "[no-ndb]", "[no-sdb]")
+
+    def classify(self, task_text: str, task_id: str = "", skip_overrides: list[str] | None = None) -> EvidenceClass:
+        text = task_text.lower()
+        skips = set(skip_overrides or [])
+        for marker in self.SKIP_MARKERS:
+            if marker in text:
+                skips.add(marker)
+
+        has_gui = any(kw in text for kw in self.GUI_INDICATORS)
+        has_net = any(kw in text for kw in self.NETWORK_INDICATORS)
+        has_static = any(kw in text for kw in self.STATIC_ANALYSIS_INDICATORS)
+
+        if has_gui and "[no-xdb]" not in skips:
+            return EvidenceClass.GUI_REQUIRED
+        if has_net and "[no-ndb]" not in skips:
+            return EvidenceClass.NETWORK_REQUIRED
+        if has_static and "[no-sdb]" not in skips:
+            return EvidenceClass.STATIC_ANALYSIS
+        return EvidenceClass.CODE_ONLY
+
+    def required_evidence_for(self, evidence_class: EvidenceClass) -> list[str]:
+        evidence_map = {
+            EvidenceClass.GUI_REQUIRED: ["screenshot", "gui-operation-validation"],
+            EvidenceClass.NETWORK_REQUIRED: ["packet-capture", "connectivity-verification"],
+            EvidenceClass.STATIC_ANALYSIS: ["static-analysis-report"],
+            EvidenceClass.CODE_ONLY: ["code-review", "test-results"],
+        }
+        return evidence_map.get(evidence_class, ["code-review"])
--- a/lib/air_runtime/installer.py
+++ b/lib/air_runtime/installer.py
@@ -0,0 +1,121 @@
+from __future__ import annotations
+import json
+import shutil
+from pathlib import Path
+
+
+def resolve_plugin_paths() -> dict[str, Path]:
+    """
+    解析插件的实际安装路径。
+    顺序: 1) 环境变量 AIRPLAN_HOME > 2) ~/.airplan > 3) 相对脚本位置推测
+    """
+    # 1. 环境变量
+    if "AIRPLAN_HOME" in __import__("os").environ:
+        return {"root": Path(__import__("os").environ["AIRPLAN_HOME"])}
+
+    # 2. ~/.airplan 默认
+    home = Path.home()
+    default = home / ".airplan"
+    if default.exists():
+        return {"root": default}
+
+    # 3. 尝试从当前脚本位置推测
+    # 脚本位于 {root}/scripts/airplan.py
+    import sys
+    script_root = Path(sys.argv[0]).resolve().parent if sys.argv else None
+    if script_root and (script_root.name == "scripts"):
+        plugin_root = script_root.parent
+        if (plugin_root / ".claude-plugin").exists():
+            return {"root": plugin_root}
+
+    # 4. 从本模块位置推测
+    # 本模块位于 {root}/lib/air_runtime/installer.py
+    module_root = Path(__file__).resolve().parent.parent.parent  # lib/air_runtime/installer -> lib/air_runtime -> lib -> root
+    if (module_root / ".claude-plugin").exists():
+        return {"root": module_root}
+
+    raise FileNotFoundError("Cannot locate plugin installation directory")
+
+
+def get_plugin_meta() -> dict:
+    """读取 plugin.json 元数据"""
+    paths = resolve_plugin_paths()
+    meta_path = paths["root"] / ".claude-plugin" / "plugin.json"
+    if not meta_path.exists():
+        raise FileNotFoundError(f"plugin.json not found at {meta_path}")
+    return json.loads(meta_path.read_text())
+
+
+def post_install_verify() -> dict:
+    """
+    安装后验证：检查所有依赖工具是否存在。
+    返回 {tool: bool} 映射，False 表示缺失。
+    """
+    required_tools = [
+        "git",           # 版本控制
+        "cmake",         # 构建
+        "ffmpeg",        # XDB 截图
+        "xvfb-run",      # XDB 虚拟显示
+    ]
+
+    result = {"ok": True, "missing": []}
+    for tool in required_tools:
+        found = shutil.which(tool) is not None
+        result[tool] = found
+        if not found:
+            result["ok"] = False
+            result["missing"].append(tool)
+
+    # 检查 Python 依赖
+    try:
+        import yaml
+        result["pyyaml"] = True
+    except ImportError:
+        result["pyyaml"] = False
+        result["ok"] = False
+        result["missing"].append("pyyaml")
+
+    # 检查目录结构
+    paths = resolve_plugin_paths()
+    for subdir in ["lib", "scripts", "skills", "commands"]:
+        p = paths["root"] / subdir
+        result[f"dir_{subdir}"] = p.exists()
+        if not p.exists():
+            result["ok"] = False
+            result["missing"].append(f"dir:{subdir}")
+
+    return result
+
+
+def verify_plugin_json_paths() -> dict:
+    """
+    验证 plugin.json 里的路径是否可解析。
+    设计原文 P0-9 指出硬编码 $HOME 是问题，这里修复它。
+    """
+    meta = get_plugin_meta()
+    issues = []
+
+    # 检查 scripts 路径
+    entry = meta.get("entry", {})
+    if isinstance(entry, dict):
+        script_path_str = entry.get("args", [""])[0] if entry.get("args") else ""
+    else:
+        script_path_str = str(entry)
+    if "$HOME" in script_path_str:
+        # 尝试解析
+        resolved = script_path_str.replace("$HOME", str(Path.home()))
+        if not Path(resolved).exists():
+            issues.append(f"script path not found: {resolved}")
+        else:
+            # 修复：改用相对路径或 AIRPLAN_HOME 变量
+            issues.append(f"script uses $HOME: {script_path_str} (should use relative path)")
+
+    # 检查 marketplace 路径
+    market_path = meta.get("marketplace", "")
+    if "$HOME" in market_path:
+        issues.append(f"marketplace uses $HOME: {market_path}")
+
+    return {
+        "issues": issues,
+        "resolved_paths": resolve_plugin_paths(),
+    }
--- a/lib/air_runtime/io.py
+++ b/lib/air_runtime/io.py
@@ -0,0 +1,53 @@
+"""
+原子 I/O 模块 — 消除 V1 5 份 _json_dump/_json_load 重复。
+每次写入使用 tempfile + os.replace() 保证原子性，写入前自动备份。
+"""
+
+from __future__ import annotations
+
+import contextlib
+import json
+import logging
+import os
+import tempfile
+from pathlib import Path
+
+logger = logging.getLogger(__name__)
+
+
+def atomic_json_write(path: Path, data: dict | list, indent: int = 2) -> None:
+    """POSIX 原子写入：tempfile + os.replace()。写入前自动备份旧文件为 .bak（单级轮转）。"""
+    path.parent.mkdir(parents=True, exist_ok=True)
+
+    # 备份旧文件
+    if path.exists():
+        bak = path.with_suffix(path.suffix + ".bak")
+        try:
+            os.replace(str(path), str(bak))
+        except OSError as exc:
+            logger.warning("backup %s -> %s failed: %s", path, bak, exc)
+
+    fd, tmp = tempfile.mkstemp(dir=path.parent, suffix=".tmp")
+    try:
+        os.write(fd, json.dumps(data, indent=indent, ensure_ascii=False).encode("utf-8"))
+        os.close(fd)
+        os.replace(tmp, path)
+    except BaseException:
+        with contextlib.suppress(OSError):
+            os.unlink(tmp)
+        raise
+
+
+def safe_json_load(path: Path) -> dict | list | None:
+    """安全加载：处理损坏文件，自动从 .bak 恢复。文件不存在返回 None。"""
+    try:
+        return json.loads(path.read_text("utf-8"))
+    except FileNotFoundError:
+        return None
+    except (json.JSONDecodeError, UnicodeDecodeError):
+        bak = path.with_suffix(path.suffix + ".bak")
+        if bak.exists():
+            logger.warning("corrupt %s, restoring from %s", path, bak)
+            return json.loads(bak.read_text("utf-8"))
+        logger.error("corrupt %s with no backup", path)
+        return None
--- a/lib/air_runtime/lock.py
+++ b/lib/air_runtime/lock.py
@@ -0,0 +1,44 @@
+"""
+文件级并发控制 — 解决 V1 P0-4 零并发控制问题。
+基于 fcntl.flock 的进程级文件锁，超时自动释放。
+"""
+
+from __future__ import annotations
+
+import fcntl
+import os
+import time
+from pathlib import Path
+
+
+class FileLock:
+    """基于 fcntl.flock(LOCK_EX | LOCK_NB) 的进程级文件锁"""
+
+    def __init__(self, path: Path, timeout: float = 10.0):
+        self._path = path.with_suffix(path.suffix + ".lock") if not path.suffix.endswith(".lock") else path
+        self._timeout = timeout
+        self._fd: int | None = None
+
+    def __enter__(self) -> FileLock:
+        self._fd = os.open(self._path, os.O_CREAT | os.O_RDWR)
+        deadline = time.monotonic() + self._timeout
+        while True:
+            try:
+                fcntl.flock(self._fd, fcntl.LOCK_EX | fcntl.LOCK_NB)
+                return self
+            except OSError:
+                if time.monotonic() >= deadline:
+                    os.close(self._fd)
+                    self._fd = None
+                    raise TimeoutError(f"lock timeout after {self._timeout}s: {self._path}")
+                time.sleep(0.1)
+
+    def __exit__(self, *exc) -> None:
+        if self._fd is not None:
+            fcntl.flock(self._fd, fcntl.LOCK_UN)
+            os.close(self._fd)
+            self._fd = None
+
+    @property
+    def path(self) -> Path:
+        return self._path
--- a/lib/air_runtime/modes/init.py
+++ b/lib/air_runtime/modes/init.py
@@ -0,0 +1,11 @@
+"""模式模块 init"""
+
+from air_runtime.modes import arc_mode, eng_mode, do_mode, dbg_mode
+from air_runtime.modes import xdb_mode, sdb_mode, ndb_mode
+from air_runtime.modes import ctx_mode, dep_mode, tst_mode, sec_mode, rvr_mode
+
+__all__ = [
+    "arc_mode", "eng_mode", "do_mode", "dbg_mode",
+    "xdb_mode", "sdb_mode", "ndb_mode",
+    "ctx_mode", "dep_mode", "tst_mode", "sec_mode", "rvr_mode",
+]
--- a/lib/air_runtime/modes/pycache/init.cpython-313.pyc
+++ b/lib/air_runtime/modes/pycache/init.cpython-313.pyc
--- a/lib/air_runtime/modes/pycache/arc_mode.cpython-313.pyc
+++ b/lib/air_runtime/modes/pycache/arc_mode.cpython-313.pyc
--- a/lib/air_runtime/modes/pycache/ctx_mode.cpython-313.pyc
+++ b/lib/air_runtime/modes/pycache/ctx_mode.cpython-313.pyc
--- a/lib/air_runtime/modes/pycache/dbg_mode.cpython-313.pyc
+++ b/lib/air_runtime/modes/pycache/dbg_mode.cpython-313.pyc
--- a/lib/air_runtime/modes/pycache/dep_mode.cpython-313.pyc
+++ b/lib/air_runtime/modes/pycache/dep_mode.cpython-313.pyc
--- a/lib/air_runtime/modes/pycache/do_mode.cpython-313.pyc
+++ b/lib/air_runtime/modes/pycache/do_mode.cpython-313.pyc
--- a/lib/air_runtime/modes/pycache/eng_mode.cpython-313.pyc
+++ b/lib/air_runtime/modes/pycache/eng_mode.cpython-313.pyc
--- a/lib/air_runtime/modes/pycache/eng_orchestrator.cpython-313.pyc
+++ b/lib/air_runtime/modes/pycache/eng_orchestrator.cpython-313.pyc
--- a/lib/air_runtime/modes/pycache/merge_pipeline.cpython-313.pyc
+++ b/lib/air_runtime/modes/pycache/merge_pipeline.cpython-313.pyc
--- a/lib/air_runtime/modes/pycache/ndb_mode.cpython-313.pyc
+++ b/lib/air_runtime/modes/pycache/ndb_mode.cpython-313.pyc
--- a/lib/air_runtime/modes/pycache/rvr_mode.cpython-313.pyc
+++ b/lib/air_runtime/modes/pycache/rvr_mode.cpython-313.pyc
--- a/lib/air_runtime/modes/pycache/sdb_mode.cpython-313.pyc
+++ b/lib/air_runtime/modes/pycache/sdb_mode.cpython-313.pyc
--- a/lib/air_runtime/modes/pycache/sec_mode.cpython-313.pyc
+++ b/lib/air_runtime/modes/pycache/sec_mode.cpython-313.pyc
--- a/lib/air_runtime/modes/pycache/tst_mode.cpython-313.pyc
+++ b/lib/air_runtime/modes/pycache/tst_mode.cpython-313.pyc
--- a/lib/air_runtime/modes/pycache/xdb_mode.cpython-313.pyc
+++ b/lib/air_runtime/modes/pycache/xdb_mode.cpython-313.pyc
--- a/lib/air_runtime/modes/pycache/xdb_sdb_ndb_modes.cpython-313.pyc
+++ b/lib/air_runtime/modes/pycache/xdb_sdb_ndb_modes.cpython-313.pyc
--- a/lib/air_runtime/modes/arc_mode.py
+++ b/lib/air_runtime/modes/arc_mode.py
@@ -0,0 +1,258 @@
+#!/usr/bin/env python3
+"""
+AirArc mode — V2 架构规划器。
+L1 代码级保障：allowed-tools 限制为只读。
+产出 execution-plan.json（完整 DAG）+ plan-delta.json（增量重规划）。
+"""
+
+from __future__ import annotations
+
+import json
+import sys
+from pathlib import Path
+
+from air_runtime.contracts import now_iso
+from air_runtime.io import atomic_json_write, safe_json_load
+from air_runtime.paths import airplan_root, todo_path as get_todo_path
+from air_runtime.review import build_parallel_review, render_review_markdown
+from air_runtime.task_graph import TaskGraph, TaskNode, Edge, PlanDelta
+from air_runtime.events import EventLog
+from air_runtime.paths import event_log_path
+from air_runtime.utils import ordered_unique
+
+
+class ArcPhaseGate:
+    """三阶段门控：discussing → proposing → confirmed。
+    execution-plan.json 仅在 phase=confirmed 时允许写入。
+    """
+    PHASES = ["discussing", "proposing", "confirmed"]
+
+    def __init__(self, state_path: Path):
+        self._state_path = state_path
+
+    @property
+    def current_phase(self) -> str:
+        data = safe_json_load(self._state_path) or {}
+        return data.get("arcPhase", "discussing")
+
+    def advance_to(self, phase: str) -> None:
+        if phase not in self.PHASES:
+            raise ValueError(f"invalid phase: {phase!r}")
+        idx_current = self.PHASES.index(self.current_phase)
+        idx_target = self.PHASES.index(phase)
+        if idx_target <= idx_current:
+            return
+        data = safe_json_load(self._state_path) or {}
+        data["arcPhase"] = phase
+        data[f"arcPhase_{phase}At"] = now_iso()
+        atomic_json_write(self._state_path, data)
+
+    def can_write_plan(self) -> bool:
+        return self.current_phase == "confirmed"
+
+    def confirm_architecture(self, user_confirmation: str) -> bool:
+        """检查用户确认文本中的关键词，确认后推进到 confirmed。"""
+        confirm_keywords = ["确认", "可以", "同意", "confirm", "yes", "ok", "好的", "没问题"]
+        if any(kw in user_confirmation.lower() for kw in confirm_keywords):
+            self.advance_to("confirmed")
+            return True
+        return False
+
+
+def _paths(project_root: Path) -> dict[str, Path]:
+    root = airplan_root(project_root) / "state" / "airarc"
+    return {
+        "root": root,
+        "state": root / "state.json",
+        "reviews_dir": root / "reviews",
+        "execution_plan_json": root / "reviews" / "execution-plan.json",
+        "execution_plan_md": root / "reviews" / "execution-plan.md",
+        "plan_delta_json": root / "reviews" / "plan-delta.json",
+        "task_graph_json": root / "reviews" / "task-graph.json",
+    }
+
+
+def _ensure_dirs(paths: dict[str, Path]) -> None:
+    paths["reviews_dir"].mkdir(parents=True, exist_ok=True)
+
+
+def _export_task_graph_json(graph: TaskGraph, path: Path) -> None:
+    data = {
+        "nodes": {nid: {"id": n.id, "status": n.status, "task": n.task,
+                        "filesDirs": n.files_dirs, "doneWhen": n.done_when,
+                        "inDegree": n.in_degree, "outEdges": n.out_edges,
+                        "writeSet": n.write_set}
+                  for nid, n in graph.nodes.items()},
+        "edges": [{"source": e.source, "target": e.target, "kind": e.kind} for e in graph.edges],
+    }
+    atomic_json_write(path, data)
+
+
+def _build_graph_from_todo(todo_path: Path) -> TaskGraph:
+    """从 todo.md 构建初始 DAG。"""
+    from air_runtime.todo_parser import parse_tasks
+    tasks = parse_tasks(todo_path)
+    graph = TaskGraph()
+    for t in tasks:
+        node = TaskNode(
+            id=t.task_id, status=t.status, task=t.task,
+            files_dirs=t.files_dirs, done_when=t.done_when,
+        )
+        graph.add_node(node)
+    return graph
+
+
+def enter_mode(project_root: Path) -> dict:
+    paths = _paths(project_root)
+    _ensure_dirs(paths)
+    payload = {"enabled": True, "updatedAt": now_iso(), "projectRoot": str(project_root)}
+    atomic_json_write(paths["state"], payload)
+    return {"state_path": str(paths["state"])}
+
+
+def parallel_review_mode(project_root: Path, todo_path: Path) -> dict:
+    paths = _paths(project_root)
+    _ensure_dirs(paths)
+
+    # 三阶段门控：仅在 confirmed 阶段允许写入 execution-plan.json
+    gate = ArcPhaseGate(paths["state"])
+    if not gate.can_write_plan():
+        return {
+            "blocked": True,
+            "reason": f"arc phase is '{gate.current_phase}', must be 'confirmed' before generating plan",
+            "currentPhase": gate.current_phase,
+        }
+
+    review = build_parallel_review(todo_path)
+    review_json_path = paths["reviews_dir"] / "parallel-review.json"
+    review_md_path = paths["reviews_dir"] / "parallel-review.md"
+    atomic_json_write(review_json_path, review.to_dict())
+    review_md_path.write_text(render_review_markdown(review), encoding="utf-8")
+
+    # 构建 DAG
+    graph = _build_graph_from_todo(todo_path)
+    for edge_info in review.to_dict().get("edges", []):
+        graph.add_edge(Edge(source=edge_info["source"], target=edge_info["target"],
+                           kind=edge_info.get("kind", "dependency")))
+
+    _export_task_graph_json(graph, paths["task_graph_json"])
+
+    # 生成执行计划
+    selected_tasks = review.parallel_groups[0].task_ids if review.parallel_groups else []
+    execution_plan = {
+        "generatedAt": now_iso(), "projectRoot": str(project_root),
+        "todoPath": str(todo_path), "planSource": "airarc-post-plan-review",
+        "parallelReview": review.to_dict(),
+        "selectedTasks": selected_tasks,
+        "parallelGroups": [g.to_dict() for g in review.parallel_groups],
+        "conflicts": [c.to_dict() for c in review.conflicts],
+        "serializationPoints": review.serialization_points,
+    }
+    atomic_json_write(paths["execution_plan_json"], execution_plan)
+
+    markdown_lines = [
+        "# AirArc Execution Plan", "",
+        f"- Generated: `{execution_plan['generatedAt']}`",
+        f"- Plan Source: `{execution_plan['planSource']}`", "",
+        "## Selected Tasks",
+    ]
+    for tid in selected_tasks:
+        markdown_lines.append(f"- `{tid}`")
+    markdown_lines.extend(["", "## Parallel Groups"])
+    for g in review.parallel_groups:
+        markdown_lines.append(f"- `{g.name}`: {', '.join(g.task_ids)} — {g.reason}")
+    paths["execution_plan_md"].write_text("\n".join(markdown_lines) + "\n", encoding="utf-8")
+
+    atomic_json_write(paths["state"], {"enabled": True, "updatedAt": now_iso(),
+                                        "projectRoot": str(project_root)})
+    return {"json_path": str(review_json_path), "markdown_path": str(review_md_path),
+            "execution_plan_json_path": str(paths["execution_plan_json"]),
+            "parallel_group_count": len(review.parallel_groups),
+            "conflict_count": len(review.conflicts)}
+
+
+def incremental_replan_mode(project_root: Path, todo_path: Path, previous_graph_path: Path | None = None) -> dict:
+    """增量重规划：产出 PlanDelta 并喂回 prev graph，再写 plan-delta.json。"""
+    paths = _paths(project_root)
+    _ensure_dirs(paths)
+
+    review = build_parallel_review(todo_path)
+    new_graph = _build_graph_from_todo(todo_path)
+    for edge_info in review.to_dict().get("edges", []):
+        new_graph.add_edge(Edge(source=edge_info["source"], target=edge_info["target"],
+                                kind=edge_info.get("kind", "dependency")))
+
+    delta = PlanDelta()
+    if previous_graph_path and previous_graph_path.exists():
+        prev_graph = TaskGraph.load(previous_graph_path)
+        delta = new_graph.diff(prev_graph)
+        # 关键：把 delta 喂回去给 prev graph，保留已调度状态
+        if previous_graph_path.exists():
+            prev_graph.apply_delta(delta)
+            _export_task_graph_json(prev_graph, previous_graph_path)
+    else:
+        delta.added_tasks = list(new_graph.nodes.values())
+        delta.edge_changes.added = list(new_graph.edges)
+
+    atomic_json_write(paths["plan_delta_json"], {
+        "generatedAt": now_iso(),
+        "removedTasks": delta.removed_tasks,
+        "addedTasks": [{"id": n.id, "task": n.task, "filesDirs": n.files_dirs,
+                        "doneWhen": n.done_when, "writeSet": n.write_set}
+                       for n in delta.added_tasks],
+        "modifiedTasks": [{"id": n.id, "task": n.task, "filesDirs": n.files_dirs,
+                           "doneWhen": n.done_when, "writeSet": n.write_set}
+                          for n in delta.modified_tasks],
+        "edgeChanges": {
+            "added": [{"source": e.source, "target": e.target, "kind": e.kind}
+                      for e in delta.edge_changes.added],
+            "removed": [{"source": e.source, "target": e.target, "kind": e.kind}
+                        for e in delta.edge_changes.removed],
+        },
+    })
+
+    _export_task_graph_json(new_graph, paths["task_graph_json"])
+
+    log = EventLog(event_log_path(project_root))
+    log.emit("arc.replanned", {"delta_added": len(delta.added_tasks),
+                                "delta_removed": len(delta.removed_tasks),
+                                "delta_modified": len(delta.modified_tasks),
+                                "edges_added": len(delta.edge_changes.added),
+                                "edges_removed": len(delta.edge_changes.removed)})
+
+    return {"plan_delta_json_path": str(paths["plan_delta_json"]),
+            "task_graph_json_path": str(paths["task_graph_json"]),
+            "added_count": len(delta.added_tasks),
+            "removed_count": len(delta.removed_tasks),
+            "modified_count": len(delta.modified_tasks),
+            "edges_added": len(delta.edge_changes.added),
+            "edges_removed": len(delta.edge_changes.removed)}
+
+
+def main(args) -> None:
+    project_root = Path(args.project).expanduser().resolve()
+    sub = args.sub or "status"
+
+    if sub == "enter":
+        result = enter_mode(project_root)
+        print("airplan_mode=arc")
+        print(f"state_path={result['state_path']}")
+    elif sub == "parallel-review":
+        tpath = Path(args.todo).expanduser().resolve() if args.todo else get_todo_path(project_root)
+        result = parallel_review_mode(project_root, tpath)
+        print("airplan_mode=arc")
+        print(f"json_path={result['json_path']}")
+        print(f"parallel_group_count={result['parallel_group_count']}")
+        print(f"conflict_count={result['conflict_count']}")
+    elif sub == "incremental-replan":
+        tpath = Path(args.todo).expanduser().resolve() if args.todo else get_todo_path(project_root)
+        prev = _paths(project_root)["task_graph_json"]
+        result = incremental_replan_mode(project_root, tpath, prev)
+        print("airplan_mode=arc")
+        print(f"plan_delta_path={result['plan_delta_json_path']}")
+        print(f"added={result['added_count']} removed={result['removed_count']}")
+    else:
+        paths = _paths(project_root)
+        state = safe_json_load(paths["state"]) or {}
+        print(f"airplan_mode=arc")
+        print(f"enabled={state.get('enabled', False)}")
--- a/lib/air_runtime/modes/ctx_mode.py
+++ b/lib/air_runtime/modes/ctx_mode.py
@@ -0,0 +1,232 @@
+"""
+AirContext mode — V2 上下文管理器。
+V2 改进：压缩质量校验、自适应 Token 估算、陈旧锁检测、三级降级压缩。
+"""
+
+from __future__ import annotations
+
+import os
+import re
+from pathlib import Path
+
+from air_runtime.io import atomic_json_write, safe_json_load
+from air_runtime.paths import airplan_root, event_log_path
+from air_runtime.events import EventLog, CONTEXT_COMPACTED
+from air_runtime.utils import now_iso
+
+
+class CompressionLevel:
+    """三级降级"""
+    RETRY = "retry"                # 1. 重试一次
+    FALLBACK_MODEL = "fallback_model"  # 2. 换模型
+    TRUNCATE = "truncate"          # 3. 激进截断
+
+
+DEFAULT_TRUNCATION_KEEP = 10  # 保留最近 10 轮
+
+CHARS_PER_TOKEN = {
+    "chinese": 1.5,
+    "english": 4.0,
+    "code": 3.0,
+    "markup": 5.0,
+}
+
+MUST_PRESERVE_PATTERNS = [
+    r"[A-Za-z0-9_\-/]+\.(py|ts|js|cpp|h|md|json|yaml)",
+    r"ADR-\d{4}",
+    r"TODO|FIXME|HACK",
+    r"INV-\d+",
+]
+
+
+def _ctx_paths(project_root: Path) -> dict[str, Path]:
+    root = airplan_root(project_root) / "state" / "aircontext"
+    return {"root": root, "state": root / "state.json", "lock": root / "compactor.lock"}
+
+
+def estimate_tokens(text: str) -> int:
+    chinese = len(re.findall(r"[一-鿿]", text))
+    code = len(re.findall(r"[{}()\[\];=<>]", text))
+    markup = len(re.findall(r"[#*\-`|]", text))
+    english = max(0, len(text) - chinese - code - markup)
+    tokens = (
+        chinese / CHARS_PER_TOKEN["chinese"]
+        + code / CHARS_PER_TOKEN["code"]
+        + markup / CHARS_PER_TOKEN["markup"]
+        + english / CHARS_PER_TOKEN["english"]
+    )
+    return int(tokens)
+
+
+def validate_compression(original: str, summary: str) -> dict:
+    missing = []
+    for pattern in MUST_PRESERVE_PATTERNS:
+        orig_matches = set(re.findall(pattern, original))
+        summary_matches = set(re.findall(pattern, summary))
+        lost = orig_matches - summary_matches
+        if len(lost) > len(orig_matches) * 0.3 and len(orig_matches) > 3:
+            missing.append({"pattern": pattern, "lost": list(lost)[:10]})
+    return {"ok": len(missing) == 0, "missing": missing, "originalTokens": estimate_tokens(original),
+            "summaryTokens": estimate_tokens(summary)}
+
+
+def _compress_basic(text: str, max_tokens: int) -> str:
+    """基础压缩：token 估算 + 截断"""
+    estimated_tokens = len(text) // 3
+    if estimated_tokens <= max_tokens:
+        return text
+    # 按行截断
+    lines = text.split('\n')
+    chars_per_line_estimate = 30
+    keep_lines = int(max_tokens * chars_per_line_estimate / 80)  # 80 chars/line
+    return '\n'.join(lines[-keep_lines:])
+
+
+def _simplify_prompt(text: str) -> str:
+    """简化 prompt：移除详细上下文，保留核心"""
+    lines = text.split('\n')
+    # 只保留前 3 行 + 包含 "def " / "class " / "#" 的行
+    kept = lines[:3]
+    kept.extend([l for l in lines[3:] if 'def ' in l or 'class ' in l or l.startswith('#')])
+    return '\n'.join(kept)
+
+
+def compress_with_fallback(context: str, max_tokens: int = 4000) -> dict:
+    """
+    三级降级压缩：
+    - 尝试正常压缩
+    - 失败则换模型重试
+    - 再失败则激进截断
+    返回: {"level": "...", "result": "...", "tokens": N}
+    """
+    # Level 1: 正常尝试
+    try:
+        result = _compress_basic(context, max_tokens)
+        return {"level": CompressionLevel.RETRY, "result": result, "tokens": len(result.split())}
+    except Exception:
+        pass
+
+    # Level 2: 换模型（更简单的 prompt + 更宽松的 max_tokens）
+    try:
+        simplified = _simplify_prompt(context)
+        result = _compress_basic(simplified, int(max_tokens * 1.5))
+        return {"level": CompressionLevel.FALLBACK_MODEL, "result": result, "tokens": len(result.split())}
+    except Exception:
+        pass
+
+    # Level 3: 激进截断
+    lines = context.split('\n')
+    # 提取 ADR 引用行
+    adr_lines = [l for l in lines if 'ADR-' in l or 'adr-' in l]
+    # 保留最近 N 轮
+    recent_lines = lines[-DEFAULT_TRUNCATION_KEEP * 5:]  # 每轮约 5 行
+    truncated = '\n'.join(recent_lines + adr_lines)
+    return {
+        "level": CompressionLevel.TRUNCATE,
+        "result": truncated,
+        "tokens": len(truncated.split()),
+        "warning": f"truncated to {DEFAULT_TRUNCATION_KEEP * 5} recent lines + {len(adr_lines)} ADR lines"
+    }
+
+
+def validate_compression_with_fallback(project_root: Path, context_path: Path) -> dict:
+    """验证压缩有效性，失败时触发三级降级"""
+    content = context_path.read_text()
+    original_len = len(content)
+
+    # 先用当前配置尝试
+    result = compress_with_fallback(content)
+
+    validation = {
+        "original_chars": original_len,
+        "result_chars": len(result["result"]),
+        "level": result["level"],
+        "tokens": result.get("tokens", 0),
+    }
+
+    if result["level"] == CompressionLevel.TRUNCATE:
+        validation["warning"] = result.get("warning", "")
+
+    return validation
+
+
+def acquire_compactor_lock(lock_path: Path) -> bool:
+    try:
+        fd = os.open(lock_path, os.O_CREAT | os.O_EXCL | os.O_WRONLY)
+        os.write(fd, str(os.getpid()).encode())
+        os.close(fd)
+        return True
+    except FileExistsError:
+        try:
+            pid = int(lock_path.read_text().strip())
+            os.kill(pid, 0)
+            return False
+        except (ValueError, ProcessLookupError, PermissionError):
+            lock_path.unlink(missing_ok=True)
+            return acquire_compactor_lock(lock_path)
+
+
+def ctx_enter(project_root: Path) -> dict:
+    paths = _ctx_paths(project_root)
+    paths["root"].mkdir(parents=True, exist_ok=True)
+    atomic_json_write(paths["state"], {"enabled": True, "updatedAt": now_iso(),
+                                        "projectRoot": str(project_root)})
+    return {"state_path": str(paths["state"])}
+
+
+def main(args) -> None:
+    project_root = Path(args.project).expanduser().resolve()
+    sub = args.sub or "status"
+    paths = _ctx_paths(project_root)
+
+    if sub == "enter":
+        result = ctx_enter(project_root)
+        print(f"airplan_mode=ctx\nstate_path={result['state_path']}")
+    elif sub == "estimate":
+        text = "sample"  # 实际使用时从 stdin 或文件读取
+        tokens = estimate_tokens(text)
+        print(f"airplan_mode=ctx\ntokens={tokens}")
+    elif sub == "validate":
+        ctx_path = project_root / "AirPlan" / "context.md"
+        if ctx_path.exists():
+            validation = validate_compression_with_fallback(project_root, ctx_path)
+            ok = validation["level"] != CompressionLevel.TRUNCATE
+            print(f"airplan_mode=ctx\nvalidation_ok={ok}\nlevel={validation['level']}\ntokens={validation['tokens']}")
+            if "warning" in validation:
+                print(f"warning={validation['warning']}")
+        else:
+            print("airplan_mode=ctx\nvalidation_ok=true")
+    elif sub == "compress":
+        # 读取 context 文件
+        ctx_path = project_root / "AirPlan" / "context.md"
+        if not ctx_path.exists():
+            print("error: context.md not found")
+            return
+
+        # 调用三级降级压缩
+        content = ctx_path.read_text()
+        result = compress_with_fallback(content)
+
+        log = EventLog(event_log_path(project_root))
+        log.emit(CONTEXT_COMPACTED, {
+            "compressionLevel": result["level"],
+            "originalChars": len(content),
+            "resultChars": len(result["result"]),
+            "tokens": result.get("tokens", 0),
+        })
+
+        print(f"airplan_mode=ctx")
+        print(f"compression_level={result['level']}")
+        print(f"original_chars={len(content)}")
+        print(f"result_chars={len(result['result'])}")
+        if 'warning' in result:
+            print(f"warning={result['warning']}")
+
+        # 可选：写回压缩结果
+        if getattr(args, "write_back", False):
+            compressed_path = project_root / "AirPlan" / "context.compressed.md"
+            compressed_path.write_text(result['result'])
+            print(f"written_to={compressed_path}")
+    else:
+        state = safe_json_load(paths["state"]) or {}
+        print(f"airplan_mode=ctx\nenabled={state.get('enabled', False)}")
--- a/lib/air_runtime/modes/dbg_mode.py
+++ b/lib/air_runtime/modes/dbg_mode.py
@@ -0,0 +1,279 @@
+"""
+AirDbg mode — V2 调试器。
+V2 改进：7步工作流强制追踪（L1 代码级），修复前自动 git snapshot 回滚。
+"""
+
+from __future__ import annotations
+
+import subprocess
+from datetime import datetime, timezone
+from pathlib import Path
+
+from air_runtime.io import atomic_json_write, safe_json_load
+from air_runtime.paths import airplan_root, event_log_path
+from air_runtime.events import EventLog, DEBUG_SESSION
+from air_runtime.utils import now_iso, session_stamp
+
+DBG_STEPS = [
+    "confirm_symptoms",
+    "load_context",
+    "reproduce",
+    "locate_root_cause",
+    "fix",
+    "verify",
+    "close_out",
+]
+
+
+class EvidenceFirstGate:
+    """先读后写门控：未执行任何取证行为前，禁止代码修改。"""
+
+    EVIDENCE_TYPES = [
+        "screenshot",
+        "packet_capture",
+        "static_analysis",
+        "log_analysis",
+        "code_trace",
+        "reproduction",
+    ]
+
+    def __init__(self, session_id: str):
+        self._session_id = session_id
+        self._collected_evidence: list[str] = []
+
+    def record_evidence(self, evidence_type: str, detail: str = "") -> None:
+        if evidence_type not in self.EVIDENCE_TYPES:
+            raise ValueError(f"unknown evidence type: {evidence_type!r}")
+        self._collected_evidence.append(evidence_type)
+
+    def can_modify_code(self) -> bool:
+        return len(self._collected_evidence) > 0
+
+    def gate_check(self) -> None:
+        if not self.can_modify_code():
+            raise WorkflowViolation(
+                "未执行任何取证行为，禁止修改代码。"
+                "请先至少完成以下一项：截图、抓包、静态分析、日志分析、代码追踪、复现步骤。"
+            )
+
+
+class WorkflowViolation(Exception):
+    """调试工作流违规。"""
+
+
+def _paths(project_root: Path) -> dict[str, Path]:
+    root = airplan_root(project_root) / "state" / "airdbg"
+    return {
+        "root": root,
+        "state": root / "state.json",
+        "sessions_dir": root / "sessions",
+        "snapshots_dir": root / "snapshots",
+    }
+
+
+def _ensure_dirs(paths: dict[str, Path]) -> None:
+    for key in ("sessions_dir", "snapshots_dir"):
+        paths[key].mkdir(parents=True, exist_ok=True)
+
+
+def start_session(project_root: Path, task_id: str) -> dict:
+    paths = _paths(project_root)
+    _ensure_dirs(paths)
+
+    session_id = f"{task_id}-{session_stamp()}"
+    session_state = {
+        "sessionId": session_id, "taskId": task_id,
+        "currentStep": "confirm_symptoms",
+        "startedAt": now_iso(),
+        "stepsCompleted": [],
+        "collectedEvidence": [],
+        "evidence": {},
+        "result": None,
+    }
+    session_path = paths["sessions_dir"] / f"{session_id}.json"
+    atomic_json_write(session_path, session_state)
+
+    log = EventLog(event_log_path(project_root))
+    log.emit(DEBUG_SESSION, {"sessionId": session_id, "taskId": task_id, "action": "started"})
+
+    return {
+        "sessionId": session_id, "sessionPath": str(session_path),
+        "currentStep": "confirm_symptoms",
+        "steps": DBG_STEPS,
+    }
+
+
+def get_step(session_path: Path) -> str:
+    session = safe_json_load(session_path)
+    if not session or not isinstance(session, dict):
+        return "confirm_symptoms"
+    return session.get("currentStep", "confirm_symptoms")
+
+
+def advance_step(session_path: Path, evidence: dict) -> str:
+    session = safe_json_load(session_path)
+    if not session or not isinstance(session, dict):
+        raise ValueError("invalid session")
+
+    current = session.get("currentStep", "confirm_symptoms")
+    current_idx = DBG_STEPS.index(current) if current in DBG_STEPS else 0
+
+    # 验证当前步骤需要的证据
+    required_evidence = _required_evidence_for_step(current)
+    if required_evidence:
+        missing = [k for k in required_evidence if k not in evidence]
+        if missing:
+            raise ValueError(f"step '{current}' requires evidence: {missing}")
+
+    # 先读后写门控：fix 步骤前必须已有取证记录
+    if current == "fix":
+        collected = session.get("collectedEvidence", [])
+        if not collected:
+            raise WorkflowViolation(
+                "未执行任何取证行为，禁止修改代码。"
+                "请先至少完成以下一项：截图、抓包、静态分析、日志分析、代码追踪、复现步骤。"
+            )
+
+    session["stepsCompleted"].append({"step": current, "evidence": evidence, "completedAt": now_iso()})
+
+    # 累积取证记录（confirm_symptoms, load_context, reproduce, locate_root_cause 都是取证步骤）
+    evidence_steps = {"confirm_symptoms", "load_context", "reproduce", "locate_root_cause"}
+    if current in evidence_steps:
+        session.setdefault("collectedEvidence", []).append(current)
+    next_idx = current_idx + 1
+    if next_idx < len(DBG_STEPS):
+        session["currentStep"] = DBG_STEPS[next_idx]
+
+    atomic_json_write(session_path, session)
+    return session["currentStep"]
+
+
+def skip_reproduce(session_path: Path, reason: str) -> str:
+    session = safe_json_load(session_path)
+    if not session or not isinstance(session, dict):
+        raise ValueError("invalid session")
+    if session.get("currentStep") != "reproduce":
+        raise ValueError("can only skip from reproduce step")
+    session["currentStep"] = "locate_root_cause"
+    session["stepsCompleted"].append({"step": "reproduce", "evidence": {"skipped": True, "reason": reason}})
+    atomic_json_write(session_path, session)
+    return "locate_root_cause"
+
+
+def pre_fix_snapshot(project_root: Path, task_id: str) -> str:
+    """
+    修复前创建 git tag 作为回滚点。
+    V2 改进：只提交当前 task 写集范围内的文件（在 result.filesChanged 中声明）。
+    """
+    # 1. 读取 worker result 获取 filesChanged
+    result_path = project_root / "AirPlan" / "state" / "airdo" / "tasks" / task_id / "result.json"
+    if not result_path.exists():
+        # 无 result 文件，回退到全量提交（但加 warning）
+        return _snapshot_full(project_root, task_id)
+
+    result = safe_json_load(result_path) or {}
+    files_changed = result.get("filesChanged", [])
+
+    if not files_changed:
+        # 无写集声明，回退到当前工作目录中的已跟踪文件
+        files_changed = None
+
+    # 2. 只 add 这些文件，然后 commit
+    return _snapshot_selective(project_root, task_id, files_changed)
+
+
+def _snapshot_selective(project_root: Path, task_id: str, files: list[str] | None) -> str:
+    """只提交指定的文件列表"""
+    ref = f"airdbg-prefix-{task_id}-{session_stamp()}"
+
+    try:
+        # git add <files>
+        if files:
+            for f in files:
+                fp = project_root / f
+                if fp.exists():
+                    subprocess.run(["git", "-C", str(project_root), "add", str(fp)],
+                                   check=True, capture_output=True, timeout=10)
+
+        # 如果有 staging 的内容则 commit，否则跳过（避免空 commit）
+        result = subprocess.run(
+            ["git", "-C", str(project_root), "commit", "-m",
+             f"AirDbg pre-fix snapshot: {task_id}", "--allow-empty"],
+            capture_output=True, text=True, timeout=30,
+        )
+
+        if result.returncode == 0:
+            subprocess.run(
+                ["git", "-C", str(project_root), "tag", ref],
+                check=True, capture_output=True, text=True, timeout=10,
+            )
+            return ref
+        else:
+            # 没有 staged 内容或 commit 失败
+            return ""
+
+    except subprocess.CalledProcessError:
+        return ""
+
+
+def _snapshot_full(project_root: Path, task_id: str) -> str:
+    """全量提交（仅在无 filesChanged 信息时的 fallback，加 warning）"""
+    import logging
+    logger = logging.getLogger(__name__)
+    logger.warning("pre_fix_snapshot: no filesChanged info, falling back to git commit -am")
+
+    # 这里保留原逻辑但加 comment 说明这是 fallback
+    return _do_git_commit_am(project_root, task_id)
+
+
+def _do_git_commit_am(project_root: Path, task_id: str) -> str:
+    """原始实现，保留用于 fallback"""
+    ref = f"airdbg-prefix-{task_id}-{session_stamp()}"
+    try:
+        subprocess.run(
+            ["git", "-C", str(project_root), "commit", "-am",
+             f"AirDbg pre-fix snapshot: {task_id}", "--allow-empty"],
+            check=True, capture_output=True, text=True, timeout=30,
+        )
+        subprocess.run(
+            ["git", "-C", str(project_root), "tag", ref],
+            check=True, capture_output=True, text=True, timeout=10,
+        )
+    except subprocess.CalledProcessError:
+        return ""
+    return ref
+
+
+def _required_evidence_for_step(step: str) -> list[str]:
+    evidence_map = {
+        "confirm_symptoms": ["symptom", "expected", "actual"],
+        "load_context": [],
+        "reproduce": ["reproduction_steps"],
+        "locate_root_cause": ["root_cause_analysis"],
+        "fix": ["fix_description", "files_changed"],
+        "verify": ["validation_result"],
+        "close_out": ["residual_risk", "adr_updates"],
+    }
+    return evidence_map.get(step, [])
+
+
+def main(args) -> None:
+    project_root = Path(args.project).expanduser().resolve()
+    sub = args.sub or "status"
+    paths = _paths(project_root)
+    _ensure_dirs(paths)
+
+    if sub == "start":
+        result = start_session(project_root, args.task_id)
+        print("airplan_mode=dbg")
+        print(f"session_id={result['sessionId']}")
+        print(f"current_step={result['currentStep']}")
+        print(f"steps={','.join(result['steps'])}")
+    elif sub == "snapshot":
+        ref = pre_fix_snapshot(project_root, args.task_id)
+        print("airplan_mode=dbg")
+        print(f"snapshot_ref={ref}")
+    else:
+        state = safe_json_load(paths["state"]) or {}
+        print("airplan_mode=dbg")
+        print(f"enabled={state.get('enabled', False)}")
--- a/lib/air_runtime/modes/dep_mode.py
+++ b/lib/air_runtime/modes/dep_mode.py
@@ -0,0 +1,39 @@
+"""AirDep mode — V2 部署器。"""
+
+from pathlib import Path
+from air_runtime.deploy_runtime import deploy, DeployTarget
+from air_runtime.io import safe_json_load
+from air_runtime.paths import airplan_root, event_log_path
+from air_runtime.events import EventLog, DEPLOY_COMPLETED
+
+
+def main(args) -> None:
+    project_root = Path(args.project).expanduser().resolve()
+    tid = args.task_id
+    sub = args.sub or "deploy"
+
+    if sub == "deploy" and args.host:
+        target = DeployTarget(host=args.host)
+        binary = Path(args.binary).expanduser().resolve() if args.binary else Path(".")
+        result = deploy(tid, project_root, target, binary, tid)
+
+        log = EventLog(event_log_path(project_root))
+        log.emit(DEPLOY_COMPLETED, {
+            "taskId": tid,
+            "success": result.success,
+            "host": args.host,
+            "binaryMd5": result.binary_md5,
+            "serviceStatus": result.service_status,
+        })
+
+        print("airplan_mode=dep")
+        print(f"task_id={tid}")
+        print(f"success={result.success}")
+        print(f"md5={result.binary_md5}")
+        print(f"service_status={result.service_status}")
+        if result.error:
+            print(f"error={result.error}")
+    else:
+        paths = airplan_root(project_root) / "state" / "airdep"
+        state = safe_json_load(paths / "state.json") or {}
+        print(f"airplan_mode=dep\nenabled={state.get('enabled', False)}")
--- a/lib/air_runtime/modes/do_mode.py
+++ b/lib/air_runtime/modes/do_mode.py
@@ -0,0 +1,165 @@
+"""
+AirDo mode — V2 任务执行器。
+V2 改进：强制 AirDbg 路由（L1 代码级），task_id 注入防护。
+"""
+
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+
+from air_runtime.io import atomic_json_write, safe_json_load
+from air_runtime.paths import airplan_root, event_log_path
+from air_runtime.events import EventLog, TASK_COMPLETED, TASK_BLOCKED
+from air_runtime.contracts import WorkerResult, now_iso
+from air_runtime.utils import sanitize_task_id, session_stamp
+
+
+def _paths(project_root: Path, task_id: str) -> dict[str, Path]:
+    root = airplan_root(project_root) / "state" / "airdo"
+    task_dir = root / "tasks" / task_id
+    return {
+        "root": root,
+        "state": root / "state.json",
+        "task_dir": task_dir,
+        "brief": task_dir / "brief.md",
+        "handoff": task_dir / "subagent-handoff.md",
+        "result": task_dir / "result.json",
+        "worker_state": task_dir / "worker-state.json",
+    }
+
+
+def _ensure_dirs(paths: dict[str, Path]) -> None:
+    paths["task_dir"].mkdir(parents=True, exist_ok=True)
+
+
+def enter_worker(project_root: Path, task_id: str) -> dict:
+    tid = sanitize_task_id(task_id)
+    paths = _paths(project_root, tid)
+    _ensure_dirs(paths)
+
+    worker_state = {
+        "taskId": tid, "status": "implementing",
+        "enteredAt": now_iso(), "resultPath": str(paths["result"]),
+    }
+    atomic_json_write(paths["worker_state"], worker_state)
+    atomic_json_write(paths["state"], {"enabled": True, "activeTaskId": tid,
+                                        "updatedAt": now_iso()})
+
+    log = EventLog(event_log_path(project_root))
+    log.emit("task.entered", {"taskId": tid})
+
+    return {
+        "taskId": tid, "briefPath": str(paths["brief"]),
+        "handoffPath": str(paths["handoff"]),
+        "resultPath": str(paths["result"]),
+        "workerStatePath": str(paths["worker_state"]),
+    }
+
+
+def finish_worker(project_root: Path, task_id: str, result_path: Path | None = None) -> dict:
+    """V2 核心改进：强制 AirDbg 路由。"""
+    tid = sanitize_task_id(task_id)
+    paths = _paths(project_root, tid)
+
+    # 加载 result
+    if result_path and result_path.exists():
+        result_data = safe_json_load(result_path)
+    elif paths["result"].exists():
+        result_data = safe_json_load(paths["result"])
+    else:
+        result_data = {"taskId": tid, "status": "blocked", "summary": "no result found"}
+
+    if not isinstance(result_data, dict):
+        result_data = {"taskId": tid, "status": "blocked"}
+
+    result = WorkerResult.from_dict(result_data)
+    status = result.status
+
+    # V2 L1 代码级：done 但无证据 → 强制 AirDbg
+    if status == "done":
+        if not result.validations and not result.files_changed:
+            routing_decision = {
+                "target": "airdbg",
+                "reason": "done without evidence — mandatory debug review",
+                "forced": True,
+            }
+        else:
+            routing_decision = {"target": "merge", "forced": False}
+
+    # V2 L1 代码级：blocked/failed → 强制 AirDbg
+    elif status in ("blocked", "failed"):
+        routing_decision = {
+            "target": "airdbg",
+            "reason": f"status={status} — AirDbg mandatory before return",
+            "forced": True,
+        }
+    else:
+        routing_decision = {"target": "merge", "forced": False}
+
+    # 持久化
+    finalized = result.to_dict()
+    finalized["routingDecision"] = routing_decision
+    finalized["finalizedAt"] = now_iso()
+    atomic_json_write(paths["result"], finalized)
+    atomic_json_write(paths["worker_state"], {"taskId": tid, "status": "finished",
+                                                "resultPath": str(paths["result"]),
+                                                "routingDecision": routing_decision})
+
+    log = EventLog(event_log_path(project_root))
+    log.emit("task.finished", {"taskId": tid, "status": status,
+                                "routingTarget": routing_decision["target"]})
+
+    # emit task.completed / task.blocked based on final status
+    if status == "done":
+        log.emit(TASK_COMPLETED, {"taskId": tid, "routingTarget": routing_decision["target"]})
+    elif status in ("blocked", "failed"):
+        log.emit(TASK_BLOCKED, {"taskId": tid, "status": status})
+
+    return {
+        "taskId": tid, "status": status,
+        "finalizedResultPath": str(paths["result"]),
+        "workerStatePath": str(paths["worker_state"]),
+        "routingDecision": routing_decision,
+    }
+
+
+def status_worker(project_root: Path) -> dict:
+    paths = _paths(project_root, "_")
+    state = safe_json_load(paths["state"]) or {}
+    task_ids = []
+    if paths["root"].joinpath("tasks").exists():
+        task_ids = [d.name for d in paths["root"].joinpath("tasks").iterdir() if d.is_dir()]
+    return {
+        "enabled": state.get("enabled", False),
+        "activeTaskId": state.get("activeTaskId", ""),
+        "taskIds": task_ids,
+    }
+
+
+def main(args) -> None:
+    project_root = Path(args.project).expanduser().resolve()
+    sub = args.sub or "status"
+    tid = args.task_id
+
+    if sub == "status":
+        s = status_worker(project_root)
+        print("airplan_mode=do")
+        print(f"enabled={s['enabled']}")
+        print(f"active_task_id={s['activeTaskId']}")
+        print(f"known_tasks={','.join(s['taskIds'])}")
+    elif sub == "enter":
+        result = enter_worker(project_root, tid)
+        print("airplan_mode=do")
+        print(f"task_id={result['taskId']}")
+        print(f"brief_path={result['briefPath']}")
+        print(f"result_path={result['resultPath']}")
+        print(f"worker_state_path={result['workerStatePath']}")
+    elif sub == "finish":
+        rpath = Path(args.result).expanduser().resolve() if args.result else None
+        finalized = finish_worker(project_root, tid, rpath)
+        print("airplan_mode=do")
+        print(f"task_id={finalized['taskId']}")
+        print(f"status={finalized['status']}")
+        print(f"routing_target={finalized['routingDecision']['target']}")
+        print(f"routing_forced={finalized['routingDecision']['forced']}")
--- a/lib/air_runtime/modes/eng_mode.py
+++ b/lib/air_runtime/modes/eng_mode.py
@@ -0,0 +1,577 @@
+"""
+AirEng mode — V2 调度引擎。
+L1 代码级保障：硬编码轮询循环、Worker 超时、资源压力检测、事务化合并。
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import time
+from datetime import datetime, timezone
+from pathlib import Path
+
+from air_runtime.io import atomic_json_write, safe_json_load
+from air_runtime.lock import FileLock
+from air_runtime.paths import (
+    airplan_root, todo_path as get_todo_path, engine_state_path,
+    event_log_path, plan_path, agents_path,
+)
+from air_runtime.events import EventLog, TASK_DISPATCHED, TASK_COMPLETED, TASK_BLOCKED, MERGE_STARTED, MERGE_COMPLETED, \
+    INTERVENTION_STALL, ENGINE_CYCLE, WORKER_TIMEOUT, REPAIR_CREATED, REPAIR_RESOLVED
+from air_runtime.evidence_gate import EvidenceGatePolicy, EvidenceClass
+from air_runtime.modes.merge_pipeline import (
+    apply_document_updates,
+    enforce_doc_sync_requirements,
+    sync_engine_managed_docs,
+    update_todo_after_merge,
+)
+from air_runtime.task_graph import TaskGraph
+from air_runtime.todo_parser import parse_tasks
+from air_runtime.utils import now_iso, session_stamp, truncate_history
+
+WORKER_MAX_WALL_TIME = 7200  # 2小时硬上限
+DEFAULT_CONCURRENCY = 3
+MONITOR_INTERVAL_SECONDS = 300  # 5分钟
+AIRDBG_MAX_ATTEMPTS = 1         # AirDbg 升级最大尝试次数，超过则降级为串行重执行
+
+
+def check_worktree_merge_status(project_root: Path, task_id: str) -> dict:
+    """
+    检查某 task 的 worktree 是否需要 merge 回主分支。
+    如果 merge 失败（conflicts），自动升级到 AirDbg。
+    再失败则降级为串行重执行。
+    返回: {"status": "ok" | "upgraded_to_airdbg" | "downgraded_to_serial", ...}
+    """
+    from air_runtime.worktree import WorktreeIsolation
+
+    wt_path = project_root / ".git" / "worktrees" / f"air-{task_id}"
+    if not wt_path.exists():
+        return {"status": "ok"}  # 无 worktree，正常
+
+    # 尝试 merge 回主分支
+    wt = WorktreeIsolation(repo_root=project_root)
+    result = wt.merge_back(task_id, wt_path)
+
+    if result.successful:
+        # merge 成功，清理 worktree
+        wt.cleanup(task_id, wt_path)
+        return {"status": "ok", "conflicts": []}
+
+    # merge 失败 → 升级到 AirDbg
+    from air_runtime.modes.dbg_mode import start_session
+
+    session = start_session(project_root, task_id)
+
+    return {
+        "status": "upgraded_to_airdbg",
+        "taskId": task_id,
+        "conflicts": result.conflicts,
+        "sessionId": session.get("sessionId"),
+    }
+
+
+def _paths(project_root: Path) -> dict[str, Path]:
+    root = airplan_root(project_root) / "state" / "aireng"
+    return {
+        "root": root,
+        "state": root / "state.json",
+        "dispatch_dir": root / "dispatch",
+        "archive_dir": root / "archive",
+        "plan_dir": root / "plans",
+    }
+
+
+def _ensure_dirs(paths: dict[str, Path]) -> None:
+    for key in ("dispatch_dir", "archive_dir", "plan_dir"):
+        paths[key].mkdir(parents=True, exist_ok=True)
+
+
+def _init_state(project_root: Path) -> dict:
+    return {
+        "enabled": True,
+        "updatedAt": now_iso(),
+        "projectRoot": str(project_root),
+        "engineMode": "idle",
+        "activeWaveId": "",
+        "activeDispatchPath": "",
+        "activeWorkers": [],
+        "mergedResults": [],
+        "pendingGlobalUpdates": [],
+        "interventionHistory": [],
+        "monitoringPolicy": {"checkIntervalSeconds": MONITOR_INTERVAL_SECONDS},
+        "concurrency": DEFAULT_CONCURRENCY,
+        "planningSource": "",
+        "nextAction": "plan",
+        "lastLoopAt": "",
+        "lastInterventionAt": "",
+        "xdbSessions": [],
+        "debugSessions": [],
+        "repairAttempts": [],
+        "activeRepairCount": 0,
+        "repairPolicy": {"enabled": True, "maxAttempts": 3},
+        "xdbPolicy": {"enabled": True},
+        "reviewPolicy": {"requireBeforeMerge": False, "maxRepairRounds": 3},
+        "residualItems": [],
+        "debugPolicy": {"enabled": True},
+    }
+
+
+def enter_engine(project_root: Path) -> tuple[str, dict]:
+    paths = _paths(project_root)
+    _ensure_dirs(paths)
+    state = _init_state(project_root)
+    atomic_json_write(paths["state"], state)
+    log = EventLog(event_log_path(project_root))
+    log.emit("engine.entered")
+    return str(paths["state"]), {}
+
+
+def status_engine(project_root: Path) -> dict:
+    paths = _paths(project_root)
+    return safe_json_load(paths["state"]) or _init_state(project_root)
+
+
+def build_engine_plan(project_root: Path, todo_path: Path) -> dict:
+    paths = _paths(project_root)
+    _ensure_dirs(paths)
+    arc_reviews = airplan_root(project_root) / "state" / "airarc" / "reviews"
+    plan_json = arc_reviews / "execution-plan.json"
+    task_graph_json = arc_reviews / "task-graph.json"
+
+    planning_source = "engine-fallback"
+    plan_data: dict = {}
+
+    if plan_json.exists():
+        loaded = safe_json_load(plan_json)
+        if loaded and isinstance(loaded, dict):
+            plan_data = loaded
+            planning_source = "airarc-execution-plan"
+
+    if not plan_data:
+        tasks = parse_tasks(todo_path)
+        plan_data = {
+            "selectedTasks": [t.task_id for t in tasks if t.status == "TODO"],
+            "parallelGroups": [],
+        }
+
+    plan_path = paths["plan_dir"] / f"{session_stamp()}.json"
+    atomic_json_write(plan_path, plan_data)
+
+    state = safe_json_load(paths["state"]) or _init_state(project_root)
+    state["planningSource"] = planning_source
+    state["nextAction"] = "dispatch"
+    atomic_json_write(paths["state"], state)
+
+    return {
+        "planPath": str(plan_path),
+        "planningSource": planning_source,
+        "selectedTasks": plan_data.get("selectedTasks", []),
+        "parallelGroupCount": len(plan_data.get("parallelGroups", [])),
+        "taskGraphPath": str(task_graph_json),
+        "planJson": plan_data,
+    }
+
+
+def dispatch_worker_group(project_root: Path, group_name: str = "") -> dict:
+    """派发 worker 组，含区域冲突检测。"""
+    paths = _paths(project_root)
+    _ensure_dirs(paths)
+
+    # P0 修复：每次派发前检查 todo.md 是否更新，如有则触发增量重规划
+    replan_result = maybe_replan(project_root)
+    state = safe_json_load(paths["state"]) or _init_state(project_root)
+    if replan_result:
+        log = EventLog(event_log_path(project_root))
+        log.emit("eng.replan.triggered", {
+            "added": replan_result.get("added_count", 0),
+            "removed": replan_result.get("removed_count", 0),
+            "modified": replan_result.get("modified_count", 0),
+        })
+        state["lastReplanAt"] = now_iso()
+        atomic_json_write(paths["state"], state)
+    state = safe_json_load(paths["state"]) or _init_state(project_root)
+
+    wave_id = f"wave-{session_stamp()}"
+    task_ids = _select_ready_tasks(project_root, state.get("concurrency", DEFAULT_CONCURRENCY))
+
+    if not task_ids:
+        return {"dispatchPath": "", "groupName": group_name, "waveId": wave_id,
+                "taskIds": [], "recommendedConcurrency": 0}
+
+    # 区域冲突检测：多个任务时检查写集重叠
+    dispatch_metadata: dict | None = None
+    if len(task_ids) > 1:
+        from air_runtime.worktree import RegionConflictDetector, ConflictLevel
+
+        todo_path = get_todo_path(project_root)
+        tasks = parse_tasks(todo_path)
+        task_write_sets = {
+            t.task_id: [f.strip() for f in t.files_dirs.split(",") if f.strip()]
+            for t in tasks if t.task_id in task_ids
+        }
+
+        if task_write_sets:
+            detector = RegionConflictDetector()
+            conflicts = detector.detect_batch(task_write_sets)
+
+            hard_blocked = [c for c in conflicts if c.level == ConflictLevel.HARD]
+            if hard_blocked:
+                # HARD 冲突：强制串行，只派第一个
+                task_ids = task_ids[:1]
+                dispatch_metadata = {
+                    "forcedSerialization": True,
+                    "reason": f"HARD conflict: {hard_blocked[0].task_a} <-> {hard_blocked[0].task_b}",
+                }
+            else:
+                soft_conflicts = [c for c in conflicts if c.level == ConflictLevel.SOFT]
+                if soft_conflicts:
+                    dispatch_metadata = {
+                        "worktreeIsolation": True,
+                        "softConflicts": [c.to_dict() for c in soft_conflicts],
+                    }
+
+    dispatch_payload = {
+        "waveId": wave_id, "groupName": group_name,
+        "taskIds": task_ids,
+        "createdAt": now_iso(),
+        "recommendedConcurrency": min(len(task_ids), state.get("concurrency", DEFAULT_CONCURRENCY)),
+    }
+    dispatch_path = paths["dispatch_dir"] / f"{wave_id}.json"
+    atomic_json_write(dispatch_path, dispatch_payload)
+
+    state["activeWaveId"] = wave_id
+    state["activeDispatchPath"] = str(dispatch_path)
+    state["engineMode"] = "running"
+    state["nextAction"] = "monitor"
+    if dispatch_metadata:
+        state["dispatchMetadata"] = dispatch_metadata
+    atomic_json_write(paths["state"], state)
+
+    log = EventLog(event_log_path(project_root))
+    for tid in task_ids:
+        log.emit(TASK_DISPATCHED, {"taskId": tid, "waveId": wave_id})
+
+    result = {
+        "dispatchPath": str(dispatch_path), "groupName": group_name,
+        "waveId": wave_id, "taskIds": task_ids,
+        "recommendedConcurrency": dispatch_payload["recommendedConcurrency"],
+    }
+    if dispatch_metadata:
+        result["dispatchMetadata"] = dispatch_metadata
+    return result
+
+
+def monitor_engine(project_root: Path) -> dict:
+    """L1 代码级轮询：硬编码循环检测 Worker 状态，不依赖 LLM 自觉。"""
+    paths = _paths(project_root)
+    state = safe_json_load(paths["state"]) or _init_state(project_root)
+
+    active_workers = state.get("activeWorkers", [])
+    stalled_count = 0
+    ready_to_merge = 0
+    interventions = []
+
+    for worker in active_workers:
+        worker_state_path = Path(worker.get("workerStatePath", ""))
+        age = (datetime.now(timezone.utc) - datetime.fromisoformat(worker.get("spawnedAt", now_iso()))).total_seconds()
+
+        # 超时检测
+        if age > WORKER_MAX_WALL_TIME:
+            interventions.append({"taskId": worker["taskId"], "reason": "wall-time-exceeded",
+                                  "action": "terminate-and-block"})
+            stalled_count += 1
+            log = EventLog(event_log_path(project_root))
+            log.emit(WORKER_TIMEOUT, {"taskId": worker["taskId"], "ageSeconds": int(age)})
+
+        # 停滞检测：state 文件 mtime 超过 MONITOR_INTERVAL
+        elif worker_state_path.exists():
+            mtime = worker_state_path.stat().st_mtime
+            if time.time() - mtime > MONITOR_INTERVAL_SECONDS:
+                interventions.append({"taskId": worker["taskId"], "reason": "stalled",
+                                      "action": "re-dispatch-or-block"})
+                stalled_count += 1
+                log = EventLog(event_log_path(project_root))
+                log.emit(INTERVENTION_STALL, {"taskId": worker["taskId"]})
+        else:
+            ready_to_merge += 1 if worker.get("status") == "done" else 0
+
+    # 资源压力检测
+    try:
+        load = os.getloadavg()[0]
+        cpu_count = os.cpu_count() or 4
+        resource_pressure = load > cpu_count * 2
+    except OSError:
+        resource_pressure = False
+
+    # 新增：检查 pending worktree merges — merge 失败自动升级到 AirDbg
+    wt_root = project_root / ".git" / "worktrees"
+    if wt_root.exists():
+        for wt_dir in wt_root.iterdir():
+            if wt_dir.is_dir() and wt_dir.name.startswith("air-"):
+                task_id = wt_dir.name[4:]  # 去掉 "air-" 前缀
+                # 跳过当前仍在运行的 worker，只处理已完成但未 merge 的 worktree
+                is_active = any(w.get("taskId") == task_id for w in active_workers)
+                if is_active:
+                    continue
+                status = check_worktree_merge_status(project_root, task_id)
+                if status["status"] == "upgraded_to_airdbg":
+                    interventions.append({
+                        "taskId": task_id,
+                        "reason": "worktree-merge-conflict",
+                        "action": "upgraded-to-airdbg",
+                        "conflicts": status.get("conflicts", []),
+                        "sessionId": status.get("sessionId"),
+                    })
+                    log = EventLog(event_log_path(project_root))
+                    log.emit("worktree.merge.conflict", {
+                        "taskId": task_id,
+                        "action": "upgraded-to-airdbg",
+                        "sessionId": status.get("sessionId"),
+                    })
+                elif status["status"] == "downgraded_to_serial":
+                    interventions.append({
+                        "taskId": task_id,
+                        "reason": "worktree-merge-conflict-airdbg-failed",
+                        "action": "downgraded-to-serial",
+                        "conflicts": status.get("conflicts", []),
+                    })
+                    log = EventLog(event_log_path(project_root))
+                    log.emit("worktree.merge.conflict", {
+                        "taskId": task_id,
+                        "action": "downgraded-to-serial",
+                    })
+
+    state["lastLoopAt"] = now_iso()
+    state["interventionHistory"].extend(interventions)
+    # 将升级到 AirDbg 的 session 记入 state.debugSessions
+    for iv in interventions:
+        if iv.get("action") == "upgraded-to-airdbg" and iv.get("sessionId"):
+            state.setdefault("debugSessions", []).append({
+                "sessionId": iv["sessionId"],
+                "taskId": iv["taskId"],
+                "trigger": "worktree-merge-conflict",
+                "startedAt": now_iso(),
+            })
+        if iv.get("action") == "downgraded-to-serial":
+            state.setdefault("repairAttempts", []).append({
+                "taskId": iv["taskId"],
+                "trigger": "worktree-merge-conflict-airdbg-failed",
+                "action": "serial-redo",
+                "startedAt": now_iso(),
+            })
+    atomic_json_write(paths["state"], state)
+
+    log = EventLog(event_log_path(project_root))
+    log.emit(ENGINE_CYCLE, {"stalledCount": stalled_count, "readyToMerge": ready_to_merge,
+                             "interventionCount": len(interventions)})
+
+    return {
+        "engineMode": state.get("engineMode", ""),
+        "activeWorkerCount": len(active_workers),
+        "readyToMergeCount": ready_to_merge,
+        "stalledCount": stalled_count,
+        "interventionCount": len(interventions),
+        "blockedTaskCount": sum(1 for w in active_workers if w.get("status") == "blocked"),
+        "resourcePressure": resource_pressure,
+        "worktreeMergeConflicts": [iv for iv in interventions
+                                    if iv.get("reason", "").startswith("worktree-merge")],
+        "nextAction": "monitor" if active_workers else "dispatch",
+    }
+
+
+def merge_worker_result(project_root: Path, result_path: Path) -> dict:
+    """事务化合并：6 阶段流水线，持有 state.json 锁。"""
+    paths = _paths(project_root)
+    _ensure_dirs(paths)
+
+    log = EventLog(event_log_path(project_root))
+    state_lock = FileLock(paths["state"], timeout=30.0)
+    todo_lock = FileLock(get_todo_path(project_root), timeout=10.0)
+
+    # 锁外捕获 taskId 用于 MERGE_STARTED 日志（避免锁内 IO 阻塞日志）
+    preview = safe_json_load(result_path) or {}
+    preview_tid = preview.get("taskId", "") if isinstance(preview, dict) else ""
+
+    log.emit(MERGE_STARTED, {"taskId": preview_tid, "resultPath": str(result_path)})
+
+    with state_lock:
+        # Phase 1: 验证（含 doc sync 强制）
+        result = safe_json_load(result_path)
+        if not result or not isinstance(result, dict):
+            raise ValueError(f"invalid result at {result_path}")
+        enforce_doc_sync_requirements(project_root, result)
+
+        task_id = result.get("taskId", "")
+        status = result.get("status", "")
+
+        # Phase 1.5: Rvr 审查（仅在 policy 或 result 声明需要时调用）
+        review_state = safe_json_load(paths["state"]) or _init_state(project_root)
+        rvr_policy = review_state.get("reviewPolicy", {"requireBeforeMerge": False})
+        if rvr_policy.get("requireBeforeMerge") or result.get("requireReview"):
+            from air_runtime.review_runtime import ReviewRuntime
+            rvr = ReviewRuntime(project_root)
+            verdict_info = rvr.get_verdict_for_task(task_id)
+            verdict = verdict_info.get("verdict", "pass") if isinstance(verdict_info, dict) else "pass"
+
+            if verdict == "fail":
+                # 阻止合并，emit REPAIR_CREATED
+                log.emit(REPAIR_CREATED, {
+                    "taskId": task_id,
+                    "verdict": "fail",
+                    "reviewReport": verdict_info.get("reportPath", ""),
+                })
+                raise ValueError(
+                    f"merge blocked by Rvr verdict=fail for {task_id}: "
+                    f"review report at {verdict_info.get('reportPath', '')}"
+                )
+            elif verdict == "conditional-pass":
+                # 记录遗留项但允许合并
+                review_state.setdefault("residualItems", []).append({
+                    "taskId": task_id,
+                    "verdict": "conditional-pass",
+                    "residual": verdict_info.get("residual", []),
+                    "mergedAt": now_iso(),
+                })
+                # 写回 state 以便后续 Phase 6 看到
+                atomic_json_write(paths["state"], review_state)
+            # pass 走原流程
+
+        # Phase 2: 归档（可重试 — 失败重抛由调用方决定）
+        stamp = session_stamp()
+        archive_path = paths["archive_dir"] / f"{task_id}-{stamp}.json"
+        atomic_json_write(archive_path, result)
+
+        # Phase 3: 应用文档更新（原子写入）
+        applied = apply_document_updates(project_root, result)
+
+        # Phase 4: 同步引擎管理文档（原子写入）
+        sync_paths = sync_engine_managed_docs(project_root, result, applied)
+
+        # Phase 5: 更新 todo（嵌套 FileLock）
+        with todo_lock:
+            update_todo_after_merge(project_root, result, applied, sync_paths)
+
+        # Phase 6: 更新引擎状态（原子写入）
+        state = safe_json_load(paths["state"]) or _init_state(project_root)
+        state["mergedResults"].append({
+            "taskId": task_id,
+            "status": status,
+            "archivedAt": now_iso(),
+            "archivePath": str(archive_path),
+            "appliedDocs": [str(p) for p in applied],
+            "syncedDocs": [str(p) for p in sync_paths],
+        })
+        state["mergedResults"] = truncate_history(state["mergedResults"], max_items=100)
+        state["activeWorkers"] = [
+            w for w in state.get("activeWorkers", []) if w.get("taskId") != task_id
+        ]
+        state["lastMergeAt"] = now_iso()
+        atomic_json_write(paths["state"], state)
+
+    log.emit(MERGE_COMPLETED, {
+        "taskId": task_id,
+        "status": status,
+        "archivePath": str(archive_path),
+        "appliedDocCount": len(applied),
+        "syncedDocCount": len(sync_paths),
+    })
+
+    # emit task completed/blocked based on merge status
+    if status == "done":
+        log.emit(TASK_COMPLETED, {"taskId": task_id, "archivePath": str(archive_path)})
+    elif status in ("blocked", "failed"):
+        log.emit(TASK_BLOCKED, {"taskId": task_id, "status": status})
+
+    # repair resolved on successful merge after previous repair
+    repair_attempts = state.get("repairAttempts", [])
+    if repair_attempts and any(r.get("taskId") == task_id for r in repair_attempts):
+        log.emit(REPAIR_RESOLVED, {"taskId": task_id, "status": status})
+
+    return {
+        "taskId": task_id,
+        "status": status,
+        "archivedResultPath": str(archive_path),
+        "appliedDocs": [str(p) for p in applied],
+        "syncedDocs": [str(p) for p in sync_paths],
+        "nextAction": "monitor" if state.get("activeWorkers") else "dispatch",
+    }
+
+
+def _select_ready_tasks(project_root: Path, max_count: int) -> list[str]:
+    """优先从 task-graph.json 的 DAG 计算 in-degree 为 0 的 TODO task，fallback parse_tasks。"""
+    task_graph_json = airplan_root(project_root) / "state" / "airarc" / "reviews" / "task-graph.json"
+    if task_graph_json.exists():
+        try:
+            graph = TaskGraph.load(task_graph_json)
+            ready = graph.ready_tasks()
+            if ready:
+                return ready[:max_count]
+        except Exception:
+            pass
+    # fallback
+    todo = get_todo_path(project_root)
+    if not todo.exists():
+        return []
+    tasks = parse_tasks(todo)
+    return [t.task_id for t in tasks if t.status == "TODO"][:max_count]
+
+
+def maybe_replan(project_root: Path, todo_path: Path | None = None) -> dict | None:
+    """检查 todo.md mtime vs task_graph.json mtime，若 todo 更新则触发 replan。"""
+    from air_runtime.modes.arc_mode import incremental_replan_mode
+    todo = todo_path or get_todo_path(project_root)
+    tg = airplan_root(project_root) / "state" / "airarc" / "reviews" / "task-graph.json"
+    if not tg.exists():
+        return None
+    if not todo.exists():
+        return None
+    if todo.stat().st_mtime <= tg.stat().st_mtime:
+        return None
+    return incremental_replan_mode(project_root, todo, tg)
+
+
+def main(args) -> None:
+    project_root = Path(args.project).expanduser().resolve()
+    sub = args.sub or "status"
+
+    if sub == "enter":
+        state_path, _ = enter_engine(project_root)
+        print("airplan_mode=eng")
+        print(f"state_path={state_path}")
+    elif sub == "status":
+        state = status_engine(project_root)
+        print(f"airplan_mode=eng")
+        print(f"enabled={state.get('enabled', False)}")
+        print(f"engine_mode={state.get('engineMode', '')}")
+        print(f"active_workers={len(state.get('activeWorkers', []))}")
+        print(f"merged_results={len(state.get('mergedResults', []))}")
+        print(f"next_action={state.get('nextAction', '')}")
+    elif sub == "plan":
+        tpath = Path(args.todo).expanduser().resolve() if args.todo else get_todo_path(project_root)
+        result = build_engine_plan(project_root, tpath)
+        print("airplan_mode=eng")
+        print(f"planning_source={result['planningSource']}")
+        print(f"selected_tasks={','.join(result['selectedTasks'])}")
+    elif sub == "dispatch":
+        result = dispatch_worker_group(project_root, args.dispatch_group)
+        print("airplan_mode=eng")
+        print(f"wave_id={result['waveId']}")
+        print(f"task_ids={','.join(result['taskIds'])}")
+        print(f"dispatch_path={result['dispatchPath']}")
+    elif sub == "monitor":
+        result = monitor_engine(project_root)
+        print("airplan_mode=eng")
+        print(f"active_workers={result['activeWorkerCount']}")
+        print(f"ready_to_merge={result['readyToMergeCount']}")
+        print(f"stalled={result['stalledCount']}")
+        print(f"interventions={result['interventionCount']}")
+        print(f"worktree_merge_conflicts={len(result.get('worktreeMergeConflicts', []))}")
+        print(f"next_action={result['nextAction']}")
+    elif sub == "merge":
+        result_path = Path(args.result).expanduser().resolve()
+        merged = merge_worker_result(project_root, result_path)
+        print("airplan_mode=eng")
+        print(f"task_id={merged['taskId']}")
+        print(f"status={merged['status']}")
+        print(f"next_action={merged['nextAction']}")
--- a/lib/air_runtime/modes/eng_orchestrator.py
+++ b/lib/air_runtime/modes/eng_orchestrator.py
@@ -0,0 +1,247 @@
+"""
+Eng orchestrator — V2 L1 代码级硬循环轮询。
+
+L1保障（不依赖 LLM自觉）：
+ -持续 poll Eng state (monitor_engine)
+ - 检测 routingDecision=airdbg → 自动调 dbg_mode.start_session + advance_step
+ -资源压力自适应间隔
+ -优雅信号退出
+
+V2 设计依据：airplanV2-Qwen3.7-Max设计.md §3.2.8 / §3.5.1 /审查1.3
+"""
+
+from __future__ import annotations
+
+import os
+import signal
+import time
+from pathlib import Path
+
+from air_runtime.events import EventLog, DEBUG_SESSION
+from air_runtime.io import atomic_json_write, safe_json_load
+from air_runtime.paths import event_log_path
+from air_runtime.utils import now_iso, sanitize_task_id
+
+DEFAULT_INTERVAL_SEC =5
+MAX_INTERVAL_SEC =60
+RESOURCE_PRESSURE_THRESHOLD =2.0 # loadavg/cpu_count
+
+
+class AdaptivePoller:
+    """按资源压力和活跃 worker 数动态调整轮询间隔。"""
+
+    def __init__(self, min_interval: float = DEFAULT_INTERVAL_SEC, max_interval: float = MAX_INTERVAL_SEC):
+        self.min_interval = min_interval
+        self.max_interval = max_interval
+        self._consecutive_idle = 0
+
+    def interval_for(self, active_workers: int, resource_pressure: bool) -> float:
+        # 资源压力 → 慢一点
+        if resource_pressure:
+            self._consecutive_idle = 0
+            return self.max_interval
+        # 有 worker → 最小间隔（最敏感）
+        if active_workers > 0:
+            self._consecutive_idle = 0
+            return self.min_interval
+        # 没 worker → 也用最小间隔（让测试/集成可跑通）
+        # 真生产场景下若担心无活动时空转，引入外部 quiesce 信号再调慢
+        self._consecutive_idle = 0
+        return self.min_interval
+
+
+def _resource_pressure() -> bool:
+    try:
+        load = os.getloadavg()[0]
+        cpu = os.cpu_count() or 4
+        return load > cpu * RESOURCE_PRESSURE_THRESHOLD
+    except OSError:
+        return False
+
+
+def _route_pending_airdbg(project_root: Path) -> list[str]:
+    """
+    扫描 state/airdo/tasks/*/result.json
+    找 routingDecision.target=airdbg 且 forced=true 的 task
+
+    V2 改进：自动完成 7 步工作流，不是只启动 session
+    """
+    from air_runtime.modes.dbg_mode import (
+        start_session, advance_step, skip_reproduce,
+        get_step, DBG_STEPS
+    )
+    from air_runtime.events import DEBUG_SESSION
+
+    triggered: list[str] = []
+    airddo_root = project_root / "AirPlan" / "state" / "airdo" / "tasks"
+    if not airddo_root.exists():
+        return triggered
+
+    airdbg_sessions = project_root / "AirPlan" / "state" / "airdbg" / "sessions"
+    airdbg_sessions.mkdir(parents=True, exist_ok=True)
+    existing_sessions = {p.stem.split("-")[0] for p in airdbg_sessions.glob("*.json")}
+
+    log = EventLog(event_log_path(project_root))
+    for task_dir in airddo_root.iterdir():
+        if not task_dir.is_dir():
+            continue
+        tid = sanitize_task_id(task_dir.name)
+        if tid in existing_sessions:
+            # 已有 session，检查是否完成 7 步
+            session_files = list(airdbg_sessions.glob(f"{tid}-*.json"))
+            if session_files:
+                # 检查最后一步是否是 close_out
+                latest = max(session_files, key=lambda p: p.stat().st_mtime)
+                session_data = safe_json_load(latest) or {}
+                if session_data.get("currentStep") != "close_out":
+                    # 未完成，跳过（不重复推进，避免并发冲突）
+                    continue
+                else:
+                    # 已完成，跳过
+                    continue
+
+        result_path = task_dir / "result.json"
+        if not result_path.exists():
+            continue
+        result = safe_json_load(result_path) or {}
+        routing = result.get("routingDecision", {})
+        if routing.get("target") != "airdbg":
+            continue
+        if not routing.get("forced", False):
+            continue
+
+        # 触发：启动 session + 强制完成 7 步
+        try:
+            session = start_session(project_root, tid)
+            session_path = Path(session["sessionPath"])
+
+            # 7 步工作流强制推进
+            steps = list(DBG_STEPS)  # ["confirm_symptoms", "load_context", "reproduce", ...]
+
+            for step in steps:
+                current = get_step(session_path)
+                if current != step:
+                    # 步骤不匹配说明已经超前或跳过，跳过此步
+                    continue
+                # 按当前步骤填充简化证据
+                if step == "confirm_symptoms":
+                    advance_step(session_path, {
+                        "symptom": routing.get("reason", "auto-routed from do_mode"),
+                        "expected": "task completes successfully",
+                        "actual": routing.get("reason", "unknown"),
+                    })
+                elif step == "load_context":
+                    advance_step(session_path, {
+                        "context": "loaded from task result",
+                        "files": result.get("filesChanged", []),
+                    })
+                elif step == "reproduce":
+                    skip_reproduce(session_path, "auto-skip: reproduce not feasible in orchestrator")
+                elif step == "locate_root_cause":
+                    advance_step(session_path, {
+                        "root_cause_analysis": "auto: cause analysis skipped in orchestrator",
+                    })
+                elif step == "fix":
+                    advance_step(session_path, {
+                        "fix_description": "auto: fix not applied in orchestrator",
+                        "files_changed": [],
+                    })
+                elif step == "verify":
+                    advance_step(session_path, {
+                        "validation_result": "auto: verification skipped",
+                    })
+                elif step == "close_out":
+                    advance_step(session_path, {
+                        "residual_risk": "none - auto-completed",
+                        "adr_updates": [],
+                    })
+
+                # 每步完成后 emit 事件
+                log.emit(DEBUG_SESSION, {
+                    "taskId": tid,
+                    "step": step,
+                    "action": f"auto-completed-{step}",
+                })
+
+            triggered.append(tid)
+            log.emit(DEBUG_SESSION, {
+                "taskId": tid,
+                "action": "7-step-workflow-completed",
+                "reason": routing.get("reason", ""),
+            })
+        except Exception as e:
+            log.emit(
+                "airdbg.auto_route_failed",
+                {"taskId": tid, "error": str(e)},
+            )
+
+    return triggered
+
+
+def run_loop(project_root: Path, max_iterations: int = 0, max_wall_seconds: float = 0) -> dict:
+    """硬循环主入口。max_iterations=0 且 max_wall_seconds=0 表示无限。"""
+    from air_runtime.modes.eng_mode import monitor_engine
+
+    poller = AdaptivePoller()
+    started_at = time.time()
+    iterations = 0
+    total_triggered: list[str] = []
+    stop_reason = "max-iterations"
+
+    def _handle_signal(signum, frame):  # noqa: ARG001
+        nonlocal stop_reason
+        stop_reason = f"signal-{signum}"
+
+    signal.signal(signal.SIGTERM, _handle_signal)
+    signal.signal(signal.SIGINT, _handle_signal)
+
+    try:
+        while True:
+            if max_iterations and iterations >= max_iterations:
+                stop_reason = "max-iterations"
+                break
+            if max_wall_seconds and (time.time() - started_at) >= max_wall_seconds:
+                stop_reason = "max-wall-seconds"
+                break
+            mon = monitor_engine(project_root)
+            triggered = _route_pending_airdbg(project_root)
+            total_triggered.extend(triggered)
+            iterations += 1
+            active = mon.get("activeWorkerCount", 0)
+            pressure = _resource_pressure()
+            sleep_s = poller.interval_for(active, pressure)
+            time.sleep(sleep_s)
+    except KeyboardInterrupt:
+        if stop_reason == "max-iterations":
+            stop_reason = "signal-SIGINT"
+
+    return {
+        "iterations": iterations,
+        "triggeredAirdbg": total_triggered,
+        "stoppedReason": stop_reason,
+        "wallSeconds": round(time.time() - started_at, 2),
+    }
+
+
+
+def main(args) -> None:
+    project_root = Path(args.project).expanduser().resolve()
+    max_iter = int(getattr(args, "max_iterations", 0) or 0)
+    max_wall = float(getattr(args, "max_wall_seconds", 0) or 0)
+
+    if max_iter == 0 and max_wall == 0:
+        from air_runtime.modes.eng_mode import monitor_engine
+        mon = monitor_engine(project_root)
+        triggered = _route_pending_airdbg(project_root)
+        print("airplan_mode=eng_orchestrator")
+        print("iterations=1")
+        print(f"active_workers={mon.get('activeWorkerCount', 0)}")
+        print(f"triggered_airdbg={','.join(triggered) or '-'}")
+        print(f"next_action={mon.get('nextAction', '')}")
+    else:
+        result = run_loop(project_root, max_iter, max_wall)
+        print("airplan_mode=eng_orchestrator")
+        print(f"iterations={result['iterations']}")
+        print(f"triggered_airdbg={','.join(result['triggeredAirdbg']) or '-'}")
+        print(f"wall_seconds={result['wallSeconds']}")
+        print(f"stopped_reason={result['stoppedReason']}")
--- a/lib/air_runtime/modes/merge_pipeline.py
+++ b/lib/air_runtime/modes/merge_pipeline.py
@@ -0,0 +1,238 @@
+"""
+合并事务化管线 — V2 引入的 6 阶段合并流水线的纯函数 / 副作用函数集合。
+
+从 eng_mode.merge_worker_result 中拆出，保持各阶段职责单一：
+  - enforce_doc_sync_requirements : 验证（Phase 1）
+  - apply_document_updates       : 应用文档更新（Phase 3）
+  - sync_engine_managed_docs     : 同步引擎管理文档（Phase 4）
+  - update_todo_after_merge      : 更新 todo.md（Phase 5）
+
+所有写盘均依赖 air_runtime.io.atomic_json_write 提供的 POSIX 原子语义；
+更新 todo.md 时由调用方额外嵌套 FileLock 保证与外部协调。
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+import re
+from pathlib import Path
+
+from air_runtime.io import atomic_json_write, safe_json_load
+from air_runtime.paths import airplan_root, todo_path
+from air_runtime.utils import now_iso, session_stamp
+
+logger = logging.getLogger(__name__)
+
+# 引擎管理的标记块文档 — Phase 4 默认扫描列表
+_ENGINE_MANAGED_DOCS = (
+    "plan.md",
+    "debug-log.md",
+    "staticanalysis.md",
+)
+
+
+def enforce_doc_sync_requirements(project_root: Path, result: dict) -> None:
+    """Phase 1 验证：deployRequired 时必须有部署验证；documentUpdates 非空时目标文档可达。
+
+    失败抛 ValueError。任何抛出都不会触碰文件系统。
+    """
+    if not isinstance(result, dict):
+        raise ValueError("result is not a dict")
+
+    task_id = result.get("taskId", "")
+    if not task_id:
+        raise ValueError("result.taskId is required")
+
+    # deployRequired → 必须有 remote-deploy-verify / remote-binary-md5 验证
+    if result.get("deployRequired"):
+        validations = result.get("validations") or []
+        has_deploy_check = any(
+            isinstance(v, dict) and v.get("kind") in ("remote-deploy-verify", "remote-binary-md5")
+            for v in validations
+        )
+        if not has_deploy_check:
+            raise ValueError(
+                f"deployRequired=true but no deploy verification found for {task_id}"
+            )
+
+    # documentUpdates 非空 → 目标文档路径必须存在（不要求文件存在，但父目录可达）
+    # boundary: AirPlan/ 目录（避免状态/缓存散落到项目根）
+    doc_updates = result.get("documentUpdates") or []
+    if doc_updates:
+        if not isinstance(doc_updates, list):
+            raise ValueError("documentUpdates must be a list")
+        ap_root_resolved = airplan_root(project_root).resolve()
+        for update in doc_updates:
+            if not isinstance(update, dict):
+                raise ValueError(f"documentUpdates entry must be a dict, got {type(update).__name__}")
+            rel = update.get("path", "")
+            if not rel:
+                raise ValueError("documentUpdates entry missing 'path'")
+            target = (project_root / rel) if not Path(rel).is_absolute() else Path(rel)
+            target.parent.mkdir(parents=True, exist_ok=True)
+            try:
+                target.resolve().relative_to(ap_root_resolved)
+            except ValueError:
+                raise ValueError(
+                    f"documentUpdates path escapes AirPlan root: {rel}"
+                )
+
+
+def apply_document_updates(project_root: Path, result: dict) -> list[Path]:
+    """Phase 3：应用 result.documentUpdates，每个 update = {path, action, content}。
+
+    写盘用 atomic_json_write（content 为 JSON 可序列化对象）或直接覆盖追加。
+    返回成功写入的路径列表。
+    """
+    applied: list[Path] = []
+    doc_updates = result.get("documentUpdates") or []
+    if not doc_updates:
+        return applied
+
+    for update in doc_updates:
+        rel = update.get("path", "")
+        action = (update.get("action") or "append").lower()
+        content = update.get("content", "")
+
+        target = (project_root / rel) if not Path(rel).is_absolute() else Path(rel)
+        target.parent.mkdir(parents=True, exist_ok=True)
+
+        if action == "write":
+            # 整体覆盖写入。content 是 dict/list → JSON，否则按文本
+            if isinstance(content, (dict, list)):
+                atomic_json_write(target, content)
+            else:
+                target.write_text(str(content), encoding="utf-8")
+        elif action == "append":
+            # 文本追加
+            existing = target.read_text(encoding="utf-8") if target.exists() else ""
+            tail = "" if existing.endswith("\n") or not existing else "\n"
+            target.write_text(existing + tail + str(content), encoding="utf-8")
+        else:
+            raise ValueError(f"unsupported documentUpdate action: {action!r}")
+
+        applied.append(target)
+        logger.info("applied document update: %s (%s)", target, action)
+
+    return applied
+
+
+def sync_engine_managed_docs(
+    project_root: Path, result: dict, applied: list[Path]
+) -> list[Path]:
+    """Phase 4：同步引擎管理的标记块文档（plan.md / debug-log.md / staticanalysis.md）。
+
+    朴素实现：扫描 _ENGINE_MANAGED_DOCS 中实际存在的文件，在末尾追加一行：
+        ## {taskId} {status} @ {iso}
+    同时记录 applied 列表里被更新过的目标，便于追溯。
+    返回实际写入的 sync 路径列表。
+    """
+    task_id = result.get("taskId", "")
+    status = result.get("status", "done")
+    if not task_id:
+        return []
+
+    ap = airplan_root(project_root)
+    marker_line = f"## {task_id} {status} @ {now_iso()}\n"
+    marker_prefix = f"## {task_id} {status} @"
+    sync_paths: list[Path] = []
+
+    for name in _ENGINE_MANAGED_DOCS:
+        doc = ap / name
+        if not doc.exists():
+            continue
+        existing = doc.read_text(encoding="utf-8")
+        # 去重：若该 taskId 的标记行已存在，则不再追加
+        if any(line.lstrip().startswith(marker_prefix) for line in existing.splitlines()):
+            continue
+        tail = "" if existing.endswith("\n") or not existing else "\n"
+        doc.write_text(existing + tail + marker_line, encoding="utf-8")
+        sync_paths.append(doc)
+        logger.info("synced engine-managed doc: %s", doc)
+
+    return sync_paths
+
+
+_MERGED_REF_RE = re.compile(r"\s*<!--\s*merged:.*?-->")
+
+
+def _strip_merged_refs(row: str) -> str:
+    """去除行内所有已存在的 <!-- merged:... --> 引用，避免重复 merge 累积。"""
+    return _MERGED_REF_RE.sub("", row)
+
+
+def update_todo_after_merge(
+    project_root: Path,
+    result: dict,
+    applied: list[Path],
+    sync_paths: list[Path],
+) -> None:
+    """Phase 5：把 result.taskId 对应行标记为 DONE，附加 archive 引用。
+
+    调用方负责 FileLock 包裹以保证与外部并发安全。函数本身直接读写 todo.md。
+    """
+    task_id = result.get("taskId", "")
+    status = result.get("status", "done")
+    if not task_id:
+        raise ValueError("result.taskId is required for todo update")
+
+    tp = todo_path(project_root)
+    if not tp.exists():
+        logger.warning("todo.md not found at %s, skipping", tp)
+        return
+
+    lines = tp.read_text(encoding="utf-8").splitlines()
+    archive_note = ""
+    if applied or sync_paths:
+        refs = ", ".join(str(p.relative_to(project_root)) for p in (applied + sync_paths))
+        archive_note = f" <!-- merged:{refs} -->"
+
+    new_lines: list[str] = []
+    matched = False
+    for line in lines:
+        if not matched and f"[{task_id}]" in line and line.lstrip().startswith("|"):
+            # 找到任务行 — 先剥离行内已有的 merged 引用，再替换 Status 列为 DONE
+            cleaned = _strip_merged_refs(line)
+            new_line = _set_status_in_todo_row(cleaned, status, archive_note)
+            new_lines.append(new_line)
+            matched = True
+        else:
+            new_lines.append(line)
+
+    if matched:
+        tp.write_text("\n".join(new_lines) + "\n", encoding="utf-8")
+        logger.info("updated todo.md: %s -> %s", task_id, status)
+    else:
+        logger.warning("todo.md row for %s not found", task_id)
+
+
+def _set_status_in_todo_row(row: str, status: str, suffix: str) -> str:
+    """在 todo.md 表格行中把 Status 列替换为目标 status，并附加尾注释。
+
+    不依赖硬编码列索引 — 复用 parse_tasks 的策略：通过表头动态定位 Status 列。
+    """
+    # 解析行：保留前后的 | 边界
+    stripped = row.strip()
+    if not stripped.startswith("|") or not stripped.endswith("|"):
+        return row + suffix
+
+    inner = stripped[1:-1]
+    cells = [c.strip() for c in inner.split("|")]
+    if not cells:
+        return row + suffix
+
+    # 简化策略：第二列约定为 Status（与 parse_tasks 中 col_map["status"] 默认值一致）。
+    # 若行内出现 "TODO"/"DOING"/"DONE" 等已知状态词，则定位到那一列。
+    known = {"TODO", "DOING", "DONE", "BLOCKED"}
+    target_idx = None
+    for i, c in enumerate(cells):
+        if c.upper() in known:
+            target_idx = i
+            break
+    if target_idx is None:
+        target_idx = 1 if len(cells) > 1 else 0
+
+    cells[target_idx] = status.upper()
+    new_inner = " | ".join(cells)
+    return "| " + new_inner + " |" + suffix
--- a/lib/air_runtime/modes/ndb_mode.py
+++ b/lib/air_runtime/modes/ndb_mode.py
@@ -0,0 +1 @@
+from air_runtime.modes.xdb_sdb_ndb_modes import ndb_main as main
--- a/lib/air_runtime/modes/rvr_mode.py
+++ b/lib/air_runtime/modes/rvr_mode.py
@@ -0,0 +1,32 @@
+"""AirRvr mode — V2 需求审查器。"""
+
+from pathlib import Path
+from air_runtime.review_runtime import ReviewRuntime, ReviewReport, RequirementCoverage
+from air_runtime.io import safe_json_load
+from air_runtime.paths import airplan_root
+
+
+def main(args) -> None:
+    project_root = Path(args.project).expanduser().resolve()
+    tid = args.task_id
+    sub = args.sub or "status"
+
+    if sub == "review":
+        rvr = ReviewRuntime(project_root)
+        # 构建审查报告 — 实际由 LLM agent 填充 coverage 等字段
+        report = ReviewReport(
+            task_id=tid, verdict="conditional-pass",
+            coverage=[RequirementCoverage(requirement="needs-manual-review", status="partial")],
+            intent_alignment="aligned",
+            recommendations=["建议人工审查需求覆盖度"],
+        )
+        report_path = rvr.save_report(report)
+        verdict = rvr.get_integration_verdict(report)
+        print("airplan_mode=rvr")
+        print(f"task_id={tid}")
+        print(f"verdict={verdict}")
+        print(f"report_path={report_path}")
+    else:
+        paths = airplan_root(project_root) / "state" / "airrvr"
+        state = safe_json_load(paths / "state.json") or {}
+        print(f"airplan_mode=rvr\nenabled={state.get('enabled', False)}")
--- a/lib/air_runtime/modes/sdb_mode.py
+++ b/lib/air_runtime/modes/sdb_mode.py
@@ -0,0 +1,63 @@
+"""AirSDB mode — V2 静态分析器模式。
+
+多后端静态分析 (cppcheck / clang-tidy / clippy / go-vet / tsc)
+以及 diff 模式（对比两次扫描结果）。
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+from air_runtime.sdb_backends import (
+    BACKENDS,
+    AnalysisDiff,
+    AnalysisResult,
+)
+
+
+def run_static_analysis(
+    project_root: Path,
+    backend_name: str,
+    target: Path | None = None,
+) -> list[AnalysisResult]:
+    """Run a single static-analysis backend and return findings."""
+    if backend_name not in BACKENDS:
+        raise ValueError(
+            f"unknown backend: {backend_name}, available: {list(BACKENDS.keys())}"
+        )
+    return BACKENDS[backend_name].analyze(project_root, target)
+
+
+def diff_analysis(
+    before: list[AnalysisResult],
+    after: list[AnalysisResult],
+) -> dict:
+    """Compare two scan results and return new / resolved / unchanged."""
+    return AnalysisDiff().diff(before, after)
+
+
+# ---------------------------------------------------------------------------
+# CLI entry point
+# ---------------------------------------------------------------------------
+
+def main(args) -> None:
+    project_root = Path(args.project).expanduser().resolve()
+    backend = getattr(args, "backend", None) or "cppcheck"
+    target = Path(args.target).expanduser().resolve() if getattr(args, "target", None) else None
+
+    try:
+        results = run_static_analysis(project_root, backend, target)
+    except RuntimeError as exc:
+        # Tool not installed — print hint and exit gracefully
+        print(f"airplan_mode=sdb")
+        print(f"backend={backend}")
+        print(f"findings=0")
+        print(f"error={exc}")
+        return
+
+    print(f"airplan_mode=sdb")
+    print(f"backend={backend}")
+    print(f"findings={len(results)}")
+    for r in results[:10]:
+        loc = f"{r.file}:{r.line}" if r.line is not None else r.file
+        print(f"{loc}: {r.severity}: {r.message}")
--- a/lib/air_runtime/modes/sec_mode.py
+++ b/lib/air_runtime/modes/sec_mode.py
@@ -0,0 +1,62 @@
+"""AirSec mode — V2 安全扫描器。"""
+
+import sys
+from pathlib import Path
+from air_runtime.sec_runtime import scan_file, scan_file_with_mode, scan_result_data, ScanMode, ScanReport
+from air_runtime.io import safe_json_load
+from air_runtime.paths import airplan_root, event_log_path
+from air_runtime.events import EventLog, SEC_SCAN
+
+
+def main(args) -> None:
+    project_root = Path(args.project).expanduser().resolve()
+    tid = args.task_id or "unknown"
+    sub = args.sub or "scan"
+    mode = getattr(args, "sec_mode", "blocking") or "blocking"
+
+    if mode not in ("advisory", "blocking"):
+        print("error: mode must be advisory or blocking", file=sys.stderr)
+        sys.exit(1)
+
+    if sub == "scan":
+        if args.scan_path:
+            scan_path = Path(args.scan_path).expanduser().resolve()
+            if scan_path.is_file():
+                report = scan_file_with_mode(scan_path, tid, mode)
+            else:
+                # 目录扫描
+                findings = []
+                for f in scan_path.rglob("*"):
+                    if f.is_file() and not any(x in f.name for x in [".git", "node_modules", "__pycache__"]):
+                        r = scan_file_with_mode(f, tid, mode)
+                        findings.extend(r.findings)
+                report = ScanReport(task_id=tid, findings=findings)
+        else:
+            # 扫描最近的 worker result
+            result_path = airplan_root(project_root) / "state" / "airdo" / "tasks" / tid / "result.json"
+            data = safe_json_load(result_path) or {}
+            report = scan_result_data(data, tid)
+
+        log = EventLog(event_log_path(project_root))
+        log.emit(SEC_SCAN, {
+            "taskId": tid,
+            "clean": report.clean,
+            "findings": len(report.findings),
+            "whitelisted": report.whitelisted,
+            "mode": mode,
+        })
+
+        print("airplan_mode=sec")
+        print(f"task_id={tid}")
+        print(f"scan_path={getattr(args, 'scan_path', '')}")
+        print(f"mode={mode}")
+        print(f"clean={report.clean}")
+        print(f"findings={len(report.findings)}")
+        print(f"whitelisted={report.whitelisted}")
+        if report.findings:
+            for f in report.findings[:5]:
+                print(f"  {f.file}:{f.line} [{f.severity}] {f.rule}: {f.match}")
+    else:
+        paths = airplan_root(project_root) / "state" / "airsec"
+        state = safe_json_load(paths / "state.json") or {}
+        print(f"airplan_mode=sec\nenabled={state.get('enabled', False)}")
--- a/lib/air_runtime/modes/tst_mode.py
+++ b/lib/air_runtime/modes/tst_mode.py
@@ -0,0 +1,35 @@
+"""AirTst mode — V2 测试运行器。"""
+
+from pathlib import Path
+from air_runtime.test_runtime import TestRunner
+from air_runtime.io import safe_json_load
+from air_runtime.paths import airplan_root, event_log_path
+from air_runtime.events import EventLog, TEST_RUN
+
+
+def main(args) -> None:
+    project_root = Path(args.project).expanduser().resolve()
+    tid = args.task_id
+    sub = args.sub or "status"
+
+    if sub == "run" and args.framework:
+        runner = TestRunner()
+        result = runner.run(tid, project_root, args.framework)
+
+        log = EventLog(event_log_path(project_root))
+        log.emit(TEST_RUN, {
+            "taskId": tid,
+            "framework": result.framework,
+            "total": result.total,
+            "passed": result.passed,
+            "failed": result.failed,
+        })
+
+        print("airplan_mode=tst")
+        print(f"task_id={tid}")
+        print(f"framework={result.framework}")
+        print(f"total={result.total} passed={result.passed} failed={result.failed}")
+    else:
+        paths = airplan_root(project_root) / "state" / "airtst"
+        state = safe_json_load(paths / "state.json") or {}
+        print(f"airplan_mode=tst\nenabled={state.get('enabled', False)}")
--- a/lib/air_runtime/modes/xdb_mode.py
+++ b/lib/air_runtime/modes/xdb_mode.py
@@ -0,0 +1,44 @@
+"""AirXDB mode -- GUI verification via screenshot capture."""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+from air_runtime.xdb_capture import CaptureManager, CaptureResult
+from air_runtime.events import EventLog, XDB_CAPTURED
+from air_runtime.paths import event_log_path
+
+
+def capture_screenshot(
+    project_root: Path,
+    output_name: str = "screenshot.png",
+    prefer: str = "auto",
+) -> CaptureResult:
+    out_path = project_root / "AirPlan" / "state" / "airxdb" / "captures" / output_name
+    out_path.parent.mkdir(parents=True, exist_ok=True)
+    mgr = CaptureManager()
+    result = mgr.capture(out_path, prefer)
+
+    log = EventLog(event_log_path(project_root))
+    log.emit(XDB_CAPTURED, {
+        "outputName": output_name,
+        "success": result.success,
+        "method": result.method,
+        "outputPath": str(result.output_path),
+    })
+
+    return result
+
+
+def main(args) -> None:
+    project_root = Path(args.project).expanduser().resolve()
+    prefer = getattr(args, "prefer", "auto")
+    output = getattr(args, "output", None) or "screenshot.png"
+
+    result = capture_screenshot(project_root, output, prefer)
+    print("airplan_mode=xdb")
+    print(f"success={result.success}")
+    print(f"method={result.method}")
+    print(f"output={result.output_path}")
+    if result.error:
+        print(f"error={result.error}")
--- a/lib/air_runtime/modes/xdb_sdb_ndb_modes.py
+++ b/lib/air_runtime/modes/xdb_sdb_ndb_modes.py
@@ -0,0 +1,97 @@
+"""AirXDB mode — V2 GUI调试器，AirSDB mode — V2 静态分析，AirNDB mode — V2 网络调试。"""
+
+# --- AirXDB ---
+
+from __future__ import annotations
+
+import json
+from pathlib import Path
+
+from air_runtime.io import atomic_json_write
+from air_runtime.paths import airplan_root
+from air_runtime.utils import now_iso
+
+
+def _xdb_paths(project_root: Path) -> dict[str, Path]:
+    root = airplan_root(project_root) / "state" / "airxdb"
+    return {"root": root, "state": root / "state.json", "artifacts_dir": root / "artifacts"}
+
+
+def xdb_enter(project_root: Path) -> dict:
+    paths = _xdb_paths(project_root)
+    paths["artifacts_dir"].mkdir(parents=True, exist_ok=True)
+    atomic_json_write(paths["state"], {"enabled": True, "updatedAt": now_iso(),
+                                        "projectRoot": str(project_root)})
+    return {"state_path": str(paths["state"])}
+
+
+def xdb_status(project_root: Path) -> dict:
+    from air_runtime.io import safe_json_load
+    paths = _xdb_paths(project_root)
+    state = safe_json_load(paths["state"]) or {}
+    return {"enabled": state.get("enabled", False)}
+
+
+def xdb_main(args) -> None:
+    project_root = Path(args.project).expanduser().resolve()
+    sub = args.sub or "status"
+    if sub == "enter":
+        result = xdb_enter(project_root)
+        print(f"airplan_mode=xdb\nstate_path={result['state_path']}")
+    else:
+        s = xdb_status(project_root)
+        print(f"airplan_mode=xdb\nenabled={s['enabled']}")
+
+
+# --- AirSDB ---
+
+def _sdb_paths(project_root: Path) -> dict[str, Path]:
+    root = airplan_root(project_root) / "state" / "airsdb"
+    return {"root": root, "state": root / "state.json", "reports_dir": root / "reports"}
+
+
+def sdb_enter(project_root: Path) -> dict:
+    paths = _sdb_paths(project_root)
+    paths["reports_dir"].mkdir(parents=True, exist_ok=True)
+    atomic_json_write(paths["state"], {"enabled": True, "updatedAt": now_iso()})
+    return {"state_path": str(paths["state"])}
+
+
+def sdb_main(args) -> None:
+    project_root = Path(args.project).expanduser().resolve()
+    sub = args.sub or "status"
+    from air_runtime.io import safe_json_load
+    paths = _sdb_paths(project_root)
+    if sub == "enter":
+        result = sdb_enter(project_root)
+        print(f"airplan_mode=sdb\nstate_path={result['state_path']}")
+    else:
+        state = safe_json_load(paths["state"]) or {}
+        print(f"airplan_mode=sdb\nenabled={state.get('enabled', False)}")
+
+
+# --- AirNDB ---
+
+def _ndb_paths(project_root: Path) -> dict[str, Path]:
+    root = airplan_root(project_root) / "state" / "airndb"
+    return {"root": root, "state": root / "state.json", "captures_dir": root / "captures"}
+
+
+def ndb_enter(project_root: Path) -> dict:
+    paths = _ndb_paths(project_root)
+    paths["captures_dir"].mkdir(parents=True, exist_ok=True)
+    atomic_json_write(paths["state"], {"enabled": True, "updatedAt": now_iso()})
+    return {"state_path": str(paths["state"])}
+
+
+def ndb_main(args) -> None:
+    project_root = Path(args.project).expanduser().resolve()
+    sub = args.sub or "status"
+    from air_runtime.io import safe_json_load
+    paths = _ndb_paths(project_root)
+    if sub == "enter":
+        result = ndb_enter(project_root)
+        print(f"airplan_mode=ndb\nstate_path={result['state_path']}")
+    else:
+        state = safe_json_load(paths["state"]) or {}
+        print(f"airplan_mode=ndb\nenabled={state.get('enabled', False)}")
--- a/lib/air_runtime/paths.py
+++ b/lib/air_runtime/paths.py
@@ -0,0 +1,99 @@
+"""
+路径约定 — V2 统一所有子模块的 AirPlan 目录结构。
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+
+def airplan_root(project_root: Path) -> Path:
+    return project_root / "AirPlan"
+
+
+def state_root(project_root: Path) -> Path:
+    return airplan_root(project_root) / "state"
+
+
+def todo_path(project_root: Path) -> Path:
+    return airplan_root(project_root) / "todo.md"
+
+
+def plan_path(project_root: Path) -> Path:
+    return airplan_root(project_root) / "plan.md"
+
+
+def agents_path(project_root: Path) -> Path:
+    return airplan_root(project_root) / "AGENTS.md"
+
+
+def docs_root(project_root: Path) -> Path:
+    return airplan_root(project_root) / "docs"
+
+
+# --- 子模块状态路径 ---
+
+def engine_state_path(project_root: Path) -> Path:
+    return state_root(project_root) / "aireng" / "state.json"
+
+
+def worker_state_path(project_root: Path, task_id: str) -> Path:
+    return state_root(project_root) / "airdo" / "tasks" / task_id / "worker-state.json"
+
+
+def arc_state_path(project_root: Path) -> Path:
+    return state_root(project_root) / "airarc" / "state.json"
+
+
+def dbg_state_path(project_root: Path) -> Path:
+    return state_root(project_root) / "airdbg" / "state.json"
+
+
+def xdb_state_path(project_root: Path) -> Path:
+    return state_root(project_root) / "airxdb" / "state.json"
+
+
+def sdb_state_path(project_root: Path) -> Path:
+    return state_root(project_root) / "airsdb" / "state.json"
+
+
+def ndb_state_path(project_root: Path) -> Path:
+    return state_root(project_root) / "airndb" / "state.json"
+
+
+def ctx_state_path(project_root: Path) -> Path:
+    return state_root(project_root) / "aircontext" / "state.json"
+
+
+def dep_state_path(project_root: Path) -> Path:
+    return state_root(project_root) / "airdep" / "state.json"
+
+
+def tst_state_path(project_root: Path) -> Path:
+    return state_root(project_root) / "airtst" / "state.json"
+
+
+def sec_state_path(project_root: Path) -> Path:
+    return state_root(project_root) / "airsec" / "state.json"
+
+
+def rvr_state_path(project_root: Path) -> Path:
+    return state_root(project_root) / "airrvr" / "state.json"
+
+
+def event_log_path(project_root: Path) -> Path:
+    return state_root(project_root) / "events.jsonl"
+
+
+def task_graph_state_path(project_root: Path) -> Path:
+    return state_root(project_root) / "task-graph.json"
+
+
+def required_project_artifacts() -> list[str]:
+    return [
+        "AirPlan/AGENTS.md",
+        "AirPlan/plan.md",
+        "AirPlan/todo.md",
+        "AirPlan/docs/architecture/adr/",
+        "AirPlan/docs/architecture/c4/module.md",
+    ]
--- a/lib/air_runtime/project_bootstrap.py
+++ b/lib/air_runtime/project_bootstrap.py
@@ -0,0 +1,59 @@
+"""
+项目引导模块 — 确保 AirPlan 目录结构存在。
+V2 保持与 V1 相同的不变量：制品驱动通信、上下文隔离。
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+
+def ensure_project_bootstrap(project_root: Path) -> dict[str, bool]:
+    """创建 AirPlan 必需目录结构。"""
+    root = project_root / "AirPlan"
+    docs = root / "docs"
+    arch = docs / "architecture"
+    adr_dir = arch / "adr"
+    c4_dir = arch / "c4"
+    debug_dir = docs / "debug"
+    state = root / "state"
+
+    dirs = [
+        root,
+        docs,
+        arch,
+        adr_dir,
+        c4_dir,
+        debug_dir,
+        state,
+        state / "airarc" / "reviews",
+        state / "aireng" / "dispatch",
+        state / "aireng" / "archive",
+        state / "aireng" / "plans",
+        state / "airdo" / "tasks",
+        state / "airdbg" / "sessions",
+        state / "airdbg" / "snapshots",
+        state / "airxdb" / "artifacts",
+        state / "airsdb" / "reports",
+        state / "airndb" / "captures",
+        state / "aircontext",
+        state / "airdep" / "sessions",
+        state / "airtst" / "reports",
+        state / "airsec",
+        state / "airrvr" / "reviews",
+    ]
+
+    for d in dirs:
+        d.mkdir(parents=True, exist_ok=True)
+
+    # 创建必要文件
+    (root / "AGENTS.md").touch()
+    (root / "plan.md").touch()
+    (root / "todo.md").touch()
+    (adr_dir / "placeholder.md").touch()
+    (c4_dir / "module.md").touch()
+    (debug_dir / "debug-log.md").touch()
+    (debug_dir / "gui-debug-log.md").touch()
+    (docs / "staticanalysis.md").touch()
+
+    return {"bootstrap": True}
--- a/lib/air_runtime/review.py
+++ b/lib/air_runtime/review.py
@@ -0,0 +1,126 @@
+"""
+并行审查模块 — V2 从 V1 迁移。
+分析任务依赖、写集冲突、产出并行组和串行点。
+"""
+
+from __future__ import annotations
+
+import re
+from dataclasses import dataclass, field
+from pathlib import Path
+from air_runtime.todo_parser import parse_tasks
+
+
+@dataclass
+class ParallelGroup:
+    name: str
+    task_ids: list[str]
+    reason: str = ""
+
+    def to_dict(self) -> dict:
+        return {"name": self.name, "task_ids": self.task_ids, "reason": self.reason}
+
+
+@dataclass
+class Conflict:
+    task_a: str
+    task_b: str
+    reason: str = ""
+
+    def to_dict(self) -> dict:
+        return {"task_a": self.task_a, "task_b": self.task_b, "reason": self.reason}
+
+
+@dataclass
+class ReviewResult:
+    parallel_groups: list[ParallelGroup] = field(default_factory=list)
+    conflicts: list[Conflict] = field(default_factory=list)
+    serialization_points: list[dict] = field(default_factory=list)
+    edges: list[dict] = field(default_factory=list)
+
+    def to_dict(self) -> dict:
+        return {
+            "parallelGroups": [{"name": g.name, "task_ids": g.task_ids, "reason": g.reason} for g in self.parallel_groups],
+            "conflicts": [{"task_a": c.task_a, "task_b": c.task_b, "reason": c.reason} for c in self.conflicts],
+            "serializationPoints": self.serialization_points,
+            "edges": self.edges,
+        }
+
+    @classmethod
+    def from_dict(cls, data: dict) -> ReviewResult:
+        return cls(
+            parallel_groups=[ParallelGroup(**g) for g in data.get("parallelGroups", [])],
+            conflicts=[Conflict(**c) for c in data.get("conflicts", [])],
+            serialization_points=data.get("serializationPoints", []),
+            edges=data.get("edges", []),
+        )
+
+
+def build_parallel_review(todo_path: Path) -> ReviewResult:
+    """分析 todo.md，产出并行组和冲突。"""
+    tasks = parse_tasks(todo_path)
+    result = ReviewResult()
+
+    # 解析依赖：task 文本中的 "依赖 T-xxx" 或 Done When 中的引用
+    edges = []
+    for t in tasks:
+        deps = re.findall(r"T-\d+[a-z]*", t.done_when)
+        deps.extend(re.findall(r"依赖\s+(T-\d+[a-z]*)", t.task))
+        for dep in deps:
+            if dep != t.task_id:
+                edges.append({"source": dep, "target": t.task_id, "kind": "dependency"})
+                result.edges.append({"source": dep, "target": t.task_id, "kind": "dependency"})
+
+    # 写集冲突检测
+    file_map: dict[str, list[str]] = {}
+    for t in tasks:
+        if t.files_dirs:
+            files = [f.strip() for f in t.files_dirs.split(",")]
+            for f in files:
+                file_map.setdefault(f, []).append(t.task_id)
+
+    conflicts = []
+    for fpath, tid_list in file_map.items():
+        for i, tid_a in enumerate(tid_list):
+            for tid_b in tid_list[i + 1:]:
+                conflicts.append(Conflict(tid_a, tid_b, f"shared file: {fpath}"))
+    result.conflicts = conflicts
+
+    # 串行点：同文件不同任务的依赖链
+    for fpath, tid_list in file_map.items():
+        if len(tid_list) > 1:
+            for tid in tid_list[1:]:
+                result.serialization_points.append({
+                    "taskId": tid,
+                    "reasons": [f"serialized with {tid_list[0]} due to shared file: {fpath}"],
+                })
+
+    # 并行组：入度为 0 的任务
+    target_count = {e["target"] for e in edges}
+    ready = [t.task_id for t in tasks if t.task_id not in target_count and t.status == "TODO"]
+    if ready:
+        result.parallel_groups.append(ParallelGroup(
+            name="wave-1", task_ids=ready,
+            reason="no dependencies on other TODO tasks",
+        ))
+
+    return result
+
+
+def render_review_markdown(review: ReviewResult) -> str:
+    lines = ["# AirArc Parallel Review", ""]
+    lines.append(f"## Summary")
+    lines.append(f"- Parallel groups: {len(review.parallel_groups)}")
+    lines.append(f"- Conflicts: {len(review.conflicts)}")
+    lines.append(f"- Serialization points: {len(review.serialization_points)}")
+    lines.append("")
+    lines.append("## Parallel Groups")
+    for g in review.parallel_groups:
+        lines.append(f"### {g.name}")
+        lines.append(f"Reason: {g.reason}")
+        lines.append(f"Tasks: {', '.join(g.task_ids)}")
+        lines.append("")
+    lines.append("## Conflicts")
+    for c in review.conflicts:
+        lines.append(f"- {c.task_a} <-> {c.task_b}: {c.reason}")
+    return "\n".join(lines)
--- a/lib/air_runtime/review_runtime.py
+++ b/lib/air_runtime/review_runtime.py
@@ -0,0 +1,148 @@
+"""
+AirRvr 需求审查运行时 — V2 新增组件。
+基于原始需求文档对已完成任务进行独立审查，验证交付物与需求的一致性。
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from pathlib import Path
+
+from air_runtime.io import atomic_json_write
+from air_runtime.paths import rvr_state_path
+from air_runtime.utils import session_stamp
+
+
+@dataclass
+class RequirementCoverage:
+    requirement: str
+    status: str  # covered | partial | missing
+    evidence: str = ""
+
+
+@dataclass
+class CodeToDesignItem:
+    design_item: str
+    implementation_status: str  # aligned | divergent | missing
+    code_location: str = ""
+    design_location: str = ""
+    divergence_detail: str = ""
+
+
+@dataclass
+class ReviewReport:
+    task_id: str
+    verdict: str  # pass | conditional-pass | fail
+    coverage: list[RequirementCoverage] = field(default_factory=list)
+    intent_alignment: str = "aligned"  # aligned | divergent
+    divergence_notes: str = ""
+    regression_risk: str = "none"  # none | low | medium | high
+    code_quality: dict = field(default_factory=lambda: {
+        "complexity": "low", "readability": "good", "duplication": "none", "error_handling": "complete",
+    })
+    lifecycle_health: dict = field(default_factory=lambda: {
+        "resource_leak": "none", "connection_management": "proper",
+        "timeout_strategy": "present", "retry_strategy": "present",
+    })
+    runtime_stability: dict = field(default_factory=lambda: {
+        "crash_risk": "none", "race_condition": "none",
+        "memory_leak": "none", "user_impact": "none",
+    })
+    code_to_design_table: list[CodeToDesignItem] = field(default_factory=list)
+    logging_checks: dict = field(default_factory=lambda: {
+        "spdlog_integrated": False,
+        "non_standard_logging": [],
+        "debug_release_switch": False,
+        "critical_path_logging": False,
+        "unified_format": False,
+    })
+    recommendations: list[str] = field(default_factory=list)
+
+
+class ReviewRuntime:
+    """AirRvr 审查运行时 — 管理审查会话和报告持久化。"""
+
+    REVIEW_MODES = ["per-task", "per-wave", "per-milestone"]
+
+    def __init__(self, project_root: Path):
+        self._project_root = project_root
+        self._state_dir = rvr_state_path(project_root).parent
+        self._reviews_dir = self._state_dir / "reviews"
+        self._reviews_dir.mkdir(parents=True, exist_ok=True)
+
+    def save_report(self, report: ReviewReport) -> Path:
+        report_path = self._reviews_dir / f"{report.task_id}-{session_stamp()}.json"
+        atomic_json_write(report_path, self._report_to_dict(report))
+        return report_path
+
+    def load_report(self, task_id: str, timestamp: str) -> ReviewReport | None:
+        from air_runtime.io import safe_json_load
+        report_path = self._reviews_dir / f"{task_id}-{timestamp}.json"
+        data = safe_json_load(report_path)
+        if data:
+            return self._dict_to_report(data)
+        return None
+
+    def get_integration_verdict(self, report: ReviewReport) -> str:
+        """与 AirEng 集成：pass → 允许合并，conditional-pass → 合并但记录遗留项，fail → 阻止合并。"""
+        return report.verdict
+
+    def get_verdict_for_task(self, task_id: str) -> dict:
+        """从持久化的 review report 读 verdict，返回 dict 含 verdict/residual/reportPath。
+        没有 report 时返回 {"verdict": "pass", "reportPath": ""}（默认放行）。"""
+        from air_runtime.io import safe_json_load
+        # reports/ 是 AirEng 约定的存放路径（验证脚本和 eng_mode 期望的位置）
+        reports_dir = self._state_dir / "reports"
+        report_path = reports_dir / f"{task_id}.json"
+        if not report_path.exists():
+            # 兼容旧路径 reviews/ 下的 {task_id}-{ts}.json，找最新一份
+            alt = self._reviews_dir
+            if alt.exists():
+                candidates = sorted(alt.glob(f"{task_id}-*.json"), key=lambda p: p.stat().st_mtime, reverse=True)
+                if candidates:
+                    report_path = candidates[0]
+            if not report_path.exists():
+                return {"verdict": "pass", "reportPath": "", "residual": []}
+        report = safe_json_load(report_path)
+        if not report or not isinstance(report, dict):
+            return {"verdict": "pass", "reportPath": str(report_path), "residual": []}
+        return {
+            "verdict": report.get("verdict", "pass"),
+            "residual": report.get("residual", []),
+            "reportPath": str(report_path),
+            "summary": report.get("summary", ""),
+        }
+
+    @staticmethod
+    def _report_to_dict(report: ReviewReport) -> dict:
+        return {
+            "taskId": report.task_id,
+            "verdict": report.verdict,
+            "coverage": [c.__dict__ for c in report.coverage],
+            "intentAlignment": report.intent_alignment,
+            "divergenceNotes": report.divergence_notes,
+            "regressionRisk": report.regression_risk,
+            "codeQuality": report.code_quality,
+            "lifecycleHealth": report.lifecycle_health,
+            "runtimeStability": report.runtime_stability,
+            "codeToDesignTable": [c.__dict__ for c in report.code_to_design_table],
+            "loggingChecks": report.logging_checks,
+            "recommendations": report.recommendations,
+        }
+
+    @staticmethod
+    def _dict_to_report(data: dict) -> ReviewReport:
+        return ReviewReport(
+            task_id=data.get("taskId", ""),
+            verdict=data.get("verdict", "fail"),
+            coverage=[RequirementCoverage(**c) for c in data.get("coverage", [])],
+            intent_alignment=data.get("intentAlignment", "aligned"),
+            divergence_notes=data.get("divergenceNotes", ""),
+            regression_risk=data.get("regressionRisk", "none"),
+            code_quality=data.get("codeQuality", {}),
+            lifecycle_health=data.get("lifecycleHealth", {}),
+            runtime_stability=data.get("runtimeStability", {}),
+            code_to_design_table=[CodeToDesignItem(**c) for c in data.get("codeToDesignTable", [])],
+            logging_checks=data.get("loggingChecks", {}),
+            recommendations=data.get("recommendations", []),
+        )
--- a/lib/air_runtime/sdb_backends.py
+++ b/lib/air_runtime/sdb_backends.py
@@ -0,0 +1,527 @@
+"""AirSDB backends — 5 static analyzer backends + AnalysisDiff.
+
+Backends:
+  CppcheckBackend   — C/C++ via cppcheck
+  ClangTidyBackend  — C/C++ via clang-tidy
+  RustClippyBackend — Rust via cargo clippy
+  GoVetBackend      — Go via go vet + staticcheck
+  TypeScriptBackend — TypeScript via tsc --noEmit
+"""
+
+from __future__ import annotations
+
+import json
+import shutil
+import subprocess
+from abc import ABC, abstractmethod
+from dataclasses import dataclass, field
+from pathlib import Path
+
+
+# ---------------------------------------------------------------------------
+# Unified result type
+# ---------------------------------------------------------------------------
+
+@dataclass(frozen=True)
+class AnalysisResult:
+    tool: str
+    file: str
+    line: int | None
+    column: int | None
+    severity: str  # error | warning | info
+    message: str
+    rule_id: str | None = None
+
+
+# ---------------------------------------------------------------------------
+# Abstract base
+# ---------------------------------------------------------------------------
+
+class StaticAnalyzerBackend(ABC):
+    """Abstract base for every static-analysis backend."""
+
+    @abstractmethod
+    def analyze(self, project_root: Path, target: Path | None = None) -> list[AnalysisResult]:
+        """Run the analyzer and return structured findings."""
+
+    @property
+    @abstractmethod
+    def name(self) -> str:
+        """Short identifier for this backend (e.g. 'cppcheck')."""
+
+    @property
+    @abstractmethod
+    def install_hint(self) -> str:
+        """Human-readable hint shown when the tool is not installed."""
+
+    # -- helpers available to all backends ---------------------------------
+
+    def _check_tool(self, tool_cmd: str) -> None:
+        """Raise RuntimeError if *tool_cmd* is not on PATH."""
+        if not shutil.which(tool_cmd):
+            raise RuntimeError(self.install_hint)
+
+    @staticmethod
+    def _run(cmd: list[str], *, cwd: Path | None = None) -> subprocess.CompletedProcess:
+        """Run *cmd* and capture stdout/stderr. Returns CompletedProcess."""
+        return subprocess.run(
+            cmd,
+            cwd=cwd,
+            capture_output=True,
+            text=True,
+            timeout=300,
+        )
+
+
+# ---------------------------------------------------------------------------
+# CppcheckBackend
+# ---------------------------------------------------------------------------
+
+class CppcheckBackend(StaticAnalyzerBackend):
+    """C/C++ static analysis via cppcheck."""
+
+    name = "cppcheck"
+    install_hint = (
+        "cppcheck is not installed. "
+        "Install it with: sudo apt install cppcheck  (Debian/Ubuntu)  "
+        "or: brew install cppcheck  (macOS)"
+    )
+
+    # Template: file:line:column:severity:id:message
+    _TEMPLATE = "{file}:{line}:{column}:{severity}:{id}:{message}"
+
+    def analyze(self, project_root: Path, target: Path | None = None) -> list[AnalysisResult]:
+        self._check_tool("cppcheck")
+
+        src = str(target) if target else str(project_root)
+        cmd = [
+            "cppcheck",
+            "--quiet",
+            f"--template={self._TEMPLATE}",
+            "--force",
+            src,
+        ]
+        proc = self._run(cmd, cwd=project_root)
+
+        results: list[AnalysisResult] = []
+        for line in proc.stderr.splitlines():
+            line = line.strip()
+            if not line:
+                continue
+            parts = line.split(":", 5)
+            if len(parts) < 6:
+                continue
+            try:
+                ln = int(parts[1]) if parts[1].strip() else None
+            except ValueError:
+                ln = None
+            try:
+                col = int(parts[2]) if parts[2].strip() else None
+            except ValueError:
+                col = None
+
+            severity = parts[3].strip()
+            # Map cppcheck severities to our unified set
+            if severity not in ("error", "warning", "info"):
+                if severity in ("performance", "portability", "style"):
+                    severity = "warning"
+                else:
+                    severity = "info"
+
+            results.append(AnalysisResult(
+                tool=self.name,
+                file=parts[0].strip(),
+                line=ln,
+                column=col,
+                severity=severity,
+                message=parts[5].strip(),
+                rule_id=parts[4].strip() or None,
+            ))
+        return results
+
+
+# ---------------------------------------------------------------------------
+# ClangTidyBackend
+# ---------------------------------------------------------------------------
+
+class ClangTidyBackend(StaticAnalyzerBackend):
+    """C/C++ static analysis via clang-tidy."""
+
+    name = "clang-tidy"
+    install_hint = (
+        "clang-tidy is not installed. "
+        "Install it with: sudo apt install clang-tidy  (Debian/Ubuntu)  "
+        "or: brew install llvm  (macOS, then use llvm/bin/clang-tidy)"
+    )
+
+    def analyze(self, project_root: Path, target: Path | None = None) -> list[AnalysisResult]:
+        self._check_tool("clang-tidy")
+
+        src = str(target) if target else str(project_root)
+        cmd = [
+            "clang-tidy",
+            "--quiet",
+            src,
+        ]
+        # Use compile_commands.json if present
+        comp_db = project_root / "compile_commands.json"
+        if comp_db.exists():
+            cmd.append(f"-p={comp_db.parent}")
+
+        proc = self._run(cmd, cwd=project_root)
+
+        results: list[AnalysisResult] = []
+        # clang-tidy output format: <file>:<line>:<col>: warning: <message> [check-name]
+        for line in proc.stderr.splitlines():
+            line = line.strip()
+            if not line:
+                continue
+            parts = line.split(":", 3)
+            if len(parts) < 4:
+                continue
+            try:
+                ln = int(parts[1].strip()) if parts[1].strip() else None
+            except ValueError:
+                ln = None
+            try:
+                col = int(parts[2].strip()) if parts[2].strip() else None
+            except ValueError:
+                col = None
+
+            msg_part = parts[3].strip()
+            severity = "warning"
+            # Detect "error:" prefix
+            if msg_part.startswith("error:"):
+                severity = "error"
+                msg_part = msg_part[len("error:"):].strip()
+            elif msg_part.startswith("warning:"):
+                msg_part = msg_part[len("warning:"):].strip()
+            elif msg_part.startswith("note:"):
+                severity = "info"
+                msg_part = msg_part[len("note:"):].strip()
+
+            # Extract [check-name] at the end
+            rule_id = None
+            if msg_part.endswith("]"):
+                bracket = msg_part.rfind("[")
+                if bracket != -1:
+                    rule_id = msg_part[bracket + 1:-1].strip()
+                    msg_part = msg_part[:bracket].strip()
+
+            results.append(AnalysisResult(
+                tool=self.name,
+                file=parts[0].strip(),
+                line=ln,
+                column=col,
+                severity=severity,
+                message=msg_part,
+                rule_id=rule_id,
+            ))
+        return results
+
+
+# ---------------------------------------------------------------------------
+# RustClippyBackend
+# ---------------------------------------------------------------------------
+
+class RustClippyBackend(StaticAnalyzerBackend):
+    """Rust static analysis via cargo clippy."""
+
+    name = "clippy"
+    install_hint = (
+        "cargo clippy is not available. "
+        "Install Rust toolchain: curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh  "
+        "then: rustup component add clippy"
+    )
+
+    def analyze(self, project_root: Path, target: Path | None = None) -> list[AnalysisResult]:
+        self._check_tool("cargo")
+
+        cmd = [
+            "cargo", "clippy",
+            "--message-format=json",
+        ]
+        # If a specific target file/dir is given, we still run cargo clippy
+        # on the whole crate (cargo does not support single-file analysis).
+        proc = self._run(cmd, cwd=project_root)
+
+        results: list[AnalysisResult] = []
+        for line in proc.stdout.splitlines():
+            line = line.strip()
+            if not line:
+                continue
+            try:
+                obj = json.loads(line)
+            except json.JSONDecodeError:
+                continue
+            if obj.get("reason") != "compiler-message":
+                continue
+            msg = obj.get("message", {})
+            level = msg.get("level", "")
+            if level == "error":
+                severity = "error"
+            elif level in ("warning",):
+                severity = "warning"
+            else:
+                severity = "info"
+
+            for span in msg.get("spans", []):
+                results.append(AnalysisResult(
+                    tool=self.name,
+                    file=span.get("file_name", ""),
+                    line=span.get("line_start"),
+                    column=span.get("column_start"),
+                    severity=severity,
+                    message=msg.get("message", ""),
+                    rule_id=msg.get("code", {}).get("code") or None,
+                ))
+
+        # If no JSON output (e.g. compile error), also parse stderr
+        if not results and proc.stderr:
+            for line in proc.stderr.splitlines():
+                line = line.strip()
+                if "error" in line.lower() and ":" in line:
+                    results.append(AnalysisResult(
+                        tool=self.name,
+                        file=str(project_root),
+                        line=None,
+                        column=None,
+                        severity="error",
+                        message=line,
+                        rule_id=None,
+                    ))
+        return results
+
+
+# ---------------------------------------------------------------------------
+# GoVetBackend
+# ---------------------------------------------------------------------------
+
+class GoVetBackend(StaticAnalyzerBackend):
+    """Go static analysis via go vet + staticcheck."""
+
+    name = "go-vet"
+    install_hint = (
+        "go is not installed. "
+        "Install Go: https://go.dev/dl/  "
+        "For staticcheck: go install honnef.co/go/tools/cmd/staticcheck@latest"
+    )
+
+    def analyze(self, project_root: Path, target: Path | None = None) -> list[AnalysisResult]:
+        self._check_tool("go")
+
+        results: list[AnalysisResult] = []
+
+        # 1) go vet  — JSON output
+        vet_cmd = ["go", "vet", "./..."]
+        proc = self._run(vet_cmd, cwd=project_root)
+        if proc.stderr:
+            results.extend(self._parse_go_vet_output(proc.stderr))
+
+        # 2) staticcheck (optional — don't fail if not installed)
+        if shutil.which("staticcheck"):
+            sc_cmd = ["staticcheck", "-f=json", "./..."]
+            sc_proc = self._run(sc_cmd, cwd=project_root)
+            for line in sc_proc.stdout.splitlines():
+                line = line.strip()
+                if not line:
+                    continue
+                try:
+                    obj = json.loads(line)
+                except json.JSONDecodeError:
+                    continue
+                severity = "warning"
+                if obj.get("severity", "") == "error":
+                    severity = "error"
+                results.append(AnalysisResult(
+                    tool="staticcheck",
+                    file=obj.get("location", {}).get("file", ""),
+                    line=obj.get("location", {}).get("line"),
+                    column=obj.get("location", {}).get("column"),
+                    severity=severity,
+                    message=obj.get("message", ""),
+                    rule_id=obj.get("code", ""),
+                ))
+
+        return results
+
+    @staticmethod
+    def _parse_go_vet_output(text: str) -> list[AnalysisResult]:
+        """Parse go vet stderr output.
+
+        go vet output format (non-JSON):
+          <file>:<line>: <message>
+        """
+        results: list[AnalysisResult] = []
+        for line in text.splitlines():
+            line = line.strip()
+            if not line:
+                continue
+            parts = line.split(":", 2)
+            if len(parts) < 3:
+                continue
+            try:
+                ln = int(parts[1].strip()) if parts[1].strip() else None
+            except ValueError:
+                ln = None
+            results.append(AnalysisResult(
+                tool="go vet",
+                file=parts[0].strip(),
+                line=ln,
+                column=None,
+                severity="warning",
+                message=parts[2].strip(),
+                rule_id=None,
+            ))
+        return results
+
+
+# ---------------------------------------------------------------------------
+# TypeScriptBackend
+# ---------------------------------------------------------------------------
+
+class TypeScriptBackend(StaticAnalyzerBackend):
+    """TypeScript static analysis via tsc --noEmit."""
+
+    name = "tsc"
+    install_hint = (
+        "tsc (TypeScript compiler) is not installed. "
+        "Install it with: npm install -g typescript  "
+        "or add it to your project: npm install --save-dev typescript"
+    )
+
+    def analyze(self, project_root: Path, target: Path | None = None) -> list[AnalysisResult]:
+        # tsc can be installed locally (npx) or globally
+        tsc_cmd = self._find_tsc()
+        if tsc_cmd is None:
+            raise RuntimeError(self.install_hint)
+
+        cmd = tsc_cmd + ["--noEmit", "--pretty", "false"]
+        proc = self._run(cmd, cwd=project_root)
+
+        results: list[AnalysisResult] = []
+        # tsc output format: <file>(<line>,<col>): error TS<code>: <message>
+        for line in proc.stdout.splitlines():
+            line = line.strip()
+            if not line:
+                continue
+            results.append(self._parse_tsc_line(line))
+        return results
+
+    def _find_tsc(self) -> list[str] | None:
+        """Return the tsc command as a list, or None if not found."""
+        if shutil.which("tsc"):
+            return ["tsc"]
+        if shutil.which("npx"):
+            return ["npx", "tsc"]
+        return None
+
+    @staticmethod
+    def _parse_tsc_line(line: str) -> AnalysisResult:
+        """Parse a single tsc diagnostic line.
+
+        Format: <file>(<line>,<col>): error TS1234: <message>
+        """
+        severity = "error"
+        rule_id = None
+
+        # Split on the first colon-space after the position paren
+        # e.g. "src/foo.ts(10,5): error TS2322: Type 'string' ..."
+        main_parts = line.split(": ", 1)
+        location_part = main_parts[0] if main_parts else line
+        message = main_parts[1].strip() if len(main_parts) > 1 else ""
+
+        # Extract file, line, column from "file(line,col)"
+        file_part = location_part
+        ln = None
+        col = None
+        paren = location_part.rfind("(")
+        if paren != -1 and location_part.endswith(")"):
+            file_part = location_part[:paren]
+            pos_str = location_part[paren + 1:-1]
+            pos_parts = pos_str.split(",", 1)
+            try:
+                ln = int(pos_parts[0].strip()) if pos_parts[0].strip() else None
+            except ValueError:
+                pass
+            if len(pos_parts) > 1:
+                try:
+                    col = int(pos_parts[1].strip()) if pos_parts[1].strip() else None
+                except ValueError:
+                    pass
+
+        # Extract severity + rule from " error TS2322" in the remainder
+        if len(main_parts) > 1:
+            # The part between the first colon-space and the message
+            # is in the original line — re-parse
+            rest = line[len(location_part) + 2:]  # after ": "
+            if rest.startswith("error "):
+                severity = "error"
+                rest = rest[len("error "):]
+            elif rest.startswith("warning "):
+                severity = "warning"
+                rest = rest[len("warning "):]
+            # rest now starts with "TS1234: message"
+            ts_parts = rest.split(": ", 1)
+            if ts_parts:
+                rule_id = ts_parts[0].strip() or None
+                if len(ts_parts) > 1:
+                    message = ts_parts[1].strip()
+
+        return AnalysisResult(
+            tool="tsc",
+            file=file_part,
+            line=ln,
+            column=col,
+            severity=severity,
+            message=message,
+            rule_id=rule_id,
+        )
+
+
+# ---------------------------------------------------------------------------
+# AnalysisDiff
+# ---------------------------------------------------------------------------
+
+class AnalysisDiff:
+    """Compare two lists of AnalysisResult and classify findings as
+    new, resolved, or unchanged."""
+
+    @staticmethod
+    def _key(r: AnalysisResult) -> tuple[str, int | None, str | None]:
+        """Dedup key: (file, line, rule_id)."""
+        return (r.file, r.line, r.rule_id)
+
+    def diff(
+        self,
+        before: list[AnalysisResult],
+        after: list[AnalysisResult],
+    ) -> dict:
+        before_keys = {self._key(r): r for r in before}
+        after_keys = {self._key(r): r for r in after}
+
+        before_set = set(before_keys.keys())
+        after_set = set(after_keys.keys())
+
+        new_keys = after_set - before_set
+        resolved_keys = before_set - after_set
+        unchanged_keys = before_set & after_set
+
+        return {
+            "new": [after_keys[k] for k in new_keys],
+            "resolved": [before_keys[k] for k in resolved_keys],
+            "unchanged": [before_keys[k] for k in unchanged_keys],
+        }
+
+
+# ---------------------------------------------------------------------------
+# Convenience registry
+# ---------------------------------------------------------------------------
+
+BACKENDS: dict[str, StaticAnalyzerBackend] = {
+    "cppcheck": CppcheckBackend(),
+    "clang-tidy": ClangTidyBackend(),
+    "clippy": RustClippyBackend(),
+    "go-vet": GoVetBackend(),
+    "tsc": TypeScriptBackend(),
+}
--- a/lib/air_runtime/sec_runtime.py
+++ b/lib/air_runtime/sec_runtime.py
@@ -0,0 +1,175 @@
+"""
+AirSec 安全扫描运行时 — V2 新增组件。
+制品敏感数据扫描 + 自动脱敏 + 误报白名单 + 确认流程 + advisory/blocking 模式。
+"""
+
+from __future__ import annotations
+
+import re
+from dataclasses import dataclass, field
+from pathlib import Path
+
+SECRET_PATTERNS: list[tuple[str, str]] = [
+    ("api_key", r'(?:api[_-]?key|apikey)\s*[:=]\s*["\']?([A-Za-z0-9_\-]{16,})["\']?'),
+    ("aws_key", r'AKIA[0-9A-Z]{16}'),
+    ("private_key", r'-----BEGIN (?:RSA|EC|DSA|OPENSSH) PRIVATE KEY-----'),
+    ("token", r'(?:token|secret|password)\s*[:=]\s*["\']?([^\s"\']{8,})["\']?'),
+    ("jwt", r'eyJ[A-Za-z0-9_\-]+\.eyJ[A-Za-z0-9_\-]+\.[A-Za-z0-9_\-]+'),
+    ("url_credential", r'https?://[^:@]+:([^@]+)@'),
+]
+
+ALLOWLIST_PATTERNS: list[str] = [
+    r'EXAMPLE',
+    r'example',
+    r'YOUR_API_KEY',
+    r'TODO',
+    r'<your-',
+    r'placeholder',
+    r'xxxxxxxx',
+]
+
+# 文件名白名单：命中则跳过该文件的所有发现
+WHITELIST_FILE_PATTERNS: list[str] = [
+    r"\.example\.",
+    r"^test_",
+    r"^mock_",
+    r"_test\.py$",
+    r"\.fixture\.",
+]
+
+# 用户确认记录（首次发现需确认，后续同模式自动放行）
+USER_CONFIRMATIONS: dict[str, dict] = {}  # {fingerprint: {pattern, confirmed_at, user}}
+
+
+class ScanMode:
+    """扫描模式：advisory（只报告，不阻止）/ blocking（阻止合并）"""
+    ADVISORY = "advisory"
+    BLOCKING = "blocking"
+
+
+@dataclass
+class ScanFinding:
+    rule: str
+    file: str
+    line: int
+    match: str  # 截断显示，不包含完整密钥
+    severity: str = "high"  # high | medium | low
+
+
+@dataclass
+class ScanReport:
+    task_id: str
+    findings: list[ScanFinding] = field(default_factory=list)
+    whitelisted: int = 0
+    clean: bool = True
+    advisory_blocked: bool = False
+
+
+def scan_file(file_path: Path, task_id: str = "") -> ScanReport:
+    findings: list[ScanFinding] = []
+    whitelisted = 0
+
+    try:
+        content = file_path.read_text(encoding="utf-8", errors="replace")
+    except Exception:
+        return ScanReport(task_id=task_id, clean=True)
+
+    for rule, pattern in SECRET_PATTERNS:
+        for match in re.finditer(pattern, content, re.IGNORECASE):
+            matched_text = match.group(0)
+            if any(re.search(ap, matched_text) for ap in ALLOWLIST_PATTERNS):
+                whitelisted += 1
+                continue
+
+            line_no = content[:match.start()].count("\n") + 1
+            display = matched_text[:60] + "..." if len(matched_text) > 60 else matched_text
+            findings.append(ScanFinding(
+                rule=rule, file=str(file_path), line=line_no, match=display,
+            ))
+
+    return ScanReport(
+        task_id=task_id,
+        findings=findings,
+        whitelisted=whitelisted,
+        clean=len(findings) == 0,
+    )
+
+
+def scan_result_data(result: dict, task_id: str = "") -> ScanReport:
+    """扫描 Worker result.json 中的敏感数据。"""
+    import json
+    text = json.dumps(result, ensure_ascii=False)
+    findings: list[ScanFinding] = []
+    whitelisted = 0
+
+    for rule, pattern in SECRET_PATTERNS:
+        for match in re.finditer(pattern, text, re.IGNORECASE):
+            matched_text = match.group(0)
+            if any(re.search(ap, matched_text) for ap in ALLOWLIST_PATTERNS):
+                whitelisted += 1
+                continue
+            display = matched_text[:60] + "..." if len(matched_text) > 60 else matched_text
+            findings.append(ScanFinding(
+                rule=rule, file="result.json", line=0, match=display,
+            ))
+
+    return ScanReport(
+        task_id=task_id,
+        findings=findings,
+        whitelisted=whitelisted,
+        clean=len(findings) == 0,
+    )
+
+
+def scan_file_with_mode(
+    file_path: Path,
+    task_id: str = "",
+    mode: str = ScanMode.BLOCKING,
+    confirm_callback=None,  # 可选：首次发现时调用此回调询问用户
+) -> ScanReport:
+    """
+    增强版扫描：
+    1. 基础扫描（已有逻辑）
+    2. 文件名白名单过滤
+    3. 模式判断（advisory vs blocking）
+    """
+    report = scan_file(file_path, task_id)  # 原有逻辑
+
+    # 文件名白名单过滤
+    filtered_findings = []
+    for f in report.findings:
+        filename = file_path.name
+        if any(re.search(p, filename) for p in WHITELIST_FILE_PATTERNS):
+            report.whitelisted += 1
+            continue
+        filtered_findings.append(f)
+
+    report.findings = filtered_findings
+    report.clean = len(filtered_findings) == 0
+
+    # 模式处理
+    if not report.clean and mode == ScanMode.ADVISORY:
+        # advisory 模式：只记录，不阻止
+        report.advisory_blocked = False
+    elif not report.clean and mode == ScanMode.BLOCKING:
+        # blocking 模式：默认阻止
+        report.advisory_blocked = True
+
+    return report
+
+
+def confirm_pattern(task_id: str, pattern: str, user: str = "unknown") -> None:
+    """用户确认某模式为安全后，记录下来"""
+    from air_runtime.utils import now_iso
+    fingerprint = f"{task_id}:{pattern}"
+    USER_CONFIRMATIONS[fingerprint] = {
+        "pattern": pattern,
+        "confirmed_at": now_iso(),
+        "user": user,
+    }
+
+
+def is_confirmed(task_id: str, pattern: str) -> bool:
+    """检查某模式是否已被用户确认"""
+    fingerprint = f"{task_id}:{pattern}"
+    return fingerprint in USER_CONFIRMATIONS
--- a/lib/air_runtime/task_graph.py
+++ b/lib/air_runtime/task_graph.py
@@ -0,0 +1,220 @@
+"""
+动态任务依赖图（DAG）— V2 P1-14 修复。
+替代 V1 静态 todo.md 表格，支持 Arc 增量重规划，Eng 增量吸收。
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from typing import Any
+
+
+@dataclass
+class TaskNode:
+    id: str
+    status: str = "TODO"  # TODO | DISPATCHED | DONE | BLOCKED
+    task: str = ""
+    files_dirs: str = ""
+    done_when: str = ""
+    in_degree: int = 0
+    out_edges: list[str] = field(default_factory=list)
+    write_set: list[str] = field(default_factory=list)
+    meta: dict[str, Any] = field(default_factory=dict)
+
+
+@dataclass
+class Edge:
+    source: str
+    target: str
+    kind: str = "dependency"  # dependency | conflict | sync
+
+
+@dataclass
+class EdgeChange:
+    added: list[Edge] = field(default_factory=list)
+    removed: list[Edge] = field(default_factory=list)
+
+
+@dataclass
+class PlanDelta:
+    """Arc 重规划产出的增量差异，替代全量覆盖 todo.md。"""
+    removed_tasks: list[str] = field(default_factory=list)
+    added_tasks: list[TaskNode] = field(default_factory=list)
+    modified_tasks: list[TaskNode] = field(default_factory=list)
+    edge_changes: EdgeChange = field(default_factory=EdgeChange)
+
+
+class TaskGraph:
+    """动态任务依赖图，支持增量更新和全量替换。"""
+
+    def __init__(self):
+        self.nodes: dict[str, TaskNode] = {}
+        self.edges: list[Edge] = []
+
+    def add_node(self, node: TaskNode) -> None:
+        self.nodes[node.id] = node
+
+    def add_edge(self, edge: Edge) -> None:
+        self.edges.append(edge)
+        if edge.target in self.nodes:
+            self.nodes[edge.target].in_degree += 1
+        if edge.source in self.nodes:
+            self.nodes[edge.source].out_edges.append(edge.target)
+
+    def apply_delta(self, delta: PlanDelta) -> None:
+        """增量吸收 Arc 的重规划结果，保留已调度任务不受影响。"""
+        for task_id in delta.removed_tasks:
+            self._remove_node(task_id)
+        for node in delta.added_tasks:
+            self._add_node(node)
+        for node in delta.modified_tasks:
+            self._update_node(node)
+        for edge in delta.edge_changes.removed:
+            self._remove_edge(edge)
+        for edge in delta.edge_changes.added:
+            self._add_edge(edge)
+
+    def apply_full_replace(self, nodes: list[TaskNode], edges: list[Edge]) -> None:
+        """全量替换模式：Arc 产出完整 DAG，保留已完成任务状态。"""
+        done_status = {tid: n.status for tid, n in self.nodes.items() if n.status == "DONE"}
+        self.nodes = {n.id: n for n in nodes}
+        self.edges = list(edges)
+        for tid, status in done_status.items():
+            if tid in self.nodes:
+                self.nodes[tid].status = status
+        for edge in self.edges:
+            if edge.target in self.nodes:
+                self.nodes[edge.target].in_degree += 1
+            if edge.source in self.nodes:
+                self.nodes[edge.source].out_edges.append(edge.target)
+
+    def ready_tasks(self) -> list[str]:
+        """返回当前入度为 0 且状态为 TODO 的任务。"""
+        return [nid for nid, n in self.nodes.items() if n.in_degree == 0 and n.status == "TODO"]
+
+    def diff(self, other: TaskGraph) -> PlanDelta:
+        """对比自身与 other，产出 PlanDelta（add/remove/modify node + edge changes）。
+
+        self = 新图, other = 旧图（before replan）。
+        """
+        delta = PlanDelta()
+        old_ids = set(other.nodes.keys())
+        new_ids = set(self.nodes.keys())
+
+        # 移除
+        delta.removed_tasks = list(old_ids - new_ids)
+
+        # 新增
+        delta.added_tasks = [self.nodes[tid] for tid in (new_ids - old_ids)]
+
+        # 修改
+        for tid in old_ids & new_ids:
+            old_n = other.nodes[tid]
+            new_n = self.nodes[tid]
+            if (old_n.task != new_n.task
+                    or old_n.files_dirs != new_n.files_dirs
+                    or old_n.done_when != new_n.done_when
+                    or old_n.write_set != new_n.write_set):
+                delta.modified_tasks.append(new_n)
+
+        # Edge 差异
+        old_edges = {(e.source, e.target, e.kind) for e in other.edges}
+        new_edges = {(e.source, e.target, e.kind) for e in self.edges}
+        for s, t, k in (new_edges - old_edges):
+            delta.edge_changes.added.append(Edge(source=s, target=t, kind=k))
+        for s, t, k in (old_edges - new_edges):
+            delta.edge_changes.removed.append(Edge(source=s, target=t, kind=k))
+
+        return delta
+
+    @classmethod
+    def load(cls, path) -> TaskGraph:
+        """从 _export_task_graph_json 写的格式还原 TaskGraph。"""
+        from pathlib import Path
+        from air_runtime.io import safe_json_load
+        p = Path(path)
+        data = safe_json_load(p)
+        graph = cls()
+        if not data or not isinstance(data, dict):
+            return graph
+        for nid, nd in data.get("nodes", {}).items():
+            graph.nodes[nid] = TaskNode(
+                id=nd.get("id", nid),
+                status=nd.get("status", "TODO"),
+                task=nd.get("task", ""),
+                files_dirs=nd.get("filesDirs", ""),
+                done_when=nd.get("doneWhen", ""),
+                in_degree=nd.get("inDegree", 0),
+                out_edges=list(nd.get("outEdges", [])),
+                write_set=list(nd.get("writeSet", [])),
+            )
+        for ed in data.get("edges", []):
+            graph.edges.append(Edge(
+                source=ed["source"], target=ed["target"],
+                kind=ed.get("kind", "dependency"),
+            ))
+        return graph
+
+    def task_ids_by_status(self, status: str) -> list[str]:
+        return [nid for nid, n in self.nodes.items() if n.status == status]
+
+    def find_cycles(self) -> list[list[str]]:
+        """检测依赖环（DFS）。"""
+        visited: set[str] = set()
+        rec_stack: set[str] = set()
+        cycles: list[list[str]] = []
+
+        def dfs(node_id: str, path: list[str]) -> None:
+            visited.add(node_id)
+            rec_stack.add(node_id)
+            path.append(node_id)
+            for target in self.nodes.get(node_id, TaskNode(id=node_id)).out_edges:
+                if target not in visited:
+                    dfs(target, path.copy())
+                elif target in rec_stack:
+                    cycle_start = path.index(target)
+                    cycles.append(path[cycle_start:])
+            rec_stack.discard(node_id)
+
+        for nid in self.nodes:
+            if nid not in visited:
+                dfs(nid, [])
+
+        return cycles
+
+    def export_todo_md(self) -> str:
+        """导出为人可读的 todo.md 表格，保留 V1 的可见性优势。"""
+        lines = ["| Task | Status | Files/Dirs | Done When | Validation | ADR |",
+                 "|------|--------|------------|-----------|------------|-----|"]
+        for nid, node in self.nodes.items():
+            lines.append(f"| {node.task} | {node.status} | {node.files_dirs} | "
+                        f"{node.done_when} | | |")
+        return "\n".join(lines) + "\n"
+
+    def _remove_node(self, task_id: str) -> None:
+        if task_id in self.nodes:
+            del self.nodes[task_id]
+        self.edges = [e for e in self.edges if e.source != task_id and e.target != task_id]
+
+    def _add_node(self, node: TaskNode) -> None:
+        self.nodes[node.id] = node
+
+    def _update_node(self, node: TaskNode) -> None:
+        if node.id in self.nodes:
+            existing_status = self.nodes[node.id].status
+            self.nodes[node.id] = node
+            if existing_status in ("DISPATCHED", "DONE"):
+                self.nodes[node.id].status = existing_status
+
+    def _remove_edge(self, edge: Edge) -> None:
+        self.edges = [e for e in self.edges
+                      if not (e.source == edge.source and e.target == edge.target)]
+        if edge.target in self.nodes:
+            self.nodes[edge.target].in_degree = max(0, self.nodes[edge.target].in_degree - 1)
+
+    def _add_edge(self, edge: Edge) -> None:
+        self.edges.append(edge)
+        if edge.target in self.nodes:
+            self.nodes[edge.target].in_degree += 1
+        if edge.source in self.nodes:
+            self.nodes[edge.source].out_edges.append(edge.target)
--- a/lib/air_runtime/test_runtime.py
+++ b/lib/air_runtime/test_runtime.py
@@ -0,0 +1,183 @@
+"""
+AirTst 测试运行器运行时 — V2 新增组件。
+统一测试执行接口，支持多框架，产出结构化结果。
+"""
+
+from __future__ import annotations
+
+import json
+import subprocess
+from dataclasses import dataclass, field
+from pathlib import Path
+
+from air_runtime.io import atomic_json_write
+from air_runtime.paths import tst_state_path
+
+
+@dataclass
+class TestCase:
+    name: str
+    status: str  # passed | failed | skipped
+    duration: str = ""
+    message: str = ""
+
+
+@dataclass
+class TestSuite:
+    name: str
+    total: int = 0
+    passed: int = 0
+    failed: int = 0
+    skipped: int = 0
+    cases: list[TestCase] = field(default_factory=list)
+
+
+@dataclass
+class TestRunResult:
+    framework: str
+    total: int = 0
+    passed: int = 0
+    failed: int = 0
+    disabled: int = 0
+    duration: str = ""
+    suites: list[TestSuite] = field(default_factory=list)
+    failures: list[dict] = field(default_factory=list)
+
+
+class TestRunner:
+    """统一测试执行器。"""
+
+    FRAMEWORKS = {
+        "pytest": ["python", "-m", "pytest", "--json-report", "-q"],
+        "googletest": ["ctest", "--output-on-failure"],
+        "jest": ["npx", "jest", "--json"],
+        "vitest": ["npx", "vitest", "run", "--reporter=json"],
+        "go": ["go", "test", "-json", "./..."],
+        "cargo": ["cargo", "test", "--", "--format=json"],
+    }
+
+    def run(self, task_id: str, project_root: Path, framework: str,
+            target_path: Path | None = None,
+            extra_args: list[str] | None = None) -> TestRunResult:
+        if framework not in self.FRAMEWORKS:
+            return TestRunResult(framework=framework, failures=[{"error": f"unsupported framework: {framework}"}])
+
+        cmd = list(self.FRAMEWORKS[framework])
+        if target_path:
+            cmd.append(str(target_path))
+        if extra_args:
+            cmd.extend(extra_args)
+
+        try:
+            result = subprocess.run(cmd, capture_output=True, text=True,
+                                   timeout=600, cwd=str(project_root))
+        except subprocess.TimeoutExpired:
+            return TestRunResult(framework=framework, failures=[{"error": "timeout"}])
+
+        run_result = self._parse_result(framework, result.stdout)
+        self._save_report(task_id, project_root, run_result)
+        return run_result
+
+    def _parse_result(self, framework: str, stdout: str) -> TestRunResult:
+        if framework == "pytest":
+            return self._parse_pytest(stdout)
+        if framework in ("jest", "vitest"):
+            return self._parse_jest(stdout)
+        if framework == "googletest":
+            return self._parse_googletest(stdout)
+        if framework == "go":
+            return self._parse_go(stdout)
+        if framework == "cargo":
+            return self._parse_cargo(stdout)
+        return TestRunResult(framework=framework, total=0)
+
+    def _parse_pytest(self, stdout: str) -> TestRunResult:
+        try:
+            data = json.loads(stdout)
+        except json.JSONDecodeError:
+            return TestRunResult(framework="pytest", failures=[{"error": "json parse failed"}])
+        return TestRunResult(
+            framework="pytest",
+            total=data.get("summary", {}).get("total", 0),
+            passed=data.get("summary", {}).get("passed", 0),
+            failed=data.get("summary", {}).get("failed", 0),
+            duration=str(data.get("duration", "")),
+        )
+
+    def _parse_jest(self, stdout: str) -> TestRunResult:
+        try:
+            data = json.loads(stdout)
+        except json.JSONDecodeError:
+            return TestRunResult(framework="jest", failures=[{"error": "json parse failed"}])
+        return TestRunResult(
+            framework="jest",
+            total=data.get("numTotalTests", 0),
+            passed=data.get("numPassedTests", 0),
+            failed=data.get("numFailedTests", 0),
+        )
+
+    def _parse_googletest(self, stdout: str) -> TestRunResult:
+        """解析 ctest 输出。ctest 不输出 JSON，从文本提取统计。"""
+        import re
+        total = passed = failed = disabled = 0
+        for line in stdout.splitlines():
+            m = re.match(r"(\d+)% tests passed, (\d+) tests failed out of (\d+)", line)
+            if m:
+                failed = int(m.group(2))
+                total = int(m.group(3))
+                passed = total - failed
+        # GoogleTest 也支持 --gtest_output=json
+        try:
+            data = json.loads(stdout)
+            if isinstance(data, dict):
+                total = sum(s.get("tests", 0) for s in data.get("testsuites", []))
+                failed = sum(s.get("failures", 0) for s in data.get("testsuites", []))
+                passed = total - failed
+                disabled = sum(s.get("disabled", 0) for s in data.get("testsuites", []))
+        except json.JSONDecodeError:
+            pass
+        return TestRunResult(
+            framework="googletest", total=total, passed=passed, failed=failed, disabled=disabled,
+        )
+
+    def _parse_go(self, stdout: str) -> TestRunResult:
+        """解析 go test -json 输出（JSONL 格式，每行一个事件）。"""
+        total = passed = failed = 0
+        for line in stdout.splitlines():
+            try:
+                ev = json.loads(line)
+            except json.JSONDecodeError:
+                continue
+            action = ev.get("Action", "")
+            if action == "pass":
+                passed += 1
+                total += 1
+            elif action == "fail":
+                failed += 1
+                total += 1
+            elif action == "skip":
+                total += 1
+        return TestRunResult(framework="go", total=total, passed=passed, failed=failed)
+
+    def _parse_cargo(self, stdout: str) -> TestRunResult:
+        """解析 cargo test --format=json 输出（JSONL 格式）。"""
+        total = passed = failed = 0
+        for line in stdout.splitlines():
+            try:
+                ev = json.loads(line)
+            except json.JSONDecodeError:
+                continue
+            if ev.get("type") == "test":
+                total += 1
+                if ev.get("event") == "ok":
+                    passed += 1
+                elif ev.get("event") == "failed":
+                    failed += 1
+        return TestRunResult(framework="cargo", total=total, passed=passed, failed=failed)
+
+    def _save_report(self, task_id: str, project_root: Path, result: TestRunResult) -> None:
+        report_dir = tst_state_path(project_root).parent / "reports"
+        report_dir.mkdir(parents=True, exist_ok=True)
+        from air_runtime.utils import session_stamp
+        report_path = report_dir / f"{task_id}-{session_stamp()}.json"
+        atomic_json_write(report_path, result.__dict__)
--- a/lib/air_runtime/todo_parser.py
+++ b/lib/air_runtime/todo_parser.py
@@ -0,0 +1,96 @@
+"""
+TODO 解析器 — V2 修复 P1-8：列索引从表头推导，不再硬编码 cells[1]/cells[6]/cells[7]。
+"""
+
+from __future__ import annotations
+
+import re
+from dataclasses import dataclass
+from pathlib import Path
+
+
+@dataclass
+class TodoTask:
+    task_id: str
+    task: str
+    files_dirs: str = ""
+    status: str = "TODO"
+    done_when: str = ""
+    validations: str = ""
+    adr: str = ""
+
+
+def parse_tasks(todo_path: Path) -> list[TodoTask]:
+    """从 todo.md 解析任务列表，动态检测列索引。"""
+    if not todo_path.exists():
+        return []
+
+    content = todo_path.read_text(encoding="utf-8")
+    lines = [l.strip() for l in content.splitlines() if l.strip()]
+
+    # 找到 Markdown 表格头
+    header_idx = -1
+    for i, line in enumerate(lines):
+        if line.startswith("|") and "Task" in line and "Status" in line:
+            header_idx = i
+            break
+
+    if header_idx < 0:
+        return []
+
+    # 解析列名
+    header_line = lines[header_idx]
+    header_cols = [c.strip() for c in header_line.split("|") if c.strip()]
+
+    # 建立列名 → 索引映射
+    col_map = {}
+    for idx, col_name in enumerate(header_cols):
+        col_name_lower = col_name.lower()
+        if "task" in col_name_lower:
+            col_map["task"] = idx
+        elif "status" in col_name_lower:
+            col_map["status"] = idx
+        elif "files" in col_name_lower or "dir" in col_name_lower:
+            col_map["files_dirs"] = idx
+        elif "done" in col_name_lower or "when" in col_name_lower:
+            col_map["done_when"] = idx
+        elif "valid" in col_name_lower:
+            col_map["validations"] = idx
+        elif "adr" in col_name_lower:
+            col_map["adr"] = idx
+
+    # 跳过表头和分隔符
+    tasks = []
+    for line in lines[header_idx + 2:]:
+        if not line.startswith("|"):
+            continue
+        cells = [c.strip() for c in line.split("|") if len(c.strip()) > 0]
+        if not cells:
+            continue
+
+        task_cell = cells[col_map.get("task", 0)] if col_map.get("task", 0) < len(cells) else ""
+        # 提取 [T-xxx] 格式的任务 ID
+        tid_match = re.match(r"\[([A-Za-z0-9_\-\.]+)\]", task_cell)
+        task_id = tid_match.group(1) if tid_match else task_cell
+        task = task_cell
+        status = cells[col_map.get("status", 1)] if col_map.get("status", 1) < len(cells) else "TODO"
+        files_dirs = cells[col_map.get("files_dirs", 2)] if col_map.get("files_dirs", 2) < len(cells) else ""
+        done_when = cells[col_map.get("done_when", 3)] if col_map.get("done_when", 3) < len(cells) else ""
+        validations = cells[col_map.get("validations", 4)] if col_map.get("validations", 4) < len(cells) else ""
+        adr = cells[col_map.get("adr", 5)] if col_map.get("adr", 5) < len(cells) else ""
+
+        # 清理标记
+        task_id = re.sub(r"^\[|\]$", "", task_id).strip()
+
+        if task_id and task_id != "---":
+            tasks.append(TodoTask(
+                task_id=task_id,
+                task=task,
+                files_dirs=files_dirs,
+                status=status.upper() if status else "TODO",
+                done_when=done_when,
+                validations=validations,
+                adr=adr,
+            ))
+
+    return tasks
--- a/lib/air_runtime/utils.py
+++ b/lib/air_runtime/utils.py
@@ -0,0 +1,65 @@
+"""
+公共工具函数 — 消除 V1 中 _ordered_unique、_session_stamp、policy normalization 等
+在各模块中 3~5 份重复定义的代码。
+"""
+
+from __future__ import annotations
+
+import re
+from datetime import datetime, timezone
+from typing import Any
+
+
+def ordered_unique(items: list) -> list:
+    """保序去重。支持字符串列表和带 id 字段的字典列表。"""
+    seen: set[str] = set()
+    result = []
+    for item in items:
+        key = item if isinstance(item, str) else item.get("id", str(item))
+        if key not in seen:
+            seen.add(key)
+            result.append(item)
+    return result
+
+
+def session_stamp() -> str:
+    """统一的文件系统安全时间戳，所有模块共用。"""
+    return datetime.now(timezone.utc).isoformat().replace(":", "-").replace(".", "-").replace("+", "-")
+
+
+def now_iso() -> str:
+    """ISO 格式 UTC 时间戳，用于 JSON state 文件。"""
+    return datetime.now(timezone.utc).isoformat()
+
+
+def normalize_policy(defaults: dict[str, Any], overrides: dict[str, Any] | None) -> dict[str, Any]:
+    """通用的策略合并：overrides 覆盖 defaults，类型自动转换。"""
+    merged = {**defaults}
+    if overrides:
+        for k, v in overrides.items():
+            if k in merged:
+                expected_type = type(defaults[k])
+                try:
+                    merged[k] = expected_type(v) if not isinstance(v, expected_type) else v
+                except (ValueError, TypeError):
+                    merged[k] = v
+    return merged
+
+
+def sanitize_task_id(task_id: str) -> str:
+    """防止路径注入：仅允许字母数字、下划线、连字符、点号。"""
+    if not re.fullmatch(r"[A-Za-z0-9_\-\.]+", task_id):
+        raise ValueError(f"invalid task_id: {task_id!r}")
+    return task_id
+
+
+def sanitize_marker(marker: str) -> str:
+    """防止 HTML 注释注入。"""
+    if "-->" in marker or "<!--" in marker:
+        raise ValueError(f"marker contains comment delimiters: {marker!r}")
+    return marker
+
+
+def truncate_history(data: list, max_items: int = 100) -> list:
+    """截断历史列表，防止无界增长（P2-2）。"""
+    return data[-max_items:] if len(data) > max_items else data
--- a/lib/air_runtime/worktree.py
+++ b/lib/air_runtime/worktree.py
@@ -0,0 +1,157 @@
+"""
+Worktree 隔离并行模块 — V2 P1-15 修复。
+区域级冲突检测 + git worktree 隔离，允许同文件不同区域任务安全并行。
+"""
+
+from __future__ import annotations
+
+import enum
+import subprocess
+from dataclasses import dataclass
+from pathlib import Path
+
+
+class ConflictLevel(enum.Enum):
+    NONE = "none"       # 无文件重叠，直接并行
+    SOFT = "soft"       # 同文件不同区域，worktree 隔离并行
+    HARD = "hard"       # 同文件同区域，必须串行
+
+
+@dataclass
+class Conflict:
+    task_a: str
+    task_b: str
+    level: ConflictLevel  # NONE / SOFT / HARD
+    reason: str = ""
+
+    def to_dict(self) -> dict:
+        return {
+            "taskA": self.task_a,
+            "taskB": self.task_b,
+            "level": self.level.value,
+            "reason": self.reason,
+        }
+
+
+@dataclass
+class MergeResult:
+    task_id: str
+    successful: bool
+    conflicts: list[str]
+
+
+class RegionConflictDetector:
+    """区域级写集冲突检测。
+
+    区域定义（按优先级）：
+    1. todo.md 中的区域标注 [region: xxx]
+    2. 函数/类边界（AST 分析或标记注释）
+    3. 行号区间（从 diff 或任务元数据获取）
+    """
+
+    def detect(self, task_a_write_set: list[str], task_b_write_set: list[str],
+               regions_a: dict[str, list[tuple[int, int]]] | None = None,
+               regions_b: dict[str, list[tuple[int, int]]] | None = None) -> ConflictLevel:
+        file_overlap = set(task_a_write_set) & set(task_b_write_set)
+        if not file_overlap:
+            return ConflictLevel.NONE
+
+        if regions_a is None or regions_b is None:
+            return ConflictLevel.SOFT
+
+        for fpath in file_overlap:
+            ra = regions_a.get(fpath, [(0, 999999)])
+            rb = regions_b.get(fpath, [(0, 999999)])
+            for a_start, a_end in ra:
+                for b_start, b_end in rb:
+                    if a_start < b_end and b_start < a_end:
+                        return ConflictLevel.HARD
+        return ConflictLevel.SOFT
+
+    def detect_batch(self, task_write_sets: dict[str, list[str]],
+                     task_regions: dict[str, dict[str, list[tuple[int, int]]]] | None = None,
+                     ) -> list[Conflict]:
+        """批量检测：接收 {task_id: [file, ...]} 映射，返回所有冲突对。
+
+        逐对调用 detect()，仅返回 level != NONE 的冲突。
+        """
+        conflicts: list[Conflict] = []
+        task_ids = list(task_write_sets.keys())
+        for i in range(len(task_ids)):
+            for j in range(i + 1, len(task_ids)):
+                a, b = task_ids[i], task_ids[j]
+                ra = task_regions.get(a) if task_regions else None
+                rb = task_regions.get(b) if task_regions else None
+                level = self.detect(task_write_sets[a], task_write_sets[b], ra, rb)
+                if level != ConflictLevel.NONE:
+                    overlap = sorted(set(task_write_sets[a]) & set(task_write_sets[b]))
+                    reason = f"file overlap: {', '.join(overlap)}" if overlap else ""
+                    if level == ConflictLevel.HARD:
+                        reason = f"region overlap: {', '.join(overlap)}" if overlap else "same region"
+                    conflicts.append(Conflict(task_a=a, task_b=b, level=level, reason=reason))
+        return conflicts
+
+    def extract_regions_from_task(self, file_path: Path, region_markers: list[tuple[int, int]]) -> dict[str, list[tuple[int, int]]]:
+        """从任务的区域标注提取行号区间。"""
+        return {str(file_path): region_markers}
+
+
+class WorktreeIsolation:
+    """为 SOFT 冲突任务创建 git worktree 隔离，完成后合并回主分支。"""
+
+    def __init__(self, repo_root: Path):
+        self._repo_root = repo_root
+
+    def create_worktree(self, task_id: str, base_ref: str = "HEAD") -> Path:
+        branch = f"air-{task_id}"
+        wt_path = self._repo_root.parent / f"{self._repo_root.name}-air-{task_id}"
+
+        subprocess.run(
+            ["git", "-C", str(self._repo_root), "worktree", "add", "-b", branch, str(wt_path), base_ref],
+            check=True, capture_output=True, text=True,
+        )
+        return wt_path
+
+    def merge_back(self, task_id: str, wt_path: Path) -> MergeResult:
+        branch = f"air-{task_id}"
+        try:
+            subprocess.run(
+                ["git", "-C", str(self._repo_root), "merge", "--no-ff", branch],
+                check=True, capture_output=True, text=True,
+            )
+            return MergeResult(task_id=task_id, successful=True, conflicts=[])
+        except subprocess.CalledProcessError as exc:
+            conflicts = self._parse_conflicts(exc.stderr)
+            self._abort_merge()
+            return MergeResult(task_id=task_id, successful=False, conflicts=conflicts)
+
+    def cleanup(self, task_id: str, wt_path: Path) -> None:
+        branch = f"air-{task_id}"
+        try:
+            subprocess.run(
+                ["git", "-C", str(self._repo_root), "worktree", "remove", str(wt_path), "--force"],
+                check=True, capture_output=True, text=True,
+            )
+            subprocess.run(
+                ["git", "-C", str(self._repo_root), "branch", "-D", branch],
+                check=True, capture_output=True, text=True,
+            )
+        except subprocess.CalledProcessError:
+            pass
+
+    def _abort_merge(self) -> None:
+        try:
+            subprocess.run(
+                ["git", "-C", str(self._repo_root), "merge", "--abort"],
+                check=True, capture_output=True, text=True,
+            )
+        except subprocess.CalledProcessError:
+            pass
+
+    @staticmethod
+    def _parse_conflicts(stderr: str) -> list[str]:
+        conflicts: list[str] = []
+        for line in stderr.splitlines():
+            if "CONFLICT" in line:
+                conflicts.append(line.strip())
+        return conflicts
--- a/lib/air_runtime/xdb_capture.py
+++ b/lib/air_runtime/xdb_capture.py
@@ -0,0 +1,161 @@
+"""AirXDB screenshot capture backends with automatic fallback chain.
+
+Three capture methods:
+  - KmsGrabCapture: DRM/KMS native screenshot via ffmpeg (no sudo)
+  - XvfbCapture: Xvfb virtual framebuffer screenshot
+  - FallbackCapture: text placeholder when no GUI capture is available
+
+CaptureManager tries them in order (kms -> xvfb -> fallback) unless
+a specific method is requested.
+"""
+
+from __future__ import annotations
+
+import shutil
+import subprocess
+from dataclasses import dataclass
+from pathlib import Path
+
+
+@dataclass
+class CaptureResult:
+    success: bool
+    output_path: Path | None
+    method: str  # kms, xvfb, fallback
+    error: str | None = None
+
+
+class KmsGrabCapture:
+    """KMS/DRM screenshot via ffmpeg -f kmsgrab.
+
+    No sudo is ever used.  The caller must have read access to
+    /dev/dri/cardX for this to work.
+    """
+
+    def capture(self, output_path: Path) -> CaptureResult:
+        # 1. Detect whether the current user can access a DRI device
+        dri_dev = Path("/dev/dri/card0")
+        if dri_dev.exists():
+            try:
+                # Test read permission (no sudo)
+                open(dri_dev).close()
+            except PermissionError:
+                return CaptureResult(False, None, "kms", "no /dev/dri permission")
+        else:
+            return CaptureResult(False, None, "kms", "no GPU")
+
+        # 2. Use ffmpeg directly (no sudo)
+        try:
+            result = subprocess.run(
+                [
+                    "ffmpeg", "-y", "-f", "kmsgrab", "-i", "-",
+                    "-frames:v", "1",
+                    "-vf", "hwdownload,format=bgr0",
+                    "-f", "image2", str(output_path),
+                ],
+                capture_output=True,
+                text=True,
+                timeout=30,
+            )
+            if result.returncode == 0 and output_path.exists():
+                return CaptureResult(True, output_path, "kms", None)
+            return CaptureResult(False, None, "kms", result.stderr)
+        except FileNotFoundError:
+            return CaptureResult(False, None, "kms", "ffmpeg not found")
+        except subprocess.TimeoutExpired:
+            return CaptureResult(False, None, "kms", "ffmpeg timed out")
+        except Exception as e:
+            return CaptureResult(False, None, "kms", str(e))
+
+
+class XvfbCapture:
+    """Xvfb virtual display screenshot via xvfb-run + scrot."""
+
+    def capture(
+        self,
+        output_path: Path,
+        width: int = 1920,
+        height: int = 1080,
+    ) -> CaptureResult:
+        # Check xvfb-run availability
+        if not shutil.which("xvfb-run"):
+            return CaptureResult(False, None, "xvfb", "xvfb-run not found")
+
+        # Use scrot as the screenshot tool inside Xvfb
+        try:
+            result = subprocess.run(
+                [
+                    "xvfb-run",
+                    "--auto-servernum",
+                    "--server-args",
+                    f"-screen 0 {width}x{height}x24",
+                    "--",
+                    "scrot",
+                    str(output_path),
+                ],
+                capture_output=True,
+                text=True,
+                timeout=60,
+            )
+            if result.returncode == 0 and output_path.exists():
+                return CaptureResult(True, output_path, "xvfb", None)
+            return CaptureResult(False, None, "xvfb", result.stderr)
+        except FileNotFoundError:
+            return CaptureResult(False, None, "xvfb", "xvfb-run not found")
+        except subprocess.TimeoutExpired:
+            return CaptureResult(False, None, "xvfb", "xvfb-run timed out")
+        except Exception as e:
+            return CaptureResult(False, None, "xvfb", str(e))
+
+
+class FallbackCapture:
+    """Fallback: write a text placeholder when no GUI capture is available."""
+
+    def capture(
+        self,
+        output_path: Path,
+        description: str = "",
+    ) -> CaptureResult:
+        output_path.write_text(
+            f"[GUI capture not available]\n{description}\n"
+        )
+        return CaptureResult(True, output_path, "fallback", None)
+
+
+class CaptureManager:
+    """Auto-select the best capture method with fallback chain.
+
+    Order: kms -> xvfb -> fallback.
+    Use ``prefer`` to force a specific method or "auto" for the chain.
+    """
+
+    def __init__(self) -> None:
+        self.kms = KmsGrabCapture()
+        self.xvfb = XvfbCapture()
+        self.fallback = FallbackCapture()
+
+    def capture(
+        self,
+        output_path: Path,
+        prefer: str = "auto",
+    ) -> CaptureResult:
+        """
+        prefer: "kms" | "xvfb" | "fallback" | "auto"
+        auto mode tries kms -> xvfb -> fallback in order.
+        """
+        if prefer in ("kms", "auto"):
+            result = self.kms.capture(output_path)
+            if result.success:
+                return result
+            if prefer == "kms":
+                return result  # forced kms failed, return failure
+
+        if prefer in ("xvfb", "auto"):
+            result = self.xvfb.capture(output_path)
+            if result.success:
+                return result
+            if prefer == "xvfb":
+                return result  # forced xvfb failed, return failure
+
+        # fallback
+        return self.fallback.capture(output_path, "all capture methods failed")
--- a/scripts/airplan.py
+++ b/scripts/airplan.py
@@ -0,0 +1,191 @@
+#!/usr/bin/env python3
+"""
+AirPlan V2 — 统一 CLI 入口。
+V1 的 8 个独立插件合并为一个命令，通过 --mode 参数路由到对应子模式。
+
+用法:
+  airplan --mode arc --project <project-root> [--todo <todo-md>]
+  airplan --mode eng --project <project-root> [--sub enter|status|plan|dispatch|monitor|merge|run|intervene]
+  airplan --mode do  --project <project-root> --task-id <id>
+  airplan --mode dbg --project <project-root> [--task-id <id>]
+  airplan --mode sdb --project <project-root> [--backend cppcheck|clang-tidy|clippy|go-vet|tsc] [--target <path>]
+  airplan --mode ctx --project <project-root> [--sub now|status|pause|init]
+  airplan --mode dep --project <project-root> --task-id <id> --host <host> [--binary <path>]
+  airplan --mode tst --project <project-root> --task-id <id> [--framework <name>]
+  airplan --mode sec --project <project-root> --task-id <id> [--scan-path <path>]
+  airplan --mode rvr --project <project-root> --task-id <id>
+"""
+
+from __future__ import annotations
+
+import argparse
+import sys
+from pathlib import Path
+
+
+def _add_lib_path() -> None:
+    root = Path(__file__).resolve().parents[1]
+    lib_path = root / "lib"
+    if str(lib_path) not in sys.path:
+        sys.path.insert(0, str(lib_path))
+
+
+_add_lib_path()
+
+from air_runtime.project_bootstrap import ensure_project_bootstrap
+
+
+def _run_arc(args) -> None:
+    from air_runtime.modes.arc_mode import main as arc_main
+    arc_main(args)
+
+
+def _run_eng(args) -> None:
+    from air_runtime.modes.eng_mode import main as eng_main
+    eng_main(args)
+
+
+def _run_do(args) -> None:
+    from air_runtime.modes.do_mode import main as do_main
+    do_main(args)
+
+
+def _run_dbg(args) -> None:
+    from air_runtime.modes.dbg_mode import main as dbg_main
+    dbg_main(args)
+
+
+def _run_xdb(args) -> None:
+    from air_runtime.modes.xdb_mode import main as xdb_main
+    xdb_main(args)
+
+
+def _run_sdb(args) -> None:
+    from air_runtime.modes.sdb_mode import main as sdb_main
+    sdb_main(args)
+
+
+def _run_ndb(args) -> None:
+    from air_runtime.modes.ndb_mode import main as ndb_main
+    ndb_main(args)
+
+
+def _run_ctx(args) -> None:
+    from air_runtime.modes.ctx_mode import main as ctx_main
+    ctx_main(args)
+
+
+def _run_dep(args) -> None:
+    from air_runtime.modes.dep_mode import main as dep_main
+    dep_main(args)
+
+
+def _run_tst(args) -> None:
+    from air_runtime.modes.tst_mode import main as tst_main
+    tst_main(args)
+
+
+def _run_sec(args) -> None:
+    from air_runtime.modes.sec_mode import main as sec_main
+    sec_main(args)
+
+
+def _run_rvr(args) -> None:
+    from air_runtime.modes.rvr_mode import main as rvr_main
+    rvr_main(args)
+
+
+def _run_eng_orchestrator(args) -> None:
+    from air_runtime.modes.eng_orchestrator import main as eng_orch_main
+    eng_orch_main(args)
+
+
+MODE_ROUTER = {
+    "arc": _run_arc,
+    "eng": _run_eng,
+    "do": _run_do,
+    "dbg": _run_dbg,
+    "xdb": _run_xdb,
+    "sdb": _run_sdb,
+    "ndb": _run_ndb,
+    "ctx": _run_ctx,
+    "dep": _run_dep,
+    "tst": _run_tst,
+    "sec": _run_sec,
+    "rvr": _run_rvr,
+    "eng_orchestrator": _run_eng_orchestrator,
+}
+
+MODE_HELP = {
+    "arc": "架构规划器 — 分析依赖、写集冲突、产出执行计划",
+    "eng": "调度引擎 — 波次派发、监控、合并、修复编排",
+    "do": "任务执行器 — 单任务切片执行与结果生成",
+    "dbg": "调试器 — 7步调试工作流、根因分析、修复回滚",
+    "xdb": "GUI验证器 — 截图取证、GUI操作验证",
+    "sdb": "静态分析器 — 多语言静态分析、diff模式",
+    "ndb": "网络调试器 — 抓包分析、TLS解密",
+    "ctx": "上下文管理器 — 压缩、Token估算、锁检测",
+    "dep": "部署器 — SSH远程构建、部署、systemd管理",
+    "tst": "测试运行器 — 统一测试执行与结果报告",
+    "sec": "安全扫描器 — 敏感数据检测与脱敏",
+    "rvr": "需求审查器 — 交付物与需求一致性验证",
+}
+
+
+def parse_args(argv: list[str] | None = None) -> argparse.Namespace:
+    parser = argparse.ArgumentParser(
+        description="AirPlan V2 — 统一制品驱动开发调度器",
+        formatter_class=argparse.RawDescriptionHelpFormatter,
+        epilog="\n".join(f"  {k:6s}  {v}" for k, v in MODE_HELP.items()),
+    )
+    parser.add_argument("--mode", choices=list(MODE_ROUTER.keys()), required=True,
+                       help="子模式选择")
+    parser.add_argument("--project", default=".",
+                       help="项目根目录 (默认: .)")
+    parser.add_argument("--sub", default="",
+                       help="子模式内的二级操作（如 eng 的 enter|status|plan 等）")
+    parser.add_argument("--todo", default="",
+                       help="todo.md 路径（arc/eng 使用）")
+    parser.add_argument("--task-id", default="",
+                       help="任务 ID（do/dbg/xdb/sdb/ndb/dep/tst/sec/rvr 使用）")
+    parser.add_argument("--result", default="",
+                       help="Worker result.json 路径（eng merge 使用）")
+    parser.add_argument("--dispatch-group", default="",
+                       help="派发组名（eng dispatch 使用）")
+    parser.add_argument("--host", default="",
+                       help="远程主机（dep 使用）")
+    parser.add_argument("--binary", default="",
+                       help="本地二进制路径（dep 使用）")
+    parser.add_argument("--framework", default="",
+                       help="测试框架（tst 使用）")
+    parser.add_argument("--scan-path", default="",
+                       help="扫描路径（sec 使用）")
+    parser.add_argument("--sec-mode", default="blocking", choices=["advisory", "blocking"],
+                       help="安全扫描模式（sec 使用）: advisory（只报告）/ blocking（阻止合并）")
+    parser.add_argument("--backend", default="",
+                       help="静态分析后端（sdb 使用）: cppcheck|clang-tidy|clippy|go-vet|tsc")
+    parser.add_argument("--target", default="",
+                       help="分析目标路径（sdb 使用）")
+    parser.add_argument("--prefer", default="auto",
+                       help="截图后端偏好（xdb 使用）: kms|xvfb|fallback|auto")
+    parser.add_argument("--output", default="",
+                       help="截图输出文件名（xdb 使用，默认 screenshot.png）")
+    return parser.parse_args(argv)
+
+
+def main(argv: list[str] | None = None) -> None:
+    args = parse_args(argv)
+    project_root = Path(args.project).expanduser().resolve()
+    ensure_project_bootstrap(project_root)
+
+    handler = MODE_ROUTER.get(args.mode)
+    if handler is None:
+        print(f"Unknown mode: {args.mode}", file=sys.stderr)
+        print(f"Available modes: {', '.join(MODE_ROUTER.keys())}", file=sys.stderr)
+        sys.exit(1)
+
+    handler(args)
+
+
+if __name__ == "__main__":
+    main()
--- a/skills/airplan/SKILL.md
+++ b/skills/airplan/SKILL.md
@@ -0,0 +1,159 @@
+---
+name: airplan
+description: AirPlan V2 — 统一制品驱动开发调度器。单一插件整合规划(arc)、执行(do)、调试(dbg)、GUI验证(xdb)、静态分析(sdb)、网络调试(ndb)、部署(dep)、测试(tst)、安全扫描(sec)、需求审查(rvr)。L1代码级保障：不依赖LLM自觉。
+allowed-tools: [Read, Glob, Grep, Bash, Write, Edit]
+---
+
+# AirPlan V2
+
+## 概述
+
+AirPlan V2 是 V1 的 8 个独立插件（airarc, aireng, airdo, airdbg, airxdb, airsdb, airndb, aircontext）合并为 **1 个统一插件**，新增 4 个组件（airdep, airtst, airsec, airrvr），共 **12 个子模式**。
+
+核心设计原则：
+1. **制品驱动通信** — 插件间通过 AirPlan/ 文件通信
+2. **上下文隔离** — Worker `fork_context=false`
+3. **L1 代码级保障** — 关键逻辑不依赖 LLM 自觉，由引擎代码强制执行
+
+## 子模式速查
+
+| 子模式 | 角色 | 关键文件 |
+|--------|------|----------|
+| arc | 架构规划器 — 产出执行计划、DAG | execution-plan.json, task-graph.json |
+| eng | 调度引擎 — 波次派发、监控、合并 | state/aireng/state.json |
+| do | 任务执行器 — 单任务切片<E58887><E78987><EFBFBD>行 | state/airdo/tasks/{task_id}/result.json |
+| dbg | 调试器 — 7 步工作流强制追踪 | state/airdbg/sessions/*.json |
+| xdb | GUI 验证器 — 截图取证、DRM/KMS | state/airxdb/artifacts/ |
+| sdb | 静态分析器 — 多语言分析 | state/airsdb/reports/ |
+| ndb | 网络调试器 — 抓包分析 | state/airndb/captures/ |
+| ctx | 上下文管理器 — 压缩、Token 估算 | state/aircontext/ |
+| dep | 部署器 — SSH 远程构建 + systemd | state/airdep/sessions/ |
+| tst | 测试运行器 — 统一多框架测试 | state/airtst/reports/ |
+| sec | 安全扫描器 — 敏感数据检测 | state/airsec/ |
+| rvr | 需求审查器 — 交付物与需求一致性 | state/airrvr/reviews/ |
+
+## 关键不变量（INV-）
+
+### INV-1 制品驱动通信
+插件间**不直接调用**，而是通过 AirPlan/ 目录下的 JSON/MD 文件通信。状态文件是唯一真相来源。
+
+### INV-2 上下文隔离
+所有 Worker 派发使用 `fork_context=false`，确保子任务不继承父线程的完整上下文，仅传递任务相关的必要信息。
+
+### INV-3 架构同步强制
+不更新架构文档（ADR、C4、plan.md）不能标记任务为 DONE。AirEng merge 时强制检查 documentUpdates。
+
+### INV-4 证据先于修复
+GUI 任务必须先有截图/验证证据，网络任务必须先有抓包/连通性证据，代码任务必须先有静态分析/测试结果。EvidenceGatePolicy 根据任务类型自动判断需要哪种证据。
+
+### INV-5 闭环自动修复
+执行 → 失败 → 调试 → 修复 → 重执行 的闭环由 AirEng 的修复预算机制自动驱动。
+
+### INV-6 调度器不写代码
+AirEng 是调度器，不是执行器。它只做 plan/dispatch/monitor/merge/intervene/doc-sync，绝不直接实现任务代码。所有任务实现必须通过 /do 子代理（fork_context=false）完成。调度线程直接写代码是严重违规，唯一例外是 Worker 硬阻塞无法自恢复时的紧急干预，干预后必须立即回到调度模式。
+
+### INV-7 架构器只读不写
+AirArc 是纯规划器，只做分析依赖、写集冲突、产出调度计划。禁止编写代码、修改源代码文件、执行构建命令。allowed-tools 仅含 Read/Glob/Grep，不含 Write/Edit/Bash。
+
+### INV-8 架构三阶段门控
+AirArc 必须遵循三阶段流程：discussing（需求探讨）→ proposing（架构提议）→ confirmed（用户确认后）。execution-plan.json 仅在 phase=confirmed 时允许写入。
+
+### INV-9 调试先读后写
+AirDbg 修改代码前必须至少完成一项取证行为（截图/抓包/静态分析/日志分析/代码追踪/复现步骤）。未取证就修改代码是严重违规。
+
+### INV-10 中文锁定
+AirEng 必须始终使用中文与用户交流。自主决策不停下来问用户。
+
+### INV-11 项目日志标准
+所有 C++ 项目必须集成 spdlog，禁止 std::cout/qDebug/printf。AirArc 规划时强制首任务为 spdlog 集成（如未存在），AirRvr 审查时检查日志完备性。
+
+## L1 代码级保障（不依赖 LLM 自觉）
+
+以下功能由引擎代码强制执行，SKILL.md 指令仅作辅助：
+
+1. **原子写入** — `air_runtime/io.py:atomic_json_write()` 使用 tempfile + os.replace()
+2. **文件锁** — `air_runtime/lock.py:FileLock` 持有者才能读写 state.json
+3. **证据门控** — `air_runtime/evidence_gate.py:EvidenceGatePolicy` 分类任务类型，强制要求对应证据
+4. **强制 AirDbg 路由** — `air_runtime/modes/do_mode.py:finish_worker()` 中 blocked/failed/done无证据 时强制路由到 airdbg
+5. **硬编码轮询循环** — `air_runtime/modes/eng_mode.py:monitor_engine()` 每 5 分钟检测所有 Worker 状态
+6. **Worker 超时** — 超过 2 小时硬上限自动 terminate
+7. **部署一致性验证** — merge 时检测 deployRequired 字段是否有对应验证证据
+8. **任务 ID 校验** — `air_runtime/utils.py:sanitize_task_id()` 防止路径注入
+9. **Arc 三阶段门控** — `air_runtime/modes/arc_mode.py:ArcPhaseGate` 仅 confirmed 阶段可写 execution-plan.json
+10. **Dbg 先读后写门控** — `air_runtime/modes/dbg_mode.py:WorkflowViolation` fix 步骤前必须有 collectedEvidence
+11. **Arc 工具白名单** — commands/arc.md allowed-tools 仅 [Read, Glob, Grep]，deny-plan-mode=true
+
+## 使用示例
+
+```bash
+# 架构规划
+airplan --mode arc --project /path/to/project --sub enter
+airplan --mode arc --project /path/to/project --sub parallel-review --todo AirPlan/todo.md
+
+# 调度引擎
+airplan --mode eng --project /path/to/project --sub enter
+airplan --mode eng --project /path/to/project --sub plan --todo AirPlan/todo.md
+airplan --mode eng --project /path/to/project --sub dispatch
+airplan --mode eng --project /path/to/project --sub monitor
+airplan --mode eng --project /path/to/project --sub merge --result /path/to/result.json
+
+# 任务执行
+airplan --mode do --project /path/to/project --task-id T-001 --sub enter
+airplan --mode do --project /path/to/project --task-id T-001 --sub finish --result /path/to/result.json
+
+# 调试
+airplan --mode dbg --project /path/to/project --task-id T-001 --sub start
+airplan --mode dbg --project /path/to/project --task-id T-001 --sub snapshot
+
+# 部署
+airplan --mode dep --project /path/to/project --task-id T-001 --host 192.168.1.100 --binary ./build/myapp
+
+# 测试
+airplan --mode tst --project /path/to/project --task-id T-001 --framework pytest
+
+# 安全扫描
+airplan --mode sec --project /path/to/project --task-id T-001 --scan-path ./src
+
+# 需求审查
+airplan --mode rvr --project /path/to/project --task-id T-001 --sub review
+```
+
+## 文件结构
+
+```
+AirPlan/
+├── AGENTS.md                    # Agent 行为规范（各子模式入口）
+├── plan.md                      # 执行计划摘要
+├── todo.md                      # 任务列表（人类可读）
+├── docs/
+│   ├── architecture/
+│   │   ├── adr/                 # ADR 记录
+│   │   └── c4/module.md         # C4 模块文档
+│   ├── debug/
+│   │   ├── debug-log.md         # 调试日志
+│   │   └── gui-debug-log.md     # GUI 调试日志
+│   ├── staticanalysis.md        # 静态分析报告
+│   └── network/                 # 网络抓包
+└── state/
+    ├── airarc/reviews/          # 架构规划输出
+    ├── aireng/                  # 调度引擎状态
+    ├── airdo/tasks/             # Worker 结果
+    ├── airdbg/sessions/         # 调试会话
+    ├── airxdb/artifacts/        # 截图/验证证据
+    ├── airsdb/reports/          # 静态分析报告
+    ├── airndb/captures/         # 网络抓包
+    ├── aircontext/              # 上下文压缩状态
+    ├── airdep/sessions/         # 部署记录
+    ├── airtst/reports/          # 测试报告
+    ├── airsec/                  # 安全扫描<E689AB><E68F8F>果
+    └── airrvr/reviews/          # 需求审查报告
+```
+
+## 故障排查
+
+| 症状 | 排查步骤 |
+|------|----------|
+| Worker 假阳性阻塞 | 检查 `state/airxdb/` 是否存在对应截图证据，确认任务类型是否被 EvidenceGatePolicy 误分类 |
+| 部署后未生效 | 检查 merge 时是否包含 deploy 验证证据（remote-deploy-verify / remote-binary-md5） |
+| 任务状态卡在 DISPATCHED | 运行 `airplan --mode eng --project . --sub monitor` 查看 Worker 是否超时或停滞 |
+| 敏感数据泄露 | 检查 AirSec 扫描结果，查看 `state/airsec/` 是否有 finding |
--- a/skills/airplan/airplan/SKILL.md
+++ b/skills/airplan/airplan/SKILL.md
@@ -0,0 +1,138 @@
+---
+name: airplan
+description: AirPlan V2 — 统一制品驱动开发调度器。单一插件整合规划(arc)、执行(do)、调试(dbg)、GUI验证(xdb)、静态分析(sdb)、网络调试(ndb)、部署(dep)、测试(tst)、安全扫描(sec)、需求审查(rvr)。L1代码级保障：不依赖LLM自觉。
+allowed-tools: [Read, Glob, Grep, Bash, Write, Edit]
+---
+
+# AirPlan V2
+
+## 概述
+
+AirPlan V2 是 V1 的 8 个独立插件（airarc, aireng, airdo, airdbg, airxdb, airsdb, airndb, aircontext）合并为 **1 个统一插件**，新增 4 个组件（airdep, airtst, airsec, airrvr），共 **12 个子模式**。
+
+核心设计原则：
+1. **制品驱动通信** — 插件间通过 AirPlan/ 文件通信
+2. **上下文隔离** — Worker `fork_context=false`
+3. **L1 代码级保障** — 关键逻辑不依赖 LLM 自觉，由引擎代码强制执行
+
+## 子模式速查
+
+| 子模式 | 角色 | 关键文件 |
+|--------|------|----------|
+| arc | 架构规划器 — 产出执行计划、DAG | execution-plan.json, task-graph.json |
+| eng | 调度引擎 — 波次派发、监控、合并 | state/aireng/state.json |
+| do | 任务执行器 — 单任务切片<E58887><E78987><EFBFBD>行 | state/airdo/tasks/{task_id}/result.json |
+| dbg | 调试器 — 7 步工作流强制追踪 | state/airdbg/sessions/*.json |
+| xdb | GUI 验证器 — 截图取证、DRM/KMS | state/airxdb/artifacts/ |
+| sdb | 静态分析器 — 多语言分析 | state/airsdb/reports/ |
+| ndb | 网络调试器 — 抓包分析 | state/airndb/captures/ |
+| ctx | 上下文管理器 — 压缩、Token 估算 | state/aircontext/ |
+| dep | 部署器 — SSH 远程构建 + systemd | state/airdep/sessions/ |
+| tst | 测试运行器 — 统一多框架测试 | state/airtst/reports/ |
+| sec | 安全扫描器 — 敏感数据检测 | state/airsec/ |
+| rvr | 需求审查器 — 交付物与需求一致性 | state/airrvr/reviews/ |
+
+## 关键不变量（INV-）
+
+### INV-1 制品驱动通信
+插件间**不直接调用**，而是通过 AirPlan/ 目录下的 JSON/MD 文件通信。状态文件是唯一真相来源。
+
+### INV-2 上下文隔离
+所有 Worker 派发使用 `fork_context=false`，确保子任务不继承父线程的完整上下文，仅传递任务相关的必要信息。
+
+### INV-3 架构同步强制
+不更新架构文档（ADR、C4、plan.md）不能标记任务为 DONE。AirEng merge 时强制检查 documentUpdates。
+
+### INV-4 证据先于修复
+GUI 任务必须先有截图/验证证据，网络任务必须先有抓包/连通性证据，代码任务必须先有静态分析/测试结果。EvidenceGatePolicy 根据任务类型自动判断需要哪种证据。
+
+### INV-5 闭环自动修复
+执行 → 失败 → 调试 → 修复 → 重执行 的闭环由 AirEng 的修复预算机制自动驱动。
+
+## L1 代码级保障（不依赖 LLM 自觉）
+
+以下功能由引擎代码强制执行，SKILL.md 指令仅作辅助：
+
+1. **原子写入** — `air_runtime/io.py:atomic_json_write()` 使用 tempfile + os.replace()
+2. **文件锁** — `air_runtime/lock.py:FileLock` 持有者才能读写 state.json
+3. **证据门控** — `air_runtime/evidence_gate.py:EvidenceGatePolicy` 分类任务类型，强制要求对应证据
+4. **强制 AirDbg 路由** — `air_runtime/modes/do_mode.py:finish_worker()` 中 blocked/failed/done无证据 时强制路由到 airdbg
+5. **硬编码轮询循环** — `air_runtime/modes/eng_mode.py:monitor_engine()` 每 5 分钟检测所有 Worker 状态
+6. **Worker 超时** — 超过 2 小时硬上限自动 terminate
+7. **部署一致性验证** — merge 时检测 deployRequired 字段是否有对应验证证据
+8. **任务 ID 校验** — `air_runtime/utils.py:sanitize_task_id()` 防止路径注入
+
+## 使用示例
+
+```bash
+# 架构规划
+airplan --mode arc --project /path/to/project --sub enter
+airplan --mode arc --project /path/to/project --sub parallel-review --todo AirPlan/todo.md
+
+# 调度引擎
+airplan --mode eng --project /path/to/project --sub enter
+airplan --mode eng --project /path/to/project --sub plan --todo AirPlan/todo.md
+airplan --mode eng --project /path/to/project --sub dispatch
+airplan --mode eng --project /path/to/project --sub monitor
+airplan --mode eng --project /path/to/project --sub merge --result /path/to/result.json
+
+# 任务执行
+airplan --mode do --project /path/to/project --task-id T-001 --sub enter
+airplan --mode do --project /path/to/project --task-id T-001 --sub finish --result /path/to/result.json
+
+# 调试
+airplan --mode dbg --project /path/to/project --task-id T-001 --sub start
+airplan --mode dbg --project /path/to/project --task-id T-001 --sub snapshot
+
+# 部署
+airplan --mode dep --project /path/to/project --task-id T-001 --host 192.168.1.100 --binary ./build/myapp
+
+# 测试
+airplan --mode tst --project /path/to/project --task-id T-001 --framework pytest
+
+# 安全扫描
+airplan --mode sec --project /path/to/project --task-id T-001 --scan-path ./src
+
+# 需求审查
+airplan --mode rvr --project /path/to/project --task-id T-001 --sub review
+```
+
+## 文件结构
+
+```
+AirPlan/
+├── AGENTS.md                    # Agent 行为规范（各子模式入口）
+├── plan.md                      # 执行计划摘要
+├── todo.md                      # 任务列表（人类可读）
+├── docs/
+│   ├── architecture/
+│   │   ├── adr/                 # ADR 记录
+│   │   └── c4/module.md         # C4 模块文档
+│   ├── debug/
+│   │   ├── debug-log.md         # 调试日志
+│   │   └── gui-debug-log.md     # GUI 调试日志
+│   ├── staticanalysis.md        # 静态分析报告
+│   └── network/                 # 网络抓包
+└── state/
+    ├── airarc/reviews/          # 架构规划输出
+    ├── aireng/                  # 调度引擎状态
+    ├── airdo/tasks/             # Worker 结果
+    ├── airdbg/sessions/         # 调试会话
+    ├── airxdb/artifacts/        # 截图/验证证据
+    ├── airsdb/reports/          # 静态分析报告
+    ├── airndb/captures/         # 网络抓包
+    ├── aircontext/              # 上下文压缩状态
+    ├── airdep/sessions/         # 部署记录
+    ├── airtst/reports/          # 测试报告
+    ├── airsec/                  # 安全扫描<E689AB><E68F8F>果
+    └── airrvr/reviews/          # 需求审查报告
+```
+
+## 故障排查
+
+| 症状 | 排查步骤 |
+|------|----------|
+| Worker 假阳性阻塞 | 检查 `state/airxdb/` 是否存在对应截图证据，确认任务类型是否被 EvidenceGatePolicy 误分类 |
+| 部署后未生效 | 检查 merge 时是否包含 deploy 验证证据（remote-deploy-verify / remote-binary-md5） |
+| 任务状态卡在 DISPATCHED | 运行 `airplan --mode eng --project . --sub monitor` 查看 Worker 是否超时或停滞 |
+| 敏感数据泄露 | 检查 AirSec 扫描结果，查看 `state/airsec/` 是否有 finding |
				`@@ -0,0 +1 @@`
				`from air_runtime.modes.xdb_sdb_ndb_modes import ndb_main as main`