dsh-switch
模型与数据dongsheng123132/dsh-switch
DeepSeek Harness 的模型控制平面。提供工具检查已注册的提供商路由、验证特定提供商/模型,并为新会话设置默认模型。附带 CLI 探测 OpenAI 兼容端点的健康与延迟。
- 2origin
- ai-agent
- deepseek-harness
- dsh
- dsh-plugin
- model-routing
- provider-health
验证与兼容性
这里展示目录实际采集到的证据;未声明的信息会明确标为未知。
- 当前版本兼容性
- 已在当前目录版本验证
- 声明的 Harness 范围
- 未声明
- 声明的平台
- 未声明
- 适用 Profile
- web
- 构建授权
- 未知
- 权限声明
- 未声明
- 外部服务
- 未声明
- 遥测声明
- 未知
这不是安全背书;安装前仍应查看源码、权限和配置。
查看证据与判定范围
验证仅覆盖标出的来源、版本和 Harness 环境,不代表未来版本仍然兼容。
github:dongsheng123132/dsh-switch#b37a63cae2654044942a5dd44b6f2bc8591f5ec6- 插件通过清单与静态结构检查,尚未完成运行时加载检查。
README
dsh-switch
Evidence-first model control plane for DeepSeek Harness.
dsh-switch does two deliberately separate jobs:
- The DSH bundle exposes model-facing tools that inspect registered provider routes, validate an exact provider/model through the official
ctx.llmseam, and save the default for new sessions throughctx.agentDefaultModel. - The standalone CLI probes OpenAI-compatible
/modelsendpoints and reports reachability, HTTP status, latency, model count, and whether a credential environment variable is configured.
It does not intercept model requests, implement another OpenAI adapter, or silently reroute an active session. Projects such as dsh-polyglot and dsh-llm-fallbacks own adapter/fallback behavior; this project is the observable control plane around routes DSH already knows.
Install the DSH bundle
Pin a reviewed commit when installing from GitHub:
dsh plugin --profile web add github:dongsheng123132/dsh-switch#<commit>
dsh --profile web --dump-config
The bundle ships JavaScript and has no install/build lifecycle script.
Tools
dsh_switch_status({ includeModels? })
- shows the default selection used by newly created sessions;
- lists registered provider routes;
- optionally asks each adapter for its advertised models;
- performs no provider network health request itself.
dsh_switch_default({ provider, model, reasoningEffort?, expectedProvider?, expectedModel? })
- validates the exact route with
ctx.llm.resolveModelInfo(); - rejects an unsupported reasoning effort;
- supports an optimistic-concurrency expectation pair;
- saves through
ctx.agentDefaultModel.saveSelection(); - reads the value back and fails if the write is not observable;
- affects new sessions only. Existing sessions keep the route recorded in their logs.
Probe provider endpoints
Copy providers.example.json to a private local file such as providers.json. Reference secrets by environment-variable name—never place keys in JSON.
node bin/dsh-switch.mjs validate --config providers.json
node bin/dsh-switch.mjs check --config providers.json
node bin/dsh-switch.mjs best --config providers.json
JSON output:
node bin/dsh-switch.mjs check --config providers.json --json
Example output:
✓ deepseek-official 83.4 ms HTTP 200, 2 models credential:configured
✗ local-vllm 1.2 ms network-error credential:not-required
The probe sends GET <baseURL><modelsPath>. It does not send a chat completion and therefore does not consume inference tokens. A successful /models response proves endpoint reachability, not end-to-end generation correctness.
Security and evidence boundaries
apiKeyfields and credentials embedded in URLs are rejected.- Output contains only
configured,missing, ornot-required; secret values never enter results. - A 401/403 endpoint is reachable but not healthy.
- DSH tool switching validates registered adapter metadata, not live provider I/O.
- No automatic fallback or cost-based routing is enabled in v0.1.
Development
npm test
npm run check
The plugin smoke test requires the official DSH runtime packages to be resolvable:
npm run smoke:plugin
License
MIT © 2026 hfshfg