captchaforge 0.2.35

use super::*;

// ─── VLM provider abstraction ────────────────────────────────────────────────
//
// Three back-ends supported, picked in this priority order at construction
// time:
//
//   1. **Anthropic Claude** — when `ANTHROPIC_API_KEY` is set in the env.
//      Strongest visual reasoning of the three; recommended for image-grid
//      captchas that need spatial understanding ("select all crosswalks").
//      Default model: `claude-haiku-4-5` (fast + cheap; switch to
//      `claude-sonnet-4-6` via `CAPTCHAFORGE_VLM_MODEL` for harder grids).
//
//   2. **OpenAI** — when `OPENAI_API_KEY` is set and Anthropic isn't.
//      Default model: `gpt-4o-mini`.
//
//   3. **Ollama (local)** — fallback when no API key is present. The
//      "anyone-can-run-this-bench" path: works with a locally-installed
//      Ollama running `llama3.2-vision:11b` (the bench's reproducibility
//      contract — no API keys, no paid SaaS, just `ollama serve` +
//      `ollama pull llama3.2-vision:11b`). Endpoint defaults to
//      `http://localhost:11434`.
//
// The provider abstraction is honoured even when `with_endpoint` /
// `with_model` are called explicitly — those override the per-provider
// defaults but don't change the back-end. Use `with_provider` to switch
// back-ends from code (e.g. force Ollama for an integration test).

const DEFAULT_OLLAMA_BASE: &str = "http://localhost:11434";
const DEFAULT_OLLAMA_MODEL: &str = "llama3.2-vision:11b";
const DEFAULT_ANTHROPIC_BASE: &str = "https://api.anthropic.com";
const DEFAULT_ANTHROPIC_MODEL: &str = "claude-haiku-4-5";
const DEFAULT_OPENAI_BASE: &str = "https://api.openai.com";
const DEFAULT_OPENAI_MODEL: &str = "gpt-4o-mini";

const ENV_VLM_ENDPOINT: &str = "CAPTCHAFORGE_VLM_ENDPOINT";
const ENV_VLM_MODEL: &str = "CAPTCHAFORGE_VLM_MODEL";
const ENV_ANTHROPIC_KEY: &str = "ANTHROPIC_API_KEY";
const ENV_OPENAI_KEY: &str = "OPENAI_API_KEY";

/// Which VLM back-end to call. See module-level docs for the
/// auto-detection priority + per-provider defaults.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum VlmProvider {
    /// Local Ollama instance — the default-everyone-can-reproduce path.
    /// No API key required; talks to `/api/generate`.
    Ollama,
    /// Anthropic Claude vision API — picked when `ANTHROPIC_API_KEY` is set.
    /// Talks to `/v1/messages` with image content blocks.
    Anthropic,
    /// OpenAI GPT-4o vision — picked when `OPENAI_API_KEY` is set
    /// (and Anthropic is not). Talks to `/v1/chat/completions` with
    /// image_url content parts.
    OpenAI,
}

impl VlmProvider {
    fn default_endpoint(self) -> &'static str {
        match self {
            Self::Ollama => DEFAULT_OLLAMA_BASE,
            Self::Anthropic => DEFAULT_ANTHROPIC_BASE,
            Self::OpenAI => DEFAULT_OPENAI_BASE,
        }
    }
    fn default_model(self) -> &'static str {
        match self {
            Self::Ollama => DEFAULT_OLLAMA_MODEL,
            Self::Anthropic => DEFAULT_ANTHROPIC_MODEL,
            Self::OpenAI => DEFAULT_OPENAI_MODEL,
        }
    }
}

/// Raw Ollama `/api/generate` request body.
#[derive(Debug, Serialize)]
struct OllamaRequest<'a> {
    model: &'a str,
    prompt: &'a str,
    images: Vec<String>,
    stream: bool,
}

/// Relevant fields from Ollama `/api/generate` response.
#[derive(Debug, Deserialize)]
struct OllamaResponse {
    response: String,
}

/// Send a screenshot (base64 PNG/JPEG) and a text prompt to a VLM endpoint.
/// Returns the model's raw text reply. Dispatches to the appropriate
/// per-provider adapter; the calling code stays provider-agnostic.
async fn vlm_query(
    client: &reqwest::Client,
    provider: VlmProvider,
    endpoint: &str,
    model: &str,
    api_key: Option<&str>,
    image_b64: &str,
    prompt: &str,
    timeout_ms: u64,
) -> Result<String> {
    match provider {
        VlmProvider::Ollama => vlm_query_ollama(client, endpoint, model, image_b64, prompt, timeout_ms).await,
        VlmProvider::Anthropic => {
            let key = api_key.ok_or_else(|| anyhow!("anthropic provider needs api_key"))?;
            vlm_query_anthropic(client, endpoint, model, key, image_b64, prompt, timeout_ms).await
        }
        VlmProvider::OpenAI => {
            let key = api_key.ok_or_else(|| anyhow!("openai provider needs api_key"))?;
            vlm_query_openai(client, endpoint, model, key, image_b64, prompt, timeout_ms).await
        }
    }
}

async fn vlm_query_ollama(
    client: &reqwest::Client,
    endpoint: &str,
    model: &str,
    image_b64: &str,
    prompt: &str,
    timeout_ms: u64,
) -> Result<String> {
    let req = OllamaRequest {
        model,
        prompt,
        images: vec![image_b64.to_string()],
        stream: false,
    };
    let resp = client
        .post(format!("{}/api/generate", endpoint.trim_end_matches('/')))
        .json(&req)
        .timeout(Duration::from_millis(timeout_ms))
        .send()
        .await
        .map_err(|e| anyhow!("ollama request failed: {}", e))?;
    if !resp.status().is_success() {
        let status = resp.status();
        let body = resp.text().await.unwrap_or_default();
        return Err(anyhow!("ollama returned {}: {}", status, body));
    }
    let ollama_resp: OllamaResponse = resp
        .json()
        .await
        .map_err(|e| anyhow!("ollama json parse: {}", e))?;
    Ok(ollama_resp.response)
}

/// Anthropic Messages API call with a single image content block.
/// Image must be base64-encoded PNG (the page-screenshot format the
/// solver already produces); we hardcode `image/png` accordingly.
async fn vlm_query_anthropic(
    client: &reqwest::Client,
    endpoint: &str,
    model: &str,
    api_key: &str,
    image_b64: &str,
    prompt: &str,
    timeout_ms: u64,
) -> Result<String> {
    let body = serde_json::json!({
        "model": model,
        "max_tokens": 512,
        "messages": [{
            "role": "user",
            "content": [
                { "type": "image", "source": { "type": "base64", "media_type": "image/png", "data": image_b64 } },
                { "type": "text", "text": prompt },
            ],
        }],
    });
    let resp = client
        .post(format!("{}/v1/messages", endpoint.trim_end_matches('/')))
        .header("x-api-key", api_key)
        .header("anthropic-version", "2023-06-01")
        .header("content-type", "application/json")
        .json(&body)
        .timeout(Duration::from_millis(timeout_ms))
        .send()
        .await
        .map_err(|e| anyhow!("anthropic request failed: {}", e))?;
    if !resp.status().is_success() {
        let status = resp.status();
        let body = resp.text().await.unwrap_or_default();
        return Err(anyhow!("anthropic returned {}: {}", status, body));
    }
    let val: serde_json::Value = resp.json().await.map_err(|e| anyhow!("anthropic json parse: {}", e))?;
    // Response shape: { content: [{ type: "text", text: "..." }, ...] }
    val["content"]
        .as_array()
        .and_then(|arr| arr.iter().find(|c| c["type"] == "text"))
        .and_then(|c| c["text"].as_str())
        .map(|s| s.to_string())
        .ok_or_else(|| anyhow!("anthropic response missing text content: {val}"))
}

async fn vlm_query_openai(
    client: &reqwest::Client,
    endpoint: &str,
    model: &str,
    api_key: &str,
    image_b64: &str,
    prompt: &str,
    timeout_ms: u64,
) -> Result<String> {
    let data_url = format!("data:image/png;base64,{}", image_b64);
    let body = serde_json::json!({
        "model": model,
        "max_tokens": 512,
        "messages": [{
            "role": "user",
            "content": [
                { "type": "text", "text": prompt },
                { "type": "image_url", "image_url": { "url": data_url } },
            ],
        }],
    });
    let resp = client
        .post(format!("{}/v1/chat/completions", endpoint.trim_end_matches('/')))
        .header("authorization", format!("Bearer {}", api_key))
        .header("content-type", "application/json")
        .json(&body)
        .timeout(Duration::from_millis(timeout_ms))
        .send()
        .await
        .map_err(|e| anyhow!("openai request failed: {}", e))?;
    if !resp.status().is_success() {
        let status = resp.status();
        let body = resp.text().await.unwrap_or_default();
        return Err(anyhow!("openai returned {}: {}", status, body));
    }
    let val: serde_json::Value = resp.json().await.map_err(|e| anyhow!("openai json parse: {}", e))?;
    // Response shape: { choices: [{ message: { content: "..." } }] }
    val["choices"][0]["message"]["content"]
        .as_str()
        .map(|s| s.to_string())
        .ok_or_else(|| anyhow!("openai response missing content: {val}"))
}

/// Extract the first JSON object embedded in free-form text (handles markdown fences).
pub(crate) fn extract_json(text: &str) -> Option<&str> {
    if let Some(start) = text.find("```json") {
        let rest = &text[start + 7..];
        if let Some(end) = rest.find("```") {
            return Some(rest[..end].trim());
        }
    }
    if let Some(start) = text.find('{') {
        let rest = &text[start..];
        let mut depth = 0;
        let mut end = None;
        for (i, c) in rest.char_indices() {
            match c {
                '{' => depth += 1,
                '}' => {
                    depth -= 1;
                    if depth == 0 {
                        end = Some(i + 1);
                        break;
                    }
                }
                _ => {}
            }
        }
        if let Some(e) = end {
            return Some(&rest[..e]);
        }
    }
    None
}

// ─── VlmCaptchaSolver ─────────────────────────────────────────────────────────

/// Solves image-based CAPTCHAs via a vision-language model (VLM) such as
/// Ollama-hosted llama3.2-vision.
///
/// Supported sub-types:
///   * Image-grid   (reCAPTCHA / hCaptcha "select all traffic lights")
///   * Text-based   (type the characters shown)
///   * Click-target (click on a specific object / button)
pub struct VlmCaptchaSolver {
    pub(crate) client: reqwest::Client,
    pub(crate) provider: VlmProvider,
    pub(crate) endpoint: String,
    pub(crate) model: String,
    pub(crate) api_key: Option<String>,
    pub(crate) config: SolveConfig,
}

impl Default for VlmCaptchaSolver {
    fn default() -> Self {
        Self::new()
    }
}

impl VlmCaptchaSolver {
    pub fn new() -> Self {
        Self::new_with_env(|name| std::env::var(name).ok())
    }

    /// Construct a solver, using the supplied closure to look up env
    /// variables. The public [`Self::new`] uses [`std::env::var`];
    /// tests use this overload to assert env precedence without
    /// mutating real process env (which is `unsafe` in the 2024
    /// edition and forbidden by the crate's `#![forbid(unsafe_code)]`).
    pub(crate) fn new_with_env<F>(env: F) -> Self
    where
        F: Fn(&str) -> Option<String>,
    {
        let config = SolveConfig::default();
        // Provider auto-detection priority: ANTHROPIC_API_KEY > OPENAI_API_KEY
        // > Ollama. Anthropic wins because Claude's vision is the strongest
        // for the spatial-reasoning image-grid case ("select all crosswalks").
        // Users who want a specific back-end set the matching env var.
        let api_key_anthropic = env(ENV_ANTHROPIC_KEY).filter(|s| !s.is_empty());
        let api_key_openai = env(ENV_OPENAI_KEY).filter(|s| !s.is_empty());
        let (provider, api_key) = if let Some(k) = api_key_anthropic {
            (VlmProvider::Anthropic, Some(k))
        } else if let Some(k) = api_key_openai {
            (VlmProvider::OpenAI, Some(k))
        } else {
            (VlmProvider::Ollama, None)
        };
        // Endpoint + model: explicit env wins, else per-provider default.
        let endpoint = env(ENV_VLM_ENDPOINT)
            .filter(|s| !s.is_empty())
            .unwrap_or_else(|| provider.default_endpoint().to_string());
        let model = env(ENV_VLM_MODEL)
            .filter(|s| !s.is_empty())
            .unwrap_or_else(|| provider.default_model().to_string());
        Self {
            client: reqwest::Client::builder()
                .timeout(Duration::from_millis(config.client_http_timeout_ms))
                .build()
                .unwrap_or_else(|_| reqwest::Client::new()),
            provider,
            endpoint,
            model,
            api_key,
            config,
        }
    }

    /// Force a specific provider (overrides env auto-detection). Useful
    /// for integration tests that need to pin Ollama even when an env
    /// API key is present, or for callers that want to ship a specific
    /// fallback chain.
    pub fn with_provider(mut self, provider: VlmProvider) -> Self {
        self.provider = provider;
        if self.endpoint.is_empty() {
            self.endpoint = provider.default_endpoint().to_string();
        }
        self
    }

    /// Provide an explicit API key (overrides env auto-detection).
    pub fn with_api_key(mut self, key: impl Into<String>) -> Self {
        self.api_key = Some(key.into());
        self
    }

    /// Override the VLM endpoint (default: `http://localhost:11434`).
    pub fn with_endpoint(mut self, url: impl Into<String>) -> Self {
        self.endpoint = url.into();
        self
    }

    /// Override the VLM model name (default: `llama3.2-vision:11b`).
    pub fn with_model(mut self, model: impl Into<String>) -> Self {
        self.model = model.into();
        self
    }

    /// Provide a custom solve configuration.
    pub fn with_config(mut self, config: SolveConfig) -> Self {
        self.config = config;
        self
    }

    pub fn grid_prompt(task: &str) -> String {
        format!(
            r#"You are solving an image-grid CAPTCHA.
Task: {}
Respond with a JSON object exactly like this:
{{"selected": [0,1,2], "confidence": 0.95}}
where `selected` is a list of zero-based tile indices that match the task.
If none match, use an empty list.
Only respond with the JSON object, no extra text."#,
            task
        )
    }

    pub fn text_captcha_prompt() -> String {
        r#"You are solving a text CAPTCHA.
Read the text in the image and respond with a JSON object exactly like this:
{"text": "answer", "confidence": 0.95}
Only respond with the JSON object, no extra text."#
            .to_string()
    }

    pub fn click_target_prompt(target: &str) -> String {
        format!(
            r#"You are solving a click-target CAPTCHA.
Click the object described as: "{}"
Respond with a JSON object exactly like this:
{{"x": 123, "y": 456, "confidence": 0.95}}
where x and y are pixel coordinates in the screenshot.
Only respond with the JSON object, no extra text."#,
            target
        )
    }
}

#[async_trait]
impl CaptchaSolver for VlmCaptchaSolver {
    fn name(&self) -> &'static str {
        "VlmCaptchaSolver"
    }

    fn method(&self) -> SolveMethod {
        SolveMethod::VisionLLM
    }

    fn supports(&self, kind: &crate::captcha_detect::DetectedCaptcha) -> bool {
        use crate::captcha_detect::DetectedCaptcha;
        // VLM screenshot+vision can attempt any visual captcha,
        // including TOML-rule vendors whose providers recommend
        // VisionLLM as a fallback method.
        matches!(
            kind,
            DetectedCaptcha::RecaptchaV2
                | DetectedCaptcha::RecaptchaV3
                | DetectedCaptcha::HCaptcha
                | DetectedCaptcha::ImageCaptcha
                | DetectedCaptcha::Turnstile
                | DetectedCaptcha::CanvasCaptcha
                | DetectedCaptcha::ShadowDomCaptcha
                | DetectedCaptcha::MultiStepCaptcha
                | DetectedCaptcha::SliderCaptcha
                | DetectedCaptcha::Custom(_)
        )
    }

    async fn solve(
        &self,
        page: &Page,
        captcha_info: &crate::captcha_detect::CaptchaInfo,
    ) -> Result<CaptchaSolveResult> {
        let t0 = Instant::now();
        let image_b64 = screenshot_b64(page).await?;

        // Use the detected kind to pick a prompt strategy.
        match captcha_info.kind {
            crate::captcha_detect::DetectedCaptcha::RecaptchaV2
            | crate::captcha_detect::DetectedCaptcha::HCaptcha
            | crate::captcha_detect::DetectedCaptcha::ImageCaptcha => {
                // Image grid challenge. Pull the task description from
                // the page itself (e.g. "Select all icons that are
                // animals") rather than passing a generic instruction —
                // the VLM does much better when the task matches what
                // the user sees.
                let task_text = page
                    .evaluate(
                        r#"(() => {
                            const probes = [
                                '.rc-imageselect-desc-no-canonical', '.rc-imageselect-desc',
                                '.task-text', '.captcha-prompt',
                                '#widget > div', '.captcha-task'
                            ];
                            for (const sel of probes) {
                                const el = document.querySelector(sel);
                                if (el && el.textContent && el.textContent.trim().length > 0) {
                                    return el.textContent.trim();
                                }
                            }
                            return 'Select all matching images';
                        })()"#,
                    )
                    .await
                    .ok()
                    .and_then(|r| r.into_value::<String>().ok())
                    .unwrap_or_else(|| "Select all matching images".to_string());
                let prompt = Self::grid_prompt(&task_text);
                let raw = vlm_query(
                    &self.client,
                    self.provider,
                    &self.endpoint,
                    &self.model,
                    self.api_key.as_deref(),
                    &image_b64,
                    &prompt,
                    self.config.vlm_http_timeout_ms,
                )
                .await?;

                let json_str = extract_json(&raw).unwrap_or(&raw);
                let val: serde_json::Value =
                    serde_json::from_str(json_str).unwrap_or(serde_json::Value::Null);
                let confidence = val["confidence"].as_f64().unwrap_or(0.5) as f32;
                let selected: Vec<usize> = val["selected"]
                    .as_array()
                    .unwrap_or(&vec![])
                    .iter()
                    .filter_map(|v| v.as_u64().map(|n| n as usize))
                    .collect();

                // Click each selected tile via direct DOM dispatch
                // — far more reliable than computing mouse coordinates
                // across iframe boundaries. Probe the grid first to
                // get the actual tile elements + dimensions, then
                // dispatch a click event on each by index.
                //
                // Grid probes per vendor:
                //   - reCAPTCHA v2:  .rc-imageselect-tile (img wrapper)
                //   - hCaptcha:      .task-image (sometimes .image)
                //   - Generic:       td > img inside the grid table
                let click_indices_js = format!(
                    r#"
                    ((indices) => {{
                        const probes = [
                            '.rc-imageselect-tile',
                            '.task-image',
                            '.image-task',
                            '.hcaptcha-checkbox-img',
                            '.rc-imageselect-table td',
                            '.hcaptcha-table td',
                            '.captcha-grid > .tile',
                            // Generic local-grid selectors: icon picker,
                            // color picker, image picker. Order matters
                            // — first non-empty match wins.
                            '#widget .icon-item',
                            '#widget .color-item',
                            '#widget .cell',
                            '#widget .grid > .cell',
                        ];
                        let tiles = null;
                        for (const sel of probes) {{
                            const els = document.querySelectorAll(sel);
                            if (els.length > 0) {{ tiles = Array.from(els); break; }}
                        }}
                        if (!tiles || tiles.length === 0) return {{ ok: false, count: 0 }};
                        let clicked = 0;
                        for (const idx of indices) {{
                            const el = tiles[idx];
                            if (!el) continue;
                            const r = el.getBoundingClientRect();
                            const cx = r.left + r.width / 2;
                            const cy = r.top + r.height / 2;
                            ['mousedown','mouseup','click'].forEach(t => {{
                                el.dispatchEvent(new MouseEvent(t, {{
                                    bubbles: true, button: 0,
                                    clientX: cx, clientY: cy
                                }}));
                            }});
                            clicked++;
                        }}
                        /* Find + click the verify/submit button. */
                        const verifyProbes = [
                            '#recaptcha-verify-button',
                            '.rc-button-default',
                            '.button-submit',
                            '[data-pp="submit"]',
                            'button[type="submit"]',
                            // Generic verify button used by local-mock
                            // grid fixtures (icon_selection,
                            // color_match, image/grid).
                            '#widget #submit',
                            '#widget button',
                        ];
                        for (const sel of verifyProbes) {{
                            const b = document.querySelector(sel);
                            if (b) {{ b.click(); break; }}
                        }}
                        return {{ ok: true, count: clicked, total: tiles.length }};
                    }})({selected})
                    "#,
                    selected = serde_json::to_string(&selected).unwrap_or_else(|_| "[]".into())
                );
                let _ = page.evaluate(click_indices_js).await;

                let success = confidence > 0.5 && !selected.is_empty();
                let cookies = if success {
                    crate::cookies::capture_from_page(page)
                        .await
                        .unwrap_or_default()
                } else {
                    Vec::new()
                };
                Ok(CaptchaSolveResult {
                    solution: serde_json::to_string(&selected).unwrap_or_default(),
                    confidence,
                    method: SolveMethod::VisionLLM,
                    time_ms: t0.elapsed().as_millis() as u64,
                    success,
                    screenshot: None,
                    cookies,
                    verified_outcome: None,
                })
            }
            crate::captcha_detect::DetectedCaptcha::RecaptchaV3
            | crate::captcha_detect::DetectedCaptcha::CanvasCaptcha
            | crate::captcha_detect::DetectedCaptcha::ShadowDomCaptcha => {
                // Text-based challenge — read the rendered text out of
                // a canvas/SVG widget and type it into the response
                // input. CanvasCaptchaDetector covers canvas, SVG,
                // rotated-text, math, icon, and color sub-types; the
                // text path is the right default.
                let prompt = Self::text_captcha_prompt();
                let raw = vlm_query(
                    &self.client,
                    self.provider,
                    &self.endpoint,
                    &self.model,
                    self.api_key.as_deref(),
                    &image_b64,
                    &prompt,
                    self.config.vlm_http_timeout_ms,
                )
                .await?;

                let json_str = extract_json(&raw).unwrap_or(&raw);
                let val: serde_json::Value =
                    serde_json::from_str(json_str).unwrap_or(serde_json::Value::Null);
                let text = val["text"].as_str().unwrap_or("").trim().to_string();
                let confidence = val["confidence"].as_f64().unwrap_or(0.5) as f32;

                // Type the answer into the response field. Try the
                // top-level light DOM first; on miss, walk shadow roots
                // and same-origin iframes to find the input. Mirrors
                // the WaitForTokenSolver's piercing probe.
                let typed_via_pierce = page
                    .evaluate(
                        format!(
                            r#"((answer) => {{
                                function* walkAllRoots(root) {{
                                    const queue = [root];
                                    const seen = new WeakSet();
                                    while (queue.length) {{
                                        const r = queue.shift();
                                        if (seen.has(r)) continue;
                                        seen.add(r);
                                        yield r;
                                        const subtree = r.querySelectorAll ? r.querySelectorAll('*') : [];
                                        for (const el of subtree) {{
                                            if (el.shadowRoot) queue.push(el.shadowRoot);
                                            if (el.tagName === 'IFRAME') {{
                                                let inner = null;
                                                try {{ inner = el.contentDocument; }} catch (_) {{}}
                                                if (inner) queue.push(inner);
                                            }}
                                        }}
                                    }}
                                }}
                                for (const root of walkAllRoots(document)) {{
                                    let inp = null;
                                    try {{ inp = root.querySelector('input[type=text], input:not([type]), textarea, .rc-response-input, input[name="captcha"], input[name="answer"]'); }} catch(_) {{ continue; }}
                                    if (!inp) continue;
                                    inp.focus(); inp.value = answer;
                                    inp.dispatchEvent(new Event('input', {{bubbles: true}}));
                                    inp.dispatchEvent(new Event('change', {{bubbles: true}}));
                                    return true;
                                }}
                                return false;
                            }})({})"#,
                            serde_json::to_string(&text).unwrap_or_else(|_| "\"\"".into())
                        ),
                    )
                    .await
                    .and_then(|r| r.into_value::<bool>().map_err(Into::into))
                    .unwrap_or(false);
                if !typed_via_pierce {
                    // Fallback to find_element click+type if the JS pierce missed.
                    if let Ok(input) = page
                        .find_element("input[type=text], textarea, .rc-response-input")
                        .await
                    {
                        input.click().await.ok();
                        input.type_str(&text).await.ok();
                    }
                }

                let success = confidence > 0.5 && !text.is_empty();
                let cookies = if success {
                    crate::cookies::capture_from_page(page)
                        .await
                        .unwrap_or_default()
                } else {
                    Vec::new()
                };
                Ok(CaptchaSolveResult {
                    solution: text.clone(),
                    confidence,
                    method: SolveMethod::VisionLLM,
                    time_ms: t0.elapsed().as_millis() as u64,
                    success,
                    screenshot: None,
                    cookies,
                    verified_outcome: None,
                })
            }
            _ => {
                // Generic click-target fallback.
                let prompt = Self::click_target_prompt("the CAPTCHA checkbox or submit button");
                let raw = vlm_query(
                    &self.client,
                    self.provider,
                    &self.endpoint,
                    &self.model,
                    self.api_key.as_deref(),
                    &image_b64,
                    &prompt,
                    self.config.vlm_http_timeout_ms,
                )
                .await?;

                let json_str = extract_json(&raw).unwrap_or(&raw);
                let val: serde_json::Value =
                    serde_json::from_str(json_str).unwrap_or(serde_json::Value::Null);
                let x = val["x"].as_f64().unwrap_or(0.0);
                let y = val["y"].as_f64().unwrap_or(0.0);
                let confidence = val["confidence"].as_f64().unwrap_or(0.0) as f32;

                let viewport = page
                    .evaluate("window.innerWidth + ',' + window.innerHeight")
                    .await?
                    .into_value::<String>()
                    .unwrap_or_default();
                let mut parts = viewport.split(',');
                let vw: f64 = parts.next().and_then(|s| s.parse().ok()).unwrap_or(1920.0);
                let vh: f64 = parts.next().and_then(|s| s.parse().ok()).unwrap_or(1080.0);

                if x > 0.0 && y > 0.0 && x < vw && y < vh {
                    crate::behavior::click_realistic(page, x, y).await?;
                }

                let success = confidence > 0.5;
                let cookies = if success {
                    crate::cookies::capture_from_page(page)
                        .await
                        .unwrap_or_default()
                } else {
                    Vec::new()
                };
                Ok(CaptchaSolveResult {
                    solution: format!("click:{},{}:{:.2}", x as i64, y as i64, confidence),
                    confidence,
                    method: SolveMethod::VisionLLM,
                    time_ms: t0.elapsed().as_millis() as u64,
                    success,
                    screenshot: None,
                    cookies,
                    verified_outcome: None,
                })
            }
        }
    }
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn vlm_solver_defaults_when_env_returns_none() {
        let s = VlmCaptchaSolver::new_with_env(|_| None);
        assert_eq!(s.endpoint, "http://localhost:11434");
        assert_eq!(s.model, "llama3.2-vision:11b");
    }

    #[test]
    fn vlm_solver_env_overrides_defaults() {
        let s = VlmCaptchaSolver::new_with_env(|k| match k {
            "CAPTCHAFORGE_VLM_ENDPOINT" => Some("http://env-host:11434".into()),
            "CAPTCHAFORGE_VLM_MODEL" => Some("env-model:latest".into()),
            _ => None,
        });
        assert_eq!(s.endpoint, "http://env-host:11434");
        assert_eq!(s.model, "env-model:latest");
    }

    #[test]
    fn vlm_solver_builder_overrides_env() {
        let s = VlmCaptchaSolver::new_with_env(|k| match k {
            "CAPTCHAFORGE_VLM_ENDPOINT" => Some("http://env-host:11434".into()),
            "CAPTCHAFORGE_VLM_MODEL" => Some("env-model:latest".into()),
            _ => None,
        })
        .with_endpoint("http://builder:11434")
        .with_model("builder-model:7b");
        assert_eq!(s.endpoint, "http://builder:11434");
        assert_eq!(s.model, "builder-model:7b");
    }

    #[test]
    fn vlm_solver_empty_env_falls_through_to_defaults() {
        let s = VlmCaptchaSolver::new_with_env(|_| Some(String::new()));
        assert_eq!(s.endpoint, "http://localhost:11434");
        assert_eq!(s.model, "llama3.2-vision:11b");
    }

    #[test]
    fn vlm_solver_env_constants_match_documented_names() {
        // The CLAUDE.md / README contract is the env-var names. Pin
        // them so a rename is a deliberate breaking change.
        assert_eq!(ENV_VLM_ENDPOINT, "CAPTCHAFORGE_VLM_ENDPOINT");
        assert_eq!(ENV_VLM_MODEL, "CAPTCHAFORGE_VLM_MODEL");
        assert_eq!(ENV_ANTHROPIC_KEY, "ANTHROPIC_API_KEY");
        assert_eq!(ENV_OPENAI_KEY, "OPENAI_API_KEY");
    }

    #[test]
    fn vlm_solver_no_api_key_picks_ollama() {
        let s = VlmCaptchaSolver::new_with_env(|_| None);
        assert_eq!(s.provider, VlmProvider::Ollama);
        assert!(s.api_key.is_none());
        assert_eq!(s.endpoint, "http://localhost:11434");
        assert_eq!(s.model, "llama3.2-vision:11b");
    }

    #[test]
    fn vlm_solver_anthropic_api_key_picks_anthropic() {
        let s = VlmCaptchaSolver::new_with_env(|k| match k {
            "ANTHROPIC_API_KEY" => Some("sk-ant-test".into()),
            _ => None,
        });
        assert_eq!(s.provider, VlmProvider::Anthropic);
        assert_eq!(s.api_key.as_deref(), Some("sk-ant-test"));
        assert_eq!(s.endpoint, "https://api.anthropic.com");
        assert_eq!(s.model, "claude-haiku-4-5");
    }

    #[test]
    fn vlm_solver_openai_api_key_picks_openai_when_no_anthropic() {
        let s = VlmCaptchaSolver::new_with_env(|k| match k {
            "OPENAI_API_KEY" => Some("sk-openai-test".into()),
            _ => None,
        });
        assert_eq!(s.provider, VlmProvider::OpenAI);
        assert_eq!(s.api_key.as_deref(), Some("sk-openai-test"));
        assert_eq!(s.endpoint, "https://api.openai.com");
        assert_eq!(s.model, "gpt-4o-mini");
    }

    #[test]
    fn vlm_solver_anthropic_wins_over_openai_when_both_set() {
        // Anthropic is preferred — strongest visual reasoning of the three.
        // If users want OpenAI specifically, they unset ANTHROPIC_API_KEY.
        let s = VlmCaptchaSolver::new_with_env(|k| match k {
            "ANTHROPIC_API_KEY" => Some("sk-ant-test".into()),
            "OPENAI_API_KEY" => Some("sk-openai-test".into()),
            _ => None,
        });
        assert_eq!(s.provider, VlmProvider::Anthropic);
        assert_eq!(s.api_key.as_deref(), Some("sk-ant-test"));
    }

    #[test]
    fn vlm_solver_with_provider_overrides_env_detection() {
        // Force Ollama even when Anthropic env key is present — useful
        // for integration tests that need a deterministic local back-end.
        let s = VlmCaptchaSolver::new_with_env(|k| match k {
            "ANTHROPIC_API_KEY" => Some("sk-ant-test".into()),
            _ => None,
        })
        .with_provider(VlmProvider::Ollama);
        assert_eq!(s.provider, VlmProvider::Ollama);
    }

    #[test]
    fn vlm_solver_with_api_key_overrides_env_detection() {
        let s = VlmCaptchaSolver::new_with_env(|_| None).with_api_key("override-key");
        assert_eq!(s.api_key.as_deref(), Some("override-key"));
    }

    #[test]
    fn vlm_solver_empty_anthropic_key_falls_through_to_ollama() {
        // Empty string treated same as unset (matches the existing
        // CAPTCHAFORGE_VLM_ENDPOINT empty-fallthrough contract).
        let s = VlmCaptchaSolver::new_with_env(|k| match k {
            "ANTHROPIC_API_KEY" => Some(String::new()),
            _ => None,
        });
        assert_eq!(s.provider, VlmProvider::Ollama);
    }

    #[test]
    fn vlm_provider_default_endpoints_and_models_consistent() {
        // Catch a regression where the per-provider defaults drift
        // out of sync with the documented constants.
        assert_eq!(VlmProvider::Ollama.default_endpoint(), DEFAULT_OLLAMA_BASE);
        assert_eq!(VlmProvider::Ollama.default_model(), DEFAULT_OLLAMA_MODEL);
        assert_eq!(VlmProvider::Anthropic.default_endpoint(), DEFAULT_ANTHROPIC_BASE);
        assert_eq!(VlmProvider::Anthropic.default_model(), DEFAULT_ANTHROPIC_MODEL);
        assert_eq!(VlmProvider::OpenAI.default_endpoint(), DEFAULT_OPENAI_BASE);
        assert_eq!(VlmProvider::OpenAI.default_model(), DEFAULT_OPENAI_MODEL);
    }

    #[test]
    fn vlm_solver_custom_endpoint() {
        let s = VlmCaptchaSolver::new().with_endpoint("http://ollama.internal:11434");
        assert_eq!(s.endpoint, "http://ollama.internal:11434");
    }

    #[test]
    fn vlm_solver_custom_model() {
        let s = VlmCaptchaSolver::new().with_model("llava:13b");
        assert_eq!(s.model, "llava:13b");
    }

    #[test]
    fn vlm_solver_chained_builders() {
        let s = VlmCaptchaSolver::new()
            .with_endpoint("http://gpu-box:11434")
            .with_model("llama3.2-vision:72b");
        assert_eq!(s.endpoint, "http://gpu-box:11434");
        assert_eq!(s.model, "llama3.2-vision:72b");
    }

    #[test]
    fn grid_prompt_includes_task() {
        let prompt = VlmCaptchaSolver::grid_prompt("Select all traffic lights");
        assert!(prompt.contains("Select all traffic lights"));
        assert!(prompt.contains("selected"));
        assert!(prompt.contains("confidence"));
    }

    #[test]
    fn text_captcha_prompt_is_valid() {
        let prompt = VlmCaptchaSolver::text_captcha_prompt();
        assert!(prompt.contains("text"));
        assert!(prompt.contains("confidence"));
    }

    #[test]
    fn click_target_prompt_includes_target() {
        let prompt = VlmCaptchaSolver::click_target_prompt("submit button");
        assert!(prompt.contains("submit button"));
        assert!(prompt.contains('"'));
    }

    #[test]
    fn extract_json_from_plain_text() {
        let s = r#"Some preamble {"key": "value"} trailing"#;
        assert_eq!(extract_json(s), Some(r#"{"key": "value"}"#));
    }

    #[test]
    fn extract_json_from_markdown_fence() {
        let s = "```json\n{\"answer\": 42}\n```";
        assert_eq!(extract_json(s), Some("{\"answer\": 42}"));
    }

    #[test]
    fn extract_json_no_json_returns_none() {
        assert_eq!(extract_json("no json here"), None);
    }
}