From 9fa6cf4b9dd1f4508ecb7edebf483be84f1b5e50 Mon Sep 17 00:00:00 2001 From: LiangshouX Date: Thu, 29 Oct 2026 17:42:22 +0800 Subject: [PATCH] fix(dsh): probe xAI on its own endpoint and wait for a saved backend - give xAI its own protocol: its credential carries an api host instead of a base URL, so the probe no longer demands LLM_BASE_URL for a configuration ReMe accepts - say that the xAI probe is a REST-level approximation of the gRPC SDK ReMe uses for that provider - hold the LLM test back while the chosen backend is not saved, because the probe reads the .env on disk --- integrations/dsh/docs/environment.md | 17 ++++++---- integrations/dsh/docs/environment_ZH.md | 13 +++++--- .../dsh/src/environment/client/locales.ts | 3 ++ .../dsh/src/environment/client/workspace.tsx | 16 +++++++++- .../dsh/src/environment/host/workspace.ts | 24 +++++++++++--- integrations/dsh/tests/env-tab.test.mjs | 31 +++++++++++++++++++ integrations/dsh/tests/env-workspace.test.mjs | 14 +++++++-- 7 files changed, 99 insertions(+), 19 deletions(-) diff --git a/integrations/dsh/docs/environment.md b/integrations/dsh/docs/environment.md index ad8a4462..26d97c76 100644 --- a/integrations/dsh/docs/environment.md +++ b/integrations/dsh/docs/environment.md @@ -104,13 +104,18 @@ DSH settings. cannot lose unrelated keys. - **Test LLM connection** runs a one-token request from inside the workspace with the selected interpreter, and adds a hint for `401`/`403` (key), `404` (model or base URL) and network errors. The request follows the - protocol of the `LLM_BACKEND` in the `.env`: `openai`, `dashscope`, `deepseek`, `moonshot` and `xai` use the + protocol of the `LLM_BACKEND` in the `.env`: `openai`, `dashscope`, `deepseek` and `moonshot` use the OpenAI-compatible `/chat/completions` (Bearer auth), `anthropic` uses `/v1/messages` (`x-api-key`), `gemini` - uses `/v1beta/models/:generateContent` (`x-goog-api-key`) and `ollama` uses `/api/chat` (a local server - needs no key). Gemini and Ollama are probed with their own endpoint, because ReMe's credentials for them do not - read `LLM_BASE_URL`; a stale value left in the `.env` cannot send the probe to a host the service never calls. - The base URL is used as the provider's own SDK builds it, so a base that already carries the version suffix - fails in the probe exactly as it fails in the service. + uses `/v1beta/models/:generateContent` (`x-goog-api-key`), `ollama` uses `/api/chat` (a local server + needs no key) and `xai` uses the OpenAI-compatible `/chat/completions` at `api.x.ai`. +- The base URL follows ReMe's own credentials: `LLM_BASE_URL` is read for the OpenAI-compatible providers and for + Anthropic, while Gemini, Ollama and xAI carry their own endpoint and ignore it. A stale value left in the `.env` + therefore cannot send the probe to a host the service never calls, and a configuration ReMe accepts is never + reported as broken. The base URL is used as the provider's own SDK builds it, so a base that already carries the + version suffix fails in the probe exactly as it fails in the service. +- ReMe reaches xAI through its gRPC SDK, so the xAI probe is a REST-level approximation of that channel. +- The probe reads the `.env` on disk, so a backend that is only picked in the form is not tested before the file is + saved; the button says so while it waits. - Drafts are bound to the workspace they were typed in: switching the workspace drops them, so a key typed for one workspace cannot be saved into another. diff --git a/integrations/dsh/docs/environment_ZH.md b/integrations/dsh/docs/environment_ZH.md index 08869590..a6af19df 100644 --- a/integrations/dsh/docs/environment_ZH.md +++ b/integrations/dsh/docs/environment_ZH.md @@ -84,11 +84,16 @@ ReMe 在服务启动时读取 workspace 目录下的 `.env`;标签页从不把 - 写入是原子且仅属主可读写的(POSIX 上为 0600);旧文件会保留为 `.env.bak`,因此一次失败的保存不会丢掉其它键。 - **测试 LLM 连接**在 workspace 内、用所选解释器发送一次只需 1 个 token 的请求,并对 `401`/`403` (密钥)、`404`(模型名或 Base URL)与网络错误给出提示。请求按 `.env` 里的 `LLM_BACKEND` 选择协议: - `openai`、`dashscope`、`deepseek`、`moonshot`、`xai` 走 OpenAI 兼容的 `/chat/completions`(Bearer 认证), + `openai`、`dashscope`、`deepseek`、`moonshot` 走 OpenAI 兼容的 `/chat/completions`(Bearer 认证), `anthropic` 走 `/v1/messages`(`x-api-key`),`gemini` 走 `/v1beta/models/:generateContent` - (`x-goog-api-key`),`ollama` 走 `/api/chat`(本地服务无需密钥)。Gemini 与 Ollama 一律使用服务商自有地址探测, - 因为 ReMe 为它们生成的凭据并不读取 `LLM_BASE_URL`;`.env` 里残留的旧值不会把探测引到服务根本不会访问的主机。 - Base URL 的拼接方式与服务商 SDK 完全一致,因此已经带版本后缀的地址在探测与服务中都会同样失败。 + (`x-goog-api-key`),`ollama` 走 `/api/chat`(本地服务无需密钥),`xai` 走 `api.x.ai` 上 OpenAI 兼容的 + `/chat/completions`。 +- Base URL 的来源与 ReMe 自身的凭据保持一致:`LLM_BASE_URL` 只对 OpenAI 兼容族与 Anthropic 生效,Gemini、Ollama + 与 xAI 自带地址、并不读取该变量。因此 `.env` 里残留的旧值不会把探测引到服务根本不会访问的主机,ReMe 能接受的 + 配置也不会被报成故障。Base URL 的拼接方式与服务商 SDK 完全一致,因此已经带版本后缀的地址在探测与服务中都会 + 同样失败。 +- ReMe 通过 gRPC SDK 连接 xAI,因此 xAI 的探测只是该通道在 REST 层面的近似。 +- 探测读取的是磁盘上的 `.env`,所以只在表单里切换后端而未保存时不会发起测试;等待期间按钮会给出提示。 - 草稿与输入它的 workspace 绑定:切换 workspace 会丢弃草稿,因此为某个 workspace 输入的密钥不会被保存到另一个。 ## 连接测试 diff --git a/integrations/dsh/src/environment/client/locales.ts b/integrations/dsh/src/environment/client/locales.ts index 2d7183c3..765b538a 100644 --- a/integrations/dsh/src/environment/client/locales.ts +++ b/integrations/dsh/src/environment/client/locales.ts @@ -94,6 +94,8 @@ export const envEn = { saveEnvFailed: "Saving failed", testLlmButton: "Test LLM connection", testingLlm: "Testing…", + testLlmSaveFirst: + "Save the .env first: the test uses the backend that is saved on disk.", llmOk: "LLM reachable ({ms} ms)", llmFailed: "LLM test failed", serviceHint: @@ -291,6 +293,7 @@ export const envZh: typeof envEn = { saveEnvFailed: "保存失败", testLlmButton: "测试 LLM 连接", testingLlm: "正在测试…", + testLlmSaveFirst: "请先保存 .env:测试使用的是磁盘上已保存的后端。", llmOk: "LLM 连接正常({ms} ms)", llmFailed: "LLM 测试失败", serviceHint: "启动或停止本地 ReMe 服务;DSH 关闭后服务仍会保持运行。", diff --git a/integrations/dsh/src/environment/client/workspace.tsx b/integrations/dsh/src/environment/client/workspace.tsx index 427aa41b..b7df96ff 100644 --- a/integrations/dsh/src/environment/client/workspace.tsx +++ b/integrations/dsh/src/environment/client/workspace.tsx @@ -415,6 +415,13 @@ export function WorkspaceSection({ }): JSX.Element { const [open, setOpen] = useState(false); const { env, llmTest, result } = workspace; + // The probe reads the saved .env, so a backend that is only picked in the + // form would be tested against the one still on disk. The button says so + // instead of quietly testing the old backend. + const savedBackend = + env.data?.entries.find((entry) => entry.key === "LLM_BACKEND")?.value ?? ""; + const backendChanged = + workspace.value("LLM_BACKEND").trim() !== savedBackend.trim(); return (
@@ -507,7 +514,9 @@ export function WorkspaceSection({ + {backendChanged ? ( + + {t("testLlmSaveFirst")} + + ) : null} {result.kind === "saved" ? ( diff --git a/integrations/dsh/src/environment/host/workspace.ts b/integrations/dsh/src/environment/host/workspace.ts index 884f757e..00973762 100644 --- a/integrations/dsh/src/environment/host/workspace.ts +++ b/integrations/dsh/src/environment/host/workspace.ts @@ -202,10 +202,11 @@ export interface LlmProtocol { * * ReMe's credentials decide this. The OpenAI-compatible providers and * Anthropic read `LLM_BASE_URL` from the workspace `.env`, so the probe - * reads it too. Gemini and Ollama ignore that variable and use their own - * endpoint, so the probe must use that one as well: a stale value left in - * the `.env` would otherwise send the probe to a host the service never - * uses, and report a working configuration as broken. + * reads it too. Gemini, Ollama and xAI ignore that variable — their + * credentials carry their own endpoint — so the probe must use that one as + * well: a stale value left in the `.env` would otherwise send the probe to + * a host the service never uses, and report a working configuration as + * broken. */ baseSource: "env" | "provider"; /** The provider's own endpoint; used when `baseSource` is `provider`. */ @@ -238,7 +239,6 @@ export const LLM_PROTOCOLS: Record = { dashscope: OPENAI_PROTOCOL, deepseek: OPENAI_PROTOCOL, moonshot: OPENAI_PROTOCOL, - xai: OPENAI_PROTOCOL, anthropic: { path: "/v1/messages", authHeader: "x-api-key", @@ -268,6 +268,20 @@ export const LLM_PROTOCOLS: Record = { baseSource: "provider", providerBase: "http://localhost:11434", }, + xai: { + // ReMe reaches xAI through its gRPC SDK, so this OpenAI-compatible + // endpoint is a connectivity approximation for that backend. It is still + // the documented REST API of the same provider and the same key, and it + // keeps the probe from reporting a working configuration as broken. + path: "/chat/completions", + authHeader: "Authorization", + authPrefix: "Bearer ", + headers: {}, + body: OPENAI_CHAT_BODY, + keyRequired: true, + baseSource: "provider", + providerBase: "https://api.x.ai/v1", + }, }; /** The backend ReMe itself falls back to when `.env` names none. */ diff --git a/integrations/dsh/tests/env-tab.test.mjs b/integrations/dsh/tests/env-tab.test.mjs index e83de809..a2f83723 100644 --- a/integrations/dsh/tests/env-tab.test.mjs +++ b/integrations/dsh/tests/env-tab.test.mjs @@ -1508,6 +1508,37 @@ test("a running job is restored for the selected interpreter", async (t) => { assert.equal(lookups[0].input.pythonPath, SUGGESTED); }); +test("the LLM test waits for the backend that is saved on disk", async (t) => { + const { renderer, restore } = await mount(); + t.after(restore); + await act(async () => { + buttonByText(renderer, envZh.expand).props.onClick(); + }); + // The saved backend is the one the probe would use. + assert.equal( + buttonByText(renderer, envZh.testLlmButton).props.disabled, + false, + ); + + // Picking another backend in the form does not change the file on disk, so + // the test would probe the old one. + const backend = renderer.root + .findAll((node) => node.type === "select") + .find((node) => node.props.value === "dashscope"); + assert.ok(backend, "the LLM backend select"); + await act(async () => { + backend.props.onChange({ target: { value: "anthropic" } }); + }); + assert.ok( + texts(renderer.toJSON()).join(" ").includes(envZh.testLlmSaveFirst), + "the button explains that the .env has to be saved first", + ); + assert.equal( + buttonByText(renderer, envZh.testLlmButton).props.disabled, + true, + ); +}); + test("a restored job is matched by the interpreter's own path semantics", async (t) => { // The Host reports the interpreter with its own spelling; a Windows path // folds case and separators, so the job is still the one on screen. diff --git a/integrations/dsh/tests/env-workspace.test.mjs b/integrations/dsh/tests/env-workspace.test.mjs index 0ad47720..ab868222 100644 --- a/integrations/dsh/tests/env-workspace.test.mjs +++ b/integrations/dsh/tests/env-workspace.test.mjs @@ -7,6 +7,7 @@ import { join } from "node:path"; import { envSummary, LLM_PROTOCOLS, + llmProbeSnippet, llmProtocolFor, maskSecret, mergeEnvFile, @@ -294,9 +295,10 @@ test("the probe builds a valid request for every backend", async (t) => { protocol.path.startsWith("/"), `${backend} path is appended to the base`, ); - assert.equal( - protocol.trimSuffix, - undefined, + const snippet = llmProbeSnippet(protocol); + assert.doesNotMatch( + snippet, + /base\[:|endswith/, `${backend} does not rewrite the configured base`, ); } @@ -347,6 +349,12 @@ test("a backend that ignores LLM_BASE_URL is probed with its own endpoint", asyn assert.match(stale, /localhost:11434/); assert.doesNotMatch(stale, /stale\.invalid/); + // xAI is the same case: its credential carries an api host, not a base URL. + const xai = await probe("xai", "https://stale.invalid"); + assert.match(xai, /api\.x\.ai/); + assert.doesNotMatch(xai, /stale\.invalid/); + assert.doesNotMatch(xai, /env\.get\("LLM_BASE_URL"/); + // Ollama needs neither a base URL nor a key. const ollama = await probe("ollama"); assert.match(ollama, /localhost:11434/);