dsh-plugin-model-proxy
DSH 插件：按模型进行代理路由（http/https/socks5），并提供设置界面——例如，opencode/muse-spark-1.2-contributor 需要代理，而同级模型保持直连
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add dsh-plugin-model-proxy@0.1.3Plugin Model Proxy 以 dsh-plugin-model-proxy 发布，当前版本为 0.1.3。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。
Community DSH plugin: route specific provider/model pairs through specific proxies (http://, https://, socks5://, socks5h://) — with a Settings UI.
Some model endpoints are only reachable through a proxy while their siblings
connect fine — for instance opencode/muse-spark-1.2-contributor may answer
403 RegionError to direct connections that other models from the same
provider never see. The coarse fixes don't help: process-wide proxy
environment variables drag every request through the tunnel, and swapping a
whole provider's baseURL hides the real upstream. This plugin routes
specific provider/model pairs at the transport layer instead — everything
else stays direct, and the configured endpoint is never rewritten.
{provider, model, proxyUrl, enabled} — specificity: exact model > prefix muse-* > *.purpose (e.g. compaction) so chat goes through the proxy while background calls stay direct.user:password entries). Soft dependency — installs without it keep working.http://, https:// (CONNECT via undici.ProxyAgent), socks5:// / socks5h:// (via optional socks, tunnelled by undici Agent + custom connect).llm/stream uses them; in-flight streams unaffected.Settings → Plugins → Model Proxy card (also works via ~/.dsh/settings.yaml).llm.providers × settings mirror, same semantics as the built-in Models page); bare directory routes follow, and "Custom…" accepts anything — hand-written yaml rules, wildcards, gateways the catalog doesn't know. Provider and model fields are dropdowns fed by the live host catalog (llm.providers / llm.models, refreshed on llm/adapters-updated); a "Custom…" entry keeps free text for wildcards (muse-*, *) or not-yet-installed providers.The package declares a dsh.bundle layer, so the DSH CLI links the dependency
and appends the activation row in one step:
dsh plugin --profile <name> add dsh-plugin-model-proxy
# optional SOCKS support:
dsh plugin --profile <name> add socks
Verify the composed layer without booting, then run:
dsh --profile <name> --dump-config # shows a "# == dsh-plugin-model-proxy" layer
dsh --profile <name>
Uninstall removes both the dependency and the layer:
dsh plugin --profile <name> remove dsh-plugin-model-proxy
Installing from GitHub (dsh plugin --profile <name> add github:<user>/dsh-plugin-model-proxy#<sha>)
fetches sources; pnpm ≥10 asks you to allowlist the prepare build — copy the
package key it prints into the profile's pnpm-workspace.yaml under
allowBuilds:, then re-run the add. Pin a commit so later pushes cannot
change what runs on your machine.
If you manage compositions by hand, install the dependency into the profile
directory first (dsh plugin --profile <name> add ./ from a checkout works;
module resolution anchors at the profile). Then insert exactly ONE loader row:
# ~/.dsh/cordis.patch.yml (or a --patch overlay)
- insert:
- id: model-proxy/host
name: dsh-plugin-model-proxy
config: {}
Patch-file rules that trip people up:
- … rows and no []) fails to parse
and breaks every boot/dump that reads it.name: (the package name), not module:.insert: — a bare - id: row means "patch an existing entry".dsh.client declaration and served at
/plugins/dsh-plugin-model-proxy/client.js. Do NOT add a model-proxy/client
row — it would run browser code inside the Node process.dsh.bundle.patch, dsh plugin add injects the host row
automatically; keeping a hand-written copy in ~/.dsh/cordis.patch.yml
inserts the same loader id twice and the composition collides at boot.Uninstall = remove the inserted rows + pnpm remove dsh-plugin-model-proxy
from the profile.
Settings → Plugins → Model Proxy
~/.dsh/settings.yaml:
model-proxy:
enabled: true
debug: false
defaultProxy: "" # fallback when no rule matches
rules:
- provider: opencode
model: muse-spark-1.2-contributor
proxyUrl: socks5://127.0.0.1:1080
enabled: true
- provider: opencode
model: "*"
proxyUrl: "" # direct for the rest of this provider
proxyUrl: "" means direct (exempt). socks5h:// resolves DNS at the proxy.
model-proxy settings namespace (installSettingsSection, live).globalThis.fetch reversibly.llm/stream waterfall, resolves proxyUrl for (provider, model) via AsyncLocalStorage, then injects a dispatcher into the adapter's fetch — for proxied requests it routes through undici.fetch (Node's native global fetch ignores custom dispatchers), with undici.ProxyAgent for http(s) and an undici Agent + socks connect (with TLS) for socks5/h.settings.plugin.item with key model-proxy — automatically paired by the Plugins tab (served ∩ registered).No Adapter fork, no baseURL rewrite. See DESIGN.md for full design and invasiveness analysis.
| Check | Result |
|---|---|
Modifies packages/* | No |
Requires fork of dsh-llm | No |
| Global side effect reversible | Yes (ctx.effect dispose restores fetch) |
| Client bundle purity gate | Passes (only type-only slot import) |
pnpm install
pnpm run build
Fully offline (local origin + CONNECT-capable toy proxy; no real network or proxy needed):
pnpm test # builds host + client, then runs node --test tests/
pnpm run typecheck
MIT