@vellumai/assistant 0.4.41 → 0.4.43

package/src/memory/conversation-crud.ts

@@ -11,7 +11,8 @@ import { getLogger } from "../util/logger.js";
 import { createRowMapper } from "../util/row-mapper.js";
 import { deleteOrphanAttachments } from "./attachments-store.js";
 import { projectAssistantMessage } from "./conversation-attention-store.js";
-import { getDb, rawExec, rawGet } from "./db.js";
+import { ensureDisplayOrderMigration } from "./conversation-display-order-migration.js";
+import { getDb, rawExec, rawGet, rawRun } from "./db.js";
 import { indexMessageNow } from "./indexer.js";
 import {
   channelInboundEvents,
@@ -893,3 +894,54 @@ export function getConversationRecentProvenanceTrustClass(
     return undefined;
   }
 }
+
+// ---------------------------------------------------------------------------
+// CRUD functions for display_order and is_pinned
+// ---------------------------------------------------------------------------
+
+export function batchSetDisplayOrders(
+  updates: Array<{
+    id: string;
+    displayOrder: number | null;
+    isPinned: boolean;
+  }>,
+): void {
+  ensureDisplayOrderMigration();
+  rawExec("BEGIN");
+  try {
+    for (const update of updates) {
+      rawRun(
+        "UPDATE conversations SET display_order = ?, is_pinned = ? WHERE id = ?",
+        update.displayOrder,
+        update.isPinned ? 1 : 0,
+        update.id,
+      );
+    }
+    rawExec("COMMIT");
+  } catch (err) {
+    rawExec("ROLLBACK");
+    throw err;
+  }
+}
+
+export function getDisplayMetaForConversations(
+  conversationIds: string[],
+): Map<string, { displayOrder: number | null; isPinned: boolean }> {
+  ensureDisplayOrderMigration();
+  const result = new Map<
+    string,
+    { displayOrder: number | null; isPinned: boolean }
+  >();
+  if (conversationIds.length === 0) return result;
+  for (const id of conversationIds) {
+    const row = rawGet<{
+      display_order: number | null;
+      is_pinned: number | null;
+    }>("SELECT display_order, is_pinned FROM conversations WHERE id = ?", id);
+    result.set(id, {
+      displayOrder: row?.display_order ?? null,
+      isPinned: (row?.is_pinned ?? 0) === 1,
+    });
+  }
+  return result;
+}

package/src/memory/conversation-display-order-migration.ts

@@ -1,9 +1,8 @@
 /**
  * Runtime migration for display_order and is_pinned columns on the
  * conversations table. Extracted into its own module to avoid circular
- * dependencies between conversation-store.ts (which re-exports from
- * conversation-queries.ts) and conversation-queries.ts (which needs
- * the migration to run before ORDER BY display_order).
+ * dependencies between conversation-crud.ts and conversation-queries.ts
+ * (which needs the migration to run before ORDER BY display_order).
  */
 
 import { getLogger } from "../util/logger.js";

package/src/memory/conversation-queries.ts

@@ -1,4 +1,4 @@
-import { and, asc, count, desc, eq, gte, lt, ne, or, sql } from "drizzle-orm";
+import { and, asc, count, desc, eq, lt, or, sql } from "drizzle-orm";
 
 import { getLogger } from "../util/logger.js";
 import type { ConversationRow, MessageRow } from "./conversation-crud.js";
@@ -55,34 +55,6 @@ export function getLatestConversation(): ConversationRow | null {
   return row ? parseConversation(row) : null;
 }
 
-/**
- * Get the next message in a conversation after a given message.
- * Uses gte + ne(id) instead of gt on timestamp so that messages sharing the
- * same millisecond (common in legacy conversations where an assistant turn and
- * the following user tool_result are saved in the same tick) are not skipped.
- */
-export function getNextMessage(
-  conversationId: string,
-  afterTimestamp: number,
-  excludeMessageId: string,
-): MessageRow | null {
-  const db = getDb();
-  const row = db
-    .select()
-    .from(messages)
-    .where(
-      and(
-        eq(messages.conversationId, conversationId),
-        gte(messages.createdAt, afterTimestamp),
-        ne(messages.id, excludeMessageId),
-      ),
-    )
-    .orderBy(asc(messages.createdAt), asc(messages.id))
-    .limit(1)
-    .get();
-  return row ? parseMessage(row) : null;
-}
-
 export interface PaginatedMessagesResult {
   messages: MessageRow[];
   /** Whether older messages exist beyond the returned page. */

package/src/memory/conversation-title-service.ts

@@ -12,8 +12,12 @@ import { getConfiguredProvider } from "../providers/provider-send-message.js";
 import type { Provider } from "../providers/types.js";
 import { getLogger } from "../util/logger.js";
 import { truncate } from "../util/truncate.js";
-import type { MessageRow } from "./conversation-store.js";
-import * as conversationStore from "./conversation-store.js";
+import {
+  getConversation,
+  getMessages,
+  type MessageRow,
+  updateConversationTitle,
+} from "./conversation-crud.js";
 
 const log = getLogger("conversation-title-service");
 
@@ -110,7 +114,7 @@ export async function generateAndPersistConversationTitle(
   } = params;
 
   // Check current title is replaceable
-  const conversation = conversationStore.getConversation(conversationId);
+  const conversation = getConversation(conversationId);
   if (conversation && !isReplaceableTitle(conversation.title)) {
     return { title: conversation.title!, updated: false };
   }
@@ -119,7 +123,7 @@ export async function generateAndPersistConversationTitle(
   if (!provider) {
     // No provider available — fall back to context-derived title or untitled
     const fallback = deriveFallbackTitle(context) ?? UNTITLED_FALLBACK;
-    conversationStore.updateConversationTitle(conversationId, fallback, 1);
+    updateConversationTitle(conversationId, fallback, 1);
     onTitleUpdated?.(fallback);
     return { title: fallback, updated: true };
   }
@@ -136,7 +140,10 @@ export async function generateAndPersistConversationTitle(
     [],
     undefined,
     {
-      config: { max_tokens: config.daemon.titleGenerationMaxTokens },
+      config: {
+        max_tokens: config.daemon.titleGenerationMaxTokens,
+        modelIntent: "latency-optimized",
+      },
       signal: combinedSignal,
     },
   );
@@ -149,12 +156,12 @@ export async function generateAndPersistConversationTitle(
     }
 
     // Re-check replaceability before persisting (race guard)
-    const current = conversationStore.getConversation(conversationId);
+    const current = getConversation(conversationId);
     if (current && !isReplaceableTitle(current.title)) {
       return { title: current.title!, updated: false };
     }
 
-    conversationStore.updateConversationTitle(conversationId, title, 1);
+    updateConversationTitle(conversationId, title, 1);
     onTitleUpdated?.(title);
     log.info({ conversationId, title }, "Auto-generated conversation title");
     return { title, updated: true };
@@ -165,13 +172,13 @@ export async function generateAndPersistConversationTitle(
   // text-response path above). A concurrent custom rename may have landed
   // while the LLM request was in-flight; writing unconditionally would
   // clobber the user's intent.
-  const currentForFallback = conversationStore.getConversation(conversationId);
+  const currentForFallback = getConversation(conversationId);
   if (currentForFallback && !isReplaceableTitle(currentForFallback.title)) {
     return { title: currentForFallback.title!, updated: false };
   }
 
   const fallback = deriveFallbackTitle(context) ?? UNTITLED_FALLBACK;
-  conversationStore.updateConversationTitle(conversationId, fallback, 1);
+  updateConversationTitle(conversationId, fallback, 1);
   onTitleUpdated?.(fallback);
   return { title: fallback, updated: true };
 }
@@ -191,16 +198,11 @@ export function queueGenerateConversationTitle(
     );
     // Replace loading placeholder with stable fallback
     try {
-      const conversation = conversationStore.getConversation(
-        params.conversationId,
-      );
+      const conversation = getConversation(params.conversationId);
       if (conversation && conversation.title === GENERATING_TITLE) {
         const fallback =
           deriveFallbackTitle(params.context) ?? UNTITLED_FALLBACK;
-        conversationStore.updateConversationTitle(
-          params.conversationId,
-          fallback,
-        );
+        updateConversationTitle(params.conversationId, fallback);
         params.onTitleUpdated?.(fallback);
       }
     } catch {
@@ -228,7 +230,7 @@ export async function regenerateConversationTitle(
 ): Promise<{ title: string; updated: boolean }> {
   const { conversationId, onTitleUpdated, signal } = params;
 
-  const conversation = conversationStore.getConversation(conversationId);
+  const conversation = getConversation(conversationId);
   if (!conversation || !conversation.isAutoTitle) {
     return { title: conversation?.title ?? UNTITLED_FALLBACK, updated: false };
   }
@@ -238,7 +240,7 @@ export async function regenerateConversationTitle(
     return { title: conversation.title ?? UNTITLED_FALLBACK, updated: false };
   }
 
-  const allMessages = conversationStore.getMessages(conversationId);
+  const allMessages = getMessages(conversationId);
   const recentMessages = allMessages.slice(-3);
   if (recentMessages.length === 0) {
     return { title: conversation.title ?? UNTITLED_FALLBACK, updated: false };
@@ -256,7 +258,10 @@ export async function regenerateConversationTitle(
     [],
     undefined,
     {
-      config: { max_tokens: config.daemon.titleGenerationMaxTokens },
+      config: {
+        max_tokens: config.daemon.titleGenerationMaxTokens,
+        modelIntent: "latency-optimized",
+      },
       signal: combinedSignal,
     },
   );
@@ -269,12 +274,12 @@ export async function regenerateConversationTitle(
     }
 
     // Re-check isAutoTitle before persisting (race guard against manual rename)
-    const current = conversationStore.getConversation(conversationId);
+    const current = getConversation(conversationId);
     if (!current || !current.isAutoTitle) {
       return { title: current?.title ?? UNTITLED_FALLBACK, updated: false };
     }
 
-    conversationStore.updateConversationTitle(conversationId, title, 1);
+    updateConversationTitle(conversationId, title, 1);
     onTitleUpdated?.(title);
     log.info(
       { conversationId, title },

package/src/memory/db-connection.ts

@@ -2,12 +2,7 @@ import { Database } from "bun:sqlite";
 
 import { drizzle } from "drizzle-orm/bun-sqlite";
 
-import {
-  ensureDataDir,
-  getDbPath,
-  migrateToDataLayout,
-  migrateToWorkspaceLayout,
-} from "../util/platform.js";
+import { ensureDataDir, getDbPath } from "../util/platform.js";
 import * as schema from "./schema.js";
 
 export type DrizzleDb = ReturnType<typeof drizzle<typeof schema>>;
@@ -16,8 +11,6 @@ let db: DrizzleDb | null = null;
 
 export function getDb(): DrizzleDb {
   if (!db) {
-    migrateToDataLayout();
-    migrateToWorkspaceLayout();
     ensureDataDir();
     const sqlite = new Database(getDbPath());
     sqlite.exec("PRAGMA journal_mode=WAL");

package/src/memory/db-init.ts

@@ -35,6 +35,7 @@ import {
   migrateAssistantContactMetadata,
   migrateBackfillContactInteractionStats,
   migrateBackfillGuardianPrincipalId,
+  migrateBackfillUsageCacheAccounting,
   migrateCallSessionMode,
   migrateCanonicalGuardianDeliveriesDestinationIndex,
   migrateCanonicalGuardianRequesterChatId,
@@ -63,6 +64,10 @@ import {
   migrateNormalizePhoneIdentities,
   migrateNotificationDeliveryThreadDecision,
   migrateReminderRoutingIntent,
+  migrateRenameGuardianVerificationValues,
+  migrateRenameVerificationSessionIdColumn,
+  migrateRenameVerificationTable,
+  migrateRenameVoiceToPhone,
   migrateSchemaIndexesAndColumns,
   migrateUsageDashboardIndexes,
   migrateVoiceInviteColumns,
@@ -304,6 +309,21 @@ export function initializeDb(): void {
   // 43. Drop all composite usage indexes — they don't eliminate temp B-trees for GROUP BY
   migrateDropUsageCompositeIndexes(database);
 
+  // 44. Backfill historical Anthropic usage rows from request-log truth before dashboard reads
+  migrateBackfillUsageCacheAccounting(database);
+
+  // 45. Rename channel_guardian_verification_challenges → channel_verification_sessions
+  migrateRenameVerificationTable(database);
+
+  // 46. Rename guardian_verification_session_id → verification_session_id in call_sessions
+  migrateRenameVerificationSessionIdColumn(database);
+
+  // 47. Rename persisted guardian_verification call_mode and event_type values
+  migrateRenameGuardianVerificationValues(database);
+
+  // 48. Rename stored "voice" channel values to "phone" across all channel text columns
+  migrateRenameVoiceToPhone(database);
+
   validateMigrationState(database);
 
   if (process.env.BUN_TEST === "1") {

package/src/memory/delivery-crud.ts

@@ -9,11 +9,7 @@ import { and, desc, eq, isNotNull } from "drizzle-orm";
 import { v4 as uuid } from "uuid";
 
 import { DAEMON_INTERNAL_ASSISTANT_ID } from "../runtime/assistant-scope.js";
-import {
-  getConversationByKey,
-  getOrCreateConversation,
-  setConversationKeyIfAbsent,
-} from "./conversation-key-store.js";
+import { getOrCreateConversation } from "./conversation-key-store.js";
 import { getDb } from "./db.js";
 import { channelInboundEvents, conversations } from "./schema.js";
 
@@ -65,35 +61,9 @@ export function recordInbound(
     };
   }
 
-  const assistantId = options?.assistantId;
-  const legacyKey = `${sourceChannel}:${externalChatId}`;
-  const scopedKey = assistantId
-    ? `asst:${assistantId}:${sourceChannel}:${externalChatId}`
-    : legacyKey;
-
-  // Resolve conversation mapping with assistant-scoped keying:
-  // 1. If scoped key exists, use it directly.
-  // 2. If assistantId is "self" and legacy key exists, reuse the legacy
-  //    conversation and create a scoped alias to prevent future bleed.
-  // 3. Otherwise, create/get conversation from the scoped key.
-  let mapping: { conversationId: string; created: boolean };
-  const scopedMapping = assistantId ? getConversationByKey(scopedKey) : null;
-  if (scopedMapping) {
-    mapping = { conversationId: scopedMapping.conversationId, created: false };
-  } else if (assistantId === DAEMON_INTERNAL_ASSISTANT_ID) {
-    const legacyMapping = getConversationByKey(legacyKey);
-    if (legacyMapping) {
-      mapping = {
-        conversationId: legacyMapping.conversationId,
-        created: false,
-      };
-      setConversationKeyIfAbsent(scopedKey, legacyMapping.conversationId);
-    } else {
-      mapping = getOrCreateConversation(scopedKey);
-    }
-  } else {
-    mapping = getOrCreateConversation(scopedKey);
-  }
+  const assistantId = options?.assistantId ?? DAEMON_INTERNAL_ASSISTANT_ID;
+  const scopedKey = `asst:${assistantId}:${sourceChannel}:${externalChatId}`;
+  const mapping = getOrCreateConversation(scopedKey);
   const now = Date.now();
   const eventId = uuid();
 

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

@@ -1,6 +1,6 @@
 /**
  * Store for external conversation bindings — maps internal conversation IDs
- * to external channel identifiers (e.g. Telegram chat ID, SMS thread ID).
+ * to external channel identifiers (e.g. Telegram chat ID, voice session).
  *
  * This enables the system to track which conversations originated from
  * external channels and expose channel metadata in session/conversation

package/src/memory/format-recall.ts

@@ -0,0 +1,47 @@
+import { estimateTextTokens } from "../context/token-estimator.js";
+import { buildInjectedText } from "./search/formatting.js";
+import { markItemUsage, trimToTokenBudget } from "./search/ranking.js";
+import type { Candidate } from "./search/types.js";
+
+export interface FormatRecallTextOptions {
+  /** Injection format: 'markdown' or 'structured_v1'. */
+  format: string;
+  /** Maximum token budget for the formatted output. */
+  maxTokens: number;
+}
+
+export interface FormatRecallTextResult {
+  /** The formatted text ready for injection. */
+  text: string;
+  /** Candidates that fit within the token budget. */
+  selected: Candidate[];
+  /** Token count of the final injected text. */
+  tokenCount: number;
+}
+
+/**
+ * Format scored recall candidates into injectable text.
+ *
+ * Trims candidates to the token budget, groups by section with temporal
+ * grounding, applies "Lost in the Middle" ordering, and marks item usage.
+ *
+ * Extracted from `formatRecallResult()` in `retriever.ts` so both the
+ * auto-injection path and the on-demand memory_recall tool can reuse it.
+ */
+export function formatRecallText(
+  candidates: Candidate[],
+  opts: FormatRecallTextOptions,
+): FormatRecallTextResult {
+  const { format, maxTokens } = opts;
+
+  const selected = trimToTokenBudget(candidates, maxTokens, format);
+  markItemUsage(selected);
+
+  const text = buildInjectedText(selected, format);
+
+  return {
+    text,
+    selected,
+    tokenCount: estimateTextTokens(text),
+  };
+}

package/src/memory/guardian-action-store.ts

@@ -2,7 +2,7 @@
  * Store for cross-channel guardian action requests and deliveries.
  *
  * Guardian action requests are created when a voice call's ASK_GUARDIAN
- * marker fires, and deliveries track per-channel dispatch (telegram, sms, mac).
+ * marker fires, and deliveries track per-channel dispatch (telegram, mac).
  * Resolution uses first-response-wins semantics: the first channel to
  * answer resolves the request and all other deliveries are marked answered.
  */
@@ -301,11 +301,10 @@ export function resolveGuardianActionRequest(
 
 /**
  * Expire a guardian action request and all its deliveries.
- * When reason is not provided, defaults to 'sweep_timeout' for backward compatibility.
  */
 export function expireGuardianActionRequest(
   id: string,
-  reason?: ExpiredReason,
+  reason: ExpiredReason,
 ): void {
   const db = getDb();
   const now = Date.now();
@@ -313,7 +312,7 @@ export function expireGuardianActionRequest(
   db.update(guardianActionRequests)
     .set({
       status: "expired",
-      expiredReason: reason ?? "sweep_timeout",
+      expiredReason: reason,
       updatedAt: now,
     })
     .where(
@@ -864,7 +863,7 @@ export function getExpiredDeliveriesByConversation(
 /**
  * Look up deliveries for requests in `awaiting_guardian_choice` follow-up state.
  * Used by inbound message routing to intercept guardian follow-up replies
- * on channel paths (Telegram, SMS).
+ * on channel paths (Telegram, WhatsApp).
  */
 export function getFollowupDeliveriesByDestination(
   channel: string,

package/src/memory/guardian-rate-limits.ts

@@ -151,9 +151,6 @@ export function recordInvalidAttempt(
     channel,
     actorExternalUserId,
     actorChatId,
-    // Legacy columns kept for backward compatibility with upgraded databases
-    invalidAttempts: 0,
-    windowStartedAt: 0,
     attemptTimestampsJson: JSON.stringify(timestamps),
     lockedUntil,
     createdAt: now,

package/src/memory/invite-store.ts

@@ -331,7 +331,7 @@ export function findActiveVoiceInvites(params: {
     .from(assistantIngressInvites)
     .where(
       and(
-        eq(assistantIngressInvites.sourceChannel, "voice"),
+        eq(assistantIngressInvites.sourceChannel, "phone"),
         eq(assistantIngressInvites.status, "active"),
         eq(
           assistantIngressInvites.expectedExternalUserId,

package/src/memory/job-handlers/backfill.ts

@@ -11,7 +11,7 @@ import {
 import {
   getConversationMemoryScopeId,
   messageMetadataSchema,
-} from "../conversation-store.js";
+} from "../conversation-crud.js";
 import { getDb } from "../db.js";
 import { indexMessageNow } from "../indexer.js";
 import {
@@ -169,6 +169,7 @@ export function backfillEntityRelationsJob(
   const batch = db
     .select({
       id: messages.id,
+      conversationId: messages.conversationId,
       role: messages.role,
       createdAt: messages.createdAt,
       metadata: messages.metadata,
@@ -180,6 +181,7 @@ export function backfillEntityRelationsJob(
     .all();
   if (batch.length === 0) return;
 
+  const scopeCache = new Map<string, string>();
   let queuedExtractEntityJobs = 0;
   let skippedUntrusted = 0;
   for (const message of batch) {
@@ -190,7 +192,12 @@ export function backfillEntityRelationsJob(
       skippedUntrusted += 1;
       continue;
     }
-    enqueueMemoryJob("extract_entities", { messageId: message.id });
+    let scopeId = scopeCache.get(message.conversationId);
+    if (scopeId === undefined) {
+      scopeId = getConversationMemoryScopeId(message.conversationId);
+      scopeCache.set(message.conversationId, scopeId);
+    }
+    enqueueMemoryJob("extract_entities", { messageId: message.id, scopeId });
     queuedExtractEntityJobs += 1;
   }
 

package/src/memory/job-handlers/extraction.ts

@@ -22,7 +22,8 @@ const log = getLogger("memory-jobs-worker");
 
 export async function extractItemsJob(job: MemoryJob): Promise<void> {
   const messageId = asString(job.payload.messageId);
-  if (!messageId) return;
+  const scopeId = asString(job.payload.scopeId);
+  if (!messageId || !scopeId) return;
 
   // If the conversation that owns this message has been marked as failed,
   // skip extraction entirely. This prevents async extraction jobs from
@@ -41,12 +42,6 @@ export async function extractItemsJob(job: MemoryJob): Promise<void> {
     return;
   }
 
-  // Backward compat: old payloads may lack scopeId — default to 'default'
-  const scopeId =
-    typeof job.payload.scopeId === "string" && job.payload.scopeId
-      ? job.payload.scopeId
-      : "default";
-
   await extractAndUpsertMemoryItemsForMessage(
     messageId,
     scopeId,

package/src/memory/job-handlers/summarization.ts

@@ -10,7 +10,7 @@ import {
   userMessage,
 } from "../../providers/provider-send-message.js";
 import { getLogger } from "../../util/logger.js";
-import { getConversationMemoryScopeId } from "../conversation-store.js";
+import { getConversationMemoryScopeId } from "../conversation-crud.js";
 import { getDb } from "../db.js";
 import {
   asString,

package/src/memory/llm-usage-store.ts

@@ -90,6 +90,7 @@ export interface UsageTimeRange {
 
 /** Aggregate totals across a time range. */
 export interface UsageTotals {
+  /** Direct input tokens only; cache traffic is reported separately below. */
   totalInputTokens: number;
   totalOutputTokens: number;
   totalCacheCreationTokens: number;
@@ -104,6 +105,7 @@ export interface UsageTotals {
 export interface UsageDayBucket {
   /** ISO date string (YYYY-MM-DD) in UTC. */
   date: string;
+  /** Direct input tokens only; cache traffic is tracked separately in totals. */
   totalInputTokens: number;
   totalOutputTokens: number;
   totalEstimatedCostUsd: number;
@@ -113,8 +115,11 @@ export interface UsageDayBucket {
 /** A grouped breakdown row (by actor, provider, or model). */
 export interface UsageGroupBreakdown {
   group: string;
+  /** Direct input tokens only; cache traffic is reported separately below. */
   totalInputTokens: number;
   totalOutputTokens: number;
+  totalCacheCreationTokens: number;
+  totalCacheReadTokens: number;
   totalEstimatedCostUsd: number;
   eventCount: number;
 }
@@ -144,6 +149,8 @@ interface GroupRow {
   group_key: string;
   total_input_tokens: number;
   total_output_tokens: number;
+  total_cache_creation_tokens: number;
+  total_cache_read_tokens: number;
   total_estimated_cost_usd: number | null;
   event_count: number;
 }
@@ -236,6 +243,8 @@ export function getUsageGroupBreakdown(
       ${column}                                      AS group_key,
       COALESCE(SUM(input_tokens), 0)                 AS total_input_tokens,
       COALESCE(SUM(output_tokens), 0)                AS total_output_tokens,
+      COALESCE(SUM(cache_creation_input_tokens), 0)  AS total_cache_creation_tokens,
+      COALESCE(SUM(cache_read_input_tokens), 0)      AS total_cache_read_tokens,
       COALESCE(SUM(estimated_cost_usd), 0)           AS total_estimated_cost_usd,
       COUNT(*)                                        AS event_count
     FROM llm_usage_events
@@ -250,6 +259,8 @@ export function getUsageGroupBreakdown(
     group: r.group_key,
     totalInputTokens: r.total_input_tokens,
     totalOutputTokens: r.total_output_tokens,
+    totalCacheCreationTokens: r.total_cache_creation_tokens,
+    totalCacheReadTokens: r.total_cache_read_tokens,
     totalEstimatedCostUsd: r.total_estimated_cost_usd ?? 0,
     eventCount: r.event_count,
   }));