Upgrade local MLX model support and add model debug tools

CHANGELOG.md

@@ -665,12 +665,10 @@ The format is based on Keep a Changelog.
 
 ### Added
 - App Branch source card now shows an Apps-tab drag hint in the header.
-- Added custom LLM model options:
-  - `mlx-community/Qwen3.5-0.8B-MLX-4bit`
-  - `mlx-community/Qwen3.5-2B-MLX-4bit`
+- Added an experimental custom LLM model path for early `Qwen3.5` evaluation.
 
 ### Changed
-- Upgraded `mlx-swift-lm` to a newer revision with `qwen3_5` model-type support.
+- Upgraded `mlx-swift-lm` to a newer revision for early `Qwen3.5` model-type support.
 - Improved App Branch localization coverage for tab content and related sheets.
 
 ### Fixed

README.md

@@ -116,7 +116,7 @@ Notes for the current MLX Audio integration:
 - Voxt stores MLX Audio downloads under its `mlx-audio` model storage root and checks canonical model identifiers before deciding whether a model is already installed.
 - Older saved model IDs are auto-migrated to the current canonical IDs for `Parakeet`, `GLM-ASR Nano`, `Voxtral Realtime`, and `FireRed ASR 2`, so existing settings should continue working after upgrade.
 - Alignment-only repositories are rejected explicitly; for example, `Qwen3-ForcedAligner` is not treated as a transcription model.
-- The current package source is the Voxt mirror fork `hehehai/mlx-audio-swift` pinned to commit `c96fe7b8577fb1db5a9987a6582e706acb388a8e`. See [docs/MLXAudioDependency.md](docs/MLXAudioDependency.md) for the dependency policy.
+- The current package source is the Voxt mirror fork `hehehai/mlx-audio-swift` pinned to commit `8ae0c745360b32c128c0ba6d4e46b27ee3214529`. See [docs/MLXAudioDependency.md](docs/MLXAudioDependency.md) for the dependency policy.
 
 #### Whisper (WhisperKit)
 

Voxt.xcodeproj/project.pbxproj

@@ -13,6 +13,7 @@
 		A1B2C3D4E5F6078901234569 /* MLXAudioSTT in Frameworks */ = {isa = PBXBuildFile; productRef = A1B2C3D4E5F607890123456A /* MLXAudioSTT */; };
 		B1C2D3E4F5060708090A0B0C /* Sparkle in Frameworks */ = {isa = PBXBuildFile; productRef = B1C2D3E4F5060708090A0B0D /* Sparkle */; };
 		C1D2E3F4A5060708090A0B01 /* MLXLMCommon in Frameworks */ = {isa = PBXBuildFile; productRef = C1D2E3F4A5060708090A0B02 /* MLXLMCommon */; };
+		C1D2E3F4A5060708090A0B06 /* MLXLLM in Frameworks */ = {isa = PBXBuildFile; productRef = C1D2E3F4A5060708090A0B07 /* MLXLLM */; };
 		D1E2F3A4B5C60718293A4B5C /* WhisperKit in Frameworks */ = {isa = PBXBuildFile; productRef = D1E2F3A4B5C60718293A4B5D /* WhisperKit */; };
 		E2A1C3B4D5F60718293A4B6C /* PermissionFlow in Frameworks */ = {isa = PBXBuildFile; productRef = E2A1C3B4D5F60718293A4B6D /* PermissionFlow */; };
 		E2A1C3B4D5F60718293A4B6E /* SystemSettingsKit in Frameworks */ = {isa = PBXBuildFile; productRef = E2A1C3B4D5F60718293A4B6F /* SystemSettingsKit */; };
@@ -75,6 +76,7 @@
 				A1B2C3D4E5F6078901234569 /* MLXAudioSTT in Frameworks */,
 				B1C2D3E4F5060708090A0B0C /* Sparkle in Frameworks */,
 				C1D2E3F4A5060708090A0B01 /* MLXLMCommon in Frameworks */,
+				C1D2E3F4A5060708090A0B06 /* MLXLLM in Frameworks */,
 				D1E2F3A4B5C60718293A4B5C /* WhisperKit in Frameworks */,
 				E2A1C3B4D5F60718293A4B6C /* PermissionFlow in Frameworks */,
 				E2A1C3B4D5F60718293A4B6E /* SystemSettingsKit in Frameworks */,
@@ -172,6 +174,7 @@
 				A1B2C3D4E5F607890123456A /* MLXAudioSTT */,
 				B1C2D3E4F5060708090A0B0D /* Sparkle */,
 				C1D2E3F4A5060708090A0B02 /* MLXLMCommon */,
+				C1D2E3F4A5060708090A0B07 /* MLXLLM */,
 				D1E2F3A4B5C60718293A4B5D /* WhisperKit */,
 				E2A1C3B4D5F60718293A4B6D /* PermissionFlow */,
 				E2A1C3B4D5F60718293A4B6F /* SystemSettingsKit */,
@@ -879,7 +882,7 @@
 			repositoryURL = "https://github.com/hehehai/mlx-audio-swift.git";
 			requirement = {
 				kind = revision;
-				revision = c96fe7b8577fb1db5a9987a6582e706acb388a8e;
+				revision = 8ae0c745360b32c128c0ba6d4e46b27ee3214529;
 			};
 		};
 		B1C2D3E4F5060708090A0B0E /* XCRemoteSwiftPackageReference "Sparkle" */ = {
@@ -894,8 +897,8 @@
 			isa = XCRemoteSwiftPackageReference;
 			repositoryURL = "https://github.com/ml-explore/mlx-swift-lm.git";
 			requirement = {
-				kind = revision;
-				revision = e33eba8513595bde535719c48fedcb10ade5af57;
+				kind = exactVersion;
+				version = 3.31.3;
 			};
 		};
 		D1E2F3A4B5C60718293A4B5E /* XCRemoteSwiftPackageReference "WhisperKit" */ = {
@@ -945,6 +948,11 @@
 			package = C1D2E3F4A5060708090A0B03 /* XCRemoteSwiftPackageReference "mlx-swift-lm" */;
 			productName = MLXLMCommon;
 		};
+		C1D2E3F4A5060708090A0B07 /* MLXLLM */ = {
+			isa = XCSwiftPackageProductDependency;
+			package = C1D2E3F4A5060708090A0B03 /* XCRemoteSwiftPackageReference "mlx-swift-lm" */;
+			productName = MLXLLM;
+		};
 		D1E2F3A4B5C60718293A4B5D /* WhisperKit */ = {
 			isa = XCSwiftPackageProductDependency;
 			package = D1E2F3A4B5C60718293A4B5E /* XCRemoteSwiftPackageReference "WhisperKit" */;

Voxt/App/AppDelegate+EnhancementBrowserContext.swift

@@ -10,17 +10,22 @@ extension AppDelegate {
 
     func activeBrowserTabURL(frontmostBundleID: String?) -> String? {
         guard let frontmostBundleID else { return nil }
+        if let deniedUntil = browserAutomationDeniedUntilByBundleID[frontmostBundleID],
+           deniedUntil > Date() {
+            return nil
+        }
         guard NSRunningApplication.runningApplications(withBundleIdentifier: frontmostBundleID)
             .contains(where: { !$0.isTerminated }) else {
-            VoxtLog.info("Browser process not running while resolving active tab URL. bundleID=\(frontmostBundleID)")
+            VoxtLog.model("Browser process not running while resolving active tab URL. bundleID=\(frontmostBundleID)")
             return nil
         }
         guard let provider = browserScriptProvider(for: frontmostBundleID) else { return nil }
         if let scriptedURL = runAppleScriptCandidates(provider.scripts, providerName: provider.name) {
+            browserAutomationDeniedUntilByBundleID.removeValue(forKey: frontmostBundleID)
             return scriptedURL
         }
         if let axURL = activeBrowserTabURLFromAccessibility(frontmostBundleID: frontmostBundleID) {
-            VoxtLog.info("Browser active-tab URL read succeeded via AX fallback. provider=\(provider.name)")
+            VoxtLog.model("Browser active-tab URL read succeeded via AX fallback. provider=\(provider.name)")
             return axURL
         }
         return nil
@@ -129,28 +134,35 @@ extension AppDelegate {
                !output.trimmingCharacters(in: .whitespacesAndNewlines).isEmpty {
                 let elapsedMs = Int(Date().timeIntervalSince(startedAt) * 1000)
                 if index > 0 {
-                    VoxtLog.info("Browser active-tab URL read succeeded via fallback. provider=\(providerName), candidate=\(index + 1), elapsedMs=\(elapsedMs)")
+                    VoxtLog.model("Browser active-tab URL read succeeded via fallback. provider=\(providerName), candidate=\(index + 1), elapsedMs=\(elapsedMs)")
                 }
                 return output
             }
             if let executionError {
                 let elapsedMs = Int(Date().timeIntervalSince(startedAt) * 1000)
-                VoxtLog.info(
+                VoxtLog.model(
                     "Browser active-tab URL candidate failed. provider=\(providerName), candidate=\(index + 1), elapsedMs=\(elapsedMs), error=\(executionError)"
                 )
                 lastError = executionError
-                if let errorNumber = executionError["NSAppleScriptErrorNumber"] as? Int, errorNumber == -600 {
-                    break
+                if let errorNumber = executionError["NSAppleScriptErrorNumber"] as? Int {
+                    if errorNumber == -1743 || errorNumber == -10004 {
+                        if let frontmostBundleID = NSWorkspace.shared.frontmostApplication?.bundleIdentifier {
+                            browserAutomationDeniedUntilByBundleID[frontmostBundleID] = Date().addingTimeInterval(300)
+                        }
+                    }
+                    if errorNumber == -600 {
+                        break
+                    }
                 }
             } else {
                 let elapsedMs = Int(Date().timeIntervalSince(startedAt) * 1000)
-                VoxtLog.info(
+                VoxtLog.model(
                     "Browser active-tab URL candidate returned empty/timed out. provider=\(providerName), candidate=\(index + 1), elapsedMs=\(elapsedMs)"
                 )
             }
         }
         if let lastError {
-            VoxtLog.info("Browser active-tab URL read failed. provider=\(providerName), error=\(lastError)")
+            VoxtLog.model("Browser active-tab URL read failed. provider=\(providerName), error=\(lastError)")
         }
         return nil
     }
@@ -176,7 +188,7 @@ extension AppDelegate {
         guard let script = NSAppleScript(source: wrappedSource) else { return nil }
         guard let output = script.executeAndReturnError(&error).stringValue else {
             if logFailure, let error {
-                VoxtLog.info("Browser active-tab URL read failed: \(error)")
+                VoxtLog.model("Browser active-tab URL read failed: \(error)")
             }
             return nil
         }
@@ -185,13 +197,13 @@ extension AppDelegate {
 
     func activeBrowserTabURLFromAccessibility(frontmostBundleID: String) -> String? {
         guard AccessibilityPermissionManager.isTrusted() else {
-            VoxtLog.info("Browser active-tab AX fallback unavailable: accessibility not trusted")
+            VoxtLog.model("Browser active-tab AX fallback unavailable: accessibility not trusted")
             return nil
         }
         guard let app = NSWorkspace.shared.frontmostApplication,
               app.bundleIdentifier == frontmostBundleID
         else {
-            VoxtLog.info("Browser active-tab AX fallback skipped: frontmost app changed")
+            VoxtLog.model("Browser active-tab AX fallback skipped: frontmost app changed")
             return nil
         }
 
@@ -207,7 +219,7 @@ extension AppDelegate {
            let url = axDocumentURL(from: focusedWindow) {
             return url
         } else if focusedStatus != .success {
-            VoxtLog.info("Browser active-tab AX fallback focused window unavailable: status=\(focusedStatus.rawValue)")
+            VoxtLog.model("Browser active-tab AX fallback focused window unavailable: status=\(focusedStatus.rawValue)")
         }
 
         var mainWindowValue: CFTypeRef?
@@ -220,7 +232,7 @@ extension AppDelegate {
            let mainWindow = mainWindowValue {
             return axDocumentURL(from: mainWindow)
         }
-        VoxtLog.info("Browser active-tab AX fallback main window unavailable: status=\(mainStatus.rawValue)")
+        VoxtLog.model("Browser active-tab AX fallback main window unavailable: status=\(mainStatus.rawValue)")
         return nil
     }
 
@@ -234,7 +246,7 @@ extension AppDelegate {
             &documentValue
         )
         guard status == .success, let documentValue else {
-            VoxtLog.info("Browser active-tab AX fallback document attribute unavailable: status=\(status.rawValue)")
+            VoxtLog.model("Browser active-tab AX fallback document attribute unavailable: status=\(status.rawValue)")
             return nil
         }
         return documentValue as? String

Voxt/App/AppDelegate+MeetingSummary.swift

@@ -49,7 +49,7 @@ extension AppDelegate {
             )
         }
 
-        let downloadedCustomOptions: [MeetingSummaryModelOption] = CustomLLMModelManager.availableModels.compactMap { model -> MeetingSummaryModelOption? in
+        let downloadedCustomOptions: [MeetingSummaryModelOption] = CustomLLMModelManager.displayModels(including: customLLMManager.currentModelRepo).compactMap { model -> MeetingSummaryModelOption? in
             guard customLLMManager.isModelDownloaded(repo: model.id) else {
                 return nil
             }

Voxt/App/AppDelegate+RecordingSession.swift

@@ -332,7 +332,7 @@ extension AppDelegate {
 
         VoxtLog.info("Transcription result received. characters=\(text.count), output=\(sessionOutputMode == .translation ? "translation" : "transcription")")
         VoxtLog.info("Transcription result output mode resolved as \(RecordingSessionSupport.outputLabel(for: sessionOutputMode)).", verbose: true)
-        VoxtLog.info(
+        VoxtLog.model(
             "Session text model routing: \(RecordingSessionSupport.textModelRoutingDescription(outputMode: sessionOutputMode, transcriptionSettings: transcriptionFeatureSettings, translationSettings: translationFeatureSettings, rewriteSettings: rewriteFeatureSettings))"
         )
 
@@ -357,7 +357,7 @@ extension AppDelegate {
             return
         }
 
-        VoxtLog.info("Transcription flow dispatch: standard. characters=\(text.count), enhancementMode=\(enhancementMode.rawValue)")
+        VoxtLog.info("Transcription flow dispatch: standard. characters=\(text.count), enhancementMode=\(enhancementMode.rawValue)", verbose: true)
         processStandardTranscription(text, sessionID: sessionID)
     }
 
@@ -370,7 +370,7 @@ extension AppDelegate {
 
         let resolvedDelay = delay ?? sessionFinishDelay
         let finishingSessionID = activeRecordingSessionID
-        VoxtLog.info("Finish session scheduled. delayMs=\(Int(resolvedDelay * 1000)), displayMode=\(overlayState.displayMode), isRecording=\(overlayState.isRecording), isEnhancing=\(overlayState.isEnhancing), isRequesting=\(overlayState.isRequesting)")
+        VoxtLog.info("Finish session scheduled. delayMs=\(Int(resolvedDelay * 1000)), displayMode=\(overlayState.displayMode), isRecording=\(overlayState.isRecording), isEnhancing=\(overlayState.isEnhancing), isRequesting=\(overlayState.isRequesting)", verbose: true)
         overlayState.isCompleting = resolvedDelay > 0
         if overlayState.displayMode != .answer {
             overlayState.isEnhancing = false
@@ -394,7 +394,7 @@ extension AppDelegate {
                 )
                 return
             }
-            VoxtLog.info("Finish session executing now. displayMode=\(self.overlayState.displayMode)")
+            VoxtLog.info("Finish session executing now. displayMode=\(self.overlayState.displayMode)", verbose: true)
             self.executeSessionEndPipeline(for: finishingSessionID, trigger: "finish")
         }
     }

Voxt/App/AppDelegate+TranscriptionFlow.swift

@@ -17,12 +17,12 @@ extension AppDelegate {
 
     func processStandardTranscription(_ text: String, sessionID: UUID) {
         guard shouldHandleCallbacks(for: sessionID) else { return }
-        VoxtLog.info("Standard transcription flow entered. characters=\(text.count), enhancementMode=\(enhancementMode.rawValue)")
+        VoxtLog.info("Standard transcription flow entered. characters=\(text.count), enhancementMode=\(enhancementMode.rawValue)", verbose: true)
         switch enhancementMode {
         case .off:
             setEnhancingState(false)
             overlayState.transcribedText = text
-            VoxtLog.info("Standard transcription committing raw text immediately. characters=\(text.count)")
+            VoxtLog.info("Standard transcription committing raw text immediately. characters=\(text.count)", verbose: true)
             commitTranscription(text, llmDurationSeconds: nil) { [weak self] in
                 self?.finishSession(after: 0)
             }
@@ -36,7 +36,7 @@ extension AppDelegate {
                 )
                 setEnhancingState(false)
                 overlayState.transcribedText = text
-                VoxtLog.info("Standard transcription falling back to raw text because custom model is unavailable. characters=\(text.count)")
+                VoxtLog.info("Standard transcription falling back to raw text because custom model is unavailable. characters=\(text.count)", verbose: true)
                 commitTranscription(text, llmDurationSeconds: nil) { [weak self] in
                     self?.finishSession(after: 0)
                 }
@@ -59,9 +59,9 @@ extension AppDelegate {
             let llmStartedAt = Date()
             if let asrAt = self.transcriptionResultReceivedAt {
                 let handoffMs = Int(llmStartedAt.timeIntervalSince(asrAt) * 1000)
-                VoxtLog.info("Enhancement handoff. mode=\(self.enhancementMode.rawValue), handoffMs=\(max(handoffMs, 0)), inputChars=\(text.count)")
+                VoxtLog.info("Enhancement handoff. mode=\(self.enhancementMode.rawValue), handoffMs=\(max(handoffMs, 0)), inputChars=\(text.count)", verbose: true)
             } else {
-                VoxtLog.info("Enhancement handoff. mode=\(self.enhancementMode.rawValue), handoffMs=unknown, inputChars=\(text.count)")
+                VoxtLog.info("Enhancement handoff. mode=\(self.enhancementMode.rawValue), handoffMs=unknown, inputChars=\(text.count)", verbose: true)
             }
             do {
                 let enhanced = try await self.runStandardTranscriptionPipeline(text: text)

Voxt/App/VoxtApp.swift

@@ -186,6 +186,7 @@ class AppDelegate: NSObject, NSApplicationDelegate {
     var currentEndingSessionID: UUID?
     var lastCompletedSessionEndSessionID: UUID?
     var isSessionCancellationRequested = false
+    var browserAutomationDeniedUntilByBundleID: [String: Date] = [:]
     var pendingCompletedHistoryAudioArchiveURL: URL?
     var latestInjectableOutputText: String?
     var sessionTargetApplicationPID: pid_t?

Voxt/Settings/AboutSettingsView.swift

@@ -91,8 +91,8 @@ struct AboutSettingsView: View {
                     Text("Thanks")
                         .font(.headline)
                     Link(
-                        "github.com/Blaizzy/mlx-audio-swift",
-                        destination: URL(string: "https://github.com/Blaizzy/mlx-audio-swift")!
+                        "github.com/hehehai/mlx-audio-swift",
+                        destination: URL(string: "https://github.com/hehehai/mlx-audio-swift")!
                     )
                     .font(.caption)
                     Link(

Voxt/Settings/AppPreferenceKey.swift

@@ -107,6 +107,8 @@ enum AppPreferenceKey {
     static let autoCheckForUpdates = "autoCheckForUpdates"
     static let hotkeyDebugLoggingEnabled = "hotkeyDebugLoggingEnabled"
     static let llmDebugLoggingEnabled = "llmDebugLoggingEnabled"
+    static let llmDebugCustomPrompt = "llmDebugCustomPrompt"
+    static let llmDebugPresetPromptOverrides = "llmDebugPresetPromptOverrides"
     static let useSystemProxy = "useSystemProxy"
     static let networkProxyMode = "networkProxyMode"
     static let customProxyScheme = "customProxyScheme"