@vellumai/assistant 0.4.54 → 0.4.56

package/src/memory/conversation-crud.ts

@@ -6,7 +6,7 @@ import type { ChannelId, InterfaceId } from "../channels/types.js";
 import { parseChannelId, parseInterfaceId } from "../channels/types.js";
 import { CHANNEL_IDS, INTERFACE_IDS, isChannelId } from "../channels/types.js";
 import { getConfig } from "../config/loader.js";
-import type { TrustContext } from "../daemon/session-runtime-assembly.js";
+import type { TrustContext } from "../daemon/conversation-runtime-assembly.js";
 import { getLogger } from "../util/logger.js";
 import { createRowMapper } from "../util/row-mapper.js";
 import { deleteOrphanAttachments } from "./attachments-store.js";
@@ -55,7 +55,7 @@ export const messageMetadataSchema = z
      * Trust class of the actor at the time this message was persisted.
      * This is a durable snapshot -- it does NOT change if the actor's
      * trust status changes later. Used by the memory write gate (indexer)
-     * and read gate (session history loading) to enforce trust-aware access.
+     * and read gate (conversation history loading) to enforce trust-aware access.
      */
     provenanceTrustClass: z
       .enum(["guardian", "trusted_contact", "unknown"])
@@ -63,6 +63,7 @@ export const messageMetadataSchema = z
     provenanceSourceChannel: channelIdSchema.optional(),
     provenanceGuardianExternalUserId: z.string().optional(),
     provenanceRequesterIdentifier: z.string().optional(),
+    automated: z.boolean().optional(),
   })
   .passthrough();
 
@@ -97,7 +98,7 @@ export interface ConversationRow {
   contextSummary: string | null;
   contextCompactedMessageCount: number;
   contextCompactedAt: number | null;
-  threadType: string;
+  conversationType: string;
   source: string;
   memoryScopeId: string;
   originChannel: string | null;
@@ -120,7 +121,7 @@ export const parseConversation = createRowMapper<
   contextSummary: "contextSummary",
   contextCompactedMessageCount: "contextCompactedMessageCount",
   contextCompactedAt: "contextCompactedAt",
-  threadType: "threadType",
+  conversationType: "conversationType",
   source: "source",
   memoryScopeId: "memoryScopeId",
   originChannel: "originChannel",
@@ -169,7 +170,7 @@ export function createConversation(
     | string
     | {
         title?: string;
-        threadType?: "standard" | "private" | "background";
+        conversationType?: "standard" | "private" | "background";
         source?: string;
         scheduleJobId?: string;
       },
@@ -180,10 +181,11 @@ export function createConversation(
     typeof titleOrOpts === "string"
       ? { title: titleOrOpts }
       : (titleOrOpts ?? {});
-  const threadType = opts.threadType ?? "standard";
+  const conversationType = opts.conversationType ?? "standard";
   const source = opts.source ?? "user";
   const id = uuid();
-  const memoryScopeId = threadType === "private" ? `private:${id}` : "default";
+  const memoryScopeId =
+    conversationType === "private" ? `private:${id}` : "default";
   const conversation = {
     id,
     title: opts.title ?? null,
@@ -195,7 +197,7 @@ export function createConversation(
     contextSummary: null as string | null,
     contextCompactedMessageCount: 0,
     contextCompactedAt: null as number | null,
-    threadType,
+    conversationType,
     source,
     memoryScopeId,
     scheduleJobId: opts.scheduleJobId ?? null,
@@ -240,11 +242,11 @@ export function getConversation(id: string): ConversationRow | null {
   return row ? parseConversation(row) : null;
 }
 
-export function getConversationThreadType(
+export function getConversationType(
   conversationId: string,
 ): "standard" | "private" {
   const conv = getConversation(conversationId);
-  const raw = conv?.threadType;
+  const raw = conv?.conversationType;
   return raw === "private" ? "private" : "standard";
 }
 
@@ -254,22 +256,143 @@ export function getConversationMemoryScopeId(conversationId: string): string {
 }
 
 /**
- * Delete a conversation and all its messages.
- * Used for ephemeral conversations (e.g. secret-redirect placeholders)
- * that should not persist in session history.
+ * Delete a conversation and all its messages, cleaning up orphaned memory
+ * artifacts (items, embeddings). Returns segment and orphaned item IDs so
+ * callers can clean up the corresponding Qdrant vector entries.
  */
-export function deleteConversation(id: string): void {
+export function deleteConversation(id: string): DeletedMemoryIds {
   const db = getDb();
+  const result: DeletedMemoryIds = { segmentIds: [], orphanedItemIds: [] };
+
   db.transaction((tx) => {
-    tx.delete(llmRequestLogs)
-      .where(eq(llmRequestLogs.conversationId, id))
-      .run();
-    tx.delete(toolInvocations)
-      .where(eq(toolInvocations.conversationId, id))
-      .run();
-    tx.delete(messages).where(eq(messages.conversationId, id)).run();
+    // Collect all message IDs for this conversation.
+    const messageRows = tx
+      .select({ id: messages.id })
+      .from(messages)
+      .where(eq(messages.conversationId, id))
+      .all();
+    const messageIds = messageRows.map((r) => r.id);
+
+    if (messageIds.length > 0) {
+      // Collect memory segment IDs linked to these messages before cascade.
+      const linkedSegments = tx
+        .select({ id: memorySegments.id })
+        .from(memorySegments)
+        .where(inArray(memorySegments.messageId, messageIds))
+        .all();
+      result.segmentIds = linkedSegments.map((r) => r.id);
+
+      // Collect memory item IDs linked to these messages before cascade.
+      const linkedItems = tx
+        .select({ memoryItemId: memoryItemSources.memoryItemId })
+        .from(memoryItemSources)
+        .where(inArray(memoryItemSources.messageId, messageIds))
+        .all();
+      const candidateItemIds = [
+        ...new Set(linkedItems.map((r) => r.memoryItemId)),
+      ];
+
+      // Delete non-cascading tables first.
+      tx.delete(llmRequestLogs)
+        .where(eq(llmRequestLogs.conversationId, id))
+        .run();
+      tx.delete(toolInvocations)
+        .where(eq(toolInvocations.conversationId, id))
+        .run();
+      // Cascade deletes memory_segments, memory_item_sources, message_attachments.
+      tx.delete(messages).where(eq(messages.conversationId, id)).run();
+
+      // Clean up segment embeddings.
+      if (result.segmentIds.length > 0) {
+        tx.delete(memoryEmbeddings)
+          .where(
+            and(
+              eq(memoryEmbeddings.targetType, "segment"),
+              inArray(memoryEmbeddings.targetId, result.segmentIds),
+            ),
+          )
+          .run();
+      }
+
+      // Clean up orphaned memory items whose only sources were in this conversation.
+      if (candidateItemIds.length > 0) {
+        const surviving = tx
+          .select({ memoryItemId: memoryItemSources.memoryItemId })
+          .from(memoryItemSources)
+          .where(inArray(memoryItemSources.memoryItemId, candidateItemIds))
+          .all();
+        const survivingIds = new Set(surviving.map((r) => r.memoryItemId));
+        const orphanedIds = candidateItemIds.filter(
+          (itemId) => !survivingIds.has(itemId),
+        );
+        result.orphanedItemIds = orphanedIds;
+
+        if (orphanedIds.length > 0) {
+          tx.delete(memoryEmbeddings)
+            .where(
+              and(
+                eq(memoryEmbeddings.targetType, "item"),
+                inArray(memoryEmbeddings.targetId, orphanedIds),
+              ),
+            )
+            .run();
+          tx.delete(memoryItems)
+            .where(inArray(memoryItems.id, orphanedIds))
+            .run();
+        }
+      }
+    } else {
+      // No messages — just clean up non-message tables.
+      tx.delete(llmRequestLogs)
+        .where(eq(llmRequestLogs.conversationId, id))
+        .run();
+      tx.delete(toolInvocations)
+        .where(eq(toolInvocations.conversationId, id))
+        .run();
+    }
+
     tx.delete(conversations).where(eq(conversations.id, id)).run();
   });
+
+  return result;
+}
+
+/**
+ * Delete all private (temporary) conversations and their associated data.
+ * Called at daemon startup to clean up ephemeral conversations from previous sessions.
+ * Returns the count and aggregated deleted memory IDs for Qdrant cleanup.
+ */
+export function purgePrivateConversations(): {
+  count: number;
+  deletedMemory: DeletedMemoryIds;
+} {
+  const db = getDb();
+  const privateConvs = db
+    .select({ id: conversations.id })
+    .from(conversations)
+    .where(eq(conversations.conversationType, "private"))
+    .all();
+
+  if (privateConvs.length === 0) {
+    return { count: 0, deletedMemory: { segmentIds: [], orphanedItemIds: [] } };
+  }
+
+  const allSegmentIds: string[] = [];
+  const allOrphanedItemIds: string[] = [];
+
+  for (const conv of privateConvs) {
+    const deleted = deleteConversation(conv.id);
+    allSegmentIds.push(...deleted.segmentIds);
+    allOrphanedItemIds.push(...deleted.orphanedItemIds);
+  }
+
+  return {
+    count: privateConvs.length,
+    deletedMemory: {
+      segmentIds: allSegmentIds,
+      orphanedItemIds: allOrphanedItemIds,
+    },
+  };
 }
 
 export async function addMessage(
@@ -369,6 +492,7 @@ export async function addMessage(
       const provenanceTrustClass = parsed?.success
         ? parsed.data.provenanceTrustClass
         : undefined;
+      const automated = parsed?.success ? parsed.data.automated : undefined;
       await indexMessageNow(
         {
           messageId: message.id,
@@ -378,6 +502,7 @@ export async function addMessage(
           createdAt: message.createdAt,
           scopeId,
           provenanceTrustClass,
+          automated,
         },
         config.memory,
       );
@@ -842,7 +967,7 @@ export function getConversationOriginInterface(
  *
  * Used by the pointer message trust resolver to detect conversations
  * whose audience is a guardian or trusted_contact (even if the
- * conversation itself isn't a desktop-origin private thread).
+ * conversation itself isn't a desktop-origin private conversation).
  */
 export function getConversationRecentProvenanceTrustClass(
   conversationId: string,

package/src/memory/conversation-key-store.ts

@@ -2,7 +2,7 @@
  * Maps conversation keys to internal conversation IDs.
  *
  * The web UI identifies conversations by an opaque `conversationKey` (e.g.
- * a user ID, a channel thread ID).  This store resolves those keys to the
+ * a user ID, a channel chat ID).  This store resolves those keys to the
  * daemon's internal conversation IDs, creating new conversations on
  * first contact.
  */
@@ -93,6 +93,34 @@ export function setConversationKeyIfAbsent(
     .run();
 }
 
+/**
+ * Resolve a value that may be either a conversation ID or a conversation key
+ * to the daemon's internal conversation ID.
+ *
+ * Returns the internal conversation ID, or `null` if neither lookup succeeds.
+ * Useful for endpoints (regenerate, undo, seen) that receive IDs from clients
+ * which may be conversation keys rather than internal IDs.
+ */
+export function resolveConversationId(idOrKey: string): string | null {
+  const db = getDb();
+
+  // Fast path: check if it's already a valid conversation ID.
+  const direct = db
+    .select({ id: conversations.id })
+    .from(conversations)
+    .where(eq(conversations.id, idOrKey))
+    .get();
+  if (direct) return direct.id;
+
+  // Slow path: check if it's a conversation key.
+  const mapping = db
+    .select()
+    .from(conversationKeys)
+    .where(eq(conversationKeys.conversationKey, idOrKey))
+    .get();
+  return mapping?.conversationId ?? null;
+}
+
 /**
  * Get or create a conversation for the given conversationKey.
  *
@@ -100,11 +128,15 @@ export function setConversationKeyIfAbsent(
  * Otherwise, creates a new conversation and mapping atomically within a
  * single transaction to prevent race conditions and orphaned rows.
  */
-export function getOrCreateConversation(conversationKey: string): {
+export function getOrCreateConversation(
+  conversationKey: string,
+  opts?: { conversationType?: "standard" | "private" },
+): {
   conversationId: string;
   created: boolean;
 } {
   const db = getDb();
+  const conversationType = opts?.conversationType ?? "standard";
 
   return db.transaction((tx) => {
     const existing = tx
@@ -118,8 +150,8 @@ export function getOrCreateConversation(conversationKey: string): {
     }
 
     // Check if the conversationKey itself is an existing conversation ID.
-    // This happens when the client loads a thread from the conversations list
-    // and uses the server's conversationId as its local sessionId / conversationKey.
+    // This happens when the client loads a conversation from the conversations list
+    // and uses the server's conversationId as its local conversationKey.
     const existingConversation = tx
       .select({ id: conversations.id })
       .from(conversations)
@@ -140,6 +172,8 @@ export function getOrCreateConversation(conversationKey: string): {
 
     const now = Date.now();
     const conversationId = uuid();
+    const memoryScopeId =
+      conversationType === "private" ? `private:${conversationId}` : "default";
 
     tx.insert(conversations)
       .values({
@@ -153,6 +187,8 @@ export function getOrCreateConversation(conversationKey: string): {
         contextSummary: null,
         contextCompactedMessageCount: 0,
         contextCompactedAt: null,
+        conversationType,
+        memoryScopeId,
       })
       .run();
 

package/src/memory/conversation-queries.ts

@@ -33,7 +33,7 @@ export function listConversations(
   const db = getDb();
   const where = includeBackground
     ? undefined
-    : sql`${conversations.threadType} != 'background'`;
+    : sql`${conversations.conversationType} NOT IN ('background', 'private')`;
   const query = db
     .select()
     .from(conversations)
@@ -48,7 +48,7 @@ export function countConversations(includeBackground = false): number {
   const db = getDb();
   const where = includeBackground
     ? undefined
-    : sql`${conversations.threadType} != 'background'`;
+    : sql`${conversations.conversationType} NOT IN ('background', 'private')`;
   const [{ total }] = db
     .select({ total: count() })
     .from(conversations)
@@ -62,7 +62,9 @@ export function getLatestConversation(): ConversationRow | null {
   const row = db
     .select()
     .from(conversations)
-    .where(sql`${conversations.threadType} != 'background'`)
+    .where(
+      sql`${conversations.conversationType} NOT IN ('background', 'private')`,
+    )
     .orderBy(desc(conversations.updatedAt))
     .limit(1)
     .get();
@@ -164,7 +166,7 @@ export function searchConversations(
         FROM messages_fts f
         JOIN messages m ON m.id = f.message_id
         JOIN conversations c ON c.id = m.conversation_id
-        WHERE messages_fts MATCH ? AND c.thread_type != 'background'
+        WHERE messages_fts MATCH ? AND c.conversation_type NOT IN ('background', 'private')
         LIMIT 1000
       `,
         ftsMatch,
@@ -189,7 +191,7 @@ export function searchConversations(
       SELECT DISTINCT m.conversation_id
       FROM messages m
       JOIN conversations c ON c.id = m.conversation_id
-      WHERE m.content LIKE ? ESCAPE '\\' AND c.thread_type != 'background'
+      WHERE m.content LIKE ? ESCAPE '\\' AND c.conversation_type NOT IN ('background', 'private')
       LIMIT 1000
     `,
       likePattern,
@@ -203,7 +205,7 @@ export function searchConversations(
     .from(conversations)
     .where(
       and(
-        sql`${conversations.threadType} != 'background'`,
+        sql`${conversations.conversationType} NOT IN ('background', 'private')`,
         sql`${conversations.title} LIKE ${titlePattern} ESCAPE '\\'`,
       ),
     )

package/src/memory/db-init.ts

@@ -41,6 +41,7 @@ import {
   migrateCallSessionMode,
   migrateCanonicalGuardianDeliveriesDestinationIndex,
   migrateCanonicalGuardianRequesterChatId,
+  migrateCapabilityCardColumns,
   migrateChannelInboundDeliveredSegments,
   migrateChannelInteractionColumns,
   migrateContactChannelsAccessFields,
@@ -49,6 +50,7 @@ import {
   migrateContactsNotesColumn,
   migrateContactsRolePrincipal,
   migrateConversationsThreadTypeIndex,
+  migrateCreateThreadStartersTable,
   migrateDropAccountsTable,
   migrateDropAssistantIdColumns,
   migrateDropConflicts,
@@ -57,6 +59,7 @@ import {
   migrateDropLegacyMemberGuardianTables,
   migrateDropLoopbackPortColumn,
   migrateDropMemorySegmentFts,
+  migrateDropOrphanedMediaTables,
   migrateDropRemindersTable,
   migrateDropUsageCompositeIndexes,
   migrateFkCascadeRebuilds,
@@ -67,6 +70,7 @@ import {
   migrateGuardianDeliveryConversationIndex,
   migrateGuardianPrincipalIdColumns,
   migrateGuardianPrincipalIdNotNull,
+  migrateGuardianTimestampsEpochMs,
   migrateGuardianVerificationPurpose,
   migrateGuardianVerificationSessions,
   migrateInviteCodeHashColumn,
@@ -79,7 +83,14 @@ import {
   migrateOAuthProvidersPingUrl,
   migrateReminderRoutingIntent,
   migrateRemindersToSchedules,
+  migrateRenameConversationTypeColumn,
+  migrateRenameFollowupsThreadIdColumn,
+  migrateRenameGmailProviderKeyToGoogle,
   migrateRenameGuardianVerificationValues,
+  migrateRenameInboxThreadStateTable,
+  migrateRenameNotificationThreadColumns,
+  migrateRenameSequenceEnrollmentsThreadIdColumn,
+  migrateRenameSequenceStepsReplyKey,
   migrateRenameVerificationSessionIdColumn,
   migrateRenameVerificationTable,
   migrateRenameVoiceToPhone,
@@ -219,7 +230,7 @@ export function initializeDb(): void {
   // 14c3. Guardian action supersession metadata (superseded_by_request_id, superseded_at) + session lookup index
   migrateGuardianActionSupersession(database);
 
-  // 14d. Index on conversations.thread_type for frequent WHERE filters
+  // 14d. Index on conversations.conversation_type for frequent WHERE filters
   migrateConversationsThreadTypeIndex(database);
 
   // 14e. Index on guardian_action_deliveries.destination_conversation_id for conversation-based lookups
@@ -250,7 +261,7 @@ export function initializeDb(): void {
   // 22. Scoped approval grants (channel-agnostic one-time-use grants)
   createScopedApprovalGrantsTable(database);
 
-  // 23. Thread decision audit columns on notification_deliveries
+  // 23. Conversation decision audit columns on notification_deliveries
   migrateNotificationDeliveryThreadDecision(database);
 
   // 24. Canonical guardian requests and deliveries (unified cross-source guardian domain)
@@ -388,6 +399,39 @@ export function initializeDb(): void {
   // 63. Drop loopback_port column from oauth_providers (moved to code-side behavior registry)
   migrateDropLoopbackPortColumn(database);
 
+  // 64. Drop orphaned media tables (CREATE TABLE removed in #16739, clean up existing databases)
+  migrateDropOrphanedMediaTables(database);
+
+  // 65. Convert guardian timestamps from ISO 8601 text to epoch ms integers
+  migrateGuardianTimestampsEpochMs(database);
+
+  // 66. Rename assistant_inbox_thread_state → assistant_inbox_conversation_state
+  migrateRenameInboxThreadStateTable(database);
+
+  // 67. Rename thread_type → conversation_type in conversations table
+  migrateRenameConversationTypeColumn(database);
+
+  // 68. Rename notification_deliveries thread columns → conversation columns
+  migrateRenameNotificationThreadColumns(database);
+
+  // 69. Rename followups.thread_id → conversation_id
+  migrateRenameFollowupsThreadIdColumn(database);
+
+  // 70. Rename sequence_enrollments.thread_id → conversation_id
+  migrateRenameSequenceEnrollmentsThreadIdColumn(database);
+
+  // 71. Rename replyToThread → replyInSameConversation in sequence steps JSON blobs
+  migrateRenameSequenceStepsReplyKey(database);
+
+  // 72. Rename integration:gmail → integration:google across OAuth tables
+  migrateRenameGmailProviderKeyToGoogle(database);
+
+  // 73. Create thread_starters table for personalized empty-thread suggestions
+  migrateCreateThreadStartersTable(database);
+
+  // 74. Add capability card columns to thread_starters + category relevance table
+  migrateCapabilityCardColumns(database);
+
   validateMigrationState(database);
 
   if (process.env.BUN_TEST === "1") {

package/src/memory/embedding-backend.ts

@@ -2,7 +2,7 @@ import { createHash } from "node:crypto";
 
 import { getOllamaBaseUrlEnv } from "../config/env.js";
 import type { AssistantConfig } from "../config/types.js";
-import { getSecureKeyAsync } from "../security/secure-keys.js";
+import { getProviderKeyAsync } from "../security/secure-keys.js";
 import { getLogger } from "../util/logger.js";
 import { GeminiEmbeddingBackend } from "./embedding-gemini.js";
 import { OllamaEmbeddingBackend } from "./embedding-ollama.js";
@@ -203,6 +203,15 @@ function getCachedOrCreate<T extends EmbeddingBackend>(
   return instance;
 }
 
+/** Check if a backend instance already exists in the cache. */
+function getCached(
+  provider: string,
+  model: string,
+  extras?: string[],
+): EmbeddingBackend | undefined {
+  return backendCache.get(cacheKey(provider, model, extras));
+}
+
 function geminiCacheExtras(config: AssistantConfig): string[] {
   const extras: string[] = [];
   if (config.memory.embeddings.geminiTaskType) {
@@ -250,7 +259,10 @@ export async function selectEmbeddingBackend(
     };
   }
   if (requested === "ollama") {
-    const ollamaKey = (await getSecureKeyAsync("ollama")) ?? undefined;
+    // Check cache first to avoid unnecessary async key fetch on cache hits
+    const cached = getCached("ollama", config.memory.embeddings.ollamaModel);
+    if (cached) return { backend: cached, reason: null };
+    const ollamaKey = (await getProviderKeyAsync("ollama")) ?? undefined;
     return {
       backend: getCachedOrCreate(
         "ollama",
@@ -286,7 +298,13 @@ export async function selectEmbeddingBackend(
           reason: null,
         };
       case "openai": {
-        const openaiKey = await getSecureKeyAsync("openai");
+        // Check cache first to avoid unnecessary async key fetch on cache hits
+        const cachedOpenai = getCached(
+          "openai",
+          config.memory.embeddings.openaiModel,
+        );
+        if (cachedOpenai) return { backend: cachedOpenai, reason: null };
+        const openaiKey = await getProviderKeyAsync("openai");
         if (!openaiKey) continue;
         return {
           backend: getCachedOrCreate(
@@ -302,7 +320,14 @@ export async function selectEmbeddingBackend(
         };
       }
       case "gemini": {
-        const geminiKey = await getSecureKeyAsync("gemini");
+        // Check cache first to avoid unnecessary async key fetch on cache hits
+        const cachedGemini = getCached(
+          "gemini",
+          config.memory.embeddings.geminiModel,
+          geminiCacheExtras(config),
+        );
+        if (cachedGemini) return { backend: cachedGemini, reason: null };
+        const geminiKey = await getProviderKeyAsync("gemini");
         if (!geminiKey) continue;
         return {
           backend: getCachedOrCreate(
@@ -323,8 +348,14 @@ export async function selectEmbeddingBackend(
         };
       }
       case "ollama": {
+        // Check cache first to avoid unnecessary async key fetch on cache hits
+        const cachedOllama = getCached(
+          "ollama",
+          config.memory.embeddings.ollamaModel,
+        );
+        if (cachedOllama) return { backend: cachedOllama, reason: null };
         if (!(await isOllamaConfigured(config))) continue;
-        const ollamaKey = (await getSecureKeyAsync("ollama")) ?? undefined;
+        const ollamaKey = (await getProviderKeyAsync("ollama")) ?? undefined;
         return {
           backend: getCachedOrCreate(
             "ollama",
@@ -539,7 +570,16 @@ async function selectFallbackBackends(
     if (provider === exclude) continue;
     switch (provider) {
       case "openai": {
-        const openaiKey = await getSecureKeyAsync("openai");
+        // Check cache first to avoid unnecessary async key fetch on cache hits
+        const cachedOpenai = getCached(
+          "openai",
+          config.memory.embeddings.openaiModel,
+        );
+        if (cachedOpenai) {
+          backends.push(cachedOpenai);
+          break;
+        }
+        const openaiKey = await getProviderKeyAsync("openai");
         if (openaiKey) {
           backends.push(
             getCachedOrCreate(
@@ -556,7 +596,17 @@ async function selectFallbackBackends(
         break;
       }
       case "gemini": {
-        const geminiKey = await getSecureKeyAsync("gemini");
+        // Check cache first to avoid unnecessary async key fetch on cache hits
+        const cachedGemini = getCached(
+          "gemini",
+          config.memory.embeddings.geminiModel,
+          geminiCacheExtras(config),
+        );
+        if (cachedGemini) {
+          backends.push(cachedGemini);
+          break;
+        }
+        const geminiKey = await getProviderKeyAsync("gemini");
         if (geminiKey) {
           backends.push(
             getCachedOrCreate(
@@ -577,9 +627,18 @@ async function selectFallbackBackends(
         }
         break;
       }
-      case "ollama":
+      case "ollama": {
+        // Check cache first to avoid unnecessary async key fetch on cache hits
+        const cachedOllama = getCached(
+          "ollama",
+          config.memory.embeddings.ollamaModel,
+        );
+        if (cachedOllama) {
+          backends.push(cachedOllama);
+          break;
+        }
         if (await isOllamaConfigured(config)) {
-          const ollamaKey = (await getSecureKeyAsync("ollama")) ?? undefined;
+          const ollamaKey = (await getProviderKeyAsync("ollama")) ?? undefined;
           backends.push(
             getCachedOrCreate(
               "ollama",
@@ -595,6 +654,7 @@ async function selectFallbackBackends(
           );
         }
         break;
+      }
     }
   }
   return backends;
@@ -626,8 +686,8 @@ export async function selectedBackendSupportsMultimodal(
 
 async function isOllamaConfigured(config: AssistantConfig): Promise<boolean> {
   return (
-    config.provider === "ollama" ||
-    Boolean(await getSecureKeyAsync("ollama")) ||
+    config.services.inference.provider === "ollama" ||
+    Boolean(await getProviderKeyAsync("ollama")) ||
     Boolean(getOllamaBaseUrlEnv())
   );
 }

package/src/memory/external-conversation-store.ts

@@ -178,7 +178,7 @@ export function getBindingByChannelChat(
 
 /**
  * Remove an external binding by channel + external chat ID.
- * Used when disconnecting a synced thread by its channel identifiers.
+ * Used when disconnecting a synced conversation by its channel identifiers.
  */
 export function deleteBindingByChannelChat(
   sourceChannel: string,