captchaforge 0.2.28

use super::*;

use crate::backends::{SttBackend, SttKind};

// ─── AudioCaptchaSolver ───────────────────────────────────────────────────────

/// Solves reCAPTCHA v2 (and generic) audio challenges.
///
/// Backend selection (auto-detected on construction):
///   - **Local `whisper` CLI** on PATH — preferred. Audio bytes are written
///     to a tempfile and `whisper <tmp> --model tiny --output_format txt`
///     is spawned; the resulting `.txt` is parsed.
///   - **OpenAI Whisper API** if `OPENAI_API_KEY` is set.
///   - **Local HTTP server** at `http://localhost:9000/asr` (the legacy
///     [openai-whisper-asr-webservice] convention).
///
/// The first available backend wins. Override with [`Self::with_stt_backend`]
/// or [`Self::with_stt_endpoint`] for explicit configuration.
///
/// [openai-whisper-asr-webservice]: https://github.com/ahmetoner/whisper-asr-webservice
pub struct AudioCaptchaSolver {
    client: reqwest::Client,
    /// Auto-detected STT backend. `None` means no local backend was found at
    /// construction; the solver falls back to [`Self::stt_endpoint`].
    pub(crate) stt_backend: Option<SttBackend>,
    /// HTTP STT endpoint — used when [`Self::stt_backend`] is `None` or its
    /// kind is `LocalServer`. Default `http://localhost:9000/asr`.
    pub(crate) stt_endpoint: String,
    pub(crate) config: SolveConfig,
}

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

impl AudioCaptchaSolver {
    pub fn new() -> Self {
        let config = SolveConfig::default();
        // Best-effort sync probe for whisper-cli + OPENAI_API_KEY. Fast
        // (PATH walk + env lookup), no network.
        let stt_backend = Self::probe_stt_sync();
        Self {
            client: reqwest::Client::builder()
                .timeout(Duration::from_millis(config.client_http_timeout_ms))
                .build()
                .unwrap_or_else(|_| reqwest::Client::new()),
            stt_backend,
            stt_endpoint: "http://localhost:9000/asr".to_string(),
            config,
        }
    }

    /// Synchronous subset of [`crate::backends::probe`] focused on STT —
    /// safe to call from `new()` (no async, no network). Picks whisper-cli
    /// over OpenAI API when both are present (local is faster + free).
    fn probe_stt_sync() -> Option<SttBackend> {
        if let Some(path) =
            crate::backends::which("whisper").or_else(|| crate::backends::which("whisper-cli"))
        {
            return Some(SttBackend {
                kind: SttKind::WhisperCli,
                endpoint: path,
                model: Some(crate::backends::DEFAULT_WHISPER_MODEL.to_string()),
            });
        }
        if std::env::var("OPENAI_API_KEY").is_ok() {
            return Some(SttBackend {
                kind: SttKind::OpenAIApi,
                endpoint: "https://api.openai.com/v1/audio/transcriptions".to_string(),
                model: Some("whisper-1".to_string()),
            });
        }
        None
    }

    pub fn with_stt_endpoint(mut self, url: impl Into<String>) -> Self {
        self.stt_endpoint = url.into();
        self
    }

    /// Explicitly set the STT backend. Use this when probing should be
    /// skipped (tests, custom pipelines, vendored whisper builds).
    pub fn with_stt_backend(mut self, backend: SttBackend) -> Self {
        self.stt_backend = Some(backend);
        self
    }

    pub fn with_config(mut self, config: SolveConfig) -> Self {
        self.config = config;
        self
    }

    /// Transcribe `audio_bytes` using the configured backend. Public so
    /// callers can reuse the audio→text path without going through a full
    /// page solve (e.g. CLI `captchaforge transcribe foo.mp3`).
    ///
    /// Pre-pipeline: when the input is a WAV container we run it through
    /// [`crate::audio_dsp::preprocess_for_stt`] (denoise → bandpass →
    /// peak-normalise) before handing it to the STT backend. Vendors
    /// deliberately distort accessibility audio; pre-processing typically
    /// lifts whisper accuracy from ~60% to ~90% on the bench fixtures.
    /// Non-WAV inputs (MP3/OGG) pass through unchanged because the
    /// pre-pipeline doesn't ship a container demuxer yet (tracked under
    /// `audio_dsp` module-level docs).
    pub async fn transcribe(&self, audio_bytes: bytes::Bytes) -> Result<String> {
        let prepared = if audio_bytes.starts_with(b"RIFF") {
            match crate::audio_dsp::preprocess_for_stt(&audio_bytes) {
                Ok(pcm) => bytes::Bytes::from(crate::audio_dsp::encode_wav_pcm16(&pcm)),
                Err(e) => {
                    // Pre-processing is best-effort — if the WAV is
                    // malformed we'd rather hand the raw bytes to STT
                    // (whose own decoder may be more forgiving) than
                    // hard-fail the whole solve.
                    tracing::debug!(
                        error = %e,
                        "audio_dsp pre-process failed; sending raw audio to STT"
                    );
                    audio_bytes
                }
            }
        } else {
            audio_bytes
        };
        match &self.stt_backend {
            Some(b) if b.kind == SttKind::WhisperCli => {
                Self::transcribe_via_whisper_cli(b, prepared).await
            }
            Some(b) if b.kind == SttKind::OpenAIApi => {
                self.transcribe_via_openai(b, prepared).await
            }
            _ => self.transcribe_via_http(prepared).await,
        }
    }

    async fn transcribe_via_whisper_cli(
        backend: &SttBackend,
        audio_bytes: bytes::Bytes,
    ) -> Result<String> {
        let model = backend.model.as_deref().unwrap_or("tiny");
        let bin = backend.endpoint.clone();
        // tempfile crate gives a path that auto-deletes on drop; whisper
        // is happy with `.wav` / `.mp3` based on the extension. We sniff
        // the first 4 bytes for a basic format choice.
        let ext = if audio_bytes.starts_with(b"RIFF") {
            "wav"
        } else if audio_bytes.starts_with(b"OggS") {
            "ogg"
        } else {
            "mp3"
        };
        let mut tmp = tempfile::Builder::new()
            .prefix("captchaforge-stt-")
            .suffix(&format!(".{ext}"))
            .tempfile()?;
        use std::io::Write;
        tmp.write_all(&audio_bytes)?;
        tmp.flush()?;
        let audio_path = tmp.path().to_path_buf();
        let outdir = tempfile::tempdir()?;
        let outdir_path = outdir.path().to_path_buf();

        // whisper writes <stem>.txt into --output_dir; spawn synchronously
        // inside spawn_blocking so we don't stall the async runtime.
        let model_owned = model.to_string();
        let result = tokio::task::spawn_blocking(move || -> Result<String> {
            let status = std::process::Command::new(&bin)
                .arg(&audio_path)
                .args(["--model", &model_owned])
                .args(["--output_format", "txt"])
                .arg("--output_dir")
                .arg(&outdir_path)
                // English-only is faster + more accurate for digit
                // CAPTCHAs; whisper auto-detects otherwise (slower).
                .args(["--language", "en"])
                .arg("--fp16=False")
                .stdout(std::process::Stdio::null())
                .stderr(std::process::Stdio::null())
                .status()
                .map_err(|e| anyhow::anyhow!("spawning whisper: {e}"))?;
            if !status.success() {
                anyhow::bail!("whisper exited {status}");
            }
            let stem = audio_path
                .file_stem()
                .and_then(|s| s.to_str())
                .ok_or_else(|| anyhow::anyhow!("audio path has no stem"))?;
            let txt_path = outdir_path.join(format!("{stem}.txt"));
            let body = std::fs::read_to_string(&txt_path)
                .map_err(|e| anyhow::anyhow!("reading {}: {e}", txt_path.display()))?;
            Ok(body.trim().to_string())
        })
        .await
        .map_err(|e| anyhow::anyhow!("whisper task join: {e}"))??;
        Ok(result)
    }

    async fn transcribe_via_openai(
        &self,
        backend: &SttBackend,
        audio_bytes: bytes::Bytes,
    ) -> Result<String> {
        let key = std::env::var("OPENAI_API_KEY")
            .map_err(|_| anyhow::anyhow!("OPENAI_API_KEY not set"))?;
        let model = backend.model.as_deref().unwrap_or("whisper-1");
        let part = reqwest::multipart::Part::bytes(audio_bytes.to_vec())
            .file_name("audio.mp3")
            .mime_str("audio/mpeg")?;
        let form = reqwest::multipart::Form::new()
            .text("model", model.to_string())
            .text("response_format", "text")
            .part("file", part);
        let resp = self
            .client
            .post(&backend.endpoint)
            .bearer_auth(key)
            .multipart(form)
            .send()
            .await?;
        if !resp.status().is_success() {
            anyhow::bail!("openai whisper returned {}", resp.status());
        }
        Ok(resp.text().await?.trim().to_string())
    }

    async fn transcribe_via_http(&self, audio_bytes: bytes::Bytes) -> Result<String> {
        let resp = self
            .client
            .post(&self.stt_endpoint)
            .header("Content-Type", "audio/mpeg")
            .body(audio_bytes)
            .send()
            .await?;
        if !resp.status().is_success() {
            anyhow::bail!("local STT returned {}", resp.status());
        }
        Ok(resp.text().await?.trim().to_string())
    }
}

/// Normalize a whisper transcript to a CAPTCHA-friendly answer.
///
/// Captcha answer fields commonly accept either spelled digits ("four two
/// nine") or compact digits ("429"). Whisper most often returns the spelled
/// form for slowly-spoken digit audio. This helper:
///   - lowercases + trims
///   - strips punctuation
///   - converts each spelled-digit token to its digit form
///   - collapses whitespace
///
/// Pure helper — pub so callers can normalize for diffing in tests.
///
/// # Examples
///
/// ```
/// use captchaforge::solver::normalize_transcript;
/// assert_eq!(normalize_transcript("Four two nine."), "429");
/// assert_eq!(normalize_transcript(" SEVEN one zero "), "710");
/// // Mixed numerals + words pass through.
/// assert_eq!(normalize_transcript("4 two 9"), "429");
/// // Non-digit transcripts keep their words (lowercased, single-spaced).
/// assert_eq!(normalize_transcript("Hello World"), "hello world");
/// ```
pub fn normalize_transcript(s: &str) -> String {
    let lowered: String = s
        .to_lowercase()
        .chars()
        .map(|c| if c.is_alphanumeric() { c } else { ' ' })
        .collect();
    let tokens: Vec<&str> = lowered.split_whitespace().collect();
    let mut out = Vec::with_capacity(tokens.len());
    let mut all_digits = true;
    for tok in &tokens {
        let digit = match *tok {
            "zero" | "oh" => Some("0"),
            "one" => Some("1"),
            "two" | "to" | "too" => Some("2"),
            "three" => Some("3"),
            "four" | "for" => Some("4"),
            "five" => Some("5"),
            "six" => Some("6"),
            "seven" => Some("7"),
            "eight" | "ate" => Some("8"),
            "nine" => Some("9"),
            t if t.chars().all(|c| c.is_ascii_digit()) => Some(*tok),
            _ => None,
        };
        if let Some(d) = digit {
            out.push(d.to_string());
        } else {
            all_digits = false;
            out.push((*tok).to_string());
        }
    }
    if all_digits {
        out.join("")
    } else {
        out.join(" ")
    }
}

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

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

    fn supports(&self, kind: &crate::captcha_detect::DetectedCaptcha) -> bool {
        use crate::captcha_detect::DetectedCaptcha;
        matches!(
            kind,
            DetectedCaptcha::RecaptchaV2 | DetectedCaptcha::AudioCaptcha
        )
    }

    async fn solve(
        &self,
        page: &Page,
        _captcha_info: &crate::captcha_detect::CaptchaInfo,
    ) -> Result<CaptchaSolveResult> {
        let t0 = Instant::now();

        // Click the audio challenge button — only present on reCAPTCHA v2.
        // Generic audio CAPTCHAs surface the `<audio>` element directly,
        // so a missing button is non-fatal.
        if let Ok(audio_btn) = page
            .find_element("#recaptcha-audio-button, .rc-button-audio")
            .await
        {
            audio_btn.click().await?;
            tokio::time::sleep(Duration::from_millis(self.config.audio_button_delay_ms)).await;
        }

        // Extract the audio source URL. Resolves relative paths against
        // the document's origin so a fixture serving `/audio.wav` works.
        let audio_src = page
            .evaluate(
                r#"(function(){const a=document.querySelector('audio[src],audio source[src]');
                    if(!a)return '';
                    const raw=a.getAttribute('src');
                    if(!raw)return '';
                    try{return new URL(raw, document.baseURI).toString();}catch(e){return raw;}})()"#,
            )
            .await?
            .into_value::<String>()?;

        if audio_src.is_empty() {
            return Ok(CaptchaSolveResult::failure(
                SolveMethod::AudioBypass,
                t0.elapsed().as_millis() as u64,
            ));
        }

        // Download the audio via the BROWSER fetch — this preserves
        // session cookies, follows redirects the same way the page
        // would, and inherits chromium's certificate policy (so
        // self-signed dev origins work when chromium has
        // --ignore-certificate-errors set). reqwest with the default
        // policy would reject self-signed certs.
        let audio_b64 = page
            .evaluate(format!(
                r#"(async (url) => {{
                    try {{
                        const r = await fetch(url, {{ credentials: 'include', cache: 'no-store' }});
                        if (!r.ok) return '';
                        const buf = await r.arrayBuffer();
                        const bytes = new Uint8Array(buf);
                        let bin = '';
                        for (let i = 0; i < bytes.length; i++) bin += String.fromCharCode(bytes[i]);
                        return btoa(bin);
                    }} catch (e) {{ return ''; }}
                }})({})"#,
                serde_json::to_string(&audio_src).unwrap_or_else(|_| "\"\"".into())
            ))
            .await?
            .into_value::<String>()
            .unwrap_or_default();
        if audio_b64.is_empty() {
            return Ok(CaptchaSolveResult::failure(
                SolveMethod::AudioBypass,
                t0.elapsed().as_millis() as u64,
            ));
        }
        use base64::Engine as _;
        let audio_bytes: bytes::Bytes = match base64::engine::general_purpose::STANDARD
            .decode(audio_b64.as_bytes())
        {
            Ok(v) => v.into(),
            Err(_) => {
                return Ok(CaptchaSolveResult::failure(
                    SolveMethod::AudioBypass,
                    t0.elapsed().as_millis() as u64,
                ));
            }
        };

        // Transcribe via the auto-detected backend (whisper-cli > OpenAI
        // API > local HTTP server). Failures bubble up as solver failure
        // so the chain falls through to the next solver instead of
        // hanging the page.
        let transcript = match self.transcribe(audio_bytes).await {
            Ok(t) if !t.is_empty() => normalize_transcript(&t),
            _ => {
                return Ok(CaptchaSolveResult::failure(
                    SolveMethod::AudioBypass,
                    t0.elapsed().as_millis() as u64,
                ));
            }
        };

        if transcript.is_empty() {
            return Ok(CaptchaSolveResult::failure(
                SolveMethod::AudioBypass,
                t0.elapsed().as_millis() as u64,
            ));
        }

        // Type the transcript into the answer field. Use a JS pierce so
        // we don't depend on element handles (which expire when the
        // page reflows after our typed input fires the fixture's own
        // input-listeners). Selectors cover reCAPTCHA v2 (`#audio-
        // response`, `.rc-response-input`) and generic audio captchas
        // (`input[name="captcha"]`, `input[name="answer"]`,
        // `input[type="text"]` as last resort).
        let typed = page
            .evaluate(format!(
                r#"((answer) => {{
                    const inp = document.querySelector(
                        '#audio-response, .rc-response-input, ' +
                        'input[name="captcha"], input[name="answer"], ' +
                        'input[type="text"]'
                    );
                    if (!inp) return false;
                    inp.focus();
                    /* Char-by-char so the page's input listener fires per
                       keystroke — many fixtures auto-submit when the
                       value matches mid-type. */
                    inp.value = '';
                    for (let i = 0; i < answer.length; i++) {{
                        inp.value += answer[i];
                        inp.dispatchEvent(new Event('input', {{bubbles: true}}));
                    }}
                    inp.dispatchEvent(new Event('change', {{bubbles: true}}));
                    return true;
                }})({})"#,
                serde_json::to_string(&transcript).unwrap_or_else(|_| "\"\"".into())
            ))
            .await?
            .into_value::<bool>()
            .unwrap_or(false);
        if !typed {
            return Ok(CaptchaSolveResult::failure(
                SolveMethod::AudioBypass,
                t0.elapsed().as_millis() as u64,
            ));
        }

        tokio::time::sleep(Duration::from_millis(self.config.audio_submit_delay_ms)).await;

        // Click the verify button if present (reCAPTCHA v2 needs it;
        // generic captchas auto-validate on input).
        if let Ok(verify) = page
            .find_element("#recaptcha-verify-button, .rc-button-default, button[type='submit']")
            .await
        {
            verify.click().await.ok();
        }

        // Success signal: either the reCAPTCHA token is populated, OR
        // the page has transitioned (title/cookie/widget removed). The
        // chain's outcome oracle does the page-transition check at a
        // higher layer, so here we just report success when we typed
        // something and the answer field accepted it; chain will
        // downgrade to failure if the oracle disagrees.
        let has_token = page
            .evaluate(
                r#"(() => {
                    const t = document.querySelector('#g-recaptcha-response, [name="g-recaptcha-response"]');
                    if (t && t.value && t.value.length > 0) return true;
                    /* Title flip is the most common generic success marker. */
                    return /solved|verified|success/i.test(document.title || '');
                })()"#,
            )
            .await?
            .into_value::<bool>()
            .unwrap_or(false);

        let cookies = if has_token {
            crate::cookies::capture_from_page(page)
                .await
                .unwrap_or_default()
        } else {
            Vec::new()
        };

        Ok(CaptchaSolveResult {
            solution: transcript.clone(),
            confidence: if has_token { 0.9 } else { 0.5 },
            method: SolveMethod::AudioBypass,
            time_ms: t0.elapsed().as_millis() as u64,
            success: has_token,
            screenshot: None,
            cookies,
            verified_outcome: None,
        })
    }
}

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

    #[test]
    fn audio_solver_custom_endpoint() {
        let s = AudioCaptchaSolver::new().with_stt_endpoint("http://custom:9999/stt");
        assert_eq!(s.stt_endpoint, "http://custom:9999/stt");
    }
}