wtagent 0.2.2 → 0.3.0

package/src/browser/grok-web-adapter.js

@@ -0,0 +1,145 @@
+import { BaseWebAdapter, firstVisible } from "./base-web-adapter.js";
+import { isUsageLimitNotice } from "../shared/usage-limit.js";
+
+export { isConnectionLostError } from "./base-web-adapter.js";
+
+const GROK_URL = "https://grok.com/";
+
+export class GrokWebAdapter extends BaseWebAdapter {
+  constructor(options = {}) {
+    super({
+      ...options,
+      baseUrl: options.baseUrl ?? GROK_URL,
+      providerName: "Grok",
+    });
+  }
+
+  conversationUrlPattern() {
+    return /^\/c\//;
+  }
+
+  composerLocators() {
+    return [
+      this.page.locator("main textarea").first(),
+      this.page.locator("textarea[aria-label]").first(),
+      this.page.locator('[role="textbox"]').first(),
+    ];
+  }
+
+  sendButtonLocators() {
+    return [
+      this.page.locator('button[aria-label*="send" i]'),
+      this.page.locator('button[aria-label*="发送"]'),
+      this.page.locator('button[type="submit"]'),
+    ];
+  }
+
+  stopButtonLocators() {
+    return [
+      this.page.locator('button[aria-label*="停止模型响应"]'),
+      this.page.locator('button[aria-label*="stop" i]'),
+    ];
+  }
+
+  newConversationControls() {
+    return [
+      this.page.locator('[data-testid="new-chat"]'),
+      this.page.locator('a[href="/"]'),
+    ];
+  }
+
+  loginControlLocators() {
+    return [
+      this.page.getByRole("button", {
+        name: /^(log in|sign in|登录|登入|ログイン|로그인)$/i,
+      }),
+      this.page.getByRole("link", {
+        name: /^(log in|sign in|登录|登入|ログイン|로그인)$/i,
+      }),
+    ];
+  }
+
+  authUrlPattern() {
+    return /^\/(?:sign-in|signin|login|auth)(?:\/|$)/i;
+  }
+
+  authTextPattern() {
+    return /sign in to grok|log in to grok|continue with x|continue with google|登录 Grok|登入 Grok/i;
+  }
+
+  assistantMessages() {
+    return this.page.locator('[data-testid="assistant-message"]');
+  }
+
+  userMessages() {
+    return this.page.locator(
+      '[data-testid="user-message"], [role="article"][aria-label="You"]',
+    );
+  }
+
+  conversationMessages() {
+    return this.page.locator(
+      '[data-testid="user-message"], [data-testid="assistant-message"], '
+        + '[role="article"][aria-label="You"]',
+    );
+  }
+
+  async messageIdentity(message) {
+    const testId = await message.getAttribute("data-testid").catch(() => null);
+    const aria = await message.getAttribute("aria-label").catch(() => null);
+    const scope = testId === "user-message" || aria === "You"
+      ? this.userMessages()
+      : this.assistantMessages();
+    return { id: null, turn: await scope.count().catch(() => 0) };
+  }
+
+  isNewAssistantIdentity({ turn, text }) {
+    if (turn != null && turn > (this.assistantCountBeforeSend ?? 0)) {
+      return true;
+    }
+    const previous = String(this.lastAssistantTextBeforeSend ?? "");
+    return Boolean(previous && String(text ?? "") !== previous);
+  }
+
+  async assistantText(message) {
+    const response = message.locator(".response-content-markdown");
+    if (await response.count().catch(() => 0) > 0) {
+      return await response.last().innerText().catch(() => "");
+    }
+    return await message.innerText().catch(() => "");
+  }
+
+  hasReliableCompletionSignal() {
+    return true;
+  }
+
+  async isAssistantGenerating() {
+    return Boolean(await firstVisible(this.stopButtonLocators()));
+  }
+
+  async findUsageLimitMarker(message) {
+    const text = await message.innerText().catch(() => "");
+    if (!isUsageLimitNotice(text)) return null;
+    const control = await firstVisible([
+      message.getByRole("button", { name: /retry|try again|upgrade|重试|升级/i }),
+      this.page.getByRole("button", { name: /retry|try again|upgrade|重试|升级/i }),
+    ]);
+    return control ? text.trim().slice(0, 120) : null;
+  }
+
+  async findGenerationErrorMarker(message) {
+    const text = await message.innerText().catch(() => "");
+    if (!/something went wrong|failed to generate|generation failed|try again|生成失败|出错了/i.test(text)) {
+      return null;
+    }
+    const control = await firstVisible([
+      message.getByRole("button", { name: /retry|try again|重试|重新生成/i }),
+      this.page.getByRole("button", { name: /retry|try again|重试|重新生成/i }),
+    ]);
+    return control ? text.trim().slice(0, 120) : null;
+  }
+
+  deadRequestGraceMultiplier() {
+    return 3;
+  }
+}

package/src/browser/kimi-web-adapter.js

@@ -19,9 +19,8 @@ const KIMI_URL = "https://www.kimi.com/";
 //   - assistant replies may include a `.thinking-container` reasoning block
 //     before the answer; assistantText reads the answer markdown outside it
 //   - a live conversation URL is /chat/<uuid>
-//   - a model switcher (`.current-model`) opens a Naive-UI popover
-//     (`.models-container .model-item`); WTAgent defaults it to K3 — see
-//     selectMode.
+// Model choice is intentionally left to the user on kimi.com; the adapter does
+// not inspect or change the site's current model selection.
 export class KimiWebAdapter extends BaseWebAdapter {
   constructor(options = {}) {
     super({
@@ -211,75 +210,4 @@ export class KimiWebAdapter extends BaseWebAdapter {
     return 5;
   }
 
-  // Kimi has a model switcher (快速 / K3 / K3 集群). The registry's defaultMode
-  // "k3" asks to switch to K3 ("擅长对话与 Agent 任务，全能旗舰") on every fresh
-  // conversation, silently. Any other value keeps the current model.
-  //
-  // The switcher is `.current-model`; it opens a Naive-UI popover whose rows are
-  // `.models-container .model-item`, the selected one carrying `checked`. After
-  // selecting, the switcher label starts with the model name (e.g. "K3 进阶").
-  // Best-effort and non-throwing, mirroring runModeSelection's contract.
-  async selectMode(mode) {
-    this.requirePage();
-    if (mode !== "k3") {
-      return { status: "skipped", requested: mode, attempts: 0 };
-    }
-
-    const switcher = this.page.locator(".current-model").first();
-    // The switcher can mount a beat after the composer on a fresh conversation;
-    // wait briefly before concluding it is absent.
-    await switcher.waitFor({ state: "visible", timeout: 10_000 }).catch(() => null);
-    if (await switcher.count().catch(() => 0) === 0) {
-      await this.writeDiagnostics("kimi-model-switcher-not-found");
-      return {
-        status: "switcher_not_found",
-        requested: mode,
-        attempts: 0,
-        reason: "Model switcher was not found.",
-      };
-    }
-
-    // Already on K3? The switcher label starts with "K3" (but not "K3 集群").
-    const label = (await switcher.innerText().catch(() => "")).trim();
-    if (/^K3(?!\s*集群)/.test(label)) {
-      return {
-        status: "already",
-        requested: mode,
-        selectedLabel: "K3",
-        attempts: 0,
-        reason: "Already using K3.",
-      };
-    }
-
-    await switcher.click({ timeout: 5_000 }).catch(() => null);
-    await this.page.locator(".models-container .model-item")
-      .first().waitFor({ state: "visible", timeout: 5_000 }).catch(() => null);
-
-    // Click the row whose title line is exactly "K3" (not "快速" / "K3 集群").
-    const k3 = this.page.locator(".models-container .model-item").filter({
-      hasText: /^K3(?!\s*集群)/,
-    }).first();
-    const clicked = await k3.count().catch(() => 0) > 0
-      && await k3.click({ timeout: 5_000 }).then(() => true).catch(() => false);
-    await this.page.waitForTimeout(500);
-
-    const after = (await switcher.innerText().catch(() => "")).trim();
-    if (clicked && /^K3(?!\s*集群)/.test(after)) {
-      return {
-        status: "select",
-        requested: mode,
-        selectedLabel: "K3",
-        attempts: 1,
-        reason: "Selected K3.",
-      };
-    }
-    await this.page.keyboard.press("Escape").catch(() => null);
-    await this.writeDiagnostics("kimi-mode-k3-unresolved");
-    return {
-      status: "unresolved",
-      requested: mode,
-      attempts: 1,
-      reason: "Could not confirm K3 was selected.",
-    };
-  }
 }

package/src/browser/provider-registry.js

@@ -2,6 +2,7 @@ import { ChatGPTWebAdapter } from "./chatgpt-web-adapter.js";
 import { ClaudeWebAdapter } from "./claude-web-adapter.js";
 import { DeepSeekWebAdapter } from "./deepseek-web-adapter.js";
 import { GeminiWebAdapter } from "./gemini-web-adapter.js";
+import { GrokWebAdapter } from "./grok-web-adapter.js";
 import { KimiWebAdapter } from "./kimi-web-adapter.js";
 import { GLMWebAdapter } from "./glm-web-adapter.js";
 import { getProfileDir } from "../platform/paths.js";
@@ -18,15 +19,13 @@ import { getProfileDir } from "../platform/paths.js";
 //                          data dir; each provider logs in once, independently
 //   status               - "active" (has a working adapter) | "planned" (named
 //                          but not implemented yet)
-//   promptsForMode       - whether the CLI shows an interactive mode picker at
-//                          conversation start (ChatGPT has Pro/Current; most
-//                          providers do not prompt)
-//   defaultMode          - mode applied silently at conversation start when the
-//                          provider does not prompt (null = keep the site's
-//                          current setting). The adapter's selectMode()
-//                          interprets this value.
 //   adapter              - the adapter class, or null until implemented
 //
+// WTAgent deliberately does not encode provider-specific model names or choose
+// a model automatically. After login, interactive runs give the user a chance
+// to choose directly on the provider website; pressing Enter keeps whatever
+// the website currently has selected.
+//
 // ChatGPT keeps the historical "chrome-profile" basename so existing logins,
 // the logout guard, and cli.test.js keep working unchanged.
 export const PROVIDERS = Object.freeze({
@@ -36,8 +35,6 @@ export const PROVIDERS = Object.freeze({
     baseUrl: "https://chatgpt.com/",
     profileBasename: "chrome-profile",
     status: "active",
-    promptsForMode: true,
-    defaultMode: null,
     adapter: ChatGPTWebAdapter,
   }),
   deepseek: Object.freeze({
@@ -46,10 +43,6 @@ export const PROVIDERS = Object.freeze({
     baseUrl: "https://chat.deepseek.com/",
     profileBasename: "deepseek-profile",
     status: "active",
-    // DeepSeek does not prompt; every new conversation silently switches to
-    // 专家模式 (Expert) + 深度思考 (Deep Thinking) — see DeepSeekWebAdapter.selectMode.
-    promptsForMode: false,
-    defaultMode: "expert-thinking",
     adapter: DeepSeekWebAdapter,
   }),
   claude: Object.freeze({
@@ -58,10 +51,6 @@ export const PROVIDERS = Object.freeze({
     baseUrl: "https://claude.ai/",
     profileBasename: "claude-profile",
     status: "active",
-    promptsForMode: false,
-    // Keep whatever model claude.ai selects in this browser profile. WTAgent
-    // does not open the model menu or override the account/site default.
-    defaultMode: null,
     adapter: ClaudeWebAdapter,
   }),
   grok: Object.freeze({
@@ -69,10 +58,8 @@ export const PROVIDERS = Object.freeze({
     label: "Grok",
     baseUrl: "https://grok.com/",
     profileBasename: "grok-profile",
-    status: "planned",
-    promptsForMode: false,
-    defaultMode: null,
-    adapter: null,
+    status: "active",
+    adapter: GrokWebAdapter,
   }),
   kimi: Object.freeze({
     id: "kimi",
@@ -80,10 +67,6 @@ export const PROVIDERS = Object.freeze({
     baseUrl: "https://www.kimi.com/",
     profileBasename: "kimi-profile",
     status: "active",
-    // Kimi does not prompt; every new conversation silently switches to the K3
-    // flagship model — see KimiWebAdapter.selectMode.
-    promptsForMode: false,
-    defaultMode: "k3",
     adapter: KimiWebAdapter,
   }),
   glm: Object.freeze({
@@ -92,10 +75,6 @@ export const PROVIDERS = Object.freeze({
     baseUrl: "https://chat.z.ai/",
     profileBasename: "glm-profile",
     status: "active",
-    // GLM does not prompt; every new conversation silently selects the newest
-    // available model (GLM-5.3, else GLM-5.2) — see GLMWebAdapter.selectMode.
-    promptsForMode: false,
-    defaultMode: "latest",
     adapter: GLMWebAdapter,
   }),
   gemini: Object.freeze({
@@ -104,9 +83,6 @@ export const PROVIDERS = Object.freeze({
     baseUrl: "https://gemini.google.com/app",
     profileBasename: "gemini-profile",
     status: "active",
-    promptsForMode: false,
-    // Preserve the model currently selected by Gemini in this profile.
-    defaultMode: null,
     adapter: GeminiWebAdapter,
   }),
 });
@@ -138,24 +114,6 @@ export function getProvider(providerId) {
   return provider;
 }
 
-// `--mode` is ChatGPT's Pro/Current switch, but people often type
-// `--mode kimi` meaning the Kimi provider. If the value is a known provider
-// id, treat it as `--model` and clear `--mode` so ChatGPT mode parsing is not
-// applied. Conflicting `--model` + `--mode <provider>` is rejected.
-export function resolveCliProviderSelection({ model, mode } = {}) {
-  const modeAsProvider = findProvider(mode);
-  if (!modeAsProvider) {
-    return { model, mode };
-  }
-  if (model != null && getProvider(model).id !== modeAsProvider.id) {
-    throw new Error(
-      `--mode ${mode} selects ${modeAsProvider.label}, but --model ${model} is already set. `
-        + `Use --model ${modeAsProvider.id}.`,
-    );
-  }
-  return { model: modeAsProvider.id, mode: undefined };
-}
-
 // Resolves the dedicated Chrome profile directory for a provider under the
 // given app data dir. ChatGPT resolves to the historical "chrome-profile".
 export function getProviderProfileDir(appDataDir, providerId) {