dsh-gateway-provider 1.1.1 → 1.1.2

package/README.md

@@ -92,7 +92,7 @@ the first start where the plugin loads. The frequently used ones:
 
 | Key | Default | Meaning |
 | --- | --- | --- |
-| `baseURL` | `https://api.newapi.ai` | Your gateway's base URL. Env fallbacks: `NEWAPI_BASE_URL`, `NEWAPI_API_URL`. |
+| `baseURL` | `https://api.newapi.ai` | Your gateway's base URL. Env fallbacks: `NEWAPI_BASE_URL`, `NEWAPI_API_URL`. A trailing `/v1` or `/v1beta` is tolerated (stripped for discovery); Gemini requests always go to `{base}/v1beta/models/…`. |
 | `apiKeyEnv` | `NEWAPI_API_KEY` | Which env/credential variable holds the key. |
 | `label` | `NewAPI` | Route label shown in the picker. |
 | `flavor` | `newapi` | Template label only (`newapi` / `litellm` / `higress` / `openai-compatible` / `custom`). |
@@ -103,7 +103,7 @@ the first start where the plugin loads. The frequently used ones:
 | `sortModelsByRelease` | `true` | Newest models first. |
 | `catalogMode` | `auto` | `v1` (`/v1/models` only) / `management` (newapi user API) / `auto`. |
 | `endpointPriority` | responses → anthropic → openai → gemini | Which protocol to prefer when a model supports several. |
-| `openaiURL` / `responsesURL` / `anthropicURL` | — | Fully-custom gateways only: per-protocol endpoint URLs; unset = that protocol off. |
+| `openaiURL` / `responsesURL` / `anthropicURL` / `geminiURL` | — | Fully-custom gateways only: per-protocol endpoint URLs; unset = that protocol off. `geminiURL` may be a bare host, `…/v1beta`, or `…/v1beta/models` — the `/v1beta` segment is ensured (pi-ai expects it in the base). |
 | `maxTokens` / `defaultContextWindow` | `32768` / `128000` | Fallbacks when models.dev has no data. |
 | `streamIdleTimeoutMs` | `600000` | Idle timeout while streaming. |
 | `headers` | — | Extra HTTP headers sent to the gateway. |
@@ -116,7 +116,8 @@ the first start where the plugin loads. The frequently used ones:
 | `401` / auth errors on every request | Key missing or wrong: check the badge in Settings → Gateway Models, or `NEWAPI_API_KEY` in `~/.dsh/.credentials.yaml`. |
 | A model's context window looks wrong | models.dev had no match. Edit the model on the settings page (or a `models:` override). |
 | Wrong format answers / tool calls flaky for one model | That model is routed over a protocol it handles poorly. Pin `protocol` on the model (`openai`, `openai-response`, `anthropic`, `gemini`). |
-| Custom gateway with separate endpoints | Use `flavor: "custom"` and set `openaiURL` / `responsesURL` / `anthropicURL` explicitly. |
+| Custom gateway with separate endpoints | Use `flavor: "custom"` and set `openaiURL` / `responsesURL` / `anthropicURL` / `geminiURL` explicitly. |
+| Gemini model returns HTML / `Unexpected token '<'` (issue #4) | The gateway serves Gemini under `/v1beta`; fixed in 1.1.2 — a plain `baseURL` now sends Gemini to `{base}/v1beta/models/…` (a `baseURL` ending in `/v1beta` also works). No older-version workaround exists; update the plugin. |
 
 ## How it works (one minute version)
 

package/docs/README.zh.md

@@ -56,7 +56,7 @@ dsh --profile web
 
 | 键 | 默认值 | 含义 |
 | --- | --- | --- |
-| `baseURL` | `https://api.newapi.ai` | 网关 base URL。环境变量兜底：`NEWAPI_BASE_URL`、`NEWAPI_API_URL`。 |
+| `baseURL` | `https://api.newapi.ai` | 网关 base URL。环境变量兜底：`NEWAPI_BASE_URL`、`NEWAPI_API_URL`。结尾带 `/v1` 或 `/v1beta` 均可（发现时会剥掉）；Gemini 请求固定发往 `{base}/v1beta/models/…`。 |
 | `apiKeyEnv` | `NEWAPI_API_KEY` | key 存在哪个环境/凭据变量里。 |
 | `label` | `NewAPI` | 选择器里显示的路由名。 |
 | `flavor` | `newapi` | 仅作模板标签（`newapi` / `litellm` / `higress` / `openai-compatible` / `custom`）。 |
@@ -67,7 +67,7 @@ dsh --profile web
 | `sortModelsByRelease` | `true` | 最新模型排前面。 |
 | `catalogMode` | `auto` | `v1`（只用 `/v1/models`）/ `management`（newapi 用户 API）/ `auto`。 |
 | `endpointPriority` | responses → anthropic → openai → gemini | 模型支持多种协议时的优先序。 |
-| `openaiURL` / `responsesURL` / `anthropicURL` | — | 仅完全自定义网关：按协议分别填端点 URL；不填 = 该协议关闭。 |
+| `openaiURL` / `responsesURL` / `anthropicURL` / `geminiURL` | — | 仅完全自定义网关：按协议分别填端点 URL；不填 = 该协议关闭。`geminiURL` 可填裸主机名、`…/v1beta` 或 `…/v1beta/models` —— `/v1beta` 段会自动确保（pi-ai 要求它含在 base 里）。 |
 | `maxTokens` / `defaultContextWindow` | `32768` / `128000` | models.dev 无数据时的兜底。 |
 | `streamIdleTimeoutMs` | `600000` | 流式空闲超时。 |
 | `headers` | — | 发往网关的额外 HTTP 头。 |
@@ -80,7 +80,8 @@ dsh --profile web
 | 每次请求都 `401` / 鉴权错误 | key 缺失或不对：看 Settings → Gateway Models 的徽标，或 `~/.dsh/.credentials.yaml` 里的 `NEWAPI_API_KEY`。 |
 | 某模型的上下文窗口看着不对 | models.dev 没匹配上。在设置页编辑该模型（或写 `models:` 覆盖）。 |
 | 单个模型工具调用不稳、格式怪 | 该模型被路由到了它处理不好的协议。在模型上钉死 `protocol`（`openai`、`openai-response`、`anthropic`、`gemini`）。 |
-| 自定义网关各协议端点分开 | 用 `flavor: "custom"`，显式填 `openaiURL` / `responsesURL` / `anthropicURL`。 |
+| 自定义网关各协议端点分开 | 用 `flavor: "custom"`，显式填 `openaiURL` / `responsesURL` / `anthropicURL` / `geminiURL`。 |
+| Gemini 模型返回 HTML / `Unexpected token '<'`（issue #4） | 网关把 Gemini 挂在 `/v1beta` 下；1.1.2 已修复 —— 普通 `baseURL` 也会把 Gemini 请求发到 `{base}/v1beta/models/…`（`baseURL` 以 `/v1beta` 结尾同样可用）。旧版本没有任何绕过手段，请升级插件。 |
 
 ## 工作原理（一分钟版）
 

package/index.js

@@ -4,7 +4,8 @@
  * Registers one or more `gateway:*` provider routes on `ctx.llm`, each backed
  * by an OpenAI-compatible `/v1/models` gateway (newapi, LiteLLM, Higress, or
  * any generic gateway), or by a fully-custom gateway whose per-protocol
- * endpoint URLs (`openaiURL` / `responsesURL` / `anthropicURL`) are written
+ * endpoint URLs (`openaiURL` / `responsesURL` / `anthropicURL` / `geminiURL`)
+ * are written
  * out in full. Backwards-compatible with the original single-connection
  * `newapi` route: the legacy flat fields (`baseURL`, `apiKeyEnv`, …) build a
  * default gateway, while the `gateways` array adds more.
@@ -121,6 +122,12 @@ const GatewaySchema = z.object({
   responsesURL: coercedString(),
   /** Full Anthropic messages endpoint URL (custom template). */
   anthropicURL: coercedString(),
+  /**
+   * Gemini native endpoint base (custom template): a bare host, a base
+   * ending in `/v1beta`, or one with a trailing `/models`. The `/v1beta`
+   * segment is ensured — pi-ai's Gemini client expects it in the base.
+   */
+  geminiURL: coercedString(),
   /** Environment-variable name (credential ref) holding the API key. */
   apiKeyEnv: coercedString().role("credential-ref"),
   /** Gateway type label (newapi/litellm/higress/openai-compatible/custom). */
@@ -149,6 +156,8 @@ export const Config = z.object({
   responsesURL: coercedString(),
   /** Full Anthropic messages endpoint URL (custom template for the default gateway). */
   anthropicURL: coercedString(),
+  /** Gemini native endpoint base (custom template for the default gateway). */
+  geminiURL: coercedString(),
   /** Gateway type label shown in the UI (newapi/litellm/higress/openai-compatible/custom). */
   flavor: z.union(FLAVORS).default("newapi"),
   /** models.dev catalog URL (any fetch-able URL; file: works for offline mirrors). */
@@ -193,7 +202,8 @@ export const Config = z.object({
 /**
  * Normalize one gateway entry into the connection facts the adapter needs.
  * Per-gateway fields fall back to the shared defaults. Protocol URL fields
- * (`openaiURL` / `responsesURL` / `anthropicURL`, the "custom" template)
+ * (`openaiURL` / `responsesURL` / `anthropicURL` / `geminiURL`, the "custom"
+ * template)
  * redefine the gateway's protocol surface: each configured URL becomes the
  * exact SDK base for its wire protocol, unconfigured protocols are
  * unavailable, and the plain `baseURL` is no longer inherited from the
@@ -290,6 +300,7 @@ export function resolveGateways(config, environment) {
       openaiURL: config.openaiURL,
       responsesURL: config.responsesURL,
       anthropicURL: config.anthropicURL,
+      geminiURL: config.geminiURL,
     }, defaults, PROVIDER, defaultLabel) },
   ];
   // Additional gateways from the array.

package/lib/catalog.js

@@ -122,15 +122,16 @@ async function fetchManagementModels(baseURL, apiKey, userId, signal) {
 }
 
 /**
- * Strip trailing slashes and a trailing version segment (`/v1`, `/v2`, …)
- * from a gateway base URL. The catalog always appends its own full paths
- * (`/v1/models`, `/api/user/models`), so a user-supplied base that already
- * carries `/v1` (the OpenAI-SDK convention, which the chat path also accepts)
- * must not leak through as `/v1/v1/models`. Mirrors the derivation the custom
+ * Strip trailing slashes and a trailing version segment (`/v1`, `/v2`, and
+ * the Gemini-style `/v1beta`) from a gateway base URL. The catalog always
+ * appends its own full paths (`/v1/models`, `/api/user/models`), so a
+ * user-supplied base that already carries a version (the SDK conventions,
+ * which the chat path also accepts) must not leak through as
+ * `/v1/v1/models` or `/v1beta/v1/models`. Mirrors the derivation the custom
  * (protocol-URL) template performs in `protocols.js`.
  */
 function stripVersion(baseURL) {
-  return String(baseURL ?? "").replace(/\/+$/, "").replace(/\/v\d+$/, "");
+  return String(baseURL ?? "").replace(/\/+$/, "").replace(/\/v\d+(?:beta)?$/i, "");
 }
 
 /**

package/lib/client.js

@@ -26,7 +26,7 @@ window.__ModuleLoader__.load({
 		var CATALOG_MODES = ["auto", "v1", "management"];
 		var PROTOCOLS = ["openai", "openai-response", "anthropic", "gemini"];
 		/** Gateway template fields (suffix "URL") → card protocol tag. */
-		var PROTOCOL_URL_KEYS = ["openaiURL", "responsesURL", "anthropicURL"];
+		var PROTOCOL_URL_KEYS = ["openaiURL", "responsesURL", "anthropicURL", "geminiURL"];
 
 		// ---- i18n (manual zh/en switch; no locale registration needed) ----
 		var ZH = (typeof navigator !== "undefined" && navigator.language && navigator.language.startsWith("zh"));
@@ -65,19 +65,21 @@ window.__ModuleLoader__.load({
 			tplHintLitellm: "LiteLLM Proxy：/v1/chat/completions 与 /v1/responses，经 /v1/models 自动发现",
 			tplHintHigress: "Higress AI 网关：OpenAI 格式为主，新版本同端口支持 Anthropic /v1/messages",
 			tplHintOpenaiCompatible: "任意 OpenAI 兼容网关：填一个基础地址，各协议路径自动派生",
-			tplHintCustom: "三种协议地址分别填写完整 URL；留空的协议不启用，模型仅以已启用协议请求",
+			tplHintCustom: "各协议地址分别填写（Gemini 可只填主机名，自动补 /v1beta）；留空的协议不启用，模型仅以已启用协议请求",
 			tplNewapi: "NewAPI", tplLitellm: "LiteLLM", tplHigress: "Higress",
 			tplOpenaiCompatible: "OpenAI 兼容", tplCustom: "完全自定义",
-			openaiURL: "OpenAI 兼容地址", responsesURL: "Responses 地址", anthropicURL: "Anthropic 地址",
+			openaiURL: "OpenAI 兼容地址", responsesURL: "Responses 地址", anthropicURL: "Anthropic 地址", geminiURL: "Gemini 地址",
 			openaiURLHint: "完整地址，以 /chat/completions 结尾（也可只填到 /v1）",
 			responsesURLHint: "完整地址，以 /responses 结尾（也可只填到 /v1）",
 			anthropicURLHint: "完整地址，以 /v1/messages 结尾",
+			geminiURLHint: "以 /v1beta 结尾（也可只填主机名，自动补 /v1beta）",
 			customNeedOne: "至少填写一个协议地址",
 			customURLPlaceholder: "https://gw.example.com/v1/chat/completions",
 			responsesURLPlaceholder: "https://gw.example.com/v1/responses",
 			anthropicURLPlaceholder: "https://gw.example.com/v1/messages",
+			geminiURLPlaceholder: "https://gw.example.com/v1beta",
 			catalogHint: "auto：优先 /v1/models，鉴权失败时回退管理 API",
-			baseURLClearHint: "留空则回退到环境变量 NEWAPI_BASE_URL 或官方云端地址；结尾带不带 /v1 均可",
+			baseURLClearHint: "留空则回退到环境变量 NEWAPI_BASE_URL 或官方云端地址；结尾带不带 /v1 或 /v1beta 均可",
 			labelClearHint: "留空则回退为 gateway:<id>",
 			keyEnvHint: "留空则继承默认网关的 Key（NEWAPI_API_KEY）",
 			keyKeepHint: "留空保持不变；填写则覆盖存储的 Key",
@@ -125,19 +127,21 @@ window.__ModuleLoader__.load({
 			tplHintLitellm: "LiteLLM proxy: /v1/chat/completions and /v1/responses; discovery via /v1/models",
 			tplHintHigress: "Higress AI gateway: OpenAI format; newer versions also serve Anthropic /v1/messages",
 			tplHintOpenaiCompatible: "Any OpenAI-compatible gateway: one base address, protocol paths derived automatically",
-			tplHintCustom: "Enter each protocol's full URL; empty protocols stay disabled and models use enabled protocols only",
+			tplHintCustom: "Enter each protocol's URL (Gemini accepts a bare host; /v1beta is added automatically); empty protocols stay disabled and models use enabled protocols only",
 			tplNewapi: "NewAPI", tplLitellm: "LiteLLM", tplHigress: "Higress",
 			tplOpenaiCompatible: "OpenAI-compatible", tplCustom: "Fully custom",
-			openaiURL: "OpenAI-compatible URL", responsesURL: "Responses URL", anthropicURL: "Anthropic URL",
+			openaiURL: "OpenAI-compatible URL", responsesURL: "Responses URL", anthropicURL: "Anthropic URL", geminiURL: "Gemini URL",
 			openaiURLHint: "Full address ending in /chat/completions (a /v1 base also works)",
 			responsesURLHint: "Full address ending in /responses (a /v1 base also works)",
 			anthropicURLHint: "Full address ending in /v1/messages",
+			geminiURLHint: "Ending in /v1beta (a bare host also works; /v1beta is added)",
 			customNeedOne: "Fill at least one protocol URL",
 			customURLPlaceholder: "https://gw.example.com/v1/chat/completions",
 			responsesURLPlaceholder: "https://gw.example.com/v1/responses",
 			anthropicURLPlaceholder: "https://gw.example.com/v1/messages",
+			geminiURLPlaceholder: "https://gw.example.com/v1beta",
 			catalogHint: "auto: prefer /v1/models, fall back to the management API on auth failure",
-			baseURLClearHint: "Empty falls back to NEWAPI_BASE_URL or the public cloud default; a trailing /v1 is optional",
+			baseURLClearHint: "Empty falls back to NEWAPI_BASE_URL or the public cloud default; a trailing /v1 or /v1beta is optional",
 			labelClearHint: "Empty falls back to gateway:<id>",
 			keyEnvHint: "Empty inherits the default gateway's key (NEWAPI_API_KEY)",
 			keyKeepHint: "Empty keeps the stored key; type to overwrite",
@@ -166,6 +170,7 @@ window.__ModuleLoader__.load({
 			openaiURL: { label: T.openaiURL, hint: T.openaiURLHint, ph: T.customURLPlaceholder, tag: "openai" },
 			responsesURL: { label: T.responsesURL, hint: T.responsesURLHint, ph: T.responsesURLPlaceholder, tag: "responses" },
 			anthropicURL: { label: T.anthropicURL, hint: T.anthropicURLHint, ph: T.anthropicURLPlaceholder, tag: "anthropic" },
+			geminiURL: { label: T.geminiURL, hint: T.geminiURLHint, ph: T.geminiURLPlaceholder, tag: "gemini" },
 		};
 
 		/** Protocol tags a gateway card shows (configured URL fields only). */
@@ -346,6 +351,7 @@ window.__ModuleLoader__.load({
 				openaiURL: protoOf(isDefault ? props.snapshot : gw, "openaiURL"),
 				responsesURL: protoOf(isDefault ? props.snapshot : gw, "responsesURL"),
 				anthropicURL: protoOf(isDefault ? props.snapshot : gw, "anthropicURL"),
+				geminiURL: protoOf(isDefault ? props.snapshot : gw, "geminiURL"),
 				apiKey: "",
 			};
 			var _v = React.useState(initial); var values = _v[0]; var setValues = _v[1];

package/lib/pi-provider.js

@@ -52,17 +52,30 @@ const PROTOCOLS = {
  * `/responses`. A gateway baseURL without `/v1` makes the SDK request
  * `{base}/responses`, which the gateway answers with its HTML UI (HTTP 200,
  * content-type text/html) instead of an SSE stream, so the SDK receives zero
- * events. Anthropic and Google protocols build their own full paths, so they
- * need no `/v1` appended.
+ * events.
+ *
+ * pi-ai's Gemini client follows the same convention with `/v1beta`: whenever
+ * a model carries a baseUrl it sets `httpOptions.apiVersion = ""` (its own
+ * Google provider ships `…googleapis.com/v1beta` as the base), so the version
+ * segment must come from the base itself. Without it the SDK requests
+ * `{base}/models/{model}:generateContent`, which gateways serving Gemini
+ * under `/v1beta` (e.g. new-api) answer with their HTML UI fallback.
+ * The Anthropic SDK builds its own full path, so it needs no version appended.
  * @param baseURL - the gateway base URL as configured.
  * @param api - the pi-ai API identifier the model will use.
- * @returns the baseURL the OpenAI SDK should point at.
+ * @returns the baseURL the SDK should point at.
  */
-function sdkBaseURL(baseURL, api) {
+export function sdkBaseURL(baseURL, api) {
   const trimmed = String(baseURL ?? "").replace(/\/+$/, "");
   if (api === "openai-completions" || api === "openai-responses") {
     if (!/\/v\d+$/.test(trimmed)) return `${trimmed}/v1`;
   }
+  // pi-ai pins apiVersion = "" when a baseUrl is set, so the version segment
+  // must already live in the base (issue #4). Case-insensitive, matching the
+  // anthropic branch's tolerance in protocols.js.
+  if (api === "google-generative-ai" && !/\/v\d+beta$/i.test(trimmed)) {
+    return `${trimmed}/v1beta`;
+  }
   return trimmed;
 }
 

package/lib/protocols.js

@@ -24,6 +24,10 @@ const ENDPOINT_SUFFIXES = {
   "openai-completions": "/chat/completions",
   "openai-responses": "/responses",
   "anthropic-messages": "/v1/messages",
+  // Gemini has no fixed endpoint suffix (the path embeds the model id:
+  // `{base}/v1beta/models/{model}:generateContent`), so only a trailing
+  // `/models` — copied from the docs' list URL — is tolerated.
+  "google-generative-ai": "/models",
 };
 
 /** Trim a configured URL and strip one known endpoint suffix if present. */
@@ -36,14 +40,16 @@ function stripSuffix(url, suffix) {
 
 /**
  * Derive the per-protocol facts from fully-qualified endpoint URLs
- * (`openaiURL` / `responsesURL` / `anthropicURL`). Unset URLs mean the
- * protocol is unavailable on this gateway ("没填的就是没有"): the returned
- * `availableTypes` is exactly the configured set.
+ * (`openaiURL` / `responsesURL` / `anthropicURL` / `geminiURL`). Unset URLs
+ * mean the protocol is unavailable on this gateway ("没填的就是没有"): the
+ * returned `availableTypes` is exactly the configured set.
  *
  * Each URL may be the complete endpoint address (`…/v1/chat/completions`),
  * an SDK base (`…/v1`), or — for anthropic — a bare host; the derivation
  * strips the known suffix and re-adds the version segment where the SDK
- * expects it.
+ * expects it. Gemini URLs accept a bare host, a base ending in `/v1beta`,
+ * or one with a trailing `/models`; the `/v1beta` segment is always ensured
+ * (pi-ai's Gemini client disables its own apiVersion when baseUrl is set).
  * @param urls - the gateway's protocol URL fields (any may be absent).
  * @returns `{ apiBases, availableTypes, catalogBase }` where apiBases maps
  *   pi-ai API id → SDK base URL, availableTypes lists gateway endpoint
@@ -80,7 +86,19 @@ export function deriveProtocolURLs(urls) {
       availableTypes.push("anthropic");
     }
   }
+  // Gemini: pi-ai's client pins `apiVersion = ""` whenever the model carries
+  // a baseUrl (mirroring its builtin Google provider, whose baseUrl is
+  // `https://generativelanguage.googleapis.com/v1beta`), so the version
+  // segment must live in the base. Accept a bare host (append `/v1beta`), a
+  // base already ending in `/v1beta`, or one with a trailing `/models`.
+  const gm = stripSuffix(urls.geminiURL, ENDPOINT_SUFFIXES["google-generative-ai"]);
+  if (gm !== undefined) {
+    apiBases["google-generative-ai"] = /\/v\d+beta$/i.test(gm) ? gm : `${gm}/v1beta`;
+    availableTypes.push("gemini");
+  }
   if (availableTypes.length === 0) return undefined;
+  // Gemini has no OpenAI-shaped `/v1/models` list endpoint, so only the
+  // OpenAI-style URLs contribute the discovery base.
   for (const api of ["openai-completions", "openai-responses"]) {
     if (apiBases[api] !== undefined) { catalogBase = apiBases[api].replace(/\/v\d+$/, ""); break; }
   }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "dsh-gateway-provider",
-  "version": "1.1.1",
+  "version": "1.1.2",
   "description": "Generic LLM gateway model provider plugin for DeepSeek Harness: mounts newapi / LiteLLM / Higress / any OpenAI-compatible gateway as provider routes, auto-discovers the model list, enriches every model with models.dev parameters, and dispatches each model over its native wire protocol (OpenAI / Anthropic / Gemini) via the pi-ai SDK",
   "repository": {
     "type": "git",