export const MIN_MAX_SESSION_MINUTES = 1; export const MAX_MAX_SESSION_MINUTES = 120; export const OPENAI_REALTIME_MAX_SESSION_MINUTES = 60; export const MIN_INACTIVITY_SECONDS = 20; export const MAX_INACTIVITY_SECONDS = 3600; export const VOICE_MAX_DURATION_WARNING_SECONDS = 120; export const VOICE_INACTIVITY_WARNING_SECONDS = 45; export const INPUT_SPEECH_END_SILENCE_MS = 700; export const SPEAKING_END_MICRO_CAPTURE_MS = 260; export const SPEAKING_END_SHORT_CAPTURE_MS = 900; export const SPEAKING_END_FINALIZE_MICRO_MS = 420; export const SPEAKING_END_FINALIZE_SHORT_MS = 220; // Keep a pause window for normal turns, but cap it tighter for lower end-of-turn latency. export const SPEAKING_END_FINALIZE_QUICK_MS = 800; export const SPEAKING_END_FINALIZE_MIN_MS = 100; export const SPEAKING_END_ADAPTIVE_BUSY_CAPTURE_COUNT = 3; export const SPEAKING_END_ADAPTIVE_HEAVY_CAPTURE_COUNT = 5; export const SPEAKING_END_ADAPTIVE_BUSY_BACKLOG = 2; export const SPEAKING_END_ADAPTIVE_HEAVY_BACKLOG = 4; export const SPEAKING_END_ADAPTIVE_BUSY_SCALE = 0.7; export const SPEAKING_END_ADAPTIVE_HEAVY_SCALE = 0.5; export const CAPTURE_IDLE_FLUSH_MS = INPUT_SPEECH_END_SILENCE_MS + 120; export const CAPTURE_MAX_DURATION_MS = 8_000; export const CAPTURE_NEAR_SILENCE_ABORT_MIN_AGE_MS = 1_000; export const CAPTURE_NEAR_SILENCE_ABORT_ACTIVE_RATIO_MAX = 0.009; export const CAPTURE_NEAR_SILENCE_ABORT_PEAK_MAX = 0.011; export const BOT_TURN_SILENCE_RESET_MS = 1200; // clankvox reports TTS depth every 500ms; if positive buffered telemetry stops // updating for this long, treat it as stale rather than durable truth. export const CLANKVOX_TTS_TELEMETRY_STALE_MS = 1600; // Make barge-in intentionally stubborn so brief talk-over/echo does not cut bot playback. export const BARGE_IN_MIN_SPEECH_MS = 700; // File-ASR/non-realtime captures must be at least this old before barge-in fires, replacing // the old timer-based armAssertiveBargeIn delay. export const BARGE_IN_STT_MIN_CAPTURE_AGE_MS = 500; // After a successful interrupt, briefly suppress outbound audio so tail echo // from the cut-off reply does not immediately retrigger barge-in. export const BARGE_IN_SUPPRESSION_MAX_MS = 4_000; // Grace period after bot TTS audio starts before barge-in is accepted. // Prevents false barge-in from echo of the bot's own voice through user mic. export const BARGE_IN_BOT_AUDIO_ECHO_GUARD_MS = 1500; // Stricter assertiveness thresholds while bot is speaking (echo rejection). // Normal gate: activeSampleRatio > 0.01, peak > 0.012 (anything above silence). // Bot-speaking gate: require clearly audible speech, not just faint echo leak. export const BARGE_IN_BOT_SPEAKING_ACTIVE_RATIO_MIN = 0.06; export const BARGE_IN_BOT_SPEAKING_PEAK_MIN = 0.05; export const ACTIVITY_TOUCH_THROTTLE_MS = 2000; export const ACTIVITY_TOUCH_MIN_SPEECH_MS = 120; export const RESPONSE_FLUSH_DEBOUNCE_MS = 280; export const OPENAI_ACTIVE_RESPONSE_RETRY_MS = 260; export const MIN_RESPONSE_REQUEST_GAP_MS = 700; export const RESPONSE_SILENCE_RETRY_DELAY_MS = 5200; export const MAX_RESPONSE_SILENCE_RETRIES = 2; export const RESPONSE_DONE_SILENCE_GRACE_MS = 1400; // Keep realtime TTS responsive by pausing queued streamed utterances before // the local Discord playback backlog reaches runaway latency. export const REALTIME_ASSISTANT_TTS_BACKPRESSURE_PAUSE_SAMPLES = 48_000 * 4; export const REALTIME_ASSISTANT_TTS_BACKPRESSURE_RESUME_SAMPLES = 48_000 * 2.5; // Keep ASR sessions warm longer to avoid reconnect penalties between turns. export const OPENAI_ASR_SESSION_IDLE_TTL_MS = 60_000; // Return ASR text sooner once transcript updates settle. export const OPENAI_ASR_TRANSCRIPT_STABLE_MS = 120; export const OPENAI_ASR_TRANSCRIPT_WAIT_MAX_MS = 700; // Deadline for waiting on per-user/shared ASR bridge before falling back to // local turn processing so first-reply latency is not blocked by slow ASR. export const OPENAI_ASR_BRIDGE_MAX_WAIT_MS = 700; export const VOICE_INTERRUPT_BURST_QUIET_GAP_MS = 360; export const VOICE_INTERRUPT_BURST_FINAL_QUIET_GAP_MS = 80; export const VOICE_INTERRUPT_BURST_MAX_MS = 1500; export const VOICE_INTERRUPT_CLASSIFIER_TIMEOUT_MS = 900; export const VOICE_INTERRUPT_DECISION_TTL_MS = 30_000; export const VOICE_INTERRUPT_SPEECH_START_SUSTAIN_MS = 700; export const VOICE_INTERRUPT_SPEECH_START_RECHECK_MS = 200; export const VOICE_GENERATION_SOUNDBOARD_CANDIDATE_TIMEOUT_MS = 1_200; export const VOICE_GENERATION_ONLY_WATCHDOG_MS = 20_000; export const CONTEXT_COMPACTION_RECENT_WINDOW = 50; export const CONTEXT_COMPACTION_BATCH_SIZE = 10; export const CONTEXT_COMPACTION_MAX_SUMMARY_CHARS = 1_200; export const CONTEXT_COMPACTION_MAX_NOTE_CHARS = 400; export const CONTEXT_COMPACTION_MAX_OUTPUT_TOKENS = 256; export const OPENAI_TOOL_CALL_EVENT_MAX = 180; export const OPENAI_TOOL_CALL_ARGUMENTS_MAX_CHARS = 24_000; export const OPENAI_TOOL_RESPONSE_DEBOUNCE_MS = 140; export const VOICE_MEMORY_WRITE_MAX_PER_MINUTE = 5; export const LEAVE_DIRECTIVE_PLAYBACK_POLL_MS = 40; export const LEAVE_DIRECTIVE_PLAYBACK_NO_SIGNAL_GRACE_MS = 400; export const LEAVE_DIRECTIVE_REALTIME_AUDIO_START_WAIT_MS = 2200; export const LEAVE_DIRECTIVE_PLAYBACK_MAX_WAIT_MS = 5000; export const ORDERED_REALTIME_PLAYBACK_BARRIER_MAX_WAIT_MS = 30_000; export const VOICE_OUTPUT_LEASE_ASSERTIVE_MS = 1_200; export const VOICE_OUTPUT_LEASE_ATOMIC_MS = 2_400; // Post-audio speech immunity for leased replies. Once the bot starts speaking, // barge-in is blocked for this window so the reply's point can land. export const BARGE_IN_LEASE_IMMUNITY_ASSERTIVE_MS = 2_000; export const BARGE_IN_LEASE_IMMUNITY_ATOMIC_MS = 4_000; // Hard failsafe for room-coalesce holds. The primary flush trigger is capture // finalization removing the last active capture ("room went quiet"). This timeout // only fires if something is genuinely broken — a capture that never finalizes // despite the 8s max-duration cap and idle/silence abort timers. Set well above // CAPTURE_MAX_DURATION_MS so it never preempts normal room settling. export const REALTIME_TURN_COALESCE_WINDOW_MS = 10_000; // Maximum PCM bytes allowed in a coalesced multi-segment turn. export const REALTIME_TURN_COALESCE_MAX_BYTES = 24_000 * 2 * 12; // Keep only one pending realtime turn; newer finalized captures are merged into it. export const REALTIME_TURN_QUEUE_MAX = 1; export const REALTIME_TURN_STALE_SKIP_MS = 2200; export const REALTIME_TURN_PENDING_MERGE_MAX_BYTES = 24_000 * 2 * 30; // Skip trivial micro-clips from speaking_end that frequently hallucinate transcript junk. export const VOICE_TURN_MIN_ASR_CLIP_MS = 100; // Drop near-silent captures before ASR so Discord speaking blips do not become random transcripts. export const VOICE_SILENCE_GATE_MIN_CLIP_MS = 280; export const VOICE_SILENCE_GATE_RMS_MAX = 0.003; export const VOICE_SILENCE_GATE_PEAK_MAX = 0.012; export const VOICE_SILENCE_GATE_ACTIVE_SAMPLE_MIN_ABS = 180; export const VOICE_SILENCE_GATE_ACTIVE_RATIO_MAX = 0.01; export const VOICE_TURN_PROMOTION_MIN_CLIP_MS = 420; export const VOICE_TURN_PROMOTION_PEAK_MIN = 0.016; export const VOICE_TURN_PROMOTION_ACTIVE_RATIO_MIN = 0.02; export const VOICE_FALLBACK_NOISE_GATE_RMS_MAX = 0.0065; export const VOICE_FALLBACK_NOISE_GATE_PEAK_MAX = 0.02; export const VOICE_FALLBACK_NOISE_GATE_ACTIVE_RATIO_MAX = 0.03; export const VOICE_EMPTY_TRANSCRIPT_ERROR_STREAK = 3; export const BOT_TURN_DEFERRED_FLUSH_DELAY_MS = BOT_TURN_SILENCE_RESET_MS + 120; export const BOT_TURN_DEFERRED_QUEUE_MAX = 8; export const BOT_TURN_DEFERRED_COALESCE_MAX = 5; export const VOICE_THOUGHT_LOOP_MIN_SILENCE_SECONDS = 8; export const VOICE_THOUGHT_LOOP_MAX_SILENCE_SECONDS = 300; export const VOICE_THOUGHT_LOOP_MIN_INTERVAL_SECONDS = 8; export const VOICE_THOUGHT_LOOP_MAX_INTERVAL_SECONDS = 600; export const VOICE_THOUGHT_LOOP_BUSY_RETRY_MS = 1400; export const VOICE_THOUGHT_MAX_CHARS = 220; export const VOICE_THOUGHT_MEMORY_SEARCH_LIMIT = 8; export const VOICE_THOUGHT_DECISION_MAX_OUTPUT_TOKENS = 220; export const BOT_DISCONNECT_GRACE_MS = 2500; export const STT_CONTEXT_MAX_MESSAGES = 50; export const STT_TRANSCRIPT_MAX_CHARS = 700; export const STT_REPLY_MAX_CHARS = 1200; export const FILE_ASR_TURN_QUEUE_MAX = 1; export const FILE_ASR_TURN_STALE_SKIP_MS = 4500; export const FILE_ASR_TURN_COALESCE_WINDOW_MS = 1200; export const FILE_ASR_TURN_COALESCE_MAX_BYTES = 24_000 * 2 * 8; export const VOICE_TURN_ADDRESSING_TRANSCRIPT_MAX_CHARS = 260; export const VOICE_DECIDER_HISTORY_MAX_TURNS = 8; export const VOICE_TRANSCRIPT_TIMELINE_MAX_TURNS = 220; export const VOICE_DECIDER_HISTORY_MAX_CHARS = 220; export const VOICE_DECIDER_PROMPT_HISTORY_MAX_CHARS = 900; export const REALTIME_INSTRUCTION_REFRESH_DEBOUNCE_MS = 220; export const REALTIME_CONTEXT_TRANSCRIPT_MAX_CHARS = 420; export const REALTIME_CONTEXT_MEMBER_LIMIT = 12; export const VOICE_MEMBERSHIP_EVENT_MAX_TRACKED = 18; export const VOICE_MEMBERSHIP_EVENT_PROMPT_LIMIT = 6; export const VOICE_MEMBERSHIP_EVENT_FRESH_MS = 60_000; export const VOICE_CHANNEL_EFFECT_EVENT_MAX_TRACKED = 18; export const VOICE_CHANNEL_EFFECT_EVENT_PROMPT_LIMIT = 6; export const VOICE_CHANNEL_EFFECT_EVENT_FRESH_MS = 60_000; export const RECENT_ENGAGEMENT_WINDOW_MS = 35_000; export const INTERRUPTED_REALTIME_OUTPUT_IGNORE_TTL_MS = 30_000; export const SOUNDBOARD_DECISION_TRANSCRIPT_MAX_CHARS = 280; export const SOUNDBOARD_CATALOG_REFRESH_MS = 60_000; // Mean logprob threshold for ASR transcript confidence gate. // OpenAI logprobs are log-base-e: -1.0 ≈ 37% per-token confidence. // Hallucinations on noise/breathing typically score well below -2.0. export const VOICE_ASR_LOGPROB_CONFIDENCE_THRESHOLD = -1.0; export const MEMORY_SENSITIVE_PATTERN_RE = /\b(?:sk-[a-z0-9]{20,}|api[_-]?key|token|password|passphrase|authorization|secret)\b/i;