@openllmsh/dsh
DeepSeek Harness (dsh) 捆绑包，通过 OpenLLM 网关路由该 harness——在内置的 pi-ai 适配器上提供一个 `openllm` provider（通过本地守护进程网关），以及 OpenLLM MCP 服务器。纯配置；请先安装 OpenLLM（参见 README）。
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:openllmsh/dsh#2eeb1d0e1a9bec3a10570bdf865638a62b080528@openllmsh/dsh 以 @openllmsh/dsh 发布，当前版本为 0.1.0。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。
@openllmsh/dsh — route DeepSeek Harness through OpenLLM.
A DeepSeek Harness (dsh)
bundle that points the harness at the OpenLLM gateway and
hands it OpenLLM's MCP tools. It is a pure-config Cordis patch layer —
installed from the dsh side, no runtime code — and it does two things:
openllm provider to dsh's in-box llm-pi-ai
adapter (openai-completions) pointed at the daemon's local-first gateway
(http://127.0.0.1:8787/v1) and makes it the default model. No adapter code,
and no API key: that loopback surface has no auth gate; the daemon fetches
signed plans with its own ~/.openllm/.env credentials and forwards upstream.openllm mcp stdio server (openllm,
claude-context, supermemory tool groups). The openllm binary resolves
~/.openllm/.env itself, so it too needs nothing from dsh.dsh holds no key and no origin — it just talks to 127.0.0.1:8787, so the
only thing it requires is that openllmd is running. Installing OpenLLM is a
prerequisite (below) — the installer starts the daemon for you; this bundle does
not install, prompt for, or onboard it. Runs on the host DeepSeek Harness — the
services it patches
(@deepseek-ai/dsh-llm-pi-ai, @deepseek-ai/dsh-mcp-client) ship in
@deepseek-ai/dsh-base.
Do this once — dsh routes to the local OpenLLM daemon, and the installer sets it up and starts it for you (macOS / Linux):
curl -fsSL https://www.openllm.sh/install | bash
This is OpenLLM's own official installer (canonical source: https://openllm.sh/install); it digest-verifies the binaries it downloads. If your policy requires it, fetch and read the script before piping it to a shell. This bundle only documents the prerequisite — it never runs it.
There is nothing else to run, paste, or export — the installer starts the daemon
and persists your key to ~/.openllm/.env, and the daemon serves dsh from there.
Install into the profile whose surface you actually run. dsh profiles are
separate bundle stacks, so a bundle added to one profile is invisible to the
others — the browser UI lives only in the web profile (web mounts
dsh-base + dsh-web-app; headless mounts dsh-base + dsh-headless).
The bare default profile is dsh-base only — no web surface — so adding the
bundle there never reaches the web UI.
Browser UI (web profile):
dsh plugin --profile web add github:openllmsh/dsh
Headless / CLI (headless profile):
dsh plugin --profile headless add github:openllmsh/dsh
Then restart the profile. Distributed from GitHub — no npm package. The built
lib/ is committed, so the git install loads with no build step (and no pnpm
allowBuilds prompt). Pin a release with github:openllmsh/dsh#<tag>.
The router base (http://127.0.0.1:8787/v1) is fixed in cordis.patch.yml. To
route through a self-hosted OpenLLM, point the daemon at your origin when you
install it — dsh still just talks to 127.0.0.1:8787. A custom daemon port
(OPENLLM_DAEMON_PORT) means editing the baseURL in cordis.patch.yml to match.
pnpm install
pnpm build # tsdown → lib/index.mjs (+ .d.mts) — commit the result
pnpm typecheck # tsc --noEmit
dsh plugin --profile demo add . # link this bundle into a scratch profile
dsh --profile demo --dump-config # inspect the composed plugin tree
The bundle's substance is cordis.patch.yml (resolved via the dsh.bundle.patch
manifest field); src/index.ts is an export {} placeholder with no runtime API,
matching @deepseek-ai/dsh-base.
lib/is committed (GitHub is the only distribution channel), so rebuild and commit it wheneversrc/changes — a stalelib/is what git installs.
Developer preview — dsh is pre-1.0; pin and re-test against each
dshbump.
MIT © OpenLLM, INC — see LICENSE.