Fix direct QA reporting + configurable retention + orphan recording tracking

assistant/src/daemon/handlers/computer-use.ts

@@ -332,6 +332,24 @@ export function handleCuSessionFinalized(
     }
   }
 
+  // Create a file-backed attachment for recordings without a reporting session
+  // so cleanup can track orphan files.
+  if (msg.recording && !(meta?.reportToSessionId)) {
+    try {
+      createFileBackedAttachment({
+        filename: `qa-recording-${msg.sessionId}.mp4`,
+        mimeType: msg.recording.mimeType || 'video/mp4',
+        sizeBytes: msg.recording.sizeBytes,
+        filePath: msg.recording.localPath,
+        sha256: undefined,
+        expiresAt: msg.recording.expiresAt,
+      });
+      log.info({ sessionId: msg.sessionId }, 'Created orphan file-backed attachment for cleanup tracking (no reportToSessionId)');
+    } catch (err) {
+      log.error({ err, sessionId: msg.sessionId }, 'Failed to create file-backed attachment for orphan recording');
+    }
+  }
+
   // Clean up all CU session state.
   removeCuSessionReferences(ctx, msg.sessionId);
   // Delete metadata last — after it has been consumed for summary injection

assistant/src/daemon/handlers/misc.ts

@@ -76,6 +76,7 @@ export async function handleTaskSubmit(
       // Create CU session (reuse handleCuSessionCreate logic)
       const sessionId = uuid();
       const isQa = detectQaIntent(msg.task);
+      const config = getConfig();
       const cuMsg: CuSessionCreate = {
         type: 'cu_session_create',
         sessionId,
@@ -84,15 +85,19 @@ export async function handleTaskSubmit(
         screenHeight: msg.screenHeight,
         attachments: msg.attachments,
         interactionType: 'computer_use',
-        ...(isQa ? { qaMode: true } : {}),
+        ...(isQa ? { qaMode: true, reportToSessionId: msg.conversationId } : {}),
       };
       handleCuSessionCreate(cuMsg, socket, ctx);
 
       ctx.send(socket, {
         type: 'task_routed',
         sessionId,
         interactionType: 'computer_use',
-        ...(isQa ? { qaMode: true } : {}),
+        ...(isQa ? {
+          qaMode: true,
+          reportToSessionId: msg.conversationId,
+          retentionDays: config.qaRecording.defaultRetentionDays,
+        } : {}),
       });
     } else {
       // Create text QA session and immediately start processing

assistant/src/daemon/handlers/shared.ts

@@ -237,6 +237,7 @@ export function wireEscalationHandler(
 
     const cuSessionId = uuid();
     const isQa = detectQaIntent(task);
+    const config = getConfig();
     const cuMsg: CuSessionCreate = {
       type: 'cu_session_create',
       sessionId: cuSessionId,
@@ -256,7 +257,10 @@ export function wireEscalationHandler(
       task,
       escalatedFrom: sourceSessionId,
       reportToSessionId: sourceSessionId,
-      ...(isQa ? { qaMode: true } : {}),
+      ...(isQa ? {
+        qaMode: true,
+        retentionDays: config.qaRecording.defaultRetentionDays,
+      } : {}),
     });
 
     return true;

assistant/src/daemon/ipc-contract.ts

@@ -232,6 +232,8 @@ export interface TaskSubmit {
   screenHeight: number;
   attachments?: UserMessageAttachment[];
   source?: 'voice' | 'text';
+  /** The originating conversation/thread ID, if submitting from a chat context. */
+  conversationId?: string;
 }
 
 export interface RideShotgunStart {
@@ -1549,6 +1551,8 @@ export interface TaskRouted {
   qaMode?: boolean;
   /** The originating chat session ID for result injection. */
   reportToSessionId?: string;
+  /** Recording retention in days (from daemon config). */
+  retentionDays?: number;
 }
 
 export interface RideShotgunResult {

clients/macos/vellum-assistant/App/AppDelegate+Sessions.swift

@@ -56,7 +56,8 @@ extension AppDelegate {
                 notificationService: self.services.activityNotificationService,
                 screenRecorder: (routed.qaMode == true) ? ScreenRecorder() : nil,
                 reportToSessionId: routed.reportToSessionId,
-                qaMode: routed.qaMode ?? false
+                qaMode: routed.qaMode ?? false,
+                retentionDays: routed.retentionDays.flatMap { Int($0) } ?? 7
             )
             // Don't bind relatedViewModel for escalated sessions — the active view model
             // may be unrelated if the user switched threads. Tool calls for escalated
@@ -139,13 +140,17 @@ extension AppDelegate {
                     extractedText: $0.extractedText
                 )
             }
+            // Pass the active thread's conversation ID so the daemon can set reportToSessionId for QA sessions
+            let activeConversationId = self.mainWindow?.threadManager.activeViewModel?.sessionId
+
             do {
                 try self.daemonClient.send(TaskSubmitMessage(
                     task: effectiveTask,
                     screenWidth: Int(screenBounds.width),
                     screenHeight: Int(screenBounds.height),
                     attachments: ipcAttachments,
-                    source: submission.source
+                    source: submission.source,
+                    conversationId: activeConversationId
                 ))
             } catch {
                 log.error("Failed to send task submit message: \(error)")
@@ -199,7 +204,8 @@ extension AppDelegate {
                     notificationService: self.services.activityNotificationService,
                     screenRecorder: (routed.qaMode == true) ? ScreenRecorder() : nil,
                     reportToSessionId: routed.reportToSessionId,
-                    qaMode: routed.qaMode ?? false
+                    qaMode: routed.qaMode ?? false,
+                    retentionDays: routed.retentionDays.flatMap { Int($0) } ?? 7
                 )
                 // Don't bind relatedViewModel — sessions started via startSession() don't
                 // originate from a chat thread, so there's no ChatViewModel to extract

clients/macos/vellum-assistant/ComputerUse/Session.swift

@@ -40,6 +40,8 @@ final class ComputerUseSession: ObservableObject {
     let reportToSessionId: String?
     /// Whether this session is running in QA/test mode.
     let qaMode: Bool
+    /// Recording retention in days (from daemon config, default 7).
+    let retentionDays: Int
 
     /// Weak reference to the chat view model for extracting tool calls for notifications.
     weak var relatedViewModel: ChatViewModel?
@@ -85,7 +87,8 @@ final class ComputerUseSession: ObservableObject {
         notificationService: ActivityNotificationServiceProtocol? = nil,
         screenRecorder: ScreenRecording? = nil,
         reportToSessionId: String? = nil,
-        qaMode: Bool = false
+        qaMode: Bool = false,
+        retentionDays: Int = 7
     ) {
         self.id = sessionId ?? UUID().uuidString
         self.task = task
@@ -103,6 +106,7 @@ final class ComputerUseSession: ObservableObject {
         self.screenRecorder = screenRecorder
         self.reportToSessionId = reportToSessionId
         self.qaMode = qaMode
+        self.retentionDays = retentionDays
         self.verifier = ActionVerifier(maxSteps: maxSteps)
         self.logger = SessionLogger(task: task, attachments: attachments)
     }
@@ -1005,7 +1009,7 @@ final class ComputerUseSession: ObservableObject {
         if let recorder = screenRecorder, recorder.isRecording {
             do {
                 let result = try await recorder.stopRecording()
-                let expiresAtEpoch = Int(Date().addingTimeInterval(7 * 24 * 3600).timeIntervalSince1970 * 1000)
+                let expiresAtEpoch = Int(Date().addingTimeInterval(Double(retentionDays) * 24 * 3600).timeIntervalSince1970 * 1000)
                 recordingData = IPCCuSessionFinalizedRecording(
                     localPath: result.fileURL.path,
                     mimeType: result.mimeType,

clients/shared/IPC/Generated/IPCContractGenerated.swift

@@ -1744,6 +1744,8 @@ public struct IPCTaskRouted: Codable, Sendable {
     public let qaMode: Bool?
     /// The originating chat session ID for result injection.
     public let reportToSessionId: String?
+    /// Recording retention in days (from daemon config).
+    public let retentionDays: Double?
 }
 
 /// Server push — broadcast when a task run creates a conversation, so the client can show it as a chat thread.
@@ -1766,6 +1768,8 @@ public struct IPCTaskSubmit: Codable, Sendable {
     public let screenHeight: Int
     public let attachments: [IPCUserMessageAttachment]?
     public let source: String?
+    /// The originating conversation/thread ID, if submitting from a chat context.
+    public let conversationId: String?
 }
 
 public struct IPCTelegramConfigRequest: Codable, Sendable {

clients/shared/IPC/IPCMessages.swift

@@ -324,8 +324,8 @@ extension IPCUserMessage {
 public typealias TaskSubmitMessage = IPCTaskSubmit
 
 extension IPCTaskSubmit {
-    public init(task: String, screenWidth: Int, screenHeight: Int, attachments: [IPCAttachment]?, source: String?) {
-        self.init(type: "task_submit", task: task, screenWidth: screenWidth, screenHeight: screenHeight, attachments: attachments, source: source)
+    public init(task: String, screenWidth: Int, screenHeight: Int, attachments: [IPCAttachment]?, source: String?, conversationId: String? = nil) {
+        self.init(type: "task_submit", task: task, screenWidth: screenWidth, screenHeight: screenHeight, attachments: attachments, source: source, conversationId: conversationId)
     }
 }