src/voice.rs

// SPF Smart Gateway - Voice Pipeline (Block Q + Block CC)
// Copyright 2026 Joseph Stone - All Rights Reserved
//
// Voice communication over mesh with real + fallback implementations.
//
// Mesh stream type: 0x03 (framing::StreamType::VoiceAudio)
//
// Architecture:
//   AudioInput  → encode → VoiceFrame → mesh (0x03) → decode → AudioOutput
//   AudioInput  → STT → text → ChatEngine (Block P)
//   ChatEngine  → TTS → VoiceFrame → AudioOutput
//
// Implementation tiers:
//   Always available: spf-voice crate (espeak-ng FFI in-process + cpal + opus static), VoiceSession, stubs
//   Feature "voice-stt": CandleWhisperSTT (candle-transformers, no C deps)
//   Feature "voice-tts": PiperTTS (ort/ONNX)
//
// Depends on: framing.rs (Block F, StreamType::VoiceAudio)

use serde::{Deserialize, Serialize};

// ============================================================================
// SHARED VOICE SESSION STATE (Block LL)
// ============================================================================

/// Global voice session — shared between MCP handler (mcp.rs) and mesh handler.
/// Mutex::new is const since Rust 1.63 — no lazy_static needed.
/// MCP "speak" action plays TTS through this.
/// Mesh voice receive plays incoming audio through this.
pub static VOICE_SESSION: std::sync::Mutex<Option<VoiceSession>> = std::sync::Mutex::new(None);

// ============================================================================
// VOICE LINE ARCHITECTURE (Block MM)
// ============================================================================

/// Voice mode: Light (zero battery) vs Rich (full pipeline)
#[derive(Debug, Clone, PartialEq, serde::Serialize, serde::Deserialize)]
pub enum VoiceMode {
    /// spf-voice in-process TTS (espeak-ng FFI), zero session, zero battery
    Light,
    /// spf-voice full pipeline + Piper TTS (voice-tts), persistent session
    Rich,
}

impl Default for VoiceMode {
    fn default() -> Self { Self::Light }
}

/// Voice settings for both Agent Line and Peer Line
#[derive(Debug, Clone, serde::Serialize, serde::Deserialize)]
pub struct VoiceSettings {
    /// Agent Line TTS mode: Light (spf-voice/espeak-ng FFI) | Rich (spf-voice + Piper ONNX)
    pub agent_tts_mode: VoiceMode,
    /// Agent Line STT mode: Light (text input) | Rich (whisper)
    pub agent_stt_mode: VoiceMode,
    /// Peer Line audio quality: Light (12kHz) | Rich (24kHz)
    pub peer_quality: VoiceMode,
    /// Auto-answer incoming peer calls
    pub auto_accept_calls: bool,
    /// Auto end-call after this many seconds of silence
    pub idle_timeout_secs: u32,
}

impl Default for VoiceSettings {
    fn default() -> Self {
        Self {
            agent_tts_mode: VoiceMode::Light,
            agent_stt_mode: VoiceMode::Light,
            peer_quality: VoiceMode::Light,
            auto_accept_calls: false,
            idle_timeout_secs: 300, // 5 minutes
        }
    }
}

/// Call direction — who initiated the call
#[derive(Debug, Clone, PartialEq)]
pub enum CallDirection {
    /// We placed the call (outbound ring sent)
    Outgoing,
    /// Peer placed the call (inbound ring received)
    Incoming,
}

/// Call lifecycle phase
#[derive(Debug, Clone, PartialEq)]
pub enum CallPhase {
    /// Ring sent or received — waiting for accept or reject
    Ringing,
    /// Call accepted — audio flowing (or ready to flow when cpal enabled)
    Active,
    /// Call ended — CALL_STATE is cleared immediately after this phase
    Ended,
}

/// Active call state for Peer Line (SPF ↔ SPF)
#[derive(Debug, Clone)]
pub struct CallState {
    /// Full mesh peer key of the other party
    pub peer_key: String,
    /// Display name — short key prefix if peer name unknown
    pub peer_name: String,
    /// Whether we initiated (Outgoing) or received (Incoming)
    pub direction: CallDirection,
    /// Current lifecycle phase
    pub phase: CallPhase,
    /// When the call was initiated — used for duration tracking
    pub started_at: std::time::Instant,
}

impl CallState {
    /// True when audio is flowing (phase == Active)
    pub fn is_active(&self) -> bool {
        matches!(self.phase, CallPhase::Active)
    }
    /// True when waiting for accept/reject (phase == Ringing)
    pub fn is_ringing(&self) -> bool {
        matches!(self.phase, CallPhase::Ringing)
    }
    /// True when Ringing OR Active — blocks new outbound calls
    pub fn is_in_progress(&self) -> bool {
        matches!(self.phase, CallPhase::Ringing | CallPhase::Active)
    }
    /// Elapsed seconds since call initiated
    pub fn elapsed_secs(&self) -> u64 {
        self.started_at.elapsed().as_secs()
    }
}

/// Voice settings — persistent across actions (Block MM)
pub static VOICE_SETTINGS: std::sync::Mutex<VoiceSettings> = std::sync::Mutex::new(VoiceSettings {
    agent_tts_mode: VoiceMode::Light,
    agent_stt_mode: VoiceMode::Light,
    peer_quality: VoiceMode::Light,
    auto_accept_calls: false,
    idle_timeout_secs: 300,
});

/// Active peer call state — None when no call active (Block MM)
pub static CALL_STATE: std::sync::Mutex<Option<CallState>> = std::sync::Mutex::new(None);

// Feature-gated imports — excluded from compilation when features disabled
// WB-3: spf-voice crate removed (Android-only deps). Inline stub satisfies all call sites.
// pub(crate) — accessible from mcp.rs via `use crate::voice::spf_voice`
pub(crate) mod spf_voice {
    /// Status of the voice pipeline hardware — all false in stub mode.
    pub struct Status {
        pub tts_available:      bool,
        pub input_available:    bool,
        pub output_available:   bool,
        pub codec_available:    bool,
        pub pipeline_open:      bool,
    }

    /// Open the persistent audio pipeline. Stub: always succeeds (no-op).
    pub fn open() -> Result<(), String> { Ok(()) }

    /// Close the persistent audio pipeline. Stub: no-op.
    pub fn close() {}

    /// Capture `_ms` milliseconds of audio. Stub: returns empty buffer.
    pub fn listen(_ms: u64) -> Result<Vec<i16>, String> { Ok(vec![]) }

    /// Synthesise `_text` to speech. Stub: not available.
    pub fn speak(_text: &str) -> Result<(), String> {
        Err("spf-voice not available (WB-3 stub)".into())
    }

    /// Encode PCM i16 → Opus bytes. Stub: not available.
    pub fn encode(_pcm: &[i16]) -> Result<Vec<u8>, String> {
        Err("spf-voice not available (WB-3 stub)".into())
    }

    /// Decode Opus bytes → PCM i16. Stub: not available.
    pub fn decode(_data: &[u8]) -> Result<Vec<i16>, String> {
        Err("spf-voice not available (WB-3 stub)".into())
    }

    /// Play PCM i16 via speaker. Stub: not available.
    pub fn play(_pcm: &[i16]) -> Result<(), String> {
        Err("spf-voice not available (WB-3 stub)".into())
    }

    /// Query hardware pipeline status. Stub: all capabilities false.
    pub fn status() -> Status {
        Status {
            tts_available:    false,
            input_available:  false,
            output_available: false,
            codec_available:  false,
            pipeline_open:    false,
        }
    }
}

// ============================================================================
// VOICE FRAME
// ============================================================================

/// A single voice audio frame for mesh transport.
/// Payload of framing::StreamType::VoiceAudio (0x03).
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct VoiceFrame {
    /// Audio codec identifier
    pub codec: String,
    /// Sample rate in Hz
    pub sample_rate: u32,
    /// Number of audio channels (1 = mono, 2 = stereo)
    pub channels: u8,
    /// Frame duration in milliseconds
    pub frame_duration_ms: u16,
    /// Encoded audio data (codec-dependent)
    pub data: Vec<u8>,
    /// Sequence number for ordering
    pub seq: u64,
    /// Sender identity (short pub_key hash)
    pub from: String,
    /// Timestamp (RFC3339)
    pub timestamp: String,
    /// Team channel ID — None = peer-to-peer call, Some(id) = team channel frame
    /// serde(default) ensures old frames without this field still deserialize correctly
    #[serde(default)]
    pub team_id: Option<String>,
}

// ============================================================================
// VOICE CONFIGURATION
// ============================================================================

/// Voice system configuration
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct VoiceConfig {
    /// Audio codec (target: "opus")
    pub codec: String,
    /// Sample rate in Hz (target: 16000 for speech)
    pub sample_rate: u32,
    /// Channels (1 = mono — sufficient for speech)
    pub channels: u8,
    /// Frame duration in ms (20ms is standard for Opus)
    pub frame_duration_ms: u16,
    /// Bitrate in bits/sec (target: 24000 for speech quality)
    pub bitrate: u32,
    /// Enable voice activity detection
    pub vad_enabled: bool,
}

impl Default for VoiceConfig {
    fn default() -> Self {
        Self {
            codec: "opus".to_string(),
            sample_rate: 16000,
            channels: 1,
            frame_duration_ms: 20,
            bitrate: 24000,
            vad_enabled: true,
        }
    }
}

impl VoiceConfig {
    /// Bytes per frame (uncompressed PCM)
    /// sample_rate × channels × 2 (16-bit) × frame_duration / 1000
    pub fn pcm_frame_bytes(&self) -> usize {
        (self.sample_rate as usize * self.channels as usize * 2
            * self.frame_duration_ms as usize) / 1000
    }
}

// ============================================================================
// VOICE TRAITS — Interfaces for future implementation
// ============================================================================

/// Audio input source (microphone, file, stream)
pub trait AudioInput: Send + Sync {
    /// Read one frame of PCM audio. Returns raw PCM bytes.
    fn read_frame(&mut self) -> Result<Vec<u8>, VoiceError>;
    /// Check if input is available
    fn is_available(&self) -> bool;
    /// Get input configuration
    fn config(&self) -> &VoiceConfig;
}

/// Audio output sink (speaker, file, stream)
pub trait AudioOutput: Send + Sync {
    /// Write one frame of PCM audio
    fn write_frame(&mut self, pcm_data: &[u8]) -> Result<(), VoiceError>;
    /// Check if output is available
    fn is_available(&self) -> bool;
    /// Get output configuration
    fn config(&self) -> &VoiceConfig;
}

/// Speech-to-Text engine
pub trait SpeechToText: Send + Sync {
    /// Transcribe audio frames to text
    fn transcribe(&mut self, frames: &[VoiceFrame]) -> Result<String, VoiceError>;
    /// Check if STT engine is available
    fn is_available(&self) -> bool;
    /// Get supported languages
    fn supported_languages(&self) -> Vec<String>;
}

/// Text-to-Speech engine
pub trait TextToSpeech: Send + Sync {
    /// Synthesize text to audio frames
    fn synthesize(&mut self, text: &str) -> Result<Vec<VoiceFrame>, VoiceError>;
    /// Check if TTS engine is available
    fn is_available(&self) -> bool;
    /// Get available voices
    fn available_voices(&self) -> Vec<String>;
}

/// Voice system error
#[derive(Debug, Clone)]
pub enum VoiceError {
    /// Feature not yet implemented
    NotAvailable(String),
    /// Codec error (encode/decode failure)
    CodecError(String),
    /// Hardware/device error
    DeviceError(String),
    /// Configuration error
    ConfigError(String),
}

impl std::fmt::Display for VoiceError {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        match self {
            VoiceError::NotAvailable(msg) => write!(f, "Not available: {}", msg),
            VoiceError::CodecError(msg) => write!(f, "Codec error: {}", msg),
            VoiceError::DeviceError(msg) => write!(f, "Device error: {}", msg),
            VoiceError::ConfigError(msg) => write!(f, "Config error: {}", msg),
        }
    }
}

// ============================================================================
// STUB IMPLEMENTATIONS — All return "not available"
// ============================================================================

/// Stub audio input — always returns NotAvailable
pub struct StubAudioInput {
    config: VoiceConfig,
}

impl StubAudioInput {
    pub fn new() -> Self {
        Self { config: VoiceConfig::default() }
    }
}

impl AudioInput for StubAudioInput {
    fn read_frame(&mut self) -> Result<Vec<u8>, VoiceError> {
        Err(VoiceError::NotAvailable("Audio input not implemented".into()))
    }

    fn is_available(&self) -> bool {
        false
    }

    fn config(&self) -> &VoiceConfig {
        &self.config
    }
}

/// Stub audio output — always returns NotAvailable
pub struct StubAudioOutput {
    config: VoiceConfig,
}

impl StubAudioOutput {
    pub fn new() -> Self {
        Self { config: VoiceConfig::default() }
    }
}

impl AudioOutput for StubAudioOutput {
    fn write_frame(&mut self, _pcm_data: &[u8]) -> Result<(), VoiceError> {
        Err(VoiceError::NotAvailable("Audio output not implemented".into()))
    }

    fn is_available(&self) -> bool {
        false
    }

    fn config(&self) -> &VoiceConfig {
        &self.config
    }
}

/// Stub STT — always returns NotAvailable
pub struct StubSTT;

impl SpeechToText for StubSTT {
    fn transcribe(&mut self, _frames: &[VoiceFrame]) -> Result<String, VoiceError> {
        Err(VoiceError::NotAvailable("Speech-to-text not implemented".into()))
    }

    fn is_available(&self) -> bool {
        false
    }

    fn supported_languages(&self) -> Vec<String> {
        vec![] // No languages available
    }
}

/// Stub TTS — always returns NotAvailable
pub struct StubTTS;

impl TextToSpeech for StubTTS {
    fn synthesize(&mut self, _text: &str) -> Result<Vec<VoiceFrame>, VoiceError> {
        Err(VoiceError::NotAvailable("Text-to-speech not implemented".into()))
    }

    fn is_available(&self) -> bool {
        false
    }

    fn available_voices(&self) -> Vec<String> {
        vec![] // No voices available
    }
}

// EspeakTTS removed — TTS is now handled by spf-voice crate (in-process FFI, no subprocess)

// OpusCodec removed — encode/decode now handled by spf-voice crate (libopus.a static FFI)

// CpalAudioInput removed — microphone capture now handled by spf-voice crate (spf_voice::listen())

// CpalAudioOutput removed — speaker playback now handled by spf-voice crate (spf_voice::play())

// ============================================================================
// VOICE TOKENIZER — GPT2 byte-level BPE decode, zero external crates
//
// Whisper uses GPT2's byte-level BPE tokenizer. Every token string is encoded
// so that each character maps to exactly one byte via the GPT2 Unicode↔byte
// table. To decode: char → byte (via table), collect bytes, UTF-8 decode.
//
// Reads vocab.json: {"<|endoftext|>": 50256, "the": 1, "Ġthe": 262, ...}
// ============================================================================

#[cfg(feature = "voice-stt")]
pub struct VoiceTokenizer {
    id_to_token: std::collections::HashMap<u32, String>,
    token_to_id_map: std::collections::HashMap<String, u32>,
    unicode_to_byte: std::collections::HashMap<char, u8>,
    special_ids: std::collections::HashSet<u32>,
}

#[cfg(feature = "voice-stt")]
impl VoiceTokenizer {
    /// Load from vocab.json: {"token_string": token_id, ...}
    pub fn from_file(vocab_path: &std::path::Path) -> Result<Self, VoiceError> {
        let json = std::fs::read_to_string(vocab_path)
            .map_err(|e| VoiceError::ConfigError(format!("vocab.json read: {}", e)))?;
        let mut token_to_id_map: std::collections::HashMap<String, u32> =
            serde_json::from_str(&json)
                .map_err(|e| VoiceError::ConfigError(format!("vocab.json parse: {}", e)))?;

        // Merge added_tokens.json if present — Whisper special tokens live here
        if let Some(parent) = vocab_path.parent() {
            let added_path = parent.join("added_tokens.json");
            if let Ok(added_json) = std::fs::read_to_string(&added_path) {
                if let Ok(added) = serde_json::from_str::<std::collections::HashMap<String, u32>>(&added_json) {
                    token_to_id_map.extend(added);
                }
            }
        }

        let unicode_to_byte = Self::build_unicode_to_byte();
        let mut special_ids = std::collections::HashSet::new();
        let id_to_token: std::collections::HashMap<u32, String> = token_to_id_map
            .iter()
            .map(|(tok, &id)| {
                if tok.starts_with("<|") && tok.ends_with("|>") {
                    special_ids.insert(id);
                }
                (id, tok.clone())
            })
            .collect();

        Ok(Self { id_to_token, token_to_id_map, unicode_to_byte, special_ids })
    }

    /// Look up token string → ID.
    pub fn token_to_id(&self, token: &str) -> Option<u32> {
        self.token_to_id_map.get(token).copied()
    }

    /// Decode token IDs → UTF-8 text. Skips special tokens when requested.
    pub fn decode(&self, ids: &[u32], skip_special: bool) -> String {
        let bytes: Vec<u8> = ids
            .iter()
            .filter(|&&id| !skip_special || !self.special_ids.contains(&id))
            .filter_map(|id| self.id_to_token.get(id))
            .flat_map(|tok| {
                tok.chars().filter_map(|c| self.unicode_to_byte.get(&c).copied())
            })
            .collect();
        String::from_utf8_lossy(&bytes).trim().to_string()
    }

    /// GPT2 Unicode↔byte mapping (inverse, for decoding).
    /// Printable bytes (33–126, 161–172, 174–255) map to themselves.
    /// The 68 invisible bytes map to codepoints 256–323 (Ā, ā, Ă, …).
    fn build_unicode_to_byte() -> std::collections::HashMap<char, u8> {
        let visible: Vec<u8> = (b'!'..=b'~')
            .chain(b'\xa1'..=b'\xac')
            .chain(b'\xae'..=b'\xff')
            .collect();
        let visible_set: std::collections::HashSet<u8> = visible.iter().copied().collect();

        let mut map = std::collections::HashMap::with_capacity(256);
        // Visible bytes: Unicode codepoint == byte value
        for &b in &visible {
            map.insert(b as char, b);
        }
        // Invisible bytes: codepoints 256, 257, … in order
        let mut cp = 256u32;
        for b in 0u8..=255 {
            if !visible_set.contains(&b) {
                if let Some(c) = char::from_u32(cp) {
                    map.insert(c, b);
                }
                cp += 1;
            }
        }
        map
    }
}

// ============================================================================
// CANDLE WHISPER STT — Speech-to-Text (feature = "voice-stt")
// Pure Rust — no C++ deps. Uses candle-transformers already compiled in binary.
// Model dir: LIVE/MODELS/whisper-tiny/
// Required:  config.json  vocab.json  model.safetensors  melfilters.bytes
// ============================================================================

#[cfg(feature = "voice-stt")]
pub struct CandleWhisperSTT {
    model: candle_transformers::models::whisper::model::Whisper,
    tokenizer: VoiceTokenizer,
    mel_filters: Vec<f32>,
    config: candle_transformers::models::whisper::Config,
    device: candle_core::Device,
    eot_token: u32,
    sot_token: u32,
    transcribe_token: u32,
    no_timestamps_token: u32,
}

// SAFETY: CandleWhisperSTT is owned exclusively by one thread at a time.
// The Whisper model's internal KV cache (RefCell) is never accessed concurrently.
#[cfg(feature = "voice-stt")]
unsafe impl Send for CandleWhisperSTT {}

#[cfg(feature = "voice-stt")]
impl CandleWhisperSTT {
    /// Load from model directory.
    /// Expects: config.json, tokenizer.json, model.safetensors, melfilters.bytes
    pub fn new(model_dir: &std::path::Path) -> Result<Self, VoiceError> {
        // ── Config ──────────────────────────────────────────────────────────
        let config_str = std::fs::read_to_string(model_dir.join("config.json"))
            .map_err(|e| VoiceError::DeviceError(format!("Whisper config.json: {}", e)))?;
        let config: candle_transformers::models::whisper::Config =
            serde_json::from_str(&config_str)
                .map_err(|e| VoiceError::ConfigError(format!("Whisper config parse: {}", e)))?;

        // ── Tokenizer ───────────────────────────────────────────────────────
        let tokenizer = VoiceTokenizer::from_file(&model_dir.join("vocab.json"))?;

        let eot_token = tokenizer
            .token_to_id(candle_transformers::models::whisper::EOT_TOKEN)
            .ok_or_else(|| VoiceError::ConfigError("EOT token missing".into()))?;
        let sot_token = tokenizer
            .token_to_id(candle_transformers::models::whisper::SOT_TOKEN)
            .ok_or_else(|| VoiceError::ConfigError("SOT token missing".into()))?;
        let transcribe_token = tokenizer
            .token_to_id(candle_transformers::models::whisper::TRANSCRIBE_TOKEN)
            .ok_or_else(|| VoiceError::ConfigError("TRANSCRIBE token missing".into()))?;
        let no_timestamps_token = tokenizer
            .token_to_id(candle_transformers::models::whisper::NO_TIMESTAMPS_TOKEN)
            .ok_or_else(|| VoiceError::ConfigError("NO_TIMESTAMPS token missing".into()))?;

        // ── Mel filters ─────────────────────────────────────────────────────
        // melfilters.bytes: little-endian f32, 80 × 201 = 16 080 values = 64 320 bytes
        let mel_bytes = std::fs::read(model_dir.join("melfilters.bytes"))
            .map_err(|e| VoiceError::DeviceError(format!("melfilters.bytes: {}", e)))?;
        let mel_filters: Vec<f32> = mel_bytes
            .chunks_exact(4)
            .map(|b| f32::from_le_bytes(b.try_into().unwrap()))
            .collect();

        // ── Weights ─────────────────────────────────────────────────────────
        let device = candle_core::Device::Cpu;
        let vb = unsafe {
            candle_nn::VarBuilder::from_mmaped_safetensors(
                &[model_dir.join("model.safetensors")],
                candle_transformers::models::whisper::DTYPE,
                &device,
            )
            .map_err(|e| VoiceError::DeviceError(format!("Weights load: {}", e)))?
        };
        let model =
            candle_transformers::models::whisper::model::Whisper::load(&vb, config.clone())
                .map_err(|e| VoiceError::DeviceError(format!("Model init: {}", e)))?;

        eprintln!("[SPF-STT] Candle Whisper loaded from {:?}", model_dir);
        Ok(Self {
            model,
            tokenizer,
            mel_filters,
            config,
            device,
            eot_token,
            sot_token,
            transcribe_token,
            no_timestamps_token,
        })
    }

    /// Convert f32 PCM (16 kHz, mono, [-1.0, 1.0]) to mel spectrogram Tensor.
    pub fn audio_to_mel(&self, samples: &[f32]) -> Result<candle_core::Tensor, VoiceError> {
        let mel = candle_transformers::models::whisper::audio::pcm_to_mel(
            &self.config,
            samples,
            &self.mel_filters,
        );
        let mel_len = mel.len();
        let n_mel = self.config.num_mel_bins;
        candle_core::Tensor::from_vec(mel, (1, n_mel, mel_len / n_mel), &self.device)
            .map_err(|e| VoiceError::DeviceError(format!("Mel tensor: {}", e)))
    }

    /// Greedy decoder loop: mel Tensor → token sequence → text.
    pub fn decode_segment(&mut self, mel: &candle_core::Tensor) -> Result<String, VoiceError> {
        let audio_features = self.model.encoder.forward(mel, true)
            .map_err(|e| VoiceError::DeviceError(format!("Encoder: {}", e)))?;

        // Seed: <|startoftranscript|> <|transcribe|> <|notimestamps|>
        let mut tokens: Vec<u32> =
            vec![self.sot_token, self.transcribe_token, self.no_timestamps_token];
        let max_tokens = self.config.max_target_positions;

        for i in 0..max_tokens {
            let tokens_t = candle_core::Tensor::new(tokens.as_slice(), &self.device)
                .and_then(|t| t.unsqueeze(0))
                .map_err(|e| VoiceError::DeviceError(format!("Token tensor: {}", e)))?;

            let ys = self.model.decoder.forward(&tokens_t, &audio_features, i == 0)
                .map_err(|e| VoiceError::DeviceError(format!("Decoder step {}: {}", i, e)))?;

            let (_, seq_len, _) = ys.dims3()
                .map_err(|e| VoiceError::DeviceError(format!("dims3: {}", e)))?;
            if seq_len == 0 { break; }

            let logits = self.model.decoder
                .final_linear(&ys.i((..1, seq_len - 1..))
                    .map_err(|e| VoiceError::DeviceError(format!("Slice: {}", e)))?)
                .and_then(|t| t.i(0))
                .and_then(|t| t.i(0))
                .map_err(|e| VoiceError::DeviceError(format!("Logits: {}", e)))?;

            let logits_v: Vec<f32> = logits.to_vec1()
                .map_err(|e| VoiceError::DeviceError(format!("to_vec1: {}", e)))?;

            let next_token = logits_v
                .iter()
                .enumerate()
                .max_by(|(_, a), (_, b)| a.total_cmp(b))
                .map(|(idx, _)| idx as u32)
                .unwrap_or(self.eot_token);

            if next_token == self.eot_token || tokens.len() >= max_tokens {
                break;
            }
            tokens.push(next_token);
        }

        let text = self.tokenizer.decode(&tokens, true);
        Ok(text)
    }
}

#[cfg(feature = "voice-stt")]
impl SpeechToText for CandleWhisperSTT {
    fn transcribe(&mut self, frames: &[VoiceFrame]) -> Result<String, VoiceError> {
        if frames.is_empty() {
            return Ok(String::new());
        }
        // Frames carry i16 LE PCM bytes — convert to f32 [-1.0, 1.0] for Whisper
        let samples: Vec<f32> = frames
            .iter()
            .flat_map(|f| f.data.chunks_exact(2))
            .map(|pair| i16::from_le_bytes([pair[0], pair[1]]) as f32 / 32768.0)
            .collect();
        let mel = self.audio_to_mel(&samples)?;
        self.decode_segment(&mel)
    }

    fn is_available(&self) -> bool { true }

    fn supported_languages(&self) -> Vec<String> {
        vec![
            "en".into(), "es".into(), "fr".into(), "de".into(), "it".into(),
            "pt".into(), "nl".into(), "ja".into(), "ko".into(), "zh".into(),
            "ru".into(), "ar".into(), "hi".into(), "pl".into(), "sv".into(),
        ]
    }
}

// ============================================================================
// STREAMING VOICE-TO-TEXT — Continuous live transcription
//
// stream_on  → opens mic, transcribes continuously, no fixed duration
// stream_off → signals thread to exit (or say "end stream" / "stop stream")
// stream_read → drains all transcribed text accumulated since last call
// ============================================================================

/// Transcribed text segments from the streaming pipeline.
/// Drained by stream_read(). Thread-safe across all tool calls.
pub static TRANSCRIPT_QUEUE: std::sync::Mutex<Vec<String>> =
    std::sync::Mutex::new(Vec::new());

/// Controls the streaming thread lifecycle.
/// true = running.  false = thread exits after current chunk.
pub static STREAM_ACTIVE: std::sync::atomic::AtomicBool =
    std::sync::atomic::AtomicBool::new(false);

/// Start continuous voice-to-text via spf-voice (in-process cpal) — independent of VOICE_SESSION.
/// Each ~10 s chunk is captured via spf_voice::listen(), transcribed, pushed to TRANSCRIPT_QUEUE.
/// Recognises "end stream" or "stop stream" in transcript to self-terminate.
#[cfg(feature = "voice-stt")]
pub fn stream_on(model_dir: std::path::PathBuf) -> Result<(), VoiceError> {
    if STREAM_ACTIVE.load(std::sync::atomic::Ordering::SeqCst) {
        return Err(VoiceError::DeviceError(
            "Stream already active — call stream_off first".into(),
        ));
    }
    // Load model on calling thread: fail fast with a clear error message
    let mut stt = CandleWhisperSTT::new(&model_dir)?;

    STREAM_ACTIVE.store(true, std::sync::atomic::Ordering::SeqCst);
    // Clear stale transcripts from any previous stream
    if let Ok(mut q) = TRANSCRIPT_QUEUE.lock() { q.clear(); }

    std::thread::spawn(move || {
        eprintln!("[SPF-STREAM] Streaming started — via spf-voice, 16 kHz mono, continuous");

        // ~10 s chunks: Whisper works best with 10-30 s segments
        const CHUNK_MS: u64 = 10_000;

        loop {
            if !STREAM_ACTIVE.load(std::sync::atomic::Ordering::SeqCst) { break; }

            // Capture audio via spf-voice (in-process cpal, blocks for CHUNK_MS)
            match spf_voice::listen(CHUNK_MS) {
                Ok(pcm_i16) => {
                    if pcm_i16.is_empty() { continue; }

                    // Convert i16 → f32 [-1.0, 1.0] for Whisper mel spectrogram
                    let f32_samples: Vec<f32> = pcm_i16.iter()
                        .map(|&s| s as f32 / 32768.0)
                        .collect();

                    match stt.audio_to_mel(&f32_samples) {
                        Ok(mel) => match stt.decode_segment(&mel) {
                            Ok(text) if !text.is_empty() => {
                                eprintln!("[SPF-STREAM] → {}", text);
                                let lower = text.to_lowercase();
                                let trimmed = lower.trim().trim_matches(|c: char| !c.is_alphanumeric() && c != ' ');

                                // Voice command detection — specific phrases only
                                let is_end = trimmed.contains("end stream") || trimmed.contains("stop stream");
                                let is_enter = trimmed == "enter" || trimmed.ends_with(" enter")
                                    || trimmed == "new line" || trimmed == "newline";
                                let is_reply = trimmed == "reply" || trimmed.ends_with(" reply");
                                let is_clear = trimmed == "clear" || trimmed.starts_with("clear ");
                                let is_read_back = trimmed.contains("read back") || trimmed.contains("read it back");
                                let is_start = trimmed.contains("start stream");
                                let is_command = is_end || is_enter || is_reply || is_clear || is_read_back || is_start;

                                if is_command {
                                    // Commands don't go into transcript — handle action directly
                                    if is_clear {
                                        if let Ok(mut q) = TRANSCRIPT_QUEUE.lock() { q.clear(); }
                                        eprintln!("[SPF-STREAM] CMD: clear — buffer cleared");
                                    } else if is_read_back {
                                        let current = TRANSCRIPT_QUEUE.lock()
                                            .map(|q| q.join(" ")).unwrap_or_default();
                                        if !current.is_empty() {
                                            let _ = spf_voice::speak(&format!("You said: {}", current));
                                        } else {
                                            let _ = spf_voice::speak("Nothing to read back");
                                        }
                                        eprintln!("[SPF-STREAM] CMD: read back");
                                    } else if is_reply {
                                        if let Ok(mut q) = TRANSCRIPT_QUEUE.lock() {
                                            q.push("[REPLY]".to_string());
                                        }
                                        eprintln!("[SPF-STREAM] CMD: reply — marker pushed");
                                    } else if is_enter {
                                        if let Ok(mut q) = TRANSCRIPT_QUEUE.lock() {
                                            q.push("\n".to_string());
                                        }
                                        eprintln!("[SPF-STREAM] CMD: enter — newline pushed");
                                    } else if is_start {
                                        eprintln!("[SPF-STREAM] CMD: start stream — already running");
                                    }
                                    if is_end {
                                        eprintln!("[SPF-STREAM] CMD: end stream — stopping");
                                        STREAM_ACTIVE.store(false, std::sync::atomic::Ordering::SeqCst);
                                        break;
                                    }
                                } else {
                                    // Normal transcript text — push to queue
                                    if let Ok(mut q) = TRANSCRIPT_QUEUE.lock() {
                                        q.push(text);
                                    }
                                }
                            }
                            Ok(_) => {} // Silence or blank — Whisper returned nothing, continue
                            Err(e) => eprintln!("[SPF-STREAM] Transcribe error: {}", e),
                        },
                        Err(e) => eprintln!("[SPF-STREAM] Mel error: {}", e),
                    }
                }
                Err(e) => {
                    eprintln!("[SPF-STREAM] Listen error: {} — stopping stream", e);
                    STREAM_ACTIVE.store(false, std::sync::atomic::Ordering::SeqCst);
                    break;
                }
            }
        }

        STREAM_ACTIVE.store(false, std::sync::atomic::Ordering::SeqCst);
        eprintln!("[SPF-STREAM] Streaming stopped");
    });

    Ok(())
}

/// Signal the streaming thread to stop. Exits after finishing its current chunk.
pub fn stream_off() {
    STREAM_ACTIVE.store(false, std::sync::atomic::Ordering::SeqCst);
    eprintln!("[SPF-STREAM] Stop signal sent");
}

/// Drain and return all transcribed text accumulated since the last call.
/// Returns empty vec when nothing is ready or no stream is active.
pub fn stream_read() -> Vec<String> {
    TRANSCRIPT_QUEUE
        .lock()
        .map(|mut q| std::mem::take(&mut *q))
        .unwrap_or_default()
}

// ============================================================================
// PIPER TTS — Neural Text-to-Speech via ONNX Runtime (feature = "voice-tts")
// ============================================================================

#[cfg(feature = "voice-tts")]
pub struct PiperTTS {
    session: OrtSession,
    config: VoiceConfig,
    model_sample_rate: u32,
}

#[cfg(feature = "voice-tts")]
impl PiperTTS {
    pub fn new(model_path: &str, config_path: &str) -> Result<Self, VoiceError> {
        let session = OrtSession::builder()
            .and_then(|b| b.commit_from_file(model_path))
            .map_err(|e| VoiceError::DeviceError(format!("ONNX model load: {}", e)))?;

        // Read Piper config JSON for sample rate
        let config_data = std::fs::read_to_string(config_path)
            .map_err(|e| VoiceError::ConfigError(format!("Piper config read: {}", e)))?;
        let config_json: serde_json::Value = serde_json::from_str(&config_data)
            .map_err(|e| VoiceError::ConfigError(format!("Piper config parse: {}", e)))?;
        let model_sample_rate = config_json["audio"]["sample_rate"].as_u64().unwrap_or(22050) as u32;

        Ok(Self {
            session,
            config: VoiceConfig::default(),
            model_sample_rate,
        })
    }
}

#[cfg(feature = "voice-tts")]
impl TextToSpeech for PiperTTS {
    fn synthesize(&mut self, text: &str) -> Result<Vec<VoiceFrame>, VoiceError> {
        // Piper expects phoneme IDs as i64 tensor
        // Simple ASCII-to-phoneme mapping (Piper's phonemizer handles the rest)
        let phoneme_ids: Vec<i64> = text.chars()
            .map(|c| c as i64)
            .collect();

        let input_len = phoneme_ids.len();
        let input_tensor = ort::Value::from_array(
            ort::ArrayExtensions::into_dyn(
                ndarray::Array2::from_shape_vec((1, input_len), phoneme_ids)
                    .map_err(|e| VoiceError::CodecError(format!("Tensor shape: {}", e)))?
            )
        ).map_err(|e| VoiceError::CodecError(format!("Input tensor: {}", e)))?;

        let input_lengths = ort::Value::from_array(
            ndarray::Array1::from_vec(vec![input_len as i64]).into_dyn()
        ).map_err(|e| VoiceError::CodecError(format!("Length tensor: {}", e)))?;

        let scales = ort::Value::from_array(
            ndarray::Array1::from_vec(vec![0.667f32, 1.0, 0.8]).into_dyn()
        ).map_err(|e| VoiceError::CodecError(format!("Scales tensor: {}", e)))?;

        let outputs = self.session.run(ort::inputs![input_tensor, input_lengths, scales]
            .map_err(|e| VoiceError::CodecError(format!("Run inputs: {}", e)))?)
            .map_err(|e| VoiceError::CodecError(format!("ONNX inference: {}", e)))?;

        // Output is audio waveform as f32
        let audio_f32: Vec<f32> = outputs[0]
            .try_extract_tensor::<f32>()
            .map_err(|e| VoiceError::CodecError(format!("Extract audio: {}", e)))?
            .view()
            .iter()
            .copied()
            .collect();

        // Convert f32 audio to i16 PCM bytes
        let pcm_bytes: Vec<u8> = audio_f32.iter()
            .flat_map(|&s| {
                let clamped = (s * 32767.0).clamp(-32768.0, 32767.0) as i16;
                clamped.to_le_bytes()
            })
            .collect();

        // Frame the audio
        let frame_bytes = (self.model_sample_rate as usize * self.config.channels as usize * 2
            * self.config.frame_duration_ms as usize) / 1000;
        if frame_bytes == 0 {
            return Err(VoiceError::ConfigError("Frame size zero".into()));
        }

        let timestamp = chrono::Utc::now().to_rfc3339();
        let frames: Vec<VoiceFrame> = pcm_bytes.chunks(frame_bytes)
            .enumerate()
            .map(|(i, chunk)| VoiceFrame {
                codec: "pcm".into(),
                sample_rate: self.model_sample_rate,
                channels: self.config.channels,
                frame_duration_ms: self.config.frame_duration_ms,
                data: chunk.to_vec(),
                seq: i as u64,
                from: String::new(),
                timestamp: timestamp.clone(),
                team_id: None,
            })
            .collect();
        Ok(frames)
    }

    fn is_available(&self) -> bool {
        true
    }

    fn available_voices(&self) -> Vec<String> {
        vec!["piper-default".into()]
    }
}

// ============================================================================
// VOICE SESSION — Pipeline state management (always compiles)
// ============================================================================

/// Manages the active voice pipeline state.
/// Uses trait objects to abstract over stub/real implementations.
/// Construction selects the best available backend for each component.
pub struct VoiceSession {
    pub config: VoiceConfig,
    input: Option<Box<dyn AudioInput>>,
    output: Option<Box<dyn AudioOutput>>,
    stt: Option<Box<dyn SpeechToText>>,
    tts: Option<Box<dyn TextToSpeech>>,
    active: bool,
    frames_sent: u64,
    frames_received: u64,
}

impl VoiceSession {
    pub fn new(config: VoiceConfig) -> Self {
        Self {
            config,
            input: None,
            output: None,
            stt: None,
            tts: None,
            active: false,
            frames_sent: 0,
            frames_received: 0,
        }
    }

    /// Start the voice pipeline.
    /// Audio I/O and TTS delegate to spf-voice crate (in-process FFI).
    /// STT: CandleWhisperSTT if voice-stt feature enabled.
    pub fn start(&mut self) -> Result<(), VoiceError> {
        // Open persistent audio pipeline — cpal streams + espeak-ng + opus codec
        // Streams stay alive until stop() is called. All spf_voice:: functions
        // auto-route through the persistent pipeline once open.
        spf_voice::open()
            .map_err(|e| VoiceError::DeviceError(format!("spf_voice::open failed: {}", e)))?;

        // Audio I/O: placeholder stubs — real I/O goes through spf_voice:: calls
        self.input = Some(Box::new(StubAudioInput::new()));
        self.output = Some(Box::new(StubAudioOutput::new()));

        // STT: Candle Whisper if feature enabled
        #[cfg(feature = "voice-stt")]
        {
            let model_dir = crate::paths::spf_root().join("LIVE/MODELS/whisper-tiny");
            match CandleWhisperSTT::new(&model_dir) {
                Ok(stt) => {
                    eprintln!("[SPF-VOICE] STT: Candle Whisper active");
                    self.stt = Some(Box::new(stt));
                }
                Err(e) => { eprintln!("[SPF-VOICE] Whisper STT unavailable: {}", e); }
            }
        }

        // TTS + audio I/O: handled by spf-voice crate — report status
        let svs = spf_voice::status();
        eprintln!("[SPF-VOICE] spf-voice: tts={} input={} output={} codec={}",
            svs.tts_available, svs.input_available, svs.output_available, svs.codec_available);

        // Piper TTS (neural, optional — voice-tts feature)
        #[cfg(feature = "voice-tts")]
        {
            let model_dir = crate::paths::spf_root().join("LIVE/MODELS/piper");
            let model = model_dir.join("en_US-lessac-medium.onnx");
            let cfg = model_dir.join("en_US-lessac-medium.onnx.json");
            if model.exists() && cfg.exists() {
                match PiperTTS::new(&model.to_string_lossy(), &cfg.to_string_lossy()) {
                    Ok(piper) => {
                        self.tts = Some(Box::new(piper));
                        eprintln!("[SPF-VOICE] TTS: Piper neural TTS active");
                    }
                    Err(e) => { eprintln!("[SPF-VOICE] Piper TTS unavailable: {}", e); }
                }
            }
        }

        self.active = true;
        eprintln!("[SPF-VOICE] Session started");
        Ok(())
    }

    /// Stop all voice pipeline components
    pub fn stop(&mut self) {
        // Close persistent audio pipeline — drops cpal streams + opus codec
        spf_voice::close();

        // Drop trait objects — stubs release cleanly.
        self.input = None;
        self.output = None;
        self.stt = None;
        self.tts = None;
        self.active = false;
        self.frames_sent = 0;
        self.frames_received = 0;
        eprintln!("[SPF-VOICE] Session stopped");
    }

    pub fn is_active(&self) -> bool {
        self.active
    }

    /// Get current voice pipeline status — queries spf-voice crate for real hardware state
    pub fn status(&self) -> VoiceStatus {
        let svs = spf_voice::status();
        VoiceStatus {
            audio_input_available: svs.input_available,
            audio_output_available: svs.output_available,
            stt_available: self.stt.as_ref().map_or(false, |s| s.is_available()),
            // tts: spf-voice espeak-ng FFI takes priority; fall back to Piper trait object if loaded
            tts_available: svs.tts_available
                || self.tts.as_ref().map_or(false, |t| t.is_available()),
            pipeline_open: svs.pipeline_open,
            codec: self.config.codec.clone(),
            sample_rate: self.config.sample_rate,
            channels: self.config.channels,
        }
    }

    /// Capture one audio frame via spf-voice (in-process cpal), encode with opus, return VoiceFrame for mesh
    pub fn capture_frame(&mut self, from_identity: &str) -> Result<VoiceFrame, VoiceError> {
        let pcm_i16 = spf_voice::listen(self.config.frame_duration_ms as u64)
            .map_err(|e| VoiceError::DeviceError(e.to_string()))?;

        let (codec, encoded) = match spf_voice::encode(&pcm_i16) {
            Ok(data) => ("opus".to_string(), data),
            Err(_) => {
                // Fall back to raw PCM bytes if encode fails
                let pcm_bytes: Vec<u8> = pcm_i16.iter()
                    .flat_map(|s| s.to_le_bytes())
                    .collect();
                ("pcm".to_string(), pcm_bytes)
            }
        };

        self.frames_sent += 1;
        Ok(VoiceFrame {
            codec,
            sample_rate: self.config.sample_rate,
            channels: self.config.channels,
            frame_duration_ms: self.config.frame_duration_ms,
            data: encoded,
            seq: self.frames_sent,
            from: from_identity.to_string(),
            timestamp: chrono::Utc::now().to_rfc3339(),
            team_id: None,
        })
    }

    /// Receive a VoiceFrame from mesh, decode with spf-voice opus, play through spf-voice cpal
    pub fn play_frame(&mut self, frame: &VoiceFrame) -> Result<(), VoiceError> {
        let pcm_i16: Vec<i16> = if frame.codec == "opus" {
            spf_voice::decode(&frame.data)
                .map_err(|e| VoiceError::CodecError(e.to_string()))?
        } else {
            // Raw PCM bytes → i16 (little-endian passthrough)
            frame.data.chunks_exact(2)
                .map(|pair| i16::from_le_bytes([pair[0], pair[1]]))
                .collect()
        };

        spf_voice::play(&pcm_i16)
            .map_err(|e| VoiceError::DeviceError(e.to_string()))?;
        self.frames_received += 1;
        Ok(())
    }

    /// Transcribe accumulated audio frames to text (STT)
    pub fn transcribe(&mut self, frames: &[VoiceFrame]) -> Result<String, VoiceError> {
        let stt = self.stt.as_mut()
            .ok_or_else(|| VoiceError::NotAvailable("No STT engine".into()))?;
        stt.transcribe(frames)
    }

    /// Synthesize text and play audio via spf-voice (in-process espeak-ng FFI).
    /// Falls back to Piper neural TTS trait object if loaded (voice-tts feature).
    /// Returns empty vec — audio is played in-process, not transmitted as mesh frames.
    pub fn speak(&mut self, text: &str) -> Result<Vec<VoiceFrame>, VoiceError> {
        // Primary: spf-voice in-process TTS (espeak-ng FFI)
        match spf_voice::speak(text) {
            Ok(()) => return Ok(vec![]),
            Err(e) => {
                eprintln!("[SPF-VOICE] spf-voice speak failed: {} — trying fallback", e);
            }
        }
        // Fallback: Piper neural TTS if loaded (voice-tts feature)
        if let Some(ref mut tts) = self.tts {
            tts.synthesize(text)
        } else {
            Err(VoiceError::NotAvailable("No TTS engine available".into()))
        }
    }

    /// Get pipeline statistics
    pub fn stats(&self) -> serde_json::Value {
        serde_json::json!({
            "active": self.active,
            "frames_sent": self.frames_sent,
            "frames_received": self.frames_received,
            "status": self.status().to_json_value(),
        })
    }
}

// ============================================================================
// VOICE STATUS (for MCP reporting)
// ============================================================================

/// Voice system status snapshot
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct VoiceStatus {
    pub audio_input_available: bool,
    pub audio_output_available: bool,
    pub stt_available: bool,
    pub tts_available: bool,
    pub pipeline_open: bool,
    pub codec: String,
    pub sample_rate: u32,
    pub channels: u8,
}

impl VoiceStatus {
    /// Build status from current stub state — queries spf-voice for real pipeline state
    pub fn from_stubs() -> Self {
        let config = VoiceConfig::default();
        let svs = spf_voice::status();
        Self {
            audio_input_available: svs.input_available,
            audio_output_available: svs.output_available,
            stt_available: false,
            tts_available: svs.tts_available,
            pipeline_open: svs.pipeline_open,
            codec: config.codec,
            sample_rate: config.sample_rate,
            channels: config.channels,
        }
    }

    /// Build status from a VoiceSession
    pub fn from_session(session: &VoiceSession) -> Self {
        session.status()
    }

    pub fn to_json_value(&self) -> serde_json::Value {
        let mode = if self.pipeline_open && (self.audio_input_available || self.audio_output_available) {
            "pipeline_active"
        } else if self.audio_input_available || self.audio_output_available {
            "active"
        } else if self.tts_available {
            "tts_only"
        } else {
            "stubs_only"
        };
        serde_json::json!({
            "status": mode,
            "pipeline_open": self.pipeline_open,
            "audio_input": self.audio_input_available,
            "audio_output": self.audio_output_available,
            "stt": self.stt_available,
            "tts": self.tts_available,
            "config": {
                "codec": self.codec,
                "sample_rate": self.sample_rate,
                "channels": self.channels,
            }
        })
    }
}

// ============================================================================
// MESH TRANSPORT HELPERS
// ============================================================================

/// Serialize a voice frame for mesh transport (StreamType::VoiceAudio = 0x03)
pub fn frame_to_bytes(frame: &VoiceFrame) -> Result<Vec<u8>, String> {
    serde_json::to_vec(frame).map_err(|e| format!("Serialize error: {}", e))
}

/// Deserialize a voice frame from mesh transport
pub fn frame_from_bytes(data: &[u8]) -> Result<VoiceFrame, String> {
    serde_json::from_slice(data).map_err(|e| format!("Deserialize error: {}", e))
}

// ============================================================================
// TEAM VOICE CHANNEL SYSTEM (Block TM-V)
// ============================================================================

/// A member currently active (joined) in a team voice channel.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct ActiveMember {
    /// Full mesh peer key of this member
    pub peer_key: String,
    /// Display name (short key prefix if unknown)
    pub name: String,
    /// RFC3339 timestamp when this member joined
    pub joined_at: String,
}

/// A team voice channel — persistent channel with invited members and active participants.
#[derive(Debug, Clone)]
pub struct TeamChannel {
    /// Unique identifier for this channel (slugified name + short id)
    pub team_id: String,
    /// Human-readable channel name
    pub name: String,
    /// All invited peer keys (may or may not be currently active)
    pub members: Vec<String>,
    /// Currently joined members — present and receiving audio
    pub active_members: Vec<ActiveMember>,
    /// Audio configuration for this channel
    pub config: VoiceConfig,
    /// RFC3339 timestamp when channel was created
    pub created_at: String,
    /// Whether this channel is open (false = archived/closed)
    pub is_open: bool,
}

impl TeamChannel {
    /// Create a new team channel with default audio config.
    pub fn new(team_id: &str, name: &str) -> Self {
        Self {
            team_id: team_id.to_string(),
            name: name.to_string(),
            members: Vec::new(),
            active_members: Vec::new(),
            config: VoiceConfig::default(),
            created_at: chrono::Utc::now().to_rfc3339(),
            is_open: true,
        }
    }

    /// Add a peer key to the invited members list (idempotent).
    pub fn add_member(&mut self, peer_key: &str) {
        if !self.members.contains(&peer_key.to_string()) {
            self.members.push(peer_key.to_string());
        }
    }

    /// Join the channel — add to active_members if not already present.
    pub fn join(&mut self, peer_key: &str, display_name: &str) {
        let already_active = self.active_members.iter().any(|m| m.peer_key == peer_key);
        if !already_active {
            self.active_members.push(ActiveMember {
                peer_key: peer_key.to_string(),
                name: display_name.to_string(),
                joined_at: chrono::Utc::now().to_rfc3339(),
            });
        }
        // Ensure they are also in the members list
        self.add_member(peer_key);
    }

    /// Leave the channel — remove from active_members. Returns true if they were present.
    pub fn leave(&mut self, peer_key: &str) -> bool {
        let before = self.active_members.len();
        self.active_members.retain(|m| m.peer_key != peer_key);
        self.active_members.len() < before
    }

    /// Number of currently active (joined) members.
    pub fn active_count(&self) -> usize {
        self.active_members.len()
    }

    /// Number of invited members (may include inactive).
    pub fn member_count(&self) -> usize {
        self.members.len()
    }

    /// Peer keys of all currently active members.
    pub fn active_peer_keys(&self) -> Vec<String> {
        self.active_members.iter().map(|m| m.peer_key.clone()).collect()
    }

    /// One-line summary for list display.
    pub fn summary_line(&self) -> String {
        format!("[{}] {} — {} invited, {} active, {}",
            self.team_id, self.name,
            self.member_count(), self.active_count(),
            if self.is_open { "open" } else { "closed" })
    }
}

/// Config entry for a single team in voice_teams.json.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct VoiceTeamEntry {
    pub team_id: String,
    pub name: String,
    #[serde(default)]
    pub members: Vec<String>,
}

/// Top-level structure of LIVE/CONFIG/voice_teams.json.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct VoiceTeamsConfig {
    /// team_id of the channel that always exists (created at startup)
    pub default_team: String,
    /// If true, incoming calls are automatically accepted without user action
    pub auto_accept_calls: bool,
    /// Preconfigured team channels loaded at startup
    #[serde(default)]
    pub teams: Vec<VoiceTeamEntry>,
}

impl Default for VoiceTeamsConfig {
    fn default() -> Self {
        Self {
            default_team: "default".to_string(),
            auto_accept_calls: false,
            teams: vec![VoiceTeamEntry {
                team_id: "default".to_string(),
                name: "Default Channel".to_string(),
                members: Vec::new(),
            }],
        }
    }
}

/// Global team channel registry — persists across tool calls within a session.
/// Key: team_id  Value: TeamChannel
pub static VOICE_TEAMS: std::sync::LazyLock<std::sync::Mutex<std::collections::HashMap<String, TeamChannel>>> =
    std::sync::LazyLock::new(|| std::sync::Mutex::new(std::collections::HashMap::new()));

/// Initialise the team channel registry from voice_teams.json.
/// Called once at startup from mcp::run() and mcp::run_worker().
/// Safe to call multiple times — idempotent (existing channels are not overwritten).
pub fn init_voice_teams() {
    let config_path = crate::paths::spf_root()
        .join("LIVE/CONFIG/voice_teams.json");

    let cfg: VoiceTeamsConfig = if config_path.exists() {
        match std::fs::read_to_string(&config_path) {
            Ok(s) => serde_json::from_str(&s).unwrap_or_default(),
            Err(e) => {
                eprintln!("[SPF-VOICE] voice_teams.json read error: {} — using defaults", e);
                VoiceTeamsConfig::default()
            }
        }
    } else {
        eprintln!("[SPF-VOICE] voice_teams.json not found at {:?} — using defaults", config_path);
        VoiceTeamsConfig::default()
    };

    let mut teams = VOICE_TEAMS.lock().unwrap_or_else(|e| e.into_inner());
    for entry in &cfg.teams {
        // Idempotent: only insert if not already present
        teams.entry(entry.team_id.clone()).or_insert_with(|| {
            let mut ch = TeamChannel::new(&entry.team_id, &entry.name);
            for m in &entry.members {
                ch.add_member(m);
            }
            ch
        });
    }
    // Always ensure the default channel exists even if not in config
    teams.entry(cfg.default_team.clone()).or_insert_with(|| {
        TeamChannel::new(&cfg.default_team, "Default Channel")
    });

    eprintln!("[SPF-VOICE] Team channels ready: {} channel(s) loaded", teams.len());
}

// ============================================================================
// MESH STREAM HANDLER
// ============================================================================
//
// SECURITY CONTRACT — enforced unconditionally:
//   - VoiceAudio stream carries AUDIO FRAMES ONLY (VoiceFrame structs)
//   - ALL call/team signaling travels via ToolRpc stream → gate → spf_voice_mode/spf_voice_team
//   - This function NEVER calls dispatch::call(), handle_tool_call(), or any tool executor
//   - Non-audio payloads (acks, status responses) are logged and silently dropped
//   - Unknown payload types are ALWAYS silently dropped — no error propagation
//   - No code path from here can execute tools or bypass the gate
//
// ============================================================================

/// Wrap a VoiceFrame into a framing::Frame for sending over the mesh VoiceAudio stream.
/// This is the ONLY frame type that should be written to a VoiceAudio stream.
pub fn voice_audio_frame(voice_frame: &VoiceFrame) -> Result<crate::framing::Frame, String> {
    let payload = frame_to_bytes(voice_frame)?;
    Ok(crate::framing::Frame::new(crate::framing::StreamType::VoiceAudio, payload))
}

/// Handle an incoming VoiceAudio mesh frame.
///
/// SECURITY: This function carries audio data ONLY.
/// All call signaling (ring/accept/reject/end) and team signaling (join/leave/invite)
/// travels via the ToolRpc mesh stream → gate → spf_voice_mode / spf_voice_team handlers.
/// This function CANNOT and WILL NOT execute tools, call dispatch, or bypass the gate.
///
/// Routing:
///   VoiceFrame with team_id = Some(id) → routed to VOICE_TEAMS[id] session
///   VoiceFrame with team_id = None     → routed to VOICE_SESSION (peer-to-peer)
///   Non-VoiceFrame JSON (acks/status)  → logged, silently dropped, no response
///   Unknown/malformed payload          → silently dropped, no response
///
/// Called from: mesh.rs stream_router() for StreamType::VoiceAudio (0x03)
pub fn handle_mesh_voice(frame: &crate::framing::Frame, peer_key: &str) -> Option<crate::framing::Frame> {
    let peer_short = &peer_key[..8.min(peer_key.len())];

    // Attempt to deserialize payload as a VoiceFrame (audio data)
    match frame_from_bytes(&frame.payload) {
        Ok(voice_frame) => {
            eprintln!("[SPF-VOICE] Audio frame from {} seq={} codec={} {} bytes{}",
                peer_short, voice_frame.seq, voice_frame.codec, voice_frame.data.len(),
                voice_frame.team_id.as_deref().map(|id| format!(" team={}", id)).unwrap_or_default());

            // Route to the correct audio session based on team_id
            if let Some(ref team_id) = voice_frame.team_id {
                // Team channel frame — route to that team's session
                if let Ok(mut teams) = VOICE_TEAMS.lock() {
                    if let Some(channel) = teams.get_mut(team_id) {
                        // Verify the sender is an active member of this team
                        let is_member = channel.active_members.iter()
                            .any(|m| m.peer_key == peer_key);
                        if is_member {
                            // Team channels share the global VOICE_SESSION for audio output
                            // (one speaker output, multiple senders in the same channel)
                            if let Ok(mut lock) = VOICE_SESSION.lock() {
                                if let Some(ref mut vs) = *lock {
                                    if vs.is_active() {
                                        if let Err(e) = vs.play_frame(&voice_frame) {
                                            eprintln!("[SPF-VOICE] Team playback error team={} seq={}: {}",
                                                team_id, voice_frame.seq, e);
                                        }
                                    }
                                }
                            }
                        } else {
                            // Sender not in active members — reject silently
                            eprintln!("[SPF-VOICE] Rejected team audio from non-member {} team={}",
                                peer_short, team_id);
                            return None;
                        }
                    } else {
                        // Unknown team — reject silently
                        eprintln!("[SPF-VOICE] Rejected audio for unknown team={} from {}",
                            team_id, peer_short);
                        return None;
                    }
                }
            } else {
                // Peer-to-peer frame — route to global VOICE_SESSION
                // Only play if there is an active call with this specific peer
                let in_active_call = {
                    let cs = CALL_STATE.lock().unwrap_or_else(|e| e.into_inner());
                    cs.as_ref().map_or(false, |c| c.peer_key == peer_key && c.is_active())
                };
                if in_active_call {
                    if let Ok(mut lock) = VOICE_SESSION.lock() {
                        if let Some(ref mut vs) = *lock {
                            if vs.is_active() {
                                if let Err(e) = vs.play_frame(&voice_frame) {
                                    eprintln!("[SPF-VOICE] Peer playback error seq={}: {}", voice_frame.seq, e);
                                }
                            }
                        }
                    }
                } else {
                    // No active call with this peer — reject silently
                    eprintln!("[SPF-VOICE] Rejected audio from {} — no active call", peer_short);
                    return None;
                }
            }

            // Send audio acknowledgement
            let ack = serde_json::json!({
                "type": "voice_ack",
                "status": "received",
                "seq": voice_frame.seq,
                "codec": voice_frame.codec,
                "bytes_received": voice_frame.data.len(),
                "voice_available": spf_voice::status().output_available,
            });
            Some(crate::framing::Frame::new(
                crate::framing::StreamType::VoiceAudio,
                serde_json::to_vec(&ack).unwrap_or_default(),
            ))
        }

        Err(_parse_err) => {
            // Payload is not a VoiceFrame — check if it is a known response type
            // (acks and status responses are expected when we send audio to peers)
            // SECURITY: We only READ the type field. No action is taken. No tools executed.
            if let Ok(json_val) = serde_json::from_slice::<serde_json::Value>(&frame.payload) {
                match json_val.get("type").and_then(|t| t.as_str()) {
                    Some("voice_ack") => {
                        // Expected: acknowledgement of audio we sent to this peer
                        let seq = json_val.get("seq").and_then(|s| s.as_u64()).unwrap_or(0);
                        eprintln!("[SPF-VOICE] Ack from {} seq={}", peer_short, seq);
                        // No response needed — drop cleanly
                        None
                    }
                    Some("voice_status") => {
                        // Expected: status response (e.g., peer reports no voice feature)
                        let status = json_val.get("status").and_then(|s| s.as_str()).unwrap_or("unknown");
                        eprintln!("[SPF-VOICE] Status from {}: {}", peer_short, status);
                        // No response needed — drop cleanly
                        None
                    }
                    Some(other_type) => {
                        // Unknown type on VoiceAudio stream — silent drop
                        // SECURITY: unknown control types are always rejected here.
                        // All legitimate control (ring/accept/team/join) travels via ToolRpc+gate.
                        eprintln!("[SPF-VOICE] Rejected unknown type '{}' on VoiceAudio stream from {} — must use ToolRpc for signaling",
                            other_type, peer_short);
                        None
                    }
                    None => {
                        // JSON but no type field — silent drop
                        eprintln!("[SPF-VOICE] Untyped JSON on VoiceAudio stream from {} — dropped",
                            peer_short);
                        None
                    }
                }
            } else {
                // Not JSON, not VoiceFrame — binary garbage or corrupted frame — silent drop
                eprintln!("[SPF-VOICE] Non-JSON non-VoiceFrame payload from {} ({} bytes) — dropped",
                    peer_short, frame.payload.len());
                None
            }
        }
    }
}

// ============================================================================
// TESTS
// ============================================================================

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

    #[test]
    fn test_voice_config_defaults() {
        let config = VoiceConfig::default();
        assert_eq!(config.codec, "opus");
        assert_eq!(config.sample_rate, 16000);
        assert_eq!(config.channels, 1);
        assert_eq!(config.frame_duration_ms, 20);
        assert!(config.vad_enabled);
    }

    #[test]
    fn test_pcm_frame_bytes() {
        let config = VoiceConfig::default();
        // 16000 Hz × 1 ch × 2 bytes × 20ms / 1000 = 640 bytes
        assert_eq!(config.pcm_frame_bytes(), 640);
    }

    #[test]
    fn test_stub_audio_input() {
        let mut input = StubAudioInput::new();
        assert!(!input.is_available());
        assert!(input.read_frame().is_err());
    }

    #[test]
    fn test_stub_audio_output() {
        let mut output = StubAudioOutput::new();
        assert!(!output.is_available());
        assert!(output.write_frame(&[0u8; 640]).is_err());
    }

    #[test]
    fn test_stub_stt() {
        let mut stt = StubSTT;
        assert!(!stt.is_available());
        assert!(stt.transcribe(&[]).is_err());
        assert!(stt.supported_languages().is_empty());
    }

    #[test]
    fn test_stub_tts() {
        let mut tts = StubTTS;
        assert!(!tts.is_available());
        assert!(tts.synthesize("hello").is_err());
        assert!(tts.available_voices().is_empty());
    }

    #[test]
    fn test_voice_status_stubs() {
        let status = VoiceStatus::from_stubs();
        assert!(!status.audio_input_available);
        assert!(!status.stt_available);
        assert_eq!(status.codec, "opus");
        let json = status.to_json_value();
        assert_eq!(json["status"], "stubs_only");
    }

    #[test]
    fn test_voice_error_display() {
        let err = VoiceError::NotAvailable("test".into());
        assert_eq!(format!("{}", err), "Not available: test");
        let err = VoiceError::CodecError("bad".into());
        assert_eq!(format!("{}", err), "Codec error: bad");
    }

    #[test]
    fn test_voice_frame_serialize_roundtrip() {
        let frame = VoiceFrame {
            codec: "opus".into(),
            sample_rate: 16000,
            channels: 1,
            frame_duration_ms: 20,
            data: vec![0xAA, 0xBB, 0xCC],
            seq: 42,
            from: "531d83fa".into(),
            timestamp: "2026-02-28T12:00:00Z".into(),
            team_id: None,
        };
        let bytes = frame_to_bytes(&frame).unwrap();
        let loaded = frame_from_bytes(&bytes).unwrap();
        assert_eq!(loaded.codec, "opus");
        assert_eq!(loaded.seq, 42);
        assert_eq!(loaded.data, vec![0xAA, 0xBB, 0xCC]);
        assert_eq!(loaded.from, "531d83fa");
    }

    // ================================================================
    // spf-voice integration tests (replaces EspeakTTS subprocess tests)
    // ================================================================

    #[test]
    fn test_spf_voice_status_no_panic() {
        // spf_voice::status() must not panic regardless of hardware state
        let svs = spf_voice::status();
        // Fields are bool — just verify they exist and are readable
        let _ = svs.tts_available;
        let _ = svs.input_available;
        let _ = svs.output_available;
        let _ = svs.codec_available;
    }

    #[test]
    fn test_voice_session_new() {
        let config = VoiceConfig::default();
        let session = VoiceSession::new(config);
        assert!(!session.is_active());
        assert_eq!(session.frames_sent, 0);
        assert_eq!(session.frames_received, 0);
        assert!(session.input.is_none());
        assert!(session.output.is_none());
        assert!(session.stt.is_none());
        assert!(session.tts.is_none());
    }

    #[test]
    fn test_voice_session_start_stop() {
        let config = VoiceConfig::default();
        let mut session = VoiceSession::new(config);
        // Without voice features, start() uses stubs — always succeeds
        let result = session.start();
        assert!(result.is_ok());
        assert!(session.is_active());
        // Input/output set to stubs (not available but present)
        assert!(session.input.is_some());
        assert!(session.output.is_some());

        session.stop();
        assert!(!session.is_active());
        assert!(session.input.is_none());
        assert!(session.output.is_none());
        assert!(session.stt.is_none());
        assert!(session.tts.is_none());
        assert_eq!(session.frames_sent, 0);
        assert_eq!(session.frames_received, 0);
    }

    #[test]
    fn test_voice_session_status_after_start() {
        let config = VoiceConfig::default();
        let mut session = VoiceSession::new(config);
        session.start().unwrap();
        let status = session.status();
        // Audio I/O availability depends on spf-voice hardware state — just verify fields exist
        let _ = status.audio_input_available;
        let _ = status.audio_output_available;
        let _ = status.stt_available;
        assert_eq!(status.codec, "opus");
        assert_eq!(status.sample_rate, 16000);
        assert_eq!(status.channels, 1);
        session.stop();
    }

    #[test]
    fn test_voice_session_stats_json() {
        let config = VoiceConfig::default();
        let session = VoiceSession::new(config);
        let stats = session.stats();
        assert_eq!(stats["active"], false);
        assert_eq!(stats["frames_sent"], 0);
        assert_eq!(stats["frames_received"], 0);
        assert!(stats["status"].is_object());
    }

    #[test]
    fn test_voice_session_capture_without_input() {
        let config = VoiceConfig::default();
        let mut session = VoiceSession::new(config);
        // No input initialized (not started) → NotAvailable error
        let result = session.capture_frame("test_identity");
        assert!(result.is_err());
        match result {
            Err(VoiceError::NotAvailable(msg)) => assert!(msg.contains("input")),
            other => panic!("Expected NotAvailable, got: {:?}", other),
        }
    }

    #[test]
    fn test_voice_session_capture_with_stubs() {
        let config = VoiceConfig::default();
        let mut session = VoiceSession::new(config);
        session.start().unwrap();
        // Stub input → read_frame returns NotAvailable
        let result = session.capture_frame("test_identity");
        assert!(result.is_err());
        session.stop();
    }

    #[test]
    fn test_voice_session_play_without_output() {
        let config = VoiceConfig::default();
        let mut session = VoiceSession::new(config);
        let frame = VoiceFrame {
            codec: "pcm".into(),
            sample_rate: 16000,
            channels: 1,
            frame_duration_ms: 20,
            data: vec![0u8; 640],
            seq: 1,
            from: "peer".into(),
            timestamp: "2026-03-01T00:00:00Z".into(),
            team_id: None,
        };
        // No output (not started) → NotAvailable
        let result = session.play_frame(&frame);
        assert!(result.is_err());
    }

    #[test]
    fn test_voice_session_speak_without_tts() {
        let config = VoiceConfig::default();
        let mut session = VoiceSession::new(config);
        // No TTS initialized → NotAvailable
        let result = session.speak("hello world");
        assert!(result.is_err());
        match result {
            Err(VoiceError::NotAvailable(msg)) => assert!(msg.contains("TTS")),
            other => panic!("Expected NotAvailable, got: {:?}", other),
        }
    }

    #[test]
    fn test_voice_session_transcribe_without_stt() {
        let config = VoiceConfig::default();
        let mut session = VoiceSession::new(config);
        let result = session.transcribe(&[]);
        assert!(result.is_err());
        match result {
            Err(VoiceError::NotAvailable(msg)) => assert!(msg.contains("STT")),
            other => panic!("Expected NotAvailable, got: {:?}", other),
        }
    }

    #[test]
    fn test_voice_status_from_session() {
        let config = VoiceConfig::default();
        let session = VoiceSession::new(config);
        let status = VoiceStatus::from_session(&session);
        assert!(!status.audio_input_available);
        assert!(!status.audio_output_available);
        assert!(!status.stt_available);
        assert!(!status.tts_available);
        assert_eq!(status.codec, "opus");
    }

    #[test]
    fn test_voice_status_mode_active() {
        let status = VoiceStatus {
            audio_input_available: true,
            audio_output_available: true,
            stt_available: false,
            tts_available: false,
            pipeline_open: false,
            codec: "opus".into(),
            sample_rate: 16000,
            channels: 1,
        };
        let json = status.to_json_value();
        assert_eq!(json["status"], "active");
        assert_eq!(json["audio_input"], true);
        assert_eq!(json["audio_output"], true);
    }

    #[test]
    fn test_voice_status_mode_tts_only() {
        let status = VoiceStatus {
            audio_input_available: false,
            audio_output_available: false,
            stt_available: false,
            tts_available: true,
            pipeline_open: false,
            codec: "opus".into(),
            sample_rate: 16000,
            channels: 1,
        };
        let json = status.to_json_value();
        assert_eq!(json["status"], "tts_only");
        assert_eq!(json["tts"], true);
    }

    #[test]
    fn test_voice_status_mode_stubs_only() {
        let status = VoiceStatus {
            audio_input_available: false,
            audio_output_available: false,
            stt_available: false,
            tts_available: false,
            pipeline_open: false,
            codec: "opus".into(),
            sample_rate: 16000,
            channels: 1,
        };
        let json = status.to_json_value();
        assert_eq!(json["status"], "stubs_only");
    }

    #[test]
    fn test_voice_status_input_only_counts_as_active() {
        // Even one audio direction = "active" mode
        let status = VoiceStatus {
            audio_input_available: true,
            audio_output_available: false,
            stt_available: false,
            tts_available: false,
            pipeline_open: false,
            codec: "opus".into(),
            sample_rate: 16000,
            channels: 1,
        };
        let json = status.to_json_value();
        assert_eq!(json["status"], "active");
    }

    #[test]
    fn test_handle_mesh_voice_valid_frame() {
        let frame = VoiceFrame {
            codec: "pcm".into(),
            sample_rate: 16000,
            channels: 1,
            frame_duration_ms: 20,
            data: vec![0x00; 640],
            seq: 7,
            from: "test_peer".into(),
            timestamp: "2026-03-01T00:00:00Z".into(),
            team_id: None,
        };
        let serialized = frame_to_bytes(&frame).unwrap();
        let mesh_frame = crate::framing::Frame::new(
            crate::framing::StreamType::VoiceAudio,
            serialized,
        );
        let response = handle_mesh_voice(&mesh_frame, "abcdef1234567890");
        assert!(response.is_some());
        let resp = response.unwrap();
        let payload: serde_json::Value = serde_json::from_slice(&resp.payload).unwrap();
        assert_eq!(payload["type"], "voice_ack");
        assert_eq!(payload["status"], "received");
        assert_eq!(payload["seq"], 7);
        assert_eq!(payload["codec"], "pcm");
        assert_eq!(payload["bytes_received"], 640);
    }

    #[test]
    fn test_handle_mesh_voice_invalid_payload() {
        let mesh_frame = crate::framing::Frame::new(
            crate::framing::StreamType::VoiceAudio,
            vec![0xFF, 0xFE, 0xFD], // Not valid JSON
        );
        // Non-JSON payload is silently dropped — no response (security: no info disclosure)
        let response = handle_mesh_voice(&mesh_frame, "deadbeef12345678");
        assert!(response.is_none());
    }

    #[test]
    fn test_handle_mesh_voice_empty_payload() {
        let mesh_frame = crate::framing::Frame::new(
            crate::framing::StreamType::VoiceAudio,
            vec![],
        );
        // Empty payload is silently dropped — returns None
        let response = handle_mesh_voice(&mesh_frame, "1234abcd");
        assert!(response.is_none());
    }

    #[test]
    fn test_handle_mesh_voice_short_peer_key() {
        // Peer key shorter than 8 chars — should not panic
        let frame = VoiceFrame {
            codec: "pcm".into(),
            sample_rate: 16000,
            channels: 1,
            frame_duration_ms: 20,
            data: vec![0x01],
            seq: 0,
            from: "short".into(),
            timestamp: "2026-03-01T00:00:00Z".into(),
            team_id: None,
        };
        let serialized = frame_to_bytes(&frame).unwrap();
        let mesh_frame = crate::framing::Frame::new(
            crate::framing::StreamType::VoiceAudio,
            serialized,
        );
        // 4-char key — tests the min() guard in peer_short
        let response = handle_mesh_voice(&mesh_frame, "abcd");
        assert!(response.is_some());
    }

    #[test]
    fn test_voice_config_custom_stereo() {
        let config = VoiceConfig {
            codec: "pcm".into(),
            sample_rate: 48000,
            channels: 2,
            frame_duration_ms: 10,
            bitrate: 64000,
            vad_enabled: false,
        };
        // 48000 × 2 × 2 × 10 / 1000 = 1920 bytes per frame
        assert_eq!(config.pcm_frame_bytes(), 1920);
        assert!(!config.vad_enabled);
    }

    #[test]
    fn test_voice_config_8khz_mono() {
        let config = VoiceConfig {
            codec: "opus".into(),
            sample_rate: 8000,
            channels: 1,
            frame_duration_ms: 20,
            bitrate: 12000,
            vad_enabled: true,
        };
        // 8000 × 1 × 2 × 20 / 1000 = 320 bytes
        assert_eq!(config.pcm_frame_bytes(), 320);
    }

    #[test]
    fn test_voice_error_all_variants() {
        let errors = vec![
            VoiceError::NotAvailable("na".into()),
            VoiceError::CodecError("ce".into()),
            VoiceError::DeviceError("de".into()),
            VoiceError::ConfigError("cfg".into()),
        ];
        let expected = vec![
            "Not available: na",
            "Codec error: ce",
            "Device error: de",
            "Config error: cfg",
        ];
        for (err, exp) in errors.iter().zip(expected.iter()) {
            assert_eq!(format!("{}", err), *exp);
        }
    }

    #[test]
    fn test_frame_from_bytes_error() {
        let result = frame_from_bytes(b"not json");
        assert!(result.is_err());
        let err_msg = result.unwrap_err();
        assert!(err_msg.contains("Deserialize"));
    }

    #[test]
    fn test_voice_frame_clone() {
        let frame = VoiceFrame {
            codec: "opus".into(),
            sample_rate: 16000,
            channels: 1,
            frame_duration_ms: 20,
            data: vec![1, 2, 3],
            seq: 99,
            from: "sender".into(),
            timestamp: "2026-03-01T00:00:00Z".into(),
            team_id: None,
        };
        let cloned = frame.clone();
        assert_eq!(cloned.seq, 99);
        assert_eq!(cloned.data, vec![1, 2, 3]);
        assert_eq!(cloned.from, "sender");
    }
}