dsh-wsl-hostsvc
从 WSL 探测 Windows 主机上的 Ollama / LM Studio / vLLM / llama-server；比较 openAiModels 的 apiReady 与 TCP。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:173787247/dsh-wsl-hostsvc#8f2cdd2857b79fa6370edc2acb777a26ffae9097Wsl Hostsvc 以 dsh-wsl-hostsvc 发布，当前版本为 0.4.3。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
Install set: part of dsh-wsl-kit. Prefer
KIT_SET=daily|llm|github|full(see kit README). Fault tree: TROUBLESHOOTING.md.
DeepSeek Harness plugin: from WSL, probe Windows-host OpenAI-compatible LLM servers and suggest a reachable baseURL.
From WSL, probes Windows-side Ollama, LM Studio, vLLM, and llama-server. TCP open is not the same as API ready.
flowchart LR
agent["dsh agent"] --> tool["host_reach"] --> win["Windows local LLM ports"]
Suite diagram and version snapshot: dsh-wsl-kit. This plugin is 0.4.3 (llm; also in full). Do not copy that matrix into this README.
| Field | Value |
|---|---|
| Plugin | dsh-wsl-hostsvc 0.4.3 |
| Minimum dsh | ≥ 0.1.2 (web UI one-shot ?token= on Windows relay :3081) |
| Latest verified | See dsh-wsl-kit Compatibility (currently 0.1.7-alpha.2) — single source of truth for the suite |
| Kit set | llm / full (some also useful alone) |
| Cloud Flash | Use model id deepseek-flash (V4.1 Flash) in ~/.dsh/settings.yaml / llm-deepseek — not configured by this plugin |
| Agent Teams | Upstream experimental; not required here |
Suite floor versions: kit check-plugin-versions.sh. Fault tree: TROUBLESHOOTING.md.
Scope: local OpenAI-compatible servers on the Windows host (host_reach). Does not configure cloud DeepSeek; for V4.1 Flash set model id deepseek-flash in settings.
Tool host_reach (default profile=all) probes:
| id | Port | Typical server |
|---|---|---|
ollama | 11434 | Ollama |
lmstudio | 1234 | LM Studio local server |
vllm | 8000 | vLLM OpenAI server |
llama | 8080 | llama-server / Unsloth Desktop |
Hosts tried: 127.0.0.1, localhost, host.docker.internal, and the Windows IP from /etc/resolv.conf.
suggestedBaseURL / per-service URLs → llm-pi-ai.providers.*.baseURL (keep /v1)providerSnippets.yaml → paste under llm-pi-ai.providers in ~/.dsh/settings.yamlollamaModels → exact ids from ollama list / /api/tagsconnectivityPlaybook → ordered tools when HTTPS/DNS also failn_ctx vs settings contextWindow, Unsloth Desktop vs Studio, VRAMTemplate: examples/local-llm-providers.settings.yaml.
Plugin-heavy dsh prompts often need >8k tokens even for a short chat turn. This is independent of cloud deepseek-flash (V4.1 Flash), which you configure under official llm-deepseek / default model — not via host_reach.
PARAMETER num_ctx 32768 (Modelfile / recreate model), or OLLAMA_NUM_CTX on Windows before starting Ollama.contextWindow in settings to that num_ctx (not the model's theoretical max like 131072 unless Ollama actually runs it).maxTokens smaller than contextWindow (e.g. 4096) so the prompt still fits.Mismatch examples:
131072 + Ollama n_ctx=8192 → API errors8192 + fat system/tools prompt (~10k) → CONTEXT_WINDOW_EXCEEDED~/.unsloth/studio). Desktop often exposes OpenAI /v1 on 8080 (llama profile).# Recommended with the LLM set:
curl -fsSL https://raw.githubusercontent.com/173787247/dsh-wsl-kit/master/install.sh | KIT_SET=llm bash
# Or alone:
dsh plugin --profile web add github:173787247/dsh-wsl-hostsvc
Restart dsh web and open a new session.
WSL UI from Windows browser: use kit scripts/restart-dsh-web.sh and open http://127.0.0.1:3081/ (relay; dsh itself binds 127.0.0.1:3080 only).
Ask the agent to run host_reach (profile: ollama|lmstudio|vllm|llama|all, optional port, includeProviders).
If nothing opens: on Windows bind beyond loopback (e.g. OLLAMA_HOST=0.0.0.0:11434) or enable WSL mirrored networking, then re-probe.
MIT