@superfish058/dsh-llm-proxy

Model-level proxy routing + retry for DSH LLM requests, with a Web settings page (模型代理) in the DSH 设置 UI.

Problem

• DSH's LLM adapter (pi-ai → OpenAI SDK → undici fetch) exposes no per-provider proxy knob.
• NODE_USE_ENV_PROXY is only honored at process start — a plugin cannot set it at runtime.
• Result: overseas model endpoints (e.g. api.b.ai) time out from a mainland network.
• Bonus problem: free-tier quotas get exhausted by transient 429s and 5xx — there is no built-in throttle, but a retry policy keeps requests alive through rate-limit windows.

Solution

On apply, the plugin installs a custom undici RoutingDispatcher (wrapped in undici's official RetryAgent) as the global dispatcher. Every LLM request is routed by target hostname:

1. Loopback hosts (localhost, 127.0.0.1, ::1) → direct (always).
2. Hosts whose models are selected in 走代理的模型 → the configured proxy (proxyHost:proxyPort).
3. Everything else → direct (domestic APIs: DeepSeek, Xiaomi, Qwen, Zhipu, …).

The default is DIRECT — the model is the unit of intent: pick the models that need the proxy, everything else stays on the direct path.

Retry (retries / retryIntervalMs, default 3 / 1000ms): the undici RetryAgent replays the request on transport errors (ECONNRESET, ECONNREFUSED, …), HTTP 429 and 5xx (500/502/503/504). LLM calls are POST, so the method list is overridden to include POST.

Because each provider has a distinct baseURL host, this gives per-provider routing with a single mechanism — below the LLM adapter. ctx.llm providers and settings.yaml stay untouched, and no process restart is needed: the dispatcher swap is live from the first request.

Connection-pool pitfall (undici 8.x, fixed in v1.0.0): the ProxyAgent's internal proxy-side pool defaults to keep-alive. When the proxy (e.g. Clash) silently closes an idle CONNECT tunnel, undici keeps reusing the dead socket and the request hangs until timeout — surfaced as Request timed out / HeadersTimeoutError, intermittently (reproduced: 2/10 requests fail with 2.5s idle gaps while a raw tunnel succeeds 10/10). The dispatcher now constructs the proxy agent with clientFactory + pipelining: 0 so every tunnel is a fresh connection (verified 10/10 on the full RetryAgent(RoutingDispatcher) path).

Web settings page（模型代理）

The DSH 设置 page shows a 模型代理 plugin card inside 插件 → 可配置插件 (the official settings.plugin.item slot, declared by @deepseek-ai/dsh-client-ui-settings-plugins). Opening the card reveals the configurable items:

• 代理地址 / 代理端口 (proxyHost / proxyPort, default 127.0.0.1:7897) — the proxy does not have to live on this machine.
• 走代理的模型 (proxiedModels) — a dropdown populated from the configured model list (llm-pi-ai + llm-deepseek namespaces via a host bridge); add multiple models, each routes its baseURL host through the proxy.
• 重试次数 / 重试间隔 (retries / retryIntervalMs) — retry policy for failed requests.

Saving writes the settings namespace llm-proxy and re-applies the dispatcher live (no restart).

• Host half (lib/settings.js): registers the llm-proxy settings namespace (same Config schema; the cordis.patch.yml config is the composition base) and watches it for committed changes. Because rc.6 host-apiproxy only exposes a hard-coded namespace allowlist, the namespace is also re-served to the browser over a loopback-only bridge (/api/dsh-llm-proxy/settings/{describe,mutate,models}) — the same pattern the official dsh-web-ui settings plugin uses.
• Browser half (src/client/): registers the settings.plugin.item card and binds a compat settings scope (official transport first, bridge fallback), so the card works on any host. The card draws its own form; the section stacks cards and reports their count.
• When no settings seam exists (tests, bare config), the plugin falls back to applying the patch config directly.

Bridge status (rc.6 temporary): the settings bridge is a *temporary rc.6 compatibility seam — settings.register()/namespace exposure is not yet public on the host side (api-proxy serves a hard-coded allowlist), so the llm-proxy namespace is re-served loopback-only. Once the official host exposes settings.register() for third-party namespaces, the bridge routes (lib/settings.js) and the bridge fallback half of src/client/settings-scope.ts can be removed; the card itself stays, riding the official transport only.

Rebuild the browser half after editing src/client/:

npm run build        # tsdown → lib/client.js (window.__ModuleLoader__ closure)
npm run typecheck    # tsc --noEmit

Then restart dsh web (tray exit → start) to load the new host half and re-serve the client bundle.

Install

# GitHub (source, builds on install — grant the allowBuilds prompt once)
dsh plugin --profile web add github:superfish058/dsh-llm-proxy

# or from a local checkout
dsh plugin --profile web add C:/path/to/dsh-llm-proxy

Git-hosted installs run the package's prepare script (tsdown); pnpm ≥10 asks you to authorize builds once — copy the printed key into the profile's pnpm-workspace.yaml and re-run add. To skip the build allowance entirely, use a tarball instead:

pnpm pack        # → dsh-llm-proxy-1.0.0.tgz
dsh plugin --profile web add ./dsh-llm-proxy-1.0.0.tgz

Then restart dsh web (tray exit → start). No other change needed.

Config

Field	Default	Meaning
proxyHost	127.0.0.1	Proxy hostname/IP; does not have to be this machine.
proxyPort	7897	Proxy port (1–65535).
proxiedModels	[]	Model keys that route through the proxy, <providerId>/<modelId> (resolved against the llm-pi-ai / llm-deepseek model list); everything else stays direct.
retries	3	Max retry attempts on failure (transport error / 429 / 5xx), 0–10.
retryIntervalMs	1000	Delay between retry attempts, ms (0–60000).

The plugin shipped its model-level design as v1.0.0 (previous internal iterations used proxyUrl/routes/rateLimits; if you carried a user layer from those, delete the old llm-proxy: section from settings.yaml and configure the new fields).

Verify

After restart, dsh web logs show:

dsh-llm-proxy: global dispatcher → RetryAgent(RoutingDispatcher) (proxy=127.0.0.1:7897, proxiedHosts=[api.b.ai], retries=3×1000ms)

Then select the b.ai model in the model picker — streaming should work, and only its host (api.b.ai) goes through the proxy.

Test

npm test            # unit tests (routing + RetryAgent retry, no network)
npm run test:client # client-bundle build check (run npm run build first)
npm run test:smoke  # end-to-end against real undici + local servers (routing + 503 retry)
npm run test:global # proves the dispatcher also intercepts Node's GLOBAL fetch
npm run test:load   # plugin entry load with a fake cordis ctx