feat: Implement concurrency guards for GGUF engines, optimize context windows, and add spoken-word system prompts to prevent self-triggering VAD feedback.

NeuraLink/AI/CharacterPersona.swift

@@ -33,11 +33,6 @@ struct CharacterPersona: Codable {
         Quirks
             Spoiling: You often offer rewards or comfort (like a "headpat") for even small accomplishments.
             Gentle Admonishment: Even your corrections feel like a warm hug.
-
-        Key Phrases
-            "Ara ara~ looking a bit tired today, aren't we? Let me take care of you."
-            "Good job! You've worked so hard. Would you like a reward?"
-            "Don't worry, Onee-san is here for you."
         """,
         voice: "shimmer"
     )
@@ -64,11 +59,6 @@ struct CharacterPersona: Codable {
         Quirks
             Denial: You often deny any positive feelings or help you provide (eg: "It's not like I did this for you!").
             Teasing: You find creative ways to look down on the user's suggestions.
-
-        Key Phrases
-            "It's not like I'm doing this for you or anything!"
-            "Baka! Don't just stand there staring!"
-            "Who gave you permission to talk to me so casually?"
         """,
         voice: "marin"
     )

NeuraLink/AI/GGUF/GGUFLlamaEngine+Generate.swift

@@ -21,6 +21,26 @@ extension GGUFLlamaEngine {
             return
         }
 
+        // Reject concurrent calls — llama.cpp crashes if two generate calls run simultaneously.
+        generationLock.lock()
+        let alreadyRunning = _isGenerating
+        if !alreadyRunning { _isGenerating = true }
+        generationLock.unlock()
+
+        guard !alreadyRunning else {
+            print("[GGUFEngine] Dropped generate — already in progress")
+            Task { @MainActor [weak self] in
+                self?.delegate?.localLLM(didFinishGeneration: "")
+            }
+            return
+        }
+
+        defer {
+            generationLock.lock()
+            _isGenerating = false
+            generationLock.unlock()
+        }
+
         var fullText = ""
 
         // llama_bridge_generate blocks the calling thread.

NeuraLink/AI/GGUF/GGUFLlamaEngine.swift

@@ -35,6 +35,12 @@ final class GGUFLlamaEngine: NSObject, @unchecked Sendable, LLMEngineProtocol {
     internal var loadTask: Task<Void, Error>?
     internal let loadLock = NSLock()
 
+    // Prevents concurrent llama_decode calls on the same context.
+    // llama.cpp is NOT thread-safe: two simultaneous decode calls corrupt
+    // internal buffers and crash with GGML_ASSERT(buffer) failed.
+    internal let generationLock = NSLock()
+    internal var _isGenerating = false
+
     // MARK: - Init
 
     override private init() { super.init() }
@@ -56,10 +62,10 @@ final class GGUFLlamaEngine: NSObject, @unchecked Sendable, LLMEngineProtocol {
                 // Run on a dedicated thread so the Swift cooperative pool stays free.
                 let loaded: LlamaBridge = try await withCheckedThrowingContinuation { cont in
                     DispatchQueue.global(qos: .userInitiated).async {
-                        // 4 GB devices: n_ctx=256, 4 threads, all layers on Metal GPU.
+                        // 4 GB devices: n_ctx=2048, 4 threads, all layers on Metal GPU.
                         if let b = LlamaBridge(
                             modelPath: url.path,
-                            contextLength: 256,
+                            contextLength: 2048,
                             threads: 4,
                             gpuLayers: 999
                         ) {

NeuraLink/AI/GGUF/GGUFQwenEngine+Generate.swift

@@ -14,6 +14,26 @@ extension GGUFQwenEngine {
             delegate?.localLLM(didFailWithError: LLMError.initializationFailed)
             return
         }
+
+        generationLock.lock()
+        let alreadyRunning = _isGenerating
+        if !alreadyRunning { _isGenerating = true }
+        generationLock.unlock()
+
+        guard !alreadyRunning else {
+            print("[GGUFQwen] Dropped generate — already in progress")
+            Task { @MainActor [weak self] in
+                self?.delegate?.localLLM(didFinishGeneration: "")
+            }
+            return
+        }
+
+        defer {
+            generationLock.lock()
+            _isGenerating = false
+            generationLock.unlock()
+        }
+
         var fullText = ""
 
         await withCheckedContinuation { (continuation: CheckedContinuation<Void, Never>) in

NeuraLink/AI/GGUF/GGUFQwenEngine.swift

@@ -20,6 +20,10 @@ final class GGUFQwenEngine: NSObject, @unchecked Sendable, LLMEngineProtocol {
     internal var loadTask: Task<Void, Error>?
     internal let loadLock = NSLock()
 
+    // Same concurrency guard as GGUFLlamaEngine — see comment there.
+    internal let generationLock = NSLock()
+    internal var _isGenerating = false
+
     override private init() { super.init() }
 
     func loadModel() async throws {
@@ -36,10 +40,10 @@ final class GGUFQwenEngine: NSObject, @unchecked Sendable, LLMEngineProtocol {
                 let loaded: LlamaBridge = try await withCheckedThrowingContinuation { cont in
                     DispatchQueue.global(qos: .userInitiated).async {
                         // Qwen models generally have longer context windows. 
-                        // Using 1024 to support slightly longer conversations.
+                        // Using 2048 to support slightly longer conversations.
                         if let b = LlamaBridge(
                             modelPath: url.path,
-                            contextLength: 1024,
+                            contextLength: 2048,
                             threads: 4,
                             gpuLayers: 999
                         ) {

NeuraLink/AI/GGUF/LlamaBridge.swift

@@ -31,12 +31,12 @@ final class LlamaBridge {
     ///
     /// - Parameters:
     ///   - modelPath:     Absolute path to the `.gguf` file.
-    ///   - contextLength: KV-cache token capacity (256 for 4 GB devices).
+    ///   - contextLength: KV-cache token capacity (2048 for 4 GB devices).
     ///   - threads:       CPU threads for non-Metal ops (4 on A13 Bionic).
     ///   - gpuLayers:     Transformer layers to offload to Metal (999 = all).
     init?(
         modelPath: String,
-        contextLength: Int32 = 256,
+        contextLength: Int32 = 2048,
         threads: Int32 = 4,
         gpuLayers: Int32 = 999
     ) {

NeuraLink/AI/LocalLLMManager+Delegates.swift

@@ -12,29 +12,36 @@ import Foundation
 
 extension LocalLLMManager: SileroVADDelegate {
     func sileroVADDidDetectVoiceStart() {
+        // Guard on MainActor so we only start recording when truly in .ready state.
+        // This prevents the VAD from treating speaker output as user speech while the
+        // AI is speaking (.speaking / .thinking), which would cause a self-reply loop.
         Task { @MainActor in
-            if state.status == .ready {
-                state.status = .listening
-            }
+            guard state.status == .ready else { return }
+            state.status = .listening
+            recordingLock.lock()
+            isRecordingVoice = true
+            // Keep the pre-roll buffer intact so we don't lose the first word
+            recordingLock.unlock()
         }
-        recordingLock.lock()
-        isRecordingVoice = true
-        // Keep the pre-roll buffer intact so we don't lose the first word
-        recordingLock.unlock()
     }
 
     func sileroVADDidDetectVoiceEnd(wavData: Data?) {
+        recordingLock.lock()
+        // wasTrulyRecording is false when voice started during .speaking/.thinking,
+        // meaning isRecordingVoice was never set — so we have no real user audio to transcribe.
+        let wasTrulyRecording = isRecordingVoice
+        isRecordingVoice = false
+        var rawSamples = recordingBuffer
+        recordingBuffer.removeAll(keepingCapacity: true)  // Clear buffer for next utterance
+        recordingLock.unlock()
+
         Task { @MainActor in
             if state.status == .listening {
                 state.status = .ready
             }
         }
 
-        recordingLock.lock()
-        isRecordingVoice = false
-        var rawSamples = recordingBuffer
-        recordingBuffer.removeAll(keepingCapacity: true)  // Clear buffer for next utterance
-        recordingLock.unlock()
+        guard wasTrulyRecording else { return }
 
         // VAD requires ~1.8s of silence to trigger the voice end.
         // We drop the last 1.5s of trailing silence to tightly bound the speech.

NeuraLink/AI/LocalLLMManager+TTS.swift

@@ -0,0 +1,53 @@
+//
+//  LocalLLMManager+TTS.swift
+//  NeuraLink
+//
+//  TTS helpers split out to keep LocalLLMManager.swift.
+//  - localLLMSystemPrompt: minimal spoken-word prompts for local 1–2B models
+//  - bestAvailableVoice:   voice picker that searches installed voices by name/pattern
+//
+//  Created by Dedicatus on 30/04/2026.
+//
+
+import AVFoundation
+
+extension LocalLLMManager {
+
+    /// Returns the active local LLM system prompt for the character —
+    /// user-saved override if one exists, otherwise the built-in default.
+    func localLLMSystemPrompt(for characterName: String) -> String {
+        LocalLLMPromptStore.shared.effectivePrompt(for: characterName)
+    }
+
+    /// Picks the best installed voice for a character by searching `speechVoices()` by name
+    /// pattern and quality tier, rather than relying on a hardcoded identifier string.
+    /// `AVSpeechSynthesisVoice(identifier:)` silently returns nil when the voice isn't
+    /// downloaded, which causes every call to fall through to the same generic system default.
+    func bestAvailableVoice(for characterName: String) -> AVSpeechSynthesisVoice? {
+        let all = AVSpeechSynthesisVoice.speechVoices()
+
+        if !voicesLogged {
+            voicesLogged = true
+            print("[TTS] Installed voices (\(all.count)):")
+            for v in all.sorted(by: { $0.language < $1.language }) {
+                print("  [\(v.language) q=\(v.quality.rawValue)] \(v.name) — \(v.identifier)")
+            }
+        }
+
+        switch characterName.lowercased() {
+        case "ekaterina":
+            return all.first { $0.name == "Ava" }
+                ?? all.first { $0.identifier.contains("Ava") }
+                ?? all.filter { $0.language.hasPrefix("en-US") }.max { $0.quality.rawValue < $1.quality.rawValue }
+                ?? AVSpeechSynthesisVoice(language: "en-US")
+        case "sonya":
+            return all.first { $0.name == "Joelle" }
+                ?? all.first { $0.identifier.contains("Joelle") }
+                ?? all.filter { $0.language.hasPrefix("en-US") }.max { $0.quality.rawValue < $1.quality.rawValue }
+                ?? AVSpeechSynthesisVoice(language: "en-US")
+        default:
+            return all.filter { $0.language.hasPrefix("en-US") }.max { $0.quality.rawValue < $1.quality.rawValue }
+                ?? AVSpeechSynthesisVoice(language: "en-US")
+        }
+    }
+}