crawlex 1.0.4

//! Human-like interaction primitives driven through the Chrome DevTools
//! Protocol Input domain.
//!
//! Bot detectors (reCAPTCHA v3, Cloudflare Turnstile, PerimeterX, DataDome)
//! score on **behavioural** signals, not just fingerprints. This module
//! produces the signal shapes those detectors look for:
//!
//! * **Mouse trajectories** are generated by the WindMouse engine
//!   ([`crate::render::motion`]) with Fitts-derived total movement time,
//!   Ornstein-Uhlenbeck jitter, and probabilistic overshoot.
//! * **Event sequence integrity**: every click emits
//!   `mousemove…mouseover+mouseenter → mousedown → mouseup → click`. A
//!   click without a preceding move is a dead giveaway for modern ML
//!   classifiers — this primitive never dispatches one.
//! * **Keystrokes** are scheduled by [`crate::render::keyboard`] using
//!   log-normal hold times, log-logistic inter-key flights, Pareto
//!   thinking pauses, and probabilistic typos with backspace correction.
//! * **Scroll** uses wheel events (not `Element.scrollIntoView`) so
//!   Chrome dispatches real `wheel` handlers.
//!
//! The active [`MotionProfile`](crate::render::motion::MotionProfile) is
//! read from the process-wide ambient slot (`MotionProfile::active()`),
//! which the CLI / pool sets at startup from `Config::motion_profile`.

use crate::render::chrome::page::Page;
use crate::render::chrome_protocol::cdp::browser_protocol::input::{
    DispatchKeyEventParams, DispatchKeyEventType, DispatchMouseEventParams, DispatchMouseEventType,
    InsertTextParams, MouseButton,
};
use crate::render::chrome_protocol::cdp::js_protocol::runtime::EvaluateParams;
use crate::render::keyboard::{KeyEvent, TypingEngine};
use crate::render::motion::lifecycle::{HIDE_SNIPPET, SHOW_SNIPPET};
use crate::render::motion::scroll::{schedule_for_active_profile as scroll_schedule, ScrollParams};
use crate::render::motion::{
    fatigue, idle::IdleState, scroll as scroll_mod, MotionEngine, MotionProfile, Point, TimedPoint,
};
use rand::rngs::SmallRng;
use rand::RngExt;
use serde::Deserialize;
use std::sync::Arc;
use std::time::Duration;

use crate::{Error, Result};

/// Rectangle in viewport coords returned by our DOM probe.
#[derive(Debug, Clone, Copy, Deserialize)]
pub struct Rect {
    pub x: f64,
    pub y: f64,
    pub w: f64,
    pub h: f64,
}

/// Resolve a selector (full DSL — css/text/role/label/testid/xpath chains)
/// to a bounding rect. Returns `None` when not found or not visible.
async fn element_rect(page: &Page, selector: &str) -> Result<Option<Rect>> {
    crate::render::selector::resolve_rect(page, selector).await
}

/// Keep a mouse position in the page so successive actions curve off each
/// other rather than all starting at (0,0). The script runner / action
/// executor threads this through every click and scroll.
#[derive(Clone, Copy, Debug, Default)]
pub struct MousePos {
    pub x: f64,
    pub y: f64,
}

/// Dispatch a single `mouseMoved` CDP event at `(x, y)`.
async fn dispatch_move(page: &Page, x: f64, y: f64) -> Result<()> {
    let params = DispatchMouseEventParams::builder()
        .r#type(DispatchMouseEventType::MouseMoved)
        .x(x)
        .y(y)
        .build()
        .map_err(|e| Error::Render(format!("mouse params: {e}")))?;
    page.execute(params)
        .await
        .map_err(|e| Error::Render(format!("mouse move: {e}")))?;
    Ok(())
}

/// Walk a precomputed trajectory as a series of CDP `mouseMoved` events,
/// honouring the per-step delay baked into each `TimedPoint`.
async fn walk_trajectory(page: &Page, pts: &[TimedPoint]) -> Result<()> {
    for p in pts {
        dispatch_move(page, p.x, p.y).await?;
        if p.delay_ms > 0 {
            tokio::time::sleep(Duration::from_millis(p.delay_ms)).await;
        }
    }
    Ok(())
}

/// Move the virtual cursor from `from` to `(x, y)` along a WindMouse
/// trajectory. Returns the new cursor position so callers can chain.
pub async fn mouse_move_to(page: &Page, from: MousePos, x: f64, y: f64) -> Result<MousePos> {
    let profile = MotionProfile::active();
    let mut engine = MotionEngine::new(profile);
    let pts = engine.trajectory(
        Point {
            x: from.x,
            y: from.y,
        },
        Point { x, y },
        // Unknown target width: assume ~40px (typical button). Fitts MT
        // still scales with distance — the width only nudges it a little.
        40.0,
    );
    walk_trajectory(page, &pts).await?;
    Ok(MousePos { x, y })
}

/// Click a CSS selector with full event-sequence integrity:
/// WindMouse trajectory → mouseover/mouseenter → post-move pause →
/// mouseDown → hold → mouseUp. The cursor state advances to the click
/// point; callers chain subsequent moves from there.
pub async fn click_selector(page: &Page, css: &str, from: MousePos) -> Result<MousePos> {
    let rect = element_rect(page, css)
        .await?
        .ok_or_else(|| Error::Render(format!("element not found: {css}")))?;
    let mut rng = rand::make_rng::<SmallRng>();
    // Jitter inside the bbox, avoiding the exact center — ML classifiers
    // flag clicks that always land on (x+w/2, y+h/2).
    let tx = rect.x + rect.w * rng.random_range(0.25..0.75);
    let ty = rect.y + rect.h * rng.random_range(0.25..0.75);
    click_point(page, from, tx, ty, rect.w.min(rect.h).max(10.0)).await
}

/// Shared click primitive: WindMouse move, mouseover/enter burst,
/// post-move pause, mouseDown, hold, mouseUp. Exposed at crate scope so
/// `ref_resolver` can share the exact sequence without duplicating CDP
/// wire code.
pub async fn click_point(
    page: &Page,
    from: MousePos,
    x: f64,
    y: f64,
    target_width: f64,
) -> Result<MousePos> {
    let profile = MotionProfile::active();
    let params = profile.params();
    let mut engine = MotionEngine::new(profile);

    let pts = engine.trajectory(
        Point {
            x: from.x,
            y: from.y,
        },
        Point { x, y },
        target_width,
    );
    walk_trajectory(page, &pts).await?;

    // Event-sequence integrity: detectors flag mousedown-without-mouseover.
    // Real browsers fire mouseover + mouseenter the instant the cursor
    // enters the target's bounding box — replay that here by re-emitting a
    // `mouseMoved` at the final point so Chrome's hit-testing machinery
    // dispatches the over/enter events itself.
    if params.emit_mouseover {
        dispatch_move(page, x, y).await?;
    }

    let mut rng = rand::make_rng::<SmallRng>();
    let pause = u64_range(
        &mut rng,
        params.post_move_pause_ms_min,
        params.post_move_pause_ms_max,
    );
    if pause > 0 {
        tokio::time::sleep(Duration::from_millis(pause)).await;
    }

    let press = DispatchMouseEventParams::builder()
        .r#type(DispatchMouseEventType::MousePressed)
        .x(x)
        .y(y)
        .button(MouseButton::Left)
        .click_count(1)
        .build()
        .map_err(|e| Error::Render(format!("click params: {e}")))?;
    page.execute(press)
        .await
        .map_err(|e| Error::Render(format!("click: {e}")))?;

    let hold = u64_range(
        &mut rng,
        params.mouse_down_pause_ms_min,
        params.mouse_down_pause_ms_max,
    );
    if hold > 0 {
        tokio::time::sleep(Duration::from_millis(hold)).await;
    }

    let release = DispatchMouseEventParams::builder()
        .r#type(DispatchMouseEventType::MouseReleased)
        .x(x)
        .y(y)
        .button(MouseButton::Left)
        .click_count(1)
        .build()
        .map_err(|e| Error::Render(format!("click params: {e}")))?;
    page.execute(release)
        .await
        .map_err(|e| Error::Render(format!("click: {e}")))?;

    Ok(MousePos { x, y })
}

fn u64_range(rng: &mut SmallRng, lo: u64, hi: u64) -> u64 {
    if hi <= lo {
        return lo;
    }
    rng.random_range(lo..hi)
}

pub async fn type_text(page: &Page, selector: &str, text: &str) -> Result<()> {
    // Use the selector engine so callers get the full DSL here too.
    if !crate::render::selector::focus(page, selector).await? {
        return Err(Error::Render(format!("focus failed: {selector}")));
    }
    dispatch_typing(page, text).await
}

/// Dispatch `text` as a scheduled keystroke timeline on the currently
/// focused element. Public so `ref_resolver` can reuse the exact same
/// distribution family when typing into an AX-snapshot-resolved node.
pub async fn dispatch_typing(page: &Page, text: &str) -> Result<()> {
    let profile = MotionProfile::active();
    let mut engine = TypingEngine::new(profile);
    let events = engine.schedule(text);

    for ev in events {
        match ev {
            KeyEvent::Pause { ms } => {
                if ms > 0 {
                    tokio::time::sleep(Duration::from_millis(ms)).await;
                }
            }
            KeyEvent::Char { ch, hold_ms } => {
                dispatch_char(page, ch, hold_ms).await?;
            }
            KeyEvent::Typo { wrong, hold_ms } => {
                dispatch_char(page, wrong, hold_ms).await?;
            }
            KeyEvent::Backspace { hold_ms } => {
                let down = DispatchKeyEventParams::builder()
                    .r#type(DispatchKeyEventType::KeyDown)
                    .key("Backspace".to_string())
                    .code("Backspace".to_string())
                    .build()
                    .map_err(|e| Error::Render(format!("backspace params: {e}")))?;
                page.execute(down)
                    .await
                    .map_err(|e| Error::Render(format!("backspace: {e}")))?;
                if hold_ms > 0 {
                    tokio::time::sleep(Duration::from_millis(hold_ms)).await;
                }
                let up = DispatchKeyEventParams::builder()
                    .r#type(DispatchKeyEventType::KeyUp)
                    .key("Backspace".to_string())
                    .code("Backspace".to_string())
                    .build()
                    .map_err(|e| Error::Render(format!("backspace params: {e}")))?;
                page.execute(up)
                    .await
                    .map_err(|e| Error::Render(format!("backspace: {e}")))?;
            }
        }
    }
    Ok(())
}

async fn dispatch_char(page: &Page, ch: char, hold_ms: u64) -> Result<()> {
    if ch.is_ascii() && !ch.is_control() {
        let text = ch.to_string();
        // KeyDown.
        let down = DispatchKeyEventParams::builder()
            .r#type(DispatchKeyEventType::KeyDown)
            .text(text.clone())
            .build()
            .map_err(|e| Error::Render(format!("key params: {e}")))?;
        page.execute(down)
            .await
            .map_err(|e| Error::Render(format!("key: {e}")))?;
        if hold_ms > 0 {
            tokio::time::sleep(Duration::from_millis(hold_ms)).await;
        }
        let up = DispatchKeyEventParams::builder()
            .r#type(DispatchKeyEventType::KeyUp)
            .text(text)
            .build()
            .map_err(|e| Error::Render(format!("key params: {e}")))?;
        page.execute(up)
            .await
            .map_err(|e| Error::Render(format!("key: {e}")))?;
    } else {
        let p = InsertTextParams::builder()
            .text(ch.to_string())
            .build()
            .map_err(|e| Error::Render(format!("insert params: {e}")))?;
        page.execute(p)
            .await
            .map_err(|e| Error::Render(format!("insert: {e}")))?;
        if hold_ms > 0 {
            tokio::time::sleep(Duration::from_millis(hold_ms)).await;
        }
    }
    Ok(())
}

pub async fn scroll_by(page: &Page, dy: f64, from: MousePos) -> Result<()> {
    // Delegate to the scroll scheduler: bursts with bell-curve velocity
    // interleaved with Pareto-distributed reading dwells (B.1). Fast
    // profile collapses to the legacy flat 120px tick flow.
    let ticks = scroll_schedule(dy);
    for tick in ticks {
        if tick.delta_y.abs() >= 0.5 {
            let params = DispatchMouseEventParams::builder()
                .r#type(DispatchMouseEventType::MouseWheel)
                .x(from.x.max(10.0))
                .y(from.y.max(10.0))
                .delta_x(0.0)
                .delta_y(tick.delta_y)
                .build()
                .map_err(|e| Error::Render(format!("wheel params: {e}")))?;
            page.execute(params)
                .await
                .map_err(|e| Error::Render(format!("wheel: {e}")))?;
        }
        if tick.delay_ms > 0 {
            tokio::time::sleep(Duration::from_millis(tick.delay_ms)).await;
        }
    }
    Ok(())
}

/// Emit a page-lifecycle transition: hide (fires `visibilitychange` +
/// window `blur`, flips `document.hidden=true`) or show (symmetric pair).
/// Used by background drivers to inject realistic focus/tab-switch events
/// — see `motion::lifecycle::schedule` for the timing plan.
pub async fn emit_page_hidden(page: &Page) -> Result<()> {
    let _ = eval_js(page, HIDE_SNIPPET).await?;
    Ok(())
}

pub async fn emit_page_visible(page: &Page) -> Result<()> {
    let _ = eval_js(page, SHOW_SNIPPET).await?;
    Ok(())
}

/// Spawn a background tokio task that dispatches ambient idle-drift
/// `mouseMoved` events when no action is active. Drops automatically when
/// `state.action_active` flips or the page connection dies. Fast profile
/// is a no-op — we intentionally preserve throughput there.
pub fn spawn_idle_drift(page: std::sync::Arc<Page>, origin: MousePos, state: Arc<IdleState>) {
    let profile = MotionProfile::active();
    if matches!(profile, MotionProfile::Fast) {
        return;
    }
    let seed = rand::make_rng::<SmallRng>().random::<u64>();
    tokio::spawn(async move {
        let mut drift = crate::render::motion::IdleDrift::for_profile(profile, seed);
        // Bounded iterations as a belt-and-braces against leaking the task
        // if the caller forgets to flip `action_active`. 6h at ~1s cadence.
        for _ in 0..21_600u64 {
            if state.is_action_active() {
                tokio::time::sleep(Duration::from_millis(200)).await;
                continue;
            }
            let (dx, dy) = drift.next_offset();
            let x = (origin.x + dx).max(1.0);
            let y = (origin.y + dy).max(1.0);
            // Best-effort: if the move fails (page closed), bail the task.
            if dispatch_move(&page, x, y).await.is_err() {
                return;
            }
            let sleep_ms = drift.next_delay_ms();
            tokio::time::sleep(Duration::from_millis(sleep_ms)).await;
        }
    });
}

/// Consume the active session's fatigue factor and return a scroll-param
/// override with decayed per-tick velocity. Handy for callers that build
/// their own scroll schedules.
pub fn scroll_params_with_fatigue() -> ScrollParams {
    let mut p = ScrollParams::for_profile(MotionProfile::active());
    let factor = fatigue::current_velocity_factor();
    p.peak_tick_px *= factor;
    p
}

// Keep the `scroll_mod` import live even if callers only use the
// re-exported schedule helper.
#[allow(dead_code)]
fn _assert_scroll_mod_linked() -> usize {
    std::mem::size_of::<scroll_mod::ScrollTick>()
}

pub async fn wait_for_selector(page: &Page, css: &str, timeout_ms: u64) -> Result<()> {
    let deadline = std::time::Instant::now() + Duration::from_millis(timeout_ms);
    loop {
        if element_rect(page, css).await?.is_some() {
            return Ok(());
        }
        if std::time::Instant::now() >= deadline {
            return Err(Error::Render(format!("wait_for_selector timeout: {css}")));
        }
        tokio::time::sleep(Duration::from_millis(50)).await;
    }
}

pub async fn eval_js(page: &Page, script: &str) -> Result<serde_json::Value> {
    let params = EvaluateParams::builder()
        .expression(script.to_string())
        .return_by_value(true)
        .await_promise(true)
        .build()
        .map_err(|e| Error::Render(format!("eval params: {e}")))?;
    let res = page
        .evaluate_expression(params)
        .await
        .map_err(|e| Error::Render(format!("eval: {e}")))?;
    Ok(res.value().cloned().unwrap_or(serde_json::Value::Null))
}