config

type AssemblyAIProviderConfig struct {
	Enabled          bool   `toml:"enabled"`
	APIKeyEnv        string `toml:"api_key_env"`
	STTModels        string `toml:"stt_models"`
	StreamingModel   string `toml:"streaming_model"`
	StreamingBaseURL string `toml:"streaming_base_url"`
}

type AssistConfig struct {
	EnabledTools []string `toml:"enabled_tools"`

	// IncludeWindowContext controls whether the Device-Target captures the
	// foreground application name + window title and feeds them to the
	// Assist LLM as context. Window titles can be sensitive (document
	// names, chat partners, URLs), so this is an explicit opt-out. The
	// Server-Target ignores this flag — there the integrating client
	// decides whether to send the `app`/`window_title` request fields.
	// Defaults to true (see config/defaults.go).
	IncludeWindowContext bool `toml:"include_window_context"`

	// HomeAssistant configures the optional Home Assistant Conversation
	// API bridge used by the Voice-Companion skill catalog. When
	// URL+TokenEnv are both set, the HA skill is wired automatically;
	// otherwise the skill stays disabled and Voice-Companion commands
	// fall through to the LLM.
	HomeAssistant AssistHomeAssistantConfig `toml:"home_assistant"`
}

type AssistHomeAssistantConfig struct {
	// URL is the base URL of the Home Assistant instance, e.g.
	// "https://ha.kombify.io:8123". No trailing slash required —
	// the HA skill trims it.
	URL string `toml:"url"`

	// TokenEnv names the env var (resolved via internal/secrets) that
	// holds a Long-Lived Access Token created via HA → Profile →
	// Long-Lived Access Tokens. The value itself is NEVER stored in
	// the TOML file.
	TokenEnv string `toml:"token_env"`

	// Language overrides the language sent to HA's Conversation API.
	// When empty, the user's locale is used.
	Language string `toml:"language"`
}

type AudioConfig struct {
	Backend        string `toml:"backend"`
	DeviceID       string `toml:"device_id"`
	OutputDeviceID string `toml:"output_device_id"`
	SampleRate     int    `toml:"sample_rate"`
	Channels       int    `toml:"channels"`
	FrameSizeMs    int    `toml:"frame_size_ms"`
	LatencyHint    string `toml:"latency_hint"`
}

type AuditConfig struct {
	Enabled         bool   `toml:"enabled"`
	RetentionDays   int    `toml:"retention_days"`
	EventLogEnabled bool   `toml:"event_log_enabled"` // wired in P2.1 (cpv.3.1) — Windows Event Log mirror
	OTLPEndpoint    string `toml:"otlp_endpoint"`     // wired in P2.2 (cpv.3.2) — OTLP exporter
	OTLPCertFile    string `toml:"otlp_cert_file"`
	OTLPKeyFile     string `toml:"otlp_key_file"`
	OTLPCAFile      string `toml:"otlp_ca_file"`
}

type Config struct {
	General        GeneralConfig        `toml:"general"`
	Audio          AudioConfig          `toml:"audio"`
	UI             UIConfig             `toml:"ui"`
	Vocabulary     VocabularyConfig     `toml:"vocabulary"`
	Customization  CustomizationConfig  `toml:"customization"`
	Speech         SpeechDefaultsConfig `toml:"speech"`
	Assist         AssistConfig         `toml:"assist"`
	Shortcuts      ShortcutsConfig      `toml:"shortcuts"`
	ModelSelection ModelSelectionConfig `toml:"model_selection"`

	// ServerConnection points the device/local-target at a remote SpeechKit
	// Server-Target. Only consulted when at least one mode in ModelSelection
	// has mode_source = "server". Disabled by default; the desktop app runs
	// fully self-contained until a user opts a mode into server-side
	// execution (typically via onboarding or settings).
	ServerConnection ServerConnectionConfig `toml:"server_connection"`

	Local           LocalConfig           `toml:"local"`
	LocalLLM        LocalLLMConfig        `toml:"local_llm"`
	VPS             VPSConfig             `toml:"vps"`
	HuggingFace     HuggingFaceConfig     `toml:"huggingface"`
	Routing         RoutingConfig         `toml:"routing"`
	Update          UpdateConfig          `toml:"update"`
	Logging         LoggingConfig         `toml:"logging"`
	Audit           AuditConfig           `toml:"audit"`
	Telemetry       TelemetryConfig       `toml:"telemetry"`
	Feedback        FeedbackConfig        `toml:"feedback"` // legacy compat; prefer Store
	Store           StoreConfig           `toml:"store"`
	Providers       ProvidersConfig       `toml:"providers"`
	ProviderOptions ProviderOptionsConfig `toml:"provider_options"`
	TTS             TTSConfig             `toml:"tts"`
	VoiceAgent      VoiceAgentConfig      `toml:"voice_agent"`

	// Server configures the standalone Linux server binary (cmd/speechkit-server).
	// All fields are optional; the desktop app (cmd/speechkit) ignores them entirely.
	Server    ServerConfig     `toml:"server"`
	Personas  []PersonaConfig  `toml:"personas"`
	Roles     []RoleConfig     `toml:"roles"`
	Sequences []SequenceConfig `toml:"sequences"`

	// HandsFree is the user-facing activation + optional voice-output layer
	// across the three strict modes. New config writes should prefer this
	// block; Wakeword remains the low-level detector compatibility block.
	HandsFree HandsFreeConfig `toml:"hands_free"`

	// Wakeword configures the always-on "Hey Quby" activation-word listener.
	// Read by cmd/speechkit (Device-Target) and any library embedder; the
	// Server-Target ignores this block in v1.
	Wakeword WakewordConfig `toml:"wakeword"`
}

type CustomizationConfig struct {
	ActiveTemplateIDs []string `toml:"active_template_ids"`
}

type DeepgramProviderConfig struct {
	Enabled                  bool   `toml:"enabled"`
	APIKeyEnv                string `toml:"api_key_env"`
	STTModel                 string `toml:"stt_model"`
	STTLanguage              string `toml:"stt_language"`
	STTSmartFormat           bool   `toml:"stt_smart_format"`
	STTDictation             bool   `toml:"stt_dictation"`
	STTFillerWords           bool   `toml:"stt_filler_words"`
	STTNumerals              bool   `toml:"stt_numerals"`
	STTDetectLanguage        bool   `toml:"stt_detect_language"`
	STTUseVocabularyKeyterms bool   `toml:"stt_use_vocabulary_keyterms"`
	STTKeyterms              string `toml:"stt_keyterms"`
	STTEndpointingMs         int    `toml:"stt_endpointing_ms"`
	DiarizationModel         string `toml:"diarization_model"`
}

type DeepgramThinkSettings struct {
	Provider    string
	Model       string
	EndpointURL string
	APIKey      string
}

type FeedbackConfig struct {
	SaveAudio          bool   `toml:"save_audio"`
	AudioRetentionDays int    `toml:"audio_retention_days"`
	DBPath             string `toml:"db_path"`
	MaxAudioStorageMB  int    `toml:"max_audio_storage_mb"`
}

type GeneralConfig struct {
	Language                 string `toml:"language"`
	Hotkey                   string `toml:"hotkey"` // Deprecated: legacy single-hotkey field kept for config file compat. Use DictateHotkey.
	DictateHotkey            string `toml:"dictate_hotkey"`
	AssistHotkey             string `toml:"assist_hotkey"`
	VoiceAgentHotkey         string `toml:"voice_agent_hotkey"`
	DictateHotkeyBehavior    string `toml:"dictate_hotkey_behavior"`
	AssistHotkeyBehavior     string `toml:"assist_hotkey_behavior"`
	VoiceAgentHotkeyBehavior string `toml:"voice_agent_hotkey_behavior"`
	DictateEnabled           bool   `toml:"dictate_enabled"`
	AssistEnabled            bool   `toml:"assist_enabled"`
	VoiceAgentEnabled        bool   `toml:"voice_agent_enabled"`
	AutoStartOnLaunch        bool   `toml:"auto_start_on_launch"`
	EagerWarmup              bool   `toml:"eager_warmup"`
	AgentHotkey              string `toml:"agent_hotkey"`
	AgentMode                string `toml:"agent_mode"`  // "assist" or "voice_agent" — determines what agent_hotkey triggers
	ActiveMode               string `toml:"active_mode"` // legacy compat
	HotkeyMode               string `toml:"hotkey_mode"` // legacy compat for single behavior setting
	AutoStopSilenceMs        int    `toml:"auto_stop_silence_ms"`
	FastModeSilenceMs        int    `toml:"fast_mode_silence_ms"`        // silence threshold for Quick Capture auto-stop
	DictateSilenceTimeoutSec int    `toml:"dictate_silence_timeout_sec"` // total silence in seconds before dictate auto-stops; 0 disables
	ModelDownloadDir         string `toml:"model_download_dir"`          // Default directory for downloaded local model files
}

type GoogleProviderConfig struct {
	Enabled                   bool   `toml:"enabled"`
	APIKeyEnv                 string `toml:"api_key_env"`
	STTAPIKeyEnv              string `toml:"stt_api_key_env"`
	STTCredentialsJSONEnv     string `toml:"stt_credentials_json_env"`
	ApplicationCredentialsEnv string `toml:"application_credentials_env"`
	STTModel                  string `toml:"stt_model"`
	UtilityModel              string `toml:"utility_model"`
	AssistModel               string `toml:"assist_model"`
	AgentModel                string `toml:"agent_model"`
	// Region is the Google Cloud region the customer's API key / project is
	// pinned to. Default "europe-west3" (Frankfurt) reflects the EU-enterprise
	// compliance posture. US customers should explicitly set "us-central1".
	//
	// IMPORTANT: this field feeds the byok.key_updated audit event and the
	// settings UI. It does NOT redirect API traffic — the Gemini Live endpoint
	// is a single global WebSocket (generativelanguage.googleapis.com). Actual
	// data residency is controlled at the Google Cloud project level. Both the
	// project region AND this field must match for the audit event to be
	// accurate. See docs/compliance/byok-gemini-region-pinning.md.
	Region string `toml:"region"`
}

type GroqProviderConfig struct {
	Enabled      bool   `toml:"enabled"`
	APIKeyEnv    string `toml:"api_key_env"`
	STTModel     string `toml:"stt_model"`
	UtilityModel string `toml:"utility_model"`
	AssistModel  string `toml:"assist_model"`
	AgentModel   string `toml:"agent_model"`
}

type HandsFreeConfig struct {
	// Enabled gates hands-free activation. Default false (opt-in).
	Enabled bool `toml:"enabled"`

	// ActivationPhraseID picks one of wakeword.DefaultCatalog's curated
	// phrases. The low-level detector mirrors this value to Wakeword.PhraseID.
	ActivationPhraseID string `toml:"activation_phrase_id"`

	// TargetMode is one of "assist", "voice_agent", or
	// "dictation_ui_assisted". Dictation hands-free still requires a visible
	// text target or explicit commit surface.
	TargetMode string `toml:"target_mode"`

	// AutoEndSilenceCutoffSec ends wake-triggered sessions after this many
	// seconds of silence. Zero falls back to the framework default.
	AutoEndSilenceCutoffSec int `toml:"auto_end_silence_cutoff_sec"`

	// VoiceOutputEnabled allows Assist/Voice-Agent hands-free experiences to
	// speak. Dictation UI-assisted targets should keep this false.
	VoiceOutputEnabled bool `toml:"voice_output_enabled"`
}

type HuggingFaceConfig struct {
	Enabled      bool   `toml:"enabled"`
	Model        string `toml:"model"`
	UtilityModel string `toml:"utility_model"`
	AssistModel  string `toml:"assist_model"`
	AgentModel   string `toml:"agent_model"`
	TokenEnv     string `toml:"token_env"`
}

type InstallMode string

type InstallState struct {
	Mode      InstallMode `toml:"mode"`
	SetupDone bool        `toml:"setup_done"`
	DeviceID  string      `toml:"device_id"`
}

type LocalConfig struct {
	Enabled   bool   `toml:"enabled"`
	Model     string `toml:"model"`
	ModelPath string `toml:"model_path"`
	Port      int    `toml:"port"`
	GPU       string `toml:"gpu"`
}

type LocalLLMConfig struct {
	Enabled      bool   `toml:"enabled"`
	BaseURL      string `toml:"base_url"`
	Model        string `toml:"model"`
	ModelPath    string `toml:"model_path"`
	Port         int    `toml:"port"`
	GPU          string `toml:"gpu"`
	UtilityModel string `toml:"utility_model"`
	AssistModel  string `toml:"assist_model"`
	AgentModel   string `toml:"agent_model"`
}

type LoggingConfig struct {
	MaxFileSizeMB int    `toml:"max_file_size_mb"`
	MaxFiles      int    `toml:"max_files"`
	Level         string `toml:"level"` // "debug" | "info" | "warn" | "error" | "off"
}

type ModeModelSelection struct {
	PrimaryProfileID  string `toml:"primary_profile_id"`
	FallbackProfileID string `toml:"fallback_profile_id"`

	// ModeSource selects whether this mode runs locally (Framework kernel
	// in-process, default) or against a remote SpeechKit Server-Target
	// configured under [server_connection]. Empty string is treated as
	// ModeSourceLocal so existing configs keep behaving as before.
	ModeSource string `toml:"mode_source"`
}

type ModelSelectionConfig struct {
	Dictate    ModeModelSelection `toml:"dictate"`
	Assist     ModeModelSelection `toml:"assist"`
	VoiceAgent ModeModelSelection `toml:"voice_agent"`
	// TTS pins the Voice-Output provider profile + optional fallback that
	// Assist and Voice-Agent use when speaking back to the user. Same shape
	// as the three product-mode selections so the catalog API stays
	// symmetric. Added in v0.37 alongside the hands-free Voice-Companion
	// flow so Thalia + Companion-Live deployments can pick a stable voice
	// (e.g. Google Studio-O DE) without editing the lower-level
	// [tts.providers.*] blocks.
	TTS ModeModelSelection `toml:"tts"`
}

type OllamaProviderConfig struct {
	Enabled      bool   `toml:"enabled"`
	BaseURL      string `toml:"base_url"`
	STTModel     string `toml:"stt_model"`
	UtilityModel string `toml:"utility_model"`
	AssistModel  string `toml:"assist_model"`
	AgentModel   string `toml:"agent_model"`
}

type OpenAIProviderConfig struct {
	Enabled       bool   `toml:"enabled"`
	APIKeyEnv     string `toml:"api_key_env"`
	STTModel      string `toml:"stt_model"`
	UtilityModel  string `toml:"utility_model"`
	AssistModel   string `toml:"assist_model"`
	AgentModel    string `toml:"agent_model"`
	TTSModel      string `toml:"tts_model"`
	TTSVoice      string `toml:"tts_voice"`
	RealtimeModel string `toml:"realtime_model"`
}

type OpenRouterProviderConfig struct {
	Enabled      bool   `toml:"enabled"`
	APIKeyEnv    string `toml:"api_key_env"`
	STTModel     string `toml:"stt_model"`
	UtilityModel string `toml:"utility_model"`
	AssistModel  string `toml:"assist_model"`
	AgentModel   string `toml:"agent_model"`
}

type OverlayFreePosition struct {
	X int `toml:"x"`
	Y int `toml:"y"`
}

type PersonaConfig struct {
	ID              string            `toml:"id"`
	DisplayName     string            `toml:"display_name"`
	Description     string            `toml:"description"`
	Voice           string            `toml:"voice"`
	Locale          string            `toml:"locale"`
	DefaultRole     string            `toml:"default_role"`
	DefaultSequence string            `toml:"default_sequence"`
	Tags            []string          `toml:"tags"`
	Metadata        map[string]string `toml:"metadata"`
}

type PolicyValues struct {
	UpdateEnabled             *bool
	UpdateManifestURL         string
	TelemetryUpdateCheck      *bool
	ProvidersEnforceLocalOnly *bool
	VoiceAgentAllowCloud      *bool
	AuditRetentionDays        *int
	AuditEventLogEnabled      *bool
	AuditOTLPEndpoint         string

	// Origin describes which registry hive first contributed a value.
	// One of "hklm-policies" | "hklm-defaults" | "hkcu" | "none" | "non-windows".
	Origin string
	// KeysFound counts all registry values that contributed to this overlay
	// (across all hives). Used in the policy.applied audit event.
	KeysFound int
}

type ProviderModalityOptions struct {
	STT        ProviderOptionOverrides `toml:"stt"`
	TTS        ProviderOptionOverrides `toml:"tts"`
	VoiceAgent ProviderOptionOverrides `toml:"voice_agent"`
}

type ProviderOptionOverrides struct {
	Language           string   `toml:"language,omitempty"`
	DetectLanguage     *bool    `toml:"detect_language,omitempty"`
	Punctuation        *bool    `toml:"punctuation,omitempty"`
	SmartFormat        *bool    `toml:"smart_format,omitempty"`
	Dictation          *bool    `toml:"dictation,omitempty"`
	FillerWords        *bool    `toml:"filler_words,omitempty"`
	Numerals           *bool    `toml:"numerals,omitempty"`
	VocabularyBias     *bool    `toml:"vocabulary_bias,omitempty"`
	Keyterms           []string `toml:"keyterms,omitempty"`
	PromptHint         string   `toml:"prompt_hint,omitempty"`
	SpeakerDiarization *bool    `toml:"speaker_diarization,omitempty"`
	Timestamps         *bool    `toml:"timestamps,omitempty"`
	EndpointingMs      *int     `toml:"endpointing_ms,omitempty"`
	TurnDetection      *bool    `toml:"turn_detection,omitempty"`
	Voice              string   `toml:"voice,omitempty"`
	Speed              *float64 `toml:"speed,omitempty"`
	AudioFormat        string   `toml:"audio_format,omitempty"`
}

type ProviderOptionsConfig struct {
	Deepgram    ProviderModalityOptions `toml:"deepgram"`
	OpenAI      ProviderModalityOptions `toml:"openai"`
	Groq        ProviderModalityOptions `toml:"groq"`
	Google      ProviderModalityOptions `toml:"google"`
	AssemblyAI  ProviderModalityOptions `toml:"assemblyai"`
	OpenRouter  ProviderModalityOptions `toml:"openrouter"`
	HuggingFace ProviderModalityOptions `toml:"huggingface"`
	Local       ProviderModalityOptions `toml:"local"`
	Ollama      ProviderModalityOptions `toml:"ollama"`
}

type ProvidersConfig struct {
	OpenAI     OpenAIProviderConfig     `toml:"openai"`
	Groq       GroqProviderConfig       `toml:"groq"`
	Google     GoogleProviderConfig     `toml:"google"`
	Deepgram   DeepgramProviderConfig   `toml:"deepgram"`
	AssemblyAI AssemblyAIProviderConfig `toml:"assemblyai"`
	Ollama     OllamaProviderConfig     `toml:"ollama"`
	OpenRouter OpenRouterProviderConfig `toml:"openrouter"`
}

type RoleConfig struct {
	ID                          string   `toml:"id"`
	DisplayName                 string   `toml:"display_name"`
	SystemPrompt                string   `toml:"system_prompt"`
	RefinementPrompt            string   `toml:"refinement_prompt"`
	Locale                      string   `toml:"locale"`
	VocabularyHint              string   `toml:"vocabulary_hint"`
	ToolAllowlist               []string `toml:"tool_allowlist"`
	Temperature                 float64  `toml:"temperature"`
	ThinkingEnabled             bool     `toml:"thinking_enabled"`
	ThinkingLevel               string   `toml:"thinking_level"`
	IncludeThoughts             bool     `toml:"include_thoughts"`
	ThinkingBudget              int      `toml:"thinking_budget"`
	AutomaticActivityDetection  bool     `toml:"automatic_activity_detection"`
	VADStartSensitivity         string   `toml:"vad_start_sensitivity"`
	VADEndSensitivity           string   `toml:"vad_end_sensitivity"`
	VADPrefixPaddingMs          int      `toml:"vad_prefix_padding_ms"`
	VADSilenceDurationMs        int      `toml:"vad_silence_duration_ms"`
	ActivityHandling            string   `toml:"activity_handling"`
	TurnCoverage                string   `toml:"turn_coverage"`
	ContextCompressionEnabled   bool     `toml:"context_compression_enabled"`
	ContextCompressionTriggerTk int64    `toml:"context_compression_trigger_tokens"`
	ContextCompressionTargetTk  int64    `toml:"context_compression_target_tokens"`
	EnableAffectiveDialog       bool     `toml:"enable_affective_dialog"`
}

type RoutingConfig struct {
	Strategy                string  `toml:"strategy"`
	PreferLocalUnderSeconds float64 `toml:"prefer_local_under_seconds"`
	ParallelCloud           bool    `toml:"parallel_cloud"`
	ReplaceOnBetter         bool    `toml:"replace_on_better"`
}

type SequenceConfig struct {
	ID          string               `toml:"id"`
	DisplayName string               `toml:"display_name"`
	Description string               `toml:"description"`
	Completion  string               `toml:"completion"` // "all_steps" | "explicit_close" | "max_turns"
	MaxTurns    int                  `toml:"max_turns"`
	Steps       []SequenceStepConfig `toml:"steps"`
}

type SequenceStepConfig struct {
	ID           string   `toml:"id"`
	Instruction  string   `toml:"instruction"`
	ExitCriteria string   `toml:"exit_criteria"`
	RequireTools []string `toml:"require_tools"`
	MaxTurns     int      `toml:"max_turns"`
}

type ServerAdminAuthSettings struct {
	Enabled       *bool  `json:"enabled,omitempty"`
	Username      string `json:"username,omitempty"`
	PasswordHash  string `json:"password_hash,omitempty"`
	PasswordValue string `json:"password,omitempty"`
}

type ServerAssistSettings struct {
	EnabledTools []string `json:"enabled_tools,omitempty"`
}

type ServerAuthSettings struct {
	Mode           string `json:"mode,omitempty"`
	BearerTokenEnv string `json:"bearer_token_env,omitempty"`
	GenerateToken  *bool  `json:"generate_token,omitempty"`
	TokenValue     string `json:"token_value,omitempty"`
}

type ServerConfig struct {
	ListenAddr        string   `toml:"listen_addr"`          // e.g. ":8080"
	PublicURL         string   `toml:"public_url"`           // external API base URL, e.g. https://speechkit.example.com/api
	Modes             []string `toml:"modes"`                // subset of ["dictation","assist","voiceagent"]; empty = all
	AuthMode          string   `toml:"auth_mode"`            // "none" | "bearer" | "edge_hmac" | "bearer_or_edge"
	BearerTokenEnv    string   `toml:"bearer_token_env"`     // env var name holding the bearer token
	BearerRole        string   `toml:"bearer_role"`          // optional role for static bearer callers, e.g. "admin"
	AdminAuthEnabled  bool     `toml:"admin_auth_enabled"`   // enables setup/admin UI username/password login
	AdminUsername     string   `toml:"admin_username"`       // setup/admin UI username; not used by API clients
	AdminPasswordHash string   `toml:"admin_password_hash"`  // bcrypt hash for setup/admin UI login
	EdgeAuthSecretEnv string   `toml:"edge_auth_secret_env"` // env var name holding the HMAC secret
	// SmokeTokenEnv names an optional env var that holds a public-friendly
	// demo bearer token. When set, the smoke UI on `/` embeds the token in
	// the rendered HTML so visitors can run all three modes without
	// configuring credentials. The smoke identity is tagged Source="smoke"
	// (Plan="demo") so handlers and rate-limiters can distinguish demo
	// traffic. Leave empty to disable smoke-from-page entirely; operators
	// must then paste their bearer token in the UI manually.
	SmokeTokenEnv      string   `toml:"smoke_token_env"`
	PublicBaseURL      string   `toml:"public_base_url"`     // public server URL used for returned client URLs
	TrustedProxyCIDRs  []string `toml:"trusted_proxy_cidrs"` // proxies allowed to supply X-Forwarded-* headers
	CORSAllowedOrigins []string `toml:"cors_allowed_origins"`
	RateLimitRPS       float64  `toml:"rate_limit_rps"`
	RateLimitBurst     int      `toml:"rate_limit_burst"`
	// RateLimitEndpointCosts assigns per-endpoint token costs so
	// expensive handlers (LLM, transcription, voice-agent session
	// create) drain the bucket faster than cheap ones. Keys are
	// either "METHOD PATH" (e.g. "POST /v1/dictation/transcribe")
	// or bare PATH. Missing entries default to 1.0. Audit S-4.
	RateLimitEndpointCosts map[string]float64 `toml:"rate_limit_endpoint_costs"`
	// DemoDailyQuota caps how many requests a Plan="demo" identity
	// (the smoke-token surface) may make per UTC day, keyed by
	// UserID + client IP. Zero disables the quota. Audit S-5.
	DemoDailyQuota         int `toml:"demo_daily_quota"`
	MaxUploadMB            int `toml:"max_upload_mb"`
	ReadHeaderTimeoutSec   int `toml:"read_header_timeout_sec"`
	ReadTimeoutSec         int `toml:"read_timeout_sec"`
	IdleTimeoutSec         int `toml:"idle_timeout_sec"`
	MaxHeaderBytes         int `toml:"max_header_bytes"`
	MaxDecodedAudioSeconds int `toml:"max_decoded_audio_seconds"`
	MaxVoiceAgentSessions  int `toml:"max_voiceagent_sessions"` // global cap
	MaxSessionsPerUser     int `toml:"max_sessions_per_user"`
	TicketTTLSec           int `toml:"ticket_ttl_sec"` // Voice Agent WS ticket TTL
	// VoiceAgentIdleTimeoutSec terminates a Voice Agent WebSocket session
	// after N seconds without any client- or provider-side activity.
	// Defaults to 900 (15 min). Set to 0 to disable the server-side idle
	// timeout (kernel-level idle handling stays in effect either way).
	VoiceAgentIdleTimeoutSec int `toml:"voiceagent_idle_timeout_sec"`
	// WSReadLimitBytes caps the per-frame size the Voice Agent WebSocket
	// will read from a client. Zero or negative defaults to 64 KiB,
	// which leaves ample headroom over real PCM chunk sizes (well under
	// 4 KB) without giving a single frame a 1 MiB memory amplification
	// vector. Bumped only for non-standard payloads.
	WSReadLimitBytes int64                `toml:"ws_read_limit_bytes"`
	LiveKit          ServerLiveKitConfig  `toml:"livekit"`
	WhisperBinary    string               `toml:"whisper_binary"` // absolute path inside container
	WhisperPort      int                  `toml:"whisper_port"`   // loopback port for whisper.cpp server
	ModelDir         string               `toml:"model_dir"`      // persistent volume, e.g. /var/lib/speechkit/models
	LogFormat        string               `toml:"log_format"`     // "json" | "text"
	LogLevel         string               `toml:"log_level"`      // "debug" | "info" | "warn" | "error"
	Features         ServerFeaturesConfig `toml:"features"`

	// TrainingData configures the server-side wake-word activation
	// collection endpoint. Default OFF so an operator must explicitly
	// accept training-data uploads from clients. See
	// docs/wakeword-training-data.md.
	TrainingData ServerTrainingDataConfig `toml:"training_data"`

	// Security configures the HTTP security-header middleware (CSP,
	// X-Frame-Options, Referrer-Policy, optional HSTS). Headers are on by
	// default; the zero value yields a strict baseline.
	Security ServerSecurityConfig `toml:"security"`

	// Debug gates runtime debugging surfaces (pprof). Off by default.
	Debug ServerDebugConfig `toml:"debug"`

	// OIDC configures JWT validation against an external identity provider,
	// used only when auth_mode = "oidc".
	OIDC ServerOIDCConfig `toml:"oidc"`
}

type ServerConnectionConfig struct {
	// Enabled is a compatibility mirror for clients that still display a
	// top-level server toggle. Runtime routing is determined by each mode's
	// mode_source; do not use Enabled as a second execution gate.
	Enabled bool `toml:"enabled"`

	// URL is the base URL of the speechkit-server, e.g.
	// "https://speechkit.example.com" or "http://localhost:8080".
	URL string `toml:"url"`

	// BearerTokenEnv names the env var that holds the bearer token sent in
	// the Authorization header. Defaults to SPEECHKIT_SERVER_TOKEN. The
	// value is never read from the TOML file itself — only the env var name
	// is configured here.
	BearerTokenEnv string `toml:"bearer_token_env"`

	// AuthMode selects how the resolved token is attached to outbound
	// requests. "bearer" sends Authorization: Bearer <token>. "api_key"
	// sends X-Api-Key: <token> for servers that use header-based API keys.
	// Empty/missing defaults to "bearer".
	AuthMode string `toml:"auth_mode"`

	// FallbackToLocal makes the device app fall back to the in-process
	// Framework kernel if a server call fails or the server is unreachable.
	// Useful for laptop deployments that may be offline; should be false
	// for kiosks that must never silently downgrade to local processing.
	FallbackToLocal bool `toml:"fallback_to_local"`

	// RequestTimeoutSec caps non-streaming HTTP calls (Dictation, Assist).
	// 0 means no explicit timeout (the underlying http.Client default
	// applies). Voice Agent WebSocket sessions are not affected.
	RequestTimeoutSec int `toml:"request_timeout_sec"`

	// ActiveTargetID selects the registered server target copied into the
	// compatibility fields above. Empty means the top-level URL/env/auth fields
	// are an ad-hoc single target.
	ActiveTargetID string `toml:"active_target_id"`

	// Targets is the optional local registry of SpeechKit server endpoints the
	// device can switch between. These are user/operator configured; product
	// builds must not inject private gateway/origin entries here.
	Targets []ServerConnectionTargetConfig `toml:"targets"`
}

type ServerConnectionTargetConfig struct {
	ID                string `toml:"id"`
	Label             string `toml:"label"`
	URL               string `toml:"url"`
	BearerTokenEnv    string `toml:"bearer_token_env"`
	AuthMode          string `toml:"auth_mode"`
	FallbackToLocal   bool   `toml:"fallback_to_local"`
	RequestTimeoutSec int    `toml:"request_timeout_sec"`
}

type ServerCredentialSettings struct {
	OpenAI      ServerProviderCredentialSettings `json:"openai,omitempty"`
	Groq        ServerProviderCredentialSettings `json:"groq,omitempty"`
	Google      ServerProviderCredentialSettings `json:"google,omitempty"`
	HuggingFace ServerProviderCredentialSettings `json:"huggingface,omitempty"`
	OpenRouter  ServerProviderCredentialSettings `json:"openrouter,omitempty"`
}

type ServerDebugConfig struct {
	PprofEnabled bool `toml:"pprof_enabled"`
	PprofPublic  bool `toml:"pprof_public"`
}

type ServerDictationSettings struct {
	Dictionary *string `json:"dictionary,omitempty"`
}

type ServerFeaturesConfig struct {
	Catalog      bool `toml:"catalog"`
	StorageReads bool `toml:"storage_reads"`
	Vocabulary   bool `toml:"vocabulary"`
	TTSDirect    bool `toml:"tts_direct"`
}

type ServerLLMSettings struct {
	Enabled      *bool  `json:"enabled,omitempty"`
	BaseURL      string `json:"base_url,omitempty"`
	UtilityModel string `json:"utility_model,omitempty"`
	AssistModel  string `json:"assist_model,omitempty"`
	AgentModel   string `json:"agent_model,omitempty"`
	HFRepo       string `json:"hf_repo,omitempty"`
}

type ServerLiveKitConfig struct {
	Enabled      bool   `toml:"enabled"`
	URL          string `toml:"url"`            // e.g. wss://livekit.example.com
	APIKeyEnv    string `toml:"api_key_env"`    // env var name holding the LiveKit API key
	APISecretEnv string `toml:"api_secret_env"` // env var name holding the LiveKit API secret
	TokenTTLSec  int    `toml:"token_ttl_sec"`  // join-token TTL
	RoomPrefix   string `toml:"room_prefix"`    // room name prefix for SpeechKit-managed rooms
}

type ServerModeProviderSettings struct {
	Dictation  ServerModeSetting `json:"dictation,omitempty"`
	Assist     ServerModeSetting `json:"assist,omitempty"`
	VoiceAgent ServerModeSetting `json:"voice_agent,omitempty"`
}

type ServerModeSetting struct {
	Enabled      *bool  `json:"enabled,omitempty"`
	ProviderKind string `json:"provider_kind,omitempty"`
	ProfileID    string `json:"profile_id,omitempty"`
	Model        string `json:"model,omitempty"`
}

type ServerModelSettings struct {
	Version            int                        `json:"version,omitempty"`
	OnboardingComplete bool                       `json:"onboarding_complete,omitempty"`
	OnboardingVersion  string                     `json:"onboarding_version,omitempty"`
	ServerAuth         ServerAuthSettings         `json:"server_auth,omitempty"`
	AdminAuth          ServerAdminAuthSettings    `json:"admin_auth,omitempty"`
	Modes              ServerModeProviderSettings `json:"modes,omitempty"`
	Credentials        ServerCredentialSettings   `json:"credentials,omitempty"`
	Dictation          ServerDictationSettings    `json:"dictation,omitempty"`
	Assist             ServerAssistSettings       `json:"assist,omitempty"`
	STT                ServerSTTSettings          `json:"stt,omitempty"`
	LLM                ServerLLMSettings          `json:"llm,omitempty"`
	VoiceAgent         ServerVoiceAgentSettings   `json:"voice_agent,omitempty"`
	TTS                ServerOptionalTTSSettings  `json:"tts,omitempty"`
}

type ServerOIDCConfig struct {
	JWKSURL          string `toml:"jwks_url"`
	Issuer           string `toml:"issuer"`
	Audience         string `toml:"audience"`
	ClockSkewSeconds int    `toml:"clock_skew_seconds"` // tolerated exp/nbf skew; default 60
	OrgClaim         string `toml:"org_claim"`          // claim -> OrgID; default "org_id"
	RoleClaim        string `toml:"role_claim"`         // claim -> Role; default "role"
}

type ServerOptionalTTSSettings struct {
	Enabled *bool `json:"enabled,omitempty"`
}

type ServerProviderCredentialSettings struct {
	Enabled *bool  `json:"enabled,omitempty"`
	Env     string `json:"env,omitempty"`
	Value   string `json:"value,omitempty"`
}

type ServerSTTSettings struct {
	Enabled *bool  `json:"enabled,omitempty"`
	URL     string `json:"url,omitempty"`
	Model   string `json:"model,omitempty"`
}

type ServerSecurityConfig struct {
	Disabled              bool   `toml:"disabled"`                // turn the middleware off entirely (not recommended)
	ContentSecurityPolicy string `toml:"content_security_policy"` // override the default strict API CSP
	FrameOptions          string `toml:"frame_options"`           // override X-Frame-Options (default "DENY")
	ReferrerPolicy        string `toml:"referrer_policy"`         // override Referrer-Policy (default "no-referrer")
	HSTS                  bool   `toml:"hsts"`                    // emit Strict-Transport-Security (only meaningful behind TLS)
	HSTSMaxAgeSeconds     int    `toml:"hsts_max_age_seconds"`    // HSTS max-age; default 63072000 (2y) when HSTS is on
}

type ServerTrainingDataConfig struct {
	// AcceptUploads gates POST /v1/wakeword/activations. When false
	// the endpoint returns 503 with a clear "feature disabled"
	// payload so device-side uploaders back off gracefully. Default
	// false.
	AcceptUploads bool `toml:"accept_uploads"`

	// AudioDir is the filesystem root where uploaded audio files
	// are stored. Empty resolves to <data>/wakeword-activations/ in
	// the container. Files land under <audio_dir>/<org>/<user>/<id>.wav.
	AudioDir string `toml:"audio_dir"`

	// PerUserQuotaBytes caps how many bytes one user can have on
	// disk before the server rejects further uploads with 413. Zero
	// = unlimited. Default 1 GiB (1073741824).
	PerUserQuotaBytes int64 `toml:"per_user_quota_bytes"`

	// RetentionDays auto-deletes uploaded clips older than this
	// many days via the maintenance worker. Zero = no auto-delete.
	// Default 180.
	RetentionDays int `toml:"retention_days"`
}

type ServerVoiceAgentSettings struct {
	Provider        string  `json:"provider,omitempty"`
	AgentProfileID  string  `json:"agent_profile_id,omitempty"`
	AgentSequenceID string  `json:"agent_sequence_id,omitempty"`
	PromptTemplate  *string `json:"prompt_template,omitempty"`
}

type ShortcutLocaleConfig struct {
	LeadingFillers []string `toml:"leading_fillers"`
	CopyLast       []string `toml:"copy_last"`
	InsertLast     []string `toml:"insert_last"`
	Summarize      []string `toml:"summarize"`
	QuickNote      []string `toml:"quick_note"`
}

type ShortcutsConfig struct {
	Locale map[string]ShortcutLocaleConfig `toml:"locale"`
}

type SpeechDefaultsConfig struct {
	Language       string  `toml:"language"`
	DetectLanguage bool    `toml:"detect_language"`
	Punctuation    bool    `toml:"punctuation"`
	SmartFormat    bool    `toml:"smart_format"`
	VocabularyBias bool    `toml:"vocabulary_bias"`
	Timestamps     bool    `toml:"timestamps"`
	EndpointingMs  int     `toml:"endpointing_ms"`
	TurnDetection  bool    `toml:"turn_detection"`
	Voice          string  `toml:"voice"`
	Speed          float64 `toml:"speed"`
	AudioFormat    string  `toml:"audio_format"`
}

type StoreConfig struct {
	Backend            string `toml:"backend"` // "sqlite" | "postgres" | registered name
	SQLitePath         string `toml:"sqlite_path"`
	PostgresDSN        string `toml:"postgres_dsn"`
	SaveAudio          bool   `toml:"save_audio"`
	AudioRetentionDays int    `toml:"audio_retention_days"`
	MaxAudioStorageMB  int    `toml:"max_audio_storage_mb"`
}

type TTSConfig struct {
	Enabled     bool           `toml:"enabled"`
	Strategy    string         `toml:"strategy"` // "cloud-first", "local-first", "cloud-only", "local-only"
	Voice       string         `toml:"voice"`    // Global default voice override
	Speed       float64        `toml:"speed"`    // Global speed 0.25-4.0, default 1.0
	Format      string         `toml:"format"`   // "mp3", "wav", "opus", "pcm"
	OpenAI      TTSOpenAI      `toml:"openai"`
	Google      TTSGoogle      `toml:"google"`
	Deepgram    TTSDeepgram    `toml:"deepgram"`
	HuggingFace TTSHuggingFace `toml:"huggingface"`
	Local       TTSLocal       `toml:"local"`
	Piper       TTSPiper       `toml:"piper"`
}

type TTSDeepgram struct {
	Enabled bool   `toml:"enabled"`
	Model   string `toml:"model"` // Aura-2 voice id, e.g. "aura-2-thalia-en"
	Voice   string `toml:"voice"` // optional explicit voice override (alias of Model)
}

type TTSGoogle struct {
	Enabled bool   `toml:"enabled"`
	Voice   string `toml:"voice"` // e.g. "de-DE-Neural2-B"
}

type TTSHuggingFace struct {
	Enabled bool   `toml:"enabled"`
	Model   string `toml:"model"` // e.g. "Qwen/Qwen3-TTS-12Hz-1.7B-Base"
}

type TTSLocal struct {
	Enabled   bool   `toml:"enabled"`
	Model     string `toml:"model"`
	ModelPath string `toml:"model_path"`
	Port      int    `toml:"port"`
}

type TTSOpenAI struct {
	Enabled bool   `toml:"enabled"`
	Model   string `toml:"model"` // "tts-1" or "tts-1-hd"
	Voice   string `toml:"voice"` // alloy, echo, fable, onyx, nova, shimmer
}

type TTSPiper struct {
	Enabled  bool   `toml:"enabled"`
	Binary   string `toml:"binary"`    // path to piper executable; empty => "piper" on PATH
	VoiceDir string `toml:"voice_dir"` // directory holding *.onnx voice files
	// DefaultVoices maps a locale short-code ("en", "de", ...) to a voice
	// filename inside VoiceDir. Empty entries fall back to the built-in
	// defaults (en_US-amy-medium.onnx, de_DE-thorsten-medium.onnx).
	DefaultVoices map[string]string `toml:"default_voices"`
	TimeoutSec    int               `toml:"timeout_sec"` // subprocess timeout; 0 => 30 s
}

type TelemetryConfig struct {
	UpdateCheck bool `toml:"update_check"`

	// TracesOTLPEndpoint enables exporting the framework's OpenTelemetry spans
	// (STT routing, TTS, Voice Agent, server lifecycle) to a vendor-neutral
	// OTLP/HTTP traces receiver. When empty (the default) the server installs
	// no TracerProvider, so every otel.Tracer span stays a zero-cost no-op and
	// behaviour is unchanged. Point it at any OTLP/HTTP collector; for Sentry's
	// OTLP ingestion use the full URL
	// https://<org>.ingest.<region>.sentry.io/api/<project>/otlp/v1/traces.
	TracesOTLPEndpoint string `toml:"traces_otlp_endpoint"`
	// TracesSampleRate is the head sampling ratio in [0,1]. 0 (or >=1) means
	// always-sample, which is the right default for a low-traffic dogfood
	// deployment where we want to see every trace.
	TracesSampleRate float64 `toml:"traces_sample_rate"`
	// The OTLP auth secret stays out of config files. OTLPAuthHeaderName is the
	// HTTP header to attach (e.g. "x-sentry-auth"); its value is read at startup
	// from the environment variable named by OTLPAuthHeaderEnv (e.g. a
	// Doppler/Render secret holding "sentry sentry_key=<public_key>"). When
	// either is empty no auth header is sent (suitable for a local collector).
	OTLPAuthHeaderName string `toml:"otlp_auth_header_name"`
	OTLPAuthHeaderEnv  string `toml:"otlp_auth_header_env"`
	// ServiceName + Environment tag every exported span's resource so Sentry can
	// separate speechkit-server from other services and staging from prod.
	ServiceName string `toml:"service_name"`
	Environment string `toml:"environment"`
}

type UIConfig struct {
	OverlayEnabled          bool                           `toml:"overlay_enabled"`
	OverlayPosition         string                         `toml:"overlay_position"` // "top", "bottom", "left", "right"
	OverlayMovable          bool                           `toml:"overlay_movable"`
	OverlayFreeX            int                            `toml:"overlay_free_x"`
	OverlayFreeY            int                            `toml:"overlay_free_y"`
	OverlayMonitorPositions map[string]OverlayFreePosition `toml:"overlay_monitor_positions"`
	Visualizer              string                         `toml:"visualizer"`
	Design                  string                         `toml:"design"`
	AssistOverlayMode       string                         `toml:"assist_overlay_mode"`
	VoiceAgentOverlayMode   string                         `toml:"voice_agent_overlay_mode"`
}

type UpdateConfig struct {
	Enabled                bool   `toml:"enabled"`
	ManifestURL            string `toml:"manifest_url"`
	CheckIntervalHours     int    `toml:"check_interval_hours"`
	SignaturePinThumbprint string `toml:"signature_pin_thumbprint"` // optional Authenticode SHA-1 thumbprint; if set, installer signature verification additionally checks cert thumbprint matches (defense against compromised signing cert)
}

type VPSConfig struct {
	Enabled   bool   `toml:"enabled"`
	URL       string `toml:"url"`
	Model     string `toml:"model"`
	APIKeyEnv string `toml:"api_key_env"`
}

type VocabularyConfig struct {
	Dictionary string `toml:"dictionary"`
}

type VoiceAgentConfig struct {
	Enabled bool `toml:"enabled"`
	// Provider selects the backend that drives a Voice Agent session.
	// Supported values:
	//   ""          (default) — same as "gemini"
	//   "gemini"    — Google Gemini Live (cloud, GOOGLE_AI_API_KEY required)
	//   "openai"    — OpenAI Realtime API (cloud, OPENAI_API_KEY required)
	//   "deepgram"  — Deepgram Voice Agent (cloud, DEEPGRAM_API_KEY required)
	//   "cascaded"  — self-hosted whisper.cpp → Genkit agent LLM → TTS pipeline
	//                 (CPU-capable; no external realtime dependency)
	//   "moshi"     — self-hosted Kyutai Moshi Rust server (GPU required, M9b)
	//
	// The Server-Target reads this field via cmd/speechkit-server. The Device-
	// Target runs "gemini", "openai", and "deepgram" in-process; other values
	// fall back to Gemini Live (or the pipeline fallback when enabled).
	Provider      string `toml:"provider"`
	Model         string `toml:"model"`          // Real-time model ID (e.g. "gemini-3.1-flash-live-preview")
	FallbackModel string `toml:"fallback_model"` // Fallback real-time model
	Voice         string `toml:"voice"`          // Voice name for real-time model
	// Deepgram Voice Agent think-LLM overrides. The think leg reasons over the
	// transcript; listen (Nova-3) and speak (Aura-2) stay Deepgram. When unset,
	// the kernel default (Deepgram-managed open_ai/gpt-4o-mini) applies. Setting
	// DeepgramThinkEndpointURL + DeepgramThinkAPIKeyEnv switches the think leg to
	// a bring-your-own LLM deployment, with the credential resolved from the
	// named env var (env -> Doppler). Read by the Server- and Device-Target
	// Deepgram Voice Agent wiring; ignored by the Gemini/cascaded backends.
	DeepgramThinkProvider    string `toml:"deepgram_think_provider"`
	DeepgramThinkModel       string `toml:"deepgram_think_model"`
	DeepgramThinkEndpointURL string `toml:"deepgram_think_endpoint_url"`
	DeepgramThinkAPIKeyEnv   string `toml:"deepgram_think_api_key_env"`
	AgentProfileID           string `toml:"agent_profile_id"`  // Built-in Voice Agent profile ID; "default" preserves current behavior.
	AgentSequenceID          string `toml:"agent_sequence_id"` // Optional workflow sequence ID; empty uses the selected persona default.
	FrameworkPrompt          string `toml:"framework_prompt"`  // Durable host/framework instruction that defines the Voice Agent behavior
	RefinementPrompt         string `toml:"refinement_prompt"` // User-specific refinement appended to the framework prompt
	AutoStartOnLaunch        bool   `toml:"auto_start_on_launch"`
	CloseBehavior            string `toml:"close_behavior"` // "continue" keeps the conversation window in the taskbar; "new_chat" ends the current chat on close
	// BargeIn controls whether the microphone stays open while the agent is
	// speaking so the user can interrupt mid-answer:
	//   "auto"   (default) — full duplex when the active output device looks
	//            like a headset (closed acoustic path, no speaker bleed);
	//            half duplex otherwise. Evaluated at session start.
	//   "always" — full duplex on every output device. Only sensible with
	//            hardware/OS echo cancellation; without it the agent hears
	//            itself through the speakers and interrupts itself.
	//   "never"  — half duplex: the mic is muted while the agent speaks and
	//            until the buffered answer finished playing.
	BargeIn                string `toml:"barge_in"`
	ReminderAfterIdleSec   int    `toml:"reminder_after_idle_sec"`
	DeactivateAfterIdleSec int    `toml:"deactivate_after_idle_sec"`
	// HoldReleaseGraceSec controls how long the Voice Agent stays open after
	// the user releases a hold-to-talk shortcut so the model has time to
	// deliver its reply. 0 (or unset) falls back to the kernel default
	// (10 seconds). The Device-Target hard-caps this at 30 seconds; values
	// above that are silently clamped at runtime so a misconfigured profile
	// cannot strand the user in a "still active" session.
	HoldReleaseGraceSec             int                    `toml:"hold_release_grace_sec"`
	PipelineFallback                bool                   `toml:"pipeline_fallback"` // Use STT -> Agent LLM -> optional TTS when the selected Voice Agent profile is not native realtime.
	ShowPrompter                    bool                   `toml:"show_prompter"`     // Show live transcript prompter window
	EnableSessionSummary            bool                   `toml:"enable_session_summary"`
	EnableInputTranscript           bool                   `toml:"enable_input_transcript"`
	EnableOutputTranscript          bool                   `toml:"enable_output_transcript"`
	EnableAffectiveDialog           bool                   `toml:"enable_affective_dialog"`
	ThinkingEnabled                 bool                   `toml:"thinking_enabled"`
	IncludeThoughts                 bool                   `toml:"include_thoughts"`
	ThinkingBudget                  int                    `toml:"thinking_budget"`
	ThinkingLevel                   string                 `toml:"thinking_level"`
	ContextCompressionEnabled       bool                   `toml:"context_compression_enabled"`
	ContextCompressionTriggerTokens int64                  `toml:"context_compression_trigger_tokens"`
	ContextCompressionTargetTokens  int64                  `toml:"context_compression_target_tokens"`
	AutomaticActivityDetection      bool                   `toml:"automatic_activity_detection"`
	ActivityHandling                string                 `toml:"activity_handling"`
	TurnCoverage                    string                 `toml:"turn_coverage"`
	VADStartSensitivity             string                 `toml:"vad_start_sensitivity"`
	VADEndSensitivity               string                 `toml:"vad_end_sensitivity"`
	VADPrefixPaddingMs              int                    `toml:"vad_prefix_padding_ms"`
	VADSilenceDurationMs            int                    `toml:"vad_silence_duration_ms"`
	Limits                          VoiceAgentLimitsConfig `toml:"limits"`
}

type VoiceAgentLimitsConfig struct {
	MaxGlobalSessions      int `toml:"max_global_sessions"`
	MaxPerIdentitySessions int `toml:"max_per_identity_sessions"`
}

type WakewordAutoEndConfig struct {
	// SilenceCutoffSec is the duration without user audio activity (in
	// whole seconds) after which a wake-word-triggered session ends.
	// Zero falls back to the framework default (10s).
	SilenceCutoffSec int `toml:"silence_cutoff_sec"`

	// ExitPhrases is the case-insensitive substring list checked against
	// each user-transcript snippet. Empty falls back to the framework
	// default (DE+EN common closers: "danke", "tschuess", "ende", "stop",
	// "thanks", "bye", "goodbye", ...).
	ExitPhrases []string `toml:"exit_phrases"`
}

type WakewordConfig struct {
	// Enabled gates the entire feature. Default false (opt-in).
	Enabled bool `toml:"enabled"`

	// Backend selects the local detector implementation. The Windows app
	// ships Sherpa-ONNX KWS, LiveKit/openWakeWord ONNX, and STT phrase-match
	// as explicit selectable paths so test builds can compare detector
	// behaviour without silently falling back to a different implementation.
	Backend string `toml:"backend"`

	// PhraseID picks one of SpeechKit's curated wake phrases from
	// wakeword.DefaultCatalog (e.g. "hey_quby", "hey_computer",
	// "hey_jarvis", "hey_mira"). When set, the corresponding ONNX file is
	// resolved automatically and Phrase/ModelPath below are ignored. When
	// empty, the explicit Phrase + ModelPath fields are used instead
	// (custom phrase mode). Switching via this field is how users pick a
	// different wake phrase in settings without editing paths by hand.
	PhraseID string `toml:"phrase_id"`

	// Phrase is the display label of the trained wake phrase, surfaced in
	// the tray and status feed. It has NO effect on detection — the ONNX
	// model encodes the actual phrase(s). One model can be trained to
	// fire on multiple pronunciation variants (e.g. "Hey Cubi" and
	// "Hey Kubi" for the same brand "Quby") via target_phrases in the
	// training yaml; the display label here remains a single brand string.
	//
	// Ignored when PhraseID matches a catalog entry (the catalog's
	// DisplayName is used instead).
	Phrase string `toml:"phrase"`

	// ModelPath is the path to the trained phrase prediction model (.onnx).
	// Empty resolves to <data_dir>/models/wakeword/hey_quby.onnx at runtime.
	//
	// Ignored when PhraseID matches a catalog entry (the catalog's
	// FileName is resolved inside the wake-word models directory).
	ModelPath string `toml:"model_path"`

	// MelspecModelPath and EmbeddingModelPath point at the shared
	// openWakeWord upstream models. Empty values resolve to the same
	// directory as ModelPath with canonical filenames.
	MelspecModelPath   string `toml:"melspec_model_path"`
	EmbeddingModelPath string `toml:"embedding_model_path"`

	// DefaultMode is the runtime mode triggered when the wake phrase fires.
	// One of "dictate" | "assist" | "voice_agent". Defaults to voice_agent.
	DefaultMode string `toml:"default_mode"`

	// Threshold is the minimum probability to count a frame as a hit.
	// Range (0.0, 1.0]. Backend-specific defaults when this is 0:
	//   - LiveKit/openWakeWord: 0.5 (Wyoming/openWakeWord canonical)
	//   - Sherpa-onnx KWS: 0.25 (sherpa-onnx upstream default)
	//   - STT phrase match: 0 (substring match, no acoustic probability)
	// Use the in-app "Test wake word" self-test in Settings to calibrate
	// for your specific microphone + environment instead of guessing.
	Threshold float64 `toml:"threshold"`

	// MinConsecutiveFrames is the number of consecutive above-threshold
	// frames required before a trigger fires. Higher = fewer false-accepts,
	// more false-rejects. Defaults to 2.
	MinConsecutiveFrames int `toml:"min_consecutive_frames"`

	// CooldownMs is the minimum gap between two triggers, in milliseconds.
	// Defaults to 1500ms.
	CooldownMs int `toml:"cooldown_ms"`

	// DebugMode enables verbose detector diagnostics. When true the sidecars
	// emit per-decode score events (openWakeWord) or set the sherpa-onnx
	// ModelConfig.Debug flag (Sherpa KWS), and the host adapter forwards
	// those signals into the user-visible log feed. Default false — only flip
	// on while tuning a wake phrase, the score event stream is high-volume.
	DebugMode bool `toml:"debug_mode"`

	// AutoEnd controls the framework-level auto-end policy applied to any
	// session that the wake-word triggered. Wake-word-origin Voice-Agent
	// activations terminate automatically on silence after this many
	// seconds, or when the user utters one of the configured exit
	// phrases. Empty values fall back to wakeword.DefaultAutoEndConfig
	// (10s silence + DE/EN exit phrases) so a TOML without an [auto_end]
	// block still gets the framework baseline. There is intentionally no
	// hard-cap on session duration — Voice-Agent is designed for
	// multi-hour dialogs and a forced cap would break regular use.
	AutoEnd WakewordAutoEndConfig `toml:"auto_end"`

	// TrainingData controls the optional activation-capture pipeline:
	// the sidecar saves the surrounding audio of each detection to a
	// local directory and, when explicitly opted-in, uploads those
	// clips to a SpeechKit-Server for training-data collection. ALL
	// fields default to OFF — see docs/wakeword-training-data.md.
	TrainingData WakewordTrainingDataConfig `toml:"training_data"`
}

type WakewordTrainingDataConfig struct {
	// LocalCaptureEnabled toggles the sidecar ring-buffer + WAV
	// writer that persists each detection's audio to LocalCaptureDir.
	// Default false. Even when true, no network traffic happens
	// unless UploadEnabled is ALSO true.
	LocalCaptureEnabled bool `toml:"local_capture_enabled"`

	// LocalCaptureDir is the filesystem root that holds the captured
	// WAV + JSON pairs. Empty resolves to
	// %LOCALAPPDATA%/SpeechKit/wakeword-activations on Windows.
	LocalCaptureDir string `toml:"local_capture_dir"`

	// LocalMaxFiles caps how many activation pairs live on disk
	// before the oldest get deleted by the rotation worker. Default
	// 500. Set to 0 for unlimited (the retention_days limit still
	// applies).
	LocalMaxFiles int `toml:"local_max_files"`

	// LocalRetentionDays auto-deletes activation pairs older than
	// this many days at sidecar startup and every hour after. Zero
	// means "do not auto-delete by age" (LocalMaxFiles still
	// applies). Default 30.
	LocalRetentionDays int `toml:"local_retention_days"`

	// PreRollMs is the duration of audio captured BEFORE the
	// detection trigger. The sidecar keeps this many milliseconds
	// of PCM in a ring buffer so the moment the detection fires it
	// can write the leading audio that contains the actual wake
	// phrase. Default 1500 (matches the wake-phrase windows the
	// existing detectors are tuned against).
	PreRollMs int `toml:"pre_roll_ms"`

	// PostRollMs is the duration of audio captured AFTER the
	// detection trigger. Useful for catching the speaker continuing
	// past the wake phrase so a labeler can hear whether the
	// utterance was a true positive or noise. Default 500.
	PostRollMs int `toml:"post_roll_ms"`

	// UploadEnabled toggles the background uploader that pushes
	// captured clips to UploadServerURL. Requires both
	// LocalCaptureEnabled=true and UploadServerURL+UploadTokenEnv
	// to be set. Default false.
	UploadEnabled bool `toml:"upload_enabled"`

	// UploadServerURL is the base URL of the SpeechKit-Server that
	// accepts POST /v1/wakeword/activations. Empty falls back to
	// [server_connection].url when set.
	UploadServerURL string `toml:"upload_server_url"`

	// UploadTokenEnv names the env var (resolved via
	// internal/secrets) that holds the bearer token used by the
	// uploader. Default "SPEECHKIT_TRAINING_TOKEN".
	UploadTokenEnv string `toml:"upload_token_env"`

	// UploadOnlyLabeled limits uploads to clips that the user has
	// labeled (correct / false_positive). When true (default),
	// unlabeled clips stay local — privacy-friendly default that
	// only ships audio after explicit user review.
	UploadOnlyLabeled bool `toml:"upload_only_labeled"`

	// UploadIntervalMinutes is the cadence at which the uploader
	// scans LocalCaptureDir and flushes new clips to the server.
	// Default 60.
	UploadIntervalMinutes int `toml:"upload_interval_minutes"`
}