fix(docker): pin GeoServer platform manifest

Add project-bound OAuth MCP queries and controlled analysis with scoped evidence,
limits, timeouts and a nearby-model spatial query using EPSG:4326 inputs.
Default nearby searches to 200 meters and five results.

Integrate investigation messaging/presentation, health/readiness and typed history;
remove duplicate web search and fix online simulation and compressed baseline writes.
Include project migrations 023-025; keep local credentials and demo seeding excluded.

Validation: 319 focused pytest checks; real PostGIS temporary-table regression;
read-only Jinqian bridge radius comparison; cross-language MCP discovery/calls.

.env.example

@@ -29,6 +29,8 @@ TIMESCALEDB_DB_HOST="localhost"
 TIMESCALEDB_DB_PORT="5433"
 TIMESCALEDB_DB_USER="tjwater"
 TIMESCALEDB_DB_PASSWORD="password"
+# 实时快照替换事务的解压预算；仅事务内生效，禁止0（无限制）。
+TIMESCALEDB_REALTIME_MAX_DECOMPRESSED_TUPLES=1000000
 
 # ============================================
 # 元数据数据库配置 (Metadata DB)
@@ -74,17 +76,52 @@ KEYCLOAK_ALGORITHM=RS256
 KEYCLOAK_AUDIENCE="account"
 KEYCLOAK_ACCESS_TOKEN_MAX_AGE_SECONDS=900
 
+# ============================================
+# 工作台事件同步
+# ============================================
+# 认证接口必须使用明确来源，逗号分隔；不要配置为 *。
+CORS_ALLOWED_ORIGINS="http://127.0.0.1:5173,http://localhost:5173"
+WORKBENCH_EVENTS_ENABLED=true
+WORKBENCH_EVENT_RETENTION_DAYS=7
+
+# 每个项目启动一个 worker 实例。长期运行优先使用 OAuth 客户端凭据。
+OPERATIONAL_WORKER_PROJECT_ID=""
+OPERATIONAL_WORKER_SERVER_URL="http://127.0.0.1:8000"
+OPERATIONAL_WORKER_AGENT_URL="http://127.0.0.1:8787"
+OPERATIONAL_WORKER_RECONCILE_SECONDS=30
+OPERATIONAL_WORKER_MAX_CONCURRENCY=4
+OPERATIONAL_WORKER_TOKEN_URL=""
+OPERATIONAL_WORKER_CLIENT_ID=""
+OPERATIONAL_WORKER_CLIENT_SECRET=""
+OPERATIONAL_WORKER_TOKEN_REFRESH_SECONDS=300
+OPERATIONAL_DATA_DELEGATION_BEARER=""
+
 
 # ============================================
-# Bocha Web Search API
+# Tianditu shared API key and service URLs
 # ============================================
-BOCHA_API_KEY="sk-your-bocha-api-key"
-BOCHA_WEB_SEARCH_URL="https://api.bochaai.com/v1/web-search"
-BOCHA_WEB_SEARCH_TIMEOUT_SECONDS=30
-
-# ============================================
-# Tianditu Geocoding API
-# ============================================
-TIANDITU_GEOCODER_TOKEN="your-tianditu-geocoder-token"
+TIANDITU_TOKEN="your-tianditu-token"
 TIANDITU_GEOCODER_URL="https://api.tianditu.gov.cn/geocoder"
 TIANDITU_GEOCODER_TIMEOUT_SECONDS=30
+
+# 地名搜索与地理编码共用 TIANDITU_TOKEN；也可沿用 TJWATER_TIANDITU_TOKEN。
+TIANDITU_SEARCH_URL="https://api.tianditu.gov.cn/v2/search"
+TIANDITU_SEARCH_TIMEOUT_SECONDS=15
+
+# MCP query resource, off until OAuth and every project BizDB are provisioned.
+# 仅 ENVIRONMENT=development 时覆盖 worker 取令牌地址；留空沿用 OPERATIONAL_WORKER_TOKEN_URL。
+OPERATIONAL_WORKER_DEV_TOKEN_URL=
+TJWATER_MCP_ENABLED=false
+TJWATER_MCP_ISSUER_URI=
+# 仅 ENVIRONMENT=development，可填写与 issuer 完全相同的内网 HTTP IPv4 issuer。
+# 正式环境必须留空；此例外不放行远程 HTTP MCP resource。
+TJWATER_MCP_DEV_HTTP_ISSUER_URI=
+TJWATER_MCP_RESOURCE_URI=
+TJWATER_MCP_AUDIENCE=tjwater-mcp
+TJWATER_MCP_HOST_CLIENT_ID=tjwater-agent-mcp
+TJWATER_MCP_ALLOWED_HOSTS=["localhost:*","127.0.0.1:*"]
+TJWATER_MCP_ALLOWED_ORIGINS=[]
+# 普通查询预算；计算工具由 Agent 用 timeout_seconds 按次指定（默认120秒）。
+TJWATER_MCP_QUERY_TIMEOUT_SECONDS=30
+# 计算等待上限，默认1小时；管理员可配置120至86400秒，实际上限通过 tools/list 发布。
+TJWATER_MCP_MAX_CALCULATION_TIMEOUT_SECONDS=3600

Dockerfile

@@ -1,13 +1,16 @@
-FROM condaforge/miniforge3:latest AS base
+FROM condaforge/miniforge3:26.7.2-0@sha256:569afc65b5338d35e276adf9a860025cd37a677659d19dcb71fc1d4c2063a94e AS base
 
 WORKDIR /app
 
+ENV CONDA_ENV_PATH=/opt/server \
+    PATH="/opt/server/bin:${PATH}"
+
 ENV PIP_INDEX_URL=https://pypi.tuna.tsinghua.edu.cn/simple \
     PIP_TRUSTED_HOST=pypi.tuna.tsinghua.edu.cn \
     UV_INDEX_URL=https://pypi.tuna.tsinghua.edu.cn/simple
 
-# 安装 Python 3.12 和 pymetis (通过 conda-forge 避免编译问题)
-RUN mamba install -y python=3.12 pymetis && \
+# 在独立环境中安装 Python 3.12 和 pymetis，避免修改 Miniforge 基础环境。
+RUN mamba create -y -p "$CONDA_ENV_PATH" python=3.12 pymetis pip && \
     mamba clean -afy
 
 COPY requirements.txt .
@@ -20,6 +23,8 @@ COPY app ./app
 COPY contracts ./contracts
 COPY infra ./infra
 COPY resources ./resources
+COPY scripts/run_operational_worker.py scripts/sync_peripheral_work_orders.py ./scripts/
+COPY samples/peripheral ./samples/peripheral
 RUN python -c "from pathlib import Path; from zipfile import ZipFile; model_dir = Path('app/algorithms/pipe_health_prediction/model'); zip_path = model_dir / 'my_survival_forest_model_quxi.zip'; joblib_name = 'my_survival_forest_model_quxi.joblib'; joblib_path = model_dir / joblib_name; assert zip_path.exists(), f'Model archive not found: {zip_path}'; archive = ZipFile(zip_path); archive.extract(joblib_name, model_dir); archive.close(); assert joblib_path.exists(), f'Model file not extracted: {joblib_path}'" && \
     rm -f app/algorithms/pipe_health_prediction/model/my_survival_forest_model_quxi.zip
 RUN mkdir -p db_inp temp data inp
@@ -33,7 +38,7 @@ COPY tests ./tests
 RUN python -m compileall -q app && \
     python -c "import app.main" && \
     python scripts/check_openapi.py && \
-    pytest -q tests
+    pytest -q --import-mode=importlib tests
 
 FROM base AS runner
 

README.md

@@ -43,8 +43,18 @@ conda run -n server uvicorn app.main:app --host 0.0.0.0 --port 8000 --reload
 conda activate server
 ```
 
+`GET /health` 是无需鉴权的 HTTP 服务存活检查，返回 `200` 和 `{"status":"ok"}`，不查询数据库。
+`GET /ready` 检查应用启动状态、元数据库 `SELECT 1` 及已启用的 MCP 传输生命周期；
+全部就绪返回 `200`，否则返回 `503`，元数据库探测预算为2秒。MCP 关闭时标记为
+`disabled`，不阻止就绪。返回结果不包含数据库地址、凭据或原始异常，禁止缓存。
+就绪状态不等同于 OAuth、用户权限或真实工具调用成功；指定项目的业务库和时序库仍通过
+带身份与 `X-Project-Id` 的 `GET /api/v1/projects/current/database-health` 检查。
+
 ## 常用命令
 
+外围接入样板是独立进程，复用 PostgreSQL，不自动建表；启动、授权边界和契约见 [外围样板说明](docs/peripheral-sample.md)。
+外围样板提供 synthetic 测试工单和测试数据的版本化 CRUD；工单经同步器投影为 `workbench.operational_tasks` 中的外围来源任务。固定演示时窗和造数方式见 [合成样本说明](docs/peripheral-generated-data.md)，运行态读取边界见 [工作台运行态说明](docs/workbench-runtime.md)。这些数据不代表现场事实，也不会触发设备控制。
+
 ```bash
 conda run -n server python -m pytest tests -q
 conda run -n server python scripts/run_server.py
@@ -73,6 +83,16 @@ docker compose -f infra/docker/docker-compose.yml config
 - `iot_data` DSN 用于 TimescaleDB，始终使用元数据配置的完整 DSN，不再从项目代码推导数据库名。
 - 元数据、业务库和 TimescaleDB 可以部署在同一主机，也可以分别部署。
 
+在线重算会原子替换一个时刻的节点、管段基线。历史分块已经压缩时，写入事务使用
+`TIMESCALEDB_REALTIME_MAX_DECOMPRESSED_TUPLES` 设置解压预算，默认100万条；配置必须
+为正整数，不允许0。该设置仅在实时快照写入事务内生效，不改变数据库全局设置或普通读取。
+预算按需根据受影响压缩批次的规模调整，超过预算仍会失败并回滚当前快照；此前完成的其他
+时刻不会一起回滚。同步、异步及单表批量快照入口使用同一预算。
+
+压缩写入回归位于 `tests/integration/test_realtime_compression.py`；仅在设置
+`REALTIME_COMPRESSION_TEST_DSN` 时执行，并且只接受库名以 `tjw_test_realtime_` 开头的
+独立可丢弃数据库。测试会创建和删除该库内的 `realtime` schema，禁止指向项目数据库。
+
 完整新建供水项目使用 `POST /api/v1/admin/project-provisions`，以
 `multipart/form-data` 同时提交 `name`、小写 `code`、可选的
 `description`、`gs_workspace`、`map_zoom` 和 INP `file`。工作流会按顺序完成：
@@ -87,6 +107,16 @@ docker compose -f infra/docker/docker-compose.yml config
 
 使用模板复制或临时方案库的模拟功能时，`biz_data` 账号必须具备数据库创建和删除权限；只有显式删除项目时才会终止该项目的现有数据库会话，普通复制不会主动中断复制源会话。
 
+奉贤项目的原始 GIS 与水质监测点使用独立脚本维护。导入脚本读取两份交付目录并写入项目 BizDB；水质脚本调用现有灵敏度分析服务生成 50 个监测点；发布脚本创建 10 个项目图层并刷新 GeoServer 缓存：
+
+```bash
+conda run -n server python scripts/import_fengxian_gis.py --database tjwater_v2
+conda run -n server python scripts/generate_fengxian_water_quality_points.py --project tjwater_v2
+conda run -n server python scripts/publish_fengxian_map_layers.py --database tjwater_v2 --workspace tjwater_next
+```
+
+发布命令需要配置 `GEOSERVER_URL`、`GEOSERVER_USERNAME` 和 `GEOSERVER_PASSWORD`。其中 7 个基础管网图层仍由项目供应流程创建，奉贤发布脚本追加 9 个原始 GIS 图层和 1 个水质监测点图层。
+
 ## 测试与发布
 
 提交前根据改动范围运行最小有效测试：
@@ -106,3 +136,7 @@ Gitea 包工作流位于 `.gitea/workflows/package.yml`，通常由 tag 触发
 ## 安全规则
 
 不要提交 `.env`、客户数据、数据库 dump、日志、生成缓存、`db_inp/`、`temp/`、`data/` 或本地密钥。CI/CD 凭据应放在 Gitea secrets 和仓库变量中。
+
+## 产品 MCP 查询
+
+Server 可选启用 MCP 2026-07-28 的25项项目查询及10项受控分析，默认关闭。它使用专用 OAuth resource Bearer、宿主绑定上下文及现有项目/任务权限，调查查询返回前持久化证据。启用需更新依赖、应用项目 SQL 迁移和配置 Keycloak token exchange；这些生产操作尚未执行。详见 [实施与启用](docs/mcp-query-implementation.md) 和 [协议/旧 CLI 能力对照](docs/mcp-query-design.md)。

app/api/health.py

@@ -0,0 +1,58 @@
+"""Public process probes, separate from authenticated project diagnostics."""
+
+import asyncio
+
+from fastapi import APIRouter, Request
+from fastapi.responses import JSONResponse
+from sqlalchemy import text
+
+from app.infra.db.metadb.database import engine
+
+
+router = APIRouter(include_in_schema=False)
+METADATA_TIMEOUT_SECONDS = 2
+
+
+async def metadata_database_ready() -> bool:
+    try:
+        async with asyncio.timeout(METADATA_TIMEOUT_SECONDS):
+            async with engine.connect() as connection:
+                result = await connection.execute(text("SELECT 1"))
+                return result.scalar_one() == 1
+    except Exception:
+        # Public probes must not expose connection strings or driver errors.
+        return False
+
+
+@router.get("/health")
+async def health() -> dict[str, str]:
+    """Check HTTP liveness without project or database dependencies."""
+    return {"status": "ok"}
+
+
+@router.get("/ready")
+async def readiness(request: Request) -> JSONResponse:
+    """Check startup, metadata connectivity and the enabled MCP lifecycle."""
+    started = getattr(request.app.state, "startup_complete", False)
+    runtime = getattr(request.app.state, "mcp_runtime", None)
+    mcp_ready = runtime is None or runtime.ready
+    metadata_ready = await metadata_database_ready() if started else False
+    ready = started and metadata_ready and mcp_ready
+    return JSONResponse(
+        status_code=200 if ready else 503,
+        content={
+            "status": "ready" if ready else "not_ready",
+            "checks": {
+                "startup": "ready" if started else "not_ready",
+                "metadata_database": (
+                    "not_checked" if not started
+                    else "ready" if metadata_ready else "unavailable"
+                ),
+                "mcp_transport": (
+                    "disabled" if runtime is None
+                    else "ready" if mcp_ready else "not_ready"
+                ),
+            },
+        },
+        headers={"Cache-Control": "no-store"},
+    )

app/api/problem_details.py

@@ -22,6 +22,10 @@ class ProblemDetails(BaseModel):
     code: str
     trace_id: str
     errors: list[dict[str, Any]] = Field(default_factory=list)
+    retryable: bool = False
+    related_refs: list[Any] = Field(default_factory=list)
+    recovery_action: str = "none"
+    retry_after_seconds: int | None = None
 
 
 def _trace_id(request: Request) -> str:
@@ -36,6 +40,10 @@ def _problem_response(
     detail: str,
     code: str,
     errors: list[dict[str, Any]] | None = None,
+    retryable: bool = False,
+    related_refs: list[Any] | None = None,
+    recovery_action: str = "none",
+    retry_after_seconds: int | None = None,
 ) -> JSONResponse:
     problem = ProblemDetails(
         type=f"https://tjwater.example/problems/{code.replace('_', '-')}",
@@ -46,6 +54,10 @@ def _problem_response(
         code=code,
         trace_id=_trace_id(request),
         errors=errors or [],
+        retryable=retryable,
+        related_refs=related_refs or [],
+        recovery_action=recovery_action,
+        retry_after_seconds=retry_after_seconds,
     )
     return JSONResponse(
         status_code=status_code,
@@ -89,7 +101,6 @@ def install_problem_details_handlers(app: FastAPI) -> None:
 
     @app.exception_handler(HTTPException)
     async def http_error_handler(request: Request, exc: HTTPException) -> JSONResponse:
-        detail = exc.detail if isinstance(exc.detail, str) else str(exc.detail)
         code_by_status = {
             401: "unauthenticated",
             403: "forbidden",
@@ -98,12 +109,43 @@ def install_problem_details_handlers(app: FastAPI) -> None:
             422: "validation_error",
             503: "dependency_unavailable",
         }
+        fallback_code = code_by_status.get(exc.status_code, "request_error")
+        detail_payload = exc.detail if isinstance(exc.detail, dict) else None
+        code = (
+            detail_payload.get("code")
+            if detail_payload and isinstance(detail_payload.get("code"), str)
+            else fallback_code
+        )
+        detail = exc.detail if isinstance(exc.detail, str) else None
+        if detail_payload:
+            candidate = detail_payload.get("message") or detail_payload.get("detail")
+            if isinstance(candidate, str):
+                detail = candidate
+        if detail is None:
+            detail = code.replace("_", " ").title()
+        related_refs = detail_payload.get("related_refs", []) if detail_payload else []
+        if not isinstance(related_refs, list):
+            related_refs = []
+        recovery_action = (
+            detail_payload.get("recovery_action", "none") if detail_payload else "none"
+        )
+        if not isinstance(recovery_action, str):
+            recovery_action = "none"
+        retry_after_seconds = (
+            detail_payload.get("retry_after_seconds") if detail_payload else None
+        )
+        if not isinstance(retry_after_seconds, int):
+            retry_after_seconds = None
         return _problem_response(
             request,
             status_code=exc.status_code,
-            title=code_by_status.get(exc.status_code, "request_error")
-            .replace("_", " ")
-            .title(),
+            title=fallback_code.replace("_", " ").title(),
             detail=detail,
-            code=code_by_status.get(exc.status_code, "request_error"),
+            code=code,
+            retryable=(
+                bool(detail_payload.get("retryable", False)) if detail_payload else False
+            ),
+            related_refs=related_refs,
+            recovery_action=recovery_action,
+            retry_after_seconds=retry_after_seconds,
         )

app/api/v1/endpoints/agent_messages.py

@@ -0,0 +1,101 @@
+"""Actor-private messaging APIs; the Agent verifies its session ownership."""
+from uuid import UUID
+
+from fastapi import APIRouter, Depends, HTTPException, Query
+from psycopg import AsyncConnection
+
+from app.auth.permissions import OPERATIONAL_TASK_VIEW, OPERATIONAL_TASK_WRITE, require_permission
+from app.auth.project_dependencies import ProjectContext, get_project_pg_connection
+from app.domain.domain_errors import DomainError
+from app.domain.schemas.agent_messages import (
+    AgentMessage, AgentMessageCreate, AgentMessageDelivery, AgentMessagePage, AgentMessageReply,
+)
+from app.infra.db.postgresql.agent_messages import AgentMessageRepository
+from app.services.agent_messages import AgentMessageService
+
+
+router = APIRouter(prefix="/agent-messages")
+
+
+def _service(ctx: ProjectContext, conn: AsyncConnection) -> AgentMessageService:
+    return AgentMessageService(AgentMessageRepository(conn, ctx.project_id), ctx)
+
+
+def _raise_domain(exc: DomainError):
+    raise HTTPException(status_code=exc.status, detail=exc.payload()) from exc
+
+
+@router.post("", response_model=AgentMessage, status_code=201)
+async def create_agent_message(
+    command: AgentMessageCreate,
+    ctx: ProjectContext = Depends(require_permission(OPERATIONAL_TASK_WRITE)),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+):
+    try:
+        return await _service(ctx, conn).create(command)
+    except DomainError as exc:
+        _raise_domain(exc)
+
+
+@router.get("", response_model=AgentMessagePage)
+async def list_agent_messages(
+    session_id: str = Query(min_length=1, max_length=500),
+    limit: int = Query(default=50, ge=1, le=100),
+    cursor: str | None = None,
+    pending_only: bool = False,
+    ctx: ProjectContext = Depends(require_permission(OPERATIONAL_TASK_VIEW)),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+):
+    try:
+        return await _service(ctx, conn).list_session(session_id, limit, cursor, pending_only)
+    except DomainError as exc:
+        _raise_domain(exc)
+
+
+@router.get("/pending", response_model=AgentMessagePage)
+async def list_pending_agent_messages(
+    limit: int = Query(default=50, ge=1, le=100),
+    cursor: str | None = None,
+    ctx: ProjectContext = Depends(require_permission(OPERATIONAL_TASK_VIEW)),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+):
+    try:
+        return await _service(ctx, conn).pending(limit, cursor)
+    except DomainError as exc:
+        _raise_domain(exc)
+
+
+@router.get("/{message_id}", response_model=AgentMessage)
+async def get_agent_message(
+    message_id: UUID,
+    ctx: ProjectContext = Depends(require_permission(OPERATIONAL_TASK_VIEW)),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+):
+    try:
+        return await _service(ctx, conn).get(message_id)
+    except DomainError as exc:
+        _raise_domain(exc)
+
+
+@router.post("/{message_id}/reply", response_model=AgentMessage, status_code=201)
+async def reply_agent_message(
+    message_id: UUID, command: AgentMessageReply,
+    ctx: ProjectContext = Depends(require_permission(OPERATIONAL_TASK_WRITE)),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+):
+    try:
+        return await _service(ctx, conn).reply(message_id, command)
+    except DomainError as exc:
+        _raise_domain(exc)
+
+
+@router.post("/{message_id}/delivery", response_model=AgentMessage)
+async def record_agent_message_delivery(
+    message_id: UUID, command: AgentMessageDelivery,
+    ctx: ProjectContext = Depends(require_permission(OPERATIONAL_TASK_VIEW)),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+):
+    try:
+        return await _service(ctx, conn).delivery(message_id, command)
+    except DomainError as exc:
+        _raise_domain(exc)

app/api/v1/endpoints/mcp_queries.py

@@ -0,0 +1,60 @@
+"""Host-only MCP bindings: declared in REST contracts even when MCP is disabled."""
+
+from fastapi import APIRouter, Header, HTTPException, Request, Response
+from psycopg.errors import UndefinedTable, InvalidSchemaName
+from app.mcp.models import BindInput, BindResponse
+from app.infra.db.postgresql.mcp_queries import McpQueryRepository
+
+router = APIRouter(tags=["MCP Query Context"])
+
+
+def runtime_for(request: Request):
+    runtime = getattr(request.app.state, "mcp_runtime", None)
+    if runtime is None:
+        raise HTTPException(503, "Product MCP is disabled")
+    return runtime
+
+
+@router.post("/mcp-query-contexts", response_model=BindResponse)
+async def create_context(
+    command: BindInput,
+    request: Request,
+    x_project_id: str = Header(alias="X-Project-Id"),
+) -> BindResponse:
+    runtime = runtime_for(request)
+    principal = await runtime.identity(request)
+    from app.mcp.contexts import bind_context
+
+    try:
+        async with runtime.scope_factory(principal, x_project_id) as scope:
+            binding = await bind_context(scope, principal, command, runtime.settings)
+    except (UndefinedTable, InvalidSchemaName):
+        raise HTTPException(
+            503, "MCP project storage has not been provisioned"
+        ) from None
+    return BindResponse(
+        context_handle=binding.context_handle,
+        expires_at=binding.expires_at,
+        task_id=binding.task_id,
+        run_id=binding.run_id,
+    )
+
+
+@router.delete("/mcp-query-contexts/{context_handle}", status_code=204)
+async def delete_context(
+    context_handle: str,
+    request: Request,
+    x_project_id: str = Header(alias="X-Project-Id"),
+) -> Response:
+    runtime = runtime_for(request)
+    principal = await runtime.identity(request)
+    try:
+        async with runtime.scope_factory(principal, x_project_id) as scope:
+            await McpQueryRepository(scope.connection, scope.project.project_id).revoke(
+                context_handle, principal.subject, principal.client_id
+            )
+    except (UndefinedTable, InvalidSchemaName):
+        raise HTTPException(
+            503, "MCP project storage has not been provisioned"
+        ) from None
+    return Response(status_code=204)

app/api/v1/endpoints/operational_tasks.py

@@ -0,0 +1,773 @@
+"""Project-scoped APIs for operational tasks, evidence and approved actions."""
+from __future__ import annotations
+
+from datetime import datetime
+from typing import Literal
+from uuid import UUID
+
+from fastapi import APIRouter, Depends, Header, HTTPException, Query, status
+from psycopg import AsyncConnection
+
+from app.auth.permissions import (
+    OPERATIONAL_TASK_EXECUTE,
+    OPERATIONAL_TASK_EXTERNAL_SYNC,
+    OPERATIONAL_TASK_VIEW,
+    OPERATIONAL_TASK_WRITE,
+    SCADA_VIEW,
+    SIMULATION_VIEW,
+    permissions_for_context,
+    require_permission,
+)
+from app.auth.project_dependencies import (
+    ProjectContext,
+    get_project_context,
+    get_project_pg_connection,
+)
+from app.domain.domain_errors import DomainError
+from app.domain.schemas.workbench_runtime import WorkbenchDateTime
+from app.domain.schemas.operational_data import TaskDataQuery, TaskDataSnapshot
+from app.domain.schemas.operational_task import (
+    ActionDecisionCommand,
+    ActionExecuteCommand,
+    ActionProposalCommand,
+    ActionReconcileCommand,
+    WorkOrderDraftUpdateCommand,
+    WorkOrderConfirmDispatchCommand,
+    TaskAccessRequest,
+    TaskAccessRequestCreate,
+    TaskAccessRequestDecision,
+    AgentArtifactsCommand,
+    AgentArtifactsResult,
+    AgentRunCancelCommand,
+    AgentRunCommand,
+    AgentRunDispatchFailureCommand,
+    AgentRunFailureCommand,
+    AgentTaskContext,
+    DispatchDueSchedulesCommand,
+    DispatchDueSchedulesResult,
+    EvidenceRecord,
+    ExternalFeedbackCommand,
+    ExternalFeedbackResult,
+    ExternalSourceWithdrawalCommand,
+    HumanApprovalAction,
+    HumanEvidenceCommand,
+    InvestigationReportPage,
+    InvestigationReportRevision,
+    NotificationAcknowledgeCommand,
+    OperationalTaskCreate,
+    OperationalTaskDetail,
+    OperationalTaskNotification,
+    OperationalTaskNotificationPage,
+    OperationalTaskPage,
+    OperationalTaskRetryCommand,
+    OperationalTaskRun,
+    OperationalTaskRunPage,
+    OperationalTaskSchedule,
+    OperationalTaskScheduleCreate,
+    OperationalTaskSchedulePage,
+    OperationalTaskScheduleUpdate,
+    RecoverExpiredLeasesCommand,
+    RecoverExpiredLeasesResult,
+    TaskStatus,
+    TaskDataQueryResult,
+    TaskType,
+    TaskPresentationEventCommand,
+    TaskPresentationEventResult,
+    TaskPresentationCursor,
+)
+from app.infra.db.postgresql.operational_tasks import OperationalTaskRepository
+from app.infra.integrations.operational_actions import (
+    OperationalActionDispatcher,
+    get_operational_action_dispatcher,
+)
+from app.infra.integrations.operational_data import TaskDataError, read_task_data
+from app.services.operational_tasks import OperationalTaskService
+
+
+router = APIRouter()
+
+
+def _service(ctx: ProjectContext, conn: AsyncConnection) -> OperationalTaskService:
+    return OperationalTaskService(
+        OperationalTaskRepository(conn, ctx.project_id),
+        ctx,
+        _task_permissions(ctx),
+    )
+
+
+def _task_permissions(ctx: ProjectContext) -> set[str]:
+    # Investigations use preconfigured project access, including service accounts.
+    # Do not force an already-authorized reader through legacy per-task delegation.
+    return set(permissions_for_context(ctx))
+
+
+def _raise_domain(exc: DomainError):
+    raise HTTPException(status_code=exc.status, detail=exc.payload()) from exc
+
+
+@router.get("/operational-tasks/{task_id}/reports", response_model=InvestigationReportPage)
+async def list_investigation_reports(
+    task_id: UUID,
+    limit: int = Query(default=20, ge=1, le=100),
+    offset: int = Query(default=0, ge=0),
+    ctx: ProjectContext = Depends(require_permission(OPERATIONAL_TASK_VIEW)),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+):
+    try:
+        items, next_offset = await _service(ctx, conn).list_reports(task_id, limit, offset)
+        return InvestigationReportPage(items=items, next_offset=next_offset)
+    except DomainError as exc:
+        _raise_domain(exc)
+
+
+@router.get("/operational-tasks/{task_id}/reports/{revision}", response_model=InvestigationReportRevision)
+async def get_investigation_report(
+    task_id: UUID,
+    revision: int,
+    ctx: ProjectContext = Depends(require_permission(OPERATIONAL_TASK_VIEW)),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+):
+    try:
+        return await _service(ctx, conn).report(task_id, revision)
+    except DomainError as exc:
+        _raise_domain(exc)
+
+
+@router.post(
+    "/operational-tasks",
+    response_model=OperationalTaskDetail,
+    status_code=status.HTTP_201_CREATED,
+)
+async def create_operational_task(
+    command: OperationalTaskCreate,
+    ctx: ProjectContext = Depends(require_permission(OPERATIONAL_TASK_WRITE)),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+):
+    try:
+        return await _service(ctx, conn).create(command)
+    except DomainError as exc:
+        _raise_domain(exc)
+
+
+@router.get("/operational-tasks", response_model=OperationalTaskPage)
+async def list_operational_tasks(
+    task_type: TaskType | None = None,
+    task_status: TaskStatus | None = Query(default=None, alias="status"),
+    limit: int = Query(default=50, ge=1, le=100),
+    offset: int = Query(default=0, ge=0),
+    peripheral_work_order_id: UUID | None = Query(
+        default=None, description="Only tasks sourced from or linked to this peripheral work order",
+    ),
+    ctx: ProjectContext = Depends(require_permission(OPERATIONAL_TASK_VIEW)),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+):
+    try:
+        items, next_offset = await _service(ctx, conn).list(
+            task_type=task_type, status=task_status, limit=limit, offset=offset,
+            **({"peripheral_work_order_id": peripheral_work_order_id}
+               if peripheral_work_order_id is not None else {}),
+        )
+        return OperationalTaskPage(
+            items=items, limit=limit, offset=offset, next_offset=next_offset
+        )
+    except DomainError as exc:
+        _raise_domain(exc)
+
+
+@router.get("/operational-task-runs", response_model=OperationalTaskRunPage)
+async def list_operational_task_runs(
+    run_status: Literal[
+        "queued", "claimed", "running", "succeeded", "failed", "timed_out", "cancelled"
+    ] | None = Query(default=None, alias="status"),
+    task_type: TaskType | None = None,
+    task_created_from: WorkbenchDateTime | None = None,
+    limit: int = Query(default=50, ge=1, le=100),
+    cursor: str | None = None,
+    ctx: ProjectContext = Depends(require_permission(OPERATIONAL_TASK_VIEW)),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+):
+    cursor_created_at = None
+    cursor_run_id = None
+    if cursor:
+        try:
+            timestamp, run_id = cursor.rsplit("|", 1)
+            cursor_created_at = datetime.fromisoformat(timestamp)
+            cursor_run_id = UUID(run_id)
+        except (ValueError, TypeError) as exc:
+            raise HTTPException(status_code=400, detail="Invalid run cursor") from exc
+    try:
+        items, next_cursor = await _service(ctx, conn).list_runs(
+            status=run_status,
+            task_type=task_type,
+            task_created_from=task_created_from,
+            limit=limit,
+            cursor_created_at=cursor_created_at,
+            cursor_run_id=cursor_run_id,
+        )
+        return OperationalTaskRunPage(items=items, next_cursor=next_cursor)
+    except DomainError as exc:
+        _raise_domain(exc)
+
+
+@router.get("/operational-tasks/{task_id}/runs", response_model=OperationalTaskRunPage)
+async def list_task_run_history(
+    task_id: UUID,
+    limit: int = Query(default=20, ge=1, le=100),
+    offset: int = Query(default=0, ge=0),
+    ctx: ProjectContext = Depends(require_permission(OPERATIONAL_TASK_VIEW)),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+):
+    try:
+        return OperationalTaskRunPage(items=await _service(ctx, conn).task_runs(task_id, limit, offset))
+    except DomainError as exc:
+        _raise_domain(exc)
+
+
+@router.get("/operational-tasks/{task_id}/runs/{run_id}", response_model=OperationalTaskRun)
+async def get_task_run(
+    task_id: UUID, run_id: UUID,
+    ctx: ProjectContext = Depends(require_permission(OPERATIONAL_TASK_VIEW)),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+):
+    try:
+        return await _service(ctx, conn).task_run(task_id, run_id)
+    except DomainError as exc:
+        _raise_domain(exc)
+
+
+@router.post(
+    "/operational-tasks/schedules",
+    response_model=OperationalTaskSchedule,
+    status_code=status.HTTP_201_CREATED,
+)
+async def create_operational_schedule(
+    command: OperationalTaskScheduleCreate,
+    ctx: ProjectContext = Depends(require_permission(OPERATIONAL_TASK_WRITE)),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+):
+    try:
+        return await _service(ctx, conn).create_schedule(command)
+    except DomainError as exc:
+        _raise_domain(exc)
+
+
+@router.get(
+    "/operational-tasks/schedules", response_model=OperationalTaskSchedulePage
+)
+async def list_operational_schedules(
+    ctx: ProjectContext = Depends(require_permission(OPERATIONAL_TASK_VIEW)),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+):
+    try:
+        return OperationalTaskSchedulePage(items=await _service(ctx, conn).list_schedules())
+    except DomainError as exc:
+        _raise_domain(exc)
+
+
+@router.patch(
+    "/operational-tasks/schedules/{schedule_id}",
+    response_model=OperationalTaskSchedule,
+)
+async def update_operational_schedule(
+    schedule_id: UUID,
+    command: OperationalTaskScheduleUpdate,
+    ctx: ProjectContext = Depends(require_permission(OPERATIONAL_TASK_WRITE)),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+):
+    try:
+        return await _service(ctx, conn).update_schedule(schedule_id, command)
+    except DomainError as exc:
+        _raise_domain(exc)
+
+
+@router.post(
+    "/operational-tasks/schedules/dispatch-due",
+    response_model=DispatchDueSchedulesResult,
+)
+async def dispatch_due_operational_schedules(
+    command: DispatchDueSchedulesCommand,
+    ctx: ProjectContext = Depends(require_permission(OPERATIONAL_TASK_EXECUTE)),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+):
+    try:
+        return await _service(ctx, conn).dispatch_due_schedules(command.limit)
+    except DomainError as exc:
+        _raise_domain(exc)
+
+
+@router.post(
+    "/operational-tasks/leases/recover",
+    response_model=RecoverExpiredLeasesResult,
+)
+async def recover_operational_task_leases(
+    command: RecoverExpiredLeasesCommand,
+    ctx: ProjectContext = Depends(require_permission(OPERATIONAL_TASK_EXECUTE)),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+):
+    try:
+        return await _service(ctx, conn).recover_expired_leases(
+            command.limit, task_type=command.task_type,
+            task_created_from=command.task_created_from,
+        )
+    except DomainError as exc:
+        _raise_domain(exc)
+
+
+@router.get(
+    "/operational-task-notifications",
+    response_model=OperationalTaskNotificationPage,
+)
+async def list_operational_task_notifications(
+    unread_only: bool = False,
+    limit: int = Query(default=50, ge=1, le=100),
+    offset: int = Query(default=0, ge=0),
+    ctx: ProjectContext = Depends(require_permission(OPERATIONAL_TASK_VIEW)),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+):
+    try:
+        items, next_offset = await _service(ctx, conn).list_notifications(
+            unread_only=unread_only, limit=limit, offset=offset
+        )
+        return OperationalTaskNotificationPage(
+            items=items, limit=limit, offset=offset, next_offset=next_offset
+        )
+    except DomainError as exc:
+        _raise_domain(exc)
+
+
+@router.post(
+    "/operational-task-notifications/{notification_id}/acknowledge",
+    response_model=OperationalTaskNotification,
+)
+async def acknowledge_operational_task_notification(
+    notification_id: UUID,
+    command: NotificationAcknowledgeCommand,
+    ctx: ProjectContext = Depends(require_permission(OPERATIONAL_TASK_VIEW)),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+):
+    try:
+        return await _service(ctx, conn).acknowledge_notification(notification_id, command)
+    except DomainError as exc:
+        _raise_domain(exc)
+
+
+@router.get("/operational-tasks/{task_id}", response_model=OperationalTaskDetail)
+async def get_operational_task(
+    task_id: UUID,
+    ctx: ProjectContext = Depends(require_permission(OPERATIONAL_TASK_VIEW)),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+):
+    try:
+        return await _service(ctx, conn).detail(task_id)
+    except DomainError as exc:
+        _raise_domain(exc)
+
+
+@router.post("/operational-tasks/{task_id}/retry", response_model=OperationalTaskDetail)
+async def retry_operational_task(
+    task_id: UUID,
+    command: OperationalTaskRetryCommand,
+    ctx: ProjectContext = Depends(require_permission(OPERATIONAL_TASK_WRITE)),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+):
+    try:
+        return await _service(ctx, conn).retry_task(task_id, command)
+    except DomainError as exc:
+        _raise_domain(exc)
+
+
+@router.post(
+    "/operational-tasks/{task_id}/runs/{run_id}/claim",
+    response_model=OperationalTaskRun,
+)
+async def claim_operational_task_run(
+    task_id: UUID,
+    run_id: UUID,
+    command: AgentRunCommand,
+    ctx: ProjectContext = Depends(require_permission(OPERATIONAL_TASK_EXECUTE)),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+):
+    try:
+        return await _service(ctx, conn).claim_run(task_id, run_id, command)
+    except DomainError as exc:
+        _raise_domain(exc)
+
+
+@router.post(
+    "/operational-tasks/{task_id}/runs/{run_id}/heartbeat",
+    response_model=OperationalTaskRun,
+)
+async def heartbeat_operational_task_run(
+    task_id: UUID,
+    run_id: UUID,
+    command: AgentRunCommand,
+    ctx: ProjectContext = Depends(require_permission(OPERATIONAL_TASK_EXECUTE)),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+):
+    try:
+        return await _service(ctx, conn).heartbeat_run(task_id, run_id, command)
+    except DomainError as exc:
+        _raise_domain(exc)
+
+
+@router.post(
+    "/operational-tasks/{task_id}/runs/{run_id}/complete",
+    response_model=OperationalTaskDetail,
+)
+async def complete_operational_task_run(
+    task_id: UUID,
+    run_id: UUID,
+    command: AgentRunCommand,
+    ctx: ProjectContext = Depends(require_permission(OPERATIONAL_TASK_EXECUTE)),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+):
+    try:
+        return await _service(ctx, conn).complete_run(task_id, run_id, command)
+    except DomainError as exc:
+        _raise_domain(exc)
+
+
+@router.post(
+    "/operational-tasks/{task_id}/runs/{run_id}/fail",
+    response_model=OperationalTaskDetail,
+)
+async def fail_operational_task_run(
+    task_id: UUID,
+    run_id: UUID,
+    command: AgentRunFailureCommand,
+    ctx: ProjectContext = Depends(require_permission(OPERATIONAL_TASK_EXECUTE)),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+):
+    try:
+        return await _service(ctx, conn).fail_run(task_id, run_id, command)
+    except DomainError as exc:
+        _raise_domain(exc)
+
+
+@router.post(
+    "/operational-tasks/{task_id}/runs/{run_id}/cancel",
+    response_model=OperationalTaskDetail,
+)
+async def cancel_operational_task_run(
+    task_id: UUID,
+    run_id: UUID,
+    command: AgentRunCancelCommand,
+    ctx: ProjectContext = Depends(require_permission(OPERATIONAL_TASK_WRITE)),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+):
+    try:
+        return await _service(ctx, conn).cancel_run(task_id, run_id, command)
+    except DomainError as exc:
+        _raise_domain(exc)
+
+
+@router.post(
+    "/operational-tasks/{task_id}/runs/{run_id}/dispatch-failure",
+    response_model=OperationalTaskRun,
+)
+async def record_operational_task_run_dispatch_failure(
+    task_id: UUID,
+    run_id: UUID,
+    command: AgentRunDispatchFailureCommand,
+    ctx: ProjectContext = Depends(require_permission(OPERATIONAL_TASK_EXECUTE)),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+):
+    try:
+        return await _service(ctx, conn).record_dispatch_failure(task_id, run_id, command)
+    except DomainError as exc:
+        _raise_domain(exc)
+
+
+@router.get(
+    "/operational-tasks/{task_id}/agent-context", response_model=AgentTaskContext
+)
+async def get_operational_task_agent_context(
+    task_id: UUID,
+    report_revision: int | None = Query(default=None, ge=1),
+    ctx: ProjectContext = Depends(require_permission(OPERATIONAL_TASK_VIEW)),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+):
+    try:
+        return await _service(ctx, conn).agent_context(task_id, report_revision)
+    except DomainError as exc:
+        _raise_domain(exc)
+
+
+@router.post(
+    "/operational-tasks/{task_id}/agent-artifacts",
+    response_model=AgentArtifactsResult,
+)
+async def publish_operational_task_artifacts(
+    task_id: UUID,
+    command: AgentArtifactsCommand,
+    ctx: ProjectContext = Depends(require_permission(OPERATIONAL_TASK_EXECUTE)),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+):
+    try:
+        return await _service(ctx, conn).publish_artifacts(task_id, command)
+    except DomainError as exc:
+        _raise_domain(exc)
+
+
+@router.get("/operational-tasks/{task_id}/presentation-cursor", response_model=TaskPresentationCursor)
+async def get_operational_task_presentation_cursor(
+    task_id: UUID,
+    ctx: ProjectContext = Depends(require_permission(OPERATIONAL_TASK_VIEW)),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+):
+    try:
+        return await _service(ctx, conn).presentation_cursor(task_id)
+    except DomainError as exc:
+        _raise_domain(exc)
+
+
+@router.post("/operational-tasks/{task_id}/presentation-events", response_model=TaskPresentationEventResult)
+async def publish_operational_task_presentation_event(
+    task_id: UUID,
+    command: TaskPresentationEventCommand,
+    ctx: ProjectContext = Depends(require_permission(OPERATIONAL_TASK_EXECUTE)),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+):
+    try:
+        return await _service(ctx, conn).publish_presentation_event(task_id, command)
+    except DomainError as exc:
+        _raise_domain(exc)
+
+
+@router.post(
+    "/operational-tasks/{task_id}/data-query", response_model=TaskDataQueryResult
+)
+async def query_operational_task_data(
+    task_id: UUID,
+    query: TaskDataQuery,
+    authorization: str | None = Header(default=None, alias="Authorization"),
+    ctx: ProjectContext = Depends(require_permission(OPERATIONAL_TASK_EXECUTE)),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+):
+    service = _service(ctx, conn)
+    try:
+        if not await service.can_read_source(task_id, query.source):
+            raise DomainError("PERMISSION_DENIED", "Task has no approved access to this data source", status=403)
+        source_permission = SCADA_VIEW if query.source.startswith("scada_") else SIMULATION_VIEW
+        snapshot = await read_task_data(
+            context=ctx, query=query, credential=authorization,
+            delegated=source_permission not in _task_permissions(ctx),
+        )
+        evidence, stored_snapshot = await service.record_data_snapshot(task_id, snapshot)
+        return TaskDataQueryResult(evidence=evidence, snapshot=stored_snapshot)
+    except TaskDataError as exc:
+        raise HTTPException(status_code=exc.status, detail=exc.detail) from exc
+    except DomainError as exc:
+        _raise_domain(exc)
+
+
+@router.post(
+    "/operational-tasks/{task_id}/access-requests",
+    response_model=TaskAccessRequest,
+    status_code=status.HTTP_201_CREATED,
+)
+async def request_operational_task_access(
+    task_id: UUID,
+    command: TaskAccessRequestCreate,
+    ctx: ProjectContext = Depends(require_permission(OPERATIONAL_TASK_EXECUTE)),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+):
+    try:
+        return await _service(ctx, conn).request_access(task_id, command)
+    except DomainError as exc:
+        _raise_domain(exc)
+
+
+@router.post(
+    "/operational-tasks/{task_id}/access-requests/{access_request_id}/decision",
+    response_model=TaskAccessRequest,
+)
+async def decide_operational_task_access(
+    task_id: UUID,
+    access_request_id: UUID,
+    command: TaskAccessRequestDecision,
+    ctx: ProjectContext = Depends(require_permission(OPERATIONAL_TASK_VIEW)),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+):
+    try:
+        return await _service(ctx, conn).decide_access(task_id, access_request_id, command)
+    except DomainError as exc:
+        _raise_domain(exc)
+
+
+@router.post(
+    "/operational-tasks/{task_id}/human-evidence",
+    response_model=OperationalTaskDetail,
+)
+async def add_operational_task_human_evidence(
+    task_id: UUID,
+    command: HumanEvidenceCommand,
+    ctx: ProjectContext = Depends(require_permission(OPERATIONAL_TASK_WRITE)),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+):
+    try:
+        return await _service(ctx, conn).add_human_evidence(task_id, command)
+    except DomainError as exc:
+        _raise_domain(exc)
+
+
+@router.get(
+    "/operational-tasks/{task_id}/evidence/{evidence_id}/snapshot",
+    response_model=TaskDataSnapshot,
+)
+async def get_operational_task_data_snapshot(
+    task_id: UUID,
+    evidence_id: str,
+    ctx: ProjectContext = Depends(require_permission(OPERATIONAL_TASK_VIEW)),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+):
+    try:
+        return await _service(ctx, conn).data_snapshot(task_id, evidence_id)
+    except DomainError as exc:
+        _raise_domain(exc)
+
+
+@router.post(
+    "/operational-tasks/{task_id}/actions",
+    response_model=HumanApprovalAction,
+    status_code=status.HTTP_201_CREATED,
+)
+async def propose_operational_task_action(
+    task_id: UUID,
+    command: ActionProposalCommand,
+    ctx: ProjectContext = Depends(require_permission(OPERATIONAL_TASK_WRITE)),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+):
+    try:
+        return await _service(ctx, conn).propose_action(task_id, command)
+    except DomainError as exc:
+        _raise_domain(exc)
+
+
+@router.post(
+    "/operational-tasks/{task_id}/actions/{action_id}/decision",
+    response_model=HumanApprovalAction,
+)
+async def decide_operational_task_action(
+    task_id: UUID,
+    action_id: UUID,
+    command: ActionDecisionCommand,
+    ctx: ProjectContext = Depends(require_permission(OPERATIONAL_TASK_VIEW)),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+):
+    try:
+        return await _service(ctx, conn).decide_action(task_id, action_id, command)
+    except DomainError as exc:
+        _raise_domain(exc)
+
+
+@router.patch(
+    "/operational-tasks/{task_id}/actions/{action_id}/draft",
+    response_model=HumanApprovalAction,
+)
+async def update_operational_work_order_draft(
+    task_id: UUID,
+    action_id: UUID,
+    command: WorkOrderDraftUpdateCommand,
+    ctx: ProjectContext = Depends(require_permission(OPERATIONAL_TASK_VIEW)),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+):
+    try:
+        return await _service(ctx, conn).update_work_order_draft(task_id, action_id, command)
+    except DomainError as exc:
+        _raise_domain(exc)
+
+
+@router.post(
+    "/operational-tasks/{task_id}/actions/{action_id}/confirm-dispatch",
+    response_model=HumanApprovalAction,
+)
+async def confirm_operational_work_order_dispatch(
+    task_id: UUID,
+    action_id: UUID,
+    command: WorkOrderConfirmDispatchCommand,
+    authorization: str | None = Header(default=None, alias="Authorization"),
+    ctx: ProjectContext = Depends(require_permission(OPERATIONAL_TASK_VIEW)),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+    dispatcher: OperationalActionDispatcher = Depends(get_operational_action_dispatcher),
+):
+    try:
+        return await _service(ctx, conn).confirm_work_order_dispatch(
+            task_id, action_id, command, dispatcher, authorization
+        )
+    except DomainError as exc:
+        _raise_domain(exc)
+
+
+@router.post(
+    "/operational-tasks/{task_id}/actions/{action_id}/execute",
+    response_model=HumanApprovalAction,
+)
+async def execute_operational_task_action(
+    task_id: UUID,
+    action_id: UUID,
+    command: ActionExecuteCommand,
+    authorization: str | None = Header(default=None, alias="Authorization"),
+    ctx: ProjectContext = Depends(require_permission(OPERATIONAL_TASK_VIEW)),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+    dispatcher: OperationalActionDispatcher = Depends(get_operational_action_dispatcher),
+):
+    try:
+        return await _service(ctx, conn).submit_action(
+            task_id, action_id, command, dispatcher, authorization
+        )
+    except DomainError as exc:
+        _raise_domain(exc)
+
+
+@router.post(
+    "/operational-tasks/{task_id}/actions/{action_id}/reconcile",
+    response_model=HumanApprovalAction,
+)
+async def reconcile_operational_task_action(
+    task_id: UUID,
+    action_id: UUID,
+    command: ActionReconcileCommand,
+    authorization: str | None = Header(default=None, alias="Authorization"),
+    ctx: ProjectContext = Depends(require_permission(OPERATIONAL_TASK_VIEW)),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+    dispatcher: OperationalActionDispatcher = Depends(get_operational_action_dispatcher),
+):
+    try:
+        return await _service(ctx, conn).reconcile_action(
+            task_id, action_id, command, dispatcher, authorization
+        )
+    except DomainError as exc:
+        _raise_domain(exc)
+
+
+@router.post(
+    "/operational-tasks/{task_id}/external-feedback",
+    response_model=ExternalFeedbackResult,
+)
+async def synchronize_operational_task_feedback(
+    task_id: UUID,
+    command: ExternalFeedbackCommand,
+    ctx: ProjectContext = Depends(require_permission(OPERATIONAL_TASK_EXTERNAL_SYNC)),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+):
+    try:
+        return await _service(ctx, conn).external_feedback(task_id, command)
+    except DomainError as exc:
+        _raise_domain(exc)
+
+
+@router.post(
+    "/operational-tasks/{task_id}/withdraw-external-source",
+    response_model=OperationalTaskDetail,
+)
+async def withdraw_operational_task_external_source(
+    task_id: UUID,
+    command: ExternalSourceWithdrawalCommand,
+    ctx: ProjectContext = Depends(require_permission(OPERATIONAL_TASK_EXTERNAL_SYNC)),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+):
+    try:
+        return await _service(ctx, conn).withdraw_external_source(task_id, command)
+    except DomainError as exc:
+        _raise_domain(exc)

app/api/v1/endpoints/place_search.py

@@ -0,0 +1,19 @@
+from fastapi import APIRouter, HTTPException
+
+from app.domain.schemas.place_search import PlaceSearchRequest, PlaceSearchResponse
+from app.services.place_search import PlaceSearchError, search_tianditu_places
+
+router = APIRouter()
+
+
+@router.post(
+    "/place-search-requests",
+    response_model=PlaceSearchResponse,
+    summary="Search Fengxian places",
+    description="调用天地图地名搜索，固定上海市奉贤区范围，只返回前 10 个可定位候选，不提供分页。",
+)
+async def tianditu_place_search(request: PlaceSearchRequest) -> PlaceSearchResponse:
+    try:
+        return await search_tianditu_places(request)
+    except PlaceSearchError as exc:
+        raise HTTPException(status_code=exc.status_code, detail=exc.detail) from exc

app/api/v1/endpoints/project_data.py

@@ -1,47 +1,116 @@
+import logging
 from uuid import UUID
 
 from fastapi import APIRouter, Depends, HTTPException, Query
 from psycopg import AsyncConnection
 
-from app.infra.db.postgresql.analysis import AnalysisRepository
 from app.auth.project_dependencies import get_project_pg_connection
+from app.domain.schemas.gis_attributes import (
+    ElementGisProperties,
+    GisElementType,
+    SourceGisFeatureProperties,
+)
+from app.infra.db.postgresql.analysis import AnalysisRepository
+from app.infra.db.postgresql.gis_attributes import GisAttributeRepository
 
 router = APIRouter()
+logger = logging.getLogger(__name__)
 
 
-async def get_database_connection(
+@router.get(
+    "/source-gis/layers/{layer_key}/features/{source_fid}/properties",
+    response_model=SourceGisFeatureProperties,
+    summary="获取源 GIS 要素台账属性",
+)
+async def get_source_gis_feature_properties(
+    layer_key: str,
+    source_fid: int,
     conn: AsyncConnection = Depends(get_project_pg_connection),
 ):
-    """获取数据库连接"""
-    yield conn
+    normalized_layer_key = layer_key.strip()
+    if (
+        not normalized_layer_key
+        or len(normalized_layer_key) > 80
+        or source_fid < 0
+    ):
+        raise HTTPException(status_code=422, detail="无效的源 GIS 要素编号")
+    try:
+        group = await GisAttributeRepository.get_source_feature_group(
+            conn,
+            normalized_layer_key,
+            source_fid,
+        )
+        if group is None:
+            raise HTTPException(status_code=404, detail="源 GIS 要素不存在")
+        return SourceGisFeatureProperties(
+            layer_key=normalized_layer_key,
+            source_fid=source_fid,
+            group=group,
+        )
+    except HTTPException:
+        raise
+    except Exception as exc:
+        logger.warning("Source GIS feature properties are unavailable", exc_info=True)
+        raise HTTPException(status_code=500, detail="查询源 GIS 要素属性失败") from exc
+
+
+@router.get(
+    "/elements/{element_type}/{element_id}/gis-properties",
+    response_model=ElementGisProperties,
+    summary="获取元素 GIS 台账属性",
+)
+async def get_element_gis_properties(
+    element_type: GisElementType,
+    element_id: str,
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+):
+    """Return source GIS records linked to one current-project map element."""
+    normalized_id = element_id.strip()
+    if not normalized_id or len(normalized_id) > 160:
+        raise HTTPException(status_code=422, detail="无效的元素编号")
+    try:
+        groups = await GisAttributeRepository.list_element_groups(
+            conn,
+            element_type,
+            normalized_id,
+        )
+        return ElementGisProperties(
+            element_type=element_type,
+            element_id=normalized_id,
+            groups=groups,
+        )
+    except Exception as exc:
+        logger.warning("Element GIS properties are unavailable", exc_info=True)
+        raise HTTPException(status_code=500, detail="查询 GIS 台账属性失败") from exc
 
 
 @router.get("/analysis/runs", summary="获取分析运行列表")
 async def get_analysis_runs(
-    conn: AsyncConnection = Depends(get_database_connection),
+    limit: int = Query(default=100, ge=1, le=100),
+    offset: int = Query(default=0, ge=0),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
 ):
-    """
-    获取所有方案信息
-    
-    返回项目中所有方案的详细信息
-    """
+    """分页返回当前项目的持久分析运行。"""
     try:
-        runs = await AnalysisRepository.list_runs(conn)
-        return {"success": True, "data": runs, "count": len(runs)}
-    except Exception as e:
-        raise HTTPException(status_code=500, detail=f"查询分析运行时发生错误: {str(e)}")
+        runs = await AnalysisRepository.list_runs(conn, limit=limit, offset=offset)
+        return {
+            "success": True,
+            "data": runs,
+            "count": len(runs),
+            "limit": limit,
+            "offset": offset,
+        }
+    except Exception as exc:
+        logger.warning("Analysis run page is unavailable", exc_info=True)
+        raise HTTPException(status_code=500, detail="查询分析运行失败") from exc
 
 
 @router.get("/analysis/runs/{run_id}", summary="获取分析运行")
 async def get_analysis_run(
     run_id: UUID,
-    conn: AsyncConnection = Depends(get_database_connection),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
 ):
-    """
-    获取所有爆管定位结果
-    
-    返回项目中所有的爆管定位分析结果
-    """
+    """返回当前项目的一条持久分析运行。"""
     try:
         run = await AnalysisRepository.get_run(conn, run_id)
         if run is None:
@@ -49,28 +118,31 @@ async def get_analysis_run(
         return run
     except HTTPException:
         raise
-    except Exception as e:
-        raise HTTPException(
-            status_code=500, detail=f"查询分析运行时发生错误: {str(e)}"
-        )
+    except Exception as exc:
+        logger.warning("Analysis run is unavailable", exc_info=True)
+        raise HTTPException(status_code=500, detail="查询分析运行失败") from exc
 
 
 @router.get("/analysis/runs/{run_id}/results", summary="获取分析结果")
 async def get_analysis_results(
     run_id: UUID,
     result_type: str | None = Query(default=None, description="结果类型"),
-    conn: AsyncConnection = Depends(get_database_connection),
+    limit: int = Query(default=100, ge=1, le=500),
+    offset: int = Query(default=0, ge=0),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
 ):
-    """
-    根据爆管事件ID查询爆管定位结果
-    
-    参数:
-        burst_incident: 爆管事件的唯一标识符
-    """
+    """分页返回一条分析运行的持久结果。"""
     try:
-        return await AnalysisRepository.list_results(conn, run_id, result_type)
-    except Exception as e:
+        return await AnalysisRepository.list_results(
+            conn,
+            run_id,
+            result_type,
+            limit=limit,
+            offset=offset,
+        )
+    except Exception as exc:
+        logger.warning("Analysis result page is unavailable", exc_info=True)
         raise HTTPException(
             status_code=500,
-            detail=f"查询分析结果时发生错误: {str(e)}",
-        )
+            detail="查询分析结果失败",
+        ) from exc

app/api/v1/endpoints/sensor_placement.py

@@ -146,6 +146,7 @@ async def optimize_sensor_placement_scheme(
             sensor_count=payload.sensor_count,
             min_diameter=payload.min_diameter,
             created_by=current_user.username,
+            sensor_type=payload.sensor_type,
         )
         run = get_sensor_placement_run(network, created["run_id"])
         return {**run, "can_edit": True}

app/api/v1/endpoints/simulation.py

@@ -1,5 +1,5 @@
 from typing import Any, List, Literal, Optional
-from datetime import datetime, timedelta
+from datetime import datetime
 from fastapi import APIRouter, Body, Depends, HTTPException, Query
 from fastapi.responses import PlainTextResponse
 from app.auth.keycloak_dependencies import get_current_keycloak_username
@@ -18,7 +18,6 @@ from app.services.simulation_scenarios import (
 from app.services.valve_isolation import analyze_valve_isolation
 from app.domain.time import (
     parse_aware_time,
-    parse_clock_duration_seconds,
     parse_utc_time,
 )
 from pydantic import BaseModel, Field, field_validator
@@ -50,27 +49,9 @@ class PressureRegulation(BaseModel):
 def run_simulation_manually_by_date(
     network_name: str, start_time: datetime, duration: int
 ) -> None:
-    end_datetime = start_time + timedelta(minutes=duration)
-    time_properties = simulation.get_time(network_name)
-    hydraulic_step_seconds = parse_clock_duration_seconds(
-        time_properties["HYDRAULIC TIMESTEP"],
-        field_name="HYDRAULIC TIMESTEP",
+    simulation.run_online_simulation(
+        name=network_name, start_time=start_time, duration_seconds=duration * 60,
     )
-    if hydraulic_step_seconds <= 0:
-        raise ValueError("HYDRAULIC TIMESTEP must be greater than 0.")
-    hydraulic_step = timedelta(seconds=hydraulic_step_seconds)
-    scada_mappings = simulation.query_corresponding_element_id_and_query_id(
-        network_name
-    )
-    current_time = start_time
-    while current_time < end_datetime:
-        simulation.run_simulation(
-            name=network_name,
-            simulation_type="realtime",
-            modify_pattern_start_time=current_time.isoformat(timespec="seconds"),
-            scada_mappings=scada_mappings,
-        )
-        current_time += hydraulic_step
 
 
 # 必须用这个PlainTextResponse，不然每个key都有引号

app/api/v1/endpoints/timeseries/composite.py

@@ -11,9 +11,10 @@ from app.services.timeseries_history import TimeseriesHistoryService
 from .dependencies import get_timescale_connection, get_postgres_connection
 
 router = APIRouter()
+history_router = APIRouter()
 
 
-@router.post(
+@history_router.post(
     "/timeseries/views/element-history/query",
     summary="批量查询管网元素历史数据",
     response_model=ElementHistoryResponse,

app/api/v1/endpoints/web_search.py

@@ -1,29 +0,0 @@
-from typing import Any
-
-from fastapi import APIRouter, HTTPException, status
-
-from app.services.web_search import (
-    BochaSearchAPIError,
-    BochaSearchConfigError,
-    WebSearchRequest,
-    search_bocha_web,
-)
-
-router = APIRouter()
-
-
-@router.post(
-    "/web-searches",
-    summary="Web Search",
-    description="调用 Bocha Web Search API 获取实时网页搜索结果",
-)
-async def web_search(request: WebSearchRequest) -> dict[str, Any]:
-    try:
-        return await search_bocha_web(request)
-    except BochaSearchConfigError as exc:
-        raise HTTPException(
-            status_code=status.HTTP_503_SERVICE_UNAVAILABLE,
-            detail=str(exc),
-        ) from exc
-    except BochaSearchAPIError as exc:
-        raise HTTPException(status_code=exc.status_code, detail=exc.detail) from exc

app/api/v1/endpoints/workbench_events.py

@@ -0,0 +1,156 @@
+"""Authenticated workbench snapshot and resumable Server-Sent Events."""
+from __future__ import annotations
+
+import asyncio
+import json
+from datetime import datetime, timezone
+
+from fastapi import APIRouter, Depends, Header, HTTPException, Query
+from fastapi.responses import StreamingResponse
+from psycopg import AsyncConnection
+from psycopg.conninfo import conninfo_to_dict
+
+from app.api.v1.endpoints.workbench_runtime import workbench_runtime
+from app.auth.permissions import OPERATIONAL_TASK_VIEW, permissions_for_context, require_permission
+from app.auth.project_dependencies import (
+    ProjectContext,
+    get_metadata_repository,
+    get_project_business_routing,
+    get_project_pg_connection,
+)
+from app.core.config import settings
+from app.domain.schemas.workbench_events import WorkbenchEvent, WorkbenchSyncSnapshot
+from app.infra.db.postgresql.operational_tasks import OperationalTaskRepository
+from app.infra.db.postgresql.workbench_events import WorkbenchEventRepository
+from app.infra.db.metadb.repositories.metadata_repository import MetadataRepository
+from app.infra.db.project_routing import ActiveProjectRouting
+from app.services.operational_tasks import OperationalTaskService
+from app.infra.db.postgresql.workbench_event_hub import workbench_event_hubs
+
+
+router = APIRouter()
+
+
+def _service(ctx: ProjectContext, conn: AsyncConnection) -> OperationalTaskService:
+    return OperationalTaskService(
+        OperationalTaskRepository(conn, ctx.project_id),
+        ctx,
+        permissions_for_context(ctx),
+    )
+
+
+@router.get("/workbench/sync-snapshot", response_model=WorkbenchSyncSnapshot)
+async def workbench_sync_snapshot(
+    ctx: ProjectContext = Depends(require_permission(OPERATIONAL_TASK_VIEW)),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+):
+    async with conn.transaction():
+        await conn.execute(
+            "SET TRANSACTION ISOLATION LEVEL REPEATABLE READ READ ONLY"
+        )
+        cursor = await WorkbenchEventRepository(conn, ctx.project_id).current_cursor()
+        runtime = await workbench_runtime(ctx=ctx, conn=conn)
+        tasks, _ = await _service(ctx, conn).list(
+            task_type=None, status=None, limit=50, offset=0
+        )
+        notifications, _ = await _service(ctx, conn).list_notifications(
+            unread_only=False, limit=100, offset=0
+        )
+        return WorkbenchSyncSnapshot(
+            schema="workbench-sync-snapshot@1",
+            event_cursor=cursor,
+            runtime=runtime,
+            tasks=tasks,
+            notifications=notifications,
+            generated_at=datetime.now(timezone.utc),
+        )
+
+
+@router.get("/workbench/events")
+async def workbench_events(
+    after: int | None = Query(default=None, ge=0),
+    last_event_id: str | None = Header(default=None, alias="Last-Event-ID"),
+    ctx: ProjectContext = Depends(require_permission(OPERATIONAL_TASK_VIEW)),
+    routing: ActiveProjectRouting = Depends(get_project_business_routing),
+    metadata_repo: MetadataRepository = Depends(get_metadata_repository),
+):
+    if not settings.WORKBENCH_EVENTS_ENABLED:
+        raise HTTPException(status_code=503, detail="Workbench events are disabled")
+    cursor = after
+    if cursor is None and last_event_id:
+        try:
+            cursor = int(last_event_id)
+        except ValueError as exc:
+            raise HTTPException(status_code=400, detail="Invalid Last-Event-ID") from exc
+    cursor = cursor or 0
+    current_identity = _database_identity(routing.business_dsn)
+    try:
+        active_routings = await metadata_repo.list_active_project_db_routings("biz_data")
+    finally:
+        # FastAPI keeps yield dependencies alive for a StreamingResponse. The
+        # stream only needs immutable routing values, so release the shared
+        # metadata checkout before the five-minute SSE body begins.
+        await metadata_repo.session.close()
+    if any(
+        item.project_id != ctx.project_id
+        and _database_identity(item.dsn) == current_identity
+        for item in active_routings
+    ):
+        raise HTTPException(
+            status_code=503,
+            detail="Workbench events require one dedicated business database per project",
+        )
+
+    async def stream():
+        started = asyncio.get_running_loop().time()
+        last_keepalive = started
+        async with workbench_event_hubs.subscribe(ctx.project_id, routing.business_dsn) as hub:
+            oldest = await hub.oldest_cursor()
+            latest = await hub.current_cursor()
+            if cursor > 0 and (
+                oldest is None or cursor < oldest - 1 or cursor > latest
+            ):
+                payload = {"schema": "workbench-event@1", "cursor": cursor}
+                yield f"event: reset_required\ndata: {json.dumps(payload, separators=(',', ':'))}\n\n"
+                return
+            current = cursor
+            wake_generation = hub.generation
+            while asyncio.get_running_loop().time() - started < 300:
+                rows = await hub.read_after(current)
+                for row in rows:
+                    event = WorkbenchEvent(
+                        schema="workbench-event@1",
+                        **row,
+                    )
+                    current = event.event_seq
+                    data = event.model_dump_json(by_alias=True)
+                    yield f"id: {current}\nevent: workbench_event\ndata: {data}\n\n"
+                if rows:
+                    continue
+                now = asyncio.get_running_loop().time()
+                if now - last_keepalive >= 15:
+                    yield f": keepalive {int(now)}\n\n"
+                    last_keepalive = now
+                # The Hub invalidates shared reads on NOTIFY and every 30s.
+                # Per-client timeout only keeps the transport alive, not SQL.
+                wake_generation = await hub.wait_for_wake(wake_generation, timeout=15)
+            yield "event: reconnect\ndata: {}\n\n"
+
+    return StreamingResponse(
+        stream(),
+        media_type="text/event-stream",
+        headers={
+            "Cache-Control": "no-cache, no-transform",
+            "X-Accel-Buffering": "no",
+            "Content-Encoding": "identity",
+        },
+    )
+
+
+def _database_identity(dsn: str) -> tuple[str, str, str]:
+    values = conninfo_to_dict(dsn)
+    return (
+        values.get("host") or values.get("hostaddr") or "localhost",
+        values.get("port") or "5432",
+        values.get("dbname") or values.get("user") or "",
+    )

app/api/v1/endpoints/workbench_runtime.py

@@ -0,0 +1,155 @@
+"""Read-only workbench runtime aggregation."""
+
+import logging
+from contextlib import asynccontextmanager
+from uuid import UUID
+
+from fastapi import APIRouter, Depends, HTTPException, Query, status
+from psycopg import AsyncConnection
+from pydantic import ValidationError
+
+from app.auth.project_dependencies import (
+    ProjectContext,
+    get_project_context,
+    get_project_pg_connection,
+)
+from app.domain.schemas.workbench_runtime import (
+    WorkbenchAnalysisHistoryPage,
+    WorkbenchPersistedAlert,
+    WorkbenchRuntimeLatestAnalysis,
+    WorkbenchRuntimeSnapshot,
+)
+from app.domain.schemas.task_alert import TaskAlertCommand, TaskAlertEvent, TaskAlertState
+from app.infra.db.postgresql.analysis import AnalysisRepository
+from app.infra.db.postgresql.workbench_operations import (
+    WorkbenchOperationsRepository,
+)
+from app.infra.db.postgresql.task_alerts import TaskAlertRepository
+from app.services.task_alerts import TaskAlertService
+from app.domain.domain_errors import DomainError
+from app.services.workbench_runtime import (
+    build_workbench_runtime_snapshot,
+    project_analysis_history_item,
+    project_latest_analysis,
+    project_sensor_location_map_payload,
+    sensor_location_ids_for_map_projection,
+)
+
+router = APIRouter()
+logger = logging.getLogger(__name__)
+
+
+@asynccontextmanager
+async def _projection_transaction(conn: AsyncConnection):
+    """Isolate an optional projection so one SQL failure cannot poison others."""
+    transaction = getattr(conn, "transaction", None)
+    if transaction is None:  # Lightweight endpoint test doubles.
+        yield
+        return
+    async with transaction():
+        yield
+
+
+@router.get(
+    "/workbench/analysis-runs",
+    response_model=WorkbenchAnalysisHistoryPage,
+)
+async def workbench_analysis_runs(
+    limit: int = Query(default=20, ge=1, le=100),
+    offset: int = Query(default=0, ge=0),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+):
+    """Return a bounded, sanitized analysis history page for the workbench."""
+    rows = await AnalysisRepository.list_run_summaries(
+        conn,
+        limit=limit + 1,
+        offset=offset,
+    )
+    has_more = len(rows) > limit
+    items = rows[:limit]
+    return {
+        "items": [project_analysis_history_item(run) for run in items],
+        "limit": limit,
+        "offset": offset,
+        "next_offset": offset + limit if has_more else None,
+    }
+
+
+@router.get(
+    "/workbench/analysis-runs/{run_id}",
+    response_model=WorkbenchRuntimeLatestAnalysis,
+)
+async def workbench_analysis_run(
+    run_id: UUID,
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+):
+    """Return one authorized project's analysis run as a typed workbench artifact."""
+    run = await AnalysisRepository.get_run(conn, run_id)
+    if run is None:
+        raise HTTPException(
+            status_code=status.HTTP_404_NOT_FOUND,
+            detail="Analysis run not found",
+        )
+    results = await AnalysisRepository.list_recent_results(conn, run_id, limit=100)
+    try:
+        referenced_map_payload = await _sensor_location_map_payload(conn, results)
+    except Exception:
+        referenced_map_payload = None
+        logger.warning(
+            "Workbench analysis spatial references are unavailable",
+            exc_info=True,
+        )
+    return project_latest_analysis(
+        run,
+        results,
+        referenced_map_payload=referenced_map_payload,
+    )
+
+
+@router.post("/workbench/task-alert-events", response_model=TaskAlertState)
+async def push_task_alert_event(
+    command: TaskAlertCommand,
+    ctx: ProjectContext = Depends(get_project_context),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+):
+    try:
+        event = TaskAlertEvent(
+            **command.model_dump(),
+            producer_id=f"principal:{ctx.user_id}",
+        )
+        return await TaskAlertService(
+            TaskAlertRepository(conn, ctx.project_id), ctx
+        ).push(event)
+    except DomainError as exc:
+        raise HTTPException(status_code=exc.status, detail=exc.message) from exc
+
+
+@router.get("/workbench/runtime", response_model=WorkbenchRuntimeSnapshot)
+async def workbench_runtime(
+    ctx: ProjectContext = Depends(get_project_context),
+    conn: AsyncConnection = Depends(get_project_pg_connection),
+):
+    from app.services.workbench_runtime import read_workbench_runtime
+    return await read_workbench_runtime(ctx, conn)
+
+
+async def _sensor_location_map_payload(
+    conn: AsyncConnection, results: list[dict]
+) -> dict | None:
+    node_ids = sensor_location_ids_for_map_projection(results)
+    if not node_ids:
+        return None
+    map_points = await AnalysisRepository.list_junction_map_points(conn, node_ids)
+    return project_sensor_location_map_payload(node_ids, map_points)
+
+
+def _validate_documents(documents, model, label: str) -> tuple[list[dict], int]:
+    valid: list[dict] = []
+    rejected = 0
+    for document in documents:
+        try:
+            valid.append(model.model_validate(document).model_dump())
+        except ValidationError:
+            rejected += 1
+            logger.warning("Workbench %s document was rejected", label)
+    return valid, rejected

app/api/v1/router.py

@@ -4,6 +4,7 @@ from app.api.v1.endpoints import (
     access,
     admin_metadata,
     agent_auth,
+    agent_messages,
     audit,
     burst_detection,
     burst_location,
@@ -11,12 +12,15 @@ from app.api.v1.endpoints import (
     leakage,
     meta,
     model_import,
+    operational_tasks,
+    place_search,
     project,
     project_data,
     scada,
     sensor_placement,
     simulation,
-    web_search,
+    workbench_runtime,
+    workbench_events,
 )
 from app.api.v1.endpoints.components import (
     controls,
@@ -85,8 +89,22 @@ burst_run_access = Depends(require_permission(BURST_RUN))
 optimization_run_access = Depends(require_permission(OPTIMIZATION_RUN))
 
 # Core services
+api_router.include_router(
+    workbench_runtime.router,
+    tags=["Workbench Runtime"],
+    dependencies=[simulation_access],
+)
+api_router.include_router(
+    workbench_events.router,
+    tags=["Workbench Events"],
+)
+api_router.include_router(
+    operational_tasks.router,
+    tags=["Operational Tasks"],
+)
 api_router.include_router(access.router, tags=["Access Control"])
 api_router.include_router(agent_auth.router, tags=["Agent Auth"])
+api_router.include_router(agent_messages.router, tags=["Agent Messages"])
 api_router.include_router(
     admin_metadata.router,
     tags=["Metadata Admin"],
@@ -143,13 +161,13 @@ api_router.include_router(
     dependencies=[optimization_run_access],
 )
 api_router.include_router(
-    web_search.router,
-    tags=["Web Search"],
+    geocoding.router,
+    tags=["Geocoding"],
     dependencies=[webgis_view_access],
 )
 api_router.include_router(
-    geocoding.router,
-    tags=["Geocoding"],
+    place_search.router,
+    tags=["Place Search"],
     dependencies=[webgis_view_access],
 )
 api_router.include_router(
@@ -179,6 +197,12 @@ for endpoint_router, tag in (
         dependencies=[simulation_access],
     )
 
+api_router.include_router(
+    ts_composite.history_router,
+    tags=["TimescaleDB - Composite"],
+    dependencies=[Depends(require_permission(SCADA_VIEW))],
+)
+
 for endpoint_router, tag in (
     (ts_scada.router, "TimescaleDB - SCADA"),
     (ts_composite.router, "TimescaleDB - Composite"),

app/auth/keycloak_dependencies.py

@@ -6,6 +6,9 @@ from fastapi.security import OAuth2PasswordBearer
 from jose import JWTError, jwt
 
 from app.core.config import settings
+from app.mcp.config import McpSettings
+
+mcp_resource_settings = McpSettings()
 
 oauth2_optional = OAuth2PasswordBearer(
     tokenUrl="keycloak", auto_error=False
@@ -29,6 +32,13 @@ def _decode_keycloak_token(token: str) -> dict:
         algorithms=[settings.KEYCLOAK_ALGORITHM],
         audience=settings.KEYCLOAK_AUDIENCE or None,
     )
+    # A resource-bound MCP token must never authorize product REST APIs,
+    # including when another audience overlaps the configured API audience.
+    audience = payload.get("aud", [])
+    audiences = [audience] if isinstance(audience, str) else audience
+    if (isinstance(audiences, list) and mcp_resource_settings.audience in audiences
+            or payload.get("azp") == mcp_resource_settings.host_client_id):
+        raise JWTError("MCP resource tokens cannot authorize product REST APIs")
     if settings.KEYCLOAK_ACCESS_TOKEN_MAX_AGE_SECONDS <= 0:
         return payload
 

app/auth/permissions.py

@@ -21,6 +21,15 @@ MODEL_IMPORT = "model.import"
 AUDIT_VIEW = "audit.view"
 ENVIRONMENT_MANAGE = "environment.manage"
 MEMBERSHIP_MANAGE = "membership.manage"
+OPERATIONAL_TASK_VIEW = "operational_task.view"
+OPERATIONAL_TASK_WRITE = "operational_task.write"
+OPERATIONAL_TASK_EXECUTE = "operational_task.execute"
+OPERATIONAL_TASK_EXTERNAL_SYNC = "operational_task.external_sync"
+OPERATIONAL_TASK_ACCESS_APPROVE = "operational_task.access.approve"
+OPERATIONAL_ACTION_WORK_ORDER_APPROVE = "operational_action.work_order.approve"
+OPERATIONAL_ACTION_WORK_ORDER_EXECUTE = "operational_action.work_order.execute"
+OPERATIONAL_ACTION_COMMAND_APPROVE = "operational_action.command.approve"
+OPERATIONAL_ACTION_COMMAND_EXECUTE = "operational_action.command.execute"
 
 PROJECT_MEMBER_PERMISSIONS = frozenset(
     {
@@ -36,6 +45,11 @@ PROJECT_MEMBER_PERMISSIONS = frozenset(
         RISK_RUN,
         OPTIMIZATION_VIEW,
         OPTIMIZATION_RUN,
+        OPERATIONAL_TASK_VIEW,
+        OPERATIONAL_TASK_WRITE,
+        OPERATIONAL_TASK_ACCESS_APPROVE,
+        OPERATIONAL_ACTION_WORK_ORDER_APPROVE,
+        OPERATIONAL_ACTION_WORK_ORDER_EXECUTE,
     }
 )
 
@@ -44,6 +58,7 @@ PROJECT_VIEWER_PERMISSIONS = frozenset(
         WEBGIS_VIEW,
         SCADA_VIEW,
         SIMULATION_VIEW,
+        OPERATIONAL_TASK_VIEW,
     }
 )
 
@@ -53,6 +68,15 @@ SYSTEM_ADMIN_PERMISSIONS = frozenset(
         AUDIT_VIEW,
         ENVIRONMENT_MANAGE,
         MEMBERSHIP_MANAGE,
+        OPERATIONAL_TASK_EXECUTE,
+        OPERATIONAL_TASK_EXTERNAL_SYNC,
+    }
+)
+
+SYSTEM_SERVICE_PERMISSIONS = frozenset(
+    {
+        OPERATIONAL_TASK_EXECUTE,
+        OPERATIONAL_TASK_EXTERNAL_SYNC,
     }
 )
 
@@ -71,6 +95,15 @@ def resolve_permissions(
     permissions = set(PROJECT_ROLE_PERMISSIONS.get(project_role or "", frozenset()))
     if is_superuser or system_role == "admin":
         permissions.update(SYSTEM_ADMIN_PERMISSIONS)
+        if project_role == "member":
+            permissions.update(
+                {
+                    OPERATIONAL_ACTION_COMMAND_APPROVE,
+                    OPERATIONAL_ACTION_COMMAND_EXECUTE,
+                }
+            )
+    elif system_role == "service" and project_role == "member":
+        permissions.update(SYSTEM_SERVICE_PERMISSIONS)
     return frozenset(permissions)
 
 

app/auth/project_dependencies.py

@@ -129,6 +129,7 @@ async def resolve_project_business_routing(
     return ActiveProjectRouting(
         project_code=ctx.project_code,
         business_dsn=business.dsn,
+        project_id=ctx.project_id,
     )
 
 
@@ -161,6 +162,7 @@ async def resolve_project_simulation_routing(
         project_code=ctx.project_code,
         business_dsn=business.dsn,
         timescale_dsn=timescale.dsn,
+        project_id=ctx.project_id,
     )
 
 

app/core/config.py

@@ -1,6 +1,7 @@
 from pathlib import Path
 from typing import Optional
 from urllib.parse import quote_plus
+from pydantic import AliasChoices, Field
 from pydantic_settings import BaseSettings, SettingsConfigDict
 
 
@@ -9,7 +10,6 @@ class Settings(BaseSettings):
     ENVIRONMENT: str = "production"
     API_V1_STR: str = "/api/v1"
     NETWORK_NAME: str = "default_network"
-
     # 敏感配置加密密钥 (Fernet)
     DATABASE_ENCRYPTION_KEY: str = ""
 
@@ -26,6 +26,9 @@ class Settings(BaseSettings):
     TIMESCALEDB_DB_PORT: str = "5433"
     TIMESCALEDB_DB_USER: str = "postgres"
     TIMESCALEDB_DB_PASSWORD: str = "password"
+    TIMESCALEDB_REALTIME_MAX_DECOMPRESSED_TUPLES: int = Field(
+        default=1_000_000, ge=1, le=2_147_483_647,
+    )
     # Metadata Database Config (PostgreSQL)
     METADATA_DB_NAME: str = "system_hub"
     METADATA_DB_HOST: str = "localhost"
@@ -63,16 +66,34 @@ class Settings(BaseSettings):
     KEYCLOAK_AUDIENCE: str = ""
     KEYCLOAK_ACCESS_TOKEN_MAX_AGE_SECONDS: int = 900
 
-    # Bocha Web Search API
-    BOCHA_API_KEY: str = ""
-    BOCHA_WEB_SEARCH_URL: str = "https://api.bochaai.com/v1/web-search"
-    BOCHA_WEB_SEARCH_TIMEOUT_SECONDS: float = 30.0
+    # Authenticated workbench synchronization.
+    WORKBENCH_EVENTS_ENABLED: bool = True
+    WORKBENCH_EVENT_RETENTION_DAYS: int = 7
+    CORS_ALLOWED_ORIGINS: str = (
+        "http://127.0.0.1:5173,http://localhost:5173,"
+        "http://127.0.0.1:15173,http://localhost:15173"
+    )
 
-    # Tianditu Geocoding API
-    TIANDITU_GEOCODER_TOKEN: str = ""
+    @property
+    def cors_allowed_origins(self) -> list[str]:
+        origins = [value.strip() for value in self.CORS_ALLOWED_ORIGINS.split(",") if value.strip()]
+        if "*" in origins:
+            raise ValueError("CORS_ALLOWED_ORIGINS must list explicit origins")
+        return origins
+
+    # One shared Tianditu key for geocoding, place search and other services.
+    TIANDITU_TOKEN: str = Field(
+        "",
+        validation_alias=AliasChoices(
+            "TIANDITU_TOKEN", "TJWATER_TIANDITU_TOKEN", "TIANDITU_GEOCODER_TOKEN"
+        ),
+    )
     TIANDITU_GEOCODER_URL: str = "https://api.tianditu.gov.cn/geocoder"
     TIANDITU_GEOCODER_TIMEOUT_SECONDS: float = 30.0
 
+    TIANDITU_SEARCH_URL: str = "https://api.tianditu.gov.cn/v2/search"
+    TIANDITU_SEARCH_TIMEOUT_SECONDS: float = 15.0
+
     @property
     def SQLALCHEMY_DATABASE_URI(self) -> str:
         db_password = quote_plus(self.DB_PASSWORD)

app/core/tianditu_logging.py

@@ -0,0 +1,17 @@
+import logging
+import re
+
+
+class TiandituTokenFilter(logging.Filter):
+    def filter(self, record: logging.LogRecord) -> bool:
+        message = record.getMessage()
+        if "tianditu.gov.cn" in message:
+            record.msg = re.sub(r"([?&]tk=)[^&\s\"']+", r"\1[redacted]", message)
+            record.args = ()
+        return True
+
+
+def install_tianditu_token_filter() -> None:
+    logger = logging.getLogger("httpx")
+    if not any(isinstance(item, TiandituTokenFilter) for item in logger.filters):
+        logger.addFilter(TiandituTokenFilter())

app/domain/domain_errors.py

@@ -0,0 +1,22 @@
+from __future__ import annotations
+
+
+class DomainError(Exception):
+    """A domain rejection with a stable API error payload."""
+
+    def __init__(self, code: str, message: str, *, refs: list | None = None,
+                 recovery: str = "none", status: int = 422):
+        super().__init__(message)
+        self.code, self.message, self.status = code, message, status
+        self.refs, self.recovery = refs or [], recovery
+
+    def payload(self) -> dict:
+        return {"code": self.code, "message": self.message,
+                "retryable": self.recovery == "retry_same_request",
+                "related_refs": self.refs, "recovery_action": self.recovery,
+                "retry_after_seconds": None}
+
+
+def require(condition: bool, code: str, message: str, **kwargs) -> None:
+    if not condition:
+        raise DomainError(code, message, **kwargs)

app/domain/schemas/admin_metadata.py

@@ -5,7 +5,7 @@ from uuid import UUID
 from pydantic import BaseModel, ConfigDict, Field, model_validator
 
 
-BusinessRole = Literal["admin", "user"]
+BusinessRole = Literal["admin", "service", "user"]
 ProjectRole = Literal["member", "viewer"]
 ProjectStatus = Literal["active", "inactive", "archived"]
 ProjectDbRole = Literal["biz_data", "iot_data"]

app/domain/schemas/agent_messages.py

@@ -0,0 +1,59 @@
+"""Private, durable messages between an ordinary chat and one investigation run."""
+from typing import Annotated, Literal
+from uuid import UUID
+
+from pydantic import Field
+
+from app.domain.schemas.shared import NonEmptyStr, StrictModel
+from app.domain.schemas.workbench_runtime import WorkbenchDateTime
+
+
+MessageText = Annotated[str, Field(min_length=1, max_length=12000, pattern=r"\S")]
+MessageStatus = Literal["pending", "enqueued", "delivered", "failed"]
+
+
+class AgentMessageCreate(StrictModel):
+    request_id: UUID
+    task_id: UUID
+    run_id: UUID
+    source_session_id: NonEmptyStr
+    text: MessageText
+
+
+class AgentMessageReply(StrictModel):
+    request_id: UUID
+    source_session_id: NonEmptyStr
+    text: MessageText
+
+
+class AgentMessageDelivery(StrictModel):
+    status: Literal["enqueued", "delivered", "failed"]
+    target_session_id: NonEmptyStr | None = None
+    native_message_id: NonEmptyStr | None = None
+    failure_detail: Annotated[str, Field(max_length=500)] | None = None
+
+
+class AgentMessage(StrictModel):
+    project_id: UUID
+    message_id: UUID
+    request_id: UUID
+    task_id: UUID
+    run_id: UUID
+    source_session_id: NonEmptyStr
+    source_owner_id: UUID
+    conversation_owner_id: UUID
+    target_session_id: NonEmptyStr | None = None
+    direction: Literal["to_task", "to_chat"]
+    in_reply_to: UUID | None = None
+    text: MessageText
+    status: MessageStatus
+    native_message_id: NonEmptyStr
+    failure_detail: str | None = None
+    created_at: WorkbenchDateTime
+    updated_at: WorkbenchDateTime
+    delivered_at: WorkbenchDateTime | None = None
+
+
+class AgentMessagePage(StrictModel):
+    items: list[AgentMessage]
+    next_cursor: str | None = None

app/domain/schemas/gis_attributes.py

@@ -0,0 +1,47 @@
+"""Schemas for project-scoped source GIS attributes."""
+
+from datetime import date
+from typing import Any, Literal
+
+from pydantic import BaseModel, ConfigDict, Field
+
+
+GisElementType = Literal[
+    "pipes",
+    "junctions",
+    "valves",
+    "reservoirs",
+    "pumps",
+    "tanks",
+    "scada",
+]
+
+
+class GisAttributeGroup(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+
+    dataset_key: str
+    dataset_title: str
+    source_date: date
+    layer_key: str
+    layer_name: str
+    source_fid: int
+    match_method: Literal["node_pair", "geometry", "source"]
+    match_distance_m: float | None = None
+    attributes: dict[str, Any] = Field(default_factory=dict)
+
+
+class ElementGisProperties(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+
+    element_type: GisElementType
+    element_id: str
+    groups: list[GisAttributeGroup] = Field(default_factory=list)
+
+
+class SourceGisFeatureProperties(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+
+    layer_key: str
+    source_fid: int
+    group: GisAttributeGroup

app/domain/schemas/operational_data.py

@@ -0,0 +1,83 @@
+"""Finite read scopes available to operational investigations."""
+from datetime import timedelta
+from typing import Literal
+
+from pydantic import Field, JsonValue, model_validator
+
+from app.domain.schemas.shared import NonEmptyStr, StrictModel
+from app.domain.schemas.workbench_runtime import WorkbenchDateTime
+
+
+TaskDataSource = Literal[
+    "workbench_runtime", "scada_devices", "scada_readings", "peripheral_points",
+    "peripheral_work_orders", "peripheral_observations", "peripheral_test_data",
+
+]
+
+
+McpSnapshotSource = Literal[
+    'run_simulation',
+    'run_contaminant_simulation',
+    'analyze_valve_closure',
+    'analyze_flushing',
+    'analyze_pressure_regulation',
+    'optimize_sensor_placement',
+    'run_burst_detection',
+    'run_burst_location',
+    'analyze_valve_isolation',
+    'run_leakage_identification',
+    'get_element_properties',
+    'list_elements',
+    'get_network_options',
+    'get_network_schema',
+    'list_scada_devices',
+    'get_scada_device',
+    'query_scada_readings',
+    'query_model_baseline_series',
+    'get_model_baseline_at_time',
+    'query_model_baseline_field',
+    'query_analysis_series',
+    'query_analysis_values',
+    'list_analysis_runs',
+    'get_analysis_run',
+    'list_analysis_results',
+    'list_sensor_placement_runs',
+    'get_sensor_placement_run',
+    'query_element_history',
+    'get_element_gis_properties',
+    'get_source_gis_properties',
+    'get_workbench_runtime',
+    'get_task_context',
+    'query_peripheral_data',
+    'get_task_evidence',
+]
+TaskSnapshotSource = TaskDataSource | McpSnapshotSource
+
+class TaskDataQuery(StrictModel):
+    source: TaskDataSource
+    device_ids: list[NonEmptyStr] = Field(default_factory=list, max_length=20)
+    point_id: NonEmptyStr | None = None
+    start: WorkbenchDateTime | None = None
+    end: WorkbenchDateTime | None = None
+
+    @model_validator(mode="after")
+    def validate_window(self):
+        if self.source in {"scada_readings", "peripheral_observations"}:
+            if self.start is None or self.end is None:
+                raise ValueError("time series queries require start and end")
+            if not timedelta(0) < self.end - self.start <= timedelta(days=1):
+                raise ValueError("query window must be positive and at most 24 hours")
+        if self.source == "scada_readings" and not self.device_ids:
+            raise ValueError("SCADA queries require device_ids")
+        if self.source == "peripheral_observations" and not self.point_id:
+            raise ValueError("peripheral queries require point_id")
+        return self
+
+
+class TaskDataSnapshot(StrictModel):
+    source: TaskSnapshotSource
+    source_ref: NonEmptyStr
+    queried_at: WorkbenchDateTime
+    origin: Literal["system_data", "synthetic"]
+    data: JsonValue
+    truncated: bool

app/domain/schemas/operational_task.py

@@ -0,0 +1,880 @@
+"""Contracts for project-scoped operational tasks and human-approved actions."""
+from __future__ import annotations
+
+import re
+from typing import Annotated, Literal
+from uuid import UUID
+
+from pydantic import Field, JsonValue, model_validator
+
+from app.domain.schemas.shared import NonEmptyStr, PositiveInt, StrictBool, StrictModel
+from app.domain.schemas.operational_data import TaskDataSnapshot, TaskDataSource
+from app.domain.schemas.workbench_runtime import (
+    WorkbenchDateTime,
+    WorkOrderActivity,
+)
+
+
+TaskType = Literal["scheduled_flow", "external_work_order", "event_investigation"]
+TaskStatus = Literal[
+    "queued", "running", "waiting_human", "completed", "failed", "cancelled"
+]
+TaskLifecycleStatus = Literal[
+    "open", "waiting_human", "completed", "failed", "cancelled"
+]
+OperationalTaskRunStatus = Literal[
+    "queued", "claimed", "running", "succeeded", "failed", "timed_out", "cancelled"
+]
+TaskPriority = Literal["normal", "urgent"]
+TaskSourceKind = Literal["schedule", "external_system", "operator", "agent"]
+TriggerKind = Literal["schedule", "external_sync", "event", "manual"]
+EvidenceKind = Literal["system_data", "external_feedback", "human_input", "synthetic"]
+PlanNodeStatus = Literal[
+    "pending", "waiting", "running", "completed", "failed", "skipped", "cancelled"
+]
+ActionKind = Literal["create_work_order", "issue_command"]
+ApprovalStatus = Literal["pending", "approved", "rejected"]
+ExecutionStatus = Literal[
+    "not_submitted", "submitted", "succeeded", "failed", "unsupported", "unknown"
+]
+
+
+class TaskSource(StrictModel):
+    kind: TaskSourceKind
+    external_id: NonEmptyStr | None = None
+    external_revision: PositiveInt | None = None
+    event_kind: NonEmptyStr | None = None
+    deleted_at: WorkbenchDateTime | None = None
+
+
+class TaskTrigger(StrictModel):
+    kind: TriggerKind
+    occurred_at: WorkbenchDateTime
+    detail: NonEmptyStr
+
+
+class OperationalTaskCreate(StrictModel):
+    request_id: UUID
+    task_type: TaskType
+    workflow: NonEmptyStr
+    title: NonEmptyStr
+    summary: NonEmptyStr
+    priority: TaskPriority = "normal"
+    source: TaskSource
+    trigger: TaskTrigger
+    scheduled_for: WorkbenchDateTime | None = None
+    agent_session_id: NonEmptyStr | None = None
+
+
+class ExecutionPlanNode(StrictModel):
+    id: NonEmptyStr
+    title: NonEmptyStr
+    owner: Literal["agent", "human"]
+    status: PlanNodeStatus
+    detail: NonEmptyStr
+    evidence_refs: list[NonEmptyStr] = Field(default_factory=list)
+    action_refs: list[UUID] = Field(default_factory=list)
+    kind: Literal["activity", "decision", "parallel_split", "parallel_join", "gate"] | None = None
+    selected_edge_id: NonEmptyStr | None = None
+    gate_ref: UUID | None = None
+    revisits_node_id: NonEmptyStr | None = None
+
+
+class ExecutionPlanEdge(StrictModel):
+    id: NonEmptyStr | None = None
+    from_node: NonEmptyStr = Field(alias="from")
+    to_node: NonEmptyStr = Field(alias="to")
+    condition: NonEmptyStr | None = None
+
+
+class ExecutionPlanInput(StrictModel):
+    schema_name: Literal["execution-plan@1", "execution-plan@2"] = Field(alias="schema")
+    title: NonEmptyStr
+    reason: NonEmptyStr
+    nodes: Annotated[list[ExecutionPlanNode], Field(min_length=1, max_length=100)]
+    edges: Annotated[list[ExecutionPlanEdge], Field(max_length=300)] = Field(
+        default_factory=list
+    )
+
+    @model_validator(mode="after")
+    def validate_graph(self):
+        node_ids = [node.id for node in self.nodes]
+        if len(set(node_ids)) != len(node_ids):
+            raise ValueError("execution plan node ids must be unique")
+        known = set(node_ids)
+        graph = {node_id: [] for node_id in node_ids}
+        indegree = {node_id: 0 for node_id in node_ids}
+        seen_edges: set[tuple[str, str]] = set()
+        edge_ids: set[str] = set()
+        for edge in self.edges:
+            pair = (edge.from_node, edge.to_node)
+            if edge.from_node not in known or edge.to_node not in known:
+                raise ValueError("execution plan edges must reference known nodes")
+            if edge.from_node == edge.to_node or pair in seen_edges:
+                raise ValueError("execution plan edges must be unique and non-reflexive")
+            seen_edges.add(pair)
+            if self.schema_name == "execution-plan@2":
+                if edge.id is None or edge.id in edge_ids:
+                    raise ValueError("v2 execution plan edge ids must be unique")
+                edge_ids.add(edge.id)
+            graph[edge.from_node].append(edge.to_node)
+            indegree[edge.to_node] += 1
+        pending = [node_id for node_id, value in indegree.items() if value == 0]
+        visited = 0
+        while pending:
+            current = pending.pop()
+            visited += 1
+            for target in graph[current]:
+                indegree[target] -= 1
+                if indegree[target] == 0:
+                    pending.append(target)
+        if visited != len(node_ids):
+            raise ValueError("execution plan must be acyclic")
+        if self.schema_name == "execution-plan@2" and sum(
+            not any(edge.to_node == node_id for edge in self.edges) for node_id in node_ids
+        ) != 1:
+            raise ValueError("v2 execution plan requires one start node")
+        if self.schema_name == "execution-plan@2":
+            by_id = {node.id: node for node in self.nodes}
+            for node in self.nodes:
+                if node.kind is None:
+                    raise ValueError("v2 execution plan nodes require kind")
+                outgoing = [edge for edge in self.edges if edge.from_node == node.id]
+                incoming = [edge for edge in self.edges if edge.to_node == node.id]
+                if node.kind == "decision":
+                    if len(outgoing) < 2 or any(not edge.condition for edge in outgoing):
+                        raise ValueError("decision requires at least two labeled routes")
+                    if node.selected_edge_id is not None and node.selected_edge_id not in {edge.id for edge in outgoing}:
+                        raise ValueError("decision selected_edge_id must reference an outgoing route")
+                    if node.status == "completed" and node.selected_edge_id is None:
+                        raise ValueError("completed decision requires a selected route")
+                    if node.selected_edge_id is not None:
+                        for edge in outgoing:
+                            target_status = by_id[edge.to_node].status
+                            if edge.id == node.selected_edge_id and target_status in {"skipped", "cancelled"}:
+                                raise ValueError("selected route cannot be skipped")
+                            if edge.id != node.selected_edge_id and target_status in {"running", "completed", "waiting"}:
+                                raise ValueError("unchosen route cannot be active")
+                elif node.selected_edge_id is not None:
+                    raise ValueError("only decision nodes may select routes")
+                if node.kind not in {"decision", "parallel_split"} and len(outgoing) > 1:
+                    raise ValueError("only decision or parallel split may branch")
+                if node.kind == "parallel_split" and len(outgoing) < 2:
+                    raise ValueError("parallel split requires at least two routes")
+                if node.kind == "parallel_join" and len(incoming) < 2:
+                    raise ValueError("parallel join requires at least two inputs")
+                if node.kind == "gate" and node.gate_ref is None:
+                    raise ValueError("gate requires a persisted reference")
+                if node.kind != "gate" and node.gate_ref is not None:
+                    raise ValueError("only gate nodes may have gate_ref")
+                if node.revisits_node_id is not None and node.revisits_node_id not in by_id:
+                    raise ValueError("revisited node must exist in the plan")
+        return self
+
+
+class EvidenceInput(StrictModel):
+    evidence_id: NonEmptyStr
+    kind: EvidenceKind
+    summary: NonEmptyStr
+    source_label: NonEmptyStr
+    source_ref: NonEmptyStr | None = None
+    observed_at: WorkbenchDateTime
+
+
+class EvidenceRecord(EvidenceInput):
+    recorded_at: WorkbenchDateTime
+    recorded_by: NonEmptyStr
+    has_snapshot: bool = False
+
+
+class StoredEvidenceRecord(EvidenceRecord):
+    data_snapshot: TaskDataSnapshot | None = None
+
+
+class InvestigationDataGap(StrictModel):
+    gap_id: NonEmptyStr
+    node_id: NonEmptyStr | None = None
+    missing_data: NonEmptyStr
+    scope: NonEmptyStr
+    impact: NonEmptyStr
+    required_data: NonEmptyStr
+    evidence_refs: list[NonEmptyStr] = Field(min_length=1)
+    blocking: StrictBool
+
+
+TaskMapLayer = Literal["junctions", "pipes", "valves", "pumps", "tanks", "reservoirs", "scada"]
+TaskFeatureId = Annotated[str, Field(min_length=1, max_length=128)]
+
+
+class InvestigationMapTarget(StrictModel):
+    source_id: TaskMapLayer
+    feature_id: TaskFeatureId
+
+
+class InvestigationDiagramNodeRef(StrictModel):
+    node_id: Annotated[str, Field(pattern=r"^[A-Za-z_][A-Za-z0-9_-]{0,79}$")]
+    label: NonEmptyStr
+    evidence_refs: list[NonEmptyStr] = Field(default_factory=list)
+    action_refs: list[UUID] = Field(default_factory=list)
+    status: Literal[
+        "pending", "running", "completed", "waiting_human", "failed", "skipped", "cancelled"
+    ] | None = None
+    map_targets: list[InvestigationMapTarget] = Field(default_factory=list, max_length=100)
+
+
+class InvestigationDiagram(StrictModel):
+    title: NonEmptyStr
+    mermaid: Annotated[str, Field(min_length=1, max_length=20000)]
+    node_refs: list[InvestigationDiagramNodeRef] = Field(default_factory=list, max_length=100)
+
+    @model_validator(mode="after")
+    def validate_description(self):
+        if not re.match(r"^\s*(?:flowchart|graph)\s+(?:TD|TB|BT|LR|RL)\b", self.mermaid):
+            raise ValueError("investigation diagrams must be Mermaid flowcharts")
+        if "```" in self.mermaid or "%%{" in self.mermaid or re.search(r"(?:^|[;\n])\s*click\b", self.mermaid):
+            raise ValueError("diagram descriptions cannot contain fences, directives or click commands")
+        ids = [node.node_id for node in self.node_refs]
+        if len(ids) != len(set(ids)):
+            raise ValueError("diagram node references must be unique")
+        return self
+
+
+class InvestigationReport(StrictModel):
+    """Read model: absent historical facts remain unknown, never inferred."""
+
+    title: NonEmptyStr
+    current_conclusion: NonEmptyStr
+    conclusion_boundary: NonEmptyStr
+    evidence_refs: list[NonEmptyStr] = Field(default_factory=list)
+    recommendations: list[NonEmptyStr] = Field(default_factory=list)
+    pending_items: list[NonEmptyStr] = Field(default_factory=list)
+    stage: Literal["interim", "final"] | None = None
+    outcome: Literal["concluded", "insufficient_data"] | None = None
+    stop_reason: NonEmptyStr | None = None
+    data_gaps: list[InvestigationDataGap] | None = None
+    diagram: InvestigationDiagram | None = None
+
+    @model_validator(mode="after")
+    def validate_result(self):
+        if self.outcome is None:
+            if self.stage is not None or self.stop_reason is not None or self.data_gaps is not None:
+                raise ValueError("structured report fields must be supplied together")
+            return self
+        if self.stage is None or self.data_gaps is None:
+            raise ValueError("structured reports require stage and data_gaps")
+        ids = [gap.gap_id for gap in self.data_gaps]
+        if len(ids) != len(set(ids)):
+            raise ValueError("data gap ids must be unique")
+        blocking = any(gap.blocking for gap in self.data_gaps)
+        if self.outcome == "insufficient_data":
+            if self.stage != "final" or not self.stop_reason or not blocking:
+                raise ValueError("insufficient data requires a final report, stop reason and blocking gap")
+            if self.recommendations or self.pending_items:
+                raise ValueError("insufficient data ends the investigation without downstream recommendations or pending work")
+        elif self.stop_reason is not None or blocking:
+            raise ValueError("concluded report cannot contain a stop reason or blocking gap")
+        return self
+
+
+class InvestigationReportInput(InvestigationReport):
+    """Write model: every new report explicitly declares its business result."""
+
+    stage: Literal["interim", "final"]
+    outcome: Literal["concluded", "insufficient_data"]
+    data_gaps: list[InvestigationDataGap]
+
+
+class InvestigationProgress(StrictModel):
+    """A complete reading snapshot, not a plan or a business result."""
+
+    summary: Annotated[str, Field(min_length=1, max_length=20000)]
+    diagram: InvestigationDiagram | None = None
+
+
+class InvestigationProgressSnapshot(StrictModel):
+    revision: PositiveInt
+    run_id: UUID
+    execution_generation: PositiveInt
+    published_at: WorkbenchDateTime
+    source_session_id: NonEmptyStr
+    progress: InvestigationProgress
+
+
+class ExecutionPlanSnapshot(StrictModel):
+    revision: PositiveInt
+    execution_generation: PositiveInt | None = None
+    source_session_id: NonEmptyStr
+    published_at: WorkbenchDateTime
+    published_by: NonEmptyStr
+    plan: ExecutionPlanInput
+
+
+class InvestigationReportRevision(StrictModel):
+    revision: PositiveInt
+    execution_generation: PositiveInt | None = None
+    source_session_id: NonEmptyStr
+    published_at: WorkbenchDateTime
+    published_by: NonEmptyStr
+    report: InvestigationReport
+
+
+class InvestigationReportPage(StrictModel):
+    items: list[InvestigationReportRevision]
+    next_offset: int | None = Field(default=None, ge=0)
+
+
+class InvestigationReportReference(StrictModel):
+    execution_generation: PositiveInt
+    revision: PositiveInt
+
+
+class AgentArtifactsCommand(StrictModel):
+    request_id: UUID
+    source_session_id: NonEmptyStr
+    run_id: UUID | None = None
+    evidence: list[EvidenceInput] = Field(default_factory=list, max_length=100)
+    plan: ExecutionPlanInput | None = None
+    report: InvestigationReportInput | None = None
+    progress: InvestigationProgress | None = None
+
+    @model_validator(mode="after")
+    def require_artifact(self):
+        if not self.evidence and self.plan is None and self.report is None and self.progress is None:
+            raise ValueError("at least one artifact must be supplied")
+        return self
+
+
+class TaskAccessRequestCreate(StrictModel):
+    request_id: UUID
+    source_session_id: NonEmptyStr
+    node_id: NonEmptyStr
+    scopes: Annotated[list[TaskDataSource], Field(min_length=1, max_length=7)]
+    reason: NonEmptyStr
+
+
+class TaskAccessRequestDecision(StrictModel):
+    request_id: UUID
+    expected_version: PositiveInt
+    decision: Literal["approved", "rejected"]
+    approved_scopes: list[TaskDataSource] = Field(default_factory=list)
+    comment: NonEmptyStr | None = None
+
+
+class TaskAccessRequest(StrictModel):
+    access_request_id: UUID
+    task_id: UUID
+    node_id: NonEmptyStr
+    scopes: list[TaskDataSource]
+    reason: NonEmptyStr
+    status: Literal["pending", "approved", "rejected"]
+    approved_scopes: list[TaskDataSource]
+    version: PositiveInt
+    requested_by: NonEmptyStr
+    requested_at: WorkbenchDateTime
+    decided_by: NonEmptyStr | None = None
+    decided_at: WorkbenchDateTime | None = None
+    comment: NonEmptyStr | None = None
+
+
+class CreateWorkOrderPayload(StrictModel):
+    title: NonEmptyStr
+    description: NonEmptyStr
+    work_order_type: NonEmptyStr
+    priority: TaskPriority = "normal"
+    scheduled_at: WorkbenchDateTime | None = None
+    location: NonEmptyStr | None = None
+    assignee: NonEmptyStr | None = None
+
+
+class CommandTarget(StrictModel):
+    kind: NonEmptyStr
+    id: NonEmptyStr
+
+
+class IssueCommandPayload(StrictModel):
+    command_type: NonEmptyStr
+    target: CommandTarget
+    parameters: dict[str, JsonValue]
+    impact_summary: NonEmptyStr
+    preconditions: list[NonEmptyStr] = Field(min_length=1)
+    expires_at: WorkbenchDateTime
+
+
+ActionPayload = CreateWorkOrderPayload | IssueCommandPayload
+
+
+class ActionProposalCommand(StrictModel):
+    request_id: UUID
+    kind: ActionKind
+    reason: NonEmptyStr
+    payload: ActionPayload
+    evidence_refs: list[NonEmptyStr] = Field(default_factory=list)
+    node_id: NonEmptyStr | None = None
+
+    @model_validator(mode="after")
+    def validate_payload_kind(self):
+        expected = CreateWorkOrderPayload if self.kind == "create_work_order" else IssueCommandPayload
+        if not isinstance(self.payload, expected):
+            raise ValueError("action payload does not match its kind")
+        return self
+
+
+class ActionDecisionCommand(StrictModel):
+    request_id: UUID
+    expected_version: PositiveInt
+    decision: Literal["approved", "rejected"]
+    comment: NonEmptyStr | None = None
+
+
+class WorkOrderDraftUpdateCommand(StrictModel):
+    request_id: UUID
+    expected_version: PositiveInt
+    payload: CreateWorkOrderPayload
+
+
+class WorkOrderConfirmDispatchCommand(StrictModel):
+    request_id: UUID
+    expected_version: PositiveInt
+
+
+class ActionExecuteCommand(StrictModel):
+    request_id: UUID
+    expected_version: PositiveInt
+
+
+class ActionReconcileCommand(StrictModel):
+    request_id: UUID
+    expected_version: PositiveInt
+
+
+class ActionReceipt(StrictModel):
+    adapter: NonEmptyStr
+    receipt_id: NonEmptyStr
+    submitted_at: WorkbenchDateTime
+    completed_at: WorkbenchDateTime | None = None
+    external_status: NonEmptyStr
+    detail: NonEmptyStr
+
+
+class HumanApprovalAction(StrictModel):
+    action_id: UUID
+    task_id: UUID
+    version: PositiveInt
+    kind: ActionKind
+    reason: NonEmptyStr
+    payload: ActionPayload
+    evidence_refs: list[NonEmptyStr]
+    node_id: NonEmptyStr | None = None
+    required_approval_permission: NonEmptyStr
+    required_execution_permission: NonEmptyStr
+    approval_status: ApprovalStatus
+    execution_status: ExecutionStatus
+    execution_request_id: UUID | None = None
+    proposed_by: NonEmptyStr
+    proposed_at: WorkbenchDateTime
+    edited_by: NonEmptyStr | None = None
+    edited_at: WorkbenchDateTime | None = None
+    decided_by: NonEmptyStr | None = None
+    decided_at: WorkbenchDateTime | None = None
+    decision_comment: NonEmptyStr | None = None
+    executed_by: NonEmptyStr | None = None
+    executed_at: WorkbenchDateTime | None = None
+    receipt: ActionReceipt | None = None
+    execution_detail: NonEmptyStr | None = None
+
+
+class ExternalWorkOrderSnapshot(StrictModel):
+    work_order_id: UUID
+    revision: PositiveInt
+    title: NonEmptyStr
+    summary: NonEmptyStr
+    status: Literal[
+        "draft", "open", "in_progress", "awaiting_reply", "completed", "cancelled"
+    ]
+    priority: Literal["low", "normal", "high", "urgent"]
+    assignee: NonEmptyStr | None = None
+    scheduled_at: WorkbenchDateTime | None = None
+    work_order_type: Literal["general", "water_quality_complaint", "burst_report"] = "general"
+    location: NonEmptyStr | None = None
+    scenario_code: NonEmptyStr | None = None
+    feedback_requirements: list[NonEmptyStr] = Field(default_factory=list)
+    activity: list[WorkOrderActivity] = Field(default_factory=list)
+    origin: Literal["synthetic"]
+    updated_at: WorkbenchDateTime
+
+    @model_validator(mode="before")
+    @classmethod
+    def accept_legacy_snapshot(cls, value):
+        if isinstance(value, dict):
+            return {key: item for key, item in value.items() if key != "subtasks"}
+        return value
+
+
+class OperationalTaskRun(StrictModel):
+    run_id: UUID
+    task_id: UUID
+    generation: PositiveInt
+    execution_request_id: UUID
+    status: OperationalTaskRunStatus
+    revision: PositiveInt
+    agent_session_id: NonEmptyStr | None = None
+    lease_expires_at: WorkbenchDateTime | None = None
+    heartbeat_at: WorkbenchDateTime | None = None
+    started_at: WorkbenchDateTime | None = None
+    finished_at: WorkbenchDateTime | None = None
+    failure_code: NonEmptyStr | None = None
+    failure_detail: NonEmptyStr | None = None
+    dispatch_attempts: int = Field(default=0, ge=0)
+    next_dispatch_at: WorkbenchDateTime
+    created_at: WorkbenchDateTime
+    updated_at: WorkbenchDateTime
+
+
+class OperationalTaskSummary(StrictModel):
+    project_id: UUID
+    task_id: UUID
+    task_type: TaskType
+    workflow: NonEmptyStr
+    lifecycle_status: TaskLifecycleStatus = "open"
+    active_run: OperationalTaskRun | None = None
+    latest_run: OperationalTaskRun | None = None
+    # Compatibility fields for the staged 014 -> 015 rollout. New consumers
+    # must use lifecycle_status and active_run.
+    status: TaskStatus
+    priority: TaskPriority
+    revision: PositiveInt
+    execution_generation: PositiveInt
+    execution_request_id: UUID
+    title: NonEmptyStr
+    summary: NonEmptyStr
+    source: TaskSource
+    trigger: TaskTrigger
+    scheduled_for: WorkbenchDateTime | None
+    agent_session_id: NonEmptyStr | None
+    agent_lease_expires_at: WorkbenchDateTime | None
+    current_plan_revision: PositiveInt | None
+    current_report_revision: PositiveInt | None
+    created_at: WorkbenchDateTime
+    updated_at: WorkbenchDateTime
+
+
+class OperationalTaskPermissions(StrictModel):
+    can_publish_artifacts: bool
+    can_propose_actions: bool
+    can_sync_external_feedback: bool
+    can_decide_work_orders: bool
+    can_execute_work_orders: bool
+    can_decide_commands: bool
+    can_execute_commands: bool
+    can_decide_access: bool = False
+
+
+class OperationalTaskDetail(OperationalTaskSummary):
+    latest_plan: ExecutionPlanSnapshot | None
+    latest_report: InvestigationReportRevision | None
+    latest_progress: InvestigationProgressSnapshot | None = None
+    evidence: list[EvidenceRecord]
+    actions: list[HumanApprovalAction]
+    access_requests: list[TaskAccessRequest] = Field(default_factory=list)
+    permissions: OperationalTaskPermissions
+    work_order_snapshot: ExternalWorkOrderSnapshot | None = None
+
+
+class OperationalTaskPage(StrictModel):
+    items: list[OperationalTaskSummary]
+    limit: PositiveInt
+    offset: int = Field(ge=0)
+    next_offset: int | None = Field(default=None, ge=0)
+
+
+class OperationalTaskRunPage(StrictModel):
+    items: list[OperationalTaskRun]
+    next_cursor: NonEmptyStr | None = None
+
+
+class AgentTaskContext(StrictModel):
+    task: OperationalTaskDetail
+    approved_data_scopes: list[NonEmptyStr]
+    selected_report: InvestigationReportRevision | None = None
+
+
+class AgentRunCommand(StrictModel):
+    request_id: UUID
+    expected_revision: PositiveInt
+    source_session_id: NonEmptyStr
+
+
+class AgentRunFailureCommand(AgentRunCommand):
+    failure_code: NonEmptyStr
+    failure_detail: NonEmptyStr
+
+
+class AgentRunCancelCommand(StrictModel):
+    request_id: UUID
+    expected_revision: PositiveInt
+    reason: NonEmptyStr
+
+
+class AgentRunDispatchFailureCommand(StrictModel):
+    request_id: UUID
+    expected_revision: PositiveInt
+    failure_code: NonEmptyStr
+    failure_detail: NonEmptyStr
+    retryable: bool = True
+
+
+class OperationalTaskRetryCommand(StrictModel):
+    request_id: UUID
+    reason: NonEmptyStr
+
+
+class RecoverExpiredLeasesCommand(StrictModel):
+    limit: int = Field(default=20, ge=1, le=100)
+    task_type: TaskType | None = None
+    task_created_from: WorkbenchDateTime | None = None
+
+
+class RecoverExpiredLeasesResult(StrictModel):
+    task_ids: list[UUID]
+    recovered_at: WorkbenchDateTime
+
+
+class AgentArtifactsResult(StrictModel):
+    task: OperationalTaskDetail
+    plan_revision: PositiveInt | None
+    report_revision: PositiveInt | None
+    progress_revision: PositiveInt | None = None
+
+
+class TaskSafeChartSeries(StrictModel):
+    name: Annotated[str, Field(max_length=128)]
+    data: Annotated[list[Annotated[float, Field(strict=True, allow_inf_nan=False)]], Field(min_length=1, max_length=100)]
+
+
+class TaskSafeChart(StrictModel):
+    title: Annotated[str, Field(max_length=200)]
+    chart_type: Literal["line", "bar"]
+    x_data: Annotated[list[Annotated[str, Field(max_length=128)]], Field(min_length=1, max_length=100)]
+    series: Annotated[list[TaskSafeChartSeries], Field(min_length=1, max_length=8)]
+
+    @model_validator(mode="after")
+    def validate_alignment(self):
+        if any(len(series.data) != len(self.x_data) for series in self.series):
+            raise ValueError("chart series must align with x_data")
+        return self
+
+
+class TaskHighlightFeatures(StrictModel):
+    action: Literal["highlight_features"]
+    layer: TaskMapLayer
+    feature_ids: Annotated[list[TaskFeatureId], Field(min_length=1, max_length=100)]
+    effect: Literal["steady", "pulse"] = "steady"
+
+
+class TaskClearHighlights(StrictModel):
+    action: Literal["clear_highlights"]
+
+
+class TaskOpenChart(StrictModel):
+    action: Literal["open_chart"]
+    chart_id: TaskFeatureId
+    chart: TaskSafeChart
+    evidence_refs: Annotated[list[NonEmptyStr], Field(min_length=1, max_length=100)]
+
+
+class TaskCloseChart(StrictModel):
+    action: Literal["close_chart"]
+    chart_id: TaskFeatureId
+
+
+class TaskSetOverlay(StrictModel):
+    action: Literal["set_overlay"]
+    overlay_id: TaskFeatureId
+    title: Annotated[str, Field(min_length=1, max_length=200)]
+    mode: Literal["geometry", "heatmap"]
+    evidence_id: NonEmptyStr
+
+
+class TaskClearOverlay(StrictModel):
+    action: Literal["clear_overlay"]
+    overlay_id: TaskFeatureId
+
+
+TaskPresentationCommand = Annotated[
+    TaskHighlightFeatures | TaskClearHighlights | TaskOpenChart | TaskCloseChart | TaskSetOverlay | TaskClearOverlay,
+    Field(discriminator="action"),
+]
+
+
+class TaskPresentationEventCommand(StrictModel):
+    request_id: UUID
+    run_id: UUID
+    source_session_id: NonEmptyStr
+    command: TaskPresentationCommand
+
+
+class TaskPresentationEventResult(StrictModel):
+    recorded: Literal[True] = True
+    event_seq: PositiveInt
+    run_id: UUID
+    execution_generation: PositiveInt
+
+
+class TaskPresentationActiveRun(StrictModel):
+    run_id: UUID
+    generation: PositiveInt
+    status: OperationalTaskRunStatus
+
+
+class TaskPresentationCursor(StrictModel):
+    task_id: UUID
+    execution_generation: PositiveInt
+    active_run: TaskPresentationActiveRun | None
+    event_cursor: int = Field(ge=0)
+
+
+class TaskDataQueryResult(StrictModel):
+    evidence: EvidenceRecord
+    snapshot: TaskDataSnapshot
+
+
+class ExternalFeedbackEvidence(StrictModel):
+    evidence_id: NonEmptyStr
+    kind: Literal["external_feedback", "synthetic"] = "external_feedback"
+    summary: NonEmptyStr
+    source_label: NonEmptyStr
+    source_ref: NonEmptyStr | None = None
+    observed_at: WorkbenchDateTime
+
+
+class ExternalFeedbackCommand(StrictModel):
+    request_id: UUID
+    source_event_id: NonEmptyStr
+    source_revision: PositiveInt
+    status: Literal[
+        "scheduled", "assigned", "in_progress", "awaiting_feedback",
+        "completed", "cancelled", "failed"
+    ]
+    occurred_at: WorkbenchDateTime
+    detail: NonEmptyStr
+    evidence: ExternalFeedbackEvidence | None = None
+    work_order_snapshot: ExternalWorkOrderSnapshot | None = None
+
+    @model_validator(mode="after")
+    def snapshot_matches_feedback_revision(self):
+        if self.work_order_snapshot is not None:
+            work_order_id = str(self.work_order_snapshot.work_order_id)
+            if self.source_event_id not in {
+                work_order_id, f"peripheral_sample:{work_order_id}"
+            }:
+                raise ValueError("work_order_snapshot id must match source_event_id")
+            if self.work_order_snapshot.revision != self.source_revision:
+                raise ValueError("work_order_snapshot revision must match source_revision")
+            expected_status = {
+                "draft": "scheduled",
+                "open": "assigned",
+                "in_progress": "in_progress",
+                "awaiting_reply": "awaiting_feedback",
+                "completed": "completed",
+                "cancelled": "cancelled",
+            }[self.work_order_snapshot.status]
+            if self.status != expected_status:
+                raise ValueError("work_order_snapshot status must match feedback status")
+        return self
+
+
+class ExternalFeedbackResult(StrictModel):
+    task: OperationalTaskDetail
+    source_event_id: NonEmptyStr
+    source_revision: PositiveInt
+    accepted_at: WorkbenchDateTime
+
+
+class ExternalSourceWithdrawalCommand(StrictModel):
+    request_id: UUID
+    source_event_id: NonEmptyStr
+
+
+class HumanEvidenceCommand(StrictModel):
+    request_id: UUID
+    evidence: EvidenceInput
+
+
+class OperationalTaskNotification(StrictModel):
+    notification_id: UUID
+    task_id: UUID
+    kind: Literal[
+        "task_created", "action_pending", "report_updated", "task_completed",
+        "task_failed", "run_delivery_degraded", "access_approval_pending",
+        "action_approval_pending", "human_input_required",
+        "external_coordination_failed"
+    ]
+    title: NonEmptyStr
+    description: NonEmptyStr
+    required_permission: NonEmptyStr | None = None
+    lifecycle_status: Literal["active", "resolved"] = "active"
+    subject_type: Literal["task", "run", "access_request", "action"] = "task"
+    subject_id: UUID | None = None
+    report_reference: InvestigationReportReference | None = None
+    resolved_at: WorkbenchDateTime | None = None
+    resolution_reason: NonEmptyStr | None = None
+    status: Literal["unread", "read"]
+    version: PositiveInt
+    created_at: WorkbenchDateTime
+    read_at: WorkbenchDateTime | None = None
+    read_by: NonEmptyStr | None = None
+
+
+class OperationalTaskNotificationPage(StrictModel):
+    items: list[OperationalTaskNotification]
+    limit: PositiveInt
+    offset: int = Field(ge=0)
+    next_offset: int | None = Field(default=None, ge=0)
+
+
+class NotificationAcknowledgeCommand(StrictModel):
+    request_id: UUID
+    expected_version: PositiveInt
+
+
+class OperationalTaskScheduleCreate(StrictModel):
+    request_id: UUID
+    schedule_key: NonEmptyStr
+    title: NonEmptyStr
+    summary: NonEmptyStr
+    interval_seconds: int = Field(ge=60, le=31_536_000)
+    next_run_at: WorkbenchDateTime
+    priority: TaskPriority = "normal"
+
+
+class OperationalTaskSchedule(StrictModel):
+    schedule_id: UUID
+    schedule_key: NonEmptyStr
+    title: NonEmptyStr
+    summary: NonEmptyStr
+    interval_seconds: int = Field(ge=60, le=31_536_000)
+    next_run_at: WorkbenchDateTime
+    priority: TaskPriority
+    enabled: bool
+    version: PositiveInt
+    created_at: WorkbenchDateTime
+    updated_at: WorkbenchDateTime
+
+
+class OperationalTaskSchedulePage(StrictModel):
+    items: list[OperationalTaskSchedule]
+
+
+class OperationalTaskScheduleUpdate(StrictModel):
+    request_id: UUID
+    expected_version: PositiveInt
+    enabled: bool
+
+
+class DispatchDueSchedulesCommand(StrictModel):
+    limit: int = Field(default=20, ge=1, le=100)
+
+
+class DispatchDueSchedulesResult(StrictModel):
+    task_ids: list[UUID]
+    dispatched_at: WorkbenchDateTime

app/domain/schemas/place_search.py

@@ -0,0 +1,28 @@
+from typing import Literal
+
+from pydantic import BaseModel, ConfigDict, Field
+
+MAX_PLACE_SEARCH_RESULTS = 10
+
+
+class PlaceSearchRequest(BaseModel):
+    model_config = ConfigDict(str_strip_whitespace=True, extra="forbid")
+
+    keyword: str = Field(..., min_length=1, max_length=120)
+
+
+class PlaceSearchItem(BaseModel):
+    id: str = Field(..., min_length=1)
+    name: str = Field(..., min_length=1)
+    address: str
+    longitude: float = Field(..., ge=-180, le=180, allow_inf_nan=False)
+    latitude: float = Field(..., ge=-90, le=90, allow_inf_nan=False)
+    category: str | None = None
+    district: Literal["奉贤区"] = "奉贤区"
+
+
+class PlaceSearchResponse(BaseModel):
+    items: list[PlaceSearchItem] = Field(..., max_length=MAX_PLACE_SEARCH_RESULTS)
+    total: int = Field(..., ge=0, le=MAX_PLACE_SEARCH_RESULTS)
+    truncated: bool = False
+    region: Literal["奉贤区"] = "奉贤区"

app/domain/schemas/sensor_placement.py

@@ -25,7 +25,7 @@ class SensorPlacementOptimizeRequest(BaseModel):
         pattern=r"^[^/\\\x00]+$",
     )
     run_name: str = Field(..., min_length=1, max_length=64)
-    sensor_type: Literal["pressure"]
+    sensor_type: Literal["pressure", "water_quality"]
     method: Literal["sensitivity", "kmeans"]
     sensor_count: int = Field(..., gt=0, le=200)
     min_diameter: int = Field(default=0, ge=0)
@@ -89,6 +89,8 @@ class SensorPlacementSchemeResponse(BaseModel):
     created_by: str
     created_at: datetime
     status: str
+    sensor_type: Literal["pressure", "water_quality"] = "pressure"
+    method: Literal["sensitivity", "kmeans"] | None = None
     sensor_locations: list[str]
     sensor_points: list[SensorPointResponse]
     can_edit: bool = False

app/domain/schemas/shared.py

@@ -0,0 +1,58 @@
+"""Small strict value types shared by independent backend features."""
+
+from datetime import datetime
+from typing import Annotated, Literal
+from uuid import UUID
+
+from pydantic import BaseModel, BeforeValidator, ConfigDict, Field, StringConstraints, model_validator
+
+NonEmptyStr = Annotated[str, StringConstraints(strip_whitespace=True, min_length=1)]
+StrictBool = Annotated[bool, Field(strict=True)]
+PositiveInt = Annotated[int, Field(strict=True, ge=1)]
+NonNegativeInt = Annotated[int, Field(strict=True, ge=0)]
+
+
+def _datetime_wire_value(value: object) -> datetime | str:
+    if not isinstance(value, (datetime, str)):
+        raise ValueError("datetime input must be an RFC3339 string or datetime instance")
+    return value
+
+
+WireDateTime = Annotated[datetime, BeforeValidator(_datetime_wire_value)]
+
+
+class StrictModel(BaseModel):
+    model_config = ConfigDict(extra="forbid", allow_inf_nan=False)
+
+
+ObjectType = Literal[
+    "analysis_task",
+    "goal",
+    "state_snapshot",
+    "action_contract",
+    "validation",
+    "run",
+    "outcome",
+    "evidence",
+    "artifact",
+    "operation",
+]
+
+
+class ObjectRef(StrictModel):
+    object_type: ObjectType
+    object_id: UUID
+    revision: PositiveInt
+
+
+class TimeRange(StrictModel):
+    start: WireDateTime
+    end: WireDateTime
+
+    @model_validator(mode="after")
+    def validate_range(self) -> "TimeRange":
+        if self.start.tzinfo is None or self.end.tzinfo is None:
+            raise ValueError("timestamps must include a timezone")
+        if self.start >= self.end:
+            raise ValueError("start must be before end")
+        return self

app/domain/schemas/task_alert.py

@@ -0,0 +1,38 @@
+"""Idempotent alert events emitted by scheduled and other backend tasks."""
+from __future__ import annotations
+
+from typing import Literal
+from uuid import UUID
+
+from app.domain.schemas.shared import NonEmptyStr, PositiveInt, StrictModel
+from app.domain.schemas.workbench_runtime import WorkbenchAlertSource, WorkbenchDateTime
+
+
+class TaskAlertCommand(StrictModel):
+    event_id: UUID
+    alert_id: NonEmptyStr
+    event_type: Literal["raised", "acknowledged", "resolved"]
+    occurred_at: WorkbenchDateTime
+    condition_id: NonEmptyStr | None
+    source: WorkbenchAlertSource
+    title: NonEmptyStr
+    description: NonEmptyStr
+    severity: Literal["warning", "critical"]
+
+
+class TaskAlertEvent(TaskAlertCommand):
+    producer_id: NonEmptyStr
+
+
+class TaskAlertState(StrictModel):
+    alert_id: NonEmptyStr
+    revision: PositiveInt
+    producer_id: NonEmptyStr
+    condition_id: NonEmptyStr | None
+    source: WorkbenchAlertSource
+    title: NonEmptyStr
+    description: NonEmptyStr
+    occurred_at: WorkbenchDateTime
+    severity: Literal["warning", "critical"]
+    status: Literal["active", "acknowledged", "resolved"]
+    updated_at: WorkbenchDateTime

app/domain/schemas/timeseries_history.py

@@ -86,3 +86,4 @@ class ElementHistorySeries(BaseModel):
 
 class ElementHistoryResponse(BaseModel):
     series: list[ElementHistorySeries]
+    truncated: bool = Field(default=False, exclude=True)

app/domain/schemas/workbench_events.py

@@ -0,0 +1,51 @@
+"""Stable workbench event-stream and synchronization contracts."""
+from __future__ import annotations
+
+from typing import Literal
+from uuid import UUID
+
+from pydantic import Field, JsonValue
+
+from app.domain.schemas.operational_task import (
+    OperationalTaskNotification,
+    OperationalTaskSummary,
+)
+from app.domain.schemas.shared import NonEmptyStr, PositiveInt, StrictModel
+from app.domain.schemas.workbench_runtime import WorkbenchDateTime, WorkbenchRuntimeSnapshot
+
+
+WorkbenchEventTopic = Literal[
+    "operational.task.changed",
+    "operational.run.queued",
+    "operational.run.changed",
+    "operational.report.published",
+    "operational.progress.changed",
+    "operational.presentation.requested",
+    "operational.action.changed",
+    "operational.notification.changed",
+    "workbench.alert.changed",
+    "workbench.work_order.changed",
+    "analysis.run.changed",
+]
+
+
+class WorkbenchEvent(StrictModel):
+    schema_name: Literal["workbench-event@1"] = Field(alias="schema")
+    event_seq: PositiveInt
+    event_id: UUID
+    project_id: UUID
+    topic: WorkbenchEventTopic
+    aggregate_type: NonEmptyStr
+    aggregate_id: NonEmptyStr
+    aggregate_revision: PositiveInt
+    occurred_at: WorkbenchDateTime
+    metadata: dict[str, JsonValue]
+
+
+class WorkbenchSyncSnapshot(StrictModel):
+    schema_name: Literal["workbench-sync-snapshot@1"] = Field(alias="schema")
+    event_cursor: int = Field(ge=0)
+    runtime: WorkbenchRuntimeSnapshot
+    tasks: list[OperationalTaskSummary]
+    notifications: list[OperationalTaskNotification]
+    generated_at: WorkbenchDateTime

app/domain/schemas/workbench_runtime.py

@@ -0,0 +1,192 @@
+"""Typed read-only projections consumed by the supply workbench."""
+
+from typing import Annotated, Literal
+
+from pydantic import AwareDatetime, BeforeValidator, Field, JsonValue, model_validator
+
+from app.domain.schemas.shared import (
+    NonEmptyStr,
+    NonNegativeInt,
+    PositiveInt,
+    StrictBool,
+    StrictModel,
+)
+
+
+def _workbench_datetime_wire_value(value: object):
+    from datetime import datetime
+
+    if not isinstance(value, (datetime, str)):
+        raise ValueError("datetime input must be an RFC3339 string or datetime instance")
+    return value
+
+
+WorkbenchDateTime = Annotated[
+    AwareDatetime, BeforeValidator(_workbench_datetime_wire_value)
+]
+ProjectionAvailability = Literal["ready", "not_configured", "degraded"]
+
+
+class WorkbenchRuntimeCondition(StrictModel):
+    id: NonEmptyStr
+    task_id: Literal["backend-schedule"]
+    session_id: NonEmptyStr
+    title: NonEmptyStr
+    summary: NonEmptyStr
+    status: Literal["pending", "running", "completed", "warning", "error"]
+    risk_level: Literal["normal", "attention", "critical"]
+    scheduled_at: WorkbenchDateTime
+    updated_at: WorkbenchDateTime
+    detail: NonEmptyStr
+    recommendation: NonEmptyStr | None
+    evidence: list[NonEmptyStr]
+    model_name: Literal["TJWater scheduler"]
+
+
+class WorkbenchAlertSource(StrictModel):
+    task_type: Literal["scheduled_task", "analysis_task", "other_task"]
+    task_id: NonEmptyStr
+    run_id: NonEmptyStr
+
+
+class WorkbenchRuntimeAlert(StrictModel):
+    id: NonEmptyStr
+    condition_id: NonEmptyStr | None
+    source: WorkbenchAlertSource
+    title: NonEmptyStr
+    description: NonEmptyStr
+    occurred_at: WorkbenchDateTime
+    severity: Literal["warning", "critical"]
+
+
+class WorkbenchPersistedAlert(StrictModel):
+    alert_id: NonEmptyStr
+    condition_id: NonEmptyStr | None
+    source: WorkbenchAlertSource
+    title: NonEmptyStr
+    description: NonEmptyStr
+    occurred_at: WorkbenchDateTime
+    severity: Literal["warning", "critical"]
+    status: Literal["active", "acknowledged", "resolved"]
+    producer_id: NonEmptyStr | None = None
+    updated_at: WorkbenchDateTime | None = None
+
+
+class WorkbenchRuntimeScenario(StrictModel):
+    id: NonEmptyStr
+    run_id: NonEmptyStr
+    name: NonEmptyStr
+    description: NonEmptyStr
+    status: Literal["draft", "active", "review"]
+    run_type: NonEmptyStr
+    run_type_label: NonEmptyStr
+    run_status: NonEmptyStr
+    created_at: WorkbenchDateTime
+
+
+class WorkbenchRuntimeAnalysisFact(StrictModel):
+    label: NonEmptyStr
+    value: NonEmptyStr
+
+
+class WorkbenchRuntimeAnalysisResult(StrictModel):
+    result_type: NonEmptyStr
+    created_at: WorkbenchDateTime
+    facts: list[WorkbenchRuntimeAnalysisFact]
+
+
+class WorkbenchRuntimeLatestAnalysis(StrictModel):
+    run_id: NonEmptyStr
+    name: NonEmptyStr
+    run_type: NonEmptyStr
+    run_type_label: NonEmptyStr
+    status: NonEmptyStr
+    status_label: NonEmptyStr
+    created_by: NonEmptyStr
+    created_at: WorkbenchDateTime
+    started_at: WorkbenchDateTime
+    results: list[WorkbenchRuntimeAnalysisResult]
+    map_payload: dict[str, JsonValue] | None
+
+
+class WorkbenchAnalysisHistoryItem(StrictModel):
+    run_id: NonEmptyStr
+    name: NonEmptyStr
+    run_type: NonEmptyStr
+    run_type_label: NonEmptyStr
+    status: NonEmptyStr
+    status_label: NonEmptyStr
+    created_by: NonEmptyStr
+    created_at: WorkbenchDateTime
+    started_at: WorkbenchDateTime
+    has_results: StrictBool
+
+
+class WorkbenchAnalysisHistoryPage(StrictModel):
+    items: list[WorkbenchAnalysisHistoryItem]
+    limit: PositiveInt
+    offset: NonNegativeInt
+    next_offset: NonNegativeInt | None
+
+
+class WorkOrderActivity(StrictModel):
+    sequence: PositiveInt
+    occurred_at: WorkbenchDateTime
+    actor: NonEmptyStr
+    kind: Literal["created", "updated", "feedback"]
+    title: NonEmptyStr
+    detail: NonEmptyStr
+
+
+class WorkbenchRuntimeWorkOrder(StrictModel):
+    id: NonEmptyStr
+    code: NonEmptyStr
+    scheduled_at: WorkbenchDateTime
+    title: NonEmptyStr
+    summary: NonEmptyStr
+    status: Literal["pending", "running", "awaiting_reply", "completed"]
+    risk_level: Literal["normal", "attention", "critical"]
+    updated_at: WorkbenchDateTime
+    detail: NonEmptyStr
+    recommendation: NonEmptyStr | None
+    duration_minutes: PositiveInt
+    source: NonEmptyStr
+    dispatcher: NonEmptyStr
+    location: NonEmptyStr | None = None
+    assignee: NonEmptyStr | None = None
+    priority: Literal["normal", "urgent"]
+    feedback_requirements: list[NonEmptyStr] = Field(default_factory=list)
+    activity: list[WorkOrderActivity] = Field(default_factory=list)
+
+    @model_validator(mode="before")
+    @classmethod
+    def accept_legacy_document(cls, value):
+        if not isinstance(value, dict) or "subtasks" not in value:
+            return value
+        document = {key: item for key, item in value.items() if key != "subtasks"}
+        subtasks = value.get("subtasks") or []
+        first = subtasks[0] if subtasks else {}
+        document.setdefault("location", first.get("location"))
+        document.setdefault("assignee", first.get("assignee"))
+        document.setdefault("feedback_requirements", list(dict.fromkeys(
+            requirement for subtask in subtasks
+            for requirement in subtask.get("reply_requirements", [])
+        )))
+        document.setdefault("activity", [])
+        return document
+
+
+class WorkbenchRuntimeSnapshot(StrictModel):
+    availability: ProjectionAvailability
+    source: Literal["none", "scheduled_tasks"]
+    scenario_availability: ProjectionAvailability
+    artifact_availability: ProjectionAvailability
+    schedule_availability: ProjectionAvailability
+    work_order_availability: ProjectionAvailability
+    alert_availability: ProjectionAvailability
+    generated_at: WorkbenchDateTime
+    conditions: list[WorkbenchRuntimeCondition]
+    alerts: list[WorkbenchRuntimeAlert]
+    scenarios: list[WorkbenchRuntimeScenario]
+    latest_analysis: WorkbenchRuntimeLatestAnalysis | None
+    work_orders: list[WorkbenchRuntimeWorkOrder]

app/infra/audit/middleware.py

@@ -61,7 +61,10 @@ class AuditMiddleware(BaseHTTPMiddleware):
         "/api/v1/audit/session-events",
         "/audit/session-events",
     }
+    # Query handles and source arguments are not legacy REST audit payloads.
+    # Task queries persist scoped evidence/receipts in their own transaction.
     EXCLUDED_PATH_PREFIXES = (
+        "/api/v1/mcp-query-contexts",
     )
 
     async def dispatch(self, request: Request, call_next: Callable) -> Response:

app/infra/db/metadb/repositories/metadata_repository.py

@@ -525,6 +525,39 @@ class MetadataRepository:
             pool_max_size=record.pool_max_size,
         )
 
+    async def list_active_project_db_routings(
+        self, db_role: str
+    ) -> List[ProjectDbRouting]:
+        result = await self.session.execute(
+            select(models.ProjectDatabase)
+            .join(models.Project, models.Project.id == models.ProjectDatabase.project_id)
+            .where(
+                models.ProjectDatabase.db_role == db_role,
+                models.Project.status == "active",
+            )
+            .order_by(models.ProjectDatabase.project_id)
+        )
+        if not is_database_encryption_configured():
+            raise ValueError("DATABASE_ENCRYPTION_KEY is not configured")
+        encryptor = get_database_encryptor()
+        try:
+            return [
+                ProjectDbRouting(
+                    project_id=record.project_id,
+                    db_role=record.db_role,
+                    db_type=record.db_type,
+                    dsn=_normalize_postgres_dsn(encryptor.decrypt(record.dsn_encrypted)),
+                    pool_min_size=record.pool_min_size,
+                    pool_max_size=record.pool_max_size,
+                )
+                for record in result.scalars().all()
+            ]
+        except InvalidToken as exc:
+            raise ValueError(
+                "Failed to decrypt project DB DSN: DATABASE_ENCRYPTION_KEY mismatch "
+                "or invalid dsn_encrypted value"
+            ) from exc
+
     async def list_projects_for_user(self, user_id: UUID) -> List[ProjectSummary]:
         stmt = (
             select(models.Project, models.UserProjectMembership.project_role)

app/infra/db/postgresql/agent_messages.py

@@ -0,0 +1,88 @@
+"""Project scoped private message storage, sharing task transaction locks."""
+from psycopg import AsyncConnection
+
+from app.infra.db.postgresql.operational_tasks import OperationalTaskRepository
+
+
+class AgentMessageRepository(OperationalTaskRepository):
+    def __init__(self, connection: AsyncConnection, project_id):
+        super().__init__(connection, project_id)
+
+    async def get_message(self, message_id: str) -> dict | None:
+        cursor = await self.connection.execute(
+            "SELECT * FROM workbench.agent_messages WHERE project_id=%s AND message_id=%s",
+            (self.project_id, message_id),
+        )
+        row = await cursor.fetchone()
+        return self._message(row) if row else None
+
+    async def get_message_request(self, owner_id: str, session_id: str, request_id: str) -> dict | None:
+        cursor = await self.connection.execute(
+            """SELECT * FROM workbench.agent_messages WHERE project_id=%s
+               AND source_owner_id=%s AND source_session_id=%s AND request_id=%s""",
+            (self.project_id, owner_id, session_id, request_id),
+        )
+        row = await cursor.fetchone()
+        return self._message(row) if row else None
+
+    @staticmethod
+    def _message(row) -> dict:
+        return dict(row)
+
+    async def insert_message(self, message: dict) -> None:
+        keys = [key for key in message if key != "project_id"]
+        await self.connection.execute(
+            f"INSERT INTO workbench.agent_messages (project_id,{','.join(keys)}) "
+            f"VALUES ({','.join(['%s'] * (len(keys) + 1))})",
+            (self.project_id, *(message[key] for key in keys)),
+        )
+
+    async def list_session_messages(
+        self, owner_id: str, session_id: str, limit: int, *, cursor_at=None, cursor_id=None, pending_only=False,
+    ) -> list[dict]:
+        cursor = await self.connection.execute(
+            """SELECT * FROM workbench.agent_messages WHERE project_id=%s
+               AND conversation_owner_id=%s AND (source_session_id=%s OR
+                    (direction='to_chat' AND target_session_id=%s))
+               AND (%s::timestamptz IS NULL OR (created_at,message_id)<(%s::timestamptz,%s::uuid))
+               AND (NOT %s OR status IN ('pending','enqueued'))
+               ORDER BY created_at DESC,message_id DESC LIMIT %s""",
+            (self.project_id, owner_id, session_id, session_id, cursor_at, cursor_at, cursor_id, pending_only, limit),
+        )
+        return [self._message(row) for row in await cursor.fetchall()]
+
+    async def list_run_session_messages(
+        self, session_id: str, limit: int, *, cursor_at=None, cursor_id=None, pending_only=False,
+    ) -> list[dict]:
+        cursor = await self.connection.execute(
+            """SELECT m.* FROM workbench.agent_messages m
+               JOIN workbench.operational_task_runs r ON r.project_id=m.project_id AND r.run_id=m.run_id
+               WHERE m.project_id=%s AND r.agent_session_id=%s
+                 AND r.status IN ('claimed','running')
+                 AND (%s::timestamptz IS NULL OR (m.created_at,m.message_id)<(%s::timestamptz,%s::uuid))
+                 AND (NOT %s OR (m.direction='to_task' AND m.status IN ('pending','enqueued')))
+               ORDER BY m.created_at DESC,m.message_id DESC LIMIT %s""",
+            (self.project_id, session_id, cursor_at, cursor_at, cursor_id, pending_only, limit),
+        )
+        return [self._message(row) for row in await cursor.fetchall()]
+
+    async def list_pending_messages(self, limit: int, *, cursor_at=None, cursor_id=None) -> list[dict]:
+        cursor = await self.connection.execute(
+            """SELECT m.* FROM workbench.agent_messages m
+               JOIN workbench.operational_task_runs r
+                 ON r.project_id=m.project_id AND r.run_id=m.run_id
+               WHERE m.project_id=%s AND m.status IN ('pending','enqueued')
+                 AND (m.direction='to_chat' OR r.status IN ('claimed','running'))
+                 AND (%s::timestamptz IS NULL OR (m.created_at,m.message_id)>(%s::timestamptz,%s::uuid))
+               ORDER BY m.created_at,m.message_id LIMIT %s""",
+            (self.project_id, cursor_at, cursor_at, cursor_id, limit),
+        )
+        return [self._message(row) for row in await cursor.fetchall()]
+
+    async def update_message_delivery(self, message_id: str, update: dict) -> None:
+        keys = list(update)
+        await self.connection.execute(
+            f"UPDATE workbench.agent_messages SET {','.join(key+'=%s' for key in keys)} "
+            "WHERE project_id=%s AND message_id=%s",
+            (*(update[key] for key in keys), self.project_id, message_id),
+        )

app/infra/db/postgresql/analysis.py

@@ -4,6 +4,7 @@ from uuid import UUID, uuid4
 
 from psycopg import AsyncConnection, Connection
 from psycopg.types.json import Jsonb
+from app.infra.db.postgresql.workbench_events import lock_event_transaction_sync
 
 
 class AnalysisRepository:
@@ -19,6 +20,7 @@ class AnalysisRepository:
         parameters: dict[str, Any],
         run_id: UUID | None = None,
     ) -> dict[str, Any]:
+        lock_event_transaction_sync(conn)
         execution_id = run_id or uuid4()
         with conn.cursor() as cur:
             cur.execute(
@@ -53,18 +55,53 @@ class AnalysisRepository:
         status: str,
         created_by: str,
         parameters: dict[str, Any],
-    ) -> None:
+    ) -> int:
+        lock_event_transaction_sync(conn)
         with conn.cursor() as cur:
             cur.execute(
                 """
                 UPDATE analysis.runs
-                SET created_by = %s, status = %s, parameters = %s
+                SET created_by = %s, status = %s, parameters = %s,
+                    revision = revision + 1
                 WHERE run_id = %s
+                RETURNING revision
                 """,
                 (created_by, status, Jsonb(parameters), run_id),
             )
-            if cur.rowcount != 1:
+            row = cur.fetchone()
+            if row is None:
                 raise LookupError(f"analysis run {run_id} does not exist")
+            return int(row["revision"])
+
+    @staticmethod
+    def append_run_changed_event_sync(
+        conn: Connection,
+        *,
+        project_id: UUID,
+        run_id: UUID,
+        revision: int,
+        status: str,
+        occurred_at: datetime,
+        results_committed: bool,
+    ) -> None:
+        with conn.cursor() as cur:
+            cur.execute(
+                """INSERT INTO workbench.domain_event_log
+                   (project_id,event_id,topic,aggregate_type,aggregate_id,
+                    aggregate_revision,occurred_at,metadata)
+                   VALUES (%s,%s,'analysis.run.changed','analysis_run',%s,%s,%s,%s)
+                   RETURNING event_seq""",
+                (
+                    project_id, uuid4(), str(run_id), revision, occurred_at,
+                    Jsonb({"run_id": str(run_id), "status": status,
+                           "results_committed": results_committed}),
+                ),
+            )
+            row = cur.fetchone()
+            cur.execute(
+                "SELECT pg_notify('workbench_events', %s)",
+                (f'{project_id}:{row["event_seq"]}',),
+            )
 
     @staticmethod
     def insert_result_sync(
@@ -101,7 +138,11 @@ class AnalysisRepository:
             return cur.fetchone()
 
     @staticmethod
-    async def list_runs(conn: AsyncConnection) -> list[dict]:
+    async def list_runs(
+        conn: AsyncConnection, *, limit: int = 100, offset: int = 0
+    ) -> list[dict]:
+        if not 1 <= limit <= 100 or offset < 0:
+            raise ValueError("analysis run page is out of range")
         async with conn.cursor() as cur:
             await cur.execute(
                 """
@@ -109,7 +150,52 @@ class AnalysisRepository:
                        started_at, status, parameters
                 FROM analysis.runs
                 ORDER BY created_at DESC, run_id
+                LIMIT %s OFFSET %s
+                """,
+                (limit, offset),
+            )
+            return await cur.fetchall()
+
+    @staticmethod
+    async def list_recent_runs(
+        conn: AsyncConnection, *, limit: int = 100
+    ) -> list[dict]:
+        if not 1 <= limit <= 100:
+            raise ValueError("recent analysis run limit must be between 1 and 100")
+        async with conn.cursor() as cur:
+            await cur.execute(
                 """
+                SELECT run_id, name, run_type, created_by, created_at,
+                       started_at, status, parameters
+                FROM analysis.runs
+                ORDER BY created_at DESC, run_id
+                LIMIT %s
+                """,
+                (limit,),
+            )
+            return await cur.fetchall()
+
+    @staticmethod
+    async def list_run_summaries(
+        conn: AsyncConnection, *, limit: int = 21, offset: int = 0
+    ) -> list[dict]:
+        if not 1 <= limit <= 101 or offset < 0:
+            raise ValueError("analysis run summary page is out of range")
+        async with conn.cursor() as cur:
+            await cur.execute(
+                """
+                SELECT r.run_id, r.name, r.run_type, r.created_by,
+                       r.created_at, r.started_at, r.status,
+                       EXISTS (
+                           SELECT 1
+                           FROM analysis.results AS result
+                           WHERE result.run_id = r.run_id
+                       ) AS has_results
+                FROM analysis.runs AS r
+                ORDER BY r.created_at DESC, r.run_id
+                LIMIT %s OFFSET %s
+                """,
+                (limit, offset),
             )
             return await cur.fetchall()
 
@@ -129,20 +215,76 @@ class AnalysisRepository:
 
     @staticmethod
     async def list_results(
-        conn: AsyncConnection, run_id: UUID, result_type: str | None = None
+        conn: AsyncConnection,
+        run_id: UUID,
+        result_type: str | None = None,
+        *,
+        limit: int = 100,
+        offset: int = 0,
+        lookahead: bool = False,
     ) -> list[dict]:
+        if not 1 <= limit <= 500 or offset < 0:
+            raise ValueError("analysis result page is out of range")
         query = """
             SELECT result_id, run_id, result_type, node_id, link_id,
                    payload, created_at
             FROM analysis.results
             WHERE run_id = %s
         """
-        params: tuple[UUID] | tuple[UUID, str] = (run_id,)
+        params: tuple = (run_id,)
         if result_type is not None:
             query += " AND result_type = %s"
             params = (run_id, result_type)
-        query += " ORDER BY created_at, result_id"
+        query += " ORDER BY created_at, result_id LIMIT %s OFFSET %s"
+        params += (limit + int(lookahead), offset)
 
         async with conn.cursor() as cur:
             await cur.execute(query, params)
             return await cur.fetchall()
+
+    @staticmethod
+    async def list_recent_results(
+        conn: AsyncConnection, run_id: UUID, *, limit: int = 100
+    ) -> list[dict]:
+        if not 1 <= limit <= 100:
+            raise ValueError("recent analysis result limit must be between 1 and 100")
+        async with conn.cursor() as cur:
+            await cur.execute(
+                """
+                SELECT result_id, run_id, result_type, node_id, link_id,
+                       payload, created_at
+                FROM (
+                    SELECT result_id, run_id, result_type, node_id, link_id,
+                           payload, created_at
+                    FROM analysis.results
+                    WHERE run_id = %s
+                    ORDER BY created_at DESC, result_id DESC
+                    LIMIT %s
+                ) AS recent_results
+                ORDER BY created_at, result_id
+                """,
+                (run_id, limit),
+            )
+            return await cur.fetchall()
+
+    @staticmethod
+    async def list_junction_map_points(
+        conn: AsyncConnection, node_ids: list[str]
+    ) -> list[dict]:
+        if not node_ids:
+            return []
+        if len(node_ids) > 500:
+            raise ValueError("analysis map projection supports at most 500 junctions")
+        async with conn.cursor() as cur:
+            await cur.execute(
+                """
+                SELECT g.id AS node_id,
+                       ST_X(ST_Transform(g.geom, 4326)) AS longitude,
+                       ST_Y(ST_Transform(g.geom, 4326)) AS latitude
+                FROM gis.junctions AS g
+                WHERE g.id = ANY(%s)
+                ORDER BY g.id
+                """,
+                (node_ids,),
+            )
+            return await cur.fetchall()

app/infra/db/postgresql/gis_attributes.py

@@ -0,0 +1,80 @@
+"""Read-only access to imported source GIS attributes."""
+
+from typing import Any
+
+from psycopg import AsyncConnection
+
+
+class GisAttributeRepository:
+    @staticmethod
+    async def list_element_groups(
+        conn: AsyncConnection,
+        element_type: str,
+        element_id: str,
+    ) -> list[dict[str, Any]]:
+        async with conn.cursor() as cur:
+            await cur.execute(
+                "SELECT to_regclass('source_gis.element_links') AS relation"
+            )
+            availability = await cur.fetchone()
+            if not availability or availability["relation"] is None:
+                return []
+            await cur.execute(
+                """
+                SELECT
+                    dataset.dataset_key,
+                    dataset.title AS dataset_title,
+                    dataset.source_date,
+                    layer.layer_key,
+                    layer.display_name AS layer_name,
+                    feature.source_fid,
+                    link.match_method,
+                    link.distance_m AS match_distance_m,
+                    feature.attributes
+                FROM source_gis.element_links AS link
+                JOIN source_gis.features AS feature
+                  ON feature.layer_key = link.layer_key
+                 AND feature.source_fid = link.source_fid
+                JOIN source_gis.layers AS layer
+                  ON layer.layer_key = feature.layer_key
+                JOIN source_gis.datasets AS dataset
+                  ON dataset.dataset_key = layer.dataset_key
+                WHERE link.element_type = %s
+                  AND link.element_id = %s
+                ORDER BY dataset.source_date DESC, layer.display_name, feature.source_fid
+                """,
+                (element_type, element_id),
+            )
+            return [dict(row) for row in await cur.fetchall()]
+
+    @staticmethod
+    async def get_source_feature_group(
+        conn: AsyncConnection,
+        layer_key: str,
+        source_fid: int,
+    ) -> dict[str, Any] | None:
+        async with conn.cursor() as cur:
+            await cur.execute(
+                """
+                SELECT
+                    dataset.dataset_key,
+                    dataset.title AS dataset_title,
+                    dataset.source_date,
+                    layer.layer_key,
+                    layer.display_name AS layer_name,
+                    feature.source_fid,
+                    'source' AS match_method,
+                    NULL::double precision AS match_distance_m,
+                    feature.attributes
+                FROM source_gis.features AS feature
+                JOIN source_gis.layers AS layer
+                  ON layer.layer_key = feature.layer_key
+                JOIN source_gis.datasets AS dataset
+                  ON dataset.dataset_key = layer.dataset_key
+                WHERE feature.layer_key = %s
+                  AND feature.source_fid = %s
+                """,
+                (layer_key, source_fid),
+            )
+            row = await cur.fetchone()
+            return dict(row) if row else None

app/infra/db/postgresql/mcp_queries.py

@@ -0,0 +1,91 @@
+"""Persistent MCP context and receipt storage in the selected project's BizDB."""
+
+from datetime import datetime, timezone
+from psycopg import AsyncConnection
+from psycopg.types.json import Jsonb
+from app.mcp.models import BoundContext
+
+
+class McpQueryRepository:
+    def __init__(self, conn: AsyncConnection, project_id):
+        self.conn, self.project_id = conn, project_id
+
+    async def create(self, context: BoundContext):
+        async with self.conn.transaction(), self.conn.cursor() as cur:
+            # Bounded maintenance; expired/revoked handles remain invalid even if cleanup fails.
+            await cur.execute(
+                """DELETE FROM workbench.mcp_query_contexts WHERE context_handle IN
+                (SELECT context_handle FROM workbench.mcp_query_contexts
+                 WHERE project_id=%s AND (expires_at < now() OR revoked_at IS NOT NULL)
+                 ORDER BY expires_at LIMIT 100 FOR UPDATE SKIP LOCKED)""",
+                (self.project_id,),
+            )
+            await cur.execute(
+                """INSERT INTO workbench.mcp_query_contexts
+                (context_handle,project_id,subject,client_id,document,expires_at)
+                VALUES (%s,%s,%s,%s,%s,%s)""",
+                (
+                    context.context_handle,
+                    self.project_id,
+                    context.subject,
+                    context.client_id,
+                    Jsonb(context.model_dump(mode="json")),
+                    context.expires_at,
+                ),
+            )
+
+    async def get(self, handle: str) -> BoundContext | None:
+        async with self.conn.transaction(), self.conn.cursor() as cur:
+            await cur.execute(
+                """SELECT document,revoked_at FROM workbench.mcp_query_contexts
+                WHERE context_handle=%s AND project_id=%s""",
+                (handle, self.project_id),
+            )
+            row = await cur.fetchone()
+        if not row:
+            return None
+        document = dict(row["document"])
+        document["revoked_at"] = row["revoked_at"]
+        return BoundContext.model_validate(document)
+
+    async def revoke(self, handle: str, subject: str, client_id: str):
+        async with self.conn.transaction(), self.conn.cursor() as cur:
+            await cur.execute(
+                """UPDATE workbench.mcp_query_contexts SET revoked_at=COALESCE(revoked_at,%s)
+                WHERE context_handle=%s AND project_id=%s AND subject=%s AND client_id=%s""",
+                (
+                    datetime.now(timezone.utc),
+                    handle,
+                    self.project_id,
+                    subject,
+                    client_id,
+                ),
+            )
+
+    async def receipt(self, scope_key: str, call_id: str) -> dict | None:
+        async with self.conn.transaction(), self.conn.cursor() as cur:
+            await cur.execute(
+                """SELECT parameters,result FROM workbench.mcp_query_receipts
+                WHERE project_id=%s AND scope_key=%s AND call_id=%s""",
+                (self.project_id, scope_key, call_id),
+            )
+            return await cur.fetchone()
+
+    async def save_receipt(
+        self, scope_key: str, call_id: str, parameters: dict, result: dict
+    ):
+        # Caller owns the short evidence + receipt transaction.
+        async with self.conn.cursor() as cur:
+            await cur.execute(
+                """INSERT INTO workbench.mcp_query_receipts
+                (project_id,scope_key,call_id,parameters,result) VALUES (%s,%s,%s,%s,%s)
+                ON CONFLICT (project_id,scope_key,call_id) DO NOTHING""",
+                (self.project_id, scope_key, call_id, Jsonb(parameters), Jsonb(result)),
+            )
+
+    async def lock_receipt(self, scope_key: str, call_id: str):
+        async with self.conn.cursor() as cur:
+            await cur.execute(
+                "SELECT pg_advisory_xact_lock(hashtextextended(%s,0))",
+                (f"mcp:{self.project_id}:{scope_key}:{call_id}",),
+            )

app/infra/db/postgresql/operational_tasks.py

@@ -0,0 +1,823 @@
+"""PostgreSQL persistence for durable operational tasks."""
+from __future__ import annotations
+
+from contextlib import asynccontextmanager
+from datetime import datetime
+from typing import Any
+
+from psycopg import AsyncConnection
+from psycopg.errors import UndefinedTable
+from psycopg.types.json import Jsonb
+from uuid import uuid4
+
+from app.domain.domain_errors import DomainError
+from app.infra.db.postgresql.workbench_events import lock_event_transaction
+from app.infra.db.postgresql.workbench_events import WorkbenchEventRepository
+
+
+class OperationalTaskRepository:
+    def __init__(self, connection: AsyncConnection, project_id):
+        self.connection = connection
+        self.project_id = str(project_id)
+
+    @asynccontextmanager
+    async def transaction(self):
+        try:
+            async with self.connection.transaction():
+                await lock_event_transaction(self.connection)
+                yield self
+        except UndefinedTable as exc:
+            raise DomainError(
+                "DEPENDENCY_UNAVAILABLE",
+                "Project operational-task storage is not provisioned",
+                status=503,
+            ) from exc
+
+    async def lock(self, key: str) -> None:
+        await self.connection.execute(
+            "SELECT pg_advisory_xact_lock(hashtextextended(%s, 0))",
+            (f"tjwater-operational:{self.project_id}:{key}",),
+        )
+
+    async def try_lock(self, key: str) -> bool:
+        cursor = await self.connection.execute(
+            "SELECT pg_try_advisory_xact_lock(hashtextextended(%s, 0)) AS acquired",
+            (f"tjwater-operational:{self.project_id}:{key}",),
+        )
+        return bool((await cursor.fetchone())["acquired"])
+
+    async def get_request(self, scope: str, request_id: str) -> dict | None:
+        cursor = await self.connection.execute(
+            """SELECT request_digest,response
+               FROM workbench.operational_task_requests
+               WHERE project_id=%s AND scope=%s AND request_id=%s""",
+            (self.project_id, scope, request_id),
+        )
+        row = await cursor.fetchone()
+        return dict(row) if row else None
+
+    async def save_request(
+        self, scope: str, request_id: str, request_digest: str, response: dict
+    ) -> None:
+        cursor = await self.connection.execute(
+            """INSERT INTO workbench.operational_task_requests
+               (project_id,scope,request_id,request_digest,response)
+               VALUES (%s,%s,%s,%s,%s)
+               ON CONFLICT(project_id,scope,request_id) DO UPDATE SET
+                 response=excluded.response
+               WHERE operational_task_requests.request_digest=excluded.request_digest""",
+            (self.project_id, scope, request_id, request_digest, Jsonb(response)),
+        )
+        if cursor.rowcount != 1:
+            raise DomainError(
+                "IDEMPOTENCY_CONFLICT",
+                "request_id was used with different content",
+                status=409,
+            )
+
+    async def get_task(self, task_id: str) -> dict | None:
+        cursor = await self.connection.execute(
+            """SELECT document FROM workbench.operational_tasks
+               WHERE project_id=%s AND task_id=%s""",
+            (self.project_id, task_id),
+        )
+        row = await cursor.fetchone()
+        return row["document"] if row else None
+
+    async def get_active_run(self, task_id: str) -> dict | None:
+        cursor = await self.connection.execute(
+            """SELECT run_id,task_id,generation,execution_request_id,status,revision,
+                      agent_session_id,lease_expires_at,heartbeat_at,started_at,finished_at,
+                      failure_code,failure_detail,dispatch_attempts,next_dispatch_at,
+                      created_at,updated_at
+               FROM workbench.operational_task_runs
+               WHERE project_id=%s AND task_id=%s
+                 AND status IN ('queued','claimed','running')
+               ORDER BY generation DESC LIMIT 1""",
+            (self.project_id, task_id),
+        )
+        row = await cursor.fetchone()
+        return dict(row) if row else None
+
+    async def get_run(self, run_id: str) -> dict | None:
+        cursor = await self.connection.execute(
+            """SELECT run_id,task_id,generation,execution_request_id,status,revision,
+                      agent_session_id,lease_expires_at,heartbeat_at,started_at,finished_at,
+                      failure_code,failure_detail,dispatch_attempts,next_dispatch_at,
+                      created_at,updated_at
+               FROM workbench.operational_task_runs
+               WHERE project_id=%s AND run_id=%s""",
+            (self.project_id, run_id),
+        )
+        row = await cursor.fetchone()
+        return dict(row) if row else None
+
+    async def list_task_runs(self, task_id: str, limit: int, offset: int) -> list[dict]:
+        cursor = await self.connection.execute(
+            """SELECT * FROM workbench.operational_task_runs
+               WHERE project_id=%s AND task_id=%s
+               ORDER BY generation DESC LIMIT %s OFFSET %s""",
+            (self.project_id, task_id, limit, offset),
+        )
+        return [{key: value for key, value in row.items() if key != "project_id"}
+                for row in await cursor.fetchall()]
+
+    async def get_latest_run(self, task_id: str) -> dict | None:
+        rows = await self.list_task_runs(task_id, 1, 0)
+        return rows[0] if rows else None
+
+    async def require_execution_storage(self) -> None:
+        """Reject claims before changing state if the inbox cannot be used."""
+        cursor = await self.connection.execute(
+            "SELECT to_regclass('workbench.agent_messages') IS NOT NULL AS available"
+        )
+        if not (await cursor.fetchone())["available"]:
+            raise DomainError(
+                "DEPENDENCY_UNAVAILABLE",
+                "Agent message storage is not provisioned; apply "
+                "resources/sql/024_agent_messages.sql to this project database before starting runs",
+                status=503,
+            )
+
+    async def pending_run_messages(self, run_id: str) -> list[str]:
+        cursor = await self.connection.execute(
+            """SELECT message_id FROM workbench.agent_messages
+               WHERE project_id=%s AND run_id=%s AND direction='to_task'
+                 AND status IN ('pending','enqueued') ORDER BY created_at LIMIT 100""",
+            (self.project_id, run_id),
+        )
+        return [str(row["message_id"]) for row in await cursor.fetchall()]
+
+    async def fail_run_messages(self, run_id: str, failure_detail: str, now: datetime) -> None:
+        await self.connection.execute(
+            """UPDATE workbench.agent_messages SET status='failed',failure_detail=%s,updated_at=%s
+               WHERE project_id=%s AND run_id=%s AND direction='to_task'
+                 AND status IN ('pending','enqueued')""",
+            (failure_detail, now, self.project_id, run_id),
+        )
+
+    async def insert_run(self, run: dict) -> None:
+        await self.connection.execute(
+            """INSERT INTO workbench.operational_task_runs
+               (project_id,run_id,task_id,generation,execution_request_id,status,revision,
+                agent_session_id,lease_expires_at,heartbeat_at,started_at,finished_at,
+                failure_code,failure_detail,dispatch_attempts,next_dispatch_at,
+                created_at,updated_at)
+               VALUES (%s,%s,%s,%s,%s,%s,%s,%s,%s,%s,%s,%s,%s,%s,%s,%s,%s,%s)""",
+            (
+                self.project_id, run["run_id"], run["task_id"], run["generation"],
+                run["execution_request_id"], run["status"], run["revision"],
+                run.get("agent_session_id"), run.get("lease_expires_at"),
+                run.get("heartbeat_at"), run.get("started_at"), run.get("finished_at"),
+                run.get("failure_code"), run.get("failure_detail"),
+                run.get("dispatch_attempts", 0), run["next_dispatch_at"],
+                run["created_at"], run["updated_at"],
+            ),
+        )
+
+    async def update_run(self, run: dict, expected_revision: int) -> None:
+        cursor = await self.connection.execute(
+            """UPDATE workbench.operational_task_runs SET
+                 status=%s,revision=%s,agent_session_id=%s,lease_expires_at=%s,
+                 heartbeat_at=%s,started_at=%s,finished_at=%s,failure_code=%s,
+                 failure_detail=%s,dispatch_attempts=%s,next_dispatch_at=%s,updated_at=%s
+               WHERE project_id=%s AND run_id=%s AND revision=%s""",
+            (
+                run["status"], run["revision"], run.get("agent_session_id"),
+                run.get("lease_expires_at"), run.get("heartbeat_at"),
+                run.get("started_at"), run.get("finished_at"), run.get("failure_code"),
+                run.get("failure_detail"), run.get("dispatch_attempts", 0),
+                run["next_dispatch_at"], run["updated_at"], self.project_id,
+                run["run_id"], expected_revision,
+            ),
+        )
+        if cursor.rowcount != 1:
+            raise DomainError("VERSION_CONFLICT", "Run revision changed", status=409)
+
+    async def list_runs(
+        self, *, status: str | None, limit: int, cursor_created_at=None, cursor_run_id=None,
+        task_type: str | None = None, task_created_from: datetime | None = None,
+    ) -> list[dict]:
+        result = await self.connection.execute(
+            """SELECT run_id,task_id,generation,execution_request_id,status,revision,
+                      agent_session_id,lease_expires_at,heartbeat_at,started_at,finished_at,
+                      failure_code,failure_detail,dispatch_attempts,next_dispatch_at,
+                      created_at,updated_at
+               FROM workbench.operational_task_runs
+               WHERE project_id=%s
+                 AND (%s::text IS NULL OR status=%s::text)
+                 AND (%s::timestamptz IS NULL OR (created_at,run_id)>(%s::timestamptz,%s::uuid))
+                 AND (status<>'queued' OR next_dispatch_at<=now())
+                 AND EXISTS (
+                     SELECT 1 FROM workbench.operational_tasks t
+                     WHERE t.project_id=operational_task_runs.project_id
+                       AND t.task_id=operational_task_runs.task_id
+                       AND (%s::text IS NULL OR t.task_type=%s::text)
+                       AND (%s::timestamptz IS NULL OR t.created_at>=%s::timestamptz)
+                 )
+               ORDER BY created_at,run_id LIMIT %s""",
+            (
+                self.project_id, status, status, cursor_created_at,
+                cursor_created_at, cursor_run_id,
+                task_type, task_type, task_created_from, task_created_from, limit,
+            ),
+        )
+        return [dict(row) for row in await result.fetchall()]
+
+    async def claim_expired_runs(
+        self, limit: int, *, task_type: str | None = None, task_created_from: datetime | None = None,
+    ) -> list[dict]:
+        cursor = await self.connection.execute(
+            """SELECT run_id,task_id,generation,execution_request_id,status,revision,
+                      agent_session_id,lease_expires_at,heartbeat_at,started_at,finished_at,
+                      failure_code,failure_detail,dispatch_attempts,next_dispatch_at,
+                      created_at,updated_at
+               FROM workbench.operational_task_runs
+               WHERE project_id=%s AND status IN ('claimed','running')
+                 AND lease_expires_at IS NOT NULL AND lease_expires_at<=now()
+                 AND EXISTS (
+                     SELECT 1 FROM workbench.operational_tasks t
+                     WHERE t.project_id=operational_task_runs.project_id
+                       AND t.task_id=operational_task_runs.task_id
+                       AND (%s::text IS NULL OR t.task_type=%s::text)
+                       AND (%s::timestamptz IS NULL OR t.created_at>=%s::timestamptz)
+                 )
+               ORDER BY lease_expires_at,run_id FOR UPDATE SKIP LOCKED LIMIT %s""",
+            (self.project_id, task_type, task_type, task_created_from, task_created_from, limit),
+        )
+        return [dict(row) for row in await cursor.fetchall()]
+
+    async def consecutive_timed_out_runs(self, task_id: str) -> int:
+        cursor = await self.connection.execute(
+            """SELECT status FROM workbench.operational_task_runs
+               WHERE project_id=%s AND task_id=%s ORDER BY generation DESC""",
+            (self.project_id, task_id),
+        )
+        count = 0
+        for row in await cursor.fetchall():
+            if row["status"] != "timed_out":
+                break
+            count += 1
+        return count
+
+    async def append_event(
+        self, *, topic: str, aggregate_type: str, aggregate_id: str,
+        aggregate_revision: int, occurred_at, metadata: dict | None = None,
+    ) -> int:
+        event_id = uuid4()
+        cursor = await self.connection.execute(
+            """INSERT INTO workbench.domain_event_log
+               (project_id,event_id,topic,aggregate_type,aggregate_id,
+                aggregate_revision,occurred_at,metadata)
+               VALUES (%s,%s,%s,%s,%s,%s,%s,%s) RETURNING event_seq""",
+            (
+                self.project_id, event_id, topic, aggregate_type, aggregate_id,
+                aggregate_revision, occurred_at, Jsonb(metadata or {}),
+            ),
+        )
+        row = await cursor.fetchone()
+        event_seq = int(row["event_seq"])
+        await self.connection.execute(
+            "SELECT pg_notify('workbench_events', %s)",
+            (f'{self.project_id}:{event_seq}',),
+        )
+        return event_seq
+
+    async def current_event_cursor(self) -> int:
+        return await WorkbenchEventRepository(self.connection, self.project_id).current_cursor()
+
+    async def get_task_by_create_request(self, request_id: str) -> dict | None:
+        cursor = await self.connection.execute(
+            """SELECT document FROM workbench.operational_tasks
+               WHERE project_id=%s AND create_request_id=%s""",
+            (self.project_id, request_id),
+        )
+        row = await cursor.fetchone()
+        return row["document"] if row else None
+
+    async def get_task_by_external_source(
+        self, source_kind: str, external_id: str
+    ) -> dict | None:
+        cursor = await self.connection.execute(
+            """SELECT document FROM workbench.operational_tasks
+               WHERE project_id=%s
+                 AND document#>>'{source,kind}'=%s
+                 AND document#>>'{source,external_id}'=%s""",
+            (self.project_id, source_kind, external_id),
+        )
+        row = await cursor.fetchone()
+        return row["document"] if row else None
+
+    async def list_tasks(
+        self,
+        *,
+        task_type: str | None,
+        status: str | None,
+        limit: int,
+        offset: int,
+        peripheral_work_order_id: str | None = None,
+    ) -> list[dict]:
+        cursor = await self.connection.execute(
+            """SELECT task.document,
+                      CASE WHEN active_run.run_id IS NULL THEN NULL
+                           ELSE to_jsonb(active_run) - 'project_id' END AS active_run,
+                      CASE WHEN latest_run.run_id IS NULL THEN NULL
+                           ELSE to_jsonb(latest_run) - 'project_id' END AS latest_run
+               FROM workbench.operational_tasks AS task
+               LEFT JOIN LATERAL (
+                   SELECT run.* FROM workbench.operational_task_runs AS run
+                   WHERE run.project_id=task.project_id AND run.task_id=task.task_id
+                     AND run.status IN ('queued','claimed','running')
+                   ORDER BY run.generation DESC LIMIT 1
+               ) AS active_run ON true
+               LEFT JOIN LATERAL (
+                   SELECT run.* FROM workbench.operational_task_runs AS run
+                   WHERE run.project_id=task.project_id AND run.task_id=task.task_id
+                   ORDER BY run.generation DESC LIMIT 1
+               ) AS latest_run ON true
+               WHERE task.project_id=%s
+                 AND (task.document#>>'{source,deleted_at}') IS NULL
+                 AND (%s::text IS NULL OR task.task_type=%s::text)
+                 AND (%s::text IS NULL OR task.status=%s::text)
+                 AND (%s::text IS NULL
+                      OR (task.document#>>'{source,kind}'='external_system'
+                          AND task.document#>>'{source,external_id}' IN (%s,%s))
+                      OR EXISTS (
+                          SELECT 1 FROM workbench.operational_task_actions AS action
+                          WHERE action.project_id=task.project_id AND action.task_id=task.task_id
+                            AND action.document->>'kind'='create_work_order'
+                            AND action.document#>>'{receipt,adapter}'='peripheral_sample'
+                            AND action.document#>>'{receipt,receipt_id}'=%s
+                      ))
+               ORDER BY task.updated_at DESC,task.task_id
+               LIMIT %s OFFSET %s""",
+            (
+                self.project_id,
+                task_type,
+                task_type,
+                status,
+                status,
+                peripheral_work_order_id,
+                peripheral_work_order_id,
+                f"peripheral_sample:{peripheral_work_order_id}" if peripheral_work_order_id else None,
+                peripheral_work_order_id,
+                limit,
+                offset,
+            ),
+        )
+        return [
+            {**row["document"], "active_run": row["active_run"], "latest_run": row["latest_run"]}
+            for row in await cursor.fetchall()
+        ]
+
+    async def insert_task(self, task: dict, create_request_id: str) -> None:
+        await self.connection.execute(
+            """INSERT INTO workbench.operational_tasks
+               (project_id,task_id,create_request_id,task_type,workflow,status,lifecycle_status,priority,
+                revision,execution_generation,execution_request_id,agent_session_id,
+                agent_lease_expires_at,current_plan_revision,current_report_revision,
+                scheduled_for,created_at,updated_at,document)
+               VALUES (%s,%s,%s,%s,%s,%s,%s,%s,%s,%s,%s,%s,%s,%s,%s,%s,%s,%s,%s)""",
+            self._task_params(task, create_request_id),
+        )
+
+    async def update_task(self, task: dict) -> None:
+        cursor = await self.connection.execute(
+            """UPDATE workbench.operational_tasks SET
+                 status=%s,lifecycle_status=%s,revision=%s,execution_generation=%s,execution_request_id=%s,
+                 agent_session_id=%s,agent_lease_expires_at=%s,current_plan_revision=%s,
+                 current_report_revision=%s,updated_at=%s,document=%s
+               WHERE project_id=%s AND task_id=%s AND revision=%s""",
+            (
+                task["status"], task["lifecycle_status"], task["revision"], task["execution_generation"],
+                task["execution_request_id"], task.get("agent_session_id"),
+                task.get("agent_lease_expires_at"), task.get("current_plan_revision"),
+                task.get("current_report_revision"), task["updated_at"], Jsonb(task),
+                self.project_id, task["task_id"], task["revision"] - 1,
+            ),
+        )
+        if cursor.rowcount != 1:
+            raise DomainError("VERSION_CONFLICT", "Task revision changed", status=409)
+
+    def _task_params(self, task: dict, create_request_id: str) -> tuple:
+        return (
+            self.project_id, task["task_id"], create_request_id, task["task_type"],
+            task["workflow"], task["status"], task["lifecycle_status"], task["priority"], task["revision"],
+            task["execution_generation"], task["execution_request_id"],
+            task.get("agent_session_id"), task.get("agent_lease_expires_at"),
+            task.get("current_plan_revision"), task.get("current_report_revision"),
+            task.get("scheduled_for"), task["created_at"], task["updated_at"], Jsonb(task),
+        )
+
+    async def get_evidence(self, task_id: str, evidence_id: str) -> dict | None:
+        cursor = await self.connection.execute(
+            """SELECT document FROM workbench.operational_task_evidence
+               WHERE project_id=%s AND task_id=%s AND evidence_id=%s""",
+            (self.project_id, task_id, evidence_id),
+        )
+        row = await cursor.fetchone()
+        return row["document"] if row else None
+
+    async def list_evidence(self, task_id: str) -> list[dict]:
+        cursor = await self.connection.execute(
+            """SELECT document FROM workbench.operational_task_evidence
+               WHERE project_id=%s AND task_id=%s ORDER BY created_at,evidence_id""",
+            (self.project_id, task_id),
+        )
+        return [row["document"] for row in await cursor.fetchall()]
+
+    async def insert_evidence(self, task_id: str, evidence: dict) -> None:
+        await self.connection.execute(
+            """INSERT INTO workbench.operational_task_evidence
+               (project_id,task_id,evidence_id,document,created_at)
+               VALUES (%s,%s,%s,%s,%s)""",
+            (self.project_id, task_id, evidence["evidence_id"], Jsonb(evidence),
+             evidence["recorded_at"]),
+        )
+
+    async def access_requests_available(self) -> bool:
+        # Older project databases can serve existing tasks before migration 019.
+        # Keep approval writes fail-closed until the table is provisioned.
+        availability = await self.connection.execute(
+            "SELECT to_regclass('workbench.operational_task_access_requests') AS relation"
+        )
+        return (await availability.fetchone())["relation"] is not None
+
+    async def list_access_requests(self, task_id: str) -> list[dict]:
+        if not await self.access_requests_available():
+            return []
+        cursor = await self.connection.execute(
+            """SELECT document FROM workbench.operational_task_access_requests
+               WHERE project_id=%s AND task_id=%s
+               ORDER BY requested_at,access_request_id""",
+            (self.project_id, task_id),
+        )
+        return [row["document"] for row in await cursor.fetchall()]
+
+    async def get_access_request(self, task_id: str, access_request_id: str) -> dict | None:
+        cursor = await self.connection.execute(
+            """SELECT document FROM workbench.operational_task_access_requests
+               WHERE project_id=%s AND task_id=%s AND access_request_id=%s""",
+            (self.project_id, task_id, access_request_id),
+        )
+        row = await cursor.fetchone()
+        return row["document"] if row else None
+
+    async def insert_access_request(self, task_id: str, document: dict) -> None:
+        await self.connection.execute(
+            """INSERT INTO workbench.operational_task_access_requests
+               (project_id,task_id,access_request_id,node_id,status,version,document,requested_at,updated_at)
+               VALUES (%s,%s,%s,%s,%s,%s,%s,%s,%s)""",
+            (self.project_id, task_id, document["access_request_id"], document["node_id"],
+             document["status"], document["version"], Jsonb(document),
+             document["requested_at"], document["requested_at"]),
+        )
+
+    async def update_access_request(self, task_id: str, document: dict, expected_version: int) -> None:
+        cursor = await self.connection.execute(
+            """UPDATE workbench.operational_task_access_requests
+               SET status=%s,version=%s,document=%s,updated_at=%s
+               WHERE project_id=%s AND task_id=%s AND access_request_id=%s AND version=%s""",
+            (document["status"], document["version"], Jsonb(document), document["decided_at"],
+             self.project_id, task_id, document["access_request_id"], expected_version),
+        )
+        if cursor.rowcount != 1:
+            raise DomainError("VERSION_CONFLICT", "Access request version changed", status=409)
+
+    async def get_latest_plan(self, task_id: str) -> dict | None:
+        return await self._latest_artifact("operational_task_plans", task_id)
+
+    async def get_latest_report(self, task_id: str) -> dict | None:
+        return await self._latest_artifact("operational_task_reports", task_id)
+
+    async def get_report(self, task_id: str, revision: int) -> dict | None:
+        cursor = await self.connection.execute(
+            """SELECT document FROM workbench.operational_task_reports
+               WHERE project_id=%s AND task_id=%s AND revision=%s""",
+            (self.project_id, task_id, revision),
+        )
+        row = await cursor.fetchone()
+        return row["document"] if row else None
+
+    async def list_reports(self, task_id: str, limit: int, offset: int) -> list[dict]:
+        cursor = await self.connection.execute(
+            """SELECT document FROM workbench.operational_task_reports
+               WHERE project_id=%s AND task_id=%s ORDER BY revision DESC
+               LIMIT %s OFFSET %s""",
+            (self.project_id, task_id, limit, offset),
+        )
+        return [row["document"] for row in await cursor.fetchall()]
+
+    async def _latest_artifact(self, table: str, task_id: str) -> dict | None:
+        if table not in {"operational_task_plans", "operational_task_reports"}:
+            raise ValueError("unsupported artifact table")
+        cursor = await self.connection.execute(
+            f"""SELECT document FROM workbench.{table}
+                WHERE project_id=%s AND task_id=%s ORDER BY revision DESC LIMIT 1""",
+            (self.project_id, task_id),
+        )
+        row = await cursor.fetchone()
+        return row["document"] if row else None
+
+    async def insert_plan(self, task_id: str, snapshot: dict) -> None:
+        await self._insert_artifact("operational_task_plans", task_id, snapshot)
+
+    async def insert_report(self, task_id: str, snapshot: dict) -> None:
+        await self._insert_artifact("operational_task_reports", task_id, snapshot)
+
+    async def get_latest_progress(self, task_id: str, generation: int) -> dict | None:
+        # Read compatibility before the coordinated progress migration is applied.
+        availability = await self.connection.execute(
+            "SELECT to_regclass('workbench.operational_task_progress') AS relation"
+        )
+        if not (await availability.fetchone())["relation"]:
+            return None
+        cursor = await self.connection.execute(
+            """SELECT progress.document FROM workbench.operational_task_progress AS progress
+               JOIN workbench.operational_task_runs AS run
+                 ON run.project_id=progress.project_id AND run.run_id=progress.run_id
+                    AND run.task_id=progress.task_id
+               WHERE progress.project_id=%s AND progress.task_id=%s AND run.generation=%s""",
+            (self.project_id, task_id, generation),
+        )
+        row = await cursor.fetchone()
+        return row["document"] if row else None
+
+    async def save_progress(self, task_id: str, snapshot: dict) -> None:
+        cursor = await self.connection.execute(
+            """INSERT INTO workbench.operational_task_progress
+               (project_id,task_id,run_id,revision,document,published_at)
+               VALUES (%s,%s,%s,%s,%s,%s)
+               ON CONFLICT(project_id,task_id,run_id) DO UPDATE SET
+                 revision=excluded.revision,document=excluded.document,published_at=excluded.published_at
+               WHERE workbench.operational_task_progress.revision=excluded.revision-1""",
+            (self.project_id, task_id, snapshot["run_id"], snapshot["revision"],
+             Jsonb(snapshot), snapshot["published_at"]),
+        )
+        if cursor.rowcount != 1:
+            raise DomainError("VERSION_CONFLICT", "Investigation progress revision changed", status=409)
+
+    async def _insert_artifact(self, table: str, task_id: str, snapshot: dict) -> None:
+        if table not in {"operational_task_plans", "operational_task_reports"}:
+            raise ValueError("unsupported artifact table")
+        await self.connection.execute(
+            f"""INSERT INTO workbench.{table}
+                (project_id,task_id,revision,document,created_at)
+                VALUES (%s,%s,%s,%s,%s)""",
+            (self.project_id, task_id, snapshot["revision"], Jsonb(snapshot),
+             snapshot["published_at"]),
+        )
+
+    async def get_action(self, task_id: str, action_id: str) -> dict | None:
+        cursor = await self.connection.execute(
+            """SELECT document FROM workbench.operational_task_actions
+               WHERE project_id=%s AND task_id=%s AND action_id=%s""",
+            (self.project_id, task_id, action_id),
+        )
+        row = await cursor.fetchone()
+        return row["document"] if row else None
+
+    async def list_actions(self, task_id: str) -> list[dict]:
+        cursor = await self.connection.execute(
+            """SELECT document FROM workbench.operational_task_actions
+               WHERE project_id=%s AND task_id=%s ORDER BY created_at,action_id""",
+            (self.project_id, task_id),
+        )
+        return [row["document"] for row in await cursor.fetchall()]
+
+    async def insert_action(self, task_id: str, action: dict) -> None:
+        await self.connection.execute(
+            """INSERT INTO workbench.operational_task_actions
+               (project_id,task_id,action_id,version,approval_status,execution_status,
+                document,created_at,updated_at)
+               VALUES (%s,%s,%s,%s,%s,%s,%s,%s,%s)""",
+            (self.project_id, task_id, action["action_id"], action["version"],
+             action["approval_status"], action["execution_status"], Jsonb(action),
+             action["proposed_at"], action["proposed_at"]),
+        )
+        await self._insert_action_revision(task_id, action)
+
+    async def update_action(self, task_id: str, action: dict) -> None:
+        await self._insert_action_revision(task_id, action)
+        await self.connection.execute(
+            """UPDATE workbench.operational_task_actions SET
+                 version=%s,approval_status=%s,execution_status=%s,document=%s,updated_at=%s
+               WHERE project_id=%s AND task_id=%s AND action_id=%s""",
+            (action["version"], action["approval_status"], action["execution_status"],
+             Jsonb(action), action.get("executed_at") or action.get("decided_at") or action.get("edited_at") or
+             action["proposed_at"], self.project_id, task_id, action["action_id"]),
+        )
+
+    async def _insert_action_revision(self, task_id: str, action: dict) -> None:
+        await self.connection.execute(
+            """INSERT INTO workbench.operational_task_action_revisions
+               (project_id,task_id,action_id,version,document,recorded_at)
+               VALUES (%s,%s,%s,%s,%s,%s)""",
+            (
+                self.project_id, task_id, action["action_id"], action["version"],
+                Jsonb(action), action.get("executed_at") or action.get("decided_at") or action.get("edited_at")
+                or action["proposed_at"],
+            ),
+        )
+
+    async def get_feedback(self, task_id: str, source_event_id: str) -> dict | None:
+        cursor = await self.connection.execute(
+            """SELECT source_revision,document,accepted_at
+               FROM workbench.operational_task_external_feedback
+               WHERE project_id=%s AND task_id=%s AND source_event_id=%s""",
+            (self.project_id, task_id, source_event_id),
+        )
+        row = await cursor.fetchone()
+        return dict(row) if row else None
+
+    async def upsert_feedback(self, task_id: str, feedback: dict, accepted_at) -> None:
+        await self.connection.execute(
+            """INSERT INTO workbench.operational_task_external_feedback
+               (project_id,task_id,source_event_id,source_revision,document,accepted_at)
+               VALUES (%s,%s,%s,%s,%s,%s)
+               ON CONFLICT(project_id,task_id,source_event_id) DO UPDATE SET
+                 source_revision=excluded.source_revision,document=excluded.document,
+                 accepted_at=excluded.accepted_at""",
+            (self.project_id, task_id, feedback["source_event_id"],
+             feedback["source_revision"], Jsonb(feedback), accepted_at),
+        )
+
+    async def insert_notification(self, notification: dict) -> None:
+        await self.connection.execute(
+            """INSERT INTO workbench.operational_task_notifications
+               (project_id,notification_id,task_id,created_at,document)
+               VALUES (%s,%s,%s,%s,%s)""",
+            (self.project_id, notification["notification_id"], notification["task_id"],
+             notification["created_at"], Jsonb(notification)),
+        )
+
+    async def resolve_notifications(
+        self,
+        task_id: str,
+        *,
+        kinds: list[str],
+        subject_type: str,
+        subject_id: str,
+        resolved_at,
+        resolution_reason: str,
+    ) -> list[dict]:
+        cursor = await self.connection.execute(
+            """UPDATE workbench.operational_task_notifications
+               SET document=document || jsonb_build_object(
+                   'lifecycle_status','resolved',
+                   'resolved_at',to_jsonb(%s::timestamptz),
+                   'resolution_reason',%s::text
+               )
+               WHERE project_id=%s AND task_id=%s
+                 AND COALESCE(document->>'lifecycle_status','active')='active'
+                 AND document->>'kind'=ANY(%s)
+                 AND document->>'subject_type'=%s
+                 AND document->>'subject_id'=%s
+               RETURNING document""",
+            (
+                resolved_at,
+                resolution_reason,
+                self.project_id,
+                task_id,
+                kinds,
+                subject_type,
+                subject_id,
+            ),
+        )
+        return [row["document"] for row in await cursor.fetchall()]
+
+    async def get_notification(self, notification_id: str, user_id: str) -> dict | None:
+        cursor = await self.connection.execute(
+            """SELECT n.document || jsonb_build_object(
+                     'status',CASE WHEN r.read_at IS NULL THEN 'unread' ELSE 'read' END,
+                     'version',COALESCE(r.version,1),
+                     'read_at',r.read_at,'read_by',
+                         CASE WHEN r.user_id IS NULL THEN NULL ELSE 'principal:' || r.user_id::text END
+                   ) AS document
+               FROM workbench.operational_task_notifications n
+               LEFT JOIN workbench.operational_task_notification_reads r
+                 ON r.project_id=n.project_id AND r.notification_id=n.notification_id
+                AND r.user_id=%s
+               WHERE n.project_id=%s AND n.notification_id=%s""",
+            (user_id, self.project_id, notification_id),
+        )
+        row = await cursor.fetchone()
+        return row["document"] if row else None
+
+    async def mark_notification_read(
+        self, notification_id: str, user_id: str, version: int, read_at
+    ) -> None:
+        await self.connection.execute(
+            """INSERT INTO workbench.operational_task_notification_reads
+               (project_id,notification_id,user_id,version,read_at)
+               VALUES (%s,%s,%s,%s,%s)
+               ON CONFLICT(project_id,notification_id,user_id) DO UPDATE SET
+                 version=excluded.version,read_at=excluded.read_at""",
+            (self.project_id, notification_id, user_id, version, read_at),
+        )
+
+    async def list_notifications(
+        self, *, user_id: str, allowed_permissions: list[str],
+        unread_only: bool, limit: int, offset: int
+    ) -> list[dict]:
+        cursor = await self.connection.execute(
+            """SELECT n.document || jsonb_build_object(
+                     'status',CASE WHEN r.read_at IS NULL THEN 'unread' ELSE 'read' END,
+                     'version',COALESCE(r.version,1),
+                     'read_at',r.read_at,'read_by',
+                         CASE WHEN r.user_id IS NULL THEN NULL ELSE 'principal:' || r.user_id::text END
+                   ) AS document
+               FROM workbench.operational_task_notifications n
+               LEFT JOIN workbench.operational_task_notification_reads r
+                 ON r.project_id=n.project_id AND r.notification_id=n.notification_id
+                AND r.user_id=%s
+               WHERE n.project_id=%s
+                 AND COALESCE(n.document->>'lifecycle_status','active')='active'
+                 AND (%s=false OR r.read_at IS NULL)
+                 AND (n.document->>'required_permission' IS NULL
+                      OR n.document->>'required_permission'=ANY(%s))
+               ORDER BY n.created_at DESC,n.notification_id LIMIT %s OFFSET %s""",
+            (user_id, self.project_id, unread_only, allowed_permissions, limit, offset),
+        )
+        return [row["document"] for row in await cursor.fetchall()]
+
+    async def get_schedule_by_request(self, request_id: str) -> dict | None:
+        cursor = await self.connection.execute(
+            """SELECT document FROM workbench.operational_task_schedules
+               WHERE project_id=%s AND create_request_id=%s""",
+            (self.project_id, request_id),
+        )
+        row = await cursor.fetchone()
+        return row["document"] if row else None
+
+    async def insert_schedule(self, schedule: dict, request_id: str) -> None:
+        await self.connection.execute(
+            """INSERT INTO workbench.operational_task_schedules
+               (project_id,schedule_id,create_request_id,schedule_key,next_run_at,
+                interval_seconds,enabled,version,document,created_at,updated_at)
+               VALUES (%s,%s,%s,%s,%s,%s,%s,%s,%s,%s,%s)""",
+            (self.project_id, schedule["schedule_id"], request_id,
+             schedule["schedule_key"], schedule["next_run_at"],
+             schedule["interval_seconds"], schedule["enabled"], schedule["version"],
+             Jsonb(schedule), schedule["created_at"], schedule["updated_at"]),
+        )
+
+    async def list_schedules(self) -> list[dict]:
+        cursor = await self.connection.execute(
+            """SELECT document FROM workbench.operational_task_schedules
+               WHERE project_id=%s ORDER BY next_run_at,schedule_id""",
+            (self.project_id,),
+        )
+        return [row["document"] for row in await cursor.fetchall()]
+
+    async def get_schedule(self, schedule_id: str) -> dict | None:
+        cursor = await self.connection.execute(
+            """SELECT document FROM workbench.operational_task_schedules
+               WHERE project_id=%s AND schedule_id=%s""",
+            (self.project_id, schedule_id),
+        )
+        row = await cursor.fetchone()
+        return row["document"] if row else None
+
+    async def update_schedule(self, schedule: dict) -> None:
+        cursor = await self.connection.execute(
+            """UPDATE workbench.operational_task_schedules SET
+                 enabled=%s,version=%s,updated_at=%s,document=%s
+               WHERE project_id=%s AND schedule_id=%s AND version=%s""",
+            (
+                schedule["enabled"], schedule["version"], schedule["updated_at"],
+                Jsonb(schedule), self.project_id, schedule["schedule_id"],
+                schedule["version"] - 1,
+            ),
+        )
+        if cursor.rowcount != 1:
+            raise DomainError("VERSION_CONFLICT", "Schedule version changed", status=409)
+
+    async def claim_due_schedules(self, limit: int) -> list[dict]:
+        cursor = await self.connection.execute(
+            """SELECT document FROM workbench.operational_task_schedules
+               WHERE project_id=%s AND enabled=true AND next_run_at <= now()
+               ORDER BY next_run_at,schedule_id FOR UPDATE SKIP LOCKED LIMIT %s""",
+            (self.project_id, limit),
+        )
+        return [row["document"] for row in await cursor.fetchall()]
+
+    async def advance_schedule(self, schedule: dict) -> None:
+        await self.connection.execute(
+            """UPDATE workbench.operational_task_schedules SET
+                 next_run_at=%s,version=%s,updated_at=%s,document=%s
+               WHERE project_id=%s AND schedule_id=%s""",
+            (schedule["next_run_at"], schedule["version"], schedule["updated_at"],
+             Jsonb(schedule), self.project_id, schedule["schedule_id"]),
+        )
+
+    async def insert_occurrence(self, schedule_id: str, due_at, task_id: str, created_at) -> bool:
+        cursor = await self.connection.execute(
+            """INSERT INTO workbench.operational_schedule_occurrences
+               (project_id,schedule_id,due_at,task_id,created_at)
+               VALUES (%s,%s,%s,%s,%s)
+               ON CONFLICT(project_id,schedule_id,due_at) DO NOTHING
+               RETURNING task_id""",
+            (self.project_id, schedule_id, due_at, task_id, created_at),
+        )
+        return await cursor.fetchone() is not None

app/infra/db/postgresql/read_queries.py

@@ -0,0 +1,406 @@
+"""Bounded read primitives over existing project views and time-series tables."""
+
+from datetime import datetime
+from typing import Any
+from uuid import UUID
+
+from psycopg import AsyncConnection
+from psycopg import sql
+from psycopg.rows import dict_row
+from fastapi import HTTPException
+from app.infra.db.postgresql.scada import _SCADA_VIEW_SELECT, _device
+from app.infra.db.timescaledb.repositories.realtime import RealtimeRepository
+from app.infra.db.timescaledb.repositories.analysis import AnalysisResultsRepository
+
+ELEMENT_VIEWS = {
+    v: v + "s" for v in ("junction", "pipe", "reservoir", "tank", "pump", "valve")
+}
+MAX_QUERY_BYTES = 8 * 1024 * 1024
+
+
+async def read_rows(conn, statement, params=(), *, timeout_ms=30000):
+    async with conn.transaction(), conn.cursor(row_factory=dict_row) as cur:
+        await cur.execute(
+            "SELECT set_config('statement_timeout',%s,true)", (str(timeout_ms),)
+        )
+        await cur.execute(statement, params)
+        return await cur.fetchall()
+
+
+async def asset_rows(conn, element_type, *, element_id=None, limit=100, offset=0):
+    view = ELEMENT_VIEWS[element_type]
+    where = sql.SQL("WHERE v.id=%s") if element_id else sql.SQL("")
+    params = (element_id,) if element_id else ()
+    statement = sql.SQL(
+        "SELECT to_jsonb(v)-'geom' AS record FROM gis.{} v {} ORDER BY v.id LIMIT %s OFFSET %s"
+    ).format(sql.Identifier(view), where)
+    rows = await read_rows(conn, statement, params + (limit, offset))
+    return [row["record"] for row in rows]
+
+
+async def scada_devices(conn, *, device_id=None, limit=100, offset=0):
+    where = " WHERE id=%s" if device_id else ""
+    params = (device_id,) if device_id else ()
+    rows = await read_rows(
+        conn,
+        _SCADA_VIEW_SELECT + where + " ORDER BY device_id LIMIT %s OFFSET %s",
+        params + (limit, offset),
+    )
+    return [_device(row) for row in rows]
+
+
+async def series_rows(
+    conn,
+    *,
+    source,
+    element_type,
+    ids,
+    start=None,
+    end=None,
+    time=None,
+    field=None,
+    run_id=None,
+    limit=1000,
+    offset=0,
+    end_exclusive=False,
+):
+    node = element_type == "junction"
+    if source == "scada":
+        table, id_column, fields = (
+            "measurements",
+            "device_id",
+            {"monitored_value", "cleaned_value"},
+        )
+    else:
+        table, id_column = (
+            ("node_results", "node_id") if node else ("link_results", "link_id")
+        )
+        fields = (
+            (
+                AnalysisResultsRepository.NODE_FIELDS
+                if node
+                else AnalysisResultsRepository.LINK_FIELDS
+            )
+            if source == "analysis"
+            else (
+                RealtimeRepository.NODE_FIELDS
+                if node
+                else RealtimeRepository.LINK_FIELDS
+            )
+        )
+    if field is not None and field not in fields:
+        raise ValueError("unsupported field for this data source and element type")
+    columns = (
+        sql.SQL("{},time,{} AS value").format(
+            sql.Identifier(id_column), sql.Identifier(field)
+        )
+        if field
+        else sql.SQL("*")
+    )
+    filters = []
+    params = []
+    if ids:
+        filters.append(sql.SQL("{}=ANY(%s)").format(sql.Identifier(id_column)))
+        params.append(ids)
+    if run_id:
+        filters.append(sql.SQL("run_id=%s"))
+        params.append(run_id)
+    if start is not None:
+        filters.append(
+            sql.SQL(
+                "time >= %s AND time < %s"
+                if end_exclusive
+                else "time >= %s AND time <= %s"
+            )
+        )
+        params.extend((start, end))
+    else:
+        # Existing point-query semantics use +/- one second.
+        filters.append(
+            sql.SQL("time BETWEEN %s-interval '1 second' AND %s+interval '1 second'")
+        )
+        params.extend((time, time))
+    statement = sql.SQL(
+        "SELECT {} FROM {}.{} WHERE {} ORDER BY time,{} LIMIT %s OFFSET %s"
+    ).format(
+        columns,
+        sql.Identifier(source),
+        sql.Identifier(table),
+        sql.SQL(" AND ").join(filters),
+        sql.Identifier(id_column),
+    )
+    return await read_rows(conn, statement, tuple(params) + (limit, offset))
+
+
+async def sensor_runs(conn, *, run_id=None, limit=100, offset=0):
+    where = " AND r.run_id=%s" if run_id else ""
+    params = (run_id,) if run_id else ()
+    return await read_rows(
+        conn,
+        """SELECT r.run_id,r.name,r.created_by,r.created_at,r.status,
+        result.payload FROM analysis.runs r JOIN LATERAL
+        (SELECT payload FROM analysis.results WHERE run_id=r.run_id
+         AND result_type='sensor_placement' ORDER BY created_at DESC,result_id DESC LIMIT 1) result ON true
+        WHERE r.run_type='sensor_placement'"""
+        + where
+        + " ORDER BY r.created_at DESC,r.run_id LIMIT %s OFFSET %s",
+        params + (limit, offset),
+    )
+
+
+async def history_rows(
+    conn: AsyncConnection, descriptors: list[dict[str, Any]], *,
+    start: datetime, end: datetime, run_id: UUID | None, limit: int, offset: int,
+) -> list[dict[str, Any]]:
+    """Page points across raw, cleaned and model series in one SQL statement."""
+    parts, params = [], []
+    for index, item in enumerate(descriptors):
+        source, field = item["source"], item["field"]
+        if source == "scada":
+            table, id_column, fields = "measurements", "device_id", {"monitored_value", "cleaned_value"}
+        elif source in {"realtime", "analysis"}:
+            node = item["element_type"] == "junction"
+            table, id_column = ("node_results", "node_id") if node else ("link_results", "link_id")
+            fields = {"pressure"} if node else {"flow"}
+        else:
+            raise ValueError("unsupported history source")
+        if field not in fields or (source == "analysis" and run_id is None):
+            raise ValueError("unsupported history field or missing run")
+        parts.append(sql.SQL(
+            "SELECT %s::integer AS series_index,time,{} AS value FROM {}.{} "
+            "WHERE {}=%s AND time >= %s AND time <= %s{}"
+        ).format(
+            sql.Identifier(field), sql.Identifier(source), sql.Identifier(table),
+            sql.Identifier(id_column), sql.SQL(" AND run_id=%s" if source == "analysis" else ""),
+        ))
+        params.extend((index, item["id"], start, end))
+        if source == "analysis":
+            params.append(run_id)
+    if not parts:
+        return []
+    statement = sql.SQL(
+        "SELECT series_index,time,value FROM ({}) points "
+        "ORDER BY series_index, time LIMIT %s OFFSET %s"
+    ).format(sql.SQL(" UNION ALL ").join(parts))
+    return await read_rows(conn, statement, tuple(params) + (limit, offset))
+
+
+async def evidence_page(
+    conn: AsyncConnection, *, project_id: UUID, task_id: UUID, evidence_id: str, data_path: list[str],
+    limit: int, offset: int, mcp_sources: list[str], full_snapshot: bool = False,
+) -> dict[str, Any]:
+    """Project the immutable snapshot in PostgreSQL before transferring JSON."""
+    return await _snapshot_page(conn, stored_sql="""
+        SELECT document->'data_snapshot' AS snapshot
+        FROM workbench.operational_task_evidence
+        WHERE project_id=%s AND task_id=%s AND evidence_id=%s
+          AND jsonb_typeof(document->'data_snapshot')='object'
+        """, stored_params=(project_id, task_id, evidence_id), data_path=data_path,
+        limit=limit, offset=offset, mcp_sources=mcp_sources, full_snapshot=full_snapshot)
+
+
+async def _snapshot_page(
+    conn: AsyncConnection, *, stored_sql: str, stored_params: tuple, data_path: list[str],
+    limit: int, offset: int, mcp_sources: list[str], full_snapshot: bool = False,
+    max_bytes: int = MAX_QUERY_BYTES,
+) -> dict[str, Any]:
+    """Shared JSONB projection; callers supply only fixed, scoped SQL sources."""
+    def descriptor(value, path):
+        return f"""jsonb_build_object(
+            'collection_type',jsonb_typeof({value}),
+            'count',CASE jsonb_typeof({value}) WHEN 'array' THEN jsonb_array_length({value})
+                WHEN 'string' THEN char_length({value} #>> '{{}}')
+                ELSE (SELECT count(*) FROM jsonb_object_keys({value})) END,
+            'data_path',to_jsonb({path}))"""
+
+    collection = descriptor("p.value", "p.path")
+    child = descriptor("child.item", "p.path || ARRAY[child.key]")
+    statement = """
+        WITH stored AS MATERIALIZED (
+            STORED_QUERY
+        ), captured AS (
+            SELECT snapshot,
+                snapshot->>'source'=ANY(%s::text[])
+                    AND jsonb_typeof(snapshot->'data')='object' AS is_mcp
+            FROM stored
+        ), selected AS MATERIALIZED (
+            SELECT snapshot,
+                CASE WHEN is_mcp THEN (snapshot->'data')-'data' END AS query_metadata,
+                (CASE WHEN is_mcp THEN snapshot->'data'->'data'
+                      ELSE snapshot->'data' END) #> %s::text[] AS value
+            FROM captured
+        ), entries AS (
+            SELECT ordinality AS position, NULL::text AS key, item AS value
+            FROM selected, jsonb_array_elements(
+                CASE WHEN jsonb_typeof(value)='array' THEN value ELSE '[]'::jsonb END
+            ) WITH ORDINALITY AS a(item, ordinality)
+            UNION ALL
+            SELECT row_number() OVER (ORDER BY e.key COLLATE "C"), e.key, e.item
+            FROM selected, jsonb_each(
+                CASE WHEN jsonb_typeof(value)='object' THEN value ELSE '{}'::jsonb END
+            ) AS e(key,item)
+        ), paged AS MATERIALIZED (
+            SELECT * FROM entries ORDER BY position LIMIT %s OFFSET %s
+        ), paths AS (
+            SELECT paged.*, jsonb_typeof(selected.value) AS parent_type,
+                %s::text[] || ARRAY[CASE WHEN jsonb_typeof(selected.value)='array'
+                    THEN (position-1)::text ELSE key END] AS path
+            FROM paged,selected
+        ), summaries AS (
+            SELECT p.position,p.key,p.value AS original,
+                CASE WHEN jsonb_typeof(p.value) IN ('array','object') OR
+                    (jsonb_typeof(p.value)='string' AND char_length(p.value #>> '{}')>256) THEN
+                    CASE WHEN p.parent_type='object' OR jsonb_typeof(p.value) IN ('array','string')
+                        THEN COLLECTION_PREVIEW
+                        ELSE COALESCE((SELECT jsonb_object_agg(child.key,
+                            CASE WHEN jsonb_typeof(child.item) IN ('array','object') OR
+                                (jsonb_typeof(child.item)='string' AND char_length(child.item #>> '{}')>256)
+                                THEN CHILD_PREVIEW ELSE child.item END)
+                            FROM jsonb_each(p.value) AS child(key,item)), '{}'::jsonb)
+                    END ELSE p.value END AS value
+            FROM paths p
+        ), projected AS (
+            SELECT CASE WHEN %s THEN snapshot ELSE jsonb_build_object(
+                'snapshot_metadata',snapshot-'data', 'query_metadata',query_metadata,
+                'nested',EXISTS(SELECT 1 FROM summaries WHERE original<>value),
+                'value_type',jsonb_typeof(value),
+                'total',CASE jsonb_typeof(value)
+                    WHEN 'array' THEN jsonb_array_length(value)
+                    WHEN 'object' THEN (SELECT count(*) FROM jsonb_object_keys(value))
+                    WHEN 'string' THEN char_length(value #>> '{}') END,
+                'value',CASE jsonb_typeof(value)
+                    WHEN 'array' THEN COALESCE((SELECT jsonb_agg(value ORDER BY position) FROM summaries),'[]'::jsonb)
+                    WHEN 'object' THEN COALESCE((SELECT jsonb_object_agg(key,value) FROM summaries),'{}'::jsonb)
+                    WHEN 'string' THEN to_jsonb(substring(value #>> '{}' FROM %s FOR %s))
+                    ELSE value END
+                ) END AS payload, value IS NULL AS missing_path
+            FROM selected
+        ), sized AS MATERIALIZED (
+            SELECT payload,missing_path,octet_length(payload::text)>%s AS too_large
+            FROM projected
+        )
+        SELECT CASE WHEN NOT too_large THEN payload END AS payload,too_large,missing_path
+        FROM sized
+    """
+    statement = statement.replace("STORED_QUERY", stored_sql).replace("COLLECTION_PREVIEW", collection).replace("CHILD_PREVIEW", child)
+    rows = await read_rows(conn, statement, (
+        *stored_params, mcp_sources, data_path, limit, offset,
+        data_path, full_snapshot, offset + 1, limit, max_bytes,
+    ))
+    if not rows:
+        raise HTTPException(404, "Stored query document was not found")
+    row = rows[0]
+    if row["too_large"]:
+        raise HTTPException(413, "Query exceeds the byte budget; select a deeper data_path or smaller page")
+    if row.get("missing_path") and not full_snapshot:
+        raise HTTPException(422, "Evidence data_path does not name a stored value")
+    return row["payload"]
+
+
+async def task_context_summary(conn: AsyncConnection, *, project_id: UUID, task_id: UUID, report_revision: int | None) -> dict:
+    """Read fixed-size metadata and explicit previews, never a complete report."""
+    report = """jsonb_build_object('revision',r.revision,
+        'execution_generation',r.document->'execution_generation',
+        'stage',r.document#>'{report,stage}','outcome',r.document#>'{report,outcome}',
+        'current_conclusion_preview',left(r.document#>>'{report,current_conclusion}',200),
+        'preview_only',true,'detail',jsonb_build_object('section','report','report_revision',r.revision))"""
+    rows = await read_rows(conn, """
+        SELECT jsonb_build_object('task',jsonb_build_object(
+            'task_id',t.task_id,'task_type',t.task_type,'workflow',t.document->>'workflow','status',t.status,
+            'lifecycle_status',t.lifecycle_status,'priority',t.priority,'revision',t.revision,
+            'execution_generation',t.execution_generation,
+            'current_report_revision',t.current_report_revision,
+            'title_preview',left(t.document->>'title',160),'preview_only',true,
+            'detail',jsonb_build_object('section','task')),
+            'latest_report',(SELECT REPORT_SUMMARY FROM workbench.operational_task_reports r
+                WHERE r.project_id=t.project_id AND r.task_id=t.task_id ORDER BY r.revision DESC LIMIT 1),
+            'selected_report',CASE WHEN %s::int IS NULL THEN NULL ELSE
+                (SELECT REPORT_SUMMARY FROM workbench.operational_task_reports r
+                 WHERE r.project_id=t.project_id AND r.task_id=t.task_id AND r.revision=%s) END
+            ) AS payload
+        FROM workbench.operational_tasks t WHERE t.project_id=%s AND t.task_id=%s
+        """.replace("REPORT_SUMMARY", report), (report_revision, report_revision, project_id, task_id))
+    if not rows:
+        raise HTTPException(404, "Task was not found in this project")
+    result = rows[0]["payload"]
+    if report_revision is not None and result["selected_report"] is None:
+        raise HTTPException(404, "Selected report revision was not found")
+    return result
+
+
+async def task_context_page(
+    conn: AsyncConnection, *, project_id: UUID, task_id: UUID, section: str,
+    report_revision: int | None, data_path: list[str], limit: int, offset: int,
+) -> dict:
+    """Page only the requested task section inside PostgreSQL, before transfer."""
+    collections = {
+        "evidence": ("operational_task_evidence", "evidence_id"),
+        "actions": ("operational_task_actions", "action_id"),
+        "access_requests": ("operational_task_access_requests", "access_request_id"),
+    }
+    documents = {"task": "operational_tasks", "report": "operational_task_reports", "plan": "operational_task_plans"}
+    if section in {"access_requests", "progress"}:
+        table = "operational_task_access_requests" if section == "access_requests" else "operational_task_progress"
+        available = await read_rows(conn, "SELECT to_regclass(%s) AS relation", ("workbench." + table,))
+        if not available[0]["relation"]:
+            return {"snapshot_metadata": {"section": section, "available": False}, "value": None, "value_type": "null", "total": None, "nested": False}
+    projection_offset = offset
+    if section in collections:
+        table, key = collections[section]
+        # These identifiers come exclusively from the fixed table above.
+        relation = sql.Identifier("workbench", table).as_string()
+        identifier = sql.Identifier(key).as_string()
+        if data_path:
+            stored = f"""SELECT jsonb_build_object('section',%s::text,'data',jsonb_build_object({identifier}::text,document)) AS snapshot
+                FROM {relation} WHERE project_id=%s AND task_id=%s AND {identifier}::text=%s"""
+            params = (section, project_id, task_id, data_path[0])
+        else:
+            # Fetch just this page of rows. Collection keys remain stable across
+            # paging, so every nested data_path uses the real evidence/action ID.
+            stored = f"""WITH items AS MATERIALIZED (
+                SELECT {identifier}::text AS key,document FROM {relation}
+                WHERE project_id=%s AND task_id=%s
+                ORDER BY {identifier}::text COLLATE "C" LIMIT %s OFFSET %s)
+                SELECT jsonb_build_object('section',%s::text,
+                    'total',(SELECT count(*) FROM {relation} WHERE project_id=%s AND task_id=%s),
+                    'data',COALESCE(jsonb_object_agg(key,document),'{{}}'::jsonb)) AS snapshot FROM items"""
+            params = (project_id, task_id, limit + 1, offset, section, project_id, task_id)
+            projection_offset = 0
+    elif section in documents:
+        relation = sql.Identifier("workbench", documents[section]).as_string()
+        where = ""
+        params = (section, project_id, task_id)
+        if section == "report":
+            where = " AND (%s::int IS NULL OR revision=%s) ORDER BY revision DESC LIMIT 1"
+            params += (report_revision, report_revision)
+        elif section == "plan":
+            where = " ORDER BY revision DESC LIMIT 1"
+        stored = f"SELECT jsonb_build_object('section',%s::text,'data',document) AS snapshot FROM {relation} WHERE project_id=%s AND task_id=%s{where}"
+    elif section == "run":
+        stored = """SELECT jsonb_build_object('section','run','data',to_jsonb(r)-'project_id') AS snapshot
+            FROM workbench.operational_task_runs r WHERE project_id=%s AND task_id=%s ORDER BY generation DESC LIMIT 1"""
+        params = (project_id, task_id)
+    elif section == "progress":
+        stored = """SELECT jsonb_build_object('section','progress','data',p.document) AS snapshot
+            FROM workbench.operational_task_progress p
+            JOIN workbench.operational_task_runs r ON r.project_id=p.project_id AND r.task_id=p.task_id AND r.run_id=p.run_id
+            JOIN workbench.operational_tasks t ON t.project_id=p.project_id AND t.task_id=p.task_id AND t.execution_generation=r.generation
+            WHERE p.project_id=%s AND p.task_id=%s"""
+        params = (project_id, task_id)
+    else:
+        raise HTTPException(422, "Unknown task context section")
+    return await _snapshot_page(conn, stored_sql=stored, stored_params=params,
+        data_path=data_path, limit=limit, offset=projection_offset, mcp_sources=[], max_bytes=8000)
+
+
+async def task_data_scopes(conn: AsyncConnection, *, project_id: UUID, task_id: UUID, permissions: set[str]) -> list[str]:
+    """Bound scope discovery to seven known sources, without loading task detail."""
+    known = ["workbench_runtime", "peripheral_points", "peripheral_work_orders", "peripheral_observations", "peripheral_test_data", "scada_devices", "scada_readings"]
+    scopes = {name for name in known if ("scada.view" if name.startswith("scada_") else "simulation.view") in permissions}
+    available = await read_rows(conn, "SELECT to_regclass('workbench.operational_task_access_requests') AS relation")
+    if available[0]["relation"]:
+        rows = await read_rows(conn, """SELECT DISTINCT scope FROM workbench.operational_task_access_requests,
+            LATERAL jsonb_array_elements_text(COALESCE(document->'approved_scopes','[]'::jsonb)) AS granted(scope)
+            WHERE project_id=%s AND task_id=%s AND status='approved' AND scope=ANY(%s::text[])""", (project_id, task_id, known))
+        scopes.update(row["scope"] for row in rows)
+    return [name for name in known if name in scopes]

app/infra/db/postgresql/scada.py

@@ -111,14 +111,16 @@ class ScadaInfoRepository:
         conn: AsyncConnection,
         node_ids: list[str],
         link_ids: list[str],
+        *, row_limit: int | None = None,
     ) -> list[dict[str, Any]]:
         if not node_ids and not link_ids:
             return []
         async with conn.cursor() as cur:
             await cur.execute(
                 _SCADA_VIEW_SELECT
-                + " WHERE node_id = ANY(%s) OR link_id = ANY(%s) ORDER BY device_id",
-                (node_ids, link_ids),
+                + " WHERE node_id = ANY(%s) OR link_id = ANY(%s) ORDER BY device_id"
+                + (" LIMIT %s" if row_limit is not None else ""),
+                (node_ids, link_ids) + ((row_limit,) if row_limit is not None else ()),
             )
             return [_device(record) for record in await cur.fetchall()]
 

app/infra/db/postgresql/sensor_placement.py

@@ -24,6 +24,8 @@ def _placement_row(row: dict[str, Any]) -> dict[str, Any]:
         "created_by": row["created_by"],
         "created_at": row["created_at"],
         "status": row["status"],
+        "sensor_type": payload.get("sensor_type", "pressure"),
+        "method": payload.get("method"),
         "sensor_locations": locations,
     }
 
@@ -57,11 +59,15 @@ def create_sensor_placement(
     min_diameter: int,
     created_by: str,
     sensor_locations: list[str],
+    sensor_type: str = "pressure",
+    method: str | None = None,
 ) -> dict[str, Any]:
     run_id = uuid4()
     payload = {
         "sensor_number": len(sensor_locations),
         "minimum_diameter": min_diameter,
+        "sensor_type": sensor_type,
+        "method": method,
         "sensor_locations": sensor_locations,
     }
     with project_connection(name) as conn, conn.transaction():

app/infra/db/postgresql/spatial_queries.py

@@ -0,0 +1,71 @@
+"""Bounded nearest-model queries on the already authorized project connection."""
+
+from math import asin, cos, degrees, radians, sin
+from typing import Any
+
+from psycopg import AsyncConnection, sql
+
+from app.infra.db.postgresql.read_queries import ELEMENT_VIEWS, read_rows
+
+_MODEL_RELATIONS = {
+    kind: sql.Identifier("gis", view) for kind, view in ELEMENT_VIEWS.items()
+}
+
+
+def _search_boxes(
+    longitude: float, latitude: float, radius_m: float,
+) -> list[tuple[float, float, float, float]]:
+    # A deliberately smaller-than-WGS84 curvature radius gives a conservative
+    # geographic envelope. It is only a GiST candidate filter; geography below
+    # determines the actual radius and distances. Split at the antimeridian.
+    angle = (radius_m + 10) / 6_300_000
+    low, high = latitude - degrees(angle), latitude + degrees(angle)
+    if low < -85 or high > 85:
+        # Avoid Mercator's polar singularity; retain exact geography filtering.
+        return []
+    width = degrees(asin(min(1.0, sin(angle) / cos(radians(latitude)))))
+    left, right = longitude - width, longitude + width
+    if left < -180:
+        return [(-180, low, right, high), (left + 360, low, 180, high)]
+    if right > 180:
+        return [(left, low, 180, high), (-180, low, right - 360, high)]
+    return [(left, low, right, high)]
+
+
+async def nearby_elements(
+    conn: AsyncConnection,
+    element_type: str,
+    *,
+    center: tuple[float, float],
+    radius_m: float,
+    limit: int,
+    offset: int,
+) -> list[dict[str, Any]]:
+    """Rank complete model geometries by spheroidal distance, never endpoints."""
+    relation = _MODEL_RELATIONS[element_type]
+    filters = []
+    parameters: list[Any] = [*center]
+    for box in _search_boxes(*center, radius_m):
+        filters.append(sql.SQL(
+            "v.geom && ST_Transform(ST_MakeEnvelope(%s,%s,%s,%s,4326),3857)"
+        ))
+        parameters.extend(box)
+    candidate_filter = sql.SQL(" OR ").join(filters) if filters else sql.SQL("TRUE")
+    statement = sql.SQL("""
+        WITH target AS (
+            SELECT ST_SetSRID(ST_MakePoint(%s,%s),4326)::geography AS point
+        ), candidates AS MATERIALIZED (
+            SELECT v.id, to_jsonb(v)-'geom' AS properties,
+                   ST_Transform(v.geom,4326)::geography AS shape
+            FROM {} v
+            WHERE v.geom IS NOT NULL AND NOT ST_IsEmpty(v.geom) AND ({})
+        )
+        SELECT c.id AS element_id, c.properties,
+               ST_Distance(c.shape,t.point,true) AS distance_m
+        FROM candidates c CROSS JOIN target t
+        WHERE ST_DWithin(c.shape,t.point,%s,true)
+        ORDER BY distance_m,c.id
+        LIMIT %s OFFSET %s
+    """).format(relation, candidate_filter)
+    parameters.extend((radius_m, limit, offset))
+    return await read_rows(conn, statement, tuple(parameters))

app/infra/db/postgresql/task_alerts.py

@@ -0,0 +1,109 @@
+"""PostgreSQL persistence for task-produced alert events and current state."""
+from __future__ import annotations
+
+from contextlib import asynccontextmanager
+
+from psycopg import AsyncConnection
+from psycopg.errors import UndefinedTable
+from psycopg.types.json import Jsonb
+from uuid import uuid4
+
+from app.domain.domain_errors import DomainError
+from app.infra.db.postgresql.workbench_events import lock_event_transaction
+
+
+class TaskAlertRepository:
+    def __init__(self, connection: AsyncConnection, project_id):
+        self.connection = connection
+        self.project_id = str(project_id)
+
+    @asynccontextmanager
+    async def transaction(self):
+        try:
+            async with self.connection.transaction():
+                await lock_event_transaction(self.connection)
+                yield self
+        except UndefinedTable as exc:
+            raise DomainError(
+                "DEPENDENCY_UNAVAILABLE",
+                "Project alert storage is not provisioned",
+                status=503,
+            ) from exc
+
+    async def lock_alert(self, alert_id: str) -> None:
+        await self.connection.execute(
+            "SELECT pg_advisory_xact_lock(hashtextextended(%s, 0))",
+            (f"tjwater-alert:{self.project_id}:{alert_id}",),
+        )
+
+    async def get_event(self, producer_id: str, event_id: str) -> dict | None:
+        cursor = await self.connection.execute(
+            """SELECT document FROM workbench.alert_events
+               WHERE project_id=%s AND producer_id=%s AND event_id=%s""",
+            (self.project_id, producer_id, event_id),
+        )
+        row = await cursor.fetchone()
+        return row["document"] if row else None
+
+    async def get_alert(self, alert_id: str) -> tuple[int, dict] | None:
+        cursor = await self.connection.execute(
+            """SELECT revision,document FROM workbench.alerts
+               WHERE project_id=%s AND alert_id=%s""",
+            (self.project_id, alert_id),
+        )
+        row = await cursor.fetchone()
+        return (row["revision"], row["document"]) if row else None
+
+    async def save_event(self, event: dict, resulting_state: dict) -> None:
+        document = event | {"resulting_state": resulting_state}
+        await self.connection.execute(
+            """INSERT INTO workbench.alert_events
+               (project_id,producer_id,event_id,alert_id,event_type,occurred_at,document)
+               VALUES (%s,%s,%s,%s,%s,%s,%s)""",
+            (
+                self.project_id,
+                event["producer_id"],
+                event["event_id"],
+                event["alert_id"],
+                event["event_type"],
+                event["occurred_at"],
+                Jsonb(document),
+            ),
+        )
+
+    async def save_alert(self, state: dict) -> None:
+        await self.connection.execute(
+            """INSERT INTO workbench.alerts
+               (project_id,alert_id,revision,status,occurred_at,updated_at,document)
+               VALUES (%s,%s,%s,%s,%s,%s,%s)
+               ON CONFLICT(project_id,alert_id) DO UPDATE SET
+                 revision=excluded.revision,status=excluded.status,
+                 updated_at=excluded.updated_at,document=excluded.document""",
+            (
+                self.project_id,
+                state["alert_id"],
+                state["revision"],
+                state["status"],
+                state["occurred_at"],
+                state["updated_at"],
+                Jsonb({key: value for key, value in state.items() if key != "revision"}),
+            ),
+        )
+
+    async def append_changed_event(self, state: dict) -> None:
+        cursor = await self.connection.execute(
+            """INSERT INTO workbench.domain_event_log
+               (project_id,event_id,topic,aggregate_type,aggregate_id,
+                aggregate_revision,occurred_at,metadata)
+               VALUES (%s,%s,'workbench.alert.changed','workbench_alert',%s,%s,%s,%s)
+               RETURNING event_seq""",
+            (
+                self.project_id, uuid4(), state["alert_id"], state["revision"],
+                state["updated_at"], Jsonb({"alert_id": state["alert_id"]}),
+            ),
+        )
+        row = await cursor.fetchone()
+        await self.connection.execute(
+            "SELECT pg_notify('workbench_events', %s)",
+            (f'{self.project_id}:{row["event_seq"]}',),
+        )

app/infra/db/postgresql/workbench_event_hub.py

@@ -0,0 +1,221 @@
+"""Per-process, project-scoped PostgreSQL LISTEN hubs for workbench SSE."""
+from __future__ import annotations
+
+import asyncio
+import contextlib
+import logging
+from collections import OrderedDict
+from contextlib import asynccontextmanager
+from uuid import UUID
+
+from psycopg import AsyncConnection
+from psycopg_pool import AsyncConnectionPool
+from psycopg.rows import dict_row
+
+from app.infra.db.postgresql.workbench_events import WorkbenchEventRepository
+
+
+logger = logging.getLogger(__name__)
+
+
+class WorkbenchEventHub:
+    """Own one LISTEN connection and fan wake-ups out to SSE subscribers."""
+
+    def __init__(self, project_id: UUID, dsn: str) -> None:
+        self.project_id = project_id
+        self.dsn = dsn
+        self.subscribers = 0
+        self._connection: AsyncConnection | None = None
+        self._pool: AsyncConnectionPool | None = None
+        self._connection_lock = asyncio.Lock()
+        self._read_lock = asyncio.Lock()
+        self._reads: OrderedDict[tuple, tuple[int, object]] = OrderedDict()
+        self._condition = asyncio.Condition()
+        self._generation = 0
+        self._ready = asyncio.Event()
+        self._closed = False
+        self._listener_task: asyncio.Task | None = None
+        self._idle_task: asyncio.Task | None = None
+
+    async def start(self) -> None:
+        if self._listener_task is None or self._listener_task.done():
+            self._closed = False
+            self._ready.clear()
+            self._listener_task = asyncio.create_task(
+                self._listen_loop(), name=f"workbench-events:{self.project_id}"
+            )
+        await asyncio.wait_for(self._ready.wait(), timeout=10)
+
+    async def _connect(self) -> AsyncConnection:
+        if self._pool is None:
+            self._pool = AsyncConnectionPool(
+                conninfo=self.dsn,
+                min_size=0,
+                max_size=1,
+                open=False,
+                kwargs={"autocommit": True, "row_factory": dict_row},
+            )
+            await self._pool.open()
+        connection = await self._pool.getconn()
+        await connection.execute("LISTEN workbench_events")
+        return connection
+
+    async def _listen_loop(self) -> None:
+        retry_seconds = 1
+        next_reconcile = asyncio.get_running_loop().time() + 30
+        while not self._closed:
+            try:
+                if self._connection is None or self._connection.closed:
+                    async with self._connection_lock:
+                        if self._connection is None or self._connection.closed:
+                            self._connection = await self._connect()
+                            self._ready.set()
+                    retry_seconds = 1
+                notified = False
+                async with self._connection_lock:
+                    connection = self._connection
+                    if connection is None:
+                        continue
+                    async for notification in connection.notifies(timeout=1, stop_after=1):
+                        if notification.payload.startswith(f"{self.project_id}:"):
+                            notified = True
+                        break
+                now = asyncio.get_running_loop().time()
+                if notified or now >= next_reconcile:
+                    await self._wake()
+                if now >= next_reconcile:
+                    next_reconcile = now + 30
+                await asyncio.sleep(0)
+            except asyncio.CancelledError:
+                raise
+            except Exception:
+                logger.exception("Workbench event listener failed for project %s", self.project_id)
+                await self._drop_connection()
+                await self._wake()
+                await asyncio.sleep(retry_seconds)
+                retry_seconds = min(retry_seconds * 2, 30)
+
+    async def _wake(self) -> None:
+        async with self._condition:
+            self._generation += 1
+            self._condition.notify_all()
+
+    async def wait_for_wake(self, generation: int, timeout: float) -> int:
+        async with self._condition:
+            if self._generation != generation:
+                return self._generation
+            with contextlib.suppress(asyncio.TimeoutError):
+                await asyncio.wait_for(
+                    self._condition.wait_for(lambda: self._generation != generation),
+                    timeout=timeout,
+                )
+            return self._generation
+
+    @property
+    def generation(self) -> int:
+        return self._generation
+
+    async def oldest_cursor(self) -> int | None:
+        return await self._read(("oldest",), lambda repository: repository.oldest_cursor())
+
+    async def current_cursor(self) -> int:
+        return await self._read(("current",), lambda repository: repository.current_cursor())
+
+    async def read_after(self, after: int) -> list[dict]:
+        return await self._read(("after", after), lambda repository: repository.read_after(after))
+
+    async def _read(self, key: tuple, operation):
+        await self.start()
+        async with self._read_lock:
+            cached = self._reads.get(key)
+            if cached is not None and cached[0] == self.generation:
+                self._reads.move_to_end(key)
+                return cached[1]
+            # Record the generation before the query: a concurrent NOTIFY must
+            # invalidate even an empty result from an older database snapshot.
+            generation = self.generation
+            async with self._connection_lock:
+                if self._connection is None or self._connection.closed:
+                    self._connection = await self._connect()
+                    self._ready.set()
+                repository = WorkbenchEventRepository(self._connection, self.project_id)
+                result = await operation(repository)
+            self._reads[key] = (generation, result)
+            self._reads.move_to_end(key)
+            while len(self._reads) > 32:
+                self._reads.popitem(last=False)
+            return result
+
+    async def _drop_connection(self) -> None:
+        async with self._connection_lock:
+            connection, self._connection = self._connection, None
+            if connection is not None and self._pool is not None:
+                with contextlib.suppress(Exception):
+                    await self._pool.putconn(connection)
+
+    async def close(self) -> None:
+        self._closed = True
+        self._reads.clear()
+        if self._idle_task is not None:
+            self._idle_task.cancel()
+            self._idle_task = None
+        if self._listener_task is not None:
+            self._listener_task.cancel()
+            with contextlib.suppress(asyncio.CancelledError):
+                await self._listener_task
+            self._listener_task = None
+        await self._drop_connection()
+        if self._pool is not None:
+            await self._pool.close()
+            self._pool = None
+
+
+class WorkbenchEventHubManager:
+    def __init__(self, idle_seconds: float = 60) -> None:
+        self.idle_seconds = idle_seconds
+        self._hubs: dict[tuple[UUID, str], WorkbenchEventHub] = {}
+        self._lock = asyncio.Lock()
+
+    @asynccontextmanager
+    async def subscribe(self, project_id: UUID, dsn: str):
+        key = (project_id, dsn)
+        async with self._lock:
+            hub = self._hubs.get(key)
+            if hub is None:
+                hub = WorkbenchEventHub(project_id, dsn)
+                self._hubs[key] = hub
+            if hub._idle_task is not None:
+                hub._idle_task.cancel()
+                hub._idle_task = None
+            hub.subscribers += 1
+        try:
+            await hub.start()
+            yield hub
+        finally:
+            async with self._lock:
+                hub.subscribers = max(0, hub.subscribers - 1)
+                if hub.subscribers == 0 and hub._idle_task is None:
+                    hub._idle_task = asyncio.create_task(self._close_when_idle(key, hub))
+
+    async def _close_when_idle(
+        self, key: tuple[UUID, str], hub: WorkbenchEventHub
+    ) -> None:
+        try:
+            await asyncio.sleep(self.idle_seconds)
+            async with self._lock:
+                if hub.subscribers or self._hubs.get(key) is not hub:
+                    return
+                self._hubs.pop(key, None)
+                hub._idle_task = None
+            await hub.close()
+        except asyncio.CancelledError:
+            return
+
+    async def close_all(self) -> None:
+        async with self._lock:
+            hubs = list(self._hubs.values())
+            self._hubs.clear()
+        await asyncio.gather(*(hub.close() for hub in hubs), return_exceptions=True)
+
+
+workbench_event_hubs = WorkbenchEventHubManager()

app/infra/db/postgresql/workbench_events.py

@@ -0,0 +1,61 @@
+"""Cursor reads for the durable project-scoped workbench event log."""
+from __future__ import annotations
+
+from psycopg import AsyncConnection, Connection
+
+
+EVENT_ORDER_LOCK = "tjwater-workbench:event-order"
+
+
+async def lock_event_transaction(connection: AsyncConnection) -> None:
+    """Serialize v2 event-producing transactions before business locks/writes.
+
+    PostgreSQL sequence allocation is not commit order. All producers must use
+    this same transaction lock so a committed higher cursor cannot hide an
+    earlier, still-uncommitted event. Other BizDBs remain rollout opt-out.
+    """
+    if getattr(getattr(connection, "info", None), "dbname", None) == "tjwater_v2":
+        await connection.execute(
+            "SELECT pg_advisory_xact_lock(hashtextextended(%s,0))", (EVENT_ORDER_LOCK,)
+        )
+
+
+def lock_event_transaction_sync(connection: Connection) -> None:
+    """Synchronous counterpart using the identical PostgreSQL lock key."""
+    if getattr(getattr(connection, "info", None), "dbname", None) == "tjwater_v2":
+        connection.execute(
+            "SELECT pg_advisory_xact_lock(hashtextextended(%s,0))", (EVENT_ORDER_LOCK,)
+        )
+
+
+class WorkbenchEventRepository:
+    def __init__(self, connection: AsyncConnection, project_id):
+        self.connection = connection
+        self.project_id = str(project_id)
+
+    async def current_cursor(self) -> int:
+        cursor = await self.connection.execute(
+            "SELECT COALESCE(max(event_seq),0) AS cursor FROM workbench.domain_event_log WHERE project_id=%s",
+            (self.project_id,),
+        )
+        row = await cursor.fetchone()
+        return int(row["cursor"])
+
+    async def oldest_cursor(self) -> int | None:
+        cursor = await self.connection.execute(
+            "SELECT min(event_seq) AS cursor FROM workbench.domain_event_log WHERE project_id=%s",
+            (self.project_id,),
+        )
+        row = await cursor.fetchone()
+        return int(row["cursor"]) if row and row["cursor"] is not None else None
+
+    async def read_after(self, after: int, limit: int = 200) -> list[dict]:
+        cursor = await self.connection.execute(
+            """SELECT event_seq,event_id,project_id,topic,aggregate_type,aggregate_id,
+                      aggregate_revision,occurred_at,metadata
+               FROM workbench.domain_event_log
+               WHERE project_id=%s AND event_seq>%s
+               ORDER BY event_seq LIMIT %s""",
+            (self.project_id, after, limit),
+        )
+        return [dict(row) for row in await cursor.fetchall()]

app/infra/db/postgresql/workbench_operations.py

@@ -0,0 +1,49 @@
+"""Optional project-local alert projection."""
+
+from dataclasses import dataclass
+from typing import Any
+
+from psycopg import AsyncConnection
+
+
+@dataclass(frozen=True)
+class WorkbenchOperationsAvailability:
+    alerts: bool
+    schedules: bool = False
+
+
+class WorkbenchOperationsRepository:
+    """Read provisioned workbench records without creating or mutating storage."""
+
+    @staticmethod
+    async def availability(
+        conn: AsyncConnection,
+    ) -> WorkbenchOperationsAvailability:
+        async with conn.cursor() as cur:
+            await cur.execute(
+                """
+                SELECT to_regclass('workbench.alerts') IS NOT NULL AS alerts
+                """
+            )
+            row = await cur.fetchone()
+        return WorkbenchOperationsAvailability(
+            alerts=bool(row and row["alerts"]),
+        )
+
+    @staticmethod
+    async def list_alerts(
+        conn: AsyncConnection, project_id: str, *, limit: int = 100
+    ) -> list[dict[str, Any]]:
+        async with conn.cursor() as cur:
+            await cur.execute(
+                """
+                SELECT document
+                FROM workbench.alerts
+                WHERE project_id = %s
+                  AND status = 'active'
+                ORDER BY occurred_at DESC, alert_id
+                LIMIT %s
+                """,
+                (project_id, limit),
+            )
+            return [row["document"] for row in await cur.fetchall()]

app/infra/db/project_routing.py

@@ -4,6 +4,7 @@ from contextlib import contextmanager
 from contextvars import ContextVar, Token
 from dataclasses import dataclass
 from typing import Iterator
+from uuid import UUID
 
 from psycopg.conninfo import conninfo_to_dict, make_conninfo
 
@@ -15,6 +16,7 @@ class ActiveProjectRouting:
     project_code: str
     business_dsn: str
     timescale_dsn: str | None = None
+    project_id: UUID | None = None
 
     @property
     def business_database_name(self) -> str:

app/infra/db/timescaledb/repositories/analysis.py

@@ -184,6 +184,7 @@ class AnalysisResultsRepository:
         start_time: datetime,
         end_time: datetime,
         field: str,
+        *, row_limit: int | None = None,
     ) -> dict[str, list[dict[str, Any]]]:
         if element_type == "node":
             table_name, id_column, valid_fields = (
@@ -214,8 +215,10 @@ class AnalysisResultsRepository:
             sql.Identifier(id_column),
             sql.Identifier(id_column),
         )
+        if row_limit is not None:
+            query += sql.SQL(" LIMIT %s")
         async with conn.cursor() as cur:
-            await cur.execute(query, (run_id, element_ids, start_time, end_time))
+            await cur.execute(query, (run_id, element_ids, start_time, end_time) + ((row_limit,) if row_limit is not None else ()))
             for row in await cur.fetchall():
                 result.setdefault(str(row["element_id"]), []).append(
                     {"time": row["time"], "value": row["value"]}

app/infra/db/timescaledb/repositories/realtime.py

@@ -3,9 +3,28 @@ from datetime import datetime, timedelta
 from collections import defaultdict
 from psycopg import AsyncConnection, Connection, sql
 from app.domain.time import parse_utc_time
+from app.core.config import settings
 
 
 class RealtimeRepository:
+    # The third set_config argument makes this transaction-local, so a pooled
+    # connection returns to its original guard after commit or rollback.
+    SNAPSHOT_BUDGET_SQL = "SELECT set_config('timescaledb.max_tuples_decompressed_per_dml_transaction', %s, true)"
+
+    @staticmethod
+    async def _configure_snapshot_write(cur) -> None:
+        await cur.execute(
+            RealtimeRepository.SNAPSHOT_BUDGET_SQL,
+            (str(settings.TIMESCALEDB_REALTIME_MAX_DECOMPRESSED_TUPLES),),
+        )
+
+    @staticmethod
+    def _configure_snapshot_write_sync(cur) -> None:
+        cur.execute(
+            RealtimeRepository.SNAPSHOT_BUDGET_SQL,
+            (str(settings.TIMESCALEDB_REALTIME_MAX_DECOMPRESSED_TUPLES),),
+        )
+
     LINK_FIELDS = frozenset(
         {
             "flow",
@@ -93,6 +112,7 @@ class RealtimeRepository:
         # 使用事务确保原子性
         async with conn.transaction():
             async with conn.cursor() as cur:
+                await RealtimeRepository._configure_snapshot_write(cur)
                 await cur.execute(
                     "SELECT pg_advisory_xact_lock(hashtextextended(%s::text, 1))",
                     (target_time,),
@@ -116,6 +136,7 @@ class RealtimeRepository:
         # 使用事务确保原子性
         with conn.transaction():
             with conn.cursor() as cur:
+                RealtimeRepository._configure_snapshot_write_sync(cur)
                 cur.execute(
                     "SELECT pg_advisory_xact_lock(hashtextextended(%s::text, 1))",
                     (target_time,),
@@ -163,6 +184,7 @@ class RealtimeRepository:
         end_time: datetime,
         link_ids: list[str],
         field: str,
+        *, row_limit: int | None = None,
     ) -> dict[str, list[dict[str, Any]]]:
         if field not in RealtimeRepository.LINK_FIELDS:
             raise ValueError(f"Invalid field: {field}")
@@ -174,8 +196,10 @@ class RealtimeRepository:
             "WHERE time BETWEEN %s AND %s AND link_id = ANY(%s) "
             "ORDER BY link_id, time"
         ).format(sql.Identifier(field))
+        if row_limit is not None:
+            query += sql.SQL(" LIMIT %s")
         async with conn.cursor() as cur:
-            await cur.execute(query, (start_time, end_time, link_ids))
+            await cur.execute(query, (start_time, end_time, link_ids) + ((row_limit,) if row_limit is not None else ()))
             for row in await cur.fetchall():
                 result.setdefault(str(row["link_id"]), []).append(
                     {"time": row["time"].isoformat(), "value": row[field]}
@@ -282,6 +306,7 @@ class RealtimeRepository:
         # 使用事务确保原子性
         async with conn.transaction():
             async with conn.cursor() as cur:
+                await RealtimeRepository._configure_snapshot_write(cur)
                 await cur.execute(
                     "SELECT pg_advisory_xact_lock(hashtextextended(%s::text, 1))",
                     (target_time,),
@@ -304,6 +329,7 @@ class RealtimeRepository:
         # 使用事务确保原子性
         with conn.transaction():
             with conn.cursor() as cur:
+                RealtimeRepository._configure_snapshot_write_sync(cur)
                 cur.execute(
                     "SELECT pg_advisory_xact_lock(hashtextextended(%s::text, 1))",
                     (target_time,),
@@ -351,6 +377,7 @@ class RealtimeRepository:
         end_time: datetime,
         node_ids: list[str],
         field: str,
+        *, row_limit: int | None = None,
     ) -> dict[str, list[dict[str, Any]]]:
         if field not in RealtimeRepository.NODE_FIELDS:
             raise ValueError(f"Invalid field: {field}")
@@ -362,8 +389,10 @@ class RealtimeRepository:
             "WHERE time BETWEEN %s AND %s AND node_id = ANY(%s) "
             "ORDER BY node_id, time"
         ).format(sql.Identifier(field))
+        if row_limit is not None:
+            query += sql.SQL(" LIMIT %s")
         async with conn.cursor() as cur:
-            await cur.execute(query, (start_time, end_time, node_ids))
+            await cur.execute(query, (start_time, end_time, node_ids) + ((row_limit,) if row_limit is not None else ()))
             for row in await cur.fetchall():
                 result.setdefault(str(row["node_id"]), []).append(
                     {"time": row["time"].isoformat(), "value": row[field]}
@@ -482,6 +511,7 @@ class RealtimeRepository:
         # table. Copy-only helpers avoid repeating replacement SQL.
         async with conn.transaction():
             async with conn.cursor() as cur:
+                await RealtimeRepository._configure_snapshot_write(cur)
                 await cur.execute(
                     "SELECT pg_advisory_xact_lock(hashtextextended(%s::text, 1))",
                     (simulation_time,),
@@ -559,6 +589,7 @@ class RealtimeRepository:
         # table. Copy-only helpers avoid repeating replacement SQL.
         with conn.transaction():
             with conn.cursor() as cur:
+                RealtimeRepository._configure_snapshot_write_sync(cur)
                 cur.execute(
                     "SELECT pg_advisory_xact_lock(hashtextextended(%s::text, 1))",
                     (simulation_time,),

app/infra/db/timescaledb/repositories/scada.py

@@ -34,13 +34,15 @@ class ScadaRepository:
         device_ids: List[str],
         start_time: datetime,
         end_time: datetime,
+        *, row_limit: int | None = None,
     ) -> List[dict]:
         async with conn.cursor() as cur:
             await cur.execute(
                 f"SELECT {ScadaRepository.RESULT_COLUMNS} FROM scada.measurements "
                 "WHERE device_id = ANY(%s) "
-                "AND time >= %s AND time <= %s ORDER BY device_id, time",
-                (device_ids, start_time, end_time),
+                "AND time >= %s AND time <= %s ORDER BY device_id, time"
+                + (" LIMIT %s" if row_limit is not None else ""),
+                (device_ids, start_time, end_time) + ((row_limit,) if row_limit is not None else ()),
             )
             return await cur.fetchall()
 

app/infra/geoserver/client.py

@@ -94,6 +94,89 @@ class GeoServerAdminClient:
             )
         return True
 
+    def feature_type_exists(
+        self,
+        *,
+        workspace: str,
+        datastore: str,
+        layer: str,
+    ) -> bool:
+        response = self._client.get(
+            f"/rest/workspaces/{workspace}/datastores/{datastore}/featuretypes/{layer}.json"
+        )
+        if response.status_code == 404:
+            return False
+        if response.is_error:
+            raise GeoServerProvisioningError(
+                f"GeoServer feature type check failed with HTTP {response.status_code}"
+            )
+        return True
+
+    def publish_feature_type(
+        self,
+        *,
+        workspace: str,
+        datastore: str,
+        layer: str,
+        title: str,
+        map_bbox: tuple[float, float, float, float],
+    ) -> None:
+        minx, miny, maxx, maxy = map_bbox
+        payload = {
+            "featureType": {
+                "name": layer,
+                "nativeName": layer,
+                "title": title,
+                "srs": "EPSG:3857",
+                "projectionPolicy": "FORCE_DECLARED",
+                "enabled": True,
+                "advertised": True,
+                "nativeBoundingBox": {
+                    "minx": minx,
+                    "miny": miny,
+                    "maxx": maxx,
+                    "maxy": maxy,
+                    "crs": "EPSG:3857",
+                },
+                "latLonBoundingBox": {
+                    "minx": _longitude(minx),
+                    "miny": _latitude(miny),
+                    "maxx": _longitude(maxx),
+                    "maxy": _latitude(maxy),
+                    "crs": "EPSG:4326",
+                },
+            }
+        }
+        if self.feature_type_exists(
+            workspace=workspace,
+            datastore=datastore,
+            layer=layer,
+        ):
+            self._request(
+                "PUT",
+                f"/rest/workspaces/{workspace}/datastores/{datastore}/featuretypes/{layer}",
+                json=payload,
+            )
+        else:
+            self._request(
+                "POST",
+                f"/rest/workspaces/{workspace}/datastores/{datastore}/featuretypes",
+                json=payload,
+            )
+        self._set_client_cache(workspace, layer)
+
+    def truncate_layer_cache(self, workspace: str, layer: str) -> None:
+        self._request(
+            "POST",
+            "/gwc/rest/masstruncate",
+            content=(
+                "<truncateLayer><layerName>"
+                f"{workspace}:{layer}"
+                "</layerName></truncateLayer>"
+            ),
+            headers={"Content-Type": "text/xml"},
+        )
+
     def create_project_workspace(
         self,
         *,

app/infra/integrations/operational_actions.py

@@ -0,0 +1,218 @@
+"""Typed external action adapters. Credentials are request-scoped, never persisted."""
+from __future__ import annotations
+
+from dataclasses import dataclass
+from datetime import datetime, timezone
+from pathlib import Path
+from typing import Literal, Protocol
+from uuid import UUID
+
+import httpx
+from pydantic import BaseModel, ConfigDict, Field, ValidationError
+from pydantic_settings import BaseSettings, SettingsConfigDict
+
+from app.auth.project_dependencies import ProjectContext
+from app.domain.schemas.operational_task import (
+    ActionReceipt, CreateWorkOrderPayload, HumanApprovalAction,
+    IssueCommandPayload, OperationalTaskDetail,
+)
+
+
+@dataclass(frozen=True)
+class ActionDispatchOutcome:
+    status: Literal["submitted", "succeeded", "failed", "unsupported", "unknown"]
+    detail: str
+    receipt: ActionReceipt | None = None
+
+
+class OperationalActionDispatcher(Protocol):
+    async def dispatch(
+        self, *, context: ProjectContext, task: OperationalTaskDetail,
+        action: HumanApprovalAction, credential: str | None,
+    ) -> ActionDispatchOutcome: ...
+
+    async def reconcile(
+        self, *, context: ProjectContext, task: OperationalTaskDetail,
+        action: HumanApprovalAction, credential: str | None,
+    ) -> ActionDispatchOutcome: ...
+
+
+class OperationalActionSettings(BaseSettings):
+    peripheral_url: str = ""
+    command_gateway_url: str = ""
+    timeout_seconds: float = Field(default=10, ge=1, le=60)
+    model_config = SettingsConfigDict(
+        env_prefix="OPERATIONAL_ACTION_", extra="ignore",
+        env_file=Path(__file__).resolve().parents[3] / ".env",
+    )
+
+
+class SampleWorkOrderReceipt(BaseModel):
+    model_config = ConfigDict(extra="ignore")
+    work_order_id: UUID
+    project_id: UUID
+    created_at: datetime
+    origin: Literal["synthetic"]
+    status: Literal[
+        "draft", "open", "in_progress", "awaiting_reply", "completed", "cancelled"
+    ]
+
+
+class CommandGatewayReceipt(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+    request_id: UUID
+    project_id: UUID
+    receipt_id: str = Field(min_length=1)
+    status: Literal["accepted", "running", "completed", "failed"]
+    submitted_at: datetime
+    completed_at: datetime | None = None
+    detail: str = Field(min_length=1)
+
+
+class HttpOperationalActionDispatcher:
+    def __init__(
+        self, settings: OperationalActionSettings | None = None,
+        *, transport: httpx.AsyncBaseTransport | None = None,
+    ):
+        self.settings = settings or OperationalActionSettings()
+        self.transport = transport
+
+    async def dispatch(
+        self, *, context: ProjectContext, task: OperationalTaskDetail,
+        action: HumanApprovalAction, credential: str | None,
+    ) -> ActionDispatchOutcome:
+        base = (self.settings.peripheral_url if action.kind == "create_work_order"
+                else self.settings.command_gateway_url)
+        if not base:
+            return ActionDispatchOutcome("unsupported", "尚未配置该动作的外部执行适配器")
+        if not credential or not credential.lower().startswith("bearer "):
+            return ActionDispatchOutcome("failed", "缺少当前用户的产品身份，未提交外部动作")
+        if action.task_id != task.task_id or task.project_id != context.project_id:
+            return ActionDispatchOutcome("failed", "动作与当前项目任务不匹配，未提交")
+        if action.approval_status != "approved":
+            return ActionDispatchOutcome("failed", "动作尚未通过人工确认，未提交")
+        request_id = action.execution_request_id
+        if request_id is None:
+            return ActionDispatchOutcome("failed", "缺少持久化执行编号，未提交")
+        headers = {"Authorization": credential, "X-Project-Id": str(context.project_id)}
+        try:
+            if action.kind == "create_work_order":
+                payload = CreateWorkOrderPayload.model_validate(action.payload)
+                body = {
+                    "request_id": str(request_id), "title": payload.title,
+                    "summary": payload.description, "status": "open",
+                    "priority": payload.priority, "assignee": payload.assignee,
+                    "scheduled_at": payload.scheduled_at.isoformat() if payload.scheduled_at else None,
+                    "location": payload.location,
+                    # An approved action creates a child execution order, not a
+                    # new inbound complaint/burst event for the sync loop.
+                    "work_order_type": "general",
+                }
+                path = "/api/sample/v1/work-orders"
+                recovery_path = f"/api/sample/v1/requests/{request_id}"
+            else:
+                payload = IssueCommandPayload.model_validate(action.payload)
+                if payload.expires_at <= datetime.now(timezone.utc):
+                    return ActionDispatchOutcome("failed", "命令确认窗口已过期，未提交")
+                body = {
+                    "request_id": str(request_id), "project_id": str(context.project_id),
+                    "task_id": str(task.task_id), "action_id": str(action.action_id),
+                    "action_version": action.version,
+                    "approved_by": action.decided_by,
+                    "approved_at": action.decided_at.isoformat() if action.decided_at else None,
+                    "command": payload.model_dump(mode="json"),
+                }
+                path = "/api/v1/commands"
+                recovery_path = f"/api/v1/commands/requests/{request_id}"
+        except (ValidationError, TypeError, ValueError):
+            return ActionDispatchOutcome("failed", "动作参数不符合执行契约，未提交")
+
+        async with httpx.AsyncClient(
+            base_url=base.rstrip("/"), headers=headers, timeout=self.settings.timeout_seconds,
+            follow_redirects=False, trust_env=False, transport=self.transport,
+        ) as client:
+            try:
+                response = await client.post(path, json=body)
+                if response.status_code in {400, 401, 403, 404, 409, 422}:
+                    # Sample responses explicitly state whether any write occurred.
+                    # An untyped rejection is not proof that no effect took place.
+                    error = response.json()
+                    if isinstance(error, dict) and error.get("effect") == "none":
+                        return ActionDispatchOutcome("failed", "外部系统拒绝了动作，未生效")
+                response.raise_for_status()
+                return self._receipt(action.kind, response.json(), request_id, context.project_id)
+            except (httpx.HTTPError, ValueError, KeyError, ValidationError):
+                # A lost response can follow a successful write. Recover using the
+                # same idempotency key, never manufacture success or blindly retry.
+                try:
+                    recovered = await client.get(recovery_path)
+                    recovered.raise_for_status()
+                    value = recovered.json()
+                    if action.kind == "create_work_order":
+                        if not isinstance(value, dict) or value.get("request_id") != str(request_id):
+                            raise ValueError("request receipt mismatch")
+                        value = value["result"]
+                    return self._receipt(action.kind, value, request_id, context.project_id)
+                except (httpx.HTTPError, ValueError, KeyError, ValidationError):
+                    return ActionDispatchOutcome("unknown", "外部执行结果待核实，请通过回执核对，勿重复下发")
+
+    async def reconcile(
+        self, *, context: ProjectContext, task: OperationalTaskDetail,
+        action: HumanApprovalAction, credential: str | None,
+    ) -> ActionDispatchOutcome:
+        request_id = action.execution_request_id
+        if not credential or request_id is None or action.task_id != task.task_id or task.project_id != context.project_id:
+            return ActionDispatchOutcome("unknown", "无法核实当前动作的身份或执行编号")
+        base = self.settings.peripheral_url if action.kind == "create_work_order" else self.settings.command_gateway_url
+        if not base:
+            return ActionDispatchOutcome("unknown", "回执查询服务尚未配置")
+        path = (f"/api/sample/v1/requests/{request_id}" if action.kind == "create_work_order"
+                else f"/api/v1/commands/requests/{request_id}")
+        try:
+            async with httpx.AsyncClient(
+                base_url=base.rstrip("/"), timeout=self.settings.timeout_seconds,
+                follow_redirects=False, trust_env=False, transport=self.transport,
+            ) as client:
+                response = await client.get(path, headers={
+                    "Authorization": credential, "X-Project-Id": str(context.project_id),
+                })
+                response.raise_for_status()
+                value = response.json()
+                if action.kind == "create_work_order":
+                    if not isinstance(value, dict) or value.get("request_id") != str(request_id):
+                        raise ValueError("request receipt mismatch")
+                    value = value["result"]
+                return self._receipt(action.kind, value, request_id, context.project_id)
+        except (httpx.HTTPError, ValueError, KeyError, ValidationError):
+            return ActionDispatchOutcome("unknown", "尚未取得可核实的执行回执")
+
+    @staticmethod
+    def _receipt(kind: str, value: object, request_id: UUID, project_id: UUID) -> ActionDispatchOutcome:
+        if kind == "create_work_order":
+            order = SampleWorkOrderReceipt.model_validate(value)
+            if order.project_id != project_id or order.created_at.tzinfo is None:
+                raise ValueError("invalid external project or timestamp")
+            receipt = ActionReceipt(
+                adapter="peripheral_sample", receipt_id=str(order.work_order_id),
+                submitted_at=order.created_at, completed_at=order.created_at,
+                external_status=order.status, detail="合成外围工单已创建，现场工作进度由外围系统继续同步",
+            )
+            return ActionDispatchOutcome("succeeded", receipt.detail, receipt)
+        result = CommandGatewayReceipt.model_validate(value)
+        if result.project_id != project_id or result.request_id != request_id:
+            raise ValueError("command receipt does not match submission")
+        if result.submitted_at.tzinfo is None or (result.completed_at and result.completed_at.tzinfo is None):
+            raise ValueError("command receipt timestamps must include timezone")
+        if result.status == "completed" and result.completed_at is None:
+            raise ValueError("command has no completion receipt")
+        receipt = ActionReceipt(
+            adapter="command_gateway", receipt_id=result.receipt_id,
+            submitted_at=result.submitted_at, completed_at=result.completed_at,
+            external_status=result.status, detail=result.detail,
+        )
+        status = "succeeded" if result.status == "completed" else "failed" if result.status == "failed" else "submitted"
+        return ActionDispatchOutcome(status, result.detail, receipt)
+
+
+def get_operational_action_dispatcher() -> OperationalActionDispatcher:
+    return HttpOperationalActionDispatcher()

app/infra/integrations/operational_data.py

@@ -0,0 +1,185 @@
+"""Read-only source adapters with bounded queries and explicit provenance."""
+from __future__ import annotations
+
+import hashlib
+import json
+from datetime import datetime, timezone
+from pathlib import Path
+
+import httpx
+from pydantic import Field
+from pydantic_settings import BaseSettings, SettingsConfigDict
+
+from app.auth.permissions import SCADA_VIEW, SIMULATION_VIEW, permissions_for_context
+from app.auth.project_dependencies import ProjectContext
+from app.domain.schemas.operational_data import TaskDataQuery, TaskDataSnapshot
+
+
+class TaskDataSettings(BaseSettings):
+    server_url: str = "http://127.0.0.1:8000"
+    peripheral_url: str = ""
+    timeout_seconds: float = Field(default=15, ge=1, le=60)
+    delegation_bearer: str = ""
+    model_config = SettingsConfigDict(
+        env_prefix="OPERATIONAL_DATA_", extra="ignore",
+        env_file=Path(__file__).resolve().parents[3] / ".env",
+    )
+
+
+class TaskDataError(Exception):
+    def __init__(self, status: int, detail: str):
+        super().__init__(detail)
+        self.status = status
+        self.detail = detail
+
+
+def _matches_project(value: object, project_id: str) -> bool:
+    if isinstance(value, list):
+        return all(_matches_project(item, project_id) for item in value)
+    if isinstance(value, dict):
+        if "project_id" in value and value["project_id"] != project_id:
+            return False
+        return all(_matches_project(item, project_id) for item in value.values())
+    return True
+
+
+def _select_demo_observations(
+    payload: object, query: TaskDataQuery, project_id: str,
+) -> dict[str, object]:
+    """Project a bounded DEMO-* observation query from the stored demo dataset."""
+    if not isinstance(payload, dict) or not _matches_project(payload, project_id):
+        raise TaskDataError(502, "数据源返回了其他项目的数据")
+    points = payload.get("points")
+    observations = payload.get("observations")
+    if not isinstance(points, list) or not isinstance(observations, list):
+        raise TaskDataError(502, "外围合成样本格式无效")
+    if not any(isinstance(point, dict) and point.get("point_id") == query.point_id for point in points):
+        raise TaskDataError(404, "合成样本测点不存在")
+    assert query.start is not None and query.end is not None
+    items: list[object] = []
+    try:
+        for item in observations:
+            if not isinstance(item, dict) or item.get("point_id") != query.point_id:
+                continue
+            observed_at = datetime.fromisoformat(str(item.get("observed_at")))
+            if query.start <= observed_at < query.end:
+                items.append(item)
+    except (TypeError, ValueError) as exc:
+        raise TaskDataError(502, "外围合成样本格式无效") from exc
+    return {
+        "origin": "synthetic",
+        "source_id": payload.get("source_id"),
+        "source_revision": payload.get("source_revision"),
+        "point_id": query.point_id,
+        "window": {
+            "start": query.start.isoformat(),
+            "end": query.end.isoformat(),
+        },
+        "items": items,
+        "truncated": False,
+    }
+
+
+async def read_task_data(
+    *, context: ProjectContext, query: TaskDataQuery, credential: str | None,
+    delegated: bool = False,
+    settings: TaskDataSettings | None = None,
+    transport: httpx.AsyncBaseTransport | None = None,
+) -> TaskDataSnapshot:
+    config = settings or TaskDataSettings()
+    permission = SCADA_VIEW if query.source.startswith("scada_") else SIMULATION_VIEW
+    if permission not in permissions_for_context(context) and not delegated:
+        raise TaskDataError(403, "当前身份没有该数据源的读取权限")
+    if delegated:
+        credential = config.delegation_bearer
+        if not credential:
+            raise TaskDataError(503, "任务授权已批准，但数据源委托身份尚未配置")
+        if not credential.lower().startswith("bearer "):
+            credential = f"Bearer {credential}"
+    if not credential or not credential.lower().startswith("bearer "):
+        raise TaskDataError(401, "数据查询需要当前产品身份")
+    peripheral = query.source.startswith("peripheral_")
+    base = config.peripheral_url if peripheral else config.server_url
+    if not base:
+        raise TaskDataError(503, "调查数据源尚未配置")
+    paths = {
+        "workbench_runtime": "/api/v1/workbench/runtime",
+        "scada_devices": "/api/v1/scada-devices",
+        "scada_readings": "/api/v1/timeseries/scada-readings",
+        "peripheral_points": "/api/sample/v1/monitoring/points",
+        "peripheral_work_orders": "/api/sample/v1/work-orders",
+        "peripheral_observations": "/api/sample/v1/monitoring/observations",
+        "peripheral_test_data": "/api/sample/v1/test-data",
+    }
+    demo_observations = (
+        query.source == "peripheral_observations"
+        and query.point_id is not None
+        and query.point_id.startswith("DEMO-")
+    )
+    path = "/api/sample/v1/demo-data" if demo_observations else paths[query.source]
+    params: dict[str, str | int] = {"limit": 100}
+    if query.source == "scada_readings":
+        params.update(device_ids=",".join(query.device_ids), start_time=query.start.isoformat(), end_time=query.end.isoformat())
+    elif query.source == "peripheral_observations" and not demo_observations:
+        params = {"point_id": query.point_id, "start": query.start.isoformat(), "end": query.end.isoformat()}
+    elif demo_observations:
+        params = {}
+    try:
+        async with httpx.AsyncClient(
+            base_url=base.rstrip("/"), timeout=config.timeout_seconds, follow_redirects=False,
+            trust_env=False, transport=transport,
+        ) as client:
+            async with client.stream("GET", path, params=params, headers={
+                "Authorization": credential, "X-Project-Id": str(context.project_id),
+            }) as response:
+                if response.status_code in {401, 403}:
+                    raise TaskDataError(response.status_code, "数据源拒绝当前身份或项目权限")
+                response.raise_for_status()
+                body = bytearray()
+                async for chunk in response.aiter_bytes():
+                    body.extend(chunk)
+                    if len(body) > 2_000_000:
+                        raise TaskDataError(422, "查询结果过大，请缩短时间范围或减少设备数量")
+        data = json.loads(body)
+    except (httpx.HTTPError, ValueError) as exc:
+        raise TaskDataError(502, "调查数据查询失败，不能据此形成已核实结论") from exc
+    if demo_observations:
+        data = _select_demo_observations(data, query, str(context.project_id))
+    elif not _matches_project(data, str(context.project_id)):
+        raise TaskDataError(502, "数据源返回了其他项目的数据")
+    origin = "synthetic" if peripheral else "system_data"
+    historical_replay = False
+    if query.source in {"peripheral_points", "peripheral_observations"}:
+        source_origin = data.get("origin") if isinstance(data, dict) else None
+        if source_origin not in {"historical_replay", "synthetic"}:
+            raise TaskDataError(502, "监测数据缺少明确来源，不能作为已核实证据")
+        historical_replay = source_origin == "historical_replay"
+        origin = "system_data" if historical_replay else "synthetic"
+    truncated = isinstance(data, dict) and bool(data.get("truncated") or data.get("next_cursor") or data.get("next_offset"))
+    if isinstance(data, list) and len(data) > 1000:
+        data, truncated = data[:1000], True
+    if isinstance(data, dict) and isinstance(data.get("items"), list):
+        if len(data["items"]) > 1000:
+            data = {**data, "items": data["items"][:1000]}
+            truncated = True
+        if isinstance(data.get("total"), int) and data["total"] > len(data["items"]):
+            truncated = True
+    # The digest identifies the returned snapshot, not merely a mutable URL.
+    digest = hashlib.sha256(json.dumps(data, ensure_ascii=False, sort_keys=True).encode()).hexdigest()
+    if demo_observations:
+        assert query.start is not None and query.end is not None
+        source_ref = (
+            f"peripheral_demo_data:{query.point_id}:"
+            f"{query.start.isoformat()}:{query.end.isoformat()}:sha256:{digest}"
+        )
+    elif historical_replay:
+        # Use a distinct identity: old snapshots may have incorrectly labeled
+        # this same historical response synthetic. Immutable evidence stays intact.
+        source_ref = f"peripheral_historical_replay:{query.source}:sha256:{digest}"
+    else:
+        source_ref = f"{query.source}:sha256:{digest}"
+    return TaskDataSnapshot(
+        source=query.source, source_ref=source_ref,
+        queried_at=datetime.now(timezone.utc), origin=origin,
+        data=data, truncated=truncated,
+    )

app/main.py

@@ -1,13 +1,15 @@
 from fastapi import FastAPI
-from contextlib import asynccontextmanager
+from contextlib import asynccontextmanager, AsyncExitStack
 from fastapi.middleware.gzip import GZipMiddleware
 from fastapi.middleware.cors import CORSMiddleware
 import logging
 from datetime import datetime
 
 from app.api.problem_details import install_problem_details_handlers
+from app.api.health import router as health_router
 from app.api.v1.rest_router import api_router
 from app.infra.db.dynamic_manager import project_connection_manager
+from app.infra.db.postgresql.workbench_event_hub import workbench_event_hubs
 from app.infra.db.metadb.database import close_metadata_engine
 from app.infra.db.timescaledb.sync_pool import close_all_timescale_pools
 from app.native.wndb.core.connection import close_all_project_pools
@@ -23,20 +25,38 @@ logger.setLevel(logging.INFO)
 
 @asynccontextmanager
 async def lifespan(app: FastAPI):
+    app.state.startup_complete = False
     logger.info("**********************************************************")
     logger.info(str(datetime.now()))
     logger.info("TJWater CloudService is starting...")
     logger.info("**********************************************************")
 
-    yield
-    # 清理资源
-    await project_connection_manager.close_all()
-    close_all_timescale_pools()
-    close_all_project_pools()
-    await close_metadata_engine()
-    logger.info("Database connections closed")
+    try:
+        async with AsyncExitStack() as stack:
+            if mcp_runtime is not None:
+                await stack.enter_async_context(mcp_runtime.lifespan())
+            app.state.startup_complete = True
+            try:
+                yield
+            finally:
+                app.state.startup_complete = False
+    finally:
+        await workbench_event_hubs.close_all()
+        await project_connection_manager.close_all()
+        close_all_timescale_pools()
+        close_all_project_pools()
+        await close_metadata_engine()
+        logger.info("Database connections closed")
 
 
+# Disabled deployments do not import the MCP SDK or acquire business credentials.
+from app.mcp.config import McpSettings
+mcp_settings = McpSettings()
+mcp_runtime = None
+if mcp_settings.enabled:
+    from app.mcp.application import McpQueryApplication
+    mcp_runtime = McpQueryApplication(mcp_settings)
+
 # 根据环境配置决定是否启用文档
 is_production = settings.ENVIRONMENT.lower() == "production"
 
@@ -53,20 +73,28 @@ app = FastAPI(
 
 
 # Include Routers
+app.state.startup_complete = False
+app.include_router(health_router)
 app.include_router(api_router, prefix="/api/v1")
 install_problem_details_handlers(app)
-# Legcy Routers without version prefix
+from app.api.v1.endpoints.mcp_queries import router as mcp_context_router
+app.state.mcp_runtime = mcp_runtime
+app.include_router(mcp_context_router, prefix="/api/v1")
+# Legacy routers without version prefix
 # app.include_router(api_router)
 
 # 配置中间件
 app.add_middleware(GZipMiddleware, minimum_size=1000)
 # 添加审计中间件（可选，记录关键操作）
 app.add_middleware(AuditMiddleware)
+if mcp_runtime is not None:
+    from app.mcp.dispatch import McpDispatchMiddleware
+    app.add_middleware(McpDispatchMiddleware, runtime=mcp_runtime)
 # 配置 CORS 中间件
 # 确保这是你最后一个添加的 app.add_middleware
 app.add_middleware(
     CORSMiddleware,
-    allow_origins=["*"],  # 允许所有来源
+    allow_origins=settings.cors_allowed_origins,
     allow_credentials=True,
     allow_methods=["*"],
     allow_headers=["*"],

app/mcp/analysis_models.py

@@ -0,0 +1,139 @@
+"""Bounded calculation inputs; identity, routing and output paths are host-owned."""
+
+from datetime import datetime, timedelta
+from typing import Annotated, Literal
+from uuid import UUID
+
+from pydantic import Field, model_validator
+
+from app.mcp.models import Id, StrictInput
+
+
+MAX_CALCULATION_TIMEOUT_SECONDS = 86400
+
+
+class CalculationInput(StrictInput):
+    timeout_seconds: int = Field(
+        default=120, ge=1, le=MAX_CALCULATION_TIMEOUT_SECONDS, strict=True,
+        description="本次计算等待预算，秒；由Agent按模型规模和任务时长选择，受工具目录中的服务端上限约束；与模拟duration不同，超时不保证已启动的计算停止或回滚",
+    )
+
+
+class BurstDetectionInput(CalculationInput):
+    target_time: datetime = Field(description="带时区的侦测时刻；读取此前15天同周期压力数据，不是单小时定位")
+    sensor_nodes: list[Id] = Field(min_length=5, max_length=20, description="已核实的压力监测节点ID，至少5个完整传感器")
+    sampling_interval_minutes: int = Field(default=15, ge=1, le=60)
+
+    @model_validator(mode="after")
+    def valid_target(self):
+        if self.target_time.tzinfo is None:
+            raise ValueError("target_time requires explicit timezone")
+        if 1440 % self.sampling_interval_minutes:
+            raise ValueError("sampling interval must divide a day")
+        if len(set(self.sensor_nodes)) != len(self.sensor_nodes):
+            raise ValueError("sensor_nodes must be distinct")
+        return self
+
+
+class BurstLocationInput(CalculationInput):
+    data_source: Literal["monitoring", "simulation"] = "monitoring"
+    simulation_run_id: UUID | None = Field(default=None, description="已有分析运行UUID，不是方案名称；simulation模式只读取已有异常结果及模型基线")
+    burst_leakage: float = Field(gt=0, allow_inf_nan=False, description="有依据的假定漏水流量，WNTR SI单位m³/s；不是已知爆管管段，不能随意猜测")
+    pressure_scada_ids: list[Id] | None = Field(default=None, min_length=1, max_length=20, description="压力监测设备关联的模型节点ID，不是设备查询编号；省略时自动读取当前项目SCADA配置中的压力测点，不需要猜测节点ID")
+    flow_scada_ids: list[Id] | None = Field(default=None, min_length=1, max_length=20, description="流量监测设备关联的模型要素ID，不是设备查询编号")
+    scada_burst_start: datetime
+    scada_burst_end: datetime
+    scada_normal_start: datetime | None = Field(default=None, description="monitoring模式正常对照窗开始，须与结束时间成对提供；均省略时用爆管窗前一天相同时段，不代表该时段已确认正常")
+    scada_normal_end: datetime | None = Field(default=None, description="monitoring模式正常对照窗结束，须与开始时间成对提供；simulation模式不填写，使用同时间窗模型基线")
+
+    @model_validator(mode="after")
+    def valid_windows(self):
+        pairs = [(self.scada_burst_start, self.scada_burst_end)]
+        if self.data_source == "monitoring":
+            if self.simulation_run_id is not None:
+                raise ValueError("monitoring cannot use simulation_run_id")
+            if self.scada_normal_start is not None or self.scada_normal_end is not None:
+                pairs.append((self.scada_normal_start, self.scada_normal_end))
+        elif self.simulation_run_id is None or self.scada_normal_start is not None or self.scada_normal_end is not None:
+            raise ValueError("simulation requires run_id and uses the same-window model baseline")
+        for start, end in pairs:
+            if start is None or end is None or start.tzinfo is None or end.tzinfo is None:
+                raise ValueError("explicit timezone-aware burst and normal windows are required")
+            if not timedelta(0) < end - start <= timedelta(days=1):
+                raise ValueError("each window must be positive and at most 24 hours")
+        for ids in (self.pressure_scada_ids, self.flow_scada_ids):
+            if ids is not None and len(set(ids)) != len(ids):
+                raise ValueError("sensor IDs must be distinct")
+        return self
+
+
+class ValveIsolationInput(CalculationInput):
+    accident_element: list[Id] = Field(min_length=1, max_length=20, description="经核实或明确作为候选的事故管段/节点ID")
+    disabled_valves: list[Id] = Field(default_factory=list, max_length=200, description="已知无法关闭的阀门；未提供不代表现场确认全部可用")
+
+
+class LeakageIdentificationInput(CalculationInput):
+    scada_start: datetime
+    scada_end: datetime
+    sensor_nodes: list[Id] = Field(min_length=1, max_length=20)
+    q_sum: float = Field(gt=0, allow_inf_nan=False, description="有依据的总漏损流量约束，m³/s，不得使用无依据默认值")
+    start_time: float = Field(ge=0, lt=24, allow_inf_nan=False, description="模型起算小时；需与SCADA时间窗及模型时钟对齐")
+    timestep: int = Field(default=15, ge=5, le=60, description="计算步长，分钟；需与观测采样对应")
+    dma_count: int = Field(ge=1, le=20)
+
+    @model_validator(mode="after")
+    def valid_window(self):
+        if self.scada_start.tzinfo is None or self.scada_end.tzinfo is None:
+            raise ValueError("timestamps require explicit timezone")
+        duration = self.scada_end - self.scada_start
+        if not timedelta(0) < duration <= timedelta(days=1):
+            raise ValueError("window must be positive and at most 24 hours")
+        if duration.total_seconds() % (self.timestep * 60):
+            raise ValueError("window must contain whole calculation steps")
+        if len(set(self.sensor_nodes)) != len(self.sensor_nodes):
+            raise ValueError("sensor_nodes must be distinct")
+        return self
+
+
+class ScenarioInput(CalculationInput):
+    start_time: datetime
+    duration: int = Field(default=900, ge=60, le=86400, description="模拟时长，秒；首次优先短时间窗")
+    output_node_ids: list[Id] = Field(min_length=1, max_length=20, description="需要返回结果的节点；不把局部结果解释为全网无影响")
+    output_link_ids: list[Id] = Field(default_factory=list, max_length=20)
+
+    @model_validator(mode="after")
+    def timezone_required(self):
+        if self.start_time.tzinfo is None:
+            raise ValueError("start_time requires explicit timezone")
+        return self
+
+
+class ContaminantSimulationInput(ScenarioInput):
+    source: Id = Field(description="假设污染源节点；这是正向传播模拟，不是自动反演污染源")
+    concentration: float = Field(gt=0, allow_inf_nan=False, description="假设源浓度，mg/L")
+    pattern: Id | None = Field(default=None, description="已存在的污染源时间模式ID；不填为持续恒定源")
+
+
+class ValveClosureInput(ScenarioInput):
+    valves: list[Id] = Field(min_length=1, max_length=20, description="在临时模型中关闭的阀门，不下发现场控制")
+
+
+class FlushingInput(ScenarioInput):
+    drainage_node_id: Id
+    flushing_flow_m3h: float = Field(gt=0, allow_inf_nan=False, description="明确的冲洗流量，m³/h；不使用无依据的自动排放量")
+    closed_valves: list[Id] = Field(default_factory=list, max_length=20)
+
+
+NonnegativeValue = Annotated[float, Field(ge=0, allow_inf_nan=False)]
+PumpSeries = Annotated[list[NonnegativeValue], Field(min_length=1, max_length=96)]
+
+
+class PressureRegulationInput(ScenarioInput):
+    pump_control: dict[Id, PumpSeries] = Field(min_length=1, max_length=20, description="已登记泵ID及运行序列；工频值为0/1，变频值为频率Hz，每项对应模型水力步长，不自动优化")
+    tank_init_level: dict[Id, Annotated[float, Field(gt=0, allow_inf_nan=False)]] = Field(default_factory=dict, max_length=20, description="水箱初始水位，须在模型允许范围内")
+
+
+class SensorPlacementInput(CalculationInput):
+    method: Literal["sensitivity", "kmeans"]
+    sensor_count: int = Field(ge=1, le=20)
+    min_diameter: int = Field(default=0, ge=0, le=5000, description="最小管径，mm")

app/mcp/analysis_service.py

@@ -0,0 +1,69 @@
+"""Side-effect-limited adapters for existing domain calculations."""
+
+import asyncio
+from threading import BoundedSemaphore
+from typing import Any
+
+from fastapi import HTTPException
+from pydantic import BaseModel
+from app.mcp.scenario_service import SCENARIO_TOOLS, calculate_scenario
+
+# These services synchronously read SCADA/model series or write online results.
+# Topology-only isolation and model-based sensor placement need only BizDB.
+TIMESCALE_CALCULATIONS = SCENARIO_TOOLS | frozenset({
+    "run_burst_detection", "run_burst_location", "run_leakage_identification",
+})
+
+# A timed-out HTTP request cannot kill a native algorithm thread. Keep its slot
+# until the calculation really exits, so retries cannot accumulate CPU jobs.
+_slots = BoundedSemaphore(2)
+
+
+def _calculate(name: str, parameters: dict[str, Any], network: str, username: str) -> dict[str, Any]:
+    if not _slots.acquire(blocking=False):
+        raise HTTPException(503, "Calculation capacity is occupied")
+    try:
+        if name in SCENARIO_TOOLS:
+            return calculate_scenario(name, parameters, network=network, username=username)
+        if name == "optimize_sensor_placement":
+            from app.services.sensor_placement import optimize_sensor_placement_by_kmeans, optimize_sensor_placement_by_sensitivity
+            parameters = dict(parameters)
+            method = parameters.pop("method")
+            optimizer = optimize_sensor_placement_by_sensitivity if method == "sensitivity" else optimize_sensor_placement_by_kmeans
+            return optimizer(project_code=network, run_name="MCP preview", created_by=username,
+                             persist_results=False, **parameters)
+        if name == "run_leakage_identification":
+            from app.services.dma_leakage_estimation import run_leakage_identification
+
+            duration = (parameters["scada_end"] - parameters["scada_start"]).total_seconds() / 3600
+            return run_leakage_identification(
+                network=network, username=username, scheme_name=None,
+                duration=duration, q_sum_unit="m3/s", output_flow_unit="m3/s",
+                n_workers=1, pop_size=50, max_gen=100, **parameters,
+            )
+        if name == "run_burst_detection":
+            from app.services.burst_detection import run_burst_detection
+
+            return run_burst_detection(network=network, username=username, scheme_name=None, **parameters)
+        if name == "run_burst_location":
+            from app.services.burst_location import run_burst_location_by_network
+
+            return run_burst_location_by_network(
+                network=network, username=username, scheme_name=None,
+                final_candidates_csv_path=None, **parameters,
+            )
+        if name == "analyze_valve_isolation":
+            from app.services.valve_isolation import analyze_valve_isolation
+
+            return analyze_valve_isolation(network=network, **parameters)
+        raise ValueError("Calculation is not registered")
+    finally:
+        _slots.release()
+
+
+async def calculate(name: str, args: BaseModel, *, network: str, username: str) -> dict[str, Any]:
+    # to_thread copies the validated project's routing context. Never accept
+    # database routes, actor identity, arbitrary measurements or paths from MCP.
+    return await asyncio.to_thread(
+        _calculate, name, args.model_dump(exclude={"context_handle", "timeout_seconds"}, exclude_none=True), network, username,
+    )

app/mcp/application.py

@@ -0,0 +1,505 @@
+"""Official SDK transport with product authorization, persistence and query handlers."""
+
+import asyncio
+import json
+import logging
+from contextlib import asynccontextmanager
+from urllib.parse import urlsplit
+from uuid import uuid5, NAMESPACE_URL
+from typing import Any
+
+from fastapi import HTTPException
+from mcp import types
+from mcp.server import Server
+from mcp.server.auth.middleware.auth_context import get_access_token
+from mcp.server.auth.settings import AuthSettings
+from mcp.server.auth.routes import create_protected_resource_routes
+from mcp.server.caching import CacheHint
+from mcp.server.transport_security import TransportSecuritySettings
+from mcp.shared.exceptions import MCPError
+from psycopg.errors import UndefinedTable, InvalidSchemaName
+from pydantic import ValidationError
+from starlette.responses import JSONResponse
+
+from app.mcp.auth import (
+    McpTokenVerifier,
+    QueryScopeMiddleware,
+    QUERY_SCOPE,
+    CONTEXT_SCOPE,
+    challenge,
+)
+from app.mcp.catalog import TOOLS
+from app.mcp.config import McpSettings
+from app.mcp.contexts import resolve_context, verify_task
+from app.mcp.models import QueryResult
+from app.mcp.scope import project_scope
+from app.mcp.query_service import execute_query, task_service, authorize_query
+from app.infra.db.postgresql.mcp_queries import McpQueryRepository
+from app.domain.schemas.operational_data import TaskDataSnapshot
+
+logger = logging.getLogger(__name__)
+
+
+def tool_error(code, message):
+    return types.CallToolResult(
+        isError=True,
+        content=[types.TextContent(type="text", text=message)],
+        structuredContent={"error": {"code": code, "message": message}},
+    )
+
+
+class ModernOnlyMiddleware:
+    def __init__(self, app):
+        self.app = app
+
+    async def __call__(self, scope, receive, send):
+        if scope["type"] != "http" or scope.get("path") != "/mcp":
+            return await self.app(scope, receive, send)
+        if scope["method"] in {"GET", "DELETE"}:
+            return await JSONResponse(
+                {"error": "modern_protocol_required"},
+                status_code=405,
+                headers={"Allow": "POST"},
+            )(scope, receive, send)
+        headers = {
+            k.decode("latin1").lower(): v.decode("latin1")
+            for k, v in scope.get("headers", [])
+        }
+        version = headers.get("mcp-protocol-version")
+        if version != "2026-07-28":
+            body = bytearray()
+            request_id = None
+            body_version = None
+            while True:
+                message = await receive()
+                if message["type"] == "http.disconnect":
+                    return
+                body.extend(message.get("body", b""))
+                if len(body) > 1024 * 1024:
+                    return await JSONResponse(
+                        {"error": "request_too_large"}, status_code=413
+                    )(scope, receive, send)
+                if not message.get("more_body"):
+                    break
+            try:
+                payload = json.loads(body)
+                request_id = payload.get("id")
+                body_version = payload.get("params", {}).get("_meta", {}).get(
+                    types.PROTOCOL_VERSION_META_KEY
+                )
+            except (ValueError, AttributeError):
+                pass
+            mismatch = not version or (
+                isinstance(body_version, str) and body_version != version
+            )
+            code = -32020 if mismatch else -32022
+            result = {
+                "jsonrpc": "2.0",
+                "id": request_id,
+                "error": {
+                    "code": code,
+                    "message": (
+                        "MCP-Protocol-Version header is missing or differs from request metadata"
+                        if mismatch else "MCP 2026-07-28 is required"
+                    ),
+                    "data": {"supported": ["2026-07-28"], "requested": version},
+                },
+            }
+            return await JSONResponse(result, status_code=400)(scope, receive, send)
+        await self.app(scope, receive, send)
+
+
+class McpQueryApplication:
+    def __init__(
+        self,
+        settings: McpSettings,
+        *,
+        verifier=None,
+        scope_factory=project_scope,
+        executor=execute_query,
+    ):
+        self.settings = settings
+        self.verifier = verifier or McpTokenVerifier(settings)
+        self.scope_factory = scope_factory
+        self.executor = executor
+        self.active = {}
+        self.ready = False
+        self.server = Server(
+            "tjwater-query",
+            version="1.0.0",
+            instructions="Project-bound water queries. Model baselines are not observations. Truncated results do not cover the full range.",
+            cache_hints={
+                "server/discover": CacheHint(300000, "public"),
+                "tools/list": CacheHint(300000, "public"),
+            },
+            on_list_tools=self.list_tools,
+            on_call_tool=self.call_tool,
+            get_tool_input_schema=self.tool_input_schema,
+            on_list_resources=self.list_resources,
+            on_list_resource_templates=self.list_templates,
+            on_read_resource=self.read_resource,
+        )
+        self.server.add_request_handler(
+            "server/discover", types.RequestParams, self.discover
+        )
+        app = self.server.streamable_http_app(
+            stateless_http=True,
+            json_response=True,
+            max_request_body_size=1024 * 1024,
+            auth=AuthSettings(
+                issuer_url=settings.issuer_uri,
+                resource_server_url=settings.resource_uri,
+                required_scopes=["mcp.catalog.read"],
+                validate_token_resource=True,
+            ),
+            token_verifier=self.verifier,
+            transport_security=TransportSecuritySettings(
+                allowed_hosts=settings.allowed_hosts,
+                allowed_origins=settings.allowed_origins,
+            ),
+        )
+        # Catalog authentication and advertised resource scopes are different concerns.
+        metadata = create_protected_resource_routes(
+            resource_url=settings.resource_uri,
+            authorization_servers=[settings.issuer_uri],
+            scopes_supported=["mcp.catalog.read", QUERY_SCOPE, CONTEXT_SCOPE],
+            resource_name="TJWater Query",
+        )
+        metadata_path = metadata[0].path
+        app.router.routes = [
+            route for route in app.router.routes if route.path != metadata_path
+        ] + metadata
+        self.asgi = ModernOnlyMiddleware(
+            QueryScopeMiddleware(app, self.verifier, settings)
+        )
+        from app.api.v1.endpoints.mcp_queries import router
+
+        self.router = router
+
+    @asynccontextmanager
+    async def lifespan(self):
+        async with self.server.session_manager.run():
+            self.ready = True
+            try:
+                yield
+            finally:
+                self.ready = False
+
+    async def discover(self, ctx, params):
+        return types.DiscoverResult(
+            supported_versions=["2026-07-28"],
+            capabilities=self.server.get_capabilities(
+                protocol_version=ctx.protocol_version,
+                extensions={"io.modelcontextprotocol/oauth-client-credentials": {}},
+            ),
+            instructions=self.server.instructions,
+        )
+
+    def tool_input_schema(self, name: str) -> dict[str, Any] | None:
+        tool = TOOLS.get(name)
+        if tool is None:
+            return None
+        schema = tool.model.model_json_schema()
+        if tool.computation:
+            schema["properties"]["timeout_seconds"]["maximum"] = self.settings.max_calculation_timeout_seconds
+        return schema
+
+    async def list_tools(self, ctx, params):
+        if params and params.cursor:
+            raise MCPError(-32602, "The query catalog has no continuation page")
+        tools = []
+        for name in sorted(TOOLS):
+            tool = TOOLS[name].protocol_tool()
+            tool.input_schema.update(self.tool_input_schema(name))
+            tools.append(tool)
+        return types.ListToolsResult(
+            tools=tools,
+            ttlMs=300000,
+            cacheScope="public",
+        )
+
+    async def list_resources(self, ctx, params):
+        return types.ListResourcesResult(resources=[], ttlMs=0, cacheScope="private")
+
+    async def list_templates(self, ctx, params):
+        return types.ListResourceTemplatesResult(
+            resourceTemplates=[
+                types.ResourceTemplate(
+                    name="task-evidence",
+                    uriTemplate="tjwater://tasks/{task_id}/evidence/{evidence_id}",
+                    mimeType="application/json",
+                )
+            ],
+            ttlMs=300000,
+            cacheScope="public",
+        )
+
+    async def read_resource(self, ctx, params):
+        meta = ctx.meta or {}
+        handle = meta.get("io.tjwater/contextHandle")
+        url = urlsplit(str(params.uri))
+        parts = url.path.strip("/").split("/")
+        if (
+            url.scheme != "tjwater"
+            or url.netloc != "tasks"
+            or len(parts) != 3
+            or parts[1] != "evidence"
+            or url.query
+            or url.fragment
+        ):
+            raise MCPError(-32602, "Unknown query resource")
+        result = await self._query(
+            ctx,
+            "get_task_evidence",
+            {"context_handle": handle, "evidence_id": parts[2]},
+            resource_task=parts[0],
+            include_full=True,
+        )
+        if result.isError:
+            raise MCPError(
+                -32602, "Evidence is unavailable in this authorization context"
+            )
+        return types.ReadResourceResult(
+            contents=[
+                types.TextResourceContents(
+                    uri=str(params.uri),
+                    mimeType="application/json",
+                    text=json.dumps(result.structuredContent, ensure_ascii=False),
+                )
+            ],
+            ttlMs=0,
+            cacheScope="private",
+        )
+
+    async def call_tool(self, ctx, params):
+        return await self._query(ctx, params.name, params.arguments or {})
+
+    async def _query(
+        self, ctx, name, arguments, resource_task=None, include_full=False
+    ):
+        tool = TOOLS.get(name)
+        if tool is None:
+            raise MCPError(-32602, "Unknown query tool")
+        principal = get_access_token()
+        if principal is None or QUERY_SCOPE not in principal.scopes:
+            return tool_error("PERMISSION_DENIED", "Query permission is required")
+        try:
+            args = tool.model.model_validate(arguments)
+        except ValidationError:
+            return tool_error(
+                "INVALID_QUERY", "Query parameters do not match the tool schema"
+            )
+        timeout_seconds = (
+            args.timeout_seconds if tool.computation else self.settings.query_timeout_seconds
+        )
+        if tool.computation and timeout_seconds > self.settings.max_calculation_timeout_seconds:
+            return tool_error("INVALID_QUERY", "Calculation timeout exceeds the server limit")
+        request = ctx.request
+        project = request.headers.get("x-project-id", "") if request else ""
+        meta = ctx.meta or {}
+        call_id = meta.get("io.tjwater/callId")
+        if not isinstance(call_id, str) or not 0 < len(call_id) <= 128:
+            return tool_error(
+                "CALL_ID_REQUIRED", "A host-owned call identifier is required"
+            )
+        key = (principal.subject, project)
+        if self.active.get(key, 0) >= 4:
+            return tool_error(
+                "RATE_LIMITED", "At most four queries may run concurrently"
+            )
+        self.active[key] = self.active.get(key, 0) + 1
+        try:
+            async with (
+                asyncio.timeout(timeout_seconds),
+                self.scope_factory(principal, project) as scope,
+            ):
+                binding = await resolve_context(scope, principal, args.context_handle)
+                if resource_task is not None and str(binding.task_id) != resource_task:
+                    raise HTTPException(403, "Resource task mismatch")
+                await authorize_query(scope, binding, tool, args)
+                repo = McpQueryRepository(scope.connection, scope.project.project_id)
+                parameters = {
+                    "tool": name,
+                    "arguments": args.model_dump(
+                        mode="json", exclude={"context_handle"}
+                    ),
+                }
+                scope_key = json.dumps(
+                    [
+                        binding.subject,
+                        binding.client_id,
+                        binding.native_session_id,
+                        str(binding.task_id),
+                        str(binding.run_id),
+                        str(binding.execution_request_id),
+                    ],
+                    separators=(",", ":"),
+                )
+                previous = (
+                    await repo.receipt(scope_key, call_id) if binding.task_id else None
+                )
+                if previous:
+                    if previous["parameters"] != parameters:
+                        raise HTTPException(
+                            409, "Call identifier has different query parameters"
+                        )
+                    result = QueryResult.model_validate(previous["result"])
+                else:
+                    result = (
+                        await self.executor(
+                            scope, binding, tool, args, full_snapshot=True
+                        )
+                        if include_full
+                        else await self.executor(scope, binding, tool, args)
+                    )
+                    # Never serialize an unbounded response into a model message or database receipt.
+                    if len(result.model_dump_json().encode()) > 8 * 1024 * 1024:
+                        return tool_error(
+                            "RESULT_TOO_LARGE",
+                            "Narrow the query or request a smaller page",
+                        )
+                    if binding.task_id and name not in {
+                        "get_task_context",
+                        "get_task_evidence",
+                    }:
+                        async with scope.connection.transaction():
+                            await repo.lock_receipt(scope_key, call_id)
+                            previous = await repo.receipt(scope_key, call_id)
+                            if previous:
+                                if previous["parameters"] != parameters:
+                                    raise HTTPException(
+                                        409,
+                                        "Call identifier has different query parameters",
+                                    )
+                                result = QueryResult.model_validate(previous["result"])
+                            else:
+                                task_repo = task_service(scope).repo
+                                # Lock the task row only after external/data-source work has finished.
+                                await task_repo.lock(f"task:{binding.task_id}")
+                                await verify_task(scope, binding)
+                                result.source_ref = "mcp:" + str(
+                                    uuid5(NAMESPACE_URL, scope_key + ":" + call_id)
+                                )
+                                snapshot = TaskDataSnapshot(
+                                    source=name,
+                                    source_ref=result.source_ref,
+                                    queried_at=result.queried_at,
+                                    origin=result.origin,
+                                    data=result.model_dump(mode="json"),
+                                    truncated=result.truncated,
+                                )
+                                evidence, _ = await task_service(
+                                    scope
+                                ).record_data_snapshot(binding.task_id, snapshot)
+                                result.evidence_id = evidence.evidence_id
+                                await repo.save_receipt(
+                                    scope_key,
+                                    call_id,
+                                    parameters,
+                                    result.model_dump(mode="json"),
+                                )
+                output = result.model_dump(mode="json")
+                content = []
+                if (
+                    not include_full
+                    and len(json.dumps(output, ensure_ascii=False).encode()) > 12000
+                ):
+                    evidence_id = result.evidence_id or (
+                        args.evidence_id if name == "get_task_evidence" else None
+                    )
+                    if binding.task_id and evidence_id:
+                        uri = (
+                            f"tjwater://tasks/{binding.task_id}/evidence/{evidence_id}"
+                        )
+                        output["data"] = {
+                            "resource_uri": uri,
+                            "preview": "The complete query snapshot is saved in task evidence",
+                        }
+                        output["truncated"] = True
+                        output["data_gaps"] = [
+                            *result.data_gaps,
+                            "Inline result abbreviated; read the saved evidence resource",
+                        ]
+                        content.append(
+                            types.ResourceLink(
+                                type="resource_link",
+                                uri=uri,
+                                name="query-evidence",
+                                mimeType="application/json",
+                            )
+                        )
+                    else:
+                        return tool_error(
+                            "RESULT_TOO_LARGE",
+                            "Request a smaller page or narrower window",
+                        )
+                content.insert(
+                    0,
+                    types.TextContent(
+                        type="text", text=json.dumps(output, ensure_ascii=False)
+                    ),
+                )
+                return types.CallToolResult(content=content, structuredContent=output)
+        except HTTPException as exc:
+            codes = {
+                401: "AUTH_REQUIRED",
+                403: "QUERY_CONTEXT_DENIED",
+                404: "NOT_FOUND",
+                409: "QUERY_CONTEXT_STALE",
+                413: "RESULT_TOO_LARGE",
+                422: "INVALID_QUERY",
+                503: "SOURCE_UNAVAILABLE",
+            }
+            return tool_error(
+                codes.get(exc.status_code, "QUERY_FAILED"),
+                str(exc.detail)
+                if isinstance(exc.detail, str)
+                and exc.status_code in {403, 404, 409, 413, 422}
+                else "Query is unavailable",
+            )
+        except (UndefinedTable, InvalidSchemaName):
+            return tool_error(
+                "MCP_STORAGE_UNAVAILABLE",
+                "MCP project storage has not been provisioned",
+            )
+        except TimeoutError:
+            return tool_error("QUERY_TIMEOUT", "Query execution timed out")
+        except ValueError:
+            return tool_error(
+                "INVALID_QUERY", "Invalid query value or incompatible source data"
+            )
+        except Exception:
+            logger.warning(
+                "MCP query unavailable", extra={"tool": name, "project_id": project}
+            )
+            return tool_error("SOURCE_UNAVAILABLE", "Query source is unavailable")
+        finally:
+            remaining = self.active[key] - 1
+            if remaining:
+                self.active[key] = remaining
+            else:
+                self.active.pop(key, None)
+
+    async def identity(self, request):
+        auth = request.headers.get("authorization", "")
+        try:
+            principal = (
+                await self.verifier.verify_token(auth[7:])
+                if auth.lower().startswith("bearer ")
+                else None
+            )
+        except Exception:
+            raise HTTPException(503, "OAuth key verification is unavailable") from None
+        if not principal:
+            raise HTTPException(
+                401,
+                "MCP authentication required",
+                headers=challenge(self.settings, CONTEXT_SCOPE),
+            )
+        if CONTEXT_SCOPE not in principal.scopes:
+            raise HTTPException(
+                403,
+                "MCP context scope required",
+                headers=challenge(self.settings, CONTEXT_SCOPE, insufficient=True),
+            )
+        return principal

app/mcp/auth.py

@@ -0,0 +1,169 @@
+"""Verify resource-bound Keycloak tokens without passing them to business APIs."""
+
+import asyncio
+import time
+from uuid import UUID
+
+import httpx
+from jose import JWTError, jwt
+from mcp.server.auth.provider import AccessToken
+from starlette.responses import JSONResponse
+
+from app.mcp.config import McpSettings
+
+CATALOG_SCOPE = "mcp.catalog.read"
+QUERY_SCOPE = "mcp.query"
+CONTEXT_SCOPE = "mcp.context.bind"
+
+
+class McpTokenVerifier:
+    def __init__(self, settings: McpSettings, client: httpx.AsyncClient | None = None):
+        self.settings = settings
+        self.client = client
+        self._keys: list[dict] = []
+        self._expires = 0.0
+        self._last_refresh = 0.0
+        self._lock = asyncio.Lock()
+
+    async def _load_keys(self, kid: str) -> list[dict]:
+        now = time.monotonic()
+        if self._expires > now and any(k.get("kid") == kid for k in self._keys):
+            return self._keys
+        async with self._lock:
+            now = time.monotonic()
+            if self._expires > now and any(k.get("kid") == kid for k in self._keys):
+                return self._keys
+            if self._expires > now and now - self._last_refresh < 20:
+                return self._keys
+            url = (
+                self.settings.issuer_uri.rstrip("/") + "/protocol/openid-connect/certs"
+            )
+            self._last_refresh = now
+            if self.client:
+                response = await self.client.get(url)
+            else:
+                async with httpx.AsyncClient(
+                    timeout=5, trust_env=False, follow_redirects=False
+                ) as client:
+                    response = await client.get(url)
+            response.raise_for_status()
+            try:
+                payload = response.json()
+            except ValueError:
+                raise httpx.RequestError("invalid OAuth key set response") from None
+            keys = payload.get("keys") if isinstance(payload, dict) else None
+            if not isinstance(keys, list) or len(keys) > 100:
+                raise httpx.RequestError("invalid OAuth key set")
+            self._keys = [
+                k
+                for k in keys
+                if isinstance(k, dict)
+                and k.get("kty") == "RSA"
+                and k.get("use", "sig") == "sig"
+                and k.get("alg", "RS256") == "RS256"
+            ]
+            self._expires = now + 300
+            return self._keys
+
+    async def verify_token(self, token: str) -> AccessToken | None:
+        if not token or len(token) > 16384:
+            return None
+        try:
+            header = jwt.get_unverified_header(token)
+            kid = header.get("kid")
+            if (
+                header.get("alg") != "RS256"
+                or not isinstance(kid, str)
+                or not 0 < len(kid) <= 128
+            ):
+                return None
+            keys = await self._load_keys(kid)
+            key = next((k for k in keys if k.get("kid") == kid), None)
+            if key is None:
+                return None
+            claims = jwt.decode(
+                token,
+                key,
+                algorithms=["RS256"],
+                audience=self.settings.audience,
+                issuer=self.settings.issuer_uri,
+                options={"require_exp": True, "require_iat": True, "require_sub": True},
+            )
+            UUID(claims["sub"])
+            client_id = claims.get("azp")
+            if not isinstance(client_id, str) or not client_id:
+                return None
+            if (
+                not isinstance(claims.get("iat"), (int, float))
+                or not time.time() - 900 < claims["iat"] <= time.time() + 30
+                or claims["exp"] > claims["iat"] + 900
+            ):
+                return None
+            return AccessToken(
+                token=token,
+                client_id=client_id,
+                subject=claims["sub"],
+                scopes=str(claims.get("scope", "")).split(),
+                expires_at=int(claims["exp"]),
+                resource=self.settings.resource_uri,
+                claims=claims,
+            )
+        except (JWTError, ValueError, KeyError, TypeError):
+            return None
+        # JWKS/network failures propagate as unavailable, never as empty data.
+
+
+def challenge(
+    settings: McpSettings, scope: str, *, insufficient: bool = False
+) -> dict[str, str]:
+    from mcp.server.auth.routes import build_resource_metadata_url
+
+    metadata = build_resource_metadata_url(settings.resource_uri)
+    suffix = ', error="insufficient_scope"' if insufficient else ""
+    return {
+        "WWW-Authenticate": f'Bearer resource_metadata="{metadata}", scope="{scope}"{suffix}'
+    }
+
+
+class QueryScopeMiddleware:
+    """HTTP scope challenges before JSON-RPC dispatch; SDK validates mirrored headers."""
+
+    def __init__(self, app, verifier: McpTokenVerifier, settings: McpSettings):
+        self.app, self.verifier, self.settings = app, verifier, settings
+
+    async def __call__(self, scope, receive, send):
+        if scope["type"] != "http" or scope.get("path") != "/mcp":
+            return await self.app(scope, receive, send)
+        headers = {
+            k.decode("latin1").lower(): v.decode("latin1")
+            for k, v in scope.get("headers", [])
+        }
+        auth = headers.get("authorization", "")
+        try:
+            token = (
+                await self.verifier.verify_token(auth[7:])
+                if auth.lower().startswith("bearer ")
+                else None
+            )
+        except (httpx.HTTPError, ValueError):
+            return await JSONResponse(
+                {"error": "authentication_unavailable"}, status_code=503
+            )(scope, receive, send)
+        required = (
+            QUERY_SCOPE
+            if headers.get("mcp-method") in {"tools/call", "resources/read"}
+            else CATALOG_SCOPE
+        )
+        if token is None:
+            return await JSONResponse(
+                {"error": "invalid_token"},
+                status_code=401,
+                headers=challenge(self.settings, required),
+            )(scope, receive, send)
+        if required not in token.scopes:
+            return await JSONResponse(
+                {"error": "insufficient_scope"},
+                status_code=403,
+                headers=challenge(self.settings, required, insufficient=True),
+            )(scope, receive, send)
+        await self.app(scope, receive, send)

app/mcp/catalog.py

@@ -0,0 +1,221 @@
+"""One finite query catalog; names, schemas and permissions come from this registry."""
+
+from dataclasses import dataclass
+from pydantic import BaseModel
+from mcp import types
+from app.mcp import models as m
+from app.mcp.analysis_models import BurstDetectionInput, BurstLocationInput, ValveIsolationInput, LeakageIdentificationInput
+from app.mcp.analysis_models import ScenarioInput, ContaminantSimulationInput, ValveClosureInput, FlushingInput, PressureRegulationInput, SensorPlacementInput
+
+
+@dataclass(frozen=True)
+class QueryTool:
+    name: str
+    description: str
+    model: type[BaseModel]
+    permission: str
+    semantics: str
+    task_only: bool = False
+    computation: bool = False
+
+    def protocol_tool(self) -> types.Tool:
+        if self.name == "run_simulation":
+            suffix = "。任务调用复用证据留存。"
+        elif self.computation:
+            suffix = "。仅计算建议，不保存方案、不控制设备；任务调用复用证据留存。"
+        elif self.name == "query_nearby_elements":
+            suffix = "。仅查询当前已授权项目的模型管网，按距离和元素ID稳定排序；按需分页，不自动遍历全量；空间邻近不代表管网连通或供水关系。"
+        else:
+            suffix = (
+                "。先限定对象和时间范围，再读取明细；集合默认20条，按需逐页扩大，"
+                "不要自动遍历全量；截断或非零offset不能证明完整范围无异常。"
+            )
+        return types.Tool(
+            name=self.name,
+            description=self.description + suffix,
+            inputSchema=self.model.model_json_schema(),
+            outputSchema=m.QueryResult.model_json_schema(),
+            annotations=types.ToolAnnotations(
+                readOnlyHint=False, destructiveHint=self.name == "run_simulation", openWorldHint=False
+            ),
+        )
+
+
+_specs = [
+    ("run_simulation", "按水力步长在在线业务模型上执行指定时间段的实时模拟，不复制数据库；更新该时段的模型条件及实时基线，返回指定节点和管段结果；duration单位为秒，不创建方案或正式分析运行，结果不是实测", ScenarioInput, "simulation.run", "analysis", False, True),
+    ("run_contaminant_simulation", "在隔离模型中模拟假设污染源的浓度传播；返回指定节点/管段结果，不反演污染源，不代表实测污染范围", ContaminantSimulationInput, "simulation.run", "analysis", False, True),
+    ("analyze_valve_closure", "在隔离模型中模拟关阀的水力影响，返回指定节点/管段结果；不操作真实阀门", ValveClosureInput, "simulation.run", "analysis", False, True),
+    ("analyze_flushing", "在隔离模型中评估指定排放节点、冲洗流量和关阀组合的水力结果；不是现场冲洗或消毒合格证明", FlushingInput, "simulation.run", "analysis", False, True),
+    ("analyze_pressure_regulation", "在隔离模型中评估给定调泵序列和水箱初始水位的压力影响；不自动优化调度，不实际控制泵", PressureRegulationInput, "simulation.run", "analysis", False, True),
+    ("optimize_sensor_placement", "使用灵敏度或K-means算法推荐压力监测点；仅返回候选节点，不保存正式方案，不保证污染溯源可观测性", SensorPlacementInput, "optimization.run", "analysis", False, True),
+    ("run_leakage_identification", "基于SCADA压力与模型进行DMA漏损分配识别，需有依据的总漏损流量和时间对齐；返回区域估计，不是已确认漏点。内部执行遗传算法与水力计算", LeakageIdentificationInput, "burst.run", "analysis", False, True),
+    ("run_burst_detection", "爆管异常侦测：使用至少5个完整压力传感器此前15天数据；只提示压力异常，不定位、不证明爆管", BurstDetectionInput, "burst.run", "analysis", False, True),
+    ("run_burst_location", "爆管候选定位：对比异常与正常压力，可辅助流量；需有依据的漏水量假设。压力节点可省略，自动读取项目SCADA压力测点；monitoring正常窗可成对省略，默认前一天同时间段，须说明该对照假设。simulation须有已有分析运行UUID并读取同窗模型基线，不是实测；定位内部会执行水力计算，不创建新的事件方案", BurstLocationInput, "burst.run", "analysis", False, True),
+    ("analyze_valve_isolation", "按事故候选要素分析拓扑隔离边界、建议关闭阀门和受影响节点；不是污染扩散或水力影响范围，不执行关阀，须现场核实", ValveIsolationInput, "simulation.run", "analysis", False, True),
+    (
+        "get_element_properties",
+        "读取当前项目的单个管网元素属性",
+        m.ElementInput,
+        "webgis.view",
+        "asset",
+    ),
+    (
+        "list_elements",
+        "分页读取管网元素，不自动读取全网",
+        m.ElementPageInput,
+        "webgis.view",
+        "asset",
+    ),
+    (
+        "query_nearby_elements",
+        "查询坐标附近的模型元素，默认pipe；center为EPSG:4326 [经度,纬度]，可直接使用geocode.center；radius_m默认200米、最大10000米，limit默认5、最大100，支持offset。按点到完整几何的WGS84椭球最短距离筛选并排序，返回element_id、distance_m和属性；不查询原始GIS管线、不自动扩大半径、不解析道路交叉点",
+        m.NearbyElementsInput,
+        "webgis.view",
+        "asset",
+    ),
+    (
+        "get_network_options",
+        "读取模型时间、能源或网络选项",
+        m.OptionsInput,
+        "webgis.view",
+        "asset",
+    ),
+    (
+        "get_network_schema",
+        "读取当前支持的元素、SCADA和选项结构",
+        m.SchemaInput,
+        "webgis.view",
+        "asset",
+    ),
+    (
+        "list_scada_devices",
+        "分页读取SCADA设备和物理单位",
+        m.PageInput,
+        "scada.view",
+        "asset",
+    ),
+    ("get_scada_device", "读取一个SCADA设备", m.DeviceInput, "scada.view", "asset"),
+    (
+        "query_scada_readings",
+        "查询带时区且最多24小时的原始或清洗监测数据",
+        m.ScadaInput,
+        "scada.view",
+        "observed",
+    ),
+    (
+        "query_model_baseline_series",
+        "读取最多20个元素的模型基线，不能称为实测",
+        m.BaselineSeriesInput,
+        "simulation.view",
+        "model_baseline",
+    ),
+    (
+        "get_model_baseline_at_time",
+        "读取单个元素指定时刻的模型基线",
+        m.BaselineAtTimeInput,
+        "simulation.view",
+        "model_baseline",
+    ),
+    (
+        "query_model_baseline_field",
+        "分页读取指定时刻的模型字段",
+        m.BaselineFieldInput,
+        "simulation.view",
+        "model_baseline",
+    ),
+    (
+        "query_analysis_series",
+        "读取指定分析运行和元素字段时序",
+        m.AnalysisSeriesInput,
+        "simulation.view",
+        "analysis",
+    ),
+    (
+        "query_analysis_values",
+        "分页读取指定分析运行某时刻的元素字段",
+        m.AnalysisValuesInput,
+        "simulation.view",
+        "analysis",
+    ),
+    (
+        "list_analysis_runs",
+        "分页读取持久分析运行",
+        m.RunsInput,
+        "webgis.view",
+        "analysis",
+    ),
+    ("get_analysis_run", "读取单个持久分析运行", m.RunInput, "webgis.view", "analysis"),
+    (
+        "list_analysis_results",
+        "分页读取分析成果，可筛选result_type",
+        m.ResultsInput,
+        "webgis.view",
+        "analysis",
+    ),
+    (
+        "list_sensor_placement_runs",
+        "分页读取已有监测点选址运行",
+        m.PageInput,
+        "optimization.run",
+        "analysis",
+    ),
+    (
+        "get_sensor_placement_run",
+        "读取已有监测点选址成果",
+        m.RunInput,
+        "optimization.run",
+        "analysis",
+    ),
+    (
+        "query_element_history",
+        "比较实测与模型基线或分析结果；analysis_comparison要求run_id；limit/offset按序列、时间排序的数据点分页",
+        m.HistoryInput,
+        "scada.view",
+        "observed",
+    ),
+    (
+        "get_element_gis_properties",
+        "读取元素关联GIS台账",
+        m.ElementInput,
+        "webgis.view",
+        "asset",
+    ),
+    (
+        "get_source_gis_properties",
+        "读取指定图层源要素台账",
+        m.SourceGisInput,
+        "webgis.view",
+        "asset",
+    ),
+    (
+        "get_workbench_runtime",
+        "读取工作台已有运行态",
+        m.StrictInput,
+        "simulation.view",
+        "task",
+    ),
+    (
+        "get_task_context",
+        "读取绑定任务的有界摘要及handling_skill元数据；section=skill读取Server按任务workflow分发的业务处置Skill正文与版本，属于流程指引而非事件证据；其他section选择task/report/evidence/actions/progress/access_requests/plan/run，data_path按返回路径读取嵌套字段，limit/offset分页；report_revision保留选定报告版本",
+        m.TaskContextInput,
+        "operational_task.view",
+        "task",
+        True,
+    ),
+    (
+        "query_peripheral_data",
+        "查询有限外围数据源；synthetic是合成证据",
+        m.PeripheralInput,
+        "simulation.view",
+        "peripheral",
+    ),
+    (
+        "get_task_evidence",
+        "读取当前绑定任务已保存的证据快照",
+        m.EvidenceInput,
+        "operational_task.view",
+        "task",
+        True,
+    ),
+]
+TOOLS = {row[0]: QueryTool(*row) for row in _specs}

app/mcp/config.py

@@ -0,0 +1,76 @@
+"""MCP is opt-in; an incomplete OAuth resource must never start listening."""
+
+from pathlib import Path
+from ipaddress import IPv4Address, IPv4Network
+from urllib.parse import urlsplit
+
+from pydantic import Field, model_validator
+from pydantic_settings import BaseSettings, SettingsConfigDict
+
+
+class McpSettings(BaseSettings):
+    enabled: bool = False
+    issuer_uri: str = ""
+    dev_http_issuer_uri: str = ""
+    environment: str = Field(default="production", validation_alias="ENVIRONMENT")
+    resource_uri: str = ""
+    audience: str = "tjwater-mcp"
+    host_client_id: str = "tjwater-agent-mcp"
+    allowed_hosts: list[str] = Field(
+        default_factory=lambda: ["127.0.0.1:*", "localhost:*"]
+    )
+    allowed_origins: list[str] = Field(default_factory=list)
+    context_ttl_seconds: int = Field(default=600, ge=30, le=600)
+    query_timeout_seconds: int = Field(default=30, ge=1, le=60)
+    max_calculation_timeout_seconds: int = Field(default=3600, ge=120, le=86400)
+    model_config = SettingsConfigDict(
+        env_prefix="TJWATER_MCP_",
+        extra="ignore",
+        populate_by_name=True,
+        env_file=Path(__file__).resolve().parents[2] / ".env",
+    )
+
+    @model_validator(mode="after")
+    def validate_resource(self):
+        if self.dev_http_issuer_uri:
+            dev_url = urlsplit(self.dev_http_issuer_uri)
+            try:
+                address = IPv4Address(dev_url.hostname or "")
+                private_ipv4 = any(address in IPv4Network(network) for network in (
+                    "10.0.0.0/8", "172.16.0.0/12", "192.168.0.0/16"
+                ))
+            except ValueError:
+                private_ipv4 = False
+            if (
+                self.environment != "development"
+                or dev_url.scheme != "http"
+                or not private_ipv4
+                or dev_url.username or dev_url.password or dev_url.query or dev_url.fragment
+                or self.dev_http_issuer_uri != self.issuer_uri
+            ):
+                raise ValueError("development HTTP issuer requires development environment and an exact private IPv4 issuer")
+        if not self.enabled:
+            return self
+        for index, value in enumerate((self.issuer_uri, self.resource_uri)):
+            url = urlsplit(value)
+            if (
+                url.scheme not in {"http", "https"}
+                or not url.hostname
+                or url.username
+                or url.password
+                or url.query
+                or url.fragment
+            ):
+                raise ValueError("MCP requires explicit issuer and resource URIs")
+            if url.scheme == "http" and url.hostname not in {
+                "localhost",
+                "127.0.0.1",
+                "::1",
+            }:
+                if index != 0 or value != self.dev_http_issuer_uri:
+                    raise ValueError("non-loopback MCP and OAuth require HTTPS")
+        if urlsplit(self.resource_uri).path != "/mcp":
+            raise ValueError("MCP resource URI must identify the /mcp endpoint")
+        if not self.audience.strip() or not self.host_client_id.strip():
+            raise ValueError("MCP audience and host client ID are required")
+        return self

app/mcp/contexts.py

@@ -0,0 +1,112 @@
+"""An OAuth identity and a Server-owned binding are both required for data queries."""
+
+import secrets
+from datetime import datetime, timedelta, timezone
+from uuid import UUID
+from fastapi import HTTPException
+from app.auth.permissions import (
+    permissions_for_context,
+    OPERATIONAL_TASK_EXECUTE,
+    OPERATIONAL_TASK_VIEW,
+)
+from app.infra.db.postgresql.operational_tasks import OperationalTaskRepository
+from app.infra.db.postgresql.mcp_queries import McpQueryRepository
+from app.mcp.models import BindInput, BoundContext
+from app.mcp.scope import QueryScope
+from app.mcp.config import McpSettings
+from mcp.server.auth.provider import AccessToken
+
+
+def task_repository(scope: QueryScope) -> OperationalTaskRepository:
+    return OperationalTaskRepository(scope.connection, scope.project.project_id)
+
+
+async def verify_task(scope: QueryScope, binding: BoundContext) -> None:
+    permissions = permissions_for_context(scope.project)
+    if OPERATIONAL_TASK_VIEW not in permissions:
+        raise HTTPException(403, "Task read permission is required")
+    repo = task_repository(scope)
+    async with scope.connection.transaction():
+        task = await repo.get_task(str(binding.task_id))
+        if not task:
+            raise HTTPException(404, "Task is not in this project")
+        if binding.purpose == "task_collaboration":
+            return
+        if OPERATIONAL_TASK_EXECUTE not in permissions:
+            raise HTTPException(403, "Task execution permission is required")
+        run = await repo.get_active_run(str(binding.task_id))
+    now = datetime.now(timezone.utc)
+    owner = binding.parent_session_id or binding.native_session_id
+    if (
+        not run
+        or run["status"] not in {"claimed", "running"}
+        or run.get("agent_session_id") != owner
+        or not run.get("lease_expires_at")
+        or run["lease_expires_at"] <= now
+        or binding.run_id is not None
+        and str(run["run_id"]) != str(binding.run_id)
+        or binding.execution_request_id is not None
+        and str(run["execution_request_id"]) != str(binding.execution_request_id)
+    ):
+        raise HTTPException(
+            409, "Investigation run is no longer active for this session"
+        )
+    binding.run_id = UUID(str(run["run_id"]))
+    binding.execution_request_id = UUID(str(run["execution_request_id"]))
+
+
+async def bind_context(
+    scope: QueryScope, principal: AccessToken, command: BindInput, settings: McpSettings
+) -> BoundContext:
+    if principal.client_id != settings.host_client_id:
+        raise HTTPException(
+            403, "Only the trusted product host may bind query contexts"
+        )
+    from app.mcp.catalog import TOOLS
+
+    binding = BoundContext(
+        context_handle=secrets.token_urlsafe(32),
+        project_id=scope.project.project_id,
+        subject=principal.subject,
+        client_id=principal.client_id,
+        tool_cap=sorted(name for name, tool in TOOLS.items() if command.purpose != "task_reader" or not tool.computation),
+        permission_cap=sorted(permissions_for_context(scope.project)),
+        native_session_id=command.native_session_id,
+        purpose=command.purpose,
+        parent_session_id=command.parent_session_id,
+        task_id=command.task_id,
+        expires_at=datetime.now(timezone.utc)
+        + timedelta(seconds=settings.context_ttl_seconds),
+    )
+    if binding.task_id:
+        await verify_task(scope, binding)
+        from app.infra.db.postgresql.read_queries import task_data_scopes
+
+        async with scope.connection.transaction():
+            binding.task_scope_cap = await task_data_scopes(scope.connection,
+                project_id=binding.project_id, task_id=binding.task_id,
+                permissions=permissions_for_context(scope.project))
+    await McpQueryRepository(scope.connection, scope.project.project_id).create(binding)
+    return binding
+
+
+async def resolve_context(
+    scope: QueryScope, principal: AccessToken, handle: str
+) -> BoundContext:
+    binding = await McpQueryRepository(scope.connection, scope.project.project_id).get(
+        handle
+    )
+    if (
+        not binding
+        or binding.revoked_at
+        or binding.expires_at <= datetime.now(timezone.utc)
+        or binding.subject != principal.subject
+        or binding.client_id != principal.client_id
+        or binding.project_id != scope.project.project_id
+    ):
+        raise HTTPException(
+            403, "Query context is unavailable for this identity and project"
+        )
+    if binding.task_id:
+        await verify_task(scope, binding)
+    return binding

app/mcp/dispatch.py

@@ -0,0 +1,15 @@
+"""Delegate MCP transport/metadata with the original request path intact."""
+
+
+class McpDispatchMiddleware:
+    def __init__(self, app, runtime):
+        self.app, self.runtime = app, runtime
+
+    async def __call__(self, scope, receive, send):
+        path = scope.get("path", "")
+        if scope["type"] == "http" and (
+            path == "/mcp" or path.startswith("/.well-known/oauth-protected-resource")
+        ):
+            await self.runtime.asgi(scope, receive, send)
+        else:
+            await self.app(scope, receive, send)

app/mcp/models.py

@@ -0,0 +1,293 @@
+"""Strict business inputs and provenance envelopes shared by MCP handlers."""
+
+from datetime import datetime, timedelta
+from typing import Annotated, Literal
+from uuid import UUID
+
+from pydantic import BaseModel, ConfigDict, Field, JsonValue, model_validator
+
+Id = Annotated[str, Field(min_length=1, max_length=160, pattern=r"^\S(?:.*\S)?$")]
+ElementType = Literal["junction", "pipe", "valve", "pump", "tank", "reservoir"]
+SeriesType = Literal["junction", "pipe"]
+
+
+class StrictInput(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+    context_handle: Annotated[str, Field(min_length=32, max_length=128)]
+
+
+class PageInput(StrictInput):
+    limit: int = Field(default=20, ge=1, le=500)
+    offset: int = Field(default=0, ge=0, le=10000)
+
+
+class TaskContextInput(PageInput):
+    section: Literal["summary", "task", "report", "evidence", "actions", "progress", "access_requests", "plan", "run", "skill"] = "summary"
+    report_revision: int | None = Field(default=None, ge=1)
+    data_path: list[Id] = Field(default_factory=list, max_length=8)
+
+    @model_validator(mode="after")
+    def summary_is_fixed(self):
+        if self.section in {"summary", "skill"} and (self.data_path or self.offset):
+            raise ValueError("Select a detail section before using data_path or offset")
+        return self
+
+
+class ElementInput(StrictInput):
+    element_type: ElementType
+    element_id: Id
+
+
+class ElementPageInput(PageInput):
+    element_type: ElementType
+
+
+class NearbyElementsInput(PageInput):
+    element_type: ElementType = "pipe"
+    center: tuple[
+        Annotated[float, Field(ge=-180, le=180, allow_inf_nan=False)],
+        Annotated[float, Field(ge=-90, le=90, allow_inf_nan=False)],
+    ] = Field(description="EPSG:4326 [longitude, latitude], directly from geocode.center")
+    radius_m: float = Field(default=200, gt=0, le=10000, allow_inf_nan=False)
+    limit: int = Field(default=5, ge=1, le=100)
+
+
+class OptionsInput(StrictInput):
+    kind: Literal["time", "energy", "pump-energy", "network"]
+    pump_id: Id | None = None
+
+    @model_validator(mode="after")
+    def pump_required(self):
+        if self.kind == "pump-energy" and not self.pump_id:
+            raise ValueError("pump_id is required for pump-energy")
+        if self.kind != "pump-energy" and self.pump_id:
+            raise ValueError("pump_id is only valid for pump-energy")
+        return self
+
+
+class SchemaInput(StrictInput):
+    kind: Literal[
+        "junction",
+        "pipe",
+        "valve",
+        "pump",
+        "tank",
+        "reservoir",
+        "time",
+        "energy",
+        "pump-energy",
+        "option",
+        "scada-device",
+    ]
+
+
+class DeviceInput(StrictInput):
+    device_id: Id
+
+
+class WindowInput(PageInput):
+    start: datetime
+    end: datetime
+    limit: int = Field(default=20, ge=1, le=10000)
+
+    @model_validator(mode="after")
+    def window_valid(self):
+        if self.start.tzinfo is None or self.end.tzinfo is None:
+            raise ValueError("timestamps require explicit timezone")
+        if not timedelta(0) < self.end - self.start <= timedelta(days=1):
+            raise ValueError("time window must be positive and at most 24 hours")
+        return self
+
+
+class ScadaInput(WindowInput):
+    device_ids: list[Id] = Field(min_length=1, max_length=20)
+    field: Literal["monitored_value", "cleaned_value"] = "monitored_value"
+
+
+class BaselineSeriesInput(WindowInput):
+    element_type: SeriesType
+    element_ids: list[Id] = Field(min_length=1, max_length=20)
+
+
+class TimeInput(StrictInput):
+    time: datetime
+
+    @model_validator(mode="after")
+    def timezone_required(self):
+        if self.time.tzinfo is None:
+            raise ValueError("timestamp requires explicit timezone")
+        return self
+
+
+class BaselineAtTimeInput(TimeInput):
+    element_type: SeriesType
+    element_id: Id
+
+
+class BaselineFieldInput(TimeInput):
+    element_type: SeriesType
+    field: Id
+    limit: int = Field(default=20, ge=1, le=500)
+    offset: int = Field(default=0, ge=0, le=10000)
+
+
+class AnalysisSeriesInput(WindowInput):
+    run_id: UUID
+    element_type: SeriesType
+    element_id: Id
+    field: Id
+
+
+class AnalysisValuesInput(BaselineFieldInput):
+    run_id: UUID
+
+
+class RunInput(StrictInput):
+    run_id: UUID
+
+
+class RunsInput(PageInput):
+    limit: int = Field(default=20, ge=1, le=100)
+
+
+class ResultsInput(PageInput):
+    run_id: UUID
+    result_type: Id | None = None
+
+
+class HistoryTarget(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+    element_type: SeriesType
+    element_id: Id
+    device_ids: list[Id] | None = Field(default=None, min_length=1, max_length=20)
+
+
+class HistoryInput(WindowInput):
+    elements: list[HistoryTarget] = Field(min_length=1, max_length=20)
+    mode: Literal["observed", "realtime_comparison", "analysis_comparison"] = "observed"
+    run_id: UUID | None = None
+
+    @model_validator(mode="after")
+    def analysis_required(self):
+        if (self.mode == "analysis_comparison") != (self.run_id is not None):
+            raise ValueError("run_id is required only for analysis_comparison")
+        return self
+
+
+class SourceGisInput(StrictInput):
+    layer_key: Annotated[str, Field(min_length=1, max_length=80)]
+    source_fid: int = Field(ge=0, le=9223372036854775807)
+
+
+class PeripheralInput(PageInput):
+    source: Literal[
+        "peripheral_points",
+        "peripheral_work_orders",
+        "peripheral_observations",
+        "peripheral_test_data",
+    ]
+    point_id: Id | None = None
+    start: datetime | None = None
+    end: datetime | None = None
+
+    @model_validator(mode="after")
+    def observation_window(self):
+        if self.source == "peripheral_observations":
+            if not self.point_id or self.start is None or self.end is None:
+                raise ValueError("observations require point_id, start and end")
+            if (
+                self.start.tzinfo is None
+                or self.end.tzinfo is None
+                or not timedelta(0) < self.end - self.start <= timedelta(days=1)
+            ):
+                raise ValueError(
+                    "observation window requires timezone and at most 24 hours"
+                )
+        elif (
+            self.point_id is not None or self.start is not None or self.end is not None
+        ):
+            raise ValueError(
+                "point_id and window are only valid for peripheral_observations"
+            )
+        return self
+
+
+class EvidenceInput(PageInput):
+    evidence_id: Id
+    data_path: list[Id] = Field(default_factory=list, max_length=8)
+
+
+class BindInput(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+    native_session_id: Id
+    purpose: Literal[
+        "ordinary_chat", "task_execution", "task_reader", "task_collaboration"
+    ]
+    task_id: UUID | None = None
+    parent_session_id: Id | None = None
+
+    @model_validator(mode="after")
+    def task_scope(self):
+        if self.purpose == "ordinary_chat" and (self.task_id or self.parent_session_id):
+            raise ValueError("ordinary chat cannot bind a task")
+        if self.purpose != "ordinary_chat" and self.task_id is None:
+            raise ValueError("task context requires task_id")
+        if (self.purpose == "task_reader") != (self.parent_session_id is not None):
+            raise ValueError("only a task reader requires parent_session_id")
+        return self
+
+
+class BindResponse(BaseModel):
+    context_handle: str
+    expires_at: datetime
+    task_id: UUID | None
+    run_id: UUID | None
+
+
+class PageInfo(BaseModel):
+    limit: int
+    offset: int
+    total: int | None = None
+    has_more: bool
+
+
+class QueryResult(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+    source: str
+    source_ref: str
+    queried_at: datetime
+    origin: Literal["system_data", "synthetic"] = "system_data"
+    semantics: Literal[
+        "asset",
+        "observed",
+        "model_baseline",
+        "analysis",
+        "comparison",
+        "task",
+        "peripheral",
+    ]
+    data: JsonValue
+    page: PageInfo | None = None
+    window: dict[str, datetime] | None = None
+    units: dict[str, str] = Field(default_factory=dict)
+    truncated: bool = False
+    data_gaps: list[str] = Field(default_factory=list)
+    evidence_id: str | None = None
+
+
+class BoundContext(BaseModel):
+    context_handle: str
+    project_id: UUID
+    subject: str
+    client_id: str
+    tool_cap: list[str] = Field(default_factory=list)
+    permission_cap: list[str] = Field(default_factory=list)
+    task_scope_cap: list[str] = Field(default_factory=list)
+    native_session_id: str
+    purpose: str
+    parent_session_id: str | None = None
+    task_id: UUID | None = None
+    run_id: UUID | None = None
+    execution_request_id: UUID | None = None
+    expires_at: datetime
+    revoked_at: datetime | None = None

app/mcp/peripheral.py

@@ -0,0 +1,132 @@
+"""Read existing project-owned peripheral facts without forwarding MCP tokens."""
+
+from datetime import datetime, timezone
+from uuid import uuid4
+from fastapi import HTTPException
+from app.infra.db.postgresql.read_queries import MAX_QUERY_BYTES, read_rows, scada_devices, series_rows
+from app.domain.schemas.operational_data import TaskDataSnapshot, TaskDataQuery
+from app.services.content_digest import digest
+from samples.peripheral.store import PostgresStore
+from app.infra.integrations.operational_data import _select_demo_observations
+
+
+async def read_peripheral_query(scope, args):
+    conn = scope.connection
+    project = str(scope.project.project_id)
+    truncated = False
+    origin = "system_data"
+    if args.source in {"peripheral_work_orders", "peripheral_test_data"}:
+        table, column = (
+            ("work_orders", "work_order_id")
+            if args.source == "peripheral_work_orders"
+            else ("test_data_records", "record_id")
+        )
+        rows = await read_rows(
+            conn,
+            "SELECT document FROM peripheral_sample."
+            + table
+            + " WHERE project_id=%s AND deleted_at IS NULL ORDER BY "
+            + column
+            + " LIMIT %s OFFSET %s",
+            (project, args.limit + 1, args.offset),
+        )
+        items = [
+            PostgresStore._active_managed_document(row["document"])
+            for row in rows[: args.limit]
+        ]
+        truncated = len(rows) > args.limit
+        if any(item.get("project_id") != project for item in items):
+            raise HTTPException(502, "Peripheral project mismatch")
+        data = {
+            "items": items,
+            "truncated": truncated,
+            "limit": args.limit,
+            "offset": args.offset,
+        }
+        origin = (
+            "synthetic"
+            if items and any(item.get("origin") == "synthetic" for item in items)
+            else "system_data"
+        )
+    elif args.source == "peripheral_points":
+        rows = await scada_devices(conn, limit=args.limit + 1, offset=args.offset)
+        data = {
+            "origin": "historical_replay",
+            "items": [
+                {
+                    "point_id": row["device_id"],
+                    "label": row["device_id"] + " · " + row["device_type"],
+                    "unit": row["measurement_unit"],
+                    "status": "unknown",
+                    "origin": "historical_replay",
+                }
+                for row in rows[: args.limit]
+            ],
+            "truncated": len(rows) > args.limit,
+        }
+        truncated = data["truncated"]
+    elif args.point_id.startswith("DEMO-"):
+        rows = await read_rows(
+            conn,
+            "SELECT CASE WHEN octet_length(document::text)<=%s THEN document END AS document,"
+            "digest,octet_length(document::text)>%s AS too_large "
+            "FROM peripheral_sample.demo_datasets WHERE project_id=%s AND dataset_id='water-demo-v1'",
+            (MAX_QUERY_BYTES, MAX_QUERY_BYTES, project),
+        )
+        if rows and rows[0]["too_large"]:
+            raise HTTPException(413, "Synthetic dataset exceeds the integrity-check byte budget")
+        if not rows or digest(rows[0]["document"]) != rows[0]["digest"]:
+            raise HTTPException(503, "Synthetic dataset is unavailable or invalid")
+        query = TaskDataQuery(
+            source=args.source, point_id=args.point_id, start=args.start, end=args.end
+        )
+        data = _select_demo_observations(
+            {**rows[0]["document"], "project_id": project}, query, project
+        )
+        items = data["items"]
+        truncated = len(items) > args.limit + args.offset
+        data["items"] = items[args.offset : args.offset + args.limit]
+        data["truncated"] = truncated
+        origin = "synthetic"
+    else:
+        devices = await scada_devices(conn, device_id=args.point_id, limit=1)
+        if not devices:
+            raise HTTPException(404, "Peripheral point was not found")
+        device = devices[0]
+        if not device.get("api_query_id"):
+            raise HTTPException(503, "Peripheral history mapping is unavailable")
+        async with scope.timescale() as ts:
+            rows = await series_rows(
+                ts,
+                source="scada",
+                element_type="junction",
+                ids=[device["api_query_id"]],
+                start=args.start,
+                end=args.end,
+                field="monitored_value",
+                limit=args.limit + 1,
+                offset=args.offset,
+                end_exclusive=True,
+            )
+        truncated = len(rows) > args.limit
+        data = {
+            "origin": "historical_replay",
+            "items": [
+                {
+                    "point_id": args.point_id,
+                    "observed_at": row["time"].isoformat(),
+                    "value": row["value"],
+                    "unit": device["measurement_unit"],
+                }
+                for row in rows[: args.limit]
+            ],
+            "truncated": truncated,
+        }
+    return TaskDataSnapshot(
+        source=args.source,
+        source_ref="mcp:" + str(uuid4()),
+        queried_at=datetime.now(timezone.utc),
+        data=data,
+        origin=origin,
+        truncated=truncated,
+    )

app/mcp/query_service.py

@@ -0,0 +1,492 @@
+"""MCP domain dispatch reuses Server repositories and services, never an HTTP loopback."""
+
+import asyncio
+import math
+from contextlib import nullcontext
+from typing import Any
+from pydantic import BaseModel
+from app.mcp.scope import QueryScope
+from app.mcp.catalog import QueryTool, TOOLS
+from app.mcp.models import BoundContext
+from datetime import datetime, timezone
+from uuid import uuid4
+from fastapi import HTTPException
+from fastapi.encoders import jsonable_encoder
+from app.auth.permissions import permissions_for_context
+from app.infra.db.postgresql import read_queries as reads
+from app.infra.db.postgresql.spatial_queries import nearby_elements
+from app.infra.db.postgresql.analysis import AnalysisRepository
+from app.infra.db.postgresql.gis_attributes import GisAttributeRepository
+from app.infra.db.postgresql.network_settings import NetworkSettingsRepository
+from app.infra.db.postgresql.operational_tasks import OperationalTaskRepository
+from app.services.operational_tasks import OperationalTaskService
+from app.mcp.models import QueryResult, PageInfo
+from app.mcp.scenario_service import SCENARIO_TOOLS
+from app.mcp.work_order_skills import read_work_order_skill
+
+
+def finite_data(value: Any) -> tuple[Any, int]:
+    """JSON cannot represent infinity/NaN; preserve the gap as null with provenance."""
+    if isinstance(value, float) and not math.isfinite(value):
+        return None, 1
+    count = 0
+    if isinstance(value, dict):
+        result = {}
+        for key, item in value.items():
+            result[key], replaced = finite_data(item)
+            count += replaced
+        return result, count
+    if isinstance(value, list):
+        result = []
+        for item in value:
+            clean, replaced = finite_data(item)
+            result.append(clean)
+            count += replaced
+        return result, count
+    return value, 0
+
+
+def task_service(scope: QueryScope) -> OperationalTaskService:
+    return OperationalTaskService(
+        OperationalTaskRepository(scope.connection, scope.project.project_id),
+        scope.project,
+        set(permissions_for_context(scope.project)),
+    )
+
+
+def page_result(rows, args):
+    has_more = len(rows) > args.limit
+    return (
+        rows[: args.limit],
+        PageInfo(limit=args.limit, offset=args.offset, total=None, has_more=has_more),
+        has_more,
+    )
+
+
+async def authorize_query(
+    scope: QueryScope, binding: BoundContext, tool: QueryTool, args: BaseModel
+) -> None:
+    if tool.name not in binding.tool_cap:
+        raise HTTPException(403, "Query is outside the captured tool catalog")
+    permissions = permissions_for_context(scope.project)
+    if tool.computation:
+        if tool.name == "optimize_sensor_placement" and scope.project.project_role != "member":
+            raise HTTPException(403, "Sensor placement requires project member access")
+        if binding.purpose == "task_reader":
+            raise HTTPException(403, "Read-only task agents cannot start calculations")
+        required = {tool.permission}
+        if tool.name in SCENARIO_TOOLS:
+            required.add("scada.view")
+        if tool.name in {"run_burst_detection", "run_burst_location", "run_leakage_identification"}:
+            required.add("simulation.view" if getattr(args, "data_source", "monitoring") == "simulation" else "scada.view")
+        if not required.issubset(permissions) or not required.issubset(binding.permission_cap):
+            raise HTTPException(403, "Calculation and input data permissions are required")
+    permission = (
+        "scada.view"
+        if tool.name == "get_network_schema" and args.kind == "scada-device"
+        else tool.permission
+    )
+    if tool.task_only and binding.task_id is None:
+        raise HTTPException(403, "This query requires a bound task")
+    granted = permission in permissions and permission in binding.permission_cap
+    source = {
+        "list_scada_devices": "scada_devices",
+        "get_scada_device": "scada_devices",
+        "query_scada_readings": "scada_readings",
+        "query_element_history": "scada_readings",
+        "get_workbench_runtime": "workbench_runtime",
+    }.get(tool.name)
+    if tool.name == "query_peripheral_data":
+        source = args.source
+    if not granted and binding.task_id and source in binding.task_scope_cap:
+        async with scope.connection.transaction():
+            granted = await task_service(scope).can_read_source(binding.task_id, source)
+    if not granted:
+        raise HTTPException(403, "Data source permission is required")
+    if (
+        tool.name == "query_element_history"
+        and args.mode != "observed"
+        and (
+            "simulation.view" not in permissions
+            or "simulation.view" not in binding.permission_cap
+        )
+    ):
+        raise HTTPException(403, "Model comparison permission is required")
+
+
+async def execute_query(
+    scope: QueryScope,
+    binding: BoundContext,
+    tool: QueryTool,
+    args: BaseModel,
+    *,
+    full_snapshot: bool = False,
+) -> QueryResult:
+    await authorize_query(scope, binding, tool, args)
+    name = tool.name
+    conn = scope.connection
+    page = None
+    truncated = False
+    units = {}
+    origin = "system_data"
+    gaps = []
+    source_ref = "mcp:" + str(uuid4())
+    evidence_id = None
+    if tool.computation:
+        from app.mcp.analysis_service import TIMESCALE_CALCULATIONS, calculate
+
+        # Context/receipt reads may have started an implicit transaction. Release
+        # it before potentially long CPU/IO work; evidence uses a new transaction.
+        await conn.commit()
+        async with (
+            scope.simulation_routing()
+            if name in TIMESCALE_CALCULATIONS
+            else nullcontext()
+        ):
+            data = await calculate(
+                name, args, network=scope.project.project_code,
+                username=scope.user.username,
+            )
+        data = {
+            "parameters": args.model_dump(mode="json", exclude={"context_handle"}),
+            "result": data,
+        }
+        gaps = [{
+            "run_simulation": "Online hydraulic recalculation updated the realtime baseline; results cover requested elements only, not observed SCADA data. No standalone analysis run was saved.",
+            "run_contaminant_simulation": "Forward model simulation for an assumed source, not source inversion or observed contamination; only requested elements are returned.",
+            "analyze_valve_closure": "Simulated hydraulic impact for requested elements only; no physical valves were operated.",
+            "analyze_flushing": "Simulated flushing hydraulics only; no field action or water-quality clearance was verified.",
+            "analyze_pressure_regulation": "Evaluation of a supplied pump schedule, not automatic optimization or device control; requested elements only.",
+            "optimize_sensor_placement": "Candidate pressure monitoring locations, not a saved deployment scheme or a guarantee of contaminant source observability.",
+            "run_leakage_identification": "DMA leakage estimates depend on the total leakage constraint and aligned model/SCADA time; they do not confirm an exact leak location.",
+            "run_burst_detection": "Pressure anomalies alone do not confirm a burst or locate a pipe.",
+            "run_burst_location": "Candidate ranking depends on the leakage assumption, model and sensor coverage; field confirmation is required.",
+            "analyze_valve_isolation": "Topology-only isolation advice; valve operability, hydraulic impact and field actions are not verified or executed.",
+        }[name]]
+        if getattr(args, "data_source", None) == "simulation":
+            gaps.append("Input is an existing analysis run and model baseline, not observed SCADA data.")
+    elif name == "query_nearby_elements":
+        rows = await nearby_elements(
+            conn, args.element_type, center=args.center, radius_m=args.radius_m,
+            limit=args.limit + 1, offset=args.offset,
+        )
+        items, page, truncated = page_result(rows, args)
+        truncated = truncated or args.offset > 0
+        data = {
+            "source": "model_network", "element_type": args.element_type,
+            "center": list(args.center), "center_crs": "EPSG:4326",
+            "radius_m": args.radius_m, "distance_method": "WGS84_spheroid",
+            "items": items,
+        }
+        units = {"distance_m": "m", "radius_m": "m"}
+        gaps = ["Spatial proximity does not establish hydraulic connectivity or a supply relationship"]
+    elif name in {"get_element_properties", "list_elements"}:
+        rows = await reads.asset_rows(
+            conn,
+            args.element_type,
+            element_id=getattr(args, "element_id", None),
+            limit=getattr(args, "limit", 1) + 1,
+            offset=getattr(args, "offset", 0),
+        )
+        if name == "get_element_properties":
+            if not rows:
+                raise HTTPException(404, "Element was not found")
+            data = rows[0]
+        else:
+            data, page, truncated = page_result(rows, args)
+    elif name in {"list_scada_devices", "get_scada_device"}:
+        rows = await reads.scada_devices(
+            conn,
+            device_id=getattr(args, "device_id", None),
+            limit=getattr(args, "limit", 1) + 1,
+            offset=getattr(args, "offset", 0),
+        )
+        if name == "get_scada_device":
+            if not rows:
+                raise HTTPException(404, "SCADA device was not found")
+            data = rows[0]
+        else:
+            data, page, truncated = page_result(rows, args)
+    elif name == "get_network_schema":
+        from app.services import tjnetwork
+
+        mapping = {
+            "option": "get_option_v3_schema",
+            "scada-device": "get_scada_info_schema",
+            "pump-energy": "get_pump_energy_schema",
+        }
+        function = getattr(
+            tjnetwork, mapping.get(args.kind, "get_" + args.kind + "_schema")
+        )
+        data = await asyncio.to_thread(function, scope.project.project_code)
+    elif name == "get_network_options":
+        if args.kind == "pump-energy":
+            rows = await reads.read_rows(
+                conn,
+                "SELECT price,pattern_id AS pattern,efficiency_curve_id AS effic FROM network.pump_energy_settings WHERE pump_id=%s",
+                (args.pump_id,),
+            )
+            data = {"pump": args.pump_id, "price": None, "pattern": None, "effic": None}
+            if rows:
+                data.update(rows[0])
+        else:
+            table = {
+                "time": "time_settings",
+                "energy": "energy_settings",
+                "network": "simulation_settings",
+            }[args.kind]
+            where = " WHERE engine_version='v3'" if args.kind == "network" else ""
+            rows = await reads.read_rows(
+                conn,
+                "SELECT key,value FROM network."
+                + table
+                + where
+                + " ORDER BY key LIMIT 501",
+            )
+            data = {str(row["key"]): str(row["value"]) for row in rows[:500]}
+            truncated = len(rows) > 500
+    elif name in {
+        "query_scada_readings",
+        "query_model_baseline_series",
+        "get_model_baseline_at_time",
+        "query_model_baseline_field",
+        "query_analysis_series",
+        "query_analysis_values",
+    }:
+        source = (
+            "scada"
+            if name == "query_scada_readings"
+            else "analysis"
+            if name.startswith("query_analysis")
+            else "realtime"
+        )
+        ids = (
+            getattr(args, "device_ids", None)
+            or getattr(args, "element_ids", None)
+            or ([args.element_id] if hasattr(args, "element_id") else [])
+        )
+        async with scope.timescale() as ts:
+            rows = await reads.series_rows(
+                ts,
+                source=source,
+                element_type=getattr(args, "element_type", "junction"),
+                ids=ids,
+                start=getattr(args, "start", None),
+                end=getattr(args, "end", None),
+                time=getattr(args, "time", None),
+                field=getattr(args, "field", None),
+                run_id=getattr(args, "run_id", None),
+                limit=getattr(args, "limit", 1) + 1,
+                offset=getattr(args, "offset", 0),
+            )
+        if name == "get_model_baseline_at_time":
+            data = rows[:1]
+        else:
+            data, page, truncated = page_result(rows, args)
+        if source != "scada":
+            async with conn.transaction():
+                units = await NetworkSettingsRepository.get_result_units(conn)
+                if source == "analysis":
+                    run = await AnalysisRepository.get_run(conn, args.run_id)
+                    if not run:
+                        raise HTTPException(404, "Analysis run was not found")
+                    recorded = (run.get("parameters") or {}).get("result_units")
+                    if isinstance(recorded, dict):
+                        units = recorded
+                    else:
+                        gaps.append(
+                            "Analysis units inferred from current network settings"
+                        )
+                if hasattr(args, "field") and args.field not in units:
+                    gaps.append(f"Unit metadata is unavailable for field {args.field}")
+        else:
+            metadata = [
+                r
+                for device in ids
+                for r in await reads.scada_devices(conn, device_id=device, limit=1)
+            ]
+            units = {
+                r["device_id"]: r["measurement_unit"]
+                for r in metadata
+                if r["device_id"] in ids
+            }
+    elif name in {"list_analysis_runs", "get_analysis_run", "list_analysis_results"}:
+        async with conn.transaction():
+            if name == "get_analysis_run":
+                data = await AnalysisRepository.get_run(conn, args.run_id)
+                if not data:
+                    raise HTTPException(404, "Analysis run was not found")
+            elif name == "list_analysis_runs":
+                # Fetch one additional row with the repository's supported summary-page bound.
+                rows = await AnalysisRepository.list_run_summaries(
+                    conn, limit=args.limit + 1, offset=args.offset
+                )
+                data, page, truncated = page_result(rows, args)
+            else:
+                rows = await AnalysisRepository.list_results(
+                    conn,
+                    args.run_id,
+                    args.result_type,
+                    limit=args.limit,
+                    lookahead=True,
+                    offset=args.offset,
+                )
+                data, page, truncated = page_result(rows, args)
+    elif name in {"list_sensor_placement_runs", "get_sensor_placement_run"}:
+        rows = await reads.sensor_runs(
+            conn,
+            run_id=getattr(args, "run_id", None),
+            limit=getattr(args, "limit", 1) + 1,
+            offset=getattr(args, "offset", 0),
+        )
+        if name == "get_sensor_placement_run":
+            if not rows:
+                raise HTTPException(404, "Sensor placement run was not found")
+            data = rows[0]
+        else:
+            data, page, truncated = page_result(rows, args)
+    elif name == "query_element_history":
+        from app.domain.schemas.timeseries_history import ElementHistoryQuery
+        from app.services.timeseries_history import TimeseriesHistoryService
+
+        query = ElementHistoryQuery(
+            start_time=args.start,
+            end_time=args.end,
+            mode=args.mode,
+            run_id=args.run_id,
+            elements=[e.model_dump() for e in args.elements],
+        )
+        async with scope.timescale() as ts, conn.transaction(), ts.transaction():
+            result = await TimeseriesHistoryService.query(
+                ts, conn, query, row_limit=args.limit, row_offset=args.offset
+            )
+        data = result.model_dump(mode="json")
+        truncated = result.truncated
+        page = PageInfo(limit=args.limit, offset=args.offset, has_more=truncated)
+        truncated = truncated or args.offset > 0
+    elif name in {"get_element_gis_properties", "get_source_gis_properties"}:
+        async with conn.transaction():
+            if name == "get_element_gis_properties":
+                data = await GisAttributeRepository.list_element_groups(
+                    conn, args.element_type, args.element_id
+                )
+            else:
+                data = await GisAttributeRepository.get_source_feature_group(
+                    conn, args.layer_key, args.source_fid
+                )
+                if data is None:
+                    raise HTTPException(404, "Source GIS feature was not found")
+    elif name == "get_task_context":
+        if args.section in {"summary", "skill"}:
+            data = await reads.task_context_summary(conn, project_id=binding.project_id,
+                task_id=binding.task_id, report_revision=args.report_revision)
+            skill = read_work_order_skill(
+                data.get("task", {}).get("workflow"), include_content=args.section == "skill"
+            )
+            if args.section == "skill":
+                data = {"section": "skill", "skill": skill}
+            else:
+                data["handling_skill"] = skill
+                data["display_time_zone"] = "Asia/Shanghai"
+                data["approved_data_scopes"] = binding.task_scope_cap
+                data["bound_run_id"] = str(binding.run_id) if binding.run_id else None
+                data["available_sections"] = ["task", "report", "evidence", "actions", "progress", "access_requests", "plan", "run", "skill"]
+                truncated = True
+                gaps.append("Task context is a bounded summary; use section, data_path, limit and offset for detail. Previews are not complete reports.")
+        else:
+            projection = await reads.task_context_page(conn, project_id=binding.project_id, task_id=binding.task_id,
+                section=args.section, report_revision=args.report_revision, data_path=args.data_path, limit=args.limit, offset=args.offset)
+            total = projection["total"]
+            if args.section in {"evidence", "actions", "access_requests"} and not args.data_path:
+                total = projection["snapshot_metadata"].get("total", total)
+            if projection["value_type"] in {"array", "object", "string"}:
+                page = PageInfo(limit=args.limit, offset=args.offset, total=total, has_more=args.offset + args.limit < total)
+            elif args.offset:
+                raise HTTPException(422, "Scalar task detail cannot use an offset")
+            truncated = bool(projection.get("nested") or args.offset or page and page.has_more)
+            data = {"section": args.section, "selected_path": args.data_path, "value": projection["value"], "metadata": projection["snapshot_metadata"]}
+            if projection.get("nested"):
+                gaps.append("Nested values have data_path descriptors; select a descriptor path for its values")
+    elif name == "get_task_evidence":
+        projection = await reads.evidence_page(
+            conn, project_id=scope.project.project_id, task_id=binding.task_id,
+            evidence_id=args.evidence_id, data_path=args.data_path,
+            limit=args.limit, offset=args.offset, mcp_sources=list(TOOLS),
+            full_snapshot=full_snapshot,
+        )
+        document = projection if full_snapshot else projection["snapshot_metadata"]
+        origin = document["origin"]
+        source_ref = document["source_ref"]
+        evidence_id = args.evidence_id
+        truncated = document.get("truncated", False)
+        if full_snapshot:
+            data = document
+        else:
+            selected = projection["value"]
+            if projection["value_type"] in {"array", "object", "string"}:
+                total = projection["total"]
+                page = PageInfo(
+                    limit=args.limit,
+                    offset=args.offset,
+                    total=total,
+                    has_more=args.offset + args.limit < total,
+                )
+                truncated = truncated or page.has_more or args.offset > 0
+            elif args.offset:
+                raise HTTPException(422, "Scalar evidence cannot use an offset")
+            if projection.get("nested"):
+                truncated = True
+                gaps.append(
+                    "Nested collections have data_path descriptors; query a descriptor path for their values"
+                )
+            data = {
+                "snapshot_metadata": document,
+                "query_metadata": projection["query_metadata"],
+                "selected_path": args.data_path,
+                "value": selected,
+            }
+    elif name == "get_workbench_runtime":
+        from app.services.workbench_runtime import read_workbench_runtime
+
+        data = (await read_workbench_runtime(scope.project, conn)).model_dump(
+            mode="json"
+        )
+    elif name == "query_peripheral_data":
+        from app.mcp.peripheral import read_peripheral_query
+
+        snapshot = await read_peripheral_query(scope, args)
+        data = snapshot.data
+        origin = snapshot.origin
+        truncated = snapshot.truncated
+    else:
+        raise ValueError("query handler is not registered")
+    window = (
+        {"start": args.start, "end": args.end}
+        if hasattr(args, "start") and args.start is not None
+        else None
+    )
+    data, nonfinite = finite_data(jsonable_encoder(data))
+    semantics = (
+        "comparison"
+        if name == "query_element_history" and args.mode != "observed"
+        else tool.semantics
+    )
+    return QueryResult(
+        source=name,
+        source_ref=source_ref,
+        evidence_id=evidence_id,
+        queried_at=datetime.now(timezone.utc),
+        semantics=semantics,
+        data=data,
+        data_gaps=gaps
+        + (
+            [f"{nonfinite} non-finite numeric values were returned as null"]
+            if nonfinite
+            else []
+        ),
+        page=page,
+        window=window,
+        units=units,
+        truncated=truncated,
+        origin=origin,
+    )

app/mcp/scenario_service.py

@@ -0,0 +1,70 @@
+"""MCP online recalculation and isolated scenario adapters."""
+
+from typing import Any
+
+SCENARIO_TOOLS = frozenset({
+    "run_simulation",
+    "run_contaminant_simulation", "analyze_valve_closure",
+    "analyze_flushing", "analyze_pressure_regulation",
+})
+
+
+def calculate_scenario(name: str, parameters: dict[str, Any], *, network: str, username: str) -> dict[str, Any]:
+    from app.services import simulation_scenarios as scenarios
+
+    parameters = dict(parameters)
+    nodes = parameters.pop("output_node_ids")
+    links = parameters.pop("output_link_ids")
+    start = parameters.pop("start_time")
+    duration = parameters.pop("duration")
+    common = dict(name=network, modify_pattern_start_time=start.isoformat(),
+                  modify_total_duration=duration, persist_results=False, scheme_name=None)
+    if name == "run_simulation":
+        from app.services.simulation import run_online_simulation
+
+        result = run_online_simulation(
+            network, start, duration, output_node_ids=nodes, output_link_ids=links,
+        )
+    elif name == "run_contaminant_simulation":
+        result = scenarios.contaminant_simulation(**common, username=username,
+            source=parameters["source"], concentration=parameters["concentration"], source_pattern=parameters.get("pattern"))
+    elif name == "analyze_valve_closure":
+        result = scenarios.valve_close_analysis(**common,
+            modify_valve_opening={valve: 0.0 for valve in parameters["valves"]})
+    elif name == "analyze_flushing":
+        # The service consumes m³/h (the older REST label incorrectly says L/s).
+        # Reject unsupported conversion rather than silently ignore the flow.
+        from app.native.wndb.model.options import get_option
+        if get_option(network).get("UNITS") not in {"LPS", "CMH"}:
+            raise ValueError("Flushing currently supports LPS or CMH models only")
+        result = scenarios.flushing_analysis(**common, username=username,
+            drainage_node_ID=parameters["drainage_node_id"], flushing_flow=parameters["flushing_flow_m3h"],
+            modify_valve_opening={valve: 0.0 for valve in parameters["closed_valves"]})
+    elif name == "analyze_pressure_regulation":
+        from app.services.simulation import query_corresponding_element_id_and_query_id
+        mappings = query_corresponding_element_id_and_query_id(network)
+        controls = parameters["pump_control"]
+        if set(controls) - (set(mappings.fixed_pumps) | set(mappings.variable_pumps)):
+            raise ValueError("Pump controls must reference registered model pumps")
+        if any(value not in (0, 1) for key, values in controls.items() if key in mappings.fixed_pumps for value in values):
+            raise ValueError("Fixed-speed pump controls must be 0 or 1")
+        result = scenarios.pressure_regulation(**common,
+            modify_fixed_pump_pattern={key: value for key, value in controls.items() if key in mappings.fixed_pumps},
+            modify_variable_pump_pattern={key: value for key, value in controls.items() if key in mappings.variable_pumps},
+            modify_tank_initial_level=parameters["tank_init_level"], scada_mappings=mappings)
+    else:
+        raise ValueError("Scenario is not registered")
+    if not isinstance(result, dict):
+        raise ValueError("Scenario did not return calculation results")
+    selected_nodes = [row for row in result.get("node_results", []) if str(row.get("node")) in nodes]
+    selected_links = [row for row in result.get("link_results", []) if str(row.get("link")) in links]
+    return {
+        "input_source": "online_model_simulation" if name == "run_simulation" else "isolated_model_simulation",
+        "persisted_baseline": name == "run_simulation", "persisted_scheme": False,
+        "start_time": start.isoformat(), "duration_seconds": duration,
+        "units": result.get("units", {}), "times": result.get("times", {}),
+        "node_results": selected_nodes, "link_results": selected_links,
+        "coverage": {"requested_nodes": nodes, "requested_links": links,
+                     "missing_nodes": sorted(set(nodes) - {str(row["node"]) for row in selected_nodes}),
+                     "missing_links": sorted(set(links) - {str(row["link"]) for row in selected_links})},
+    }

app/mcp/scope.py

@@ -0,0 +1,76 @@
+"""Resolve authoritative metadata and project routing for each MCP request."""
+
+from contextlib import asynccontextmanager
+from dataclasses import dataclass
+from typing import Any, AsyncIterator
+from psycopg import AsyncConnection
+from app.auth.project_dependencies import ProjectContext
+from app.infra.db.project_routing import ActiveProjectRouting
+from fastapi import HTTPException
+from mcp.server.auth.provider import AccessToken
+
+from app.auth.metadata_dependencies import get_current_metadata_user
+from app.auth.project_dependencies import (
+    resolve_project_context,
+    resolve_project_business_routing,
+    resolve_project_simulation_routing,
+    get_project_pg_connection,
+    get_project_timescale_connection,
+)
+from app.infra.db.metadb.database import get_metadata_session
+from app.infra.db.metadb.repositories.metadata_repository import MetadataRepository
+from app.infra.db.project_routing import activate_project_routing
+
+
+class QueryMetadataRepository(MetadataRepository):
+    async def get_project_db_routing(self, project_id, role):
+        routing = await super().get_project_db_routing(project_id, role)
+        # Existing connection generators obtain routing before awaiting a pool.
+        # End that implicit SELECT transaction here, before the network wait.
+        await self.session.commit()
+        return routing
+
+
+@dataclass
+class QueryScope:
+    project: ProjectContext
+    user: Any
+    connection: AsyncConnection
+    routing: ActiveProjectRouting
+    metadata: QueryMetadataRepository
+
+    @asynccontextmanager
+    async def simulation_routing(self) -> AsyncIterator[ActiveProjectRouting]:
+        """Resolve both databases before a native calculation can modify its model."""
+        routing = await resolve_project_simulation_routing(self.project, self.metadata)
+        await self.metadata.session.commit()
+        with activate_project_routing(routing):
+            yield routing
+
+    @asynccontextmanager
+    async def timescale(self):
+        # Metadata repository here is used only to resolve a known project DB.
+        async with asynccontextmanager(get_project_timescale_connection)(
+            self.project, self.metadata
+        ) as conn:
+            await self.metadata.session.commit()
+            yield conn
+
+
+@asynccontextmanager
+async def project_scope(principal: AccessToken, project_id: str):
+    if not principal.claims:
+        raise HTTPException(401, "MCP identity is missing")
+    async with asynccontextmanager(get_metadata_session)() as session:
+        metadata = QueryMetadataRepository(session)
+        user = await get_current_metadata_user(principal.claims, metadata)
+        project = await resolve_project_context(project_id, user, metadata)
+        routing = await resolve_project_business_routing(project, metadata)
+        # End the metadata read before any external query/IO waits.
+        await session.commit()
+        async with asynccontextmanager(get_project_pg_connection)(
+            project, metadata
+        ) as conn:
+            await session.commit()
+            with activate_project_routing(routing):
+                yield QueryScope(project, user, conn, routing, metadata)

app/mcp/work_order_skills.py

@@ -0,0 +1,40 @@
+"""Server-owned business instructions, selected from the persisted task workflow."""
+
+from hashlib import sha256
+from pathlib import Path
+from typing import Any
+
+import yaml
+
+
+SKILL_ROOT = Path(__file__).resolve().parents[2] / "resources" / "work-order-skills"
+WORKFLOW_SKILLS = {
+    "burst_report": "burst-report",
+    "water_quality_complaint": "water-quality-complaint",
+    "general": "appointment-follow-up",
+}
+
+
+def read_work_order_skill(workflow: str | None, *, include_content: bool = False) -> dict[str, Any]:
+    """Read only packaged, allowlisted instructions; never interpret a task path."""
+    name = WORKFLOW_SKILLS.get(workflow)
+    if name is None:
+        return {"available": False, "workflow": workflow, "reason": "No business skill is configured for this workflow"}
+    content = (SKILL_ROOT / name / "SKILL.md").read_text(encoding="utf-8")
+    _, frontmatter, body = content.split("---", 2)
+    header = yaml.safe_load(frontmatter)
+    if header["name"] != name:
+        raise ValueError("Packaged work-order skill name does not match its directory")
+    result = {
+        "available": True,
+        "kind": "business_instructions",
+        "workflow": workflow,
+        "name": header["name"],
+        "description": header["description"],
+        "version": str(header["metadata"]["version"]),
+        "content_sha256": sha256(content.encode("utf-8")).hexdigest(),
+        "read": {"section": "skill"},
+    }
+    if include_content:
+        result["instructions"] = body.strip()
+    return result

app/native/wndb/inp/exporter.py

@@ -120,7 +120,7 @@ def dump_inp(project: str, inp: str, version: str = '3'):
             file.write('\n'.join(inp_out_demand(project)))
 
         elif name == STATUS:
-            file.write('\n'.join(inp_out_status(project)))
+            file.write('\n'.join(inp_out_status(project, version=version)))
 
         elif name == PATTERNS:
             if version == '3':
@@ -232,7 +232,7 @@ def export_inp(project: str, version: str = '3') -> ChangeSet:
             inp += '\n'.join(inp_out_demand(project))
 
         elif name == STATUS:
-            inp += '\n'.join(inp_out_status(project))
+            inp += '\n'.join(inp_out_status(project, version=version))
 
         elif name == PATTERNS:
             if version == '3':

app/native/wndb/model/status.py

@@ -93,14 +93,16 @@ def inp_in_status(line: str) -> str:
         return str(f"insert into network.link_initial_settings (link_id, status, setting) values ({sql_literal(link)}, null, {sql_literal(float(value))});")
 
 
-def inp_out_status(name: str) -> list[str]:
+def inp_out_status(name: str, version: str = '2') -> list[str]:
     lines = []
     objs = read_all(name, 'select link_id as link, status, setting from network.link_initial_settings order by link_id')
     for obj in objs:
         link = obj['link']
         status = obj['status'] if obj['status'] != None else ''
         setting = obj['setting'] if obj['setting'] != None else ''
-        if status != '':
+        # EPANET 2 valves are active at their configured setting by default.
+        # ACTIVE is an internal state, not a legal [STATUS] keyword in v2.
+        if status != '' and not (version == '2' and status.upper() == LINK_STATUS_ACTIVE):
             lines.append(f'{link} {status}')
         if setting != '':
             lines.append(f'{link} {setting}')

app/services/agent_messages.py

@@ -0,0 +1,186 @@
+"""Messages never change task generation or silently target a replacement run."""
+from datetime import datetime, timezone
+from uuid import UUID, uuid4
+
+from app.auth.permissions import OPERATIONAL_TASK_EXECUTE, permissions_for_context
+from app.domain.domain_errors import DomainError, require
+from app.domain.schemas.agent_messages import (
+    AgentMessage, AgentMessageCreate, AgentMessageDelivery, AgentMessagePage, AgentMessageReply,
+)
+
+
+def _message_cursor(cursor: str | None):
+    if not cursor:
+        return None, None
+    try:
+        stamp, identifier = cursor.rsplit("|", 1)
+        cursor_at, cursor_id = datetime.fromisoformat(stamp), UUID(identifier)
+        require(cursor_at.tzinfo is not None, "INVALID_CURSOR", "Cursor needs a timezone", status=400)
+        return cursor_at, cursor_id
+    except (ValueError, TypeError) as exc:
+        raise DomainError("INVALID_CURSOR", "Invalid message cursor", status=400) from exc
+
+
+class AgentMessageService:
+    def __init__(self, repository, context, *, clock=None):
+        self.repo = repository
+        self.context = context
+        self.clock = clock or (lambda: datetime.now(timezone.utc))
+
+    def _require_service(self) -> None:
+        require(self.context.system_role == "service" and
+                OPERATIONAL_TASK_EXECUTE in permissions_for_context(self.context), "SERVICE_REQUIRED",
+                "Message dispatch requires a project service account", status=403)
+
+    async def get(self, message_id: UUID) -> AgentMessage:
+        if self.context.system_role == "service":
+            self._require_service()
+        row = await self.repo.get_message(str(message_id))
+        require(row is not None, "MESSAGE_NOT_FOUND", "Message not found", status=404)
+        require(self.context.system_role == "service" or
+                str(row["conversation_owner_id"]) == str(self.context.user_id),
+                "MESSAGE_NOT_FOUND", "Message not found", status=404)
+        return AgentMessage.model_validate(row)
+
+    async def _active_run(self, task_id: UUID, run_id: UUID) -> dict:
+        run = await self.repo.get_run(str(run_id))
+        require(run is not None and str(run["task_id"]) == str(task_id),
+                "RUN_NOT_FOUND", "Run does not belong to the task", status=404)
+        task = await self.repo.get_task(str(task_id))
+        require(task is not None and task["task_type"] == "event_investigation",
+                "INVALID_TASK_TYPE", "Messages require an investigation task", status=409)
+        require(run["status"] in {"queued", "claimed", "running"},
+                "RUN_TERMINAL", "Run no longer accepts messages", status=409)
+        if run["status"] != "queued":
+            expires = run.get("lease_expires_at")
+            if isinstance(expires, str):
+                expires = datetime.fromisoformat(expires.replace("Z", "+00:00"))
+            require(expires is not None and expires > self.clock(),
+                    "LEASE_EXPIRED", "Run lease expired", status=409)
+        return run
+
+    async def _insert(self, *, command, owner_id, conversation_owner_id, direction, task_id, run_id,
+                      target_session_id, in_reply_to=None) -> AgentMessage:
+        prior = await self.repo.get_message_request(
+            str(owner_id), command.source_session_id, str(command.request_id),
+        )
+        expected = {
+            "task_id": str(task_id), "run_id": str(run_id),
+            "direction": direction, "in_reply_to": str(in_reply_to) if in_reply_to else None,
+            "text": command.text,
+        }
+        if prior is not None:
+            require(all((str(prior[key]) if prior[key] is not None else None) == value
+                        for key, value in expected.items()),
+                    "IDEMPOTENCY_CONFLICT", "request_id was used with different content", status=409)
+            return AgentMessage.model_validate(prior)
+        # The run lock serializes message creation with completion/cancellation.
+        await self._active_run(task_id, run_id)
+        message_id = uuid4()
+        now = self.clock()
+        message = AgentMessage(
+            project_id=self.context.project_id,
+            message_id=message_id, request_id=command.request_id,
+            task_id=task_id, run_id=run_id, source_session_id=command.source_session_id,
+            source_owner_id=owner_id, conversation_owner_id=conversation_owner_id,
+            target_session_id=target_session_id,
+            direction=direction, in_reply_to=in_reply_to, text=command.text,
+            status="pending", native_message_id=f"msg_{message_id}",
+            created_at=now, updated_at=now,
+        )
+        await self.repo.insert_message(message.model_dump(mode="json"))
+        return message
+
+    async def create(self, command: AgentMessageCreate) -> AgentMessage:
+        async with self.repo.transaction():
+            await self.repo.lock(f"run:{command.run_id}")
+            run = await self.repo.get_run(str(command.run_id))
+            return await self._insert(
+                command=command, owner_id=self.context.user_id,
+                conversation_owner_id=self.context.user_id, direction="to_task",
+                task_id=command.task_id, run_id=command.run_id,
+                target_session_id=run.get("agent_session_id") if run else None,
+            )
+
+    async def reply(self, message_id: UUID, command: AgentMessageReply) -> AgentMessage:
+        async with self.repo.transaction():
+            original = await self.get(message_id)
+            await self.repo.lock(f"run:{original.run_id}")
+            run = await self.repo.get_run(str(original.run_id))
+            if original.direction == "to_task":
+                self._require_service()
+                require(run is not None and run.get("agent_session_id") == command.source_session_id,
+                        "SESSION_MISMATCH", "Session does not own the message run", status=409)
+                direction, target = "to_chat", original.source_session_id
+            else:
+                require(original.conversation_owner_id == self.context.user_id,
+                        "MESSAGE_NOT_FOUND", "Message not found", status=404)
+                require(command.source_session_id == original.target_session_id,
+                        "SESSION_MISMATCH", "Only the original chat can reply", status=409)
+                direction = "to_task"
+                target = run.get("agent_session_id") if run else None
+            return await self._insert(
+                command=command, owner_id=self.context.user_id,
+                conversation_owner_id=original.conversation_owner_id, direction=direction,
+                task_id=original.task_id, run_id=original.run_id,
+                target_session_id=target, in_reply_to=message_id,
+            )
+
+    async def list_session(
+        self, session_id: str, limit: int, cursor: str | None, pending_only: bool = False,
+    ) -> AgentMessagePage:
+        cursor_at, cursor_id = _message_cursor(cursor)
+        kwargs = {"cursor_at": cursor_at, "cursor_id": cursor_id, "pending_only": pending_only}
+        if self.context.system_role == "service":
+            self._require_service()
+            rows = await self.repo.list_run_session_messages(session_id, limit + 1, **kwargs)
+            if not rows:
+                rows = await self.repo.list_session_messages(str(self.context.user_id), session_id, limit + 1, **kwargs)
+        else:
+            rows = await self.repo.list_session_messages(str(self.context.user_id), session_id, limit + 1, **kwargs)
+        items = [AgentMessage.model_validate(row) for row in rows[:limit]]
+        next_cursor = f"{items[-1].created_at.isoformat()}|{items[-1].message_id}" if len(rows) > limit else None
+        return AgentMessagePage(items=items, next_cursor=next_cursor)
+
+    async def pending(self, limit: int, cursor: str | None = None) -> AgentMessagePage:
+        self._require_service()
+        cursor_at, cursor_id = _message_cursor(cursor)
+        rows = await self.repo.list_pending_messages(limit + 1, cursor_at=cursor_at, cursor_id=cursor_id)
+        items = [AgentMessage.model_validate(row) for row in rows[:limit]]
+        next_cursor = f"{items[-1].created_at.isoformat()}|{items[-1].message_id}" if len(rows) > limit else None
+        return AgentMessagePage(items=items, next_cursor=next_cursor)
+
+    async def delivery(self, message_id: UUID, command: AgentMessageDelivery) -> AgentMessage:
+        async with self.repo.transaction():
+            message = await self.get(message_id)
+            if self.context.system_role == "service":
+                self._require_service()
+            if self.context.system_role != "service":
+                require(message.direction == "to_chat",
+                        "SERVICE_REQUIRED", "Only the receiver can acknowledge a chat reply", status=403)
+            await self.repo.lock(f"run:{message.run_id}")
+            await self.repo.lock(f"message:{message_id}")
+            message = await self.get(message_id)
+            if message.status in {"delivered", "failed"}:
+                require(message.status == command.status, "MESSAGE_TERMINAL",
+                        "Message already has a terminal delivery result", status=409)
+                return message
+            if message.direction == "to_task" and command.status != "failed":
+                run = await self._active_run(message.task_id, message.run_id)
+                target = run.get("agent_session_id")
+                require(target is not None and command.target_session_id == target,
+                        "SESSION_MISMATCH", "Delivery must target the owning run session", status=409)
+            else:
+                target = message.target_session_id
+                require(command.target_session_id in {None, target}, "SESSION_MISMATCH",
+                        "Delivery cannot change the reply channel", status=409)
+            require(command.native_message_id in {None, message.native_message_id},
+                    "MESSAGE_ID_MISMATCH", "Native message identity is immutable", status=409)
+            now = self.clock()
+            update = {
+                "status": command.status, "target_session_id": target,
+                "failure_detail": command.failure_detail, "updated_at": now,
+                "delivered_at": now if command.status == "delivered" else None,
+            }
+            await self.repo.update_message_delivery(str(message_id), update)
+            return message.model_copy(update=update)

app/services/burst_location.py

@@ -63,6 +63,7 @@ def run_burst_location_by_network(
     use_scada_flow: bool = False,
     scheme_name: str | None = None,
     simulation_run_id: UUID | None = None,
+    final_candidates_csv_path: str | None = "temp/burst_location/final_round_candidates.csv",
 ) -> dict[str, Any]:
     if not network:
         raise ValueError("network is required.")
@@ -312,6 +313,7 @@ def run_burst_location_by_network(
             min_dpressure=min_dpressure,
             basic_pressure=basic_pressure,
             visualize_partition=False,
+            final_candidates_csv_path=final_candidates_csv_path,
         )
 
     payload: dict[str, Any] = {

app/services/content_digest.py

@@ -0,0 +1,18 @@
+"""RFC 8785 canonical JSON hashing shared by persisted records."""
+
+from hashlib import sha256
+
+import rfc8785
+
+from app.domain.domain_errors import DomainError
+
+
+def canonical(value: object) -> bytes:
+    try:
+        return rfc8785.dumps(value)
+    except rfc8785.CanonicalizationError as exc:
+        raise DomainError("SCHEMA_INVALID", "Content is not canonicalizable JSON") from exc
+
+
+def digest(value: object) -> str:
+    return "sha256:" + sha256(canonical(value)).hexdigest()

app/services/geocoding.py

@@ -5,6 +5,9 @@ import httpx
 from pydantic import AliasChoices, BaseModel, Field
 
 from app.core.config import settings
+from app.core.tianditu_logging import install_tianditu_token_filter
+
+install_tianditu_token_filter()
 
 
 class TiandituGeocodeRequest(BaseModel):
@@ -32,12 +35,12 @@ async def geocode_tianditu(
     *,
     client: httpx.AsyncClient | None = None,
 ) -> dict[str, Any]:
-    if not settings.TIANDITU_GEOCODER_TOKEN:
-        raise TiandituGeocodingConfigError("TIANDITU_GEOCODER_TOKEN is not configured")
+    if not settings.TIANDITU_TOKEN:
+        raise TiandituGeocodingConfigError("TIANDITU_TOKEN is not configured")
 
     params = {
         "ds": json.dumps({"keyWord": request.keyword}, ensure_ascii=False),
-        "tk": settings.TIANDITU_GEOCODER_TOKEN,
+        "tk": settings.TIANDITU_TOKEN,
     }
 
     if client is not None:

app/services/operational_presentation.py

@@ -0,0 +1,99 @@
+"""Resolve browser overlays only from immutable Server-ingested snapshots."""
+from __future__ import annotations
+
+from math import isfinite
+
+from app.domain.domain_errors import require
+from app.domain.schemas.operational_data import TaskDataSnapshot
+
+
+def _number(value) -> bool:
+    return isinstance(value, (int, float)) and not isinstance(value, bool) and isfinite(value)
+
+
+def _position(value):
+    require(isinstance(value, list) and len(value) in {2, 3}
+            and all(_number(item) for item in value)
+            and -180 <= value[0] <= 180 and -90 <= value[1] <= 90,
+            "OVERLAY_GEOMETRY_INVALID", "Overlay coordinates must be finite WGS84 positions", status=422)
+    return value[:2]
+
+
+def _geometry(value):
+    require(isinstance(value, dict), "OVERLAY_GEOMETRY_INVALID", "Overlay geometry is missing", status=422)
+    kind = value.get("type")
+    depths = {"Point": 0, "MultiPoint": 1, "LineString": 1,
+              "MultiLineString": 2, "Polygon": 2, "MultiPolygon": 3}
+    require(kind in depths, "OVERLAY_GEOMETRY_INVALID", "Unsupported overlay geometry", status=422)
+    positions = 0
+
+    def coordinates(items, depth):
+        nonlocal positions
+        if depth == 0:
+            positions += 1
+            require(positions <= 20000, "OVERLAY_TOO_LARGE", "Overlay geometry is too large", status=422)
+            return _position(items)
+        require(isinstance(items, list) and 0 < len(items) <= 20000,
+                "OVERLAY_GEOMETRY_INVALID", "Overlay coordinates are empty or too large", status=422)
+        return [coordinates(item, depth - 1) for item in items]
+
+    normalized = coordinates(value.get("coordinates"), depths[kind])
+    lines = ([normalized] if kind == "LineString" else normalized if kind == "MultiLineString" else [])
+    require(all(len(line) >= 2 for line in lines), "OVERLAY_GEOMETRY_INVALID",
+            "Overlay lines need at least two positions", status=422)
+    rings = (normalized if kind == "Polygon" else [ring for polygon in normalized for ring in polygon]
+             if kind == "MultiPolygon" else [])
+    require(all(len(ring) >= 4 and ring[0] == ring[-1] for ring in rings),
+            "OVERLAY_GEOMETRY_INVALID", "Overlay polygon rings must be closed", status=422)
+    return {"type": kind, "coordinates": normalized}
+
+
+def resolve_overlay(snapshot: TaskDataSnapshot, mode: str) -> dict:
+    """Use recorded GeoJSON or recorded coordinates; never geocode or infer a boundary."""
+    data = snapshot.data
+    candidates = [data]
+    if isinstance(data, dict):
+        candidates.extend([data.get("map_payload"), data.get("drawing_payload")])
+        latest = data.get("latest_analysis")
+        if isinstance(latest, dict):
+            candidates.append(latest.get("map_payload"))
+    collection = next((item for item in candidates
+                       if isinstance(item, dict) and item.get("type") == "FeatureCollection"), None)
+    if collection is not None:
+        raw_features = collection.get("features")
+        require(isinstance(raw_features, list), "OVERLAY_GEOMETRY_INVALID",
+                "Snapshot FeatureCollection is invalid", status=422)
+    else:
+        rows = data.get("items") if isinstance(data, dict) else data
+        require(isinstance(rows, list), "OVERLAY_DATA_MISSING",
+                "Evidence snapshot has no recorded coordinates or GeoJSON", status=422)
+        raw_features = []
+        for row in rows:
+            if not isinstance(row, dict) or row.get("longitude") is None or row.get("latitude") is None:
+                continue
+            raw_features.append({
+                "type": "Feature", "geometry": {"type": "Point", "coordinates": [row["longitude"], row["latitude"]]},
+                "id": row.get("device_id") or row.get("point_id") or row.get("work_order_id") or row.get("id"),
+                "properties": {"weight": row.get("weight", 1)},
+            })
+    require(0 < len(raw_features) <= 1000, "OVERLAY_DATA_MISSING",
+            "Overlay needs between one and 1000 recorded features", status=422)
+    features = []
+    for feature in raw_features:
+        require(isinstance(feature, dict) and feature.get("type") == "Feature",
+                "OVERLAY_GEOMETRY_INVALID", "Snapshot contains an invalid feature", status=422)
+        geometry = _geometry(feature.get("geometry"))
+        require(mode != "heatmap" or geometry["type"] == "Point", "OVERLAY_HEATMAP_POINTS_REQUIRED",
+                "Heatmaps require recorded points, not inferred polygon centroids", status=422)
+        properties = feature.get("properties") or {}
+        require(isinstance(properties, dict), "OVERLAY_GEOMETRY_INVALID", "Feature properties are invalid", status=422)
+        normalized = {"type": "Feature", "geometry": geometry, "properties": dict(properties)}
+        if isinstance(feature.get("id"), (str, int)) and not isinstance(feature.get("id"), bool):
+            normalized["id"] = feature["id"]
+        if mode == "heatmap":
+            weight = properties.get("weight", 1)
+            require(_number(weight) and weight >= 0, "OVERLAY_WEIGHT_INVALID",
+                    "Heatmap weight must be a finite nonnegative number", status=422)
+            normalized["properties"]["weight"] = weight
+        features.append(normalized)
+    return {"type": "FeatureCollection", "features": features}

app/services/place_search.py

@@ -0,0 +1,115 @@
+import json
+from typing import Any
+
+import httpx
+from pydantic import ValidationError
+
+from app.core.config import settings
+from app.core.tianditu_logging import install_tianditu_token_filter
+from app.domain.schemas.place_search import (
+    MAX_PLACE_SEARCH_RESULTS,
+    PlaceSearchItem,
+    PlaceSearchRequest,
+    PlaceSearchResponse,
+)
+
+FENGXIAN_ADMIN_CODE = "156310120"
+install_tianditu_token_filter()
+
+
+class PlaceSearchError(RuntimeError):
+    def __init__(self, status_code: int, detail: str):
+        super().__init__(detail)
+        self.status_code = status_code
+        self.detail = detail
+
+
+async def search_tianditu_places(
+    request: PlaceSearchRequest,
+    *,
+    client: httpx.AsyncClient | None = None,
+) -> PlaceSearchResponse:
+    token = settings.TIANDITU_TOKEN
+    if not token:
+        raise PlaceSearchError(503, "地名搜索服务未配置")
+    params = {
+        "postStr": json.dumps(
+            {
+                "keyWord": request.keyword,
+                "queryType": 12,
+                "specify": FENGXIAN_ADMIN_CODE,
+                "start": 0,
+                "count": MAX_PLACE_SEARCH_RESULTS,
+                "show": 2,
+            },
+            ensure_ascii=False,
+        ),
+        "type": "query",
+        "tk": token,
+    }
+    try:
+        if client is not None:
+            response = await client.get(settings.TIANDITU_SEARCH_URL, params=params)
+        else:
+            async with httpx.AsyncClient(
+                timeout=settings.TIANDITU_SEARCH_TIMEOUT_SECONDS
+            ) as managed_client:
+                response = await managed_client.get(
+                    settings.TIANDITU_SEARCH_URL, params=params
+                )
+        response.raise_for_status()
+    except httpx.TimeoutException:
+        raise PlaceSearchError(504, "地名搜索超时，请稍后重试") from None
+    except httpx.HTTPError:
+        # Request URLs contain the provider key and must not reach the client.
+        raise PlaceSearchError(502, "天地图地名搜索服务暂不可用") from None
+    try:
+        return _parse_response(response.json())
+    except (ValueError, TypeError, KeyError, AttributeError, ValidationError):
+        raise PlaceSearchError(502, "天地图返回了无效的地名搜索结果") from None
+
+
+def _parse_response(payload: dict[str, Any]) -> PlaceSearchResponse:
+    code = str(payload.get("status", {}).get("infocode", ""))
+    if code == "3001":
+        return PlaceSearchResponse(items=[], total=0)
+    if code != "1000":
+        raise PlaceSearchError(502, "天地图地名搜索请求失败，请稍后重试")
+
+    result_type = str(payload.get("resultType", ""))
+    if result_type not in {"1", "3"}:
+        raise PlaceSearchError(502, "未获得可定位的地名结果，请细化关键字")
+    raw_items = payload.get("pois", []) if result_type == "1" else payload.get("area", [])
+    if not isinstance(raw_items, list) or len(raw_items) > MAX_PLACE_SEARCH_RESULTS:
+        raise ValueError("Invalid search results")
+    items = []
+    for raw in raw_items:
+        admin_code = str(raw.get("provinceCode") or raw.get("adminCode") or "")
+        if admin_code and not admin_code.startswith(("156310", "310")):
+            raise ValueError("Result outside Shanghai")
+        county_code = str(raw.get("countyCode") or (raw.get("adminCode") if result_type == "3" else "") or "")
+        county = raw.get("county")
+        if county_code and not county_code.startswith((FENGXIAN_ADMIN_CODE, "310120")):
+            raise ValueError("Result outside Fengxian")
+        if county and county != "奉贤区" or not county_code and county != "奉贤区":
+            raise ValueError("Missing or mismatched district")
+        lon, lat = str(raw["lonlat"]).split(",")
+        name = raw["name"]
+        if not isinstance(name, str) or not name.strip():
+            raise ValueError("Missing place name")
+        name = name.strip()
+        category = raw.get("typeName") or ("公交站" if str(raw.get("poiType")) == "102" else None)
+        items.append(PlaceSearchItem(
+            id=str(raw.get("hotPointID") or raw.get("adminCode") or f"{name}:{lon},{lat}"),
+            name=name,
+            address=str(raw.get("address") or name),
+            longitude=float(lon), latitude=float(lat),
+            category="行政区" if result_type == "3" else category,
+            district="奉贤区",
+        ))
+    total = int(payload.get("count", 0))
+    if total < len(items):
+        raise ValueError("Invalid search total")
+    return PlaceSearchResponse(
+        items=items, total=len(items), truncated=total > MAX_PLACE_SEARCH_RESULTS,
+    )

app/services/scheme_management.py

@@ -1,9 +1,10 @@
-from datetime import datetime
+from datetime import datetime, timezone
 from typing import Any
 from uuid import UUID, uuid4
 
 from app.infra.db.postgresql.analysis import AnalysisRepository
 from app.native.wndb.core.connection import project_connection, project_transaction
+from app.infra.db.project_routing import get_active_project_routing
 from app.domain.time import parse_utc_time
 
 
@@ -40,16 +41,21 @@ def create_analysis_run(
     started_at = parse_utc_time(scheme_start_time, field_name="scheme_start_time")
     run_id = uuid4()
     with project_connection(name) as conn:
-        AnalysisRepository.create_run_sync(
-            conn,
-            run_id=run_id,
-            name=scheme_name,
-            run_type=scheme_type,
-            created_by=username,
-            started_at=started_at,
-            status=status,
-            parameters=scheme_detail,
-        )
+        with conn.transaction():
+            AnalysisRepository.create_run_sync(
+                conn,
+                run_id=run_id,
+                name=scheme_name,
+                run_type=scheme_type,
+                created_by=username,
+                started_at=started_at,
+                status=status,
+                parameters=scheme_detail,
+            )
+            _append_run_event(
+                conn, run_id, revision=1, status=status,
+                occurred_at=started_at, results_committed=False,
+            )
     return run_id
 
 
@@ -84,6 +90,10 @@ def store_analysis_run_with_result(
             result_type=result_type,
             payload=result_payload,
         )
+        _append_run_event(
+            conn, run_id, revision=1, status="completed",
+            occurred_at=started_at, results_committed=True,
+        )
     return run_id
 
 
@@ -96,13 +106,37 @@ def update_analysis_run(
     scheme_detail: dict[str, Any],
 ) -> None:
     with project_connection(name) as conn:
-        AnalysisRepository.update_run_sync(
-            conn,
-            run_id,
-            status=status,
-            created_by=username,
-            parameters=scheme_detail,
-        )
+        with conn.transaction():
+            revision = AnalysisRepository.update_run_sync(
+                conn,
+                run_id,
+                status=status,
+                created_by=username,
+                parameters=scheme_detail,
+            )
+            _append_run_event(
+                conn, run_id, revision=revision, status=status,
+                occurred_at=datetime.now(timezone.utc),
+                results_committed=status == "completed",
+            )
+
+
+def _append_run_event(
+    conn, run_id: UUID, *, revision: int, status: str,
+    occurred_at: datetime, results_committed: bool,
+) -> None:
+    routing = get_active_project_routing()
+    if routing is None or routing.project_id is None:
+        return
+    AnalysisRepository.append_run_changed_event_sync(
+        conn,
+        project_id=routing.project_id,
+        run_id=run_id,
+        revision=revision,
+        status=status,
+        occurred_at=occurred_at,
+        results_committed=results_committed,
+    )
 
 
 def get_analysis_run(name: str, run_id: UUID) -> dict[str, Any]:

app/services/sensor_placement.py

@@ -80,14 +80,26 @@ def _create_validated_placement(
     min_diameter: int,
     created_by: str,
     sensor_locations: list[str],
+    sensor_type: str = "pressure",
+    method: str | None = None,
+    persist_results: bool = True,
 ) -> dict[str, Any]:
     validate_sensor_placement_nodes(project_code, sensor_locations)
+    if not persist_results:
+        return {
+            "sensor_locations": sensor_locations,
+            "sensor_points": _sensor_points(project_code, sensor_locations),
+            "method": method,
+            "sensor_type": sensor_type,
+        }
     return sensor_placement_repository.create_sensor_placement(
         project_code,
         run_name=run_name,
         min_diameter=min_diameter,
         created_by=created_by,
         sensor_locations=sensor_locations,
+        sensor_type=sensor_type,
+        method=method,
     )
 
 
@@ -97,6 +109,8 @@ def optimize_sensor_placement_by_sensitivity(
     sensor_count: int,
     min_diameter: int,
     created_by: str,
+    sensor_type: str = "pressure",
+    persist_results: bool = True,
 ) -> dict[str, Any]:
     """Run sensitivity placement and persist the validated result."""
 
@@ -112,6 +126,9 @@ def optimize_sensor_placement_by_sensitivity(
         min_diameter=min_diameter,
         created_by=created_by,
         sensor_locations=sensor_locations,
+        sensor_type=sensor_type,
+        method="sensitivity",
+        persist_results=persist_results,
     )
 
 
@@ -121,6 +138,8 @@ def optimize_sensor_placement_by_kmeans(
     sensor_count: int,
     min_diameter: int,
     created_by: str,
+    sensor_type: str = "pressure",
+    persist_results: bool = True,
 ) -> dict[str, Any]:
     """Export the model, run K-means placement, and persist the result."""
 
@@ -136,6 +155,9 @@ def optimize_sensor_placement_by_kmeans(
         min_diameter=min_diameter,
         created_by=created_by,
         sensor_locations=sensor_locations,
+        sensor_type=sensor_type,
+        method="kmeans",
+        persist_results=persist_results,
     )
 
 

app/services/simulation.py

@@ -11,7 +11,7 @@ from app.native.wndb.model.tanks import get_tank, set_tank
 from app.native.wndb.model.times import get_time, set_time
 
 # from get_real_status import *
-from datetime import datetime
+from datetime import datetime, timedelta
 from math import modf
 import json
 import pytz
@@ -150,6 +150,69 @@ def _apply_valve_control(
         set_status(project_name, cs)
 
 
+def run_online_simulation(
+    name: str,
+    start_time: datetime,
+    duration_seconds: int,
+    *,
+    output_node_ids: list[str] | None = None,
+    output_link_ids: list[str] | None = None,
+) -> dict:
+    """Recalculate online baseline snapshots without cloning a model database."""
+    if start_time.tzinfo is None or duration_seconds <= 0:
+        raise ValueError("Online simulation requires an aware start time and positive duration")
+    step = parse_clock_duration_seconds(
+        get_time(name)["HYDRAULIC TIMESTEP"], field_name="HYDRAULIC TIMESTEP"
+    )
+    if step <= 0:
+        raise ValueError("HYDRAULIC TIMESTEP must be greater than 0.")
+    mappings = query_corresponding_element_id_and_query_id(name)
+    selected = {
+        "node": {node: [] for node in output_node_ids or []},
+        "link": {link: [] for link in output_link_ids or []},
+    }
+    sample_times = []
+    units = {}
+    current = start_time
+    end = start_time + timedelta(seconds=duration_seconds)
+    while current < end:
+        output = run_simulation(
+            name=name,
+            simulation_type="realtime",
+            modify_pattern_start_time=current.isoformat(timespec="seconds"),
+            scada_mappings=mappings,
+        )
+        if not isinstance(output, dict):
+            raise RuntimeError("Online simulation did not return calculation results")
+        units = output.get("units", {})
+        sample_times.append(current.isoformat(timespec="seconds"))
+        for kind, elements in selected.items():
+            if not elements:
+                continue
+            rows = {
+                str(row[kind]): row.get("result", [])
+                for row in output.get(f"{kind}_results", []) if str(row[kind]) in elements
+            }
+            for element, values in elements.items():
+                samples = rows.get(element, [])
+                # Each realtime run is a single snapshot. Missing samples remain
+                # explicit so later values cannot shift to an earlier timestamp.
+                values.append(samples[0] if samples else None)
+        current += timedelta(seconds=step)
+    return {
+        "units": units,
+        "times": {
+            "report_start": 0, "report_step": step,
+            "sim_duration": duration_seconds, "num_periods": len(sample_times),
+            "sample_times": sample_times,
+        },
+        **{f"{kind}_results": [
+            {kind: element, "result": values}
+            for element, values in elements.items() if any(value is not None for value in values)
+        ] for kind, elements in selected.items()},
+    }
+
+
 # 2025/01/11
 def run_simulation(
     name: str,
@@ -170,7 +233,8 @@ def run_simulation(
     scheme_username: str = "system",
     scheme_detail: dict | None = None,
     scada_mappings: ScadaElementMappings | None = None,
-) -> UUID | None:
+    persist_results: bool = True,
+) -> UUID | dict | None:
     """
     传入需要修改的参数，改变数据库中对应位置的值，然后计算，返回结果
     :param name: 模型名称，数据库中对应的名字
@@ -189,6 +253,10 @@ def run_simulation(
     :param scheme_name：模拟方案名称
     :return:
     """
+    if not persist_results and (
+        not name.startswith("tjw_tmp_") or simulation_type.upper() != "EXTENDED"
+    ):
+        raise ValueError("Non-persistent simulation requires an isolated temporary model")
     # 记录开始时间
     time_cost_start = time.perf_counter()
 
@@ -514,6 +582,8 @@ def run_simulation(
     num_periods_result = times_info.get("num_periods")
     if num_periods_result is None:
         raise RuntimeError("run_project output missing times.num_periods")
+    if not persist_results:
+        return output_data
     print("simulation_type", simulation_type)
     print("before store result")
     # print(num_periods_result)
@@ -583,4 +653,6 @@ def run_simulation(
     # TimescaleInternalQueries.fill_scheme_simulation_result_to_SCADA(scheme_type=scheme_type, scheme_name=scheme_name)
 
     print("after store result")
+    if simulation_type.upper() == "REALTIME":
+        return output_data
     return run_id if simulation_type.upper() == "EXTENDED" else None

app/services/simulation_scenarios.py

@@ -222,7 +222,8 @@ def valve_close_analysis(
     modify_valve_opening: dict[str, float] = None,
     scheme_name: str = None,
     _temporary_project: str | None = None,
-) -> None:
+    persist_results: bool = True,
+) -> dict | str | None:
     """
     关阀模拟
     :param name: 模型名称，数据库中对应的名字
@@ -271,7 +272,7 @@ def valve_close_analysis(
     set_option(new_name, cs_options)
     # result = run_simulation_ex(new_name,'realtime', modify_pattern_start_time, modify_pattern_start_time, modify_total_duration,
     #                            downloading_prohibition=True)
-    simulation.run_simulation(
+    result = simulation.run_simulation(
         name=new_name,
         simulation_type="extended",
         modify_pattern_start_time=modify_pattern_start_time,
@@ -280,7 +281,9 @@ def valve_close_analysis(
         scheme_type="valve_close_analysis",
         scheme_name=scheme_name,
         result_db_name=name,
+        persist_results=persist_results,
     )
+    return result if not persist_results else None
     # step 3. restore the base model
     # for valve in valves:
     #     execute_undo(name)
@@ -303,7 +306,8 @@ def flushing_analysis(
     username: str | None = None,
     valve_control: dict[str, dict] = None,
     _temporary_project: str | None = None,
-) -> None:
+    persist_results: bool = True,
+) -> dict | str | None:
     """
     管道冲洗模拟
     :param name: 模型名称，数据库中对应的名字
@@ -425,7 +429,7 @@ def flushing_analysis(
     set_option(new_name, cs_options)
     # result = run_simulation_ex(new_name,'realtime', modify_pattern_start_time, modify_pattern_start_time, modify_total_duration,
     #                            downloading_prohibition=True)
-    simulation.run_simulation(
+    result = simulation.run_simulation(
         name=new_name,
         simulation_type="extended",
         modify_pattern_start_time=modify_pattern_start_time,
@@ -437,7 +441,9 @@ def flushing_analysis(
         result_db_name=name,
         scheme_username=username,
         scheme_detail=scheme_detail,
+        persist_results=persist_results,
     )
+    return result if not persist_results else None
     # step 4. restore the base model
     # return result
 
@@ -457,7 +463,8 @@ def contaminant_simulation(
     source_pattern: str = None,  # 污染源时间变化模式名称
     username: str | None = None,
     _temporary_project: str | None = None,
-) -> None:
+    persist_results: bool = True,
+) -> dict | str | None:
     """
     污染模拟
     :param name: 模型名称，数据库中对应的名字
@@ -568,7 +575,7 @@ def contaminant_simulation(
     # step 5. run simulation
     # result = run_simulation_ex(new_name,'realtime', modify_pattern_start_time, modify_pattern_start_time,  modify_total_duration,
     #                            downloading_prohibition=True)
-    simulation.run_simulation(
+    result = simulation.run_simulation(
         name=new_name,
         simulation_type="extended",
         modify_pattern_start_time=modify_pattern_start_time,
@@ -578,7 +585,9 @@ def contaminant_simulation(
         result_db_name=name,
         scheme_username=username,
         scheme_detail=scheme_detail,
+        persist_results=persist_results,
     )
+    return result if not persist_results else None
 
     # for i in range(1,operation_step):
     #     execute_undo(name)
@@ -600,7 +609,8 @@ def pressure_regulation(
     scheme_name: str = None,
     scada_mappings: simulation.ScadaElementMappings | None = None,
     _temporary_project: str | None = None,
-) -> None:
+    persist_results: bool = True,
+) -> dict | str | None:
     """
     区域调压模拟，用来模拟未来15分钟内，开关水泵对区域压力的影响
     :param name: 模型名称，数据库中对应的名字
@@ -645,7 +655,7 @@ def pressure_regulation(
     #                            pump_control=pump_control,
     #                            tank_initial_level_control=tank_initial_level_control,
     #                            downloading_prohibition=True)
-    simulation.run_simulation(
+    result = simulation.run_simulation(
         name=new_name,
         simulation_type="extended",
         modify_pattern_start_time=modify_pattern_start_time,
@@ -657,5 +667,7 @@ def pressure_regulation(
         scheme_name=scheme_name,
         result_db_name=name,
         scada_mappings=scada_mappings,
+        persist_results=persist_results,
     )
+    return result if not persist_results else None
     # return result

app/services/task_alerts.py

@@ -0,0 +1,127 @@
+"""Monotonic, idempotent state machine for task-produced alerts."""
+from __future__ import annotations
+
+from datetime import datetime, timezone
+
+from app.domain.schemas.task_alert import TaskAlertEvent, TaskAlertState
+from app.domain.domain_errors import require
+from app.services.content_digest import digest
+
+
+class TaskAlertService:
+    def __init__(
+        self,
+        repository,
+        context,
+        *,
+        clock=None,
+        internal_producer_id: str | None = None,
+    ):
+        self.repo = repository
+        self.context = context
+        self.clock = clock or (lambda: datetime.now(timezone.utc))
+        self.internal_producer_id = internal_producer_id
+
+    async def push(self, command: TaskAlertEvent) -> TaskAlertState:
+        if self.internal_producer_id is None:
+            require(
+                self.context is not None and self.context.project_role == "member",
+                "PERMISSION_DENIED",
+                "Alert events require project write permission",
+                status=403,
+            )
+            require(
+                command.producer_id == f"principal:{self.context.user_id}",
+                "PERMISSION_DENIED",
+                "Alert producer must match the authenticated project principal",
+                status=403,
+            )
+        else:
+            require(
+                command.producer_id == self.internal_producer_id,
+                "PERMISSION_DENIED",
+                "Internal alert producer identity does not match",
+                status=403,
+            )
+        event = command.model_dump(mode="json")
+        async with self.repo.transaction():
+            await self.repo.lock_alert(command.alert_id)
+            prior_event = await self.repo.get_event(
+                command.producer_id, str(command.event_id)
+            )
+            if prior_event is not None:
+                prior_command = {
+                    key: value for key, value in prior_event.items()
+                    if key != "resulting_state"
+                }
+                require(
+                    digest(prior_command) == digest(event),
+                    "IDEMPOTENCY_CONFLICT",
+                    "event_id was used with a different alert event",
+                    status=409,
+                )
+                return TaskAlertState.model_validate(prior_event["resulting_state"])
+
+            current = await self.repo.get_alert(command.alert_id)
+            state = self._transition(command, current)
+            state_document = state.model_dump(mode="json")
+            await self.repo.save_event(event, state_document)
+            await self.repo.save_alert(state_document)
+            append_changed_event = getattr(self.repo, "append_changed_event", None)
+            if append_changed_event is not None:
+                await append_changed_event(state_document)
+            return state
+
+    def _transition(
+        self, command: TaskAlertEvent, current: tuple[int, dict] | None
+    ) -> TaskAlertState:
+        now = self.clock()
+        require(now.tzinfo is not None, "SCHEMA_INVALID", "Alert clock must be timezone-aware")
+        if current is None:
+            require(command.event_type == "raised", "INVALID_TRANSITION",
+                    "An alert must be raised before it can change state", status=409)
+            revision = 1
+            status = "active"
+        else:
+            revision, document = current
+            require(document.get("producer_id") == command.producer_id,
+                    "PERMISSION_DENIED", "Only the originating producer can update an alert", status=403)
+            require(document.get("source") == command.source.model_dump(mode="json"),
+                    "SCOPE_MISMATCH", "Alert source cannot change", status=409)
+            require(
+                all(
+                    document.get(field) == event_value
+                    for field, event_value in {
+                        "condition_id": command.condition_id,
+                        "title": command.title,
+                        "description": command.description,
+                        "occurred_at": command.model_dump(mode="json")["occurred_at"],
+                        "severity": command.severity,
+                    }.items()
+                ),
+                "IDEMPOTENCY_CONFLICT",
+                "Alert facts cannot change after the alert is raised",
+                status=409,
+            )
+            allowed = {
+                ("active", "acknowledged"): "acknowledged",
+                ("active", "resolved"): "resolved",
+                ("acknowledged", "resolved"): "resolved",
+            }
+            status = allowed.get((document["status"], command.event_type))
+            require(status is not None, "INVALID_TRANSITION",
+                    "Alert state transition is not permitted", status=409)
+            revision += 1
+        return TaskAlertState(
+            alert_id=command.alert_id,
+            revision=revision,
+            producer_id=command.producer_id,
+            condition_id=command.condition_id,
+            source=command.source,
+            title=command.title,
+            description=command.description,
+            occurred_at=command.occurred_at,
+            severity=command.severity,
+            status=status,
+            updated_at=now,
+        )

app/services/timeseries_history.py

@@ -16,6 +16,7 @@ from app.domain.schemas.timeseries_history import (
 from app.infra.db.postgresql.analysis import AnalysisRepository
 from app.infra.db.postgresql.network_settings import NetworkSettingsRepository
 from app.infra.db.postgresql.scada import ScadaInfoRepository
+from app.infra.db.postgresql import read_queries
 from app.infra.db.timescaledb.repositories.analysis import AnalysisResultsRepository
 from app.infra.db.timescaledb.repositories.realtime import RealtimeRepository
 from app.infra.db.timescaledb.repositories.scada import ScadaRepository
@@ -56,13 +57,69 @@ def _points(
     ]
 
 
+async def _query_page(timescale_conn, postgres_conn, query, selected_devices, limit, offset):
+    descriptors = []
+    series = []
+
+    def add(*, source, element_type, element_id, device_id=None, field, unit, unit_inferred=False):
+        descriptors.append({
+            "source": source, "element_type": element_type,
+            "id": device_id if source == "scada" else element_id, "field": field,
+        })
+        metric = _target_metric(element_type)
+        series.append(ElementHistorySeries(
+            element_id=element_id, element_type=element_type, device_id=device_id,
+            metric=metric, source=(
+                HistorySource.SCADA_RAW if field == "monitored_value" else HistorySource.SCADA_CLEANED
+            ) if source == "scada" else (
+                HistorySource.ANALYSIS_SIMULATION if source == "analysis" else HistorySource.REALTIME_SIMULATION
+            ),
+            source_unit=unit, display_unit=display_unit_for_metric(metric.value),
+            unit_inferred=unit_inferred, points=[],
+        ))
+
+    # Stable series ordering makes a page crossing raw/cleaned/model boundaries repeatable.
+    for device_id, (device, element_type, element_id) in sorted(selected_devices.items()):
+        for field in ("monitored_value", "cleaned_value"):
+            add(source="scada", element_type=element_type, element_id=element_id,
+                device_id=device_id, field=field, unit=device["measurement_unit"])
+    if query.mode.value != "observed":
+        units = await NetworkSettingsRepository.get_result_units(postgres_conn)
+        source = "analysis" if query.mode.value == "analysis_comparison" else "realtime"
+        inferred = False
+        if source == "analysis":
+            run = await AnalysisRepository.get_run(postgres_conn, query.run_id)
+            if run is None:
+                raise ValueError(f"analysis run does not exist: {query.run_id}")
+            recorded = (run.get("parameters") or {}).get("result_units")
+            inferred = not isinstance(recorded, dict)
+            if not inferred:
+                units = recorded
+        targets = {(t.element_type, t.element_id) for t in query.elements}
+        for element_type, element_id in sorted(targets):
+            metric = _target_metric(element_type)
+            add(source=source, element_type=element_type, element_id=element_id,
+                field=metric.value, unit=str(units.get(metric.value, "")).strip(), unit_inferred=inferred)
+    rows = await read_queries.history_rows(
+        timescale_conn, descriptors, start=query.start_time, end=query.end_time,
+        run_id=query.run_id, limit=limit + 1, offset=offset,
+    )
+    for row in rows[:limit]:
+        series[row["series_index"]].points.append(HistoryPoint(time=row["time"], value=row["value"]))
+    return ElementHistoryResponse(series=[s for s in series if s.points], truncated=len(rows) > limit)
+
+
 class TimeseriesHistoryService:
     @staticmethod
     async def query(
         timescale_conn: AsyncConnection,
         postgres_conn: AsyncConnection,
         query: ElementHistoryQuery,
+        *, row_limit: int | None = None, row_offset: int = 0,
     ) -> ElementHistoryResponse:
+        if row_offset < 0 or (row_offset and row_limit is None) or (row_limit is not None and row_limit < 1):
+            raise ValueError("row_offset requires a bounded page")
+
         pipe_ids = list(
             dict.fromkeys(
                 target.element_id
@@ -78,8 +135,10 @@ class TimeseriesHistoryService:
             )
         )
         devices = await ScadaInfoRepository.get_scadas_for_elements(
-            postgres_conn, junction_ids, pipe_ids
+            postgres_conn, junction_ids, pipe_ids, **({"row_limit": 1001} if row_limit is not None else {})
         )
+        if row_limit is not None and len(devices) > 1000:
+            raise ValueError("too many device mappings; narrow the element selection")
         devices_by_element: dict[tuple[HistoryElementType, str], list[dict[str, Any]]] = (
             defaultdict(list)
         )
@@ -122,6 +181,12 @@ class TimeseriesHistoryService:
                     target.element_id,
                 )
 
+        if row_limit is not None and len(selected_devices) > 20:
+            raise ValueError("select at most 20 devices explicitly")
+        if row_limit is not None:
+            return await _query_page(
+                timescale_conn, postgres_conn, query, selected_devices, row_limit, row_offset,
+            )
         scada_rows = await ScadaRepository.get_scada_by_ids_time_range(
             timescale_conn,
             list(selected_devices),

app/services/web_search.py

@@ -1,93 +0,0 @@
-from typing import Any, Literal
-
-import httpx
-from pydantic import BaseModel, Field
-
-from app.core.config import settings
-
-
-Freshness = Literal["noLimit", "oneDay", "oneWeek", "oneMonth", "oneYear"]
-
-
-class WebSearchRequest(BaseModel):
-    query: str = Field(..., min_length=1, description="搜索关键词")
-    freshness: Freshness | str = Field(
-        default="noLimit",
-        description="时间范围：noLimit、oneDay、oneWeek、oneMonth、oneYear 或日期范围",
-    )
-    summary: bool = Field(default=True, description="是否返回网页摘要")
-    count: int = Field(default=10, ge=1, le=50, description="返回结果数量")
-    include: list[str] | None = Field(default=None, description="限定搜索域名")
-    exclude: list[str] | None = Field(default=None, description="排除搜索域名")
-
-
-class BochaSearchConfigError(RuntimeError):
-    pass
-
-
-class BochaSearchAPIError(RuntimeError):
-    def __init__(self, status_code: int, detail: Any):
-        super().__init__("Bocha Web Search API request failed")
-        self.status_code = status_code
-        self.detail = detail
-
-
-def _build_payload(request: WebSearchRequest) -> dict[str, Any]:
-    payload = request.model_dump(exclude_none=True)
-    if request.include:
-        payload["include"] = ",".join(request.include)
-    if request.exclude:
-        payload["exclude"] = ",".join(request.exclude)
-    return payload
-
-
-async def search_bocha_web(
-    request: WebSearchRequest,
-    *,
-    client: httpx.AsyncClient | None = None,
-) -> dict[str, Any]:
-    if not settings.BOCHA_API_KEY:
-        raise BochaSearchConfigError("BOCHA_API_KEY is not configured")
-
-    headers = {
-        "Authorization": f"Bearer {settings.BOCHA_API_KEY}",
-        "Content-Type": "application/json",
-    }
-    payload = _build_payload(request)
-
-    if client is not None:
-        response = await client.post(
-            settings.BOCHA_WEB_SEARCH_URL,
-            headers=headers,
-            json=payload,
-        )
-        return _parse_response(response)
-
-    async with httpx.AsyncClient(
-        timeout=settings.BOCHA_WEB_SEARCH_TIMEOUT_SECONDS
-    ) as managed_client:
-        response = await managed_client.post(
-            settings.BOCHA_WEB_SEARCH_URL,
-            headers=headers,
-            json=payload,
-        )
-        return _parse_response(response)
-
-
-def _parse_response(response: httpx.Response) -> dict[str, Any]:
-    try:
-        response.raise_for_status()
-    except httpx.HTTPStatusError as exc:
-        raise BochaSearchAPIError(
-            exc.response.status_code,
-            _response_detail(exc.response),
-        ) from exc
-
-    return response.json()
-
-
-def _response_detail(response: httpx.Response) -> Any:
-    try:
-        return response.json()
-    except ValueError:
-        return response.text

app/services/workbench_runtime.py

@@ -0,0 +1,482 @@
+"""Project persisted runtime state into the read-only supply workbench feed."""
+
+from collections.abc import Callable
+from datetime import datetime, timezone
+from math import isfinite
+from typing import Any
+
+from app.domain.schemas.workbench_runtime import WorkbenchPersistedAlert
+
+_ANALYSIS_TYPE_LABELS = {
+    "burst_analysis": "爆管分析",
+    "burst_detection": "爆管检测",
+    "burst_location": "爆管定位",
+    "contaminant_analysis": "污染物分析",
+    "contaminant_simulation": "污染物模拟",
+    "dma_leak_identification": "DMA 漏损识别",
+    "flushing_analysis": "冲洗分析",
+    "pressure_regulation": "压力调控",
+    "sensor_placement": "传感器布点",
+    "valve_close_analysis": "关阀分析",
+}
+_ANALYSIS_STATUS_LABELS = {
+    "queued": "排队中",
+    "pending": "等待中",
+    "running": "运行中",
+    "completed": "已完成",
+    "succeeded": "已成功",
+    "failed": "已失败",
+    "error": "异常",
+    "cancelled": "已取消",
+}
+
+
+async def build_workbench_runtime_snapshot(
+    *,
+    analysis_runs: list[dict[str, Any]] | None = None,
+    latest_analysis_results: list[dict[str, Any]] | None = None,
+    latest_analysis_map_payload: dict[str, Any] | None = None,
+    alert_storage_configured: bool = False,
+    persisted_alerts: list[dict[str, Any]] | None = None,
+    scenario_availability: str | None = None,
+    artifact_availability: str | None = None,
+    schedule_availability: str | None = None,
+    alert_availability: str | None = None,
+    projection_degraded: bool = False,
+    clock: Callable[[], datetime] = lambda: datetime.now(timezone.utc),
+) -> dict[str, Any]:
+    persisted_analysis_runs = (analysis_runs or [])[:100]
+    scenarios = [project_analysis_scenario(run) for run in persisted_analysis_runs[:20]]
+    latest_analysis = (
+        project_latest_analysis(
+            persisted_analysis_runs[0],
+            latest_analysis_results or [],
+            referenced_map_payload=latest_analysis_map_payload,
+        )
+        if persisted_analysis_runs
+        else None
+    )
+    stored_alerts = [
+        WorkbenchPersistedAlert.model_validate(item)
+        for item in (
+            (persisted_alerts or [])[:100]
+            if alert_storage_configured
+            else []
+        )
+    ]
+    active_stored_alerts = [
+        {
+            "id": f"backend-alert-{alert.alert_id}",
+            "condition_id": alert.condition_id,
+            "source": alert.source.model_dump(),
+            "title": alert.title,
+            "description": alert.description,
+            "occurred_at": alert.occurred_at,
+            "severity": alert.severity,
+        }
+        for alert in stored_alerts
+        if alert.status == "active"
+    ]
+    projected_alerts = active_stored_alerts
+    projected_alerts.sort(
+        key=lambda alert: _timestamp(alert["occurred_at"]),
+        reverse=True,
+    )
+    source = "none"
+    resolved_scenario_availability = scenario_availability or (
+        "ready" if scenarios else "not_configured"
+    )
+    resolved_artifact_availability = artifact_availability or (
+        "ready" if latest_analysis else "not_configured"
+    )
+    resolved_schedule_availability = schedule_availability or "not_configured"
+    resolved_work_order_availability = "not_configured"
+    resolved_alert_availability = alert_availability or (
+        "ready" if alert_storage_configured else "not_configured"
+    )
+    section_availabilities = (
+        resolved_scenario_availability,
+        resolved_artifact_availability,
+        resolved_schedule_availability,
+        resolved_work_order_availability,
+        resolved_alert_availability,
+    )
+    return {
+        "availability": (
+            "degraded"
+            if projection_degraded or "degraded" in section_availabilities
+            else
+            "ready"
+            if (
+                source != "none"
+                or alert_storage_configured
+            )
+            else "not_configured"
+        ),
+        "source": source,
+        "scenario_availability": resolved_scenario_availability,
+        "artifact_availability": resolved_artifact_availability,
+        "schedule_availability": resolved_schedule_availability,
+        "work_order_availability": resolved_work_order_availability,
+        "alert_availability": resolved_alert_availability,
+        "generated_at": clock(),
+        "conditions": [],
+        "alerts": projected_alerts,
+        "scenarios": scenarios,
+        "latest_analysis": latest_analysis,
+        "work_orders": [],
+    }
+
+
+def project_analysis_scenario(run: dict[str, Any]) -> dict[str, Any]:
+    run_id = str(run["run_id"])
+    run_type = str(run.get("run_type") or "analysis")
+    run_status = str(run.get("status") or "unknown").lower()
+    type_label = _analysis_type_label(run_type)
+    return {
+        "id": f"analysis-run-{run_id}",
+        "run_id": run_id,
+        "name": str(run.get("name") or type_label),
+        "description": (
+            f"{type_label} · {_analysis_status_label(run_status)} · "
+            f"创建人 {run.get('created_by') or 'unknown'}"
+        ),
+        "status": _analysis_scenario_state(run_status),
+        "run_type": run_type,
+        "run_type_label": type_label,
+        "run_status": run_status,
+        "created_at": run["created_at"],
+    }
+
+
+def project_analysis_history_item(run: dict[str, Any]) -> dict[str, Any]:
+    run_type = str(run.get("run_type") or "analysis")
+    run_status = str(run.get("status") or "unknown").lower()
+    created_at = run["created_at"]
+    return {
+        "run_id": str(run["run_id"]),
+        "name": str(run.get("name") or _analysis_type_label(run_type)),
+        "run_type": run_type,
+        "run_type_label": _analysis_type_label(run_type),
+        "status": run_status,
+        "status_label": _analysis_status_label(run_status),
+        "created_by": str(run.get("created_by") or "unknown"),
+        "created_at": created_at,
+        "started_at": run.get("started_at") or created_at,
+        "has_results": bool(run.get("has_results")),
+    }
+
+
+def project_latest_analysis(
+    run: dict[str, Any],
+    results: list[dict[str, Any]],
+    *,
+    referenced_map_payload: dict[str, Any] | None = None,
+) -> dict[str, Any]:
+    run_type = str(run.get("run_type") or "analysis")
+    run_status = str(run.get("status") or "unknown").lower()
+    created_at = run["created_at"]
+    return {
+        "run_id": str(run["run_id"]),
+        "name": str(run.get("name") or _analysis_type_label(run_type)),
+        "run_type": run_type,
+        "run_type_label": _analysis_type_label(run_type),
+        "status": run_status,
+        "status_label": _analysis_status_label(run_status),
+        "created_by": str(run.get("created_by") or "unknown"),
+        "created_at": created_at,
+        "started_at": run.get("started_at") or created_at,
+        "results": [
+            {
+                "result_type": str(result.get("result_type") or "analysis_result"),
+                "created_at": result.get("created_at") or created_at,
+                "facts": _analysis_result_facts(result),
+            }
+            for result in results
+        ],
+        "map_payload": _analysis_map_payload(results) or referenced_map_payload,
+    }
+
+
+def sensor_location_ids_for_map_projection(
+    results: list[dict[str, Any]],
+) -> list[str]:
+    if _analysis_map_payload(results) is not None:
+        return []
+    for result in reversed(results):
+        if result.get("result_type") != "sensor_placement":
+            continue
+        payload = result.get("payload")
+        if not isinstance(payload, dict):
+            continue
+        locations = _non_empty_text_list(payload.get("sensor_locations"))
+        if locations:
+            return list(dict.fromkeys(locations))[:500]
+    return []
+
+
+def project_sensor_location_map_payload(
+    node_ids: list[str], map_points: list[dict[str, Any]]
+) -> dict[str, Any] | None:
+    points_by_id = {
+        str(point.get("node_id")): point
+        for point in map_points
+        if point.get("node_id") is not None
+    }
+    features: list[dict[str, Any]] = []
+    for position, node_id in enumerate(node_ids, start=1):
+        point = points_by_id.get(node_id)
+        if point is None:
+            continue
+        longitude = point.get("longitude")
+        latitude = point.get("latitude")
+        if (
+            isinstance(longitude, bool)
+            or isinstance(latitude, bool)
+            or not isinstance(longitude, (int, float))
+            or not isinstance(latitude, (int, float))
+            or not isfinite(float(longitude))
+            or not isfinite(float(latitude))
+        ):
+            continue
+        features.append({
+            "type": "Feature",
+            "id": node_id,
+            "properties": {
+                "label": f"监测点 {position} · {node_id}",
+                "node_id": node_id,
+                "result_type": "sensor_placement",
+                "position": position,
+            },
+            "geometry": {
+                "type": "Point",
+                "coordinates": [float(longitude), float(latitude)],
+            },
+        })
+    if not features:
+        return None
+    return {"type": "FeatureCollection", "features": features}
+
+
+def _analysis_result_facts(result: dict[str, Any]) -> list[dict[str, str]]:
+    facts: list[dict[str, str]] = []
+    node_id = _non_empty_text(result.get("node_id"))
+    link_id = _non_empty_text(result.get("link_id"))
+    if node_id:
+        facts.append({"label": "关联节点", "value": node_id})
+    if link_id:
+        facts.append({"label": "关联管段", "value": link_id})
+
+    payload = result.get("payload")
+    if not isinstance(payload, dict):
+        return facts
+    result_type = str(result.get("result_type") or "")
+    if result_type == "sensor_placement":
+        locations = _non_empty_text_list(payload.get("sensor_locations"))
+        sensor_count = _finite_number_text(payload.get("sensor_number"))
+        if sensor_count is None and locations:
+            sensor_count = str(len(locations))
+        if sensor_count is not None:
+            facts.append({"label": "测点数量", "value": sensor_count})
+        minimum_diameter = _finite_number_text(payload.get("minimum_diameter"))
+        if minimum_diameter is not None:
+            facts.append({"label": "最小管径参数", "value": minimum_diameter})
+        if locations:
+            facts.append({
+                "label": "测点位置",
+                "value": _summarize_identifiers(locations),
+            })
+    elif result_type == "leakage_identification":
+        run_status = _non_empty_text(payload.get("run_status"))
+        if run_status:
+            facts.append({"label": "结果状态", "value": run_status})
+        sensor_nodes = _non_empty_text_list(payload.get("sensor_nodes"))
+        if sensor_nodes:
+            facts.append({"label": "传感器节点数", "value": str(len(sensor_nodes))})
+        rows = payload.get("rows")
+        if isinstance(rows, list):
+            facts.append({"label": "结果行数", "value": str(len(rows))})
+        areas = payload.get("areas")
+        if isinstance(areas, list):
+            facts.append({"label": "分区数量", "value": str(len(areas))})
+    return facts
+
+
+def _non_empty_text(value: object) -> str | None:
+    if not isinstance(value, str):
+        return None
+    normalized = value.strip()
+    return normalized or None
+
+
+def _non_empty_text_list(value: object) -> list[str]:
+    if not isinstance(value, list):
+        return []
+    return [text for item in value if (text := _non_empty_text(item)) is not None]
+
+
+def _finite_number_text(value: object) -> str | None:
+    if isinstance(value, bool) or not isinstance(value, (int, float)):
+        return None
+    if isinstance(value, float) and not isfinite(value):
+        return None
+    return str(value)
+
+
+def _summarize_identifiers(values: list[str], limit: int = 12) -> str:
+    visible = values[:limit]
+    suffix = f" 等 {len(values)} 个" if len(values) > limit else ""
+    return "、".join(visible) + suffix
+
+
+def _analysis_map_payload(
+    results: list[dict[str, Any]],
+) -> dict[str, Any] | None:
+    for result in reversed(results):
+        payload = result.get("payload")
+        if not isinstance(payload, dict):
+            continue
+        candidate = payload.get("drawing_payload")
+        if not isinstance(candidate, dict):
+            candidate = payload
+        if (
+            candidate.get("type") == "FeatureCollection"
+            and isinstance(candidate.get("features"), list)
+            and candidate["features"]
+        ):
+            return candidate
+    return None
+
+
+def _analysis_type_label(run_type: str) -> str:
+    return _ANALYSIS_TYPE_LABELS.get(run_type, run_type)
+
+
+def _analysis_status_label(status: str) -> str:
+    return _ANALYSIS_STATUS_LABELS.get(status, status)
+
+
+def _analysis_scenario_state(status: str) -> str:
+    if status == "running":
+        return "active"
+    if status in {"completed", "succeeded", "failed", "error"}:
+        return "review"
+    return "draft"
+
+
+def _timestamp(value: datetime | str) -> float:
+    parsed = (
+        value
+        if isinstance(value, datetime)
+        else datetime.fromisoformat(value.replace("Z", "+00:00"))
+    )
+    return parsed.timestamp()
+
+
+# Persisted runtime queries shared by REST and MCP.
+from contextlib import asynccontextmanager
+import logging
+from pydantic import ValidationError
+from app.auth.project_dependencies import ProjectContext
+from app.infra.db.postgresql.analysis import AnalysisRepository
+from app.infra.db.postgresql.workbench_operations import WorkbenchOperationsRepository
+from app.domain.schemas.workbench_runtime import WorkbenchRuntimeSnapshot
+
+logger = logging.getLogger(__name__)
+
+
+@asynccontextmanager
+async def _projection_transaction(conn):
+    transaction = getattr(conn, "transaction", None)
+    if transaction is None:
+        yield
+        return
+    async with transaction():
+        yield
+
+
+async def _sensor_location_map_payload(conn, results):
+    node_ids = sensor_location_ids_for_map_projection(results)
+    if not node_ids:
+        return None
+    points = await AnalysisRepository.list_junction_map_points(conn,node_ids)
+    return project_sensor_location_map_payload(node_ids,points)
+
+
+def _validate_documents(documents, model, label):
+    valid=[];rejected=0
+    for document in documents:
+        try:valid.append(model.model_validate(document).model_dump())
+        except ValidationError:
+            rejected+=1
+            logger.warning("Workbench %s document was rejected",label)
+    return valid,rejected
+
+
+async def read_workbench_runtime(ctx: ProjectContext, conn):
+    analysis_runs: list[dict] | None = None
+    latest_results: list[dict] | None = None
+    latest_analysis_map_payload: dict | None = None
+    analysis_availability = "not_configured"
+    alert_storage_configured = False
+    alerts: list[dict] | None = None
+    schedule_availability = "not_configured"
+    alert_availability = "not_configured"
+    try:
+        async with _projection_transaction(conn):
+            analysis_runs = await AnalysisRepository.list_recent_runs(conn, limit=100)
+            if analysis_runs:
+                latest_results = await AnalysisRepository.list_recent_results(
+                    conn, analysis_runs[0]["run_id"], limit=100
+                )
+        analysis_availability = "ready" if analysis_runs else "not_configured"
+    except Exception:
+        logger.warning("Workbench analysis projection is unavailable", exc_info=True)
+        analysis_runs = None
+        latest_results = None
+        analysis_availability = "degraded"
+    if analysis_runs and latest_results:
+        try:
+            async with _projection_transaction(conn):
+                latest_analysis_map_payload = await _sensor_location_map_payload(
+                    conn, latest_results
+                )
+        except Exception:
+            logger.warning(
+                "Workbench analysis spatial references are unavailable",
+                exc_info=True,
+            )
+    try:
+        async with _projection_transaction(conn):
+            operations_availability = await WorkbenchOperationsRepository.availability(conn)
+        alert_storage_configured = operations_availability.alerts
+    except Exception:
+        logger.warning("Workbench operations storage probe is unavailable", exc_info=True)
+        schedule_availability = "degraded"
+        alert_availability = "degraded"
+    if alert_storage_configured:
+        try:
+            async with _projection_transaction(conn):
+                alert_documents = await WorkbenchOperationsRepository.list_alerts(
+                    conn, str(ctx.project_id)
+                )
+            alerts, invalid_alerts = _validate_documents(
+                alert_documents, WorkbenchPersistedAlert, "alert"
+            )
+            alert_availability = "degraded" if invalid_alerts else "ready"
+        except Exception:
+            alerts = None
+            alert_availability = "degraded"
+            logger.warning("Workbench alert projection is unavailable", exc_info=True)
+    return await build_workbench_runtime_snapshot(
+        analysis_runs=analysis_runs,
+        latest_analysis_results=latest_results,
+        latest_analysis_map_payload=latest_analysis_map_payload,
+        alert_storage_configured=alert_storage_configured,
+        persisted_alerts=alerts,
+        scenario_availability=analysis_availability,
+        artifact_availability=analysis_availability,
+        schedule_availability=schedule_availability,
+        alert_availability=alert_availability,
+        projection_degraded=False,
+    )

contracts/manifest.json

@@ -3,7 +3,7 @@
   "contracts": {
     "server": {
       "file": "server-v1.openapi.json",
-      "sha256": "b565d841061c9091f48ff3118bcc0cbb1b918b0cb8c2316d177570e8b2d8ba29"
+      "sha256": "949077b0e1c711cfef78319d24cbe081e8d71444e7b25018507e12ba374018a3"
     }
   }
 }

docs/dynamic-investigation-plan.md

@@ -0,0 +1,53 @@
+# 动态调查计划数据契约
+
+调查任务从工单投诉内容或人工请求创建。Agent 在本次执行中生成 `execution-plan@2`，根据新证据追加修订；数据库保存不可变计划快照的 JSONB 文档，并不保存固定流程模板。任务、Run、证据、访问申请、审核动作和报告各有独立生命周期。计划中的 `gate_ref` 只引用已经持久化的访问申请或动作，不能把审批结果写成模型自行判断的文本。
+
+## `execution-plan@2`
+
+| 字段 | 含义 |
+| --- | --- |
+| `schema` | 固定为 `execution-plan@2`；旧快照继续按 `execution-plan@1` 读取 |
+| `title`、`reason` | 本次计划名称和制定或修订原因 |
+| `nodes[]` | 具有稳定 `id`、`title`、`owner`、`status`、`detail`、`kind`、`evidence_refs[]`、`action_refs[]` 的步骤 |
+| `nodes[].kind` | `activity` 单项执行；`decision` 条件判断；`parallel_split` 并行分叉；`parallel_join` 并行汇合；`gate` 人工审核或授权等待点 |
+| `nodes[].selected_edge_id` | 仅判断节点使用。判断完成时指向实际选择的出边；未选择时为 `null` |
+| `nodes[].gate_ref` | 仅审核节点使用。UUID 指向任务内已有的访问申请或待确认动作 |
+| `nodes[].revisits_node_id` | 再次核验时，新节点引用先前节点 ID；不创建回环连线 |
+| `edges[]` | 每条有稳定 `id`、`from`、`to` 和可空 `condition`；判断路线必须有条件文字 |
+
+图必须无环且只有一个起点。只有 `decision` 和 `parallel_split` 可以有多条出边；`parallel_join` 至少有两条入边。选中路线的直接目标不能标为跳过；未选路线的直接目标不能标为执行中、待人工或已完成。执行中和已完成的步骤，前置步骤必须完成或跳过。已完成的业务判断与活动必须引用真实证据或动作；纯分叉、汇合和审核节点不要求伪造证据。服务端在发布时再次核验引用、当前 Run 和会话租约。
+
+下面是最小示例；`gate_ref` 所指申请先通过独立接口创建，再发布包含审核节点的计划。审批后，Agent 从当前任务上下文读取结果并发布下一版计划。
+
+```json
+{
+  "schema": "execution-plan@2",
+  "title": "水质投诉调查",
+  "reason": "核对工单位置和当前风险",
+  "nodes": [
+    {"id": "risk", "title": "判断紧急风险", "owner": "agent", "status": "completed", "detail": "根据投诉记录判断", "kind": "decision", "selected_edge_id": "urgent", "evidence_refs": ["complaint-1"], "action_refs": []},
+    {"id": "field", "title": "安排现场采样", "owner": "human", "status": "pending", "detail": "待安排现场人员", "kind": "activity", "evidence_refs": [], "action_refs": []},
+    {"id": "routine", "title": "常规核查", "owner": "human", "status": "skipped", "detail": "本次选择紧急路线", "kind": "activity", "evidence_refs": [], "action_refs": []},
+    {"id": "split", "title": "并行核对", "owner": "agent", "status": "pending", "detail": "授权后同时核对现场和数据", "kind": "parallel_split", "evidence_refs": [], "action_refs": []},
+    {"id": "upstream", "title": "上游监测", "owner": "agent", "status": "pending", "detail": "待批准读取", "kind": "activity", "evidence_refs": [], "action_refs": []},
+    {"id": "network", "title": "管网运行", "owner": "agent", "status": "pending", "detail": "待核对运行事件", "kind": "activity", "evidence_refs": [], "action_refs": []},
+    {"id": "join", "title": "汇总调查证据", "owner": "agent", "status": "waiting", "detail": "等待现场与数据结果", "kind": "parallel_join", "evidence_refs": [], "action_refs": []},
+    {"id": "access", "title": "审核数据范围", "owner": "human", "status": "waiting", "detail": "等待项目成员审核", "kind": "gate", "gate_ref": "84000000-0000-4000-8000-000000000001", "evidence_refs": [], "action_refs": []}
+  ],
+  "edges": [
+    {"id": "urgent", "from": "risk", "to": "access", "condition": "有紧急风险"},
+    {"id": "normal", "from": "risk", "to": "routine", "condition": "无紧急风险"},
+    {"id": "access-split", "from": "access", "to": "split", "condition": null},
+    {"id": "field-route", "from": "split", "to": "field", "condition": "现场"},
+    {"id": "upstream-route", "from": "split", "to": "upstream", "condition": "上游"},
+    {"id": "network-route", "from": "split", "to": "network", "condition": "管网"},
+    {"id": "field-join", "from": "field", "to": "join", "condition": null},
+    {"id": "upstream-join", "from": "upstream", "to": "join", "condition": null},
+    {"id": "network-join", "from": "network", "to": "join", "condition": null}
+  ]
+}
+```
+
+实际调查可以把现场路径与 Agent 路径放在同一个并行分叉下，并在下一次修订中补充更多判断或并行组。已发布计划保留原样；拒绝授权、人工补充、复测或回访触发新 Run 时，Agent 读取当前任务事实后继续规划。旧 `execution-plan@1` 只有依赖关系，前端按原始前置条件展示，不推断它代表条件选择还是并行执行。
+
+访问申请独立存于 `workbench.operational_task_access_requests`，包括请求范围、理由、申请节点、审核人、批准范围和版本；工单草稿及其审批、回执仍存于动作表。前端默认呈现选中路线，并在明确的分叉与汇合处组成并行组；未选路线折叠保留供核对。

docs/integration-ui-keycloak-client.md

@@ -0,0 +1,22 @@
+# 15173 联调 UI Keycloak 客户端
+
+2026-09-23 已在 `tjwater` realm 验证独立 public client
+`tjwater-integration-ui`。该客户端只用于本机 15173 联调页面，不替代或修改
+现有 `next-tjwater` 客户端。
+
+已验证配置：
+
+- Standard Authorization Code Flow：开启。
+- PKCE：强制 `S256`。
+- Implicit Flow、Direct Access Grants、Service Accounts：关闭。
+- Public client：开启，不保存客户端密钥。
+- Full Scope Allowed：关闭；未授予 realm-management 或系统管理员角色。
+- Redirect URI：`http://127.0.0.1:15173/*`、`http://localhost:15173/*`。
+- Web Origin：`http://127.0.0.1:15173`、`http://localhost:15173`。
+
+可使用以下幂等脚本复核或恢复同一配置；管理员密码仅从隐藏提示或进程环境
+读取，不会写入文件：
+
+```bash
+conda run -n server python scripts/provision_integration_ui_client.py
+```

docs/investigation-progress-and-presentation.md

@@ -0,0 +1,116 @@
+# 调查进展与实时展示
+
+调查执行中可持续发布正文和完整 Mermaid 图，尚未形成结论时也能发布。
+Server 将阅读进展、业务报告和浏览器展示请求分别保存。运行结束仍要求同轮最终报告，
+进展中的节点状态不参与调度、审批或结束判断。
+
+## 进展快照
+
+`POST /api/v1/operational-tasks/{task_id}/agent-artifacts` 新增可选 `progress`：
+
+```json
+{
+  "request_id": "<uuid>",
+  "run_id": "<current-run-uuid>",
+  "source_session_id": "<claimed-session-id>",
+  "progress": {
+    "summary": "正在核对投诉时段的监测记录，尚未形成结论。",
+    "diagram": {
+      "title": "当前调查进展",
+      "mermaid": "flowchart LR\nread[核查监测记录] --> judge[判断证据范围]",
+      "node_refs": [{
+        "node_id": "read",
+        "label": "核查监测记录",
+        "status": "running",
+        "evidence_refs": [],
+        "action_refs": [],
+        "map_targets": [{"source_id": "pipes", "feature_id": "<known-pipe-id>"}]
+      }]
+    }
+  }
+}
+```
+
+示例中的编号由实际任务和资产替换。正文最长 20,000 字符，`diagram` 可空。
+图示沿用 [调查报告契约](investigation-report-contract.md) 的 Mermaid 安全限制。
+节点状态可空，或为 `pending/running/completed/waiting_human/failed/skipped/cancelled`。
+`map_targets` 只接受 `junctions/pipes/valves/pumps/tanks/reservoirs/scada` 七类来源和要素编号。
+图中节点引用在本次发布内唯一，证据和动作引用必须属于该任务。跨修订可以修改节点编号、
+分支和文字，Server 不验证依赖执行顺序或节点完工。
+
+每次传入 `progress` 都完整替换该 run 的快照，传入 `diagram: null` 清除上一版图示。
+省略 `progress` 保留原快照。同一个请求编号重试只发布一次。
+`workbench.operational_task_progress` 每个任务和 run 保存最新一份快照，独立于七天事件清理。
+任务详情和 Agent 上下文返回 `latest_progress`，包含 `revision/run_id/execution_generation/`
+`published_at/source_session_id/progress`。只读取当前执行代次，新轮次尚未发布时返回空。
+快照更新发出 `operational.progress.changed`，前端据此读取最新内容。
+
+## 浏览器实时联动
+
+`POST /api/v1/operational-tasks/{task_id}/presentation-events` 接收
+`request_id/run_id/source_session_id/command`，要求任务执行权限。
+Server 校验当前 run、绑定会话、有效租约和幂等身份，记录 `operational.presentation.requested`。
+响应为 `recorded/event_seq/run_id/execution_generation`，表示展示请求已保存，
+不表示浏览器已显示或现场动作已执行。
+
+| command.action | 参数 |
+| --- | --- |
+| highlight_features | layer、feature_ids、effect：steady 或 pulse |
+| clear_highlights | 无 |
+| open_chart | chart_id、chart、非空 evidence_refs |
+| close_chart | chart_id |
+| set_overlay | overlay_id、title、mode：geometry 或 heatmap、evidence_id |
+| clear_overlay | overlay_id |
+
+图表限定折线和柱形，包含 `title/chart_type/x_data/series`。最多 100 个横轴值和 8 组有限数值序列，
+序列长度与横轴对齐，引用证据必须属于当前任务。地图高亮来源沿用上述七类。
+覆盖层输入不接受模型提供的 `data`。Server 从证据的不可变数据快照提取现有 GeoJSON，
+或使用已记录的 `longitude/latitude` 点位，将 `FeatureCollection` 加入事件中的 `command.data`。
+热力图只接受记录中的点，权重为有限非负数，未记录权重时取 1。
+没有坐标的地址、空数据、非法几何、模型自绘区域均不能生成覆盖层。该接口不做地址解析、
+范围推算或污染传播计算。展示事件总大小上限为 512 KiB，超限返回 422。
+
+`GET /api/v1/operational-tasks/{task_id}/presentation-cursor` 在同一事务中读取
+`task_id/execution_generation/active_run/event_cursor`，其中活跃 run 仅返回 `run_id/generation/status`。
+浏览器打开 Dev 联动、切换任务或断线重连时，以此建立新基线，只消费之后的展示事件。
+跟随开关只影响当前浏览器的临时展示，正文和图示更新始终可读，调查仍由后台继续执行。
+关闭 Dev 面板不改变跟随开关。关闭跟随开关、切换任务或断线时清理临时效果，
+执行结束停止执行态动效，已显示的图表与覆盖层保留供阅读。不回放历史请求，不重新执行调查。
+
+## 人工反馈续调
+
+外围同步器根据动作回执找到关联子工单。样板创建固定为 `revision=1`，首次创建回执不作为
+新的现场反馈入队。后续版本可唤醒已结束调查，旧报告和证据保留，生成新的执行代次。
+现场备注从工单 `activity.kind=feedback` 的记录进入反馈证据，仍标明样板合成来源。
+重复请求或同一源版本重放不增加执行代次，取消调查拒绝新反馈。
+
+同步器先送达各父调查反馈，再推进预约工单投影的源版本。定向源事件和动作回执补查保留，
+处理源更新早于回执保存的情况。全量对账只为源版本变化补查已结束调查，本次不回填历史反馈。
+原投诉收到新版本后同步推进其已有来源修订号。数据不足仍正常结束本轮，
+只有新的可信业务输入进入后续轮次，不依赖图示节点制造人工等待。
+
+## 部署与验证
+
+新增 [023 迁移](../resources/sql/023_operational_task_progress.sql) 仅适用于 `tjwater_v2`，
+增加进展表和两个事件 topic 的 CHECK 取值。2026-10-10 已在目标库完成全库备份、
+迁移事务回滚演练及正式应用，其他项目和旧任务未改动。
+
+实际 worker 曾成功投递指定调查，模型发布进展和报告，但因 service 账号的任务权限过滤
+导致 `approved_data_scopes` 为空，误作数据不足结束。这不是业务数据确实不足。
+用户确认开放规则后，任务 API 已移除该额外过滤，沿用账号已有的项目读取权限。
+在线上下文返回七类来源，同一身份实际读取 SCADA 设备和外围合成工单成功；没有新增
+角色权限或放宽跨项目、审批和设备控制边界。旧委托路径保留兼容，已有权限无需委托凭据。
+用户级 `tjwater-operational-worker.service` 已重新启用。旧报告保留，未自动重新调查，
+不能把旧报告视为修复后的业务验收。任务创建时间下界和并发配置见
+[执行 worker](operational-task-workers.md)，未回填旧任务。
+进展读接口在表尚未部署时返回空，写入口须在迁移与 Server 更新后启用。
+按现有数据库发布流程应用迁移，再更新 Server、Agent 和前端，运行中的 worker 需要加载更新。
+
+聚焦检查：
+
+```bash
+conda run -n server python -m pytest tests/unit/test_operational_tasks.py tests/unit/test_operational_presentation.py tests/unit/test_peripheral_task_sync.py tests/unit/test_peripheral_push_sync.py tests/api/test_operational_task_dynamic_flow.py tests/api/test_openapi_contract.py -q
+git diff --check
+```
+
+这些检查覆盖隔离 HTTP 与状态机，不代替共享数据库迁移、真实模型运行或现场复测验收。

docs/investigation-report-contract.md

@@ -0,0 +1,166 @@
+# 调查报告数据契约
+
+本契约适用于供水事件调查，由 Server 的 `InvestigationReportInput`（写入）和
+`InvestigationReport`（读取）定义。Agent 和前端采用同样的字段和约束。
+不增加格式版本标签；原字段的类型保持稳定。`revision` 是报告修订号，
+`execution_generation` 是调查轮次，两者均不用于选择格式解析器。
+
+## 数据模型
+
+| 字段 | 类型 | 新写入规则 |
+| --- | --- | --- |
+| title | 非空字符串 | 报告标题 |
+| current_conclusion | 非空字符串 | 当前判断文字；无结论时明确写无法判断，不编造原因 |
+| conclusion_boundary | 非空字符串 | 判断范围及限制 |
+| evidence_refs | 字符串数组 | 引用该任务中已持久化、不可变的证据编号 |
+| recommendations | 字符串数组 | 保持原类型，处置建议不等于已执行动作 |
+| pending_items | 字符串数组 | 保持原类型，不从文字推断状态 |
+| stage | interim / final | 必填；阶段性报告不能用于结束运行 |
+| outcome | concluded / insufficient_data | 必填；业务调查结果，与运行成功/失败独立 |
+| stop_reason | 非空字符串或 null | 证据不足时必填；已形成结论时必须为 null |
+| data_gaps | 缺口对象数组 | 必填；空数组表示已记录且没有缺口 |
+| diagram | 调查脉络对象或 null | 可选；给人阅读的 Mermaid 说明，不控制执行或任务状态 |
+
+缺口对象除 node_id 外的字段必填：
+
+| 字段 | 类型 | 含义 |
+| --- | --- | --- |
+| gap_id | 非空字符串 | 报告内唯一的缺口编号 |
+| node_id | 非空字符串或 null | 可选的展示关联，不要求有执行计划 |
+| missing_data | 非空字符串 | 缺失数据内容 |
+| scope | 非空字符串 | 查询对象、地点及时间范围 |
+| impact | 非空字符串 | 因此无法回答的调查问题 |
+| required_data | 非空字符串 | 新一轮调查需要补齐的具体数据 |
+| evidence_refs | 至少一个证据编号 | 实际查询或可信输入记录；必须也出现在报告 evidence_refs 中 |
+| blocking | boolean | 是否阻止形成结论 |
+
+禁止未知字段、未知枚举和重复 gap_id。各引用在 Server 中校验。
+查询失败、权限配置异常不能作为“没有数据”的证明。
+
+调查脉络对象包含 `title`、`mermaid` 和可选的 `node_refs`（默认空数组）。
+`mermaid` 最长 20,000 字符，必须以 `flowchart` 或 `graph` 及方向开头，
+不包含 Markdown 围栏、初始化指令和 click 命令。前端使用 strict 安全级别渲染 SVG。
+完整图语法的解析失败只影响图示，正文、依据和审核仍可阅读。
+
+每条 node_ref 包含 `node_id`、`label`、`evidence_refs`、`action_refs`；后两项默认空数组。
+node_id 为最多 80 字符的 ASCII 标识符，必须以字母或下划线开头，在本图引用内唯一；
+最多 100 条引用。证据必须来自报告 evidence_refs，动作必须属于当前任务。
+节点可以说明候选方向与条件关系；图中的分支、状态文字不能证明实际并发、动作执行或现场完成。
+图的颜色、字体和交互由前端统一配置。正文不包含 SVG 或 HTML。
+
+## 业务约束
+
+1. `concluded` 允许 interim/final，stop_reason 为 null，不允许 blocking 缺口。
+2. `insufficient_data` 只能是 final，至少一个 blocking 缺口，必须提供 stop_reason；
+   recommendations 和 pending_items 必须为空，不发起补数据的工单动作。
+3. 报告必须来自当前执行轮次、会话和有效租约；执行计划与图示均可缺省。
+   历史计划或未结束的解释节点不阻止发布与正常收尾。运行状态来自 Server run，
+   审核和业务结果来自真实动作与回执，图示不参与计算。
+4. 尚有待审核、已批准未提交、已提交或结果未知的动作时，不能发布数据不足退出报告。
+   不自动撤销已有动作，不影响独立现场处置。
+5. 数据不足报告保存后，本执行代不再接受新的报告。补齐数据应使用已有 retry 接口
+   创建新 execution_generation，原报告、证据和回执保持可追溯。
+6. 事件调查提议动作前必须已保存本轮 outcome=concluded 的报告。
+   顺序为阶段性结论、动作提议、最终报告、运行收尾。图示可在报告中关联动作编号。
+7. 数据不足时顺序为查询证据、最终报告、运行收尾。
+   运行正常结束为 succeeded，调查生命周期为 completed；页面显示“已结束·证据不足”，
+   不显示正常水质、不自动关闭原投诉工单，不等待人工授权或补数据。
+8. 数据读取范围由系统配置；本场景的人工审核只针对具体业务动作。
+   既有数据访问审核接口仅保留兼容其他调用，不作为事件调查的产品入口。
+
+## 历史读取与兼容
+
+旧报告缺少 stage/outcome/stop_reason/data_gaps 时，Server 统一返回 null。
+null 表示未记录，不能转换成 concluded、空缺口清单或正常水质。
+前端兼容滚动发布期间的缺省/null，未知字段仍严格拒绝；一旦提供 outcome，
+其余结构化字段和业务约束必须完整满足。新写入不接受全空的历史格式。
+
+历史导入脚本的 dry-run 继续校验原始记录；缺少结构化结果的历史报告不能直接发布为新调查报告，实写在创建任务前拒绝，避免按旧正文编造结果。
+
+无需改写历史报告或新增数据库表，新增字段保存在原报告 document 中。
+上线顺序：先部署能读取可选新字段的前端，再协调更新 Server 新写入契约与 Agent。
+在切换写入端前应结束旧 Agent 在途运行，旧写入格式会被明确拒绝，不能静默推断结果。
+回退前须保留支持新字段的读取端；不能直接回退到会拒绝新字段的旧前端。
+
+## 接口与错误
+
+所有路径均以 `/api/v1` 开头，保留现有 Bearer 身份与项目范围校验。
+
+| 方法与路径 | 用途 |
+| --- | --- |
+| GET /operational-tasks/{task_id} | 读取统一的报告模型，旧结构化字段返回 null |
+| GET /operational-tasks/{task_id}/reports?limit=20&offset=0 | 按修订号倒序读取历史，返回 items 与 next_offset |
+| GET /operational-tasks/{task_id}/reports/{revision} | 读取指定报告快照，任务及项目范围校验 |
+| GET /operational-tasks/{task_id}/agent-context?report_revision=1 | selected_report 是所选报告，task 保留当前真实状态；未传则 selected_report 为 null |
+| POST /operational-tasks/{task_id}/agent-artifacts | 写入报告和计划；保留 request_id 幂等、run_id 与会话归属校验 |
+| POST /operational-tasks/{task_id}/runs/{run_id}/complete | 事件调查校验本轮最终报告后结束运行；检查 expected_revision |
+| POST /operational-tasks/{task_id}/actions | 仅本轮已保存 concluded 报告的事件调查允许提议动作 |
+| POST /operational-tasks/{task_id}/retry | 创建新调查轮次，不覆盖上一轮报告 |
+
+| 错误码 | HTTP | 含义 |
+| --- | --- | --- |
+| 请求体校验错误 | 422 | 新报告缺字段、未知枚举、无关键缺口或相互矛盾的字段 |
+| REPORT_GAP_EVIDENCE_INVALID | 422 | 缺口依据不在报告的证据集合内 |
+| REPORT_NOT_FOUND | 404 | 指定任务中的报告修订不存在 |
+| REPORT_ACTION_PENDING | 409 | 缺数退出仍有关联的未决动作 |
+| INVESTIGATION_ENDED | 409 | 本轮已经发布缺数退出报告，再写成果需要新轮次 |
+| ACTION_CONCLUSION_REQUIRED | 409 | 没有本轮已形成结论的报告，不得生成动作 |
+
+complete 接口保留已有返回方式：不完整成果会把 run 标为 failed，
+failure_code 为 ARTIFACT_INCOMPLETE；调用方须读取返回的业务状态，不能只看 HTTP 200。
+
+## 缺数报告示例
+
+以下为 agent-artifacts 请求中 report 字段的完整示例。
+`sample-query-empty` 必须是同一任务内已持久化的空查询证据；示例本身不构成真实查询事实。
+
+```json
+{
+  "title": "水质投诉调查报告",
+  "current_conclusion": "缺少投诉时段的末梢采样，无法确定异常来源。",
+  "conclusion_boundary": "已核对可用记录，本轮未确认水质异常原因。",
+  "evidence_refs": ["sample-query-empty"],
+  "recommendations": [],
+  "pending_items": [],
+  "stage": "final",
+  "outcome": "insufficient_data",
+  "stop_reason": "关键采样数据缺失，本轮调查结束。",
+  "data_gaps": [{
+    "gap_id": "terminal-sample-gap",
+    "node_id": null,
+    "missing_data": "投诉时段末梢水样检测结果",
+    "scope": "投诉小区，2026-09-24 08:00 至 09:00（北京时间）",
+    "impact": "无法判断异常来自供水管网还是用户端。",
+    "required_data": "该地点和时段的采样检测记录",
+    "evidence_refs": ["sample-query-empty"],
+    "blocking": true
+  }]
+}
+```
+
+## 验收
+
+- 新写入缺字段、不一致结果、无证据的关键缺口、未知业务证据或动作引用均被拒绝。
+- 关键数据不足保存报告并正常结束；无运行失败、无待人工通知、无后续动作。
+- 阶段性报告不能冒充最终报告；旧代结论不能授权新代动作。
+- 已有未决动作阻止矛盾的缺数退出；重试幂等，不重复生成报告。
+- 新轮次保留旧报告历史标识，不能显示成当前轮次已结束。
+- 桌面及 320/375px 展示结果、缺口、证据，审批区不出现数据授权入口。
+
+## 发布、阅读与协作
+
+执行中尚未形成结论的正文和 Mermaid 使用独立的 `progress` 快照持续更新，
+见 [调查进展与实时展示](investigation-progress-and-presentation.md)。
+报告的 interim/final 与 concluded/insufficient_data 结果契约保持不变。
+
+每次发布保留 `operational.report.published` 同步事件。首次 final、阶段报告转 final、
+新轮次 final 或判断/边界/依据/建议/缺口实质变化时，生成 `report_updated` 通知，
+其中 `report_reference = {execution_generation, revision}`。纯图示修订不生成额外成果提醒。
+新提醒取代同任务之前的成果提醒；读取相应报告后按用户确认已读。
+前端在调查列表显示“新报告”，这类提醒不进入“待办与告警”。
+
+正在阅读的报告和图示固定在所选修订，收到更新后显示“查看新报告”，不自动替换正文。
+当前待审核动作单独读取最新真实状态，即使用户正在阅读旧报告也不会审批旧版本。
+协作请求 `task_context` 可携带 report_revision、evidence_id、gap_id；轮次和报告修订号由
+Agent 服务根据 Server 返回的快照生成 `data-task_context`，随回答流和历史消息持久化。
+该引用说明回答依据，不授予访问权限，也不成为执行约束。

docs/mcp-analysis-tools.md

@@ -0,0 +1,71 @@
+# Agent 业务分析 MCP
+
+Agent 原生工具统一使用 `mcp_tjwater_` 前缀。Server目录共35项：25项查询、下表10项分析；合并6项宿主查询后共41项。工具参数来自Server实际MCP schema，不维护第二套参数契约。
+
+| 工具名（不含前缀） | 能力与边界 | 必要权限 |
+| --- | --- | --- |
+| `run_simulation` | 在线模拟：按水力步长直接计算业务库，不复制数据库；更新模型条件和实时基线，返回指定节点/管段，不创建方案或正式分析运行 | `simulation.run` + `scada.view` |
+| `run_burst_detection` | 指定时刻的15天压力异常侦测，至少5个完整监测节点；不定位 | `burst.run` + `scada.view` |
+| `run_burst_location` | 异常/正常压力、可选流量及漏水量假设的候选管段排序 | `burst.run` + 对应数据的 `scada.view` 或 `simulation.view` |
+| `analyze_valve_isolation` | 拓扑隔离边界、候选关阀及影响节点；不是水力计算或真实关阀 | `simulation.run` |
+| `run_leakage_identification` | SCADA压力约束下的DMA漏损分配估计；需总漏损量依据 | `burst.run` + `scada.view` |
+| `run_contaminant_simulation` | 已假设污染源的正向浓度传播；不是源头反演 | `simulation.run` + `scada.view` |
+| `analyze_valve_closure` | 给定关阀组合的水力影响 | `simulation.run` + `scada.view` |
+| `analyze_flushing` | 指定排放节点和流量的冲洗水力影响；不证明消毒合格 | `simulation.run` + `scada.view` |
+| `analyze_pressure_regulation` | 给定调泵序列、水箱初始水位的压力影响；不自动优化调度 | `simulation.run` + `scada.view` |
+| `optimize_sensor_placement` | 灵敏度或K-means压力监测点建议；不保存正式布点方案 | `optimization.run`，项目成员角色 |
+
+## 执行与证据
+
+标准水力模拟示例（只展示参数，不执行）：
+
+```json
+{"start_time":"2026-10-08T12:00:00+08:00","duration":3600,"output_node_ids":["<已核实节点ID>"],"output_link_ids":["540672"]}
+```
+
+工具完整 ID 为 `mcp_tjwater_run_simulation`。开始时间必须带时区，时长为60至86400秒，
+默认900秒；至少指定一个节点，节点和管段各最多20个。项目与身份由宿主绑定。
+输出沿用场景结果信封，包含 `node_results`、`link_results`、单位、时间和覆盖范围。
+它复用 REST `/api/v1/simulation-runs` 的在线计算服务：REST 时长仍为分钟，MCP 时长为秒。
+每个水力步长执行一次 `realtime` 快照，更新在线模型条件和该时刻的实时基线；不复制模型数据库。
+结果中的 `times.sample_times` 对齐采样点，窗口不包含结束时刻；例如一步长15分钟、时长3600秒时返回4个采样点。
+`persisted_baseline=true` 表示在线基线已重算，`persisted_scheme=false` 表示没有创建方案。
+该工具的 MCP annotation 为 `readOnlyHint=false`、`destructiveHint=true`，因为重算会替换对应时刻的基线。
+它不会创建 `analysis.run_id`，不能将本次计算冒充已有爆管事件运行；调查调用仍可保存证据。
+计算读取现有模型和对应SCADA条件，不从摘要虚构观测，也不自动补齐数据或执行现场操作。
+
+爆管定位 `run_burst_location` 默认使用 `data_source="monitoring"`。`pressure_scada_ids` 可省略，由后端从当前项目SCADA配置自动获取压力测点对应的模型节点；显式指定时仍限制为1至20个不同节点。正常对照时间窗 `scada_normal_start` / `scada_normal_end` 可成对省略，沿用后端默认的前一天同时间段；返回结果中的 `pressure_scada_ids` 和 `scada_window` 说明实际测点和时间窗。默认对照不等于已确认该时段正常。正常窗无有效压力数据或两窗没有共同有效测点时计算失败，不补造数据、不自动换窗。正常窗只有部分测点缺数时，沿用服务逻辑使用两窗共同有效测点，结合返回测点范围解释结果。
+
+仅提供异常时段和有依据的漏水量即可使用上述默认值，例如：
+
+```json
+{"scada_burst_start":"2026-10-08T12:00:00+08:00","scada_burst_end":"2026-10-08T13:00:00+08:00","burst_leakage":0.0274}
+```
+
+`data_source="simulation"` 仍须提供已有 `simulation_run_id`，使用同窗模型基线作正常对照，不填写正常时间窗；在线 `run_simulation` 不创建该运行UUID，不能替代已有分析运行。
+
+- 身份、项目路由、任务和运行由宿主绑定；模型不能提供network、username、凭据、文件路径、原始观测数组或方案保存名称。
+- 依赖时序数据的8项计算通过权限校验后，先从项目元数据解析并激活 BizDB 与 TimescaleDB 完整路由，再启动计算线程；时序库未配置时在修改模型前失败。普通业务查询、纯拓扑隔离及监测点选址不强制要求时序库路由。
+- 普通项目会话、任务协作和调查主Agent可调用；个人会话、学习复盘和只读reader不能调用分析。Server每次同时校验当前权限与绑定时的权限上限；任务数据读取scope不能替代计算权限。管理员显式deny继续生效。
+- 复用现有业务函数。在线模拟直接调用业务库；定位关闭共享候选CSV输出；识别不保存方案；选址使用不持久化模式。其余四项场景模拟复用项目专用`<业务库>_template`及唯一临时模型库，校验模板同步就绪，沿用临时库并发限制和清理；不是复制在线业务库。模板复制仍会短暂停止模板连接/订阅工作进程，不能称为完全没有基础设施副作用。
+- 场景`persist_results=False`仅允许隔离临时模型，跳过正式analysis运行、Timescale结果和基线写入；既有REST默认保存行为不变。仅返回请求的最多20个节点及20个管段，包含单位、时间和缺失对象；不能用局部结果断言全网无影响。
+- 数据源和计算假设必须有依据。定位漏水量和DMA总漏损约束使用m³/s；冲洗使用服务实际接收的m³/h，不沿用旧REST误标的L/s。工频泵序列0/1，变频泵序列为Hz；每项对应模型水力步长。冲洗当前限LPS/CMH模型。选址当前推荐压力点，不冒充污染溯源优化。
+- 调查结果经现有快照/证据/回执链保存，只有返回evidence_id后才引用为系统证据；普通聊天不自动保存调查证据。模拟结果始终是analysis而非实测。已知模拟事件源管段是输入背景，不是算法定位准确性的证明。
+
+## 资源边界与发布
+
+10项计算工具均提供 `timeout_seconds`，由Agent根据模型规模、模拟步数及已有耗时按次选择；它表示实际等待预算，与 `duration` 表示的模拟时间不同。未填写默认120秒。服务端上限由 `TJWATER_MCP_MAX_CALCULATION_TIMEOUT_SECONDS` 控制，默认3600秒，管理员可设120至86400秒；`tools/list` schema 发布实际有效上限，超限在计算前拒绝。
+
+例如 `duration=3600, timeout_seconds=900` 表示模拟1小时、允许等待15分钟。Agent的上游HTTP/SDK预算为所选预算加5秒，整个宿主调用另留60秒处理认证、上下文绑定和清理；OpenCode连接上限从工具目录计算并写入宿主连接文件。本地MCP请求关闭Bun空闲断连，仍受上述调用截止时间控制。普通查询继续使用原30秒Server预算与35秒宿主预算。
+
+计算在线程执行，不阻塞事件循环，单Server进程最多两项计算。超时不能杀死底层原生计算，槽位一直保留到计算真正退出；超时结果不会补写为调查成功结果，不自动反复重试。本次未增加后台计算作业队列。普通聊天大结果仍受现有12KB内联限制，调查结果可经证据分页读取。
+
+在线重算逐个时刻写入基线，超时或中途失败可能已有部分时刻更新，不能把工具失败解释为整个窗口已回滚；不要自动重试整个窗口。
+
+无需新增数据库迁移或模型配置。须同步发布Server、Agent和前端证据枚举；先更新Server目录，再重启Agent，client模式还需重载原生连接。保持既有OAuth、MCP开关及数据库路由配置，不自动更改权限或重启调查。
+
+本次验证限schema、权限、协议目录和mock业务函数，不运行真实模拟、优化或付费模型。真实项目的模板同步状态、运算耗时、SCADA覆盖和结果准确性仍需另行验收。
+
+2026-10-19 修复：EPANET 2 导出省略阀门默认 `ACTIVE` 状态行并保留数值设置，EPANET 3 导出保留原表达；覆盖文件和文本导出。MCP 在线模拟改为共用在线计算服务，方案模拟继续隔离。回归验证包含原生 EPANET 输入校验及 mock 在线基线写入，不重算真实项目历史数据。
+
+标准水力模拟接入已通过后端schema/权限/隔离适配测试、Agent目录与权限测试、前端证据解析测试；Agent和前端类型检查、前端改动文件ESLint通过。隔离HTTP smoke使用真实两端SDK发现34项Server工具及40项合并工具，并核对 `run_simulation` 输入schema和上下文绑定/查询；身份服务及数据执行为fixture，不等同生产验收。测试启动器剔除宿主MCP环境配置，避免私有issuer覆盖测试的loopback设置。运行中的服务仍需同步重载后确认新目录，未启动真实模拟。

docs/mcp-keycloak-setup.md

@@ -0,0 +1,88 @@
+# MCP 接入的 Keycloak 管理员配置清单
+
+这里创建的是 OAuth 认证配置。OpenCode 到 Agent、Agent 到 Server 的 MCP 协议客户端已实现，无需再开发或安装 MCP 客户端程序。
+
+以下使用已创建的请求方客户端 `agent-mcp`，MCP 资源的 audience 使用 `tjwater-mcp`。它们位于现有业务登录使用的同一个 realm。
+
+## 1. 请求方客户端 agent-mcp
+
+在 Keycloak 的 Clients 中打开 `agent-mcp`：
+
+| 设置 | 值 |
+| --- | --- |
+| Client type / Protocol | OpenID Connect |
+| Client authentication | On；这是 confidential 客户端 |
+| Service accounts roles | On；用于启动时读取 MCP 工具目录 |
+| Standard token exchange | On；用于按当前用户或 worker 的身份换取 MCP 令牌 |
+| Standard flow | Off；该客户端不负责浏览器登录 |
+| Direct access grants | Off |
+| Access token lifespan | 建议300秒，不超过900秒 |
+| 签名算法 | RS256 |
+
+保存后，在 Credentials 中取得 Client Secret。密钥只写入 Agent 的忽略文件 `.local.env`：
+
+```dotenv
+TJWATER_MCP_CLIENT_ID='agent-mcp'
+TJWATER_MCP_CLIENT_SECRET='<Credentials 中的 Client Secret>'
+```
+
+不要把密钥放入聊天、前端配置、OpenCode 配置或 Git。Server 只配置 `TJWATER_MCP_HOST_CLIENT_ID='agent-mcp'`，不需要保存这个客户端的密钥。
+
+## 2. 资源客户端及 scopes
+
+创建或复用表示 MCP 服务的 OIDC 客户端 `tjwater-mcp`；关闭浏览器登录、直接密码登录和服务账户。这个 ID 是目标 audience，不是另一个需要部署的服务。
+
+创建以下 OIDC Client scopes，并作为 Optional scopes 分配给 `agent-mcp`；启用 Include in token scope：
+
+- `mcp.catalog.read`
+- `mcp.query`
+- `mcp.context.bind`
+
+在这些 scopes 中添加 Audience mapper，使对应 token 的 `aud` 包含目标客户端 `tjwater-mcp`。应用按实际 grant 请求 scope：服务账户请求只有 `mcp.catalog.read`；用户/worker 的交换请求才包含查询和上下文 scope。不要给服务账户额外分配项目成员、管理或设备控制权限。
+
+## 3. 允许当前业务身份交换
+
+在原有前端登录客户端和 worker 客户端的专用 scope 中添加 Audience mapper，使源 token 的 `aud` 包含请求方 `agent-mcp`。保留原业务 API 所需的 audience；不要给源 token 添加 MCP 目标 `tjwater-mcp`。
+
+交换后必须保持用户主体 `sub` 不变，`azp=agent-mcp`，目标 `aud` 包含 `tjwater-mcp`。Keycloak 的 `audience` 参数只筛选已有映射的 audience，不能代替 Audience mapper。
+
+版本和界面名称以当前 Keycloak 为准。配置依据：[Keycloak Standard Token Exchange](https://www.keycloak.org/securing-apps/token-exchange)。仅在公开发现中看到 token exchange，并不能证明具体客户端已启用或映射正确。
+
+## 4. 地址、数据库和启用顺序
+
+正式环境提供该 realm 的 HTTPS issuer。本地开发可显式允许一个 RFC1918 IPv4 HTTP issuer：Agent 设置 `NODE_ENV=development`，Server 设置 `ENVIRONMENT=development`，两端 `TJWATER_MCP_DEV_HTTP_ISSUER_URI` 均填写与各自 `TJWATER_MCP_ISSUER_URI` 完全相同的地址（含 realm）。此例外不放行远程 HTTP MCP resource；Agent 只允许该 issuer 同源的远程 HTTP 发现及令牌端点，继续拒绝重定向。正式环境必须清空此项并使用 HTTPS 域名。配置值必须与发现文档和令牌 `iss` 一致，不能仅替换访问地址。
+
+例如，同机开发时 Server `.env` 可配置为（内网地址按实际认证服务替换）：
+
+```dotenv
+ENVIRONMENT=development
+TJWATER_MCP_ENABLED=true
+TJWATER_MCP_ISSUER_URI=http://192.168.1.114:8088/realms/tjwater
+TJWATER_MCP_DEV_HTTP_ISSUER_URI=http://192.168.1.114:8088/realms/tjwater
+TJWATER_MCP_RESOURCE_URI=http://127.0.0.1:8000/mcp
+```
+
+Agent `.local.env` 使用相同的三个 URI，并设置 `NODE_ENV=development`、
+`TJWATER_MCP_ENABLED=true`。开发模式不跳过签名、issuer、audience、scope 或项目权限校验，
+也不自动接受另一种 issuer。切换 Server 环境还会开启开发 API 文档页面。
+修改环境文件后需重启相应进程；已有进程不一定自动重读 `.env`，系统环境变量也可能覆盖文件。
+应先确认 Server 的 `/.well-known/oauth-protected-resource/mcp` 公布的新 issuer，
+再启动 Agent 检查工具目录。切回生产时，两端恢复相同 HTTPS issuer 并删除 HTTP 白名单。
+
+业务身份也必须使用同一 issuer：供水前端只在被 Git 忽略的
+`.env.development.local` 中将 `TJWATER_KEYCLOAK_ISSUER` 设为该内网地址，
+保留 `.env.local` 的 HTTPS 配置供正常 production 构建使用。
+Server 本机 `.env` 可额外设置 `OPERATIONAL_WORKER_DEV_TOKEN_URL` 为同一内网 issuer
+发现文档中的 `token_endpoint`；该值仅在 `ENVIRONMENT=development` 时生效，
+不替换原来的 `OPERATIONAL_WORKER_TOKEN_URL`。重启开发前端和 worker，并在浏览器重新登录。
+旧 HTTPS 令牌不能用于内网 issuer 的交换；工具目录读取成功不等于业务身份交换成功。
+
+接入人员随后执行：
+
+1. 安装 MCP SDK 2 对应的 Server 依赖。
+2. 对选定项目业务库备份并应用 [025_mcp_query_contexts.sql](../resources/sql/025_mcp_query_contexts.sql)。
+3. 在 Agent 和 Server 的私有环境中配置相同的 issuer、resource、audience；resource 必须指向实际 Server `/mcp`，请求方客户端 ID 均使用 `agent-mcp`。
+4. 验证带标准 `resource` 的目录 grant 和用户/worker token exchange 都成功；若身份平台拒绝，先解决兼容性，不移除参数重试。
+5. 在受控测试实例中开启两端 `TJWATER_MCP_ENABLED` 并重载；通过测试项目查询、任务证据和越权拒绝验收后，再按部署流程启用正式环境。
+
+完整环境变量与验收要求见 [MCP 实施文档](mcp-query-implementation.md)。在以上步骤完成前保持 MCP 开关关闭。

docs/mcp-query-design.md

@@ -0,0 +1,280 @@
+# NextTJWater MCP 查询与认证设计
+
+日期：2026-10-15。状态：查询与认证代码已在本地实现，认证发现、resource 参数和版本头体校验已补齐，默认关闭，未部署；真实 Keycloak 与项目数据库验收待完成。
+
+本设计将现有后端查询能力通过 MCP 开放给普通项目 Agent、调查主 Agent 和调查只读子 Agent。采用 MCP 2026-07-28、Streamable HTTP、现有 Python Server 内挂载服务、可信 Agent 宿主按请求代理认证。旧 CLI 用于查找功能和接口线索，业务实现继续由当前 Server 负责。
+
+第一版完成查询、权限、取消、分页、调查证据和凭据续期的完整闭环。已有报告发布、工单审批、设备控制、调查运行生命周期继续使用当前受控工具；不通过 MCP 自动开放旧 CLI 中的分析执行命令。
+
+## 当前系统与功能核对
+
+Agent 的独立只读查询现统一通过原生 `mcp_tjwater` 连接调用。重复的项目查询插件与自定义 HTTP 查询桥已移除；`memory_manager`、`skill_manager` 保留完整读写插件，发布、动作、消息发送和界面展示也保留插件形式。普通会话凭据保存在宿主的 `src/product/credential-refresh.ts`；调查由 worker 以服务身份投递并续期。调查子 Agent 通过父任务绑定访问数据，不持有独立产品凭据。
+
+实施前从 `app.main.app.openapi()` 核对了 266 路径的契约，其路径集合及各 HTTP operation 与工作区 `contracts/server-v1.openapi.json` 一致。公开 API 经 `app/api/v1/rest_router.py` 适配，项目字段注入和分页须看最终注册契约，不能只看底层 handler 签名。旧 CLI 的 41 条业务命令中，38 条能匹配当前 HTTP 方法与路径：29 条读命令、9 条执行命令。另有 3 条需要替换或不开放，详见附表。路径匹配不代表参数兼容、数据可用或算法已完成业务验收。
+
+实施前 `requirements.txt` 和 conda 环境均为 `mcp==1.9.4`、`fastmcp==2.9.2`，当时没有 Server MCP 注册实现。Agent 使用的 OpenCode 2.0.27 配置类型支持 `legacy`、`auto`、`2026-07-28`，宿主认证代理现已补充，见实施说明。
+
+## 协议选择
+
+官方 `specification/latest` 指向 [MCP 2026-07-28](https://modelcontextprotocol.io/specification/2026-07-28)。新实现按该正式版本建设；产品内部连接固定要求这个版本，版本不支持时返回明确错误。第一版不提供旧 `initialize` 握手兼容层。
+
+| 协议要求 | 本系统设计 |
+| --- | --- |
+| 无状态请求与逐请求能力声明 | 每次请求发送标准 `_meta` 中的版本、客户端信息和能力；不以连接识别用户或调查轮次 |
+| `server/discover` | 返回版本、服务信息和实际支持能力；不能用自报服务名建立信任 |
+| HTTP 请求镜像头 | 由 SDK发送并校验 `MCP-Protocol-Version`、`Mcp-Method`、适用时的 `Mcp-Name`；头与正文不一致拒绝 |
+| 普通结果 | 返回 `resultType=complete`、`structuredContent`、匹配的 `outputSchema`，以及简短文字说明 |
+| 列表和资源缓存提示 | 返回 `ttlMs`、`cacheScope`；缓存不作为权限证明 |
+| 取消与断线 | 取消传到 SDK、数据库和外部请求；断开的流不按旧 SSE 游标恢复 |
+
+这些改变来自[正式版变更说明](https://modelcontextprotocol.io/specification/2026-07-28/changelog)和 [Streamable HTTP 规范](https://modelcontextprotocol.io/specification/2026-07-28/basic/transports/streamable-http)。协议支持情况以实际发现与调用验收为准，不以软件版本字符串代替。
+
+第一版声明 tools 和 resources；不声明 prompts、subscriptions、MRTR、Tasks、MCP Apps、Sampling、Roots 或 Logging 能力。查询不要求模型补填身份、密码或令牌。现有产品调查任务不会替换为 MCP Tasks。
+
+Schema 使用 JSON Schema 2020-12，模型输入对象拒绝未声明字段；只使用服务自带的本地定义，不在校验时从任意外部 URL 获取 `$ref`。宿主校验模型投影，Server 再校验完整线协议参数。工具发现、结构化结果、状态句柄和执行错误遵循[正式工具规范](https://modelcontextprotocol.io/specification/2026-07-28/server/tools)。
+
+| 失败类型 | 响应与宿主处理 |
+| --- | --- |
+| 未认证或过期 | HTTP 401 和 Bearer challenge；宿主最多续期一次 |
+| scope 或产品权限不足 | HTTP 403；scope 不足时带 insufficient_scope，不自动扩大产品权限 |
+| 请求头与正文不一致 | HTTP 400、JSON-RPC -32020 |
+| 协议版本不支持 | HTTP 400、JSON-RPC -32022，列出支持版本 |
+| RPC 方法不存在 | HTTP 404、JSON-RPC -32601 |
+| 工具不存在或 RPC 参数结构无效 | JSON-RPC -32602 |
+| 业务参数、数据源或上下文失败 | resultType=complete、isError=true；返回稳定业务错误码和可操作说明 |
+
+旧客户端 GET/DELETE 到 `/mcp` 返回 405；旧 Mcp-Session-Id 和 Last-Event-ID 不建立或恢复身份。产品上下文删除使用独立 REST 控制路径，不混用协议会话删除。
+
+`server/discover` 和固定查询目录按工具名稳定排序，缓存提示取 `ttlMs=300000`、`cacheScope=public`，正文仅包含通用定义。业务结果资源取 `ttlMs=0`、`cacheScope=private`；宿主按令牌、主体、项目和上下文分别保存缓存，不跨授权上下文复用。[缓存规范](https://modelcontextprotocol.io/specification/2026-07-28/server/utilities/caching)
+
+## 服务与调用结构
+
+在现有 FastAPI Server 挂载规范路径 `/mcp`，通过官方 Python SDK v2 承担协议处理。ASGI 的生命周期纳入现有应用启动和关闭；不得用根路径挂载遮蔽 REST 路由。官方 SDK 支持[在现有 ASGI 应用中挂载](https://py.sdk.modelcontextprotocol.io/run/asgi/)，其[协议版本说明](https://py.sdk.modelcontextprotocol.io/protocol-versions/)覆盖 2026-07-28。
+
+查询适配器复用现有 service、repository、项目路由和业务权限，不从 MCP 向自己的 REST API再发一次 HTTP 请求，不转发收到的 MCP Bearer。需要从 endpoint 提取逻辑时，提取供 REST 和 MCP 共用的 service，而不直接调用带 FastAPI Depends 默认值的 endpoint 函数。
+
+```text
+模型调用查询工具
+    → OpenCode 原生 MCP（tools/call，_meta 中关联原生 sessionID）
+    → Agent 本地 MCP 服务（内部密钥认证、宿主会话和直接父子关系检查）
+    → 宿主专用产品 MCP 客户端
+        → Keycloak（按主体获取面向 MCP 的令牌）
+        → Server /mcp（标准 MCP 请求 + 宿主绑定的上下文）
+            → 项目与任务权限检查
+            → 共用查询 service
+            → BizDB / TimescaleDB / 既有外围适配器
+            → 调查证据保存和有范围的数据结果
+```
+
+OpenCode 原生发现本地 MCP 的工具目录，工具 ID 为 `mcp_tjwater_<name>`。本地 MCP 默认提供地址解析、会话搜索、调查任务列表与详情、消息读取、待办读取六项查询；启用产品 MCP 后，宿主从真实 Server `tools/list` 发现24项项目查询，投影并合并到本地目录，共30项。公开网页搜索使用 OpenCode 原生 `websearch`，Server 不提供搜索接口。可信插件只配置连接和权限策略，不注册独立查询工具或自定义执行器。待办读取经宿主 RPC 访问原插件存储，保留已有清单。
+
+本地服务使用官方 TypeScript Server SDK 2.3.1、2026-07-28 Streamable HTTP，绑定127.0.0.1临时端口，拒绝浏览器来源和错误内部密钥。OpenCode 的 `_meta["ai.opencode/sessionID"]` 仅是关联值；通过内部通道认证后，服务再解析宿主身份或已核验的直接只读子会话。未知会话、个人项目查询和学习复盘均拒绝。URL、身份、项目、网络、产品认证头和上下文句柄不属于模型参数。产品 MCP 未启用时项目查询不可用，没有旧桥回退。取消信号贯穿整个调用链。[OpenCode MCP 会话上下文](https://dev.opencode.ai/v2/docs/mcp-servers/#context)
+
+现有 OpenCode 共享 MCP 配置缺少逐会话 OAuth 提供器，不能把一份用户 Bearer 填进 `mcp.servers.headers`。产品内部连接由 Agent 宿主按请求认证，外部 MCP 继续保持原配置路径和现有策略，二者不能误注册为同名工具或形成两份不同凭据的业务连接。
+
+## 认证与上下文绑定
+
+### OAuth 资源与现有登录
+
+Keycloak 继续作为授权服务器。MCP 是独立的受保护资源，资源标识由 `TJWATER_MCP_RESOURCE_URI` 指定为部署的规范 `/mcp` URI；令牌目标 audience 取专用客户端 `tjwater-mcp`。Agent 宿主使用预注册 confidential client `tjwater-agent-mcp`。用户继续使用当前前端标准流程和 S256 PKCE 登录，不新增产品内的第二次登录。
+
+OAuth scope 分成 `mcp.catalog.read`、`mcp.query` 和 `mcp.context.bind`。宿主启动时以自身 client_credentials 取得仅含 catalog scope 的令牌，发现真实目录；该身份不获得项目业务读取能力。会话查询令牌必须具有 query scope，绑定接口还要求 context.bind。scope 只开放接口类别，具体数据权限仍由产品主体、项目成员关系和工具 registry 求交。
+
+MCP 端提供 `/.well-known/oauth-protected-resource/mcp`，401 响应包含对应的 `WWW-Authenticate` 资源发现信息。验证签名、固定 issuer、audience、exp、iat、客户端来源和查询 scope；对新 MCP 使用独立验证器，不依赖当前 REST 验证器可为空的 audience，也不通过修改全局 audience 破坏现有 REST。
+
+标准 OAuth 客户端在授权和 token 请求中携带规范资源 URI。MCP 仅接收确实面向它发行的访问令牌，令牌每次通过 Authorization 头发送，不出现在 URL、工具参数或模型文本中。[MCP 授权规范](https://modelcontextprotocol.io/specification/2026-07-28/basic/authorization)、[授权服务器发现](https://modelcontextprotocol.io/specification/2026-07-28/basic/authorization/authorization-server-discovery)
+
+产品宿主先发送无凭据的标准发现请求，从401 challenge 读取资源元数据地址；无地址时按 MCP 路径、根路径依次发现 well-known。资源 URI 必须匹配已配置资源，authorization_servers 必须包含预注册凭据所属的固定 issuer。随后使用官方 SDK 按规范顺序发现 OAuth/OIDC 元数据，严格比较 issuer，从元数据取得 token endpoint；不向其他 issuer 发送原凭据。公开发现请求不带产品 Bearer 或 secret，单份文档最多64KiB，总预算10秒；拒绝重定向和远程明文地址。发现缓存最多5分钟，MCP 401 后失效重查。根据元数据使用 client_secret_post 或 client_secret_basic。
+
+产品内部宿主基于当前产品凭据使用 Keycloak Standard Token Exchange V2 换取面向 `tjwater-mcp` 的短期访问令牌；保留原主体，限定查询 scope，避免共享一个超级服务账号访问所有普通用户数据。专用令牌只保存在 Agent 主进程内存中，刷新令牌仍由现有登录机制持有。
+
+Keycloak V2 需要 confidential requester、启用 Standard token exchange、subject token 对 requester 的 audience，以及目标 audience/scope 配置。client_credentials 与 token exchange 均发送 `resource=规范MCP URI`；exchange 另保留产品要求的 audience。身份平台必须将该资源映射到专用 audience，并核验两种 grant 返回的令牌。若部署版本拒绝 resource，应先升级或提供兼容适配再启用；不能删除标准参数后重试。不增加第三方客户端授权码流程。[Keycloak Token Exchange](https://www.keycloak.org/securing-apps/token-exchange)
+
+上述 realm/client 能力尚未在线核验，是部署前提，由身份平台配置负责人落实。达不到前提时停止启用内部 MCP；不以普通 API Bearer 转发、共享管理员 token 或关闭 audience 校验降级。
+
+### 每次调用绑定项目与调查轮次
+
+OAuth 令牌只证明主体与可用 scope，业务范围由 Server 签发的 `context_handle` 约束。它是应用上下文，不是 MCP 协议会话，也不能单独作为认证凭据。
+
+可信 Agent 宿主通过受认证的 `POST /api/v1/mcp-query-contexts` 申请上下文。请求中的 session、purpose、项目和任务信息均由原生会话及宿主已有绑定读取；仅允许 `azp=tjwater-agent-mcp` 的宿主客户端创建任务上下文。Server 从令牌确定主体、从成员关系确定项目读取权限，验证 task、run、execution_request 和父会话占用的租约；原生父子关系由可信宿主先核验；不接受自报用户、权限或数据 scope。
+
+Server 保存的绑定字段包括：主体、client ID、project ID、native session ID、purpose、可用查询工具、任务及运行标识、父会话、创建与过期时间、撤销状态。句柄使用至少 256 位随机值，10 分钟过期。普通上下文到期由宿主重新申请；任务上下文有效性另外取决于当前轮次和 lease，不能用续租让旧轮次重新获得访问权。
+
+所有业务 MCP tool 的线协议 schema 具有宿主字段 `context_handle`。Agent 宿主客户端在发布给本地 MCP 的 schema 投影中移除它，并在请求 Server 时覆盖注入；它不使用 `x-mcp-header`，也不进入聊天 transcript 或普通日志。若输入包含同名字段、身份字段或保留 scope 字段，宿主直接拒绝。resources/read 由宿主在应用命名空间 `_meta` 中注入同一上下文，URI 仅标识任务证据而不携带句柄或令牌。外部客户端若将来接入，应自己调用绑定接口并在宿主侧选择项目，不从模型内容决定项目。
+
+Server 每次同时核验 Bearer 与句柄所属的 `sub/azp`，重新校验成员关系和当前权限；任务调用还检查项目、任务、run、execution request 和 lease。句柄复制到其他用户、客户端或项目无效。切换项目新建会话，关闭/删除会话撤销绑定；调用 `DELETE /api/v1/mcp-query-contexts/{context_handle}` 时检查绑定所有者。令牌在新 scope 下续期时不会自动扩大已有上下文的工具授权。
+
+普通会话的 `purpose=ordinary_chat`，仅访问其绑定项目。调查主 Agent 的 `purpose=task_execution`，使用 worker 的现有服务主体和本轮任务范围。只读子 Agent 的 `purpose=task_reader`，宿主根据已经验证的父子关系绑定父轮次；绑定只允许查询和系统证据追加，不授予报告发布、审批、设备控制、Shell 或再次委派权限。个人会话不创建项目查询上下文，返回现有 `PROJECT_CONTEXT_REQUIRED`。
+
+### 续期和退出
+
+普通项目会话复用现有 `credential_refresh_required` 回传与前端 Keycloak `updateToken`。收到更新的产品 access token 后，宿主重新 exchange 并继续原查询，用户只在现有登录本身失效时重新登录。并发续期按主体与原凭据版本合并，最终缓存按 issuer、resource、主体、scope 和原凭据版本隔离。
+
+后台调查不依赖浏览器。worker 保持当前 client_credentials 续期；新服务 Bearer 到达后由 Agent 更新原会话产品凭据，再 exchange。MCP 声明官方 `io.modelcontextprotocol/oauth-client-credentials` 授权扩展以支持服务身份；首次范围仍由服务账号成员关系与任务授权决定。[Client Credentials 扩展](https://modelcontextprotocol.io/extensions/auth/oauth-client-credentials)
+
+MCP 401 清除本主体缓存并续期一次；403 不反复刷新、不自动申请更大 scope。认证服务不可用时返回认证暂不可用，不能返回空数组掩盖故障。注销时撤销产品绑定、取消调用并清除宿主缓存。JWT exchange 的源 token 撤销不必然撤销已交换 token，因此 MCP 令牌短期有效，Server 的绑定撤销和每次业务权限检查仍然必须执行。
+
+## 查询工具目录
+
+以下为第一版明确目录。名称为 MCP 服务端名字，OpenCode 原生发现后添加 `mcp_tjwater_` 前缀。元素枚举固定为 junction、pipe、valve、pump、tank、reservoir；时间必须带时区；身份、项目和 network 都不在模型输入中。HTTP 路径表示现有能力来源，MCP 实现调用其共用 service。
+
+| 工具 | 业务输入 | 当前能力来源与限制 |
+| --- | --- | --- |
+| `get_element_properties` | element_type、element_id | 六类 `/{type}s/properties` |
+| `list_elements` | element_type、limit、offset | 六类 `/{type}s`；返回一页，不暗中获取全网 |
+| `query_nearby_elements` | center、可选 element_type、radius_m、limit、offset | 当前项目模型管网空间邻近查询；直接接收地理编码的 EPSG:4326 `[经度,纬度]`，默认pipe、200米、5条，最大10000米、单页100条；返回element_id、distance_m及properties，不查询原始GIS源要素 |
+| `get_network_options` | kind、可选 pump_id | network-options 及 time、energy、pump-energy；泵能源必须给泵 ID |
+| `get_network_schema` | kind | 固定 network-schemas 枚举，以当前注册路径为准 |
+| `list_scada_devices` | limit、offset | `/scada-devices`，显式分页 |
+| `get_scada_device` | device_id | `/scada-devices/{device_id}` |
+| `query_scada_readings` | device_ids、start、end、field | scada-readings/fields；区分 monitored_value 和 cleaned_value |
+| `query_model_baseline_series` | element_type、start、end、元素筛选 | realtime/links、realtime/nodes；标记为模型基线，不称为实测 |
+| `get_model_baseline_at_time` | element_type、element_id、time | realtime/simulation-results |
+| `query_model_baseline_field` | element_type、field、time | realtime/records；只允许当前字段枚举 |
+| `query_analysis_series` | run_id、element_type、element_id、field、start、end | analysis/runs/{run_id}/links 或 nodes/{id} |
+| `query_analysis_values` | run_id、element_type、field、time | analysis/runs/{run_id}/values |
+| `list_analysis_runs` | limit、offset | `/analysis/runs` |
+| `get_analysis_run` | run_id | `/analysis/runs/{run_id}` |
+| `list_analysis_results` | run_id、result_type、limit、offset | `/analysis/runs/{run_id}/results`；保留 result_type |
+| `list_sensor_placement_runs` | limit、offset | sensor-placement-runs；公开 REST 已包装分页，但底层 service 读取整数组，MCP 应推动共用 service 做真实有界查询 |
+| `get_sensor_placement_run` | run_id | sensor-placement-runs/{run_id}；REST 适配层已注入 network，MCP 共用 service 也从可信项目上下文解析 |
+| `query_element_history` | elements、start、end、mode、可选 run_id | element-history/query；observed、realtime_comparison、analysis_comparison |
+| `get_element_gis_properties` | element_type、element_id | elements/{type}/{id}/gis-properties |
+| `get_source_gis_properties` | layer_key、source_fid | source-gis/layers/{key}/features/{id}/properties |
+| `get_workbench_runtime` | 已有运行态查询参数 | workbench/runtime；只读取 |
+| `get_task_context` | section、data_path、limit、offset、可选 report_revision；无模型可选 task_id | 当前绑定任务的固定摘要及分章节、路径分页详情；预览不等于完整报告 |
+| `query_peripheral_data` | source、point_id、start、end、limit、offset | 固定外围 points、work_orders、observations、test_data 适配器；保留 synthetic |
+| `get_task_evidence` | evidence_id、data_path、limit、offset | 当前任务的 evidence/{id}/snapshot；项目、任务及可读轮次由绑定约束 |
+
+初版为24项查询；新增空间邻近查询后，当前为25项查询，连同10项分析共35项Server工具。旧 CLI 的 `data pipeline-health` 实际触发模型预测，不能包装成“读取已保存结果”；它作为 `predict_pipeline_health` 的后续计算候选，先核对模型资产、执行成本和计算权限。旧 CLI 的 9 条 POST 执行命令也仅列作后续候选，不加入查询目录。
+
+空间邻近查询沿用 `webgis.view` 和宿主绑定项目，数据库连接不从工具参数选择。`app/infra/db/postgresql/spatial_queries.py` 使用既有 `gis.*` 模型视图及GiST几何索引做保守候选筛选，再转换到WGS84 geography，按点到完整几何的椭球最短距离以米筛选和排序，最后在数据库中执行LIMIT/OFFSET，不把全网几何交给Agent。跨日期变更线时拆分候选包围框，极区跳过Mercator候选筛选；空或缺失几何不参与查询。距离语义依据 [PostGIS ST_DWithin](https://postgis.net/docs/ST_DWithin.html) 和 [ST_Distance](https://postgis.net/docs/ST_Distance.html)，不把EPSG:3857平面距离直接当作实际米数。
+
+输出 `data.source=model_network`、`center_crs=EPSG:4326`、`distance_method=WGS84_spheroid`，`data.items` 包含 `element_id`、`distance_m` 和 `properties`；`page.has_more` 与 `truncated` 保留分页边界。没有命中返回空items，不自动扩大半径；空间邻近不证明拓扑连接或供水归属，道路/河流交叉位置必须由调用方先确认。查询预算仍为30秒，调查调用沿用现有证据与回执保存，不新增REST查询桥或数据库迁移。
+
+聚焦验证：`python -m pytest tests/unit/test_mcp_spatial.py tests/unit/test_mcp_query.py -q`；设置 `MCP_SPATIAL_TEST_DSN` 后运行 `python -m pytest tests/integration/test_mcp_spatial_postgis.py -q` 可验证实际PostGIS距离与分页，测试只创建事务内TEMP表并回滚，不写业务数据。
+
+权限保持当前真实规则：管网、GIS 和分析运行读取沿用 webgis.view，SCADA 沿用 scada.view，模型基线和分析时序沿用 simulation.view；监测点选址路由当前整体要求 optimization.run，MCP 不擅自改成 optimization.view。历史比较模式额外核对其中使用的模拟数据权限。拆分读权限需要单独的后端授权设计，不能借 MCP 接入顺便降级。
+
+新源 registry 明确 source、所需权限、查询参数 Schema、数据语义、结果上限、任务 scope 和是否允许证据保存。原 7 个 TaskDataSource 与新源按明确规则扩展，历史证据仍可反序列化。既有已批准 scope 只迁移到与其原意等价的查询能力，不用“只读”作为放开所有源的理由。
+
+## 分页 大结果与数据语义
+
+集合查询默认 limit=20，最大 500；分析运行最大 100。offset 最大10000，超过时缩小时间窗或目标。常规集合在 SQL 中限制返回量。带总量的集合返回 items、total、limit、offset、has_more；没有稳定总量时 total=null，不伪造总数。Agent 首次优先单对象、必要的短时间窗和小页；只有明确问题仍缺证据且 has_more=true 时继续翻页，不自动遍历全网。
+
+历史组合查询按稳定的序列顺序（设备ID、原始/清洗，再按类型和元素ID排列模型序列）及时间对数据点分页；在同一条 UNION ALL SQL 上应用全局 limit/offset，并多读一个点计算 has_more。证据查询按绑定项目、任务和证据ID，在数据库中选取 data_path 和页；嵌套集合返回路径描述，数组保持索引顺序，对象按 C 排序，字符串按 Unicode 字符分页。证据页和完整资源在返回数据库客户端前检查8MiB上限。此限制约束传输和应用进程，不代表数据库无需读取或解压 JSONB。
+
+DEMO 样本仍需校验整份 RFC 8785 摘要，不能仅验证所选页。数据库先检查整份 JSON 文本大小，超过8MiB则拒绝传输；预算内读取完整样本，校验后筛选和分页。超限须由数据源拆分样本，不能通过反复缩小查询窗口消除该限制。
+
+时序默认最多 20 个目标、24 小时，最多返回 10000 条或 8 MiB。全网模型基线接口当前没有元素筛选和结果分页，接入前必须在共用 service 增加有界筛选/游标；MCP 不直接代理当前无限制接口。批量历史采用其现有 mode 和 run_id 语义，不假定新接口完全等价于旧三类 composite 查询。
+
+结果超过内联 12000 字节时返回摘要、数量、时间范围、truncated、继续查询参数或受保护 resource_link。第一版 resources 暴露已有任务证据快照，普通查询用分页继续读取；不新增全局 result-ref 存储或任意主机文件路径 API。资源读取同样核验令牌和上下文，不能凭 URI 得到数据。
+
+共用输出对象包含 source、source_ref、queried_at、origin、data、window、units、page、truncated、data_gaps，以及调查时的 evidence_id。原 service 返回结构由各工具的 Pydantic 输出模型明确描述，不靠自由文本或一个无约束的 data 字段作为最终契约。第一版每主体/项目最多 4 个并发查询；超过并发返回可重试限流结果，队列不无限增长。查询执行预算沿用 Server 已配置的访问超时，认证续期等待单独计时。
+
+普通原始查询错误返回结构化业务错误，不改成空数据。结果标明数据来源、查询时间、源时间窗口、单位、SCADA 实测或清洗、模型基线或分析运行、synthetic、缺失及截断。`truncated=true` 不能支持“完整范围未见异常”的结论。
+
+## 调查证据与幂等
+
+共享查询入口根据已验证 purpose 选择包装方式。普通聊天返回查询结果；调查及 reader 查询在返回前调用现有 `OperationalTaskService.record_data_snapshot`，保存范围与来源，再返回 snapshot 和 evidence_id。外部 MCP 返回内容不能通过这一入口被重新标成内部系统证据。
+
+原生 callID 经宿主转为 application request key。Server 按主体、项目、会话、任务、run、callID 和规范参数记录查询回执；同 key、同参数重试读取已保存 snapshot，同 key、不同参数拒绝。MCP 断线重发使用新 JSON-RPC ID，但保持同一 application request key；证据存储提交失败就返回失败，不能先返回成功再异步补证据。
+
+调查取数前后都校验本轮仍然有效。长查询期间不持有任务行锁；完成后用短事务确认轮次并保存证据。查询途中被新轮次替代时，不把晚到结果写进新轮次，也不复活旧运行。外部网络请求与数据库读写使用短事务，不跨网络等待持锁。
+
+查询业务源是只读，但调查包装会追加证据，因此相应工具不能仅根据 GET 方法声明 `readOnlyHint=true`。固定目录内可能保存证据的工具保守标为 `readOnlyHint=false`、`destructiveHint=false`，执行策略按已审查的精确工具名允许“读源并追加系统证据”，不由 annotation 决定授权。精确 grant 不覆盖报告发布、工单或设备动作。
+
+## 客户端策略与生命周期
+
+内部业务查询在项目会话、调查执行和已验证 task_reader 中开放对应精确工具。reader 的精确白名单仅增加可信 `mcp_tjwater_*` 项目查询目录；现有外部 MCP 禁用规则继续执行，没有向模型开放通用 resources 桥或子 Agent 写权限。
+
+取消传播至 OpenCode 原生 MCP、本地 MCP、产品 MCP 客户端、Server handler 和底层工作。接口断线时停止请求，宿主不会无条件后台补跑。读查询只允许一次认证重试和一次可判定未完成的传输重试；调查依 application request key 防止重复保存。
+
+本地 MCP 采用已有内部 token 和浏览器来源拒绝规则；它只证明可信 OpenCode 到宿主的通道，不能替代 Server 的 OAuth 主体认证。日志保留主体、项目、任务、run、工具名、耗时、数量、取消与错误类别，不保存 Bearer、secret、context_handle 或原始敏感数据。
+
+MCP 服务失败时其调用返回可重试错误，现有 REST 和原生查询仍可使用；如果管理员启用 MCP 作为必需产品能力，Agent 启动预热必须验证 server/discover、实际 tools/list、Schema、身份路径和取消，不调用付费推理，不用“连接成功”代替能力就绪。
+
+## 实施范围与验收
+
+这是两仓库的中等规模改动，预计超过 8 个文件；第一版挂载于已有 Server，不新增独立部署服务。前端正常情况下继续复用现有凭据刷新事件，只有验收发现契约缺口时才扩展。
+
+| 仓库 | 实施落点 |
+| --- | --- |
+| Server | app/main.py 挂载和 lifespan；app/core/config.py；新增 app/mcp 的协议、工具与认证适配模块 |
+| Server | app/services 的共用查询与上下文绑定服务；app/domain/schemas 的查询/输出契约；任务源与证据回执的持久存储要求 |
+| Server | 现有 API service 提取、有界查询；独立 MCP 令牌验证；requirements.txt；.env.example；tests/api、tests/auth、tests/unit；OpenAPI 导出 |
+| Agent | src/product 的 MCP 客户端与令牌缓存；现有 credential-refresh；新增内部 MCP call 桥 |
+| Agent | .opencode/plugins/tjwater-tools.ts；MCP 和 task-reader 策略；工具 schema 投影；真实目录启动检查；package.json 与锁文件；聚焦测试 |
+
+升级 Server 的官方 mcp SDK 到已发布 v2，按官方迁移文档核对 lifespan、认证、协议模式和结果类型后固定实际发布版本。不能仅升级 fastmcp 并假定新协议成立；已有依赖是否删除，须先确认真实使用与兼容关系。当前 Python SDK 文档示例同时包含 legacy 配置，协议行为以 2026-07-28 规范和实际验收为准。[Python SDK v2 迁移说明](https://py.sdk.modelcontextprotocol.io/migration/)
+
+部署配置包含固定 Keycloak issuer、规范 MCP resource URI、专用 audience、Agent confidential client ID 与 secret。secret 只在 Agent 宿主环境或密钥管理中，Keycloak 签名密钥由 JWKS 获取并校验固定 issuer，不发给 OpenCode。沿用 worker 自有 client_credentials 配置，不把 worker secret 转交给 Agent。
+
+Server 需提供上下文绑定、撤销和调查查询回执的持久存储；多进程读写须一致，不能用单进程 dict 宣称重启或横向扩容可用。应用层只保存令牌派生身份和范围，不保存 token。存储由后端按现有数据库管理流程交付，本设计不执行数据库迁移。
+
+验收必须覆盖：
+
+1. 实际 `server/discover`、24 工具目录、版本固定、标准 `_meta`、resultType、缓存提示；头与正文不一致、未知工具、非法 schema、旧握手的错误行为。
+2. 两个用户两个项目并发，交错调用也不串凭据；句柄换用户、项目、client、task、run 或失效均拒绝；个人会话不得获得业务数据。
+3. 普通登录无需再次登录 MCP；401续期一次、403不升权；并发 refresh 合并；注销、成员撤销和背景服务凭据续期。
+4. worker 无浏览器完成查询；reader 只取父任务范围；不获得报告、审批、控制、Shell 或其他外部 MCP。
+5. 管网属性与分页、SCADA 原值/清洗值、模型基线、分析 run、历史三种 mode、GIS 与外围 synthetic；不支持的旧功能返回明确能力缺口。
+6. 24小时/目标数/分页边界、实际服务端限量、资源读取授权；超大结果不会全量进入内存或聊天。
+7. 查询成功且证据已经持久化；断线重试仅一个证据回执；查询中轮次被替代、取消、DB失败和认证服务失败均不误报成功。
+8. REST 与 MCP 的权限和数据结果相同；新 SDK 不改变原 API、worker 或报告读取；启动真实目录检查不调用模型推理。
+
+使用 Server 现有 conda 环境运行新增 MCP 聚焦 tests/api、tests/auth、tests/unit；Agent 执行 bun run check 与新增凭据、MCP桥、策略、目录聚焦测试。用户登录、refresh 和两项目隔离需用现有前端做一次实际浏览器验收。设计文档本身仅进行路径检查与 git diff --check，不执行业务写请求。
+
+交付可以分两次独立上线：先增加服务端只读 MCP、OAuth与上下文，保留现有产品调用；再接入 Agent 宿主客户端和 reader，验证后切换产品查询。任何一步失败都可以回到已有原生工具，不能在单次失败时自动切换凭据绕过权限。回滚停用 MCP 入口/宿主注册并撤销上下文，保留已经保存的调查证据与旧报告。提交、身份平台变更、数据库供给和部署另按实际实施授权执行。
+
+## 旧 CLI 与当前后端对照
+
+路径均为 `/api/v1` 下的当前能力；“执行候选”不加入第一版查询 MCP。此表依据当前源码生成 OpenAPI 和旧 CLI handler，而非线上算法成功证明。
+
+| 旧 CLI 命令 | 当前路径或处理 | 第一版决定 |
+| --- | --- | --- |
+| network get-junction-properties | GET /junctions/properties | get_element_properties |
+| network get-pipe-properties | GET /pipes/properties | get_element_properties |
+| network get-all-pipes-properties | GET /pipes | list_elements 分页 |
+| network get-reservoir-properties | GET /reservoirs/properties | get_element_properties |
+| network get-all-reservoirs-properties | GET /reservoirs | list_elements 分页 |
+| network get-tank-properties | GET /tanks/properties | get_element_properties |
+| network get-all-tanks-properties | GET /tanks | list_elements 分页 |
+| network get-pump-properties | GET /pumps/properties | get_element_properties |
+| network get-all-pumps-properties | GET /pumps | list_elements 分页 |
+| network get-valve-properties | GET /valves/properties | get_element_properties |
+| network get-all-valves-properties | GET /valves | list_elements 分页 |
+| component option schema | GET /network-schemas/{kind} | get_network_schema |
+| component option get | GET /network-options 及 time、energy、pump-energy | get_network_options |
+| simulation run | POST /simulation-runs | 执行候选，duration 为分钟 |
+| analysis burst | POST /burst-analyses | 执行候选 |
+| analysis valve isolation | POST /valve-isolation-analyses | 执行候选 |
+| analysis flushing | POST /flushing-analyses | 执行候选 |
+| analysis age | /water-age-analyses 未注册 | 不开放，不假定算法仍可运行 |
+| analysis contaminant | POST /contaminant-simulations | 执行候选 |
+| analysis sensor-placement run | POST /sensor-placement-runs | 执行候选 |
+| analysis sensor-placement list | GET /sensor-placement-runs | list_sensor_placement_runs，有界适配 |
+| analysis sensor-placement get | GET /sensor-placement-runs/{run_id} | get_sensor_placement_run，读取项目路由中的既有成果 |
+| analysis runs list | GET /analysis/runs | list_analysis_runs |
+| analysis runs get | GET /analysis/runs/{run_id} | get_analysis_run |
+| analysis runs results | GET /analysis/runs/{run_id}/results | list_analysis_results |
+| analysis leakage identify | POST /leakage-identifications | 执行候选 |
+| analysis burst-detection detect | POST /burst-detections | 执行候选 |
+| analysis burst-location locate | POST /burst-locations | 执行候选 |
+| data timeseries realtime links | GET /timeseries/realtime/links | query_model_baseline_series，已补最多20个元素筛选及SQL限量 |
+| data timeseries realtime nodes | GET /timeseries/realtime/nodes | query_model_baseline_series，已补最多20个元素筛选及SQL限量 |
+| data timeseries realtime simulation-by-id-time | GET /timeseries/realtime/simulation-results | get_model_baseline_at_time |
+| data timeseries realtime simulation-by-time-property | GET /timeseries/realtime/records | query_model_baseline_field |
+| data timeseries analysis link-field | GET /timeseries/analysis/runs/{run_id}/links/{link_id} | query_analysis_series |
+| data timeseries analysis node-field | GET /timeseries/analysis/runs/{run_id}/nodes/{node_id} | query_analysis_series |
+| data timeseries analysis values | GET /timeseries/analysis/runs/{run_id}/values | query_analysis_values |
+| data timeseries scada query | GET /timeseries/scada-readings 及 fields | query_scada_readings |
+| data timeseries composite | 原三个 views 路径未注册；现有 POST /timeseries/views/element-history/query | 按三种 history mode 重新适配，不直接宣称等价 |
+| data pipeline-health | GET /pipeline-health-predictions，实际进行预测 | 后续计算候选，先验证模型资产 |
+| data scada get | 原 /scada-devices/detail 未注册；现有 GET /scada-devices/{device_id} | 修正为 get_scada_device |
+| data scada list | GET /scada-devices | list_scada_devices 分页 |
+| data scada schema | GET /network-schemas/scada-device | get_network_schema |
+
+当前 Agent 的 `ProjectDataClient` 读取单个元素等能力仍较窄；其中 SCADA 设备读取仅列表，缺少单设备新路径入口。MCP 切换前以本表和当前 Schema 对齐，不能从旧 CLI 直接复制参数或依赖其全量分页行为。
+
+任务上下文查询支持可选的正整数 `report_revision`；未显式指定时宿主保留协作会话绑定的报告修订号，避免将历史报告替换为最新报告。