@agentium/browser 3.0.2 → 3.1.1

package/dist/index.d.ts

@@ -1,801 +1,5 @@
-import * as _agentium_core from '@agentium/core';
-import { ModelProvider, UnifiedMemoryConfig, ToolDef, CostTracker, LogLevel, EventBus, MemoryManager } from '@agentium/core';
-
-/**
- * Secure credential store for BrowserAgent.
- *
- * Secrets are stored in memory and NEVER sent to the LLM.
- * The model works with placeholders (e.g. `{{email}}`, `{{password}}`),
- * and the agent resolves them to real values only at execution time.
- */
-declare class CredentialVault {
-    private secrets;
-    constructor(initial?: Record<string, string>);
-    /** Store a credential. Key names become the placeholder: `{{key}}`. */
-    set(key: string, value: string): this;
-    /** Retrieve a credential value. Returns undefined if not found. */
-    get(key: string): string | undefined;
-    has(key: string): boolean;
-    /** List available placeholder names (never exposes values). */
-    keys(): string[];
-    /**
-     * Load credentials from environment variables.
-     * Maps env var names to placeholder keys.
-     *
-     * @example
-     * vault.fromEnv({ email: "LOGIN_EMAIL", password: "LOGIN_PASS" });
-     */
-    fromEnv(mapping: Record<string, string>): this;
-    /**
-     * Replace `{{key}}` placeholders in text with actual credential values.
-     * Used internally by BrowserAgent right before executing a type action.
-     */
-    resolve(text: string): string;
-    /**
-     * Replace any occurrence of real credential values in text with
-     * their `{{key}}` placeholder. Used to sanitize logs and action history.
-     */
-    mask(text: string): string;
-}
-
-/**
- * The complete set of actions a `BrowserAgent` can take during a run.
- *
- * Click / type / scroll accept EITHER an `index` (preferred — resolved
- * against the DOM tree we showed the model) OR `x`/`y` coordinates as a
- * fallback. Indexed actions are far more reliable on dynamic pages because
- * they survive layout shifts, devicePixelRatio mismatches, and "the
- * sibling 4 pixels away" failure modes.
- */
-type BrowserAction = {
-    action: "click";
-    /** Element index from the DOM snapshot. Preferred. */
-    index?: number;
-    /** Fallback x coordinate (CSS pixels). */
-    x?: number;
-    /** Fallback y coordinate (CSS pixels). */
-    y?: number;
-    /** Free-form description; if it contains a quoted label, used as a text-locator fallback. */
-    description?: string;
-} | {
-    action: "type";
-    /** Element index from the DOM snapshot. Preferred. */
-    index?: number;
-    text: string;
-    /** Fallback x coordinate (CSS pixels). */
-    x?: number;
-    /** Fallback y coordinate (CSS pixels). */
-    y?: number;
-    /** If true (default), clear the field first. */
-    clear?: boolean;
-    /** If true (default), press Enter after typing when text ends with `\n`. */
-    submit?: boolean;
-} | {
-    action: "scroll";
-    direction: "up" | "down";
-    /** Pixels to scroll. Default: 400. */
-    amount?: number;
-    /** If provided, scroll the element with this index into view instead. */
-    index?: number;
-} | {
-    action: "navigate";
-    url: string;
-} | {
-    action: "back";
-} | {
-    action: "wait";
-    ms: number;
-} | {
-    action: "screenshot";
-} | {
-    action: "send_keys";
-    keys: string;
-} | {
-    action: "find_text";
-    text: string;
-} | {
-    action: "evaluate";
-    code: string;
-} | {
-    action: "dropdown_options";
-    index: number;
-} | {
-    action: "select_dropdown";
-    index: number;
-    text: string;
-} | {
-    action: "upload_file";
-    index: number;
-    path: string;
-} | {
-    action: "extract";
-    /** Natural-language description of what to extract. */
-    query: string;
-    /** Include link hrefs in the extracted content. Default: false. */
-    extractLinks?: boolean;
-} | {
-    action: "tool";
-    /** Name of a custom tool registered on the BrowserAgent. */
-    name: string;
-    args?: Record<string, unknown>;
-} | {
-    action: "done";
-    result: string;
-} | {
-    action: "fail";
-    reason: string;
-};
-/**
- * One entry in the DOM snapshot the model sees. Returned by
- * `BrowserProvider.extractDOM()` alongside the human-readable string form.
- */
-interface DomElement {
-    /** Stable 1-based index for this step. Used by indexed actions. */
-    index: number;
-    /** Center coordinate in CSS pixels (fallback for the model if needed). */
-    cx: number;
-    cy: number;
-    /** ARIA role or tag name. */
-    role: string;
-    /** Input type, if applicable. */
-    type?: string;
-    /** Visible label (text / aria-label / placeholder / href / …). */
-    label: string;
-    /** Tag name. */
-    tag: string;
-    /** Whether it's an `<input>` / `<textarea>` / `[contenteditable]`. */
-    isInput: boolean;
-    /** Whether it's a `<select>`. */
-    isSelect: boolean;
-    /** Whether it's an `<input type="file">`. */
-    isFile: boolean;
-    /** Origin frame (`"main"` or the iframe `src`/path). */
-    frame?: string;
-}
-/**
- * Scroll-context metadata returned alongside `DomElement[]`. Gives the
- * model spatial awareness so it can decide when to scroll vs when more
- * content is below / above the fold.
- */
-interface DomScrollContext {
-    /** Approximate viewports of scrollable content above the current view. */
-    pagesAbove: number;
-    /** Approximate viewports of scrollable content below the current view. */
-    pagesBelow: number;
-    /** Total interactive elements found (visible + hidden combined). */
-    totalInteractive: number;
-    /** Count of interactive elements that exist on the page but aren't in the viewport. */
-    hiddenInteractive: number;
-}
-/** Combined return value of `BrowserProvider.extractDOM()`. */
-interface DomSnapshot {
-    /** Human-readable string fed to the model. */
-    text: string;
-    /** Structured list (stable indices) for runtime resolution. */
-    elements: DomElement[];
-    /** Spatial/scroll context. */
-    scroll: DomScrollContext;
-}
-interface BrowserAgentConfig {
-    name: string;
-    /** Vision-capable model (GPT-4o, Gemini, etc.) */
-    model: ModelProvider;
-    /**
-     * Optional secondary (usually cheaper) model used for the `extract`
-     * action and other text-only sub-tasks. Falls back to `model`.
-     */
-    pageExtractionLLM?: ModelProvider;
-    /**
-     * Fallback model used automatically when the primary `model` returns a
-     * rate-limit / auth / 5xx error or fails to produce valid JSON several
-     * times in a row. Failure-budget aware. Leave unset to disable.
-     */
-    fallbackModel?: ModelProvider;
-    /**
-     * Ask the model to emit a structured thinking/evaluation/memory/next_goal
-     * envelope around its action(s). Significantly improves accuracy and
-     * self-correction on multi-step tasks. Default: `true`. Set `false` for
-     * a `flash_mode` that just returns the raw action(s) — useful for very
-     * fast / cheap models that are bad at long outputs.
-     */
-    useThinking?: boolean;
-    /**
-     * Maximum number of recent step turns kept verbatim in the conversation
-     * history sent to the model. Older turns are compacted into a single
-     * summary line. Default: 6. Set 0 to disable conversation history
-     * entirely (each step rebuilt from scratch — v2.0 behaviour).
-     */
-    historyWindow?: number;
-    /** Extra instructions appended to the default system prompt. */
-    instructions?: string;
-    /**
-     * Append additional instructions to the default system prompt.
-     * Alias for `instructions` (browser-use parity). Both are concatenated.
-     */
-    extendSystemMessage?: string;
-    /**
-     * Completely replace the default system prompt with this string.
-     * The credentials / DOM / coordinate sections are still appended at the
-     * end automatically. Most users should use `instructions` instead.
-     */
-    overrideSystemMessage?: string;
-    /** Max vision loop iterations. Default: 30 */
-    maxSteps?: number;
-    /**
-     * Maximum number of consecutive step failures (action threw, invalid
-     * JSON from model, locator timeout) before the agent gives up. Default: 3.
-     */
-    maxFailures?: number;
-    /**
-     * Maximum number of actions the model can return in a single step.
-     * If the model returns an array, we execute them in order until the page
-     * navigates or the DOM changes substantially. Default: 3.
-     * Set to 1 to force one-action-per-step (the v2.0.x behaviour).
-     */
-    maxActionsPerStep?: number;
-    /**
-     * Actions to run before the LLM loop starts. Useful for boilerplate
-     * (cookie-banner click, login flow, scrolling) you already know the
-     * answer to — saves vision tokens.
-     */
-    initialActions?: BrowserAction[];
-    /**
-     * Vision mode. Default: `"auto"`.
-     * - `true`: always send a screenshot with every step (v2.0.x behaviour).
-     * - `false`: never send screenshots; DOM-only operation.
-     * - `"auto"`: send a screenshot on the first step and whenever the model
-     *   used the `screenshot` action in the previous step. Saves vision
-     *   tokens dramatically on DOM-only-suitable workloads.
-     */
-    useVision?: boolean | "auto";
-    /**
-     * If true (default), detect a URL in the task string and navigate to it
-     * before the first LLM call — saves one round trip on simple tasks.
-     */
-    directlyOpenUrl?: boolean;
-    /** Run browser without visible window. Default: true */
-    headless?: boolean;
-    /** Browser viewport size. Default: 1280x720 */
-    viewport?: {
-        width: number;
-        height: number;
-    };
-    /** Initial URL to navigate to before starting the task */
-    startUrl?: string;
-    /** Milliseconds to wait after each action for the page to settle. Default: 1500 */
-    waitAfterAction?: number;
-    /** Max consecutive identical actions before the agent auto-fails. Default: 3 */
-    maxRepeats?: number;
-    /**
-     * Include a simplified DOM/accessibility tree (each interactive element
-     * tagged with its exact center coordinates) alongside the screenshot.
-     * Dramatically improves click accuracy — strongly recommended.
-     * Default: true
-     */
-    useDOM?: boolean;
-    /**
-     * Allow the `evaluate` action to run arbitrary JavaScript inside the
-     * page. Default: false (security). Only enable if you trust the source
-     * of task strings — a malicious task could exfiltrate page contents.
-     */
-    allowEvaluate?: boolean;
-    /**
-     * Restrict navigation to specific domains. Wildcard patterns supported:
-     * `"example.com"`, `"*.example.com"`, `"http*://example.com"`.
-     * When set, any `navigate` action to a non-matching URL throws.
-     */
-    allowedDomains?: string[];
-    /**
-     * Block navigation to specific domains. Same pattern format as
-     * `allowedDomains`. Evaluated AFTER `allowedDomains`, so if both are
-     * set a URL must be in `allowedDomains` AND not in `prohibitedDomains`.
-     */
-    prohibitedDomains?: string[];
-    /**
-     * Path to a Playwright storageState JSON file.
-     * Restores cookies, localStorage, and sessionStorage from a previous session.
-     */
-    storageState?: string;
-    /**
-     * Connect to an existing browser via Chrome DevTools Protocol instead
-     * of launching one. Format: `"http://localhost:9222"`. When set,
-     * `headless`, `stealth.args`, `recordVideo`, etc. are ignored — the
-     * existing browser's configuration is used.
-     */
-    cdpUrl?: string;
-    /**
-     * Enable video recording of the browser session.
-     * Pass `true` for default dir (`./browser-videos`) or `{ dir: "/path" }`.
-     */
-    recordVideo?: boolean | {
-        dir: string;
-    };
-    /**
-     * Secure credential vault. The LLM never sees real values — only
-     * placeholders like `{{email}}`, `{{password}}`. Real values are
-     * injected at execution time and scrubbed from all logs.
-     */
-    credentials?: CredentialVault;
-    /**
-     * Enable stealth mode to avoid bot detection.
-     * Pass `true` for sensible defaults or a `StealthConfig` object for fine control.
-     * Patches navigator.webdriver, plugins, permissions, WebGL, and more.
-     */
-    stealth?: boolean | StealthConfig;
-    /**
-     * Simulate human-like behavior — jittered clicks, variable typing speed,
-     * mouse movement curves, random micro-pauses.
-     * Pass `true` for defaults or a `HumanizeConfig` for fine control.
-     */
-    humanize?: boolean | HumanizeConfig;
-    /**
-     * Unified memory config — persist browser sessions, decisions, and
-     * summaries of past runs. Same config as Agent and VoiceAgent.
-     */
-    memory?: UnifiedMemoryConfig;
-    /** Skills — pre-packaged or learned tool bundles. */
-    skills?: Array<_agentium_core.Skill | string>;
-    /**
-     * Custom tools that the BrowserAgent itself can invoke during a run.
-     * The agent emits `{ "action": "tool", "name": "<tool>", "args": {...} }`
-     * and we dispatch to the tool's `execute(args)`. Use this for 2FA codes,
-     * API calls, file I/O, calling out to other agents — anything the
-     * browser can't do alone.
-     */
-    tools?: ToolDef[];
-    /** Cost tracker — track vision model token usage and enforce budgets across browser runs. */
-    costTracker?: CostTracker;
-    logLevel?: LogLevel;
-    eventBus?: EventBus;
-}
-interface BrowserRunOpts {
-    /** Override startUrl from config */
-    startUrl?: string;
-    /** Per-run model API key override */
-    apiKey?: string;
-    /** Session identifier for memory persistence and event tracking */
-    sessionId?: string;
-    /** User identifier for memory personalization */
-    userId?: string;
-    /** Path to save storageState (cookies/auth) after the run completes */
-    saveStorageState?: string;
-    /** Per-run override of `maxSteps`. */
-    maxSteps?: number;
-}
-interface BrowserRunOutput {
-    /** Final text result produced by the agent */
-    result: string;
-    /** Whether the task completed successfully (vs maxSteps exhausted or fail) */
-    success: boolean;
-    /** Full action history with screenshots */
-    steps: BrowserStep[];
-    /** URL at completion */
-    finalUrl: string;
-    /** Last screenshot captured */
-    finalScreenshot: Buffer;
-    /** Total time taken in milliseconds */
-    durationMs: number;
-    /** Video file path (if recordVideo was enabled) */
-    videoPath?: string;
-    /** Extracted content from every `extract` action, in chronological order. */
-    extractedContent?: string[];
-}
-interface BrowserStep {
-    index: number;
-    action: BrowserAction;
-    /** Screenshot taken before this action was executed. May be empty if useVision=false. */
-    screenshot: Buffer;
-    pageUrl: string;
-    pageTitle: string;
-    timestamp: Date;
-    /** Simplified DOM snapshot (if useDOM is enabled) */
-    dom?: string;
-    /** Free-form result string produced by the action (e.g. extract output). */
-    output?: string;
-    /** Whether this step succeeded (vs threw / failed locator). Default: true. */
-    ok?: boolean;
-    /** Model's chain-of-thought reasoning (if `useThinking` was on). */
-    thinking?: string;
-    /** Model's evaluation of whether the previous action met its goal. */
-    evaluationPreviousGoal?: string;
-    /** Model's running memory of important state. */
-    memory?: string;
-    /** Model's stated next goal for this step. */
-    nextGoal?: string;
-}
-/**
- * Structured envelope the model returns when `useThinking: true`. Inspired
- * by browser-use's `AgentOutput`. Every field is optional from a runtime
- * standpoint — only `action` is required for execution.
- */
-interface AgentOutput {
-    thinking?: string;
-    evaluationPreviousGoal?: string;
-    memory?: string;
-    nextGoal?: string;
-    action: BrowserAction | BrowserAction[];
-}
-interface StealthConfig {
-    /**
-     * Remove `navigator.webdriver` flag and patch common detection vectors
-     * (plugins, languages, permissions, WebGL, etc.). Default when stealth=true.
-     */
-    patchFingerprint?: boolean;
-    /** Custom User-Agent string. A realistic one is used by default. */
-    userAgent?: string;
-    /** Browser locale. Default: "en-US" */
-    locale?: string;
-    /** Timezone ID (IANA). Default: "America/New_York" */
-    timezone?: string;
-    /** Fake geolocation */
-    geolocation?: {
-        latitude: number;
-        longitude: number;
-        accuracy?: number;
-    };
-    /** Ignore HTTPS certificate errors. Default: false (secure). Only enable for local testing. */
-    ignoreHTTPSErrors?: boolean;
-    /**
-     * `window.devicePixelRatio` to emulate. Default: `1`. Set to `2` to mimic
-     * a Retina display (sharper screenshots at 2× cost). Setting to 2 on a
-     * non-Retina host display can cause the headed window to look zoomed-out
-     * or stretched because the OS compositor downsamples a 2× surface.
-     */
-    deviceScaleFactor?: number;
-    /** HTTP/SOCKS proxy. Format: "http://user:pass@host:port" */
-    proxy?: {
-        server: string;
-        username?: string;
-        password?: string;
-    };
-}
-interface HumanizeConfig {
-    /** Per-character typing delay range in ms. Default: [40, 120] */
-    typingDelay?: [number, number];
-    /** Random pixel offset added to click coordinates. Default: 3 */
-    clickJitter?: number;
-    /** Extra random pause between actions in ms range. Default: [200, 800] */
-    actionDelay?: [number, number];
-    /** Simulate human-like mouse movement to target before clicking. Default: true */
-    mouseMovement?: boolean;
-}
-interface PageInfo {
-    url: string;
-    title: string;
-    viewportSize: {
-        width: number;
-        height: number;
-    };
-}
-
-declare class BrowserAgent {
-    readonly name: string;
-    readonly eventBus: EventBus;
-    private model;
-    private pageExtractionLLM;
-    private fallbackModel;
-    private useThinking;
-    private historyWindow;
-    private instructions?;
-    private extendSystemMessage?;
-    private overrideSystemMessage?;
-    private maxSteps;
-    private maxFailures;
-    private maxActionsPerStep;
-    private initialActions;
-    private useVision;
-    private directlyOpenUrl;
-    private headless;
-    private viewport;
-    private defaultStartUrl?;
-    private waitAfterAction;
-    private maxRepeats;
-    private useDOM;
-    private allowEvaluate;
-    private allowedDomains?;
-    private prohibitedDomains?;
-    private storageState?;
-    private cdpUrl?;
-    private recordVideo?;
-    private credentials?;
-    private stealth?;
-    private humanize?;
-    private tools;
-    private costTracker;
-    private memoryManager;
-    private logger;
-    /** Access the MemoryManager (if memory is configured). */
-    get memory(): MemoryManager | null;
-    constructor(config: BrowserAgentConfig);
-    run(task: string, opts?: BrowserRunOpts): Promise<BrowserRunOutput>;
-    /**
-     * Returns a ToolDef that lets a regular Agent delegate browser tasks
-     * to this BrowserAgent.
-     */
-    asTool(config?: {
-        name?: string;
-        description?: string;
-    }): ToolDef;
-    private shouldCaptureVision;
-    private detectUrlInTask;
-    private assertDomainAllowed;
-    /**
-     * Build the message array sent to the model: system prompt + a compact
-     * summary of older turns (if any) + the most recent `historyWindow`
-     * turns verbatim + the current step's user message.
-     *
-     * Inspired by browser-use's history compaction. Keeps tokens bounded
-     * while giving the model meaningful context about what it already
-     * tried.
-     */
-    private buildMessages;
-    /**
-     * Call the primary model, retrying once with `fallbackModel` on
-     * transient errors (5xx, 429, network). Returns the response or
-     * `null` if both models failed.
-     */
-    private callModelWithFallback;
-    private isTransientError;
-    /**
-     * Parse the model's raw response into an `AgentOutput`. Tolerant to
-     * three shapes:
-     *  - Full envelope: { thinking, evaluation_previous_goal, action, ... }
-     *  - Raw action object (legacy / `useThinking: false`)
-     *  - Raw action array
-     *
-     * Also strips ```json fences the model occasionally adds.
-     */
-    private parseEnvelope;
-    /**
-     * Quick post-navigation health check. If the page came back blank
-     * (no body text and no interactive elements), reload once and wait
-     * for stable. This catches the "FreightOS half-loaded font test" class
-     * of failure before the LLM ever sees it.
-     */
-    private navigationHealthCheck;
-    /**
-     * Force-finalize the run with whatever partial data the agent has.
-     * Called when:
-     *   - `maxSteps` is exhausted without an explicit `done`,
-     *   - `maxFailures` is exceeded,
-     *   - the model can't produce parseable JSON enough times to make
-     *     forward progress.
-     * The result is composed from `extractedContent` + the last few
-     * action summaries so the caller gets something useful instead of
-     * just a one-line error.
-     */
-    private forceDone;
-    private finalize;
-    /**
-     * Execute a single action. Returns `{ output?, didNavigate? }` for the
-     * caller's state tracking. May throw — the loop handles failure-budget
-     * accounting in that case.
-     */
-    private executeAction;
-    private sleep;
-    /**
-     * Parse a quoted target keyword from a click action's `description`.
-     * Returns `undefined` for generic / ambiguous labels (login buttons,
-     * close, OK, etc.) where a substring text match could fire on the
-     * wrong element.
-     */
-    private extractClickKeyword;
-}
-
-/**
- * Playwright wrapper with stealth anti-detection, human-like behavior,
- * indexed DOM element resolution, and a rich action vocabulary.
- *
- * The `BrowserProvider` is intentionally LLM-agnostic — it exposes the
- * primitives that `BrowserAgent` orchestrates via vision+DOM reasoning.
- */
-declare class BrowserProvider {
-    private browser;
-    private context;
-    private page;
-    private pages;
-    private activeTabId;
-    private tabCounter;
-    private _viewport;
-    private _videoDir?;
-    private _humanize?;
-    /**
-     * Most recent DOM snapshot (one per `extractDOM` call). Indexed actions
-     * (`clickByIndex`, `inputByIndex`, …) resolve their `index` against this.
-     */
-    private _lastDom;
-    /** True if we connected over CDP (don't tear down the browser on close). */
-    private _attached;
-    constructor();
-    launch(opts?: {
-        headless?: boolean;
-        viewport?: {
-            width: number;
-            height: number;
-        };
-        storageState?: string;
-        recordVideo?: boolean | {
-            dir: string;
-        };
-        stealth?: boolean | StealthConfig;
-        humanize?: boolean | HumanizeConfig;
-        cdpUrl?: string;
-    }): Promise<void>;
-    saveStorageState(path: string): Promise<void>;
-    navigate(url: string): Promise<void>;
-    back(): Promise<void>;
-    screenshot(): Promise<Buffer>;
-    /** Viewport size in CSS pixels (matches screenshot dimensions). */
-    get viewport(): {
-        width: number;
-        height: number;
-    };
-    /** Most recent DOM snapshot. Each entry has a stable `index`. */
-    get lastDom(): DomElement[];
-    click(x: number, y: number): Promise<void>;
-    type(text: string): Promise<void>;
-    clickAndType(x: number, y: number, text: string): Promise<void>;
-    pressKey(key: string): Promise<void>;
-    /**
-     * Send arbitrary keyboard keys / shortcuts. Accepts a single
-     * Playwright key spec (`"Enter"`, `"Control+l"`, `"Shift+ArrowDown"`)
-     * or a space-separated sequence (`"Tab Tab Enter"`).
-     */
-    sendKeys(keys: string): Promise<void>;
-    scroll(direction: "up" | "down", amount?: number): Promise<void>;
-    /**
-     * Build a Playwright locator for a DOM-snapshot index. Each `extractDOM`
-     * call tags surviving elements with `data-bua-idx="<n>"`; we resolve by
-     * that attribute. Returns null if the index is unknown.
-     */
-    private locatorForIndex;
-    /**
-     * Click an element by its DOM-snapshot index. The most reliable click
-     * path on dynamic pages — survives layout shifts and DPR oddities.
-     */
-    clickByIndex(index: number, opts?: {
-        timeout?: number;
-    }): Promise<boolean>;
-    /**
-     * Focus an indexed input, optionally clear it, and type. Returns false
-     * if the index couldn't be resolved or the input couldn't be focused.
-     */
-    inputByIndex(index: number, text: string, opts?: {
-        clear?: boolean;
-        submit?: boolean;
-        timeout?: number;
-    }): Promise<boolean>;
-    uploadFileByIndex(index: number, path: string): Promise<boolean>;
-    /** Scroll the indexed element into view (no click). */
-    scrollIntoViewByIndex(index: number): Promise<boolean>;
-    /**
-     * Deterministic, DOM-based click using Playwright's text locator.
-     *
-     * Returns `true` if a matching, visible, clickable element was found and
-     * clicked within `timeout` ms; `false` otherwise (so the caller can fall
-     * back to coordinate clicking). Substring-matches by default — e.g.
-     * `clickByText("Cheapest")` matches "Cheapest · 23-28 days · $2,550".
-     */
-    clickByText(keyword: string, opts?: {
-        timeout?: number;
-    }): Promise<boolean>;
-    /**
-     * Scroll the first occurrence of `text` into view. Returns false if no
-     * match was found within the timeout.
-     */
-    findText(text: string, opts?: {
-        timeout?: number;
-    }): Promise<boolean>;
-    /**
-     * Read the options of a native `<select>` at the given DOM-snapshot
-     * index. Returns `[]` if the element is not a `<select>`.
-     */
-    dropdownOptions(index: number): Promise<{
-        value: string;
-        label: string;
-        selected: boolean;
-    }[]>;
-    /**
-     * Select an option in a native `<select>` by its visible text or value.
-     * Returns false if the element isn't a `<select>` or no option matched.
-     */
-    selectDropdown(index: number, text: string): Promise<boolean>;
-    /**
-     * Run arbitrary JS in the page context. The caller is responsible for
-     * gating this behind a config flag — the BrowserAgent only routes the
-     * `evaluate` action here when `allowEvaluate: true`.
-     *
-     * The code is wrapped in `(async () => { ... })()` and the return value
-     * is coerced to a string for the model.
-     */
-    evaluate(code: string): Promise<string>;
-    /**
-     * Returns a clean text representation of the visible page body, with
-     * optional link extraction. Used by the BrowserAgent's `extract` action
-     * — the text is passed to a (usually cheap) LLM with the user's query.
-     */
-    pageText(opts?: {
-        extractLinks?: boolean;
-        maxChars?: number;
-    }): Promise<string>;
-    /**
-     * Snapshot the interactive elements visible in the viewport, tag each
-     * with a `data-bua-idx="<n>"` attribute (used by indexed actions), and
-     * return:
-     *   - `text`: a human-readable string fed to the model
-     *   - `elements`: the structured list with stable indices
-     *   - `scroll`: spatial context (pages above/below, hidden interactive count)
-     *
-     * Five properties matter for accuracy:
-     *  - **Hit-tested**: each listed coordinate / index actually reaches the
-     *    labeled element (overlays / occlusion skip the entry).
-     *  - **Visibility filtered (with parent chain)**: an element is dropped
-     *    if itself OR any ancestor is `display:none`, `visibility:hidden`,
-     *    `pointer-events:none`, or near-zero opacity.
-     *  - **Shadow DOM piercing**: traverses open shadow roots so custom
-     *    elements / web components are visible to the agent.
-     *  - **Same-origin iframes**: walks into each accessible iframe and
-     *    includes its interactive elements (offset by the iframe's screen
-     *    position so the coordinates the model sees are still viewport-
-     *    relative).
-     *  - **`cursor: pointer` fallback pass**: catches custom React widgets
-     *    that have no semantic role/href/onclick but are clickable.
-     */
-    extractDOM(opts?: {
-        maxElements?: number;
-    }): Promise<DomSnapshot>;
-    /**
-     * Install the `__buaExtract` global on the main page. Idempotent —
-     * subsequent calls are no-ops.
-     */
-    private installExtractorScript;
-    /**
-     * The extractor source. Lives in its own method so we can also inject
-     * it into iframes that haven't yet had it loaded.
-     *
-     * This function intentionally runs entirely in the page context. It:
-     *   - traverses the regular DOM + open shadow roots (deep)
-     *   - applies a parent-chain visibility filter
-     *   - applies a `cursor:pointer` second pass for custom widgets
-     *   - hit-tests each candidate at its center to avoid overlay collisions
-     *   - returns scroll context (pages above/below, hidden counts)
-     *   - tags survivors with `data-bua-idx` for indexed actions
-     */
-    private extractorScriptSource;
-    getPageInfo(): Promise<PageInfo>;
-    waitForStable(minWait?: number): Promise<void>;
-    newTab(url?: string): Promise<string>;
-    switchTab(tabId: string): Promise<void>;
-    closeTab(tabId: string): Promise<void>;
-    listTabs(): {
-        id: string;
-        url: string;
-        active: boolean;
-    }[];
-    get currentTabId(): string;
-    getVideoPath(tabId?: string): Promise<string | null>;
-    get videoDir(): string | undefined;
-    close(): Promise<void>;
-    /** Add small random offset to coordinates to avoid pixel-perfect bot patterns. */
-    private jitter;
-    /**
-     * Safety net: clamp coordinates returned by the vision model to the actual
-     * viewport. If a model occasionally returns image-space coordinates from a
-     * 2x screenshot (despite our `scale: "css"` fix), this prevents Playwright
-     * from clicking at e.g. (2200, 1300) and either erroring or landing on a
-     * random off-screen element.
-     */
-    private clampToViewport;
-    /**
-     * Simulate human mouse movement using smoothstep interpolation.
-     */
-    private humanMouseMove;
-    /** Small random pause after an interaction. */
-    private humanPause;
-    private randInt;
-    private ensurePage;
-    private ensureContext;
-    private sleep;
-}
-
-export { type AgentOutput, type BrowserAction, BrowserAgent, type BrowserAgentConfig, BrowserProvider, type BrowserRunOpts, type BrowserRunOutput, type BrowserStep, CredentialVault, type DomElement, type DomScrollContext, type DomSnapshot, type HumanizeConfig, type PageInfo, type StealthConfig };
+export { BrowserAgent } from "./browser-agent.js";
+export { BrowserProvider } from "./browser-provider.js";
+export { CredentialVault } from "./credential-vault.js";
+export type { AgentOutput, BrowserAction, BrowserAgentConfig, BrowserRunOpts, BrowserRunOutput, BrowserStep, DomElement, DomScrollContext, DomSnapshot, HumanizeConfig, PageInfo, StealthConfig, } from "./types.js";
+//# sourceMappingURL=index.d.ts.map