diputs18 · diputs18 · Jul 26, 2026 · Jul 19, 2026 · Jul 26, 2026 · Jul 26, 2026
diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
@@ -0,0 +1,73 @@
+name: Tianjun CI
+
+on:
+  push:
+  pull_request:
+
+jobs:
+  verify:
+    runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        python-version: ["3.11", "3.12"]
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: pip
+      - uses: actions/setup-node@v4
+        with:
+          node-version: "22"
+      - uses: actions/setup-java@v4
+        with:
+          distribution: temurin
+          java-version: "21"
+      - name: Install Node dependencies
+        run: npm ci
+      - name: Install test dependencies
+        run: python -m pip install -e ".[dev,experiments,mcp]"
+      - name: Build and smoke-test installed wheel
+        run: |
+          python -m pip wheel . --no-deps --wheel-dir wheelhouse
+          python scripts/wheel_install_smoke_test.py wheelhouse/*.whl
+      - name: Compile Python
+        run: python -m compileall -q src tests
+      - name: Check Dashboard JavaScript syntax
+        run: find src/tianjun/interfaces/dashboard/static/js -name '*.js' -print0 | xargs -0 -n1 node --check
+      - name: Run executable frontend tests
+        run: npm run test:frontend
+      - name: Run tests
+        run: python -m pytest
+      - name: Run convergence checks
+        run: python scripts/convergence_check.py
+      - name: Run offline smoke test
+        run: python scripts/smoke_test.py --port 8136
+
+  browser:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+          cache: pip
+      - uses: actions/setup-node@v4
+        with:
+          node-version: "22"
+          cache: npm
+      - name: Install Tianjun and browser test dependencies
+        run: |
+          python -m pip install -e .
+          npm ci
+      - name: Install Chromium
+        run: npx playwright install --with-deps chromium
+      - name: Run browser tests
+        run: npm run test:browser
+      - uses: actions/upload-artifact@v4
+        if: ${{ !cancelled() }}
+        with:
+          name: playwright-report
+          path: playwright-report/
+          retention-days: 14
diff --git a/.gitignore b/.gitignore
@@ -5,6 +5,11 @@ __pycache__/
 .mypy_cache/
 .ruff_cache/
 *.egg-info/
+build/
+wheelhouse/
+node_modules/
+playwright-report/
+test-results/
 
 .env
 configs/secrets.toml
@@ -15,3 +20,7 @@ tianjun.toml
 
 data/alibaba_microservices_v2022/
 runtime_report.json
+
+# Local experiment outputs and temporary runtime data
+exp_out/
+tmp/
diff --git a/README.md b/README.md
@@ -4,6 +4,8 @@ Tianjun Engine 是一个本地优先的算网调度控制平面原型。它把
 
 本项目用于研究、演示和架构实验，并非生产级云平台。资源清单、定价、拓扑和执行事实必须来自已注册节点、CloudSimPlus 桥接器或真实节点代理；LLM 可以解释和帮助解析意图，但不能在没有明确确认路径的情况下捏造控制平面事实或提交工作。
 
+当前版本已加入批任务 JSON/CSV 导入、共享快照联合分配、Pareto + Tchebycheff 十维评分、Future-Fit 碎片评估、运行碳核算、显式确认原子预留以及外部 Hermes MCP 审计。接口和核算口径见 [批任务与运行碳调度说明](docs/batch-carbon-scheduling.md)。
+
 ## 快速开始
 
 ### 0. 安装依赖
@@ -84,9 +86,13 @@ CloudSimPlus 示例节点会随注册请求上报 CPU、内存、GPU、存储、
 任务下发后的状态流转如下：
 
 1. `/tasks` 或策略提交只把任务写入控制平面并加入 `pending_queue`，此时任务会显示为“待调度”，还不会出现在执行记录。
-2. CloudSimPlus 桥接器或真实节点代理持续心跳并请求 `/leases/next`，控制平面才会做调度决策并把任务租约发给目标节点。
-3. 节点执行过程中回传 `/task-runs/progress`，Dashboard 的拓扑路径会根据 `/report` 中的 `active_runs`、最新进度、调度决策和节点 inventory 实时切换。
-4. 节点最终回传 `/task-runs/result` 后，控制平面才写入执行记录；因此“执行记录没有这个任务”通常表示任务还在待调度队列或已发租约但尚未回传结果。
+2. CloudSimPlus 桥接器或真实节点代理持续心跳并请求 `/leases/next`，取得带唯一 `lease_id` 和 TTL 的任务租约，再通过 `/leases/ack` 确认领取。
+3. 节点执行过程中回传 `/task-runs/progress`，进度会续期租约；超时且没有续期的租约会释放资源并按重试策略重新排队。
+4. 节点最终通过 `/task-runs/result` 携带 `lease_id` 与稳定的 `result_id` 回传结果；相同结果重试会返回同一收据，不会重复写执行记录。
+
+控制平面默认把 SQLite v2 状态库写入平台用户状态目录（`${TIANJUN_STATE_DIR}/tianjun-state.sqlite`）。节点最后在线时间、任务、批次、预演计划、预留账本、幂等映射与结果收据均可跨进程重启恢复；可用 `--state-db` 覆盖路径。数据库升级前会在原目录创建 `*.pre-v2-from-v*.bak` 完整备份；高于程序支持版本的数据库会拒绝打开，迁移失败会整体回滚。`/ready` 会执行完整性检查和可逆写入探测。Dashboard 中“实时遥测”“CloudSim 模拟”“配置曲线”和“分配估算”是不同来源，不应互相替代解读。
+
+HTTP 服务内置独立生命周期清理器，不依赖 Dashboard 轮询即可回收过期节点和租约。默认每秒检查一次，可通过 `server.lifecycle_sweep_interval_seconds` 调整；服务关闭时会等待清理线程安全退出。
 
 ### 5. 打开 Dashboard
 
@@ -112,6 +118,8 @@ python -B main.py mcp-server `
 
 MCP server 会把 Tianjun HTTP API 包装为工具，包括读取集群状态、开始/继续聊天会话、起草/比较/仿真/解释策略，以及带确认边界的策略提交和任务调度。
 
+顶部的 MCP 状态只统计真正带 MCP 工具标识的 HTTP 请求，启动 MCP 进程本身不算成功调用。测试套件会启动真实 stdio MCP 子进程并调用 `get_cluster_state`，验证 Dashboard 的最近调用和成功计数同步更新。
+
 ### 7. 可选：真实节点代理
 
 如需使用真实节点遥测代理，而不是模拟节点：
@@ -143,6 +151,15 @@ python scripts\smoke_test.py --port 8135
 
 该脚本会启动离线控制平面，检查 `/health`、`/report`、`/dashboard`，并验证 MCP 工具契约可导入。它用于快速验证，不代表完整启动。
 
+前端逻辑和真实浏览器回归分别运行：
+
+```powershell
+npm run test:frontend
+npm run test:browser
+```
+
+浏览器回归使用 Chromium 覆盖 1366 和 1920 两档 PC 视口，包括标签页键盘语义、轮询竞态、拓扑图层、空态/错误态和拓扑几何边界。
+
 
 ## LLM 配置
 

diff --git a/configs/batch_experiments.json b/configs/batch_experiments.json
@@ -0,0 +1,12 @@
+{
+  "schema_version": "1.0",
+  "node_counts": [20, 50, 100],
+  "batch_task_counts": [20, 100, 500],
+  "load_rates": [0.30, 0.60, 0.85, 0.95],
+  "workloads": ["cpu", "gpu", "memory", "data", "mixed"],
+  "seeds": [20260718, 20260719, 20260720, 20260721, 20260722, 20260723, 20260724, 20260725, 20260726, 20260727],
+  "online_strategies": ["B0-current", "B1-batch-greedy", "B3-batch-local-search", "B4-pareto-tchebycheff", "B6-hierarchical-batch"],
+  "offline_strategies": ["B2-milp-oracle", "B5-nsga2"],
+  "carbon_scope": "operational_only",
+  "normalization_bounds_version": "engineering-v1"
+}
diff --git a/configs/cloudsim_core_experiments.json b/configs/cloudsim_core_experiments.json
@@ -0,0 +1,19 @@
+{
+  "schema_version": "1.0",
+  "description": "CloudSim Plus execution-level comparison. Every run starts an isolated Tianjun control plane and writes measured batch metrics under exp_out/cloudsim_core.",
+  "strategies": [
+    "B0-current",
+    "B4-pareto-tchebycheff",
+    "B6-hierarchical-batch",
+    "B6-green-single-v1",
+    "B6-green-sla-85-v1"
+  ],
+  "scenarios": [
+    "normal",
+    "fault-active"
+  ],
+  "seeds": [20260527, 20260528, 20260529, 20260530, 20260531, 20260532, 20260533, 20260534, 20260535, 20260536],
+  "cloudlets": 36,
+  "carbon_scope": "operational_only",
+  "output_directory": "exp_out/cloudsim_core"
+}
diff --git a/configs/fragmentation_green_experiments.json b/configs/fragmentation_green_experiments.json
@@ -0,0 +1,59 @@
+{
+  "schema_version": "3.0",
+  "description": "High-load heterogeneous-fragmentation and green objective ablation with held-out calibration seeds.",
+  "node_counts": [20],
+  "batch_task_counts": [60],
+  "load_rates": [0.85, 0.95],
+  "workloads": ["mixed"],
+  "fragmentation_modes": ["heterogeneous"],
+  "seeds": [20260718, 20260719, 20260720, 20260721, 20260722],
+  "training_seeds": [20260718, 20260719, 20260720],
+  "validation_seeds": [20260721, 20260722],
+  "online_strategies": [
+    "B0-current",
+    "B4-pareto-tchebycheff",
+    "B6-hierarchical-batch"
+  ],
+  "offline_strategies": [],
+  "objective_experiments": {
+    "single_atomic": ["carbon", "fragmentation"],
+    "dual_atomic": [
+      ["carbon", "completion"],
+      ["carbon", "fragmentation"]
+    ],
+    "single_groups": ["green_carbon", "resource_efficiency"],
+    "dual_groups": [
+      ["green_carbon", "sla_quality"],
+      ["green_carbon", "resource_efficiency"]
+    ]
+  },
+  "weight_profiles": [
+    {
+      "label": "W0-current-green",
+      "strategy": "B6-hierarchical-batch",
+      "group_weights": {"sla_quality": 0.26, "network_coordination": 0.20, "resource_efficiency": 0.22, "economic_cost": 0.12, "green_carbon": 0.20}
+    },
+    {
+      "label": "W1-carbon-080",
+      "strategy": "B6-hierarchical-batch",
+      "group_weights": {"sla_quality": 0.25, "network_coordination": 0.10, "resource_efficiency": 0.20, "economic_cost": 0.05, "green_carbon": 0.40}
+    },
+    {
+      "label": "W2-carbon-120",
+      "strategy": "B6-hierarchical-batch",
+      "group_weights": {"sla_quality": 0.20, "network_coordination": 0.07, "resource_efficiency": 0.20, "economic_cost": 0.03, "green_carbon": 0.50}
+    },
+    {
+      "label": "W3-carbon-160",
+      "strategy": "B6-hierarchical-batch",
+      "group_weights": {"sla_quality": 0.18, "network_coordination": 0.05, "resource_efficiency": 0.15, "economic_cost": 0.02, "green_carbon": 0.60}
+    },
+    {
+      "label": "W4-carbon-200",
+      "strategy": "B6-hierarchical-batch",
+      "group_weights": {"sla_quality": 0.20, "network_coordination": 0.04, "resource_efficiency": 0.25, "economic_cost": 0.01, "green_carbon": 0.50}
+    }
+  ],
+  "carbon_scope": "operational_only",
+  "normalization_bounds_version": "engineering-v1"
+}
diff --git a/configs/objective_ablation_experiments.json b/configs/objective_ablation_experiments.json
@@ -0,0 +1,38 @@
+{
+  "schema_version": "2.0",
+  "node_counts": [20, 50],
+  "batch_task_counts": [20, 100],
+  "load_rates": [0.60, 0.85],
+  "workloads": ["mixed"],
+  "seeds": [20260718, 20260719, 20260720, 20260721, 20260722],
+  "online_strategies": [
+    "B0-current",
+    "B4-pareto-tchebycheff",
+    "B6-hierarchical-batch"
+  ],
+  "offline_strategies": [],
+  "objective_experiments": {
+    "single_atomic": [
+      "performance",
+      "completion",
+      "cost",
+      "reliability",
+      "balance",
+      "fragmentation",
+      "locality",
+      "network",
+      "carbon"
+    ],
+    "dual_atomic": "all",
+    "single_groups": [
+      "sla_quality",
+      "network_coordination",
+      "resource_efficiency",
+      "economic_cost",
+      "green_carbon"
+    ],
+    "dual_groups": "all"
+  },
+  "carbon_scope": "operational_only",
+  "normalization_bounds_version": "engineering-v1"
+}
diff --git a/configs/tianjun.example.toml b/configs/tianjun.example.toml
@@ -6,7 +6,10 @@
 host = "127.0.0.1"
 port = 8024
 heartbeat_timeout_seconds = 15
+lease_timeout_seconds = 60
+lifecycle_sweep_interval_seconds = 1
 policy_update_interval = 2
+state_db = "${TIANJUN_STATE_DIR}/tianjun-state.sqlite"
 
 [model]
 dir = "${TIANJUN_HOME}/data/trained_models"

diff --git a/docs/api.md b/docs/api.md
@@ -6,8 +6,14 @@
 
 | 方法 | 路径 | 用途 |
 | --- | --- | --- |
-| GET | `/health` | 运行时健康、模型状态、聊天运行时状态 |
-| GET | `/report` | Dashboard 和工具用的控制平面状态 |
+| GET | `/health` | 进程健康和去敏后的依赖状态；始终返回可解析状态 |
+| GET | `/ready` | 就绪检查；依赖未就绪时返回 503 |
+| GET | `/report` | 完整控制平面状态，保留给 MCP 和兼容客户端 |
+| GET | `/report/summary` | Dashboard 总览和顶部状态的精简报告 |
+| GET | `/report/scheduling` | 调度决策页所需节点与决策报告 |
+| GET | `/report/topology` | 拓扑、VM 遥测和路径报告 |
+| GET | `/report/tasks?limit=50&cursor=0` | 分页任务执行报告 |
+| GET | `/report/model` | 模型、权重与策略历史报告 |
 | GET | `/dashboard` | 静态 Dashboard 壳页面 |
 | GET | `/chat/sessions/{session_id}` | 读取聊天会话状态 |
 | POST | `/chat/sessions` | 开始聊天会话 |
@@ -35,9 +41,10 @@
 | POST | `/nodes/heartbeat` | 更新节点心跳和遥测数据 |
 | POST | `/tasks` | 提交任务 |
 | POST | `/tasks/{task_id}/schedule` | 调度待处理任务；需要明确确认 |
-| POST | `/leases/next` | 节点代理租约轮询 |
-| POST | `/task-runs/progress` | 报告任务进度 |
-| POST | `/task-runs/result` | 报告最终任务结果 |
+| POST | `/leases/next` | 节点代理租约轮询；返回 `lease_id`、签发时间与到期时间 |
+| POST | `/leases/ack` | 使用 `node_id`、`task_id`、`lease_id` 确认领取并续期 |
+| POST | `/task-runs/progress` | 报告任务进度并续期租约；新客户端应携带 `lease_id` |
+| POST | `/task-runs/result` | 幂等报告最终结果；新客户端应携带 `lease_id` 和稳定 `result_id` |
 | POST | `/task-runs/cancel` | 取消活动任务运行 |
 | POST | `/schedule/preview` | CloudSimPlus 兼容的调度预览 |
 | POST | `/schedule/commit` | CloudSimPlus 兼容的直接提交 |
@@ -68,3 +75,9 @@ HTTP 路由调用 `CentralControlPlane` facade。facade 将已迁移的行为转
 - `PolicyWorkflowService` 处理策略起草、比较、模拟、提交和反馈路由。
 
 本文只描述公开 API 行为；服务拆分不改变路由语义。
+
+Dashboard 只轮询按页面拆分的报告。所有报告视图都带有 `report_version`、`resource_snapshot_version` 和 `generated_at`，客户端可据此识别跨请求快照差异。完整 `/report` 不再用于浏览器高频轮询。
+
+CloudSimPlus 心跳中的 `telemetry.cpu_utilization`、`ram_utilization` 和 `bandwidth_utilization` 会规范化为节点的 `runtime_utilization`。未上报的指标保持 `null`，Dashboard 显示为 `--`，不会生成伪实时值。
+
+租约默认 TTL 为 60 秒，可通过 `server.lease_timeout_seconds` 配置。相同 `result_id` 的重试返回已保存的结果收据并标记 `idempotent_replay=true`，不会重复累计执行、能耗或碳数据。
diff --git a/docs/architecture.md b/docs/architecture.md
@@ -28,12 +28,13 @@ Tianjun Engine 围绕公共适配器、中央控制平面门面和可测试的
 | 服务 | 当前职责 |
 | --- | --- |
 | `NodeRegistry` | 节点注册、心跳、节点遥测变更、节点持久化 |
-| `TaskLeaseService` | 任务提交、预览、pending 调度、agent 租约轮询、租约激活 |
+| `TaskLeaseService` | 任务提交、预览、pending 调度、租约发放、ACK、TTL 续期/回收和并发幂等 |
+| `LifecycleSweeper` | 与用户流量解耦地回收过期节点和租约，并随 HTTP 服务安全启停 |
 | `RequirementDialogueService` | 需求解析、需求会话开始/继续/读取、地域可用性载荷 |
 | `PolicyWorkflowService` | 策略起草、候选比较、模拟、提交、反馈解析、反馈记录、反馈优化 |
 | `src/tianjun/cli/commands/` | 所有 CLI 命令处理器 |
 
-`CentralControlPlane` 保留 facade 方法、共享状态、报表组装、恢复/持久化协调、拓扑注册、策略权重更新以及执行进度/结果回报等跨领域逻辑。已经迁移到服务中的业务流程不应复制回门面类。
+`CentralControlPlane` 保留 facade 方法、共享状态、拓扑注册、策略权重更新以及执行进度/结果回报等跨领域逻辑。数据库模式/迁移、控制面恢复、批输入校验、聊天模型/常量与拓扑几何已拆为独立模块；已经迁移到服务中的业务流程不应复制回门面类。
 
 ## CloudSimPlus 仿真链路
 
@@ -43,7 +44,8 @@ Tianjun Engine 围绕公共适配器、中央控制平面门面和可测试的
 - 调用 `/nodes/register` 注册 CloudSimPlus 仿真 VM 节点。
 - 持续调用 `/nodes/heartbeat` 上报在线状态。
 - 通过 `/schedule/commit` 请求 Tianjun 控制平面做调度决策。
-- 在 CloudSimPlus 仿真完成后通过 `/task-runs/result` 回报执行结果。
+- 通过 `/leases/next` 和 `/leases/ack` 领取并确认带 TTL 的任务租约。
+- 在 CloudSimPlus 仿真完成后通过携带 `lease_id`/`result_id` 的 `/task-runs/result` 幂等回报执行结果。
 
 完整启动命令见 [README.md](../README.md)。