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.
这个插件用于按 provider + model 维度选择 HTTP、HTTPS 或 SOCKS5 代理。它只改变请求经过的传输通道,不修改模型适配器,也不改写上游 baseURL;没有匹配到规则的请求仍然可以保持直连。
配置界面位于:
Settings → Plugins → Model Proxy
规则按照以下优先级匹配:
精确模型 > 模型前缀通配(例如 muse-*)> * 通配
代理地址在下方“代理主机”区域统一维护。规则只需要从下拉框选择一个主机,或选择 Direct(直连)。多个规则可以复用同一个代理主机,代理密码可以通过 credentialRef 引用 DSH credentials 服务,而不必写入 profile。
选择 Direct 时不会创建代理 Dispatcher,但规则中配置的固定 Header 仍会发送,因此“直连”和“添加 Header”可以同时使用。
每条规则都支持添加请求 Header,并可以为每个 Header 选择固定值或请求上下文中的内置变量。Header 按规则保存,同一个代理主机下的不同规则可以发送不同的 Header。
以 OpenCode 实际使用的会话 Header 为例:
- provider: opencode-go
model: "*"
proxyHostId: local-socks
enabled: true
headerValueSources:
x-opencode-session: sessionId上面的配置会在每次匹配请求时读取 GenerateOptions.sessionId,将当前会话 ID 写入 x-opencode-session。其他内置变量还包括:
provider:当前 Provider;model:当前模型;purpose:当前请求用途,例如compaction或session-title。
也可以在同一规则中组合使用:
headerValueSources:
x-opencode-session: sessionId
x-provider-name: provider
x-model-name: model
x-request-purpose: purpose如果上游要求一个固定值,则使用传统写法:
headers:
x-opencode-session: "fixed-session-value"x-opencode-session 的值应替换为当前 OpenCode 会话或上游中继要求的实际值。它是附加 Header,不是 API Token;不要把 Authorization 或 Proxy-Authorization 写进规则,认证信息应由适配器或凭据服务管理。
Header 规则如下:
- Header 值可以是固定字符串,也可以选择当前请求的
Session ID、Provider、Model或Purpose;没有对应上下文值时不会注入该 Header; - 固定 Header 值必须是单行字符串,不能包含 CR/LF;
- Header 名称必须符合 HTTP token 规则;
- 每条规则最多添加 10 个 Header;
- 请求中已经存在的同名 Header 优先,规则值不会覆盖适配器生成的值;
- Header 值不会写入调试日志,日志只显示 Header 名称;
- Header 会在代理模式和直连模式下发送。
在设置界面中,可以点击 Headers → + x-opencode-session 快速添加 OpenCode 会话 Header,再选择 Current session ID;也可以选择 Current provider、Current model 或 Current purpose。如果需要固定值,则选择 Fixed value 并填写实际内容。
- Per-rule routing:
{provider, model, proxyHostId, enabled}— specificity: exact model > prefixmuse-*>*. - Reusable proxy hosts: maintain named HTTP/HTTPS/SOCKS5 host profiles once; rules select a profile from a dropdown.
Directskips the proxy without disabling the rule. - Header value strategies: per-rule
headerscan use fixed values orheaderValueSourcesto selectsessionId,provider,model, orpurpose; sent in both direct and proxied modes; existing headers always win. - Purpose filter: optional per-rule
purpose(e.g.compaction) so chat goes through the proxy while background calls stay direct. - credentialRef: keep proxy passwords in the DSH credentials service instead of the profile config; reusable proxy hosts reference them by name (
user:passwordentries). Soft dependency — installs without it keep working. - Auto probe: newly configured proxies are connectivity-tested once (CONNECT/socks handshake, no model quota); results go to the host log.
- Protocols:
http://,https://(CONNECT viaundici.ProxyAgent),socks5:///socks5h://(via optionalsocks, tunnelled by undiciAgent+ custom connect). - Zero baseURL mutation — keep the real upstream.
- Live: change rules → next
llm/streamuses them; in-flight streams unaffected. - UI:
Settings → Plugins → Model Proxycard; edits commit live to the profile patch (hand-written yaml works too — see "Via file"). - Provider picker: dropdown groups user-configured providers first (derived from
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 onllm/adapters-updated); a "Custom…" entry keeps free text for wildcards (muse-*,*) or not-yet-installed providers. - Batch & grouped management: adding rules checks multiple models for one provider at once — one rule each, sharing the selected proxy host/purpose; the list groups cards by provider with group-level host selection, enable/disable-all, and delete (cross-provider grouping never affects match order).
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 socksVerify 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-proxyInstalling 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
name: dsh-plugin-model-proxy
config: {}Patch-file rules that trip people up:
- The file must parse as a top-level YAML array of patch entries. Comments
are fine; a file of only comments (no
- …rows and no[]) fails to parse and breaks every boot/dump that reads it. - The loader-entry specifier field is
name:(the package name), notmodule:. - New rows must be nested under
insert:— a bare- id:row means "patch an existing entry". - Only the host half is a loader entry. The browser half is discovered automatically
from the package's
dsh.clientdeclaration and served at/plugins/dsh-plugin-model-proxy/client.js. Do NOT add amodel-proxy/clientrow — it would run browser code inside the Node process. - Never mix this path with the profile-managed install above. Once the
dependency declares
dsh.bundle.patch,dsh plugin addinjects the host row automatically; keeping a hand-written copy in~/.dsh/cordis.patch.ymlinserts 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
The entry's config in the active profile patch (~/.dsh/profiles/<name>/cordis.patch.yml;
the settings UI writes edits to the same row):
- id: model-proxy
config:
enabled: true
debug: false
# Named connection profiles are managed once and reused by rules.
proxyHosts:
- id: local-socks
name: Local SOCKS
proxyUrl: socks5://127.0.0.1:1080
credentialRef: local-proxy-auth # optional credentials-service entry
- id: office-http
name: Office HTTP
proxyUrl: http://127.0.0.1:7890
defaultProxyHostId: "" # fallback when no rule matches; empty = direct
defaultProxy: "" # legacy inline fallback, kept for compatibility
rules:
- provider: opencode
model: muse-spark-1.2-contributor
proxyHostId: local-socks
enabled: true
- provider: opencode-go
model: "*"
proxyHostId: office-http
enabled: true
headers:
x-opencode-session: <fixed-value> # fixed mode
# Or use the current request session ID:
# headerValueSources:
# x-opencode-session: sessionId
- provider: opencode
model: "*"
# no proxyHostId / proxyUrl = Direct
enabled: trueproxyHostId references an entry in proxyHosts; omitting it selects direct
mode. proxyUrl and per-rule credentialRef remain accepted for legacy
hand-written rules and are migrated into named hosts when the card is saved.
socks5h:// resolves DNS at the proxy. Field changes commit live (volatile
Config) without restarting the profile.
Per-rule headers hold fixed extra request headers. A headerValueSources map
can select sessionId, provider, model, or purpose for a header name; the
value is read from the current request context. Both forms are sent whether the
selected route is proxied or direct. A header already present on the outgoing
request (e.g. Authorization, attribution user-agent) always wins and is never
overwritten. Authorization / Proxy-Authorization cannot be set via rules.
Header values never appear in logs (debug prints names only).
- The plugin's volatile Config fields are the
model-proxysettings section (the loader entry id); edits commit live. - Wraps
globalThis.fetchreversibly. - Listens on
llm/streamwaterfall, resolves the selected reusable host for(provider, model)viaAsyncLocalStorage, then injects adispatcherinto the adapter'sfetch— for proxied requests it routes throughundici.fetch(Node's native global fetch ignores custom dispatchers), withundici.ProxyAgentforhttp(s)and an undiciAgent+ socksconnect(with TLS) forsocks5/h. Direct rules still receive their fixed headers through the native fetch path. - Browser registers
settings.plugin.itemwith keymodel-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 buildFully 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 typecheckMIT