Prefer multi-NE execManagedNe batch and parallelize read MCP tools.

Guide ops to use ne_ids/ume_ne_ids for sweeps, raise exec wall-clock for batches, and mark inventory/list MCP tools read_only so consecutive queries can run in parallel.

Co-authored-by: Cursor <cursoragent@cursor.com>

runtime/tools/mcp/adapter.py

@@ -21,8 +21,9 @@ def _mcp_row_env_config(row: dict[str, Any]) -> tuple[list[str], dict[str, str]]
 
 # Long-running netx tools exceed the generic MCP row timeout (often 30s).
 # Production WA ops showed execManagedNe p90/p95 glued to ~30000ms timeouts.
+# Batch multi-NE exec posts once to /exec-batch (server concurrency); allow longer wall clock.
 _MCP_TOOL_TIMEOUT_OVERRIDES_S: dict[str, float] = {
-    "execManagedNe": 320.0,
+    "execManagedNe": 620.0,
     "sqlQueryUme": 90.0,
     "findTopologyPaths": 60.0,
     "aggregateUmeAlarmsRaw": 60.0,
@@ -44,6 +45,26 @@ _MCP_LIST_CACHE_TTL_S: dict[str, float] = {
     "findTopologyPaths": 60.0,
 }
 
+# Safe to run in parallel with other consecutive read-only tools (separate MCP stdio processes).
+# Do NOT include execManagedNe: same-tool fans share one stdio lock — use ne_ids/ume_ne_ids batch instead.
+_MCP_READ_ONLY_TOOLS: frozenset[str] = frozenset(
+    {
+        "listCliTargets",
+        "listManagedNe",
+        "getManagedNe",
+        "getUmeNe",
+        "queryUmeNeInventory",
+        "queryUmeAlarms",
+        "queryUmeAlarmsRaw",
+        "aggregateUmeAlarms",
+        "aggregateUmeAlarmsRaw",
+        "listUmeAlarmFields",
+        "runUmeDiagnostics",
+        "findTopologyPaths",
+        "sqlQueryUme",
+    }
+)
+
 _MCP_LIST_CACHE_LOCK = threading.Lock()
 _MCP_LIST_CACHE: dict[str, tuple[float, float, dict[str, Any]]] = {}
 
@@ -176,6 +197,7 @@ class _McpBoundTool:
             timeout_s=self.timeout_s,
             required_permissions=self.required_permissions,
             execution_mode="subprocess",
+            read_only=self.tool_name in _MCP_READ_ONLY_TOOLS,
         )
 
 

runtime/workspaces/ops/ROLE_SYSTEM.en.md

@@ -38,7 +38,7 @@ You are the ops specialist (network operations expert).
 - Spreadsheet delivery: `ume_alarm_xlsx_report` or `write_xlsx(deliverable=true)` — never claim a file was sent without deliverable marking.
 - **Field default is English**: WhatsApp channel dispatch defaults to `lang=en`; user-visible replies must contain **zero CJK**. Translate Chinese tool fields before display.
 - Group chats default to **per-speaker session isolation** (members do not share dialogue memory within the same group).
-- Call `listCliTargets` at most once per session and reuse ids; batch `execManagedNe` commands; default `read_timeout_sec=60` — on timeout raise it, no blind retries.
+- Call `listCliTargets` at most once per session and reuse ids; for many NEs with the same show commands use one `execManagedNe(ne_ids|ume_ne_ids=..., commands=...)` (server concurrency) — do not loop one-NE calls; default `read_timeout_sec=60` — on timeout raise it, no blind retries.
 - Replies like `YES` / `confirm` / `继续` / `please continue`: continue the previous unfinished task — do **not** re-ask for confirmation or restart the query.
 - On `tool_invalid_arguments`, fix args using the returned `example`; on timeout hints, raise `read_timeout_sec` or shrink commands.
 

runtime/workspaces/ops/ROLE_SYSTEM.md

@@ -29,7 +29,7 @@
 - 用户要表格/Excel：`ume_alarm_xlsx_report` 或 `write_xlsx(deliverable=true)`；禁止只写文件不投递。
 - **现场默认英文**：WhatsApp 渠道默认 `lang=en`；英文会话回复不得含汉字；工具中文字段先翻译再展示。
 - 群聊默认按**发言人隔离会话**（同群不同人互不串上下文）；勿假设「群共享一个对话记忆」。
-- `listCliTargets` 每会话最多查一次并复用 id；`execManagedNe` 合并 commands，超时调 `read_timeout_sec`（默认 60），禁止盲重试。
+- `listCliTargets` 每会话最多查一次并复用 id；多台同命令用 `execManagedNe(ne_ids|ume_ne_ids=..., commands=...)` 一批并发，勿逐台循环；超时调 `read_timeout_sec`（默认 60），禁止盲重试。
 - 用户回复 `YES` / `confirm` / `确认` / `可以` / `继续` / `please continue`：直接承接上一未完成任务继续执行，**不要**再问一遍确认或重开查询。
 - 工具返回 `tool_invalid_arguments` 时按返回的 `example` 修正参数；返回超时 hint 时提高 `read_timeout_sec` 或减命令，禁止相同参数重试。
 

skills/_workspace/ops/ops-netx-managed-ne-playbook/SKILL.md

@@ -21,6 +21,7 @@ description: 面向 ops 专家的 netx 纳管网元（网元管理）作业手
    - **UME 清单（无需逐台纳管）**：`mcp__netx__listCliTargets`（`source=ume`）或 `queryUmeNeInventory` 取 `ne_id`，再用 `ume_ne_id` 执行 CLI（需先在 netx **UME → CLI 连接** 配置统一凭据/跳板）
 2. **登录查信息**
    - `mcp__netx__execManagedNe`：`ne_id` **或** `ume_ne_id` + `commands`（默认最多 5 条，可由 `NETX_NE_EXEC_MAX_COMMANDS` 调高，硬上限 50）
+   - **多台同命令（推荐）**：一次调用传 `ne_ids` / `ume_ne_ids`（或 `targets`）+ 共享 `commands`，服务端并发执行（默认 concurrency=4，最多 20 台）。禁止对同一 show 命令逐台循环 `execManagedNe`
    - **一次会话内**：`listCliTargets` 最多调用一次，缓存返回的 id；多条 show 合并进同一次 `commands`，禁止「list→exec→list→exec」循环
    - 超时：提高 `read_timeout_sec`（默认 60，慢命令 90–120）或减少命令条数，禁止对同一命令盲重试
 

tests/test_mcp_read_only_parallel.py

@@ -0,0 +1,65 @@
+from __future__ import annotations
+
+from runtime.tools.base import ToolRegistry, ToolSpec
+from runtime.tools.mcp.adapter import _MCP_READ_ONLY_TOOLS, _McpBoundTool, mcp_timeout_for_tool
+
+
+def test_list_query_mcp_tools_marked_read_only() -> None:
+    for name in ("listManagedNe", "queryUmeAlarmsRaw", "listCliTargets", "findTopologyPaths"):
+        assert name in _MCP_READ_ONLY_TOOLS
+        spec = _McpBoundTool(
+            server_id="netx",
+            tool_name=name,
+            description="t",
+            parameters={"type": "object", "properties": {}},
+            command=["echo"],
+        ).to_spec()
+        assert spec.read_only is True
+        assert spec.name == f"mcp__netx__{name}"
+
+
+def test_exec_managed_ne_not_read_only_use_batch_instead() -> None:
+    assert "execManagedNe" not in _MCP_READ_ONLY_TOOLS
+    spec = _McpBoundTool(
+        server_id="netx",
+        tool_name="execManagedNe",
+        description="t",
+        parameters={"type": "object", "properties": {}},
+        command=["echo"],
+    ).to_spec()
+    assert spec.read_only is False
+
+
+def test_exec_managed_ne_timeout_allows_batch_wall_clock() -> None:
+    assert mcp_timeout_for_tool("execManagedNe", 30.0) >= 600.0
+
+
+def test_partition_would_parallelize_consecutive_list_tools() -> None:
+    from runtime.chat.tool_runtime import partition_tool_use_batches
+    from svc.llm.chat_models import LLMToolCall
+
+    reg = ToolRegistry(
+        [
+            ToolSpec(
+                name="mcp__netx__listManagedNe",
+                description="a",
+                parameters={},
+                handler=lambda _a: {"ok": True},
+                read_only=True,
+            ),
+            ToolSpec(
+                name="mcp__netx__queryUmeAlarmsRaw",
+                description="b",
+                parameters={},
+                handler=lambda _a: {"ok": True},
+                read_only=True,
+            ),
+        ]
+    )
+    calls = [
+        LLMToolCall(id="1", name="mcp__netx__listManagedNe", arguments={}),
+        LLMToolCall(id="2", name="mcp__netx__queryUmeAlarmsRaw", arguments={}),
+    ]
+    batches = partition_tool_use_batches(calls, reg)
+    assert len(batches) == 1
+    assert len(batches[0]) == 2