feat: add MiniMax M3 as a cloud AI provider

Dayflow/Dayflow/Core/AI/ClaudeProvider+ActivityCards.swift

@@ -28,7 +28,7 @@ extension ClaudeProvider {
   ) async throws -> (cards: [ActivityCardData], log: LLMCall) {
     let callStart = Date()
     let prompt = buildCardsPrompt(observations: observations, context: context)
-    let modelConfiguration = Self.activityCardModelConfiguration()
+    let modelConfiguration = resolvedActivityCardModelConfiguration()
     let model = modelConfiguration.model
     let effort = modelConfiguration.reasoningEffort
     var run: ChatCLIRunResult?

Dayflow/Dayflow/Core/AI/ClaudeProvider+Text.swift

@@ -8,7 +8,7 @@ extension ClaudeProvider {
       tool: .claude,
       prompt: prompt,
       workingDirectory: config.workingDirectory,
-      model: "sonnet",
+      model: resolvedChatStreamingModel,
       reasoningEffort: nil,
       sessionId: sessionId
     )
@@ -44,9 +44,14 @@ extension ClaudeProvider {
   }
 
   func generateText(prompt: String) async throws -> (text: String, log: LLMCall) {
+    // One-off text requests (dashboard chat, ad-hoc prompts) intentionally
+    // ignore `defaultModel`: the user-selected Claude model is reserved for
+    // timeline-bound operations (transcription + card generation) so we
+    // don't accidentally let a "fast" model setting degrade richer reasoning
+    // tasks. Fall back to the same stable default as generateChatStreaming.
     try await generateText(
       prompt: prompt,
-      model: "sonnet",
+      model: resolvedChatStreamingModel,
       reasoningEffort: "high",
       disableTools: false
     )

Dayflow/Dayflow/Core/AI/ClaudeProvider+Transcription.swift

@@ -19,7 +19,7 @@ extension ClaudeProvider {
     }
 
     let callStart = Date()
-    let modelConfiguration = Self.transcriptionModelConfiguration()
+    let modelConfiguration = resolvedTranscriptionModelConfiguration()
     let model = modelConfiguration.model
     let effort = modelConfiguration.reasoningEffort
     let preparedInput: ClaudePreparedTranscriptionInput

Dayflow/Dayflow/Core/AI/ClaudeProvider.swift

@@ -5,8 +5,15 @@ final class ClaudeProvider: AgentCLISupporting {
   var cliTool: ChatCLITool { .claude }
   let runner = ChatCLIProcessRunner()
   let config = ChatCLIConfigManager.shared
+  /// When non-nil, this overrides whatever the static `*ModelConfiguration`
+  /// methods would otherwise pick as the canonical Claude model. `LLMService`
+  /// populates it from `UserDefaults` (key: `llmClaudeModel`) so the model
+  /// the provider uses stays in sync with the user's Settings selection.
+  /// When nil, providers fall back to the hardcoded default.
+  let defaultModel: String?
 
-  init() {
+  init(defaultModel: String? = nil) {
+    self.defaultModel = defaultModel
     config.ensureWorkingDirectory()
   }
 
@@ -53,6 +60,35 @@ final class ClaudeProvider: AgentCLISupporting {
     return first.adding(second)
   }
 
+  /// Resolved model for activity-card generation. Uses `defaultModel` when the
+  /// user has configured one in Settings, otherwise falls back to the
+  /// canonical Claude model returned by `activityCardModelConfiguration()`.
+  /// Reasoning effort is always taken from the static default because the
+  /// effort is a property of the Claude profile, not the model id.
+  func resolvedActivityCardModelConfiguration() -> (
+    model: String, reasoningEffort: String?
+  ) {
+    let fallback = Self.activityCardModelConfiguration()
+    return (model: defaultModel ?? fallback.model, reasoningEffort: fallback.reasoningEffort)
+  }
+
+  /// Resolved model for screenshot transcription. Same shape as
+  /// `resolvedActivityCardModelConfiguration()`; kept distinct so future per-
+  /// operation overrides (e.g. a faster transcription-only model) can be
+  /// added without touching the activity-card path.
+  func resolvedTranscriptionModelConfiguration() -> (
+    model: String, reasoningEffort: String?
+  ) {
+    let fallback = Self.transcriptionModelConfiguration()
+    return (model: defaultModel ?? fallback.model, reasoningEffort: fallback.reasoningEffort)
+  }
+
+  /// Model used for one-off `generateChatStreaming` calls (dashboard chat,
+  /// on-demand text). Falls back to a stable Claude default.
+  var resolvedChatStreamingModel: String {
+    defaultModel ?? Self.activityCardModelConfiguration().model
+  }
+
   func validateSuccessfulClaudeProcess(_ run: ChatCLIRunResult) throws {
     guard run.exitCode == 0 else {
       let message = run.stderr.trimmingCharacters(in: .whitespacesAndNewlines)

Dayflow/Dayflow/Core/AI/CodexProvider+ActivityCards.swift

@@ -28,7 +28,7 @@ extension CodexProvider {
     let basePrompt = buildCardsPrompt(observations: observations, context: context)
     var actualPromptUsed = basePrompt
 
-    let modelConfiguration = Self.activityCardModelConfiguration()
+    let modelConfiguration = resolvedActivityCardModelConfiguration()
     let legacyModelConfiguration = Self.legacyActivityCardModelConfiguration()
     var model = modelConfiguration.model
     var effort = modelConfiguration.reasoningEffort

Dayflow/Dayflow/Core/AI/CodexProvider+Text.swift

@@ -8,7 +8,7 @@ extension CodexProvider {
       tool: .codex,
       prompt: prompt,
       workingDirectory: config.workingDirectory,
-      model: "gpt-5.4",
+      model: resolvedChatStreamingModel,
       reasoningEffort: "low",
       sessionId: sessionId
     )
@@ -44,9 +44,14 @@ extension CodexProvider {
   }
 
   func generateText(prompt: String) async throws -> (text: String, log: LLMCall) {
+    // One-off text requests (dashboard chat, ad-hoc prompts) intentionally
+    // ignore `defaultModel`: the user-selected Codex model is reserved for
+    // timeline-bound operations (transcription + card generation) so we
+    // don't accidentally let a "fast" model setting degrade richer reasoning
+    // tasks. Fall back to the same stable default as generateChatStreaming.
     try await generateText(
       prompt: prompt,
-      model: "gpt-5.2",
+      model: resolvedChatStreamingModel,
       reasoningEffort: "high",
       disableTools: false
     )

Dayflow/Dayflow/Core/AI/CodexProvider+Transcription.swift

@@ -55,7 +55,7 @@ extension CodexProvider {
       )
     }
 
-    let modelConfiguration = Self.transcriptionModelConfiguration()
+    let modelConfiguration = resolvedTranscriptionModelConfiguration()
     let legacyModelConfiguration = Self.legacyTranscriptionModelConfiguration()
     var model = modelConfiguration.model
     var effort = modelConfiguration.reasoningEffort

Dayflow/Dayflow/Core/AI/CodexProvider.swift

@@ -5,12 +5,50 @@ final class CodexProvider: AgentCLISupporting {
   var cliTool: ChatCLITool { .codex }
   let runner = ChatCLIProcessRunner()
   let config = ChatCLIConfigManager.shared
+  /// When non-nil, this overrides whatever the static `*ModelConfiguration`
+  /// methods would otherwise pick as the canonical Codex model. `LLMService`
+  /// populates it from `UserDefaults` (key: `llmCodexModel`) so the model the
+  /// provider uses stays in sync with the user's Settings selection. When
+  /// nil, providers fall back to the hardcoded default.
+  let defaultModel: String?
 
-  init() {
+  init(defaultModel: String? = nil) {
+    self.defaultModel = defaultModel
     config.ensureWorkingDirectory()
   }
 }
 
+extension CodexProvider {
+  /// Resolved model for activity-card generation. Uses `defaultModel` when the
+  /// user has configured one in Settings, otherwise falls back to the
+  /// canonical Codex model returned by `activityCardModelConfiguration()`.
+  /// Reasoning effort is always taken from the static default because the
+  /// effort is a property of the Codex profile, not the model id.
+  func resolvedActivityCardModelConfiguration() -> (
+    model: String, reasoningEffort: String?
+  ) {
+    let fallback = Self.activityCardModelConfiguration()
+    return (model: defaultModel ?? fallback.model, reasoningEffort: fallback.reasoningEffort)
+  }
+
+  /// Resolved model for screenshot transcription. Same shape as
+  /// `resolvedActivityCardModelConfiguration()`; kept distinct so future per-
+  /// operation overrides (e.g. a faster transcription-only model) can be
+  /// added without touching the activity-card path.
+  func resolvedTranscriptionModelConfiguration() -> (
+    model: String, reasoningEffort: String?
+  ) {
+    let fallback = Self.transcriptionModelConfiguration()
+    return (model: defaultModel ?? fallback.model, reasoningEffort: fallback.reasoningEffort)
+  }
+
+  /// Model used for one-off `generateChatStreaming` calls (dashboard chat,
+  /// on-demand text). Falls back to a stable Codex default.
+  var resolvedChatStreamingModel: String {
+    defaultModel ?? Self.activityCardModelConfiguration().model
+  }
+}
+
 extension CodexProvider {
   static func shouldUseLegacyModel(
     after error: Error,

Dayflow/Dayflow/Core/AI/DailyRecapModels.swift

@@ -56,6 +56,8 @@ enum DailyRecapProvider: String, Codable, CaseIterable, Sendable {
       return .local
     case .openAICompatible:
       return .none
+    case .minimax:
+      return .none
     }
   }
 

Dayflow/Dayflow/Core/AI/LLMService.swift

@@ -154,10 +154,63 @@ final class LLMService: LLMServicing {
     return OllamaProvider(openAICompatible: runtimeConfiguration)
   }
 
+  private func makeMiniMaxProvider() -> MiniMaxProvider? {
+    guard let key = KeychainManager.shared.retrieve(for: MiniMaxProvider.keychainKey),
+      !key.trimmingCharacters(in: .whitespacesAndNewlines).isEmpty
+    else {
+      print("❌ [LLMService] MiniMax provider unavailable: missing API key")
+      return nil
+    }
+    return MiniMaxProvider()
+  }
+
   private func providerLabel(for providerID: LLMProviderID) -> String {
     providerID.providerLabel
   }
 
+  /// Returns the model ID the provider should be stamped with on generated
+  /// cards. Falls back to the canonical provider ID when the user hasn't
+  /// chosen a specific model yet.
+  private func providerModelId(for providerID: LLMProviderID) -> String? {
+    switch providerID {
+    case .gemini:
+      return GeminiModelPreference.load().primary.rawValue
+    case .local:
+      return UserDefaults.standard.string(forKey: "llmLocalModelId")
+    case .openAICompatible:
+      return OpenAICompatiblePreferences.load()?.modelID
+    case .chatGPT:
+      return UserDefaults.standard.string(forKey: "llmCodexModel")
+    case .claude:
+      return UserDefaults.standard.string(forKey: "llmClaudeModel")
+    case .minimax:
+      return MiniMaxModelPreference.load().modelId
+    case .dayflow:
+      return nil
+    }
+  }
+
+  /// Reads the user-selected Chat CLI model from `UserDefaults` for the
+  /// given provider and returns a trimmed, non-empty value. Returns `nil`
+  /// when the key is unset or the stored value is whitespace-only so the
+  /// provider's `defaultModel` is `nil` and the static canonical default
+  /// is used. Mirrors v1's `makeChatCLIProvider` pass-through so Settings
+  /// selections actually reach the running provider.
+  private func resolvedChatCLIDefaultModel(for providerID: LLMProviderID) -> String? {
+    let key: String
+    switch providerID {
+    case .chatGPT:
+      key = "llmCodexModel"
+    case .claude:
+      key = "llmClaudeModel"
+    default:
+      return nil
+    }
+    guard let raw = UserDefaults.standard.string(forKey: key) else { return nil }
+    let trimmed = raw.trimmingCharacters(in: .whitespacesAndNewlines)
+    return trimmed.isEmpty ? nil : trimmed
+  }
+
   private func noProviderError() -> NSError {
     NSError(
       domain: "LLMService",
@@ -246,21 +299,35 @@ final class LLMService: LLMServicing {
         ), fallbackState: nil
       )
     case .chatGPT:
-      let provider = CodexProvider()
+      let provider = CodexProvider(defaultModel: resolvedChatCLIDefaultModel(for: .chatGPT))
       return (
         actions: BatchProviderActions(
           transcribeScreenshots: provider.transcribeScreenshots,
           generateActivityCards: provider.generateActivityCards
         ), fallbackState: nil
       )
     case .claude:
-      let provider = ClaudeProvider()
+      let provider = ClaudeProvider(defaultModel: resolvedChatCLIDefaultModel(for: .claude))
       return (
         actions: BatchProviderActions(
           transcribeScreenshots: provider.transcribeScreenshots,
           generateActivityCards: provider.generateActivityCards
         ), fallbackState: nil
       )
+    case .minimax:
+      guard let provider = makeMiniMaxProvider() else { throw noProviderError() }
+      return (
+        actions: BatchProviderActions(
+          transcribeScreenshots: { [provider] screenshots, batchStartTime, batchId in
+            try await provider.transcribeScreenshots(
+              screenshots, batchStartTime: batchStartTime, batchId: batchId)
+          },
+          generateActivityCards: { [provider] observations, context, batchId in
+            try await provider.generateActivityCards(
+              observations: observations, context: context, batchId: batchId)
+          }
+        ), fallbackState: nil
+      )
     }
   }
 
@@ -540,21 +607,29 @@ final class LLMService: LLMServicing {
         generateTextStreaming: nil
       )
     case .chatGPT:
-      let provider = CodexProvider()
+      let provider = CodexProvider(defaultModel: resolvedChatCLIDefaultModel(for: .chatGPT))
       return TextProviderActions(
         generateText: { prompt in
           try await provider.generateText(prompt: prompt)
         },
         generateTextStreaming: provider.generateTextStreaming
       )
     case .claude:
-      let provider = ClaudeProvider()
+      let provider = ClaudeProvider(defaultModel: resolvedChatCLIDefaultModel(for: .claude))
       return TextProviderActions(
         generateText: { prompt in
           try await provider.generateText(prompt: prompt)
         },
         generateTextStreaming: provider.generateTextStreaming
       )
+    case .minimax:
+      guard let provider = makeMiniMaxProvider() else { throw noProviderError() }
+      return TextProviderActions(
+        generateText: { prompt in
+          try await provider.generateText(prompt: prompt)
+        },
+        generateTextStreaming: nil
+      )
     }
   }
 
@@ -853,7 +928,9 @@ final class LLMService: LLMServicing {
                 detailedSummary: card.detailedSummary,
                 distractions: card.distractions,
                 appSites: card.appSites,
-                isBackupGenerated: isBackupGenerated ? true : nil
+                isBackupGenerated: isBackupGenerated ? true : nil,
+                providerId: activeContext.id.providerLabel,
+                modelId: providerModelId(for: activeContext.id)
               )
             },
             batchId: batchId
@@ -1191,15 +1268,17 @@ final class LLMService: LLMServicing {
         history: request.history
       )
     case .codex:
-      return CodexProvider().generateChatStreaming(
-        prompt: request.prompt,
-        sessionId: request.sessionId
-      )
+      return CodexProvider(defaultModel: resolvedChatCLIDefaultModel(for: .chatGPT))
+        .generateChatStreaming(
+          prompt: request.prompt,
+          sessionId: request.sessionId
+        )
     case .claude:
-      return ClaudeProvider().generateChatStreaming(
-        prompt: request.prompt,
-        sessionId: request.sessionId
-      )
+      return ClaudeProvider(defaultModel: resolvedChatCLIDefaultModel(for: .claude))
+        .generateChatStreaming(
+          prompt: request.prompt,
+          sessionId: request.sessionId
+        )
     }
   }
 }

Dayflow/Dayflow/Core/AI/LLMTypes.swift

@@ -87,6 +87,7 @@ enum LLMProviderID: String, Codable, CaseIterable {
   case claude
   case openAICompatible = "openai_compatible"
   case local
+  case minimax
 
   var analyticsName: String {
     switch self {
@@ -100,6 +101,8 @@ enum LLMProviderID: String, Codable, CaseIterable {
       return "openai_compatible"
     case .local:
       return "ollama"
+    case .minimax:
+      return "minimax"
     }
   }
 
@@ -112,6 +115,8 @@ enum LLMProviderID: String, Codable, CaseIterable {
     case .openAICompatible: return "openai_compatible"
     case .local:
       return "local"
+    case .minimax:
+      return "minimax"
     }
   }
 }