DeepSeek HarnessDeepSeek Harness

DeepSeek HarnessDeepSeek Harness

Start with engine capabilities, then follow its installation, model sync and workspace guide. Pi remains experimental; engine capabilities differ.先了解各引擎能力，再按目标完成安装、模型同步和工作台配置。Pi 仍为实验版，各引擎功能并不相同。

DSH：运行配置与完整工作台

DeepSeek Harness 独立于 OpenClaw Gateway。按 ClawPanel 0.22.1 基线，默认受管目标是 0.2.0-rc.2，它仍为预发布候选版（RC），不是上游稳定正式版。0.2.1-alpha.2 仅作为显式 Alpha 测试通道，可按页面切回默认目标；不要把「稳定通道」标签理解为去掉了 RC 状态。

  1. 从侧栏选择 DeepSeek Harness，打开「运行与配置」，检查完整 Node.js 24.16.0+（24.x）和 npm。
  2. 安装页面给出的受管目标，再核对实际版本与日志；运行时安装在受管目录，不使用全局 npm 环境。
  3. 在「通用 → 模型渠道」保存 Provider、Base URL、API Key、模型及上下文/输出上限，显式点击「同步到 DeepSeek Harness」，按需设置默认模型。
  4. 启动服务，从「工作台」进入内嵌 DSH Web 界面，选择工作区、模型并创建会话，验证完整对话回复。

模型协议与网络

当前转换支持 OpenAI Chat Completions、OpenAI Responses、Anthropic Messages；Ollama 转到 OpenAI 兼容 /v1，Google Generative AI 暂无此同步转换。修改渠道后再次同步，配置回读不等于真实 Provider 调用成功。

DSH 服务保持回环监听，Web/headless 模式通过面板认证、短期令牌与内嵌代理使用，无需另开外部 DSH 端口。不要关闭上游认证。工作区选择和模型调用属于面板服务器；浏览器电脑的目录不会自动共享。

排查与版本边界

先看「运行与配置」的状态、版本和日志，再检查目标同步与模型选择。面板重启后若遗留 DSH 进程导致认证异常，按页面停止并由面板重新启动，不从日志提取令牌。Alpha 安装/版本回读不等于全部工具或付费模型协议已验收。

另见引擎与驱动及Web 部署。DSH 的受管生命周期不应套用为其他引擎或外部实例的管理规则。

DSH configuration and full workspace

DeepSeek Harness runs independently of OpenClaw Gateway. At the ClawPanel 0.22.1 baseline, the default managed target is 0.2.0-rc.2, still a prerelease release candidate (RC), not an upstream stable release. 0.2.1-alpha.2 is an explicit Alpha test channel with a return path to the default target. A stable-channel label does not remove the RC status.

  1. Select DeepSeek Harness, open Run & configuration and check a complete Node.js 24.16.0+ (24.x) installation with npm.
  2. Install the displayed managed target, then check actual version and logs. Its managed directory does not use global npm installation.
  3. Save Provider, Base URL, API key, models and context/output limits in General → Model channels. Explicitly select Sync to DeepSeek Harness and optionally set a default model.
  4. Start the service, enter the embedded DSH Web workspace, select a project and model, create a session and verify a complete test reply.

Protocols and networking

Current conversion covers OpenAI Chat Completions, OpenAI Responses and Anthropic Messages. Ollama uses its OpenAI-compatible /v1 interface; Google Generative AI has no conversion here. Sync again after channel edits; configuration readback is not proof of a real Provider call.

DSH stays on loopback. Web/headless access uses panel authentication, short-lived tokens and the embedded proxy, without an extra public DSH port. Keep upstream authentication enabled. Projects and model calls belong to the panel server, not the browser computer.

Troubleshooting and acceptance boundaries

Check configuration status, version and logs before target synchronization and model selection. If a leftover DSH process causes authentication errors after panel restart, stop it through the panel and start it again; do not recover tokens from logs. Alpha installation and version readback do not prove acceptance of every tool or paid model protocol.

See Engines & drivers and Web deployment. Do not apply DSH lifecycle rules to other engines or external instances.

qt.cool