wtagent 0.2.2 → 0.3.0

package/README.md

@@ -2,7 +2,7 @@
 
 Use your web AI account as a local CLI agent.
 
-WTAgent connects ChatGPT, Claude, DeepSeek, Gemini, Kimi, or GLM Web to your local project: the model reasons in the browser while WTAgent reads and writes local files and runs commands on your machine.
+WTAgent connects ChatGPT, Claude, DeepSeek, Gemini, Grok, Kimi, or GLM Web to your local project: the model reasons in the browser while WTAgent reads and writes local files and runs commands on your machine.
 
 [中文](#中文)
 
@@ -57,7 +57,7 @@ Multiline paste and `↑` / `↓` input history are supported. Press `Ctrl+C` or
 
 ## 中文
 
-WTAgent 将 ChatGPT、Claude、DeepSeek、Gemini、Kimi 或 GLM 网页聊天连接到本地项目：模型在浏览器中思考，WTAgent 在你的电脑上读写本地文件并运行命令。
+WTAgent 将 ChatGPT、Claude、DeepSeek、Gemini、Grok、Kimi 或 GLM 网页聊天连接到本地项目：模型在浏览器中思考，WTAgent 在你的电脑上读写本地文件并运行命令。
 
 ### 快速开始
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "wtagent",
-  "version": "0.2.2",
+  "version": "0.3.0",
   "description": "Turn your own web AI session into a local tool-using agent.",
   "repository": {
     "type": "git",
@@ -30,7 +30,7 @@
   ],
   "scripts": {
     "test": "node --test",
-    "test:unit": "node --test test/protocol.test.js test/policy.test.js test/tools.test.js test/prompt-builder.test.js test/session-export.test.js test/platform.test.js test/cli.test.js test/at-files.test.js test/self-update.test.js test/startup-notices.test.js test/provider-registry.test.js test/claude-adapter.test.js test/deepseek-adapter.test.js test/gemini-adapter.test.js test/kimi-adapter.test.js test/glm-adapter.test.js",
+    "test:unit": "node --test test/protocol.test.js test/policy.test.js test/tools.test.js test/prompt-builder.test.js test/session-export.test.js test/platform.test.js test/cli.test.js test/at-files.test.js test/self-update.test.js test/startup-notices.test.js test/provider-registry.test.js test/claude-adapter.test.js test/deepseek-adapter.test.js test/gemini-adapter.test.js test/grok-adapter.test.js test/kimi-adapter.test.js test/glm-adapter.test.js",
     "test:integration": "node --test test/runtime.test.js",
     "test:windows": "node --test test/windows-command-launcher.test.js test/windows-diagnostics.test.js test/windows-search-path.test.js test/atomic-write.test.js test/process-utils.test.js test/cdp-state.test.js test/windows-fixtures.test.js test/windows-package-smoke.test.js",
     "check": "node scripts/check-syntax.js",

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

@@ -5,7 +5,6 @@ import { launchAndConnectCdpChrome } from "./cdp-browser.js";
 import { discoverChromeExecutable } from "../platform/chrome-discovery.js";
 import { ensureDirectory } from "../platform/paths.js";
 import { BrowserAdapterError } from "../shared/errors.js";
-import { runModeSelection } from "./mode-selection.js";
 
 // Playwright error messages for a dead transport. The Chrome process itself is
 // usually still alive (e.g. the connection died while the Mac slept); these
@@ -76,9 +75,10 @@ function sameConversationUrl(left, right) {
 // The runtime talks to this class only through its public methods; everything
 // that varies between providers (ChatGPT, DeepSeek, …) is isolated behind the
 // overridable "primitive" methods below. A concrete provider subclass supplies
-// its base URL, DOM locators, message-identity extraction, and (optionally) a
-// model-switcher port; it must NOT re-implement the turn-completion loop, the
-// send/auth/reconnect flow, or the WTAgent <agent_response> protocol timing.
+// its base URL, DOM locators, and message-identity extraction; it must NOT
+// re-implement the turn-completion loop, send/auth/reconnect flow, or the
+// WTAgent <agent_response> protocol timing. Model selection deliberately stays
+// outside adapters: users choose their model directly on the provider website.
 //
 // IMPORTANT (JavaScript semantics): primitives that the base dispatches to a
 // provider are declared as ordinary (non-#private) methods. `#private` methods
@@ -113,7 +113,6 @@ export class BaseWebAdapter {
     this.assistantCountBeforeSend = 0;
     this.sentUserTurn = null;
     this.lastAssistantMessageId = null;
-    this.lastModeSelection = null;
   }
 
   // ---- provider primitives (override in subclasses) ----------------------
@@ -203,23 +202,6 @@ export class BaseWebAdapter {
     return null;
   }
 
-  // Port consumed by runModeSelection(). Default reports no switcher, so
-  // selectMode() resolves to "switcher_not_found" and never blocks a provider
-  // that has no model picker.
-  modeSelectionPort() {
-    return {
-      alreadyOnMode: async () => false,
-      hasSwitcher: async () => false,
-      openMenu: async () => {},
-      readOptions: async () => [],
-      clickOption: async () => false,
-      waitClosed: async () => false,
-      waitSelected: async () => false,
-      closeMenu: async () => {},
-      writeDiagnostics: async (label) => this.writeDiagnostics(label),
-    };
-  }
-
   // Best-effort composer file upload. Default: nothing attached.
   async attachFiles(_files) {
     return { attached: [], failed: [] };
@@ -521,18 +503,6 @@ export class BaseWebAdapter {
     }
   }
 
-  async selectMode(mode) {
-    this.requirePage();
-    if (!mode) {
-      return { status: "skipped", requested: mode, attempts: 0 };
-    }
-
-    const port = this.modeSelectionPort();
-    const result = await runModeSelection(port, mode);
-    this.lastModeSelection = result;
-    return result;
-  }
-
   async getConversationUrl() {
     this.requirePage();
     return this.page.url();

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

@@ -1,12 +1,6 @@
 import { BaseWebAdapter, firstVisible, hasCompleteAgentEnvelope } from "./base-web-adapter.js";
 import { BrowserAdapterError } from "../shared/errors.js";
 import { isUsageLimitNotice } from "../shared/usage-limit.js";
-import {
-  chooseModeOption,
-  labelMatchesToken,
-  slugMatchesToken,
-  normalizeToken,
-} from "./mode-selection.js";
 
 // Re-exported so existing importers (the runtime and browser-adapter tests)
 // keep importing it from here even though it now lives on the provider-agnostic
@@ -21,10 +15,10 @@ function parseConversationTurn(value) {
 }
 
 // ChatGPT-specific implementation of the WebModelAdapter contract. Only the
-// provider primitives (DOM locators, message identity, mode-switcher port,
-// upload, overlay handling, usage-limit card) live here; all orchestration —
-// launch/auth/send/reconnect and the turn-completion loop — is inherited from
-// BaseWebAdapter.
+// provider primitives (DOM locators, message identity, upload, overlay
+// handling, usage-limit card) live here; all orchestration — launch/auth/send/
+// reconnect and the turn-completion loop — is inherited from BaseWebAdapter.
+// Model choice is intentionally left to the user in the ChatGPT website.
 export class ChatGPTWebAdapter extends BaseWebAdapter {
   constructor(options = {}) {
     super({
@@ -337,211 +331,4 @@ export class ChatGPTWebAdapter extends BaseWebAdapter {
     }).catch(() => null);
   }
 
-  // DOM port for mode-selection.js. Stable attribute slugs are preferred;
-  // exact visible labels are a fallback for current menus that omit those
-  // attributes. Disabled state is still read from DOM state, never from text.
-  modeSelectionPort() {
-    const page = this.page;
-    return {
-      alreadyOnMode: async (requested) => {
-        const token = normalizeToken(requested);
-        const switcher = await this.#findModelSwitcher();
-        const switcherLabel = switcher
-          ? await switcher.innerText().catch(() => "")
-          : "";
-        if (labelMatchesToken(switcherLabel, token)) {
-          return true;
-        }
-
-        const messages = this.assistantMessages();
-        if (await messages.count() === 0) {
-          return false;
-        }
-        const slug = await messages.last()
-          .getAttribute("data-message-model-slug")
-          .catch(() => null);
-        return Boolean(slug) && slugMatchesToken(slug, token);
-      },
-      hasSwitcher: async () => Boolean(await this.#findModelSwitcher()),
-      openMenu: async (requested) => {
-        const switcher = await this.#findModelSwitcher();
-        if (switcher) {
-          const opened = await switcher.click({ timeout: 5_000 })
-            .then(() => true)
-            .catch(() => false);
-          if (!opened) {
-            await switcher.click({ force: true, timeout: 5_000 })
-              .catch(() => null);
-          }
-          // Wait for the Radix menu portal to attach before reading options.
-          await page.locator('[role="menu"]:visible [role="menuitem"]:visible, [role="menu"]:visible [role="menuitemradio"]:visible')
-            .first().waitFor({ state: "visible", timeout: 5_000 })
-            .catch(() => null);
-          await this.#revealNestedModeOption(requested);
-        }
-      },
-      readOptions: async () => this.#readModeOptions(),
-      clickOption: async (index) => {
-        const option = this.#modeOptionLocators().nth(index);
-        try {
-          await option.click({ timeout: 5_000 });
-          return true;
-        } catch {
-          return false;
-        }
-      },
-      waitClosed: async () => {
-        const ok = await page.locator('[role="menu"]').first()
-          .waitFor({ state: "hidden", timeout: 5_000 })
-          .then(() => true)
-          .catch(() => false);
-        return ok;
-      },
-      waitSelected: async (requested) => {
-        const token = normalizeToken(requested);
-        const deadline = Date.now() + 5_000;
-        while (Date.now() < deadline) {
-          const switcher = await this.#findModelSwitcher(500);
-          const label = switcher
-            ? await switcher.innerText().catch(() => "")
-            : "";
-          if (labelMatchesToken(label, token)) {
-            return true;
-          }
-
-          const selectedOption = (await this.#readModeOptions()).find(
-            (option) => option.selected
-              && (
-                slugMatchesToken(option.slug, token)
-                || labelMatchesToken(option.label, token)
-              ),
-          );
-          if (selectedOption) {
-            return true;
-          }
-          await page.waitForTimeout(100);
-        }
-        return false;
-      },
-      closeMenu: async () => {
-        await page.keyboard.press("Escape").catch(() => null);
-        await page.waitForTimeout(150);
-      },
-      writeDiagnostics: async (label) => this.writeDiagnostics(label),
-    };
-  }
-
-  // ---- ChatGPT-only mode-picker helpers (private) ------------------------
-
-  #modeOptionLocators() {
-    return this.page.locator(
-      '[role="menu"]:visible [role="menuitemradio"]:visible, '
-        + '[role="menu"]:visible [role="menuitem"]:visible',
-    );
-  }
-
-  async #revealNestedModeOption(requested) {
-    const present = (options) => {
-      const choice = chooseModeOption(options, requested);
-      return choice.status !== "unavailable";
-    };
-
-    let options = await this.#readModeOptions();
-    if (present(options)) {
-      return;
-    }
-
-    const submenuTriggers = this.page.locator(
-      '[role="menu"]:visible [role="menuitem"][aria-haspopup="menu"]:visible',
-    );
-    const count = await submenuTriggers.count().catch(() => 0);
-    for (let index = 0; index < count; index += 1) {
-      const trigger = submenuTriggers.nth(index);
-      await trigger.dispatchEvent("pointermove", { pointerType: "mouse" })
-        .catch(() => null);
-      await this.page.waitForTimeout(250);
-      options = await this.#readModeOptions();
-      if (present(options)) {
-        return;
-      }
-    }
-  }
-
-  // Enumerates the open menu's options into { index, slug, label, disabled }.
-  // `slug` is the first stable attribute found and may be empty in current
-  // ChatGPT menus; `disabled` is read from ARIA / data-state, never from text.
-  async #readModeOptions() {
-    const locator = this.#modeOptionLocators();
-    const count = await locator.count().catch(() => 0);
-    const options = [];
-    for (let index = 0; index < count; index += 1) {
-      const item = locator.nth(index);
-      const [testId, dataValue, dataTestValue, id, ariaDisabled, ariaChecked, dataDisabled, dataState, label] =
-        await Promise.all([
-          item.getAttribute("data-testid").catch(() => null),
-          item.getAttribute("data-value").catch(() => null),
-          item.getAttribute("data-test-value").catch(() => null),
-          item.getAttribute("id").catch(() => null),
-          item.getAttribute("aria-disabled").catch(() => null),
-          item.getAttribute("aria-checked").catch(() => null),
-          item.getAttribute("data-disabled").catch(() => null),
-          item.getAttribute("data-state").catch(() => null),
-          item.innerText().catch(() => ""),
-        ]);
-      const slug = [testId, dataValue, dataTestValue, id].find(Boolean) ?? "";
-      const disabled = ariaDisabled === "true"
-        || dataDisabled === "true"
-        || dataDisabled === ""
-        || dataState === "disabled"
-        || !await item.isEnabled().catch(() => true);
-      const selected = ariaChecked === "true" || dataState === "checked";
-      options.push({
-        index,
-        slug,
-        label: (label ?? "").trim(),
-        disabled,
-        selected,
-      });
-    }
-    return options;
-  }
-
-  async #findModelSwitcher(timeoutMs = 45_000) {
-    const deadline = Date.now() + timeoutMs;
-
-    while (Date.now() < deadline) {
-      const direct = await firstVisible([
-        this.page.locator('[data-testid="model-switcher-dropdown-button"]'),
-        this.page.locator(
-          'main button.__composer-pill[aria-haspopup="menu"]',
-        ),
-        this.page.locator('button[aria-label*="model" i]'),
-        this.page.locator('button[aria-label*="模式"]'),
-      ]);
-      if (direct) {
-        return direct;
-      }
-
-      const menuButtons = this.page.locator(
-        'main button[aria-haspopup="menu"]:not([data-testid^="history-item-"])',
-      );
-      const count = await menuButtons.count();
-      for (let index = 0; index < count; index += 1) {
-        const button = menuButtons.nth(index);
-        if (!await button.isVisible().catch(() => false)) continue;
-        const text = (await button.innerText().catch(() => "")).trim();
-        const label = await button.getAttribute("aria-label").catch(() => "");
-        if (
-          /^(chatgpt|gpt(?:-[\w.]+)?|auto|instant|thinking|pro)$/i.test(text)
-          || /model|模型|模式/i.test(label ?? "")
-        ) {
-          return button;
-        }
-      }
-
-      await this.page.waitForTimeout(250);
-    }
-
-    return null;
-  }
 }

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

@@ -34,112 +34,6 @@ export class DeepSeekWebAdapter extends BaseWebAdapter {
     return /^\/a\/chat\/s\//;
   }
 
-  // DeepSeek has no ChatGPT-style model dropdown; instead a new conversation
-  // exposes mode chips (快速/专家/识图) and toggles (深度思考/智能搜索). The
-  // registry's defaultMode ("expert-thinking") asks for 专家模式 (Expert) +
-  // 深度思考 (Deep Thinking) — WTAgent's preferred DeepSeek setup, applied
-  // silently on every fresh conversation (no interactive picker). Any other
-  // requested value keeps the site's current setting.
-  //
-  // Selection is LANGUAGE-INDEPENDENT and idempotent:
-  //   - the chips are role="radio" in a fixed order (fast, expert, image), so
-  //     the expert chip is the second radio; visible labels are only used to
-  //     double-check the position when they match a locale we know
-  //   - after expert is active DeepSeek shows exactly ONE .ds-toggle-button
-  //     (deep thinking); fast mode shows two (deep thinking + web search), so
-  //     the toggle count itself proves the chip switch landed. Clicking the
-  //     single remaining toggle needs no label at all.
-  // Selection is best-effort — a UI change never aborts the run; it reports
-  // "unresolved" and keeps the current mode, like runModeSelection.
-  async selectMode(mode) {
-    this.requirePage();
-    if (mode !== "expert-thinking") {
-      return { status: "skipped", requested: mode, attempts: 0 };
-    }
-
-    const steps = [];
-    const chip = await this.#findExpertChip();
-    if (!chip) {
-      steps.push({ ok: false, label: "expert-chip" });
-    } else {
-      steps.push(await this.#ensureChipChecked(chip, "expert-chip"));
-      if (steps[0].ok) {
-        steps.push(await this.#ensureSingleThinkingToggle());
-      }
-    }
-
-    const failed = steps.filter((s) => !s.ok);
-    if (failed.length > 0) {
-      await this.writeDiagnostics("deepseek-mode-partial");
-      return {
-        status: "unresolved",
-        requested: mode,
-        selectedLabel: "expert + deep-thinking",
-        attempts: 1,
-        reason: `Could not confirm: ${failed.map((s) => s.label).join(", ")}.`,
-      };
-    }
-    return {
-      status: "select",
-      requested: mode,
-      selectedLabel: "expert + deep-thinking",
-      attempts: 1,
-      reason: "Selected expert mode with deep thinking.",
-    };
-  }
-
-  // Locates the expert chip without relying on its locale: try known labels
-  // first, then fall back to the fixed chip order (fast, expert, image).
-  async #findExpertChip() {
-    const radios = this.page.locator('[role="radio"]');
-    const count = await radios.count().catch(() => 0);
-    if (count === 0) {
-      return null;
-    }
-    const labelPattern = /专家|expert/i;
-    for (let index = 0; index < count; index += 1) {
-      const text = (await radios.nth(index).innerText().catch(() => "")).trim();
-      if (labelPattern.test(text)) {
-        return radios.nth(index);
-      }
-    }
-    // Positional fallback: 快速/专家/识图 — the expert chip is the 2nd radio.
-    return count >= 2 ? radios.nth(1) : null;
-  }
-
-  // Clicks a chip unless it is already aria-checked. Returns { ok, label }.
-  async #ensureChipChecked(chip, label) {
-    if (await chip.getAttribute("aria-checked").catch(() => null) === "true") {
-      return { ok: true, label };
-    }
-    await chip.click({ timeout: 5_000 }).catch(() => null);
-    await this.page.waitForTimeout(400);
-    const checked = await chip.getAttribute("aria-checked").catch(() => null);
-    return { ok: checked === "true", label };
-  }
-
-  // In expert mode exactly ONE toggle (deep thinking) exists; in fast mode
-  // there are two. Clicking the single remaining toggle therefore never needs
-  // a label. The toggle count also verifies the chip switch actually landed.
-  async #ensureSingleThinkingToggle() {
-    const toggles = this.page.locator(".ds-toggle-button");
-    const count = await toggles.count().catch(() => 0);
-    if (count !== 1) {
-      return { ok: false, label: "deep-thinking-toggle" };
-    }
-    const toggle = toggles.nth(0);
-    const isSelected = async () => (
-      (await toggle.getAttribute("class").catch(() => "") ?? "")
-        .includes("ds-toggle-button--selected")
-    );
-    if (await isSelected()) {
-      return { ok: true, label: "deep-thinking-toggle" };
-    }
-    await toggle.click({ timeout: 5_000 }).catch(() => null);
-    await this.page.waitForTimeout(400);
-    return { ok: await isSelected(), label: "deep-thinking-toggle" };
-  }
-
   composerLocators() {
     return [
       this.page.locator('textarea[name="search"]'),

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

@@ -5,11 +5,6 @@ export { isConnectionLostError } from "./base-web-adapter.js";
 
 const GLM_URL = "https://chat.z.ai/";
 
-// Preferred models, newest first: try GLM-5.3 when present, else GLM-5.2. The
-// site sometimes exposes 5.3 and sometimes only 5.2, so selection walks this
-// list and clicks the first one present in the menu.
-const PREFERRED_MODELS = ["GLM-5.3", "GLM-5.2"];
-
 // GLM / Z.ai (chat.z.ai) adapter.
 //
 // chat.z.ai is an Open WebUI (Svelte) frontend, verified against the live app:
@@ -21,8 +16,8 @@ const PREFERRED_MODELS = ["GLM-5.3", "GLM-5.2"];
 //   - assistant answer markdown is `.markdown-prose` / `.prose`; a "思考过程"
 //     (deep-thinking) block may precede it and is excluded when reading the reply
 //   - a live conversation URL is /c/<uuid>
-//   - model switcher: `button.modelSelectorButton`; the registry defaultMode
-//     "latest" picks the newest available model (GLM-5.3, else GLM-5.2)
+//   - model choice is intentionally left to the user on chat.z.ai; WTAgent does
+//     not inspect or override the site's current model selection
 //   - Cloudflare guards the site; the base throwIfBlockedPage surfaces the
 //     window so the user can pass the check (wtagent's own CDP launch is not
 //     fingerprinted the way headless automation is)
@@ -161,86 +156,4 @@ export class GLMWebAdapter extends BaseWebAdapter {
     return 120;
   }
 
-  // Selects the newest available model. The registry's defaultMode "latest" maps
-  // to PREFERRED_MODELS (GLM-5.3, else GLM-5.2). Best-effort and non-throwing.
-  //
-  // The switcher is `button.modelSelectorButton`; opening it lists options whose
-  // visible text is the exact model name. After clicking, the switcher label
-  // becomes the selected model name.
-  async selectMode(mode) {
-    this.requirePage();
-    if (mode !== "latest") {
-      return { status: "skipped", requested: mode, attempts: 0 };
-    }
-
-    const switcher = this.page.locator("button.modelSelectorButton").first();
-    await switcher.waitFor({ state: "visible", timeout: 10_000 }).catch(() => null);
-    if (await switcher.count().catch(() => 0) === 0) {
-      await this.writeDiagnostics("glm-model-switcher-not-found");
-      return {
-        status: "switcher_not_found",
-        requested: mode,
-        attempts: 0,
-        reason: "Model switcher was not found.",
-      };
-    }
-
-    const current = (await switcher.innerText().catch(() => "")).trim();
-    // Already on the most-preferred model that exists? If the current label is
-    // the first preferred model, nothing to do.
-    if (current.startsWith(PREFERRED_MODELS[0])) {
-      return {
-        status: "already",
-        requested: mode,
-        selectedLabel: PREFERRED_MODELS[0],
-        attempts: 0,
-        reason: `Already using ${PREFERRED_MODELS[0]}.`,
-      };
-    }
-
-    for (const model of PREFERRED_MODELS) {
-      await switcher.click({ timeout: 5_000 }).catch(() => null);
-      await this.page.waitForTimeout(600);
-      const option = this.page.getByText(model, { exact: true }).first();
-      if (await option.count().catch(() => 0) === 0) {
-        // Not in the menu; close and try the next preferred model.
-        await this.page.keyboard.press("Escape").catch(() => null);
-        continue;
-      }
-      await option.click({ timeout: 5_000 }).catch(() => null);
-      await this.page.waitForTimeout(600);
-      const after = (await switcher.innerText().catch(() => "")).trim();
-      if (after.startsWith(model)) {
-        // The model menu stays open after a selection; a click in the page
-        // center dismisses it so it does not cover the composer.
-        await this.#dismissModelMenu();
-        return {
-          status: current.startsWith(model) ? "already" : "select",
-          requested: mode,
-          selectedLabel: model,
-          attempts: 1,
-          reason: `Selected ${model}.`,
-        };
-      }
-    }
-
-    await this.#dismissModelMenu();
-    await this.writeDiagnostics("glm-mode-latest-unresolved");
-    return {
-      status: "unresolved",
-      requested: mode,
-      attempts: 1,
-      reason: `Could not select any of: ${PREFERRED_MODELS.join(", ")}.`,
-    };
-  }
-
-  async #dismissModelMenu() {
-    const viewport = this.page.viewportSize?.() ?? { width: 1280, height: 800 };
-    await this.page.mouse.click(
-      Math.floor(viewport.width / 2),
-      Math.floor(viewport.height / 2),
-    ).catch(() => null);
-    await this.page.keyboard.press("Escape").catch(() => null);
-    await this.page.waitForTimeout(200);
-  }
 }