DeepSeek Harness 插件

mcp-sentinel

DeepSeek Harness plugin that acts as a sentinel between the AI agent and MCP servers — polling long-running tasks on the agent's behalf so that token-costly status loops never enter the LLM inference(英文原文)

跳到安装方式

来源信息

GitHub 仓库
GCS-ZHN/mcp-sentinel
最近更新
2026年8月16日
分类
自动化与任务
GitHub stars
5
载体类型
plugin
包路径
packages/deepseek-harness
目录证据
上游声明已找到 dsh.bundle
证据路径
packages/deepseek-harness/package.json#dsh.bundle
核对版本
0.1.0-rc.8
上游核对日期
2026-08-20

该证据由上游目录提供。本站没有安装、运行或安全审核这个插件。

安装

默认先复制一段 Prompt,让 Agent 读 GitHub 仓库和源码;需要自己装时再切到命令。

复制这段 Prompt,发给 DSH、Codex 或其他 Agent,让它先读 GitHub 仓库和源码。

请先不要安装或执行任何命令。阅读这个插件的 GitHub 仓库、README 和关键源码,然后用清楚、直接的方式回答以下问题,帮助我判断它是否适合我的需求:

1. 这个插件是什么,解决什么问题;
2. 适合哪些用户和典型使用场景;
3. 安装后如何使用,并给出一个最小使用示例;
4. 有哪些已知限制,以及隐私、安全、兼容性或维护风险;
5. 给出“推荐 / 有条件推荐 / 不推荐”的明确建议和理由。

请区分仓库明确说明、根据源码推断和未知信息。证据不足时请明确说明,不要猜测或照抄 README。

GitHub:https://github.com/GCS-ZHN/mcp-sentinel/tree/HEAD/packages/deepseek-harness
插件名:mcp-sentinel
作者:GCS-ZHN

检查来源文件

安装前先看这个插件目录里的 README 和其他文件。

文件资源管理器3 个文件
README.md来源说明 · 只读预览

@gcszhn/mcp-sentinel-deepseek-harness-plugin

A DeepSeek Harness plugin that acts as a sentinel between the AI agent and MCP servers — polling long-running tasks on the agent's behalf so that token-costly status loops never enter the LLM inference path.

This package is the DeepSeek Harness adapter for @gcszhn/mcp-sentinel-core.

Install

Install into a profile with the dsh CLI:

dsh plugin --profile <name> add @gcszhn/mcp-sentinel-deepseek-harness-plugin

The package ships a dsh.bundle manifest, so dsh plugin add appends it to the profile's dsh.profile.bundles list and activates its cordis.patch.yml layer. Verify the layer without booting:

dsh --profile <name> --dump-config

How it talks to MCP

The plugin runs in external-invoker mode: it reuses the MCP tools already registered by the harness's @deepseek-ai/dsh-mcp-client bridge and never owns MCP connections itself, so there is no separate servers config and no sentinel-specific MCP wiring. You keep configuring MCP exactly as you already do for the harness — one dsh-mcp-client instance per server:

# This is ordinary dsh-mcp-client config, not sentinel config.
- insert:
    - id: mcp-ci
      name: "@deepseek-ai/dsh-mcp-client"
      config:
        serverName: ci
        transport: stdio
        command: bun
        args: ["/path/to/ci-mcp-server.ts"]

When calling mcp_sentinel_poll, server is the mcp-client serverName and tool is the server's raw tool name; the sentinel invokes mcp__<server>__<tool> (e.g. mcp__ci__get_status) through the harness tool registry. Anything you already bridged with dsh-mcp-client is immediately pollable — no extra step.

Tools

mcp_sentinel_poll

Submit a long-running MCP tool call and poll it at regular intervals until a condition is met. Returns a sentinel ID immediately; when the sentinel resolves, the plugin pushes a completion notice into the originating agent's inbox (Agent.followup) so the driver wakes and the agent can collect the result with mcp_sentinel_attach (blocking), mcp_sentinel_status, or mcp_sentinel_read.

ParameterTypeDefaultDescription
serverstring_required_serverName of a dsh-mcp-client instance
toolstring_required_Tool name to call on the server
argsobject{}JSON object of arguments for the tool
intervalnumber5000Poll interval in milliseconds
timeoutnumber_optional_Max poll duration in ms (unset = no limit)
untilobject_required_JSON condition object

mcp_sentinel_status

Check status, list active tasks, or cancel a running task (action =

status | list | cancel).

mcp_sentinel_attach

Block the agent, waiting for a sentinel to complete. Zero token cost during the wait.

mcp_sentinel_read

Read raw poll outputs with offset/limit pagination.

Condition model

Conditions are pure declarative data:

{ "path": "status", "is": "eq", "value": "completed" }        // path leaf
{ "is": "eq", "value": "completed" }                          // no path: compare the raw result
{ "path": "tasks[0].exit_code", "is": "ne", "value": 0 }     // array-index path
{ "not": { "path": "status", "is": "eq", "value": "error" } } // negation
{ "and": [ /* conditions */ ] }                               // logical AND
{ "or": [ /* conditions */ ] }                                // logical OR

Operators: eq, ne, gt, gte, lt, lte, contains, match. Omit path (or leave it empty) to match a non-JSON tool result directly.

License

MIT