diff --git a/ios/OffgridMobile/Info.plist b/ios/OffgridMobile/Info.plist index 3e6cdd82..414d96e3 100644 --- a/ios/OffgridMobile/Info.plist +++ b/ios/OffgridMobile/Info.plist @@ -31,8 +31,6 @@ NSAllowsLocalNetworking - NSLocalNetworkUsageDescription - Off Grid scans your local network to automatically discover LLM servers such as Ollama and LM Studio. NSBonjourServices _http._tcp @@ -43,6 +41,8 @@ This app needs access to your camera to take photos and attach them to conversations. NSFaceIDUsageDescription This app may use Face ID to protect access to your stored credentials. + NSLocalNetworkUsageDescription + Off Grid scans your local network to automatically discover LLM servers such as Ollama and LM Studio. NSMicrophoneUsageDescription This app needs access to your microphone for voice-to-text transcription using Whisper. NSPhotoLibraryAddUsageDescription diff --git a/ios/Podfile.lock b/ios/Podfile.lock index 8646eef3..2674daf8 100644 --- a/ios/Podfile.lock +++ b/ios/Podfile.lock @@ -10,7 +10,7 @@ PODS: - hermes-engine (0.14.0): - hermes-engine/Pre-built (= 0.14.0) - hermes-engine/Pre-built (0.14.0) - - llama-rn (0.11.4): + - llama-rn (0.12.0-rc.5): - boost - DoubleConversion - fast_float @@ -40,15 +40,15 @@ PODS: - ReactCommon/turbomodule/core - SocketRocket - Yoga - - lottie-ios (4.6.0) - - lottie-react-native (7.3.6): + - lottie-ios (4.5.0) + - lottie-react-native (7.3.5): - boost - DoubleConversion - fast_float - fmt - glog - hermes-engine - - lottie-ios (= 4.6.0) + - lottie-ios (= 4.5.0) - RCT-Folly - RCT-Folly/Fabric - RCTRequired @@ -2094,7 +2094,7 @@ PODS: - ReactCommon/turbomodule/core - SocketRocket - Yoga - - react-native-safe-area-context (5.7.0): + - react-native-safe-area-context (5.6.2): - boost - DoubleConversion - fast_float @@ -2112,8 +2112,8 @@ PODS: - React-graphics - React-ImageManager - React-jsi - - react-native-safe-area-context/common (= 5.7.0) - - react-native-safe-area-context/fabric (= 5.7.0) + - react-native-safe-area-context/common (= 5.6.2) + - react-native-safe-area-context/fabric (= 5.6.2) - React-NativeModulesApple - React-RCTFabric - React-renderercss @@ -2124,7 +2124,7 @@ PODS: - ReactCommon/turbomodule/core - SocketRocket - Yoga - - react-native-safe-area-context/common (5.7.0): + - react-native-safe-area-context/common (5.6.2): - boost - DoubleConversion - fast_float @@ -2152,7 +2152,7 @@ PODS: - ReactCommon/turbomodule/core - SocketRocket - Yoga - - react-native-safe-area-context/fabric (5.7.0): + - react-native-safe-area-context/fabric (5.6.2): - boost - DoubleConversion - fast_float @@ -2825,7 +2825,7 @@ PODS: - ReactCommon/turbomodule/core - SocketRocket - Yoga - - RNDeviceInfo (15.0.2): + - RNDeviceInfo (15.0.1): - React-Core - RNFS (2.20.0): - React-Core @@ -2913,7 +2913,7 @@ PODS: - ReactCommon/turbomodule/core - SocketRocket - Yoga - - RNReanimated (4.2.2): + - RNReanimated (4.2.1): - boost - DoubleConversion - fast_float @@ -2940,11 +2940,11 @@ PODS: - ReactCodegen - ReactCommon/turbomodule/bridging - ReactCommon/turbomodule/core - - RNReanimated/reanimated (= 4.2.2) + - RNReanimated/reanimated (= 4.2.1) - RNWorklets - SocketRocket - Yoga - - RNReanimated/reanimated (4.2.2): + - RNReanimated/reanimated (4.2.1): - boost - DoubleConversion - fast_float @@ -2971,11 +2971,11 @@ PODS: - ReactCodegen - ReactCommon/turbomodule/bridging - ReactCommon/turbomodule/core - - RNReanimated/reanimated/apple (= 4.2.2) + - RNReanimated/reanimated/apple (= 4.2.1) - RNWorklets - SocketRocket - Yoga - - RNReanimated/reanimated/apple (4.2.2): + - RNReanimated/reanimated/apple (4.2.1): - boost - DoubleConversion - fast_float @@ -3005,7 +3005,7 @@ PODS: - RNWorklets - SocketRocket - Yoga - - RNScreens (4.24.0): + - RNScreens (4.20.0): - boost - DoubleConversion - fast_float @@ -3032,10 +3032,10 @@ PODS: - ReactCodegen - ReactCommon/turbomodule/bridging - ReactCommon/turbomodule/core - - RNScreens/common (= 4.24.0) + - RNScreens/common (= 4.20.0) - SocketRocket - Yoga - - RNScreens/common (4.24.0): + - RNScreens/common (4.20.0): - boost - DoubleConversion - fast_float @@ -3149,7 +3149,7 @@ PODS: - ReactCommon/turbomodule/core - SocketRocket - Yoga - - RNWorklets (0.7.4): + - RNWorklets (0.7.3): - boost - DoubleConversion - fast_float @@ -3176,10 +3176,10 @@ PODS: - ReactCodegen - ReactCommon/turbomodule/bridging - ReactCommon/turbomodule/core - - RNWorklets/worklets (= 0.7.4) + - RNWorklets/worklets (= 0.7.3) - SocketRocket - Yoga - - RNWorklets/worklets (0.7.4): + - RNWorklets/worklets (0.7.3): - boost - DoubleConversion - fast_float @@ -3206,10 +3206,10 @@ PODS: - ReactCodegen - ReactCommon/turbomodule/bridging - ReactCommon/turbomodule/core - - RNWorklets/worklets/apple (= 0.7.4) + - RNWorklets/worklets/apple (= 0.7.3) - SocketRocket - Yoga - - RNWorklets/worklets/apple (0.7.4): + - RNWorklets/worklets/apple (0.7.3): - boost - DoubleConversion - fast_float @@ -3603,10 +3603,10 @@ SPEC CHECKSUMS: FBLazyVector: 309703e71d3f2f1ed7dc7889d58309c9d77a95a4 fmt: a40bb5bd0294ea969aaaba240a927bd33d878cdd glog: 5683914934d5b6e4240e497e0f4a3b42d1854183 - hermes-engine: 3f74bbb07573d284e764cee0131ae769e16c53b8 - llama-rn: f673d63fdd04ce96793c60ba03be52804d94f49a - lottie-ios: 8f959969761e9c45d70353667d00af0e5b9cadb3 - lottie-react-native: 983fd0489530e8d40f173de7f04e2f88b9317a15 + hermes-engine: 8c6be38f94b3bf8b864981980e64e55f08e467ec + llama-rn: 3ae5a64b3d08ff41f9e62b214ba5004e475b9561 + lottie-ios: a881093fab623c467d3bce374367755c272bdd59 + lottie-react-native: 691b8363e8c591fb78a78254ff2517258891456b op-sqlite: bafff369cecaee4fe65c89eec47deaba26f2db95 RCT-Folly: 846fda9475e61ec7bcbf8a3fe81edfcaeb090669 RCTDeprecation: a41bbdd9af30bf2e5715796b313e44ec43eefff1 @@ -3648,7 +3648,7 @@ SPEC CHECKSUMS: react-native-document-picker: dc2d83366e47e89e7c51e8a41eab99c1d54e941c react-native-document-viewer: 8c6ed07e7e27352743fa98e8dd6d288ad925b884 react-native-image-picker: 0314366753615115fa55c3cc937ac44cb7e75702 - react-native-safe-area-context: befb5404eb8a16fdc07fa2bebab3568ecabcbb8a + react-native-safe-area-context: c00143b4823773bba23f2f19f85663ae89ceb460 react-native-slider: 34064ca1a6864d7b263e44dd76a2d794e8d26744 react-native-voice: 908a0eba96c8c3d643e4f98b7232c6557d0a6f9c React-NativeModulesApple: a2c3d2cbec893956a5b3e4060322db2984fff75b @@ -3685,16 +3685,16 @@ SPEC CHECKSUMS: ReactCodegen: 3d48510bcef445f6403c0004047d4d9cbb915435 ReactCommon: ac934cb340aee91282ecd6f273a26d24d4c55cae RNCAsyncStorage: 29f0230e1a25f36c20b05f65e2eb8958d6526e82 - RNDeviceInfo: 4c852998208b60dc192ae3529e5867817719ad1e + RNDeviceInfo: 36d7f232bfe7c9b5c494cb7793230424ed32c388 RNFS: 89de7d7f4c0f6bafa05343c578f61118c8282ed8 RNGestureHandler: cd4be101cfa17ea6bbd438710caa02e286a84381 RNKeychain: a2c134ab796272c3d605e035ab727591000b30f3 RNReactNativeHapticFeedback: be4f1b4bf0398c30b59b76ed92ecb0a2ff3a69c6 - RNReanimated: 18324d3313d6477e1d12836c20c3ee30afb5de30 - RNScreens: 7f643ee0fd1407dc5085c7795460bd93da113b8f + RNReanimated: 292cd58688552a22b3fc1cefcfbc49b336dfed68 + RNScreens: 714e10b6b554f7dc7ad9f78dcf36dc8e3fc73415 RNSVG: 595abfa0f9ac26d56afcaaedf4e37a00f54cab71 RNVectorIcons: 791f13226ec4a3fd13062eda9e892159f0981fae - RNWorklets: a3184955a41f2be46898a937e2821469c8c8da42 + RNWorklets: 944dddd0eef13006b658e653abbb3ee8365c3809 RNZipArchive: 4304f5100eab004eeb7349adc51997b3a28deb76 SocketRocket: d4aabe649be1e368d1318fdf28a022d714d65748 SSZipArchive: c69881e8ac5521f0e622291387add5f60f30f3c4 diff --git a/package-lock.json b/package-lock.json index f44f33ad..da069b35 100644 --- a/package-lock.json +++ b/package-lock.json @@ -25,7 +25,7 @@ "@shopify/flash-list": "^2.2.2", "@testing-library/react-native": "^13.3.3", "@types/react-native-vector-icons": "^6.4.18", - "llama.rn": "^0.11.2", + "llama.rn": "^0.12.0-rc.5", "lottie-react-native": "^7.3.5", "moti": "^0.30.0", "patch-package": "^8.0.1", @@ -10048,9 +10048,9 @@ } }, "node_modules/llama.rn": { - "version": "0.11.2", - "resolved": "https://registry.npmjs.org/llama.rn/-/llama.rn-0.11.2.tgz", - "integrity": "sha512-LpZ1r9iLi9YdrbbpED4gB7QNhStYghdKbr1QX85UEoqRKeQhfi6yBiIwSRANLa0xMjp6yDIL4p+y3RjgyNG3GQ==", + "version": "0.12.0-rc.5", + "resolved": "https://registry.npmjs.org/llama.rn/-/llama.rn-0.12.0-rc.5.tgz", + "integrity": "sha512-92UDVtroH4hMWekgGyjxtAM4/K5NizO4kEPnhGOloXpuH67H5GWH3sZsT617afJwVwyvLTX8WtoY/M1Ke9wjNw==", "license": "MIT", "engines": { "node": ">= 16.0.0" @@ -14156,7 +14156,7 @@ "version": "5.9.3", "resolved": "https://registry.npmjs.org/typescript/-/typescript-5.9.3.tgz", "integrity": "sha512-jl1vZzPDinLr9eUt3J/t7V6FgNEw9QjvBPdysz9KfQDD41fQrC2Y4vKQdiaUpFT4bXlb1RHhLpp8wtm6M5TgSw==", - "dev": true, + "devOptional": true, "license": "Apache-2.0", "bin": { "tsc": "bin/tsc", diff --git a/package.json b/package.json index 66e24464..6830de61 100644 --- a/package.json +++ b/package.json @@ -36,7 +36,7 @@ "@shopify/flash-list": "^2.2.2", "@testing-library/react-native": "^13.3.3", "@types/react-native-vector-icons": "^6.4.18", - "llama.rn": "^0.11.2", + "llama.rn": "^0.12.0-rc.5", "lottie-react-native": "^7.3.5", "moti": "^0.30.0", "patch-package": "^8.0.1", diff --git a/src/components/ChatMessage/utils.ts b/src/components/ChatMessage/utils.ts index cebd4e0a..12acaf74 100644 --- a/src/components/ChatMessage/utils.ts +++ b/src/components/ChatMessage/utils.ts @@ -4,12 +4,36 @@ import type { ParsedContent } from './types'; /** * Parse content that may contain thinking/reasoning sections. - * Handles two formats: - * 1. HLSL.. HLSL tags (used by llama models with thinking enabled) - * 2. <|channel|>analysis<|message|>...<|channel|>final<|message|> (used by Qwen and similar models) + * Handles three formats: + * 1. ... tags (DeepSeek-style, used by llama models with thinking enabled) + * 2. <|channel>thought\n... (Gemma 4) + * 3. <|channel|>analysis<|message|>...<|channel|>final<|message|> (Qwen and similar models) */ export function parseThinkingContent(content: string): ParsedContent { - // First, check for channel-based thinking format + // Gemma 4 thinking format: <|channel>thought\n[thinking][response] + // Note asymmetric tags: <|channel> opens (with channel name 'thought'), closes. + const gemmaOpenMatch = content.match(/<\|channel>thought\n/i); + const gemmaCloseMatch = content.match(//i); + + if (gemmaOpenMatch) { + const thinkStart = gemmaOpenMatch.index! + gemmaOpenMatch[0].length; + if (gemmaCloseMatch && gemmaCloseMatch.index! >= thinkStart) { + const thinkEnd = gemmaCloseMatch.index!; + return { + thinking: content.slice(thinkStart, thinkEnd).trim(), + response: content.slice(thinkEnd + gemmaCloseMatch[0].length).trim(), + isThinkingComplete: true, + }; + } + // Still streaming — thinking not yet closed + return { + thinking: content.slice(thinkStart).trim(), + response: '', + isThinkingComplete: false, + }; + } + + // Check for channel-based thinking format // Format: <|channel|>analysis<|message|>[thinking content]<|channel|>final<|message|>[response] const channelAnalysisMatch = content.match(/<\|channel\|>analysis<\|message\|>/i); const channelFinalMatch = content.match(/<\|channel\|>final<\|message\|>/i); diff --git a/src/components/ModelCard.tsx b/src/components/ModelCard.tsx index 6c8145fc..9d332ba0 100644 --- a/src/components/ModelCard.tsx +++ b/src/components/ModelCard.tsx @@ -138,7 +138,7 @@ export const ModelCard: React.FC = ({ disabled={!onPress} testID={testID} > - + {compact ? ( = Dispatch>; const FALLBACK_RECENT_MESSAGE_COUNT = 2; @@ -224,6 +223,17 @@ async function injectRagContext(projectId: string | undefined, query: string, pr } return prompt; } +/** + * Gemma 4 requires <|think|> at the start of the system prompt to activate thinking mode. + * For E2B/E4B variants (the mobile-sized models), omitting this token fully disables thinking. + */ +function applyGemma4ThinkToken(prompt: string, isRemote: boolean): string { + if (!isRemote && llmService.isGemma4Model() && llmService.isThinkingEnabled()) { + return `<|think|>\n${prompt}`; + } + return prompt; +} + function resolveToolsAndPrompt(deps: GenerationDeps, conversation: any): { enabledTools: string[]; rawPrompt: string } { const project = conversation?.projectId ? useProjectStore.getState().getProject(conversation.projectId) : null; const { activeServerId, activeRemoteTextModelId } = useRemoteServerStore.getState(); @@ -253,11 +263,12 @@ export async function startGenerationFn(deps: GenerationDeps, call: StartGenerat const conversation = useChatStore.getState().conversations.find(c => c.id === targetConversationId); const { enabledTools, rawPrompt } = resolveToolsAndPrompt(deps, conversation); const basePrompt = await injectRagContext(conversation?.projectId, messageText, rawPrompt); - // Remote models use native tool_choice: 'auto' — skip heuristic gate and always pass enabled tools const isRemote = !!useRemoteServerStore.getState().activeRemoteTextModelId; - const heuristicMatch = shouldUseToolsForMessage(messageText, enabledTools); - const activeTools = (isRemote || heuristicMatch) ? enabledTools : []; - const systemPrompt = (!isRemote && activeTools.length > 0) ? `${basePrompt}${buildToolSystemPromptHint(activeTools)}` : basePrompt; + const activeTools = enabledTools; + const systemPrompt = applyGemma4ThinkToken( + (!isRemote && activeTools.length > 0) ? `${basePrompt}${buildToolSystemPromptHint(activeTools)}` : basePrompt, + isRemote, + ); logger.log(`[ChatGen][DEBUG] isRemote=${isRemote}, tools=[${activeTools.join(', ')}], path=${activeTools.length > 0 ? 'withTools' : 'generate'}`); const messagesForContext = buildMessagesForContext(targetConversationId, messageText, systemPrompt); await prepareContext(setDebugInfo, systemPrompt, messagesForContext); @@ -332,9 +343,12 @@ export async function regenerateResponseFn(deps: GenerationDeps, call: Regenerat .map(m => m.id === userMessage.id ? { ...m, content: messageText } : m); const { enabledTools, rawPrompt } = resolveToolsAndPrompt(deps, conversation); const isRemote = !!useRemoteServerStore.getState().activeRemoteTextModelId; - const activeTools = (isRemote || shouldUseToolsForMessage(messageText, enabledTools)) ? enabledTools : []; + const activeTools = enabledTools; const basePrompt = await injectRagContext(conversation?.projectId, messageText, rawPrompt); - const systemPrompt = (!isRemote && activeTools.length > 0) ? `${basePrompt}${buildToolSystemPromptHint(activeTools)}` : basePrompt; + const systemPrompt = applyGemma4ThinkToken( + (!isRemote && activeTools.length > 0) ? `${basePrompt}${buildToolSystemPromptHint(activeTools)}` : basePrompt, + isRemote, + ); const { prefix, filtered } = applyCompactionPrefix(conversation, systemPrompt, messagesUpToUser); try { await generateWithCompactionRetry({ id: targetConversationId, prompt: systemPrompt, messages: [...prefix, ...filtered] }, activeTools, conversation?.projectId); diff --git a/src/screens/ModelsScreen/importHelpers.ts b/src/screens/ModelsScreen/importHelpers.ts index 651417e4..7667b318 100644 --- a/src/screens/ModelsScreen/importHelpers.ts +++ b/src/screens/ModelsScreen/importHelpers.ts @@ -1,5 +1,7 @@ +import RNFS from 'react-native-fs'; +import { Alert } from 'react-native'; import { modelManager } from '../../services'; -import { showAlert, hideAlert, AlertState } from '../../components/CustomAlert'; +import { showAlert, AlertState } from '../../components/CustomAlert'; import { DownloadedModel } from '../../types'; export type GgufFileRef = { uri: string; name: string; size: number }; @@ -43,13 +45,21 @@ export async function importGgufFiles( deps: GgufImportDeps, ): Promise { const { setAlertState, setImportProgress, addDownloadedModel } = deps; + console.log('[IMPORT] importGgufFiles called with', files.length, 'file(s)'); if (files.length === 1) { + const resolvedFileName = files[0].name ?? 'unknown'; + console.log('[IMPORT] Single file import. uri:', files[0].uri, '| name:', JSON.stringify(files[0].name), '| resolvedFileName:', resolvedFileName, '| size:', files[0].size); const model = await modelManager.importLocalModel({ sourceUri: files[0].uri, - fileName: files[0].name ?? 'unknown', - onProgress: p => setImportProgress(p), + fileName: resolvedFileName, + sourceSize: files[0].size, + onProgress: p => { + console.log(`[IMPORT] Progress: ${(p.fraction * 100).toFixed(1)}% — ${p.fileName}`); + setImportProgress(p); + }, }); + console.log('[IMPORT] Single file import complete. model.name:', model.name, '| model.id:', model.id); addDownloadedModel(model); setAlertState(showAlert('Success', `${model.name} imported successfully!`)); return; @@ -57,30 +67,70 @@ export async function importGgufFiles( const file1: GgufFileRef = { uri: files[0].uri, name: files[0].name ?? '', size: files[0].size ?? 0 }; const file2: GgufFileRef = { uri: files[1].uri, name: files[1].name ?? '', size: files[1].size ?? 0 }; + console.log('[IMPORT] Two-file (vision) import.'); + console.log('[IMPORT] file1 — name:', JSON.stringify(file1.name), '| size:', file1.size, '| uri:', file1.uri); + console.log('[IMPORT] file2 — name:', JSON.stringify(file2.name), '| size:', file2.size, '| uri:', file2.uri); + console.log('[IMPORT] isMmProj(file1.name):', isMmProj(file1.name), '| isMmProj(file2.name):', isMmProj(file2.name)); + + // Check if files exist RIGHT AFTER picker returns — before any dialog + const file1ExistsBefore = await RNFS.exists(file1.uri.replace('file://', '')); + const file2ExistsBefore = await RNFS.exists(file2.uri.replace('file://', '')); + console.log('[IMPORT] FILE EXISTS CHECK (before dialog) — file1:', file1ExistsBefore, '| file2:', file2ExistsBefore); + console.log('[IMPORT] file1 decoded path:', decodeURIComponent(file1.uri.replace('file://', ''))); + console.log('[IMPORT] file2 decoded path:', decodeURIComponent(file2.uri.replace('file://', ''))); + const { mainFile, mmProjFile } = classifyGgufPair(file1, file2); + console.log('[IMPORT] Classification — mainFile:', mainFile.name, '| mmProjFile:', mmProjFile.name); + + const dialogOpenTime = Date.now(); + console.log('[IMPORT] Showing confirmation dialog at t=0ms'); const confirmed = await new Promise(resolve => { - setAlertState( - showAlert( - 'Import Vision Model?', - `Main model: ${mainFile.name}\nProjector: ${mmProjFile.name}\n\nIf these look wrong, cancel and rename your files.`, - [ - { text: 'Cancel', style: 'cancel', onPress: () => { setAlertState(hideAlert()); resolve(false); } }, - { text: 'Import', onPress: () => { setAlertState(hideAlert()); resolve(true); } }, - ], - ), + Alert.alert( + 'Import Vision Model?', + `Main model: ${mainFile.name}\nProjector: ${mmProjFile.name}\n\nIf these look wrong, cancel and rename your files.`, + [ + { text: 'Cancel', style: 'cancel', onPress: () => resolve(false) }, + { text: 'Import', onPress: () => resolve(true) }, + ], + { cancelable: false }, ); }); - if (!confirmed) return; + const dialogDurationMs = Date.now() - dialogOpenTime; + console.log('[IMPORT] Dialog closed after', dialogDurationMs, 'ms. confirmed:', confirmed); + + if (!confirmed) { + console.log('[IMPORT] User cancelled vision model import confirmation.'); + return; + } + + // Check if files STILL exist after dialog was dismissed — key check + const mainExistsAfter = await RNFS.exists(decodeURIComponent(mainFile.uri.replace('file://', ''))); + const mmProjExistsAfter = await RNFS.exists(decodeURIComponent(mmProjFile.uri.replace('file://', ''))); + console.log('[IMPORT] FILE EXISTS CHECK (after dialog, before copy) — mainFile:', mainExistsAfter, '| mmProjFile:', mmProjExistsAfter); + console.log('[IMPORT] mainFile path:', decodeURIComponent(mainFile.uri.replace('file://', ''))); + console.log('[IMPORT] mmProjFile path:', decodeURIComponent(mmProjFile.uri.replace('file://', ''))); + + if (!mainExistsAfter || !mmProjExistsAfter) { + console.log('[IMPORT] ⚠️ FILES GONE after dialog! iOS deleted temp inbox files during the', dialogDurationMs, 'ms dialog wait.'); + console.log('[IMPORT] This confirms the temp file eviction bug. Need keepLocalCopy() before dialog.'); + } + console.log('[IMPORT] Vision import confirmed. Starting importLocalModel...'); const model = await modelManager.importLocalModel({ sourceUri: mainFile.uri, fileName: mainFile.name, - onProgress: p => setImportProgress(p), + sourceSize: mainFile.size, + onProgress: p => { + console.log(`[IMPORT] Vision progress: ${(p.fraction * 100).toFixed(1)}% — ${p.fileName}`); + setImportProgress(p); + }, mmProjSourceUri: mmProjFile.uri, mmProjFileName: mmProjFile.name, + mmProjSourceSize: mmProjFile.size, }); + console.log('[IMPORT] Vision import complete. model.name:', model.name, '| isVisionModel:', model.isVisionModel, '| mmProjPath:', model.mmProjPath); addDownloadedModel(model); setAlertState(showAlert('Success', `${model.name} imported with vision projector!`)); } diff --git a/src/screens/ModelsScreen/useModelsScreen.ts b/src/screens/ModelsScreen/useModelsScreen.ts index 5f5570c8..c2d0b250 100644 --- a/src/screens/ModelsScreen/useModelsScreen.ts +++ b/src/screens/ModelsScreen/useModelsScreen.ts @@ -120,44 +120,81 @@ export function useModelsScreen() { importImageModelZip(sourceUri, fileName, { addDownloadedImageModel, activeImageModelId, setActiveImageModelId, setImportProgress, setAlertState }); const handleImportLocalModel = async () => { - if (isImporting) return; - setIsImporting(true); + console.log('[IMPORT] handleImportLocalModel called. isImporting =', isImporting); + if (isImporting) { + console.log('[IMPORT] BLOCKED — isImporting is true, returning early. This is the silent-exit bug.'); + return; + } try { + console.log('[IMPORT] Opening file picker...'); const result = await pick({ type: [types.allFiles], allowMultiSelection: true }); - if (!result || result.length === 0) return; + console.log('[IMPORT] Picker returned. File count:', result?.length ?? 0); + setIsImporting(true); + console.log('[IMPORT] setIsImporting(true) called AFTER picker returns'); + result?.forEach((f, i) => { + console.log(`[IMPORT] File[${i}] name=${JSON.stringify(f.name)} uri=${f.uri} size=${f.size} type=${f.type} nativeType=${f.nativeType} error=${f.error}`); + }); + + if (!result || result.length === 0) { + console.log('[IMPORT] Empty result from picker, returning.'); + return; + } - const allGguf = result.every(f => (f.name ?? '').toLowerCase().endsWith('.gguf')); - const singleZip = result.length === 1 && (result[0].name ?? '').toLowerCase().endsWith('.zip'); + // Resolve filename: use picker name if available, fall back to last path segment of URI + const resolvedFiles = result.map(f => ({ + ...f, + name: f.name?.trim() || decodeURIComponent(f.uri.split('/').pop() ?? '') || 'unknown', + })); + resolvedFiles.forEach((f, i) => { + console.log(`[IMPORT] File[${i}] resolvedName=${JSON.stringify(f.name)} (original=${JSON.stringify(result[i].name)})`); + }); + + const allGguf = resolvedFiles.every(f => f.name.toLowerCase().endsWith('.gguf')); + const singleZip = resolvedFiles.length === 1 && resolvedFiles[0].name.toLowerCase().endsWith('.zip'); + console.log('[IMPORT] Validation — allGguf:', allGguf, '| singleZip:', singleZip); + resolvedFiles.forEach((f, i) => { + console.log(`[IMPORT] File[${i}] nameForCheck=${JSON.stringify(f.name)} endsWithGguf=${f.name.toLowerCase().endsWith('.gguf')}`); + }); if (!allGguf && !singleZip) { + console.log('[IMPORT] VALIDATION FAILED — showing Invalid File alert'); setAlertState(showAlert( 'Invalid File', - result.length > 1 + resolvedFiles.length > 1 ? 'When selecting multiple files, all must be .gguf files (main model + mmproj projector).' : 'Supported formats: .gguf (text models) and .zip (image models).', )); return; } - if (result.length > 2) { + if (resolvedFiles.length > 2) { + console.log('[IMPORT] Too many files selected:', resolvedFiles.length); setAlertState(showAlert('Too Many Files', 'Select 1 file (text/zip) or 2 .gguf files (vision model + mmproj projector).')); return; } - const firstUri = result[0].uri; - const firstFileName = result[0].name ?? 'unknown'; + const firstUri = resolvedFiles[0].uri; + const firstFileName = resolvedFiles[0].name; + console.log('[IMPORT] firstUri:', firstUri, '| firstFileName:', firstFileName); setImportProgress({ fraction: 0, fileName: firstFileName }); if (singleZip) { + console.log('[IMPORT] Single ZIP detected, routing to handleImportImageModelZip'); await handleImportImageModelZip(firstUri, firstFileName); return; } - await importGgufFiles(result.slice(0, 2), { setAlertState, setImportProgress, addDownloadedModel }); + console.log('[IMPORT] Routing to importGgufFiles with', resolvedFiles.slice(0, 2).length, 'file(s)'); + await importGgufFiles(resolvedFiles.slice(0, 2), { setAlertState, setImportProgress, addDownloadedModel }); } catch (error: unknown) { - if (isErrorWithCode(error) && error.code === errorCodes.OPERATION_CANCELED) return; + if (isErrorWithCode(error) && error.code === errorCodes.OPERATION_CANCELED) { + console.log('[IMPORT] User cancelled the picker.'); + return; + } + console.log('[IMPORT] ERROR caught:', getErrorMessage(error), error); setAlertState(showAlert('Import Failed', getErrorMessage(error))); } finally { + console.log('[IMPORT] finally block — resetting isImporting and progress'); setIsImporting(false); setImportProgress(null); } diff --git a/src/services/llm.ts b/src/services/llm.ts index 55ab97dc..1fdcf145 100644 --- a/src/services/llm.ts +++ b/src/services/llm.ts @@ -150,14 +150,22 @@ class LLMService { supportsToolCalling(): boolean { return this.toolCallingSupported; } supportsThinking(): boolean { return this.thinkingSupported; } isThinkingEnabled(): boolean { return this.thinkingSupported && useAppStore.getState().settings.thinkingEnabled; } + isGemma4Model(): boolean { + const path = this.currentModelPath?.toLowerCase() ?? ''; + return path.includes('gemma-4') || path.includes('gemma4'); + } /** Disable ctx_shift on Android when GPU layers are active — the OpenCL backend SIGSEGVs on the ggml set op used by KV cache shifting. */ private shouldDisableCtxShift(): boolean { return Platform.OS === 'android' && this.activeGpuLayers > 0; } private detectToolCallingSupport(): void { if (!this.context) { this.toolCallingSupported = false; return; } try { const jinja = (this.context as any)?.model?.chatTemplates?.jinja; + logger.log('[LLM][TOOLS] Full jinja caps:', JSON.stringify(jinja)); + logger.log('[LLM][TOOLS] defaultCaps?.toolCalls:', jinja?.defaultCaps?.toolCalls); + logger.log('[LLM][TOOLS] toolUse:', jinja?.toolUse); + logger.log('[LLM][TOOLS] toolUseCaps?.toolCalls:', jinja?.toolUseCaps?.toolCalls); this.toolCallingSupported = !!(jinja?.defaultCaps?.toolCalls || jinja?.toolUse || jinja?.toolUseCaps?.toolCalls); - logger.log('[LLM] Tool calling supported:', this.toolCallingSupported); + logger.log('[LLM][TOOLS] toolCallingSupported =', this.toolCallingSupported); } catch (e) { logger.warn('[LLM] Error detecting tool calling support:', e); this.toolCallingSupported = false; } } private detectThinkingSupport(): void { @@ -196,11 +204,13 @@ class LLMService { const startTime = Date.now(); let firstTokenMs = 0, tokenCount = 0, firstReceived = false; let fullContent = '', fullReasoningContent = '', streamedContentSoFar = '', streamedReasoningSoFar = ''; - const completionParams = { messages: oaiMessages, ...buildCompletionParams(settings, { disableCtxShift: this.shouldDisableCtxShift() }), ...buildThinkingCompletionParams(this.isThinkingEnabled()) }; + const completionParams = { messages: oaiMessages, ...buildCompletionParams(settings, { disableCtxShift: this.shouldDisableCtxShift() }), ...buildThinkingCompletionParams(this.isThinkingEnabled(), this.isGemma4Model()) }; + logger.log(`[LLM][THINKING] thinkingSupported=${this.thinkingSupported}, thinkingEnabled=${useAppStore.getState().settings.thinkingEnabled}, isThinkingEnabled=${this.isThinkingEnabled()}, enable_thinking=${(completionParams as any).enable_thinking}, reasoning_format=${(completionParams as any).reasoning_format}`); const completionResult = await safeCompletion(ctx, () => ctx.completion(completionParams, (data: any) => { if (!this.isGenerating || !data.token) return; - if (!firstReceived) { firstReceived = true; firstTokenMs = Date.now() - startTime; } + if (!firstReceived) { firstReceived = true; firstTokenMs = Date.now() - startTime; logger.log(`[LLM][THINKING] First token raw data — token: ${JSON.stringify(data.token)}, content: ${JSON.stringify(data.content)}, reasoning_content: ${JSON.stringify(data.reasoning_content)}`); } tokenCount++; + if (data.reasoning_content) logger.log(`[LLM][THINKING] reasoning_content chunk received: ${JSON.stringify(data.reasoning_content)}`); const content = getStreamingDelta(data.content ?? (!data.reasoning_content ? data.token : undefined), streamedContentSoFar); const reasoningContent = getStreamingDelta(data.reasoning_content || undefined, streamedReasoningSoFar); if (data.content) streamedContentSoFar = data.content; @@ -214,6 +224,7 @@ class LLMService { this.performanceStats = recordGenerationStats(startTime, firstTokenMs, tokenCount); if (completionResult?.context_full) { logger.log('[LLM] Context full detected — signalling for compaction'); throw new Error('Context is full'); } const result = { content: cr?.content || cr?.text || fullContent, reasoningContent: cr?.reasoning_content || fullReasoningContent }; + logger.log(`[LLM][THINKING] Final result — hasContent=${!!result.content}, hasReasoningContent=${!!result.reasoningContent}, reasoningLength=${result.reasoningContent?.length ?? 0}, fullReasoningFromStream=${fullReasoningContent.length}`); onComplete?.(result); return result.content; })(); @@ -224,6 +235,7 @@ class LLMService { const work = generateWithToolsImpl({ context: this.context, isGenerating: this.isGenerating, isThinkingEnabled: this.isThinkingEnabled(), + isGemma4Model: this.isGemma4Model(), disableCtxShift: this.shouldDisableCtxShift(), manageContextWindow: (msgs, extra?) => this.manageContextWindow(msgs, extra), convertToOAIMessages: (msgs) => this.convertToOAIMessages(msgs), diff --git a/src/services/llmHelpers.ts b/src/services/llmHelpers.ts index ba8971e5..571ec24e 100644 --- a/src/services/llmHelpers.ts +++ b/src/services/llmHelpers.ts @@ -148,7 +148,12 @@ export async function initContextWithFallback( logger.error(`[LLM] Attempt 3/3 failed (CPU, ctx=2048): ${finalMsg}`); logger.error(`[LLM] All 3 init attempts failed for model: ${modelPath}`); logger.error(`[LLM] Error chain — GPU: "${gpuMsg}" | CPU: "${cpuMsg}" | min-ctx: "${finalMsg}"`); - throw new Error(`Failed to load model even at minimum context (2048). This may indicate insufficient memory, a corrupted model file, or an unsupported model format. (${finalMsg})`); + const errorParts = [ + gpuMsg && gpuMsg !== finalMsg ? `GPU: ${gpuMsg}` : null, + cpuMsg && cpuMsg !== finalMsg ? `CPU: ${cpuMsg}` : null, + `min-ctx: ${finalMsg}`, + ].filter(Boolean).join(' | '); + throw new Error(`Failed to load model even at minimum context (2048). This may indicate insufficient memory, a corrupted model file, or an unsupported model format.\n\nError chain: ${errorParts}`); } } } @@ -184,8 +189,10 @@ export function supportsNativeThinking(context: LlamaContext | null): boolean { return false; } } -export function buildThinkingCompletionParams(enableThinking: boolean): { enable_thinking: boolean; reasoning_format: 'none' | 'deepseek' } { - return { enable_thinking: enableThinking, reasoning_format: enableThinking ? 'deepseek' : 'none' }; +export function buildThinkingCompletionParams(enableThinking: boolean, isGemma4: boolean = false): { enable_thinking: boolean; reasoning_format: 'none' | 'deepseek' } { + // Gemma 4 uses its own <|channel>thought\n... format — not DeepSeek's tags. + // Set reasoning_format:'none' so llama.rn doesn't try to strip DeepSeek tags; we parse it ourselves. + return { enable_thinking: enableThinking, reasoning_format: (enableThinking && !isGemma4) ? 'deepseek' : 'none' }; } export function getStreamingDelta(nextValue: string | undefined, previousValue: string): string | undefined { if (!nextValue) return undefined; diff --git a/src/services/llmToolGeneration.ts b/src/services/llmToolGeneration.ts index 242f337d..a057f201 100644 --- a/src/services/llmToolGeneration.ts +++ b/src/services/llmToolGeneration.ts @@ -13,6 +13,68 @@ import logger from '../utils/logger'; type ToolStreamCallback = (data: StreamToken) => void; type ToolCompleteCallback = (fullResponse: string) => void; +/** + * Suppresses Gemma 4's native tool call tokens from the visible text stream. + * Gemma 4 wraps tool calls in <|tool_call>... — llama.rn parses + * the structured call fine, but the raw tokens still flow through data.token. + * This filter buffers the stream and drops everything inside those tags. + */ +class ToolCallTokenFilter { + private inBlock = false; + private buffer = ''; + + process(token: string): string { + this.buffer += token; + return this.flush(); + } + + private flush(): string { + const openTag = '<|tool_call>'; + const closeTag = ''; + let output = ''; + + while (this.buffer.length > 0) { + if (this.inBlock) { + const closeIdx = this.buffer.indexOf(closeTag); + if (closeIdx === -1) { + // Partial close tag may be at the end — hold it in the buffer + const partial = this.partialSuffix(this.buffer, closeTag); + this.buffer = partial > 0 ? this.buffer.slice(this.buffer.length - partial) : ''; + break; + } + // Drop everything up to and including the close tag + this.buffer = this.buffer.slice(closeIdx + closeTag.length); + this.inBlock = false; + } else { + const openIdx = this.buffer.indexOf(openTag); + if (openIdx === -1) { + const partial = this.partialSuffix(this.buffer, openTag); + if (partial > 0) { + output += this.buffer.slice(0, this.buffer.length - partial); + this.buffer = this.buffer.slice(this.buffer.length - partial); + } else { + output += this.buffer; + this.buffer = ''; + } + break; + } + output += this.buffer.slice(0, openIdx); + this.buffer = this.buffer.slice(openIdx + openTag.length); + this.inBlock = true; + } + } + + return output; + } + + private partialSuffix(text: string, tag: string): number { + for (let len = Math.min(tag.length - 1, text.length); len > 0; len--) { + if (text.endsWith(tag.slice(0, len))) return len; + } + return 0; + } +} + function parseToolCall(tc: any): ToolCall { const fn = tc.function || {}; let args = fn.arguments || {}; @@ -26,6 +88,7 @@ export interface ToolGenerationDeps { context: any; isGenerating: boolean; isThinkingEnabled: boolean; + isGemma4Model: boolean; disableCtxShift: boolean; manageContextWindow: (messages: Message[], extraReserve?: number) => Promise; convertToOAIMessages: (messages: Message[]) => any[]; @@ -57,13 +120,15 @@ export async function generateWithToolsImpl( let fullResponse = ''; let firstReceived = false; const collectedToolCalls: ToolCall[] = []; + // Gemma 4 emits <|tool_call>... tokens in the stream; filter them out. + const toolCallFilter = deps.isGemma4Model ? new ToolCallTokenFilter() : null; const completionParams = { messages: oaiMessages, ...buildCompletionParams(settings, { disableCtxShift: deps.disableCtxShift }), tools: options.tools, tool_choice: 'auto', - ...buildThinkingCompletionParams(deps.isThinkingEnabled), + ...buildThinkingCompletionParams(deps.isThinkingEnabled, deps.isGemma4Model), }; logger.log('[LLM-Tools] === INPUT ==='); logger.log(JSON.stringify(completionParams, null, 2)); @@ -77,8 +142,9 @@ export async function generateWithToolsImpl( if (!data.token) return; if (!firstReceived) { firstReceived = true; firstTokenMs = Date.now() - startTime; } tokenCount++; - fullResponse += data.token; - options.onStream?.({ content: data.token }); + const visibleToken = toolCallFilter ? toolCallFilter.process(data.token) : data.token; + fullResponse += visibleToken; + if (visibleToken) options.onStream?.({ content: visibleToken }); }), 'generateWithTools'); logger.log('[LLM-Tools] === OUTPUT ==='); logger.log(JSON.stringify(completionResult, null, 2)); diff --git a/src/services/modelManager/scan.ts b/src/services/modelManager/scan.ts index d3d26db7..68acda54 100644 --- a/src/services/modelManager/scan.ts +++ b/src/services/modelManager/scan.ts @@ -196,53 +196,96 @@ export interface ImportLocalModelOpts { sourceUri: string; fileName: string; modelsDir: string; + sourceSize?: number | null; onProgress?: (progress: { fraction: number; fileName: string }) => void; mmProjSourceUri?: string; mmProjFileName?: string; + mmProjSourceSize?: number | null; } -async function resolveAndroidUri(uri: string, cacheFileName: string): Promise<{ resolved: string; tempPath: string | null }> { - if (Platform.OS !== 'android' || !uri.startsWith('content://')) { - return { resolved: uri, tempPath: null }; +function resolveUri(uri: string): string { + // Android content:// URIs are passed directly to RNFS.copyFile — no cache copy needed. + // iOS file:// URIs need decoding (%20 → space) so RNFS can find the file on disk. + if (uri.startsWith('content://')) { + console.log('[IMPORT][scan] resolveUri — Android content:// URI, using as-is'); + console.log('[IMPORT][scan] uri:', uri); + return uri; } - const tempPath = `${RNFS.CachesDirectoryPath}/${Date.now()}_${cacheFileName}`; - await RNFS.copyFile(uri, tempPath); - return { resolved: tempPath, tempPath }; + const decoded = decodeURIComponent(uri); + console.log('[IMPORT][scan] resolveUri — file URI, decoded'); + console.log('[IMPORT][scan] original:', uri); + console.log('[IMPORT][scan] decoded: ', decoded); + return decoded; } export async function importLocalModel(opts: ImportLocalModelOpts): Promise { // NOSONAR - const { sourceUri, fileName, modelsDir, onProgress, mmProjSourceUri, mmProjFileName } = opts; + const { sourceUri, fileName, modelsDir, sourceSize, onProgress, mmProjSourceUri, mmProjFileName, mmProjSourceSize } = opts; + const importStart = Date.now(); + const elapsed = () => `+${Date.now() - importStart}ms`; + + console.log('[IMPORT][scan] ── importLocalModel START ──────────────────'); + console.log('[IMPORT][scan] Platform.OS:', Platform.OS); + console.log('[IMPORT][scan] fileName:', fileName); + console.log('[IMPORT][scan] sourceUri:', sourceUri); + console.log('[IMPORT][scan] sourceSize:', sourceSize ?? 'unknown'); + console.log('[IMPORT][scan] mmProjFileName:', mmProjFileName ?? 'none'); + console.log('[IMPORT][scan] mmProjSourceUri:', mmProjSourceUri ?? 'none'); + console.log('[IMPORT][scan] mmProjSourceSize:', mmProjSourceSize ?? 'unknown'); + console.log('[IMPORT][scan] modelsDir:', modelsDir); + + // Heartbeat — logs every 3s so we can see exactly where it gets stuck + let heartbeatStep = 'init'; + const heartbeat = setInterval(() => { + console.log(`[IMPORT][scan] ⏱ HEARTBEAT — still running at ${elapsed()}, current step: ${heartbeatStep}`); + }, 3000); if (!fileName.toLowerCase().endsWith('.gguf')) { + clearInterval(heartbeat); + console.log('[IMPORT][scan] ERROR — fileName does not end with .gguf:', fileName); throw new Error('Only .gguf files can be imported'); } - const { resolved: resolvedSource, tempPath: tempCachePath } = await resolveAndroidUri(sourceUri, fileName); - const { resolved: resolvedMmProjSource, tempPath: tempMmProjCachePath } = mmProjSourceUri && mmProjFileName - ? await resolveAndroidUri(mmProjSourceUri, mmProjFileName) - : { resolved: mmProjSourceUri, tempPath: null }; + heartbeatStep = 'resolving URIs'; + const resolvedSource = resolveUri(sourceUri); + const resolvedMmProjSource = mmProjSourceUri ? resolveUri(mmProjSourceUri) : undefined; + console.log(`[IMPORT][scan] ${elapsed()} resolvedSource:`, resolvedSource); + if (mmProjFileName) { + console.log(`[IMPORT][scan] ${elapsed()} resolvedMmProjSource:`, resolvedMmProjSource ?? 'none'); + } try { + heartbeatStep = 'checking dest paths'; const destPath = `${modelsDir}/${fileName}`; - if (await RNFS.exists(destPath)) throw new Error(`A model file named "${fileName}" already exists`); + console.log(`[IMPORT][scan] ${elapsed()} destPath:`, destPath); + const destExists = await RNFS.exists(destPath); + console.log(`[IMPORT][scan] ${elapsed()} dest already exists:`, destExists); + if (destExists) throw new Error(`A model file named "${fileName}" already exists`); if (mmProjFileName && await RNFS.exists(`${modelsDir}/${mmProjFileName}`)) { throw new Error(`A file named "${mmProjFileName}" already exists`); } // Copy main model: progress 0→0.5 when mmproj present, 0→1 otherwise const mainProgressScale = mmProjFileName ? 0.5 : 1; + heartbeatStep = 'copying main model'; + console.log(`[IMPORT][scan] ${elapsed()} Starting main model copy. sourceSize: ${sourceSize ?? 'unknown'} mainProgressScale: ${mainProgressScale}`); + console.log(`[IMPORT][scan] ${elapsed()} copy FROM:`, resolvedSource); + console.log(`[IMPORT][scan] ${elapsed()} copy TO: `, destPath); + const mainCopyStart = Date.now(); await copyFileWithProgress( resolvedSource, destPath, + sourceSize ?? null, onProgress ? (fraction) => onProgress({ fraction: fraction * mainProgressScale, fileName }) : undefined, ); + console.log(`[IMPORT][scan] ${elapsed()} Main model copy complete in ${Date.now() - mainCopyStart}ms`); const quantMatch = fileName.match(/[_-](Q\d+[_\w]*|f16|f32)/i); const quantization = quantMatch ? quantMatch[1].toUpperCase() : 'Unknown'; const modelName = fileName.replace(/\.gguf$/i, '').replace(/[_-]Q\d+.*/i, ''); const destStat = await RNFS.stat(destPath); const fileSize = parseSizeInt(destStat.size); + console.log('[IMPORT][scan] modelName:', modelName, '| quantization:', quantization, '| fileSize:', fileSize, 'bytes'); const pseudoFile: ModelFile = { name: fileName, size: fileSize, quantization, downloadUrl: '' }; const model = await buildDownloadedModel({ modelId: 'local_import', file: pseudoFile, resolvedLocalPath: destPath }); @@ -257,36 +300,61 @@ export async function importLocalModel(opts: ImportLocalModelOpts): Promise onProgress({ fraction: 0.5 + fraction * 0.5, fileName: mmProjFileName }) : undefined, ); + console.log(`[IMPORT][scan] ${elapsed()} mmproj copy complete in ${Date.now() - mmProjCopyStart}ms`); const mmProjStat = await RNFS.stat(mmProjDestPath); builtModel.mmProjPath = mmProjDestPath; builtModel.mmProjFileName = mmProjFileName; builtModel.mmProjFileSize = parseSizeInt(mmProjStat.size); builtModel.isVisionModel = true; + console.log(`[IMPORT][scan] ${elapsed()} mmproj linked. mmProjFileSize:`, builtModel.mmProjFileSize, 'bytes'); } + heartbeatStep = 'persisting metadata'; + console.log(`[IMPORT][scan] ${elapsed()} Persisting model metadata...`); await persistDownloadedModel(builtModel, modelsDir); + console.log(`[IMPORT][scan] ${elapsed()} ── importLocalModel COMPLETE. id: ${builtModel.id} ──`); return builtModel; + } catch (e) { + console.log(`[IMPORT][scan] ${elapsed()} ❌ importLocalModel ERROR:`, e); + throw e; } finally { - if (tempCachePath) await RNFS.unlink(tempCachePath).catch(() => {}); - if (tempMmProjCachePath) await RNFS.unlink(tempMmProjCachePath).catch(() => {}); + clearInterval(heartbeat); } } async function copyFileWithProgress( source: string, dest: string, + knownTotalBytes: number | null, onProgress?: (fraction: number) => void, ): Promise { - const sourceStat = await RNFS.stat(source); - const totalBytes = parseSizeInt(sourceStat.size); + const copyStart = Date.now(); + const copyElapsed = () => `+${Date.now() - copyStart}ms`; + const totalBytes = knownTotalBytes ?? 0; + const totalMB = totalBytes > 0 ? (totalBytes / 1024 / 1024).toFixed(1) : '?'; + + console.log(`[IMPORT][scan] copyFileWithProgress START — knownTotalBytes: ${knownTotalBytes ?? 'unknown'} (${totalMB} MB)`); + console.log('[IMPORT][scan] FROM:', source); + console.log('[IMPORT][scan] TO: ', dest); + if (!knownTotalBytes) { + console.log('[IMPORT][scan] No known size — progress will be indeterminate'); + } + let polling = true; + let lastWritten = 0; const pollInterval = setInterval(async () => { if (!polling) return; @@ -295,21 +363,37 @@ async function copyFileWithProgress( if (exists) { const stat = await RNFS.stat(dest); const written = parseSizeInt(stat.size); - onProgress?.(totalBytes > 0 ? Math.min(written / totalBytes, 0.99) : 0); + const writtenMB = (written / 1024 / 1024).toFixed(1); + const delta = written - lastWritten; + const speedMBs = ((delta / 1024 / 1024) / 0.5).toFixed(1); + lastWritten = written; + if (totalBytes > 0) { + const pct = Math.min(written / totalBytes, 0.99); + console.log(`[IMPORT][scan] ${copyElapsed()} copy poll — ${writtenMB}/${totalMB} MB (${(pct * 100).toFixed(1)}%) speed: ${speedMBs} MB/s`); + onProgress?.(pct); + } else { + console.log(`[IMPORT][scan] ${copyElapsed()} copy poll — ${writtenMB} MB written (size unknown) speed: ${speedMBs} MB/s`); + // No fraction available — don't call onProgress so UI stays indeterminate + } + } else { + console.log(`[IMPORT][scan] ${copyElapsed()} copy poll — dest not created yet`); } - } catch { - // File may not exist yet, ignore + } catch (e) { + console.log(`[IMPORT][scan] ${copyElapsed()} copy poll error:`, e); } }, 500); try { + console.log(`[IMPORT][scan] ${copyElapsed()} calling RNFS.copyFile...`); await RNFS.copyFile(source, dest); polling = false; clearInterval(pollInterval); + console.log(`[IMPORT][scan] ${copyElapsed()} RNFS.copyFile resolved — 100% done`); onProgress?.(1); } catch (error) { polling = false; clearInterval(pollInterval); + console.log(`[IMPORT][scan] ${copyElapsed()} RNFS.copyFile FAILED:`, error); await RNFS.unlink(dest).catch(() => {}); throw error; } diff --git a/src/utils/messageContent.ts b/src/utils/messageContent.ts index d23e77cb..52f9d92d 100644 --- a/src/utils/messageContent.ts +++ b/src/utils/messageContent.ts @@ -5,16 +5,29 @@ const CONTROL_TOKEN_PATTERNS: RegExp[] = [ /<\|eot_id\|>/gi, /<\/s>/gi, /[\s\S]*?<\/tool_call>\s*/g, + // Gemma 4 native tool call format: <|tool_call>... + // The streaming filter in llmToolGeneration suppresses these live; + // this catches any that slip through into stored message content. + /<\|tool_call>[\s\S]*?\s*/g, + // Gemma 4 string-delimiter token that may appear outside a tool block + /<\|">/g, ]; // Patterns for channel-based thinking format (used by some models like Qwen) const CHANNEL_ANALYSIS_START = /<\|channel\|>analysis<\|message\|>/gi; const CHANNEL_FINAL_START = /<\|channel\|>final<\|message\|>/gi; +// Gemma 4 thinking tags: <|channel>thought\n... +// These are stripped as a safety net; parseThinkingContent handles them before display. +const GEMMA4_THINK_OPEN = /<\|channel>thought\n/gi; +const GEMMA4_THINK_CLOSE = //gi; + export function stripControlTokens(content: string): string { let result = CONTROL_TOKEN_PATTERNS.reduce((acc, pattern) => acc.replace(pattern, ''), content); // Remove channel markers but preserve the content after them result = result.replace(CHANNEL_ANALYSIS_START, ''); result = result.replace(CHANNEL_FINAL_START, ''); + result = result.replace(GEMMA4_THINK_OPEN, ''); + result = result.replace(GEMMA4_THINK_CLOSE, ''); return result; } \ No newline at end of file