@vellumai/assistant 0.5.4 → 0.5.6

package/src/memory/conversation-crud.ts

@@ -47,18 +47,13 @@ import {
   conversations,
   conversationStarters,
   llmRequestLogs,
-  memoryChunks,
   memoryEmbeddings,
-  memoryEpisodes,
   memoryItems,
   memoryItemSources,
-  memoryObservations,
   memorySegments,
   memorySummaries,
   messageAttachments,
   messages,
-  openLoops,
-  timeContexts,
   toolInvocations,
 } from "./schema.js";
 import { cancelPendingJobsForConversation } from "./task-memory-cleanup.js";
@@ -177,9 +172,6 @@ export interface ConversationRow {
   forkParentMessageId: string | null;
   isAutoTitle: number;
   scheduleJobId: string | null;
-  memoryReducedThroughMessageId: string | null;
-  memoryDirtyTailSinceMessageId: string | null;
-  memoryLastReducedAt: number | null;
 }
 
 export const parseConversation = createRowMapper<
@@ -205,9 +197,6 @@ export const parseConversation = createRowMapper<
   forkParentMessageId: "forkParentMessageId",
   isAutoTitle: "isAutoTitle",
   scheduleJobId: "scheduleJobId",
-  memoryReducedThroughMessageId: "memoryReducedThroughMessageId",
-  memoryDirtyTailSinceMessageId: "memoryDirtyTailSinceMessageId",
-  memoryLastReducedAt: "memoryLastReducedAt",
 });
 
 export interface MessageRow {
@@ -555,9 +544,6 @@ export function deleteConversation(id: string): DeletedMemoryIds {
     segmentIds: [],
     orphanedItemIds: [],
     deletedSummaryIds: [],
-    deletedObservationIds: [],
-    deletedChunkIds: [],
-    deletedEpisodeIds: [],
   };
 
   // Capture createdAt before the transaction deletes the row — needed to
@@ -711,75 +697,6 @@ export function deleteConversation(id: string): DeletedMemoryIds {
       tx.delete(conversationStarters)
         .where(eq(conversationStarters.scopeId, memoryScopeId))
         .run();
-
-      // Sweep brief-state tables scoped to this private conversation.
-      tx.delete(timeContexts)
-        .where(eq(timeContexts.scopeId, memoryScopeId))
-        .run();
-      tx.delete(openLoops).where(eq(openLoops.scopeId, memoryScopeId)).run();
-    }
-
-    // Collect archive table IDs before the cascade delete removes them.
-    // Observations and episodes reference conversations with ON DELETE CASCADE,
-    // and chunks cascade from observations.
-    const observationRows = tx
-      .select({ id: memoryObservations.id })
-      .from(memoryObservations)
-      .where(eq(memoryObservations.conversationId, id))
-      .all();
-    const observationIds = observationRows.map((r) => r.id);
-
-    if (observationIds.length > 0) {
-      // Collect chunk IDs before observations cascade-delete them.
-      const chunkRows = tx
-        .select({ id: memoryChunks.id })
-        .from(memoryChunks)
-        .where(inArray(memoryChunks.observationId, observationIds))
-        .all();
-      const chunkIds = chunkRows.map((r) => r.id);
-
-      // Clean up embeddings for chunks.
-      if (chunkIds.length > 0) {
-        tx.delete(memoryEmbeddings)
-          .where(
-            and(
-              eq(memoryEmbeddings.targetType, "chunk"),
-              inArray(memoryEmbeddings.targetId, chunkIds),
-            ),
-          )
-          .run();
-        result.deletedChunkIds.push(...chunkIds);
-      }
-
-      // Clean up embeddings for observations.
-      tx.delete(memoryEmbeddings)
-        .where(
-          and(
-            eq(memoryEmbeddings.targetType, "observation"),
-            inArray(memoryEmbeddings.targetId, observationIds),
-          ),
-        )
-        .run();
-      result.deletedObservationIds.push(...observationIds);
-    }
-
-    const episodeRows = tx
-      .select({ id: memoryEpisodes.id })
-      .from(memoryEpisodes)
-      .where(eq(memoryEpisodes.conversationId, id))
-      .all();
-    const episodeIds = episodeRows.map((r) => r.id);
-
-    if (episodeIds.length > 0) {
-      tx.delete(memoryEmbeddings)
-        .where(
-          and(
-            eq(memoryEmbeddings.targetType, "episode"),
-            inArray(memoryEmbeddings.targetId, episodeIds),
-          ),
-        )
-        .run();
-      result.deletedEpisodeIds.push(...episodeIds);
     }
 
     tx.delete(conversations).where(eq(conversations.id, id)).run();
@@ -1005,9 +922,6 @@ export function purgePrivateConversations(): {
         segmentIds: [],
         orphanedItemIds: [],
         deletedSummaryIds: [],
-        deletedObservationIds: [],
-        deletedChunkIds: [],
-        deletedEpisodeIds: [],
       },
     };
   }
@@ -1015,18 +929,12 @@ export function purgePrivateConversations(): {
   const allSegmentIds: string[] = [];
   const allOrphanedItemIds: string[] = [];
   const allDeletedSummaryIds: string[] = [];
-  const allDeletedObservationIds: string[] = [];
-  const allDeletedChunkIds: string[] = [];
-  const allDeletedEpisodeIds: string[] = [];
 
   for (const conv of privateConvs) {
     const deleted = deleteConversation(conv.id);
     allSegmentIds.push(...deleted.segmentIds);
     allOrphanedItemIds.push(...deleted.orphanedItemIds);
     allDeletedSummaryIds.push(...deleted.deletedSummaryIds);
-    allDeletedObservationIds.push(...deleted.deletedObservationIds);
-    allDeletedChunkIds.push(...deleted.deletedChunkIds);
-    allDeletedEpisodeIds.push(...deleted.deletedEpisodeIds);
   }
 
   return {
@@ -1035,9 +943,6 @@ export function purgePrivateConversations(): {
       segmentIds: allSegmentIds,
       orphanedItemIds: allOrphanedItemIds,
       deletedSummaryIds: allDeletedSummaryIds,
-      deletedObservationIds: allDeletedObservationIds,
-      deletedChunkIds: allDeletedChunkIds,
-      deletedEpisodeIds: allDeletedEpisodeIds,
     },
   };
 }
@@ -1120,13 +1025,6 @@ export async function addMessage(
       throw err;
     }
   }
-
-  // Mark the conversation dirty for delayed memory reduction. This runs
-  // after the insert transaction succeeds so the reducer knows which
-  // conversations have unprocessed messages. The helper preserves the
-  // earliest unreduced boundary (no-op when already dirty).
-  markConversationMemoryDirty(conversationId, messageId);
-
   const message = {
     id: messageId,
     conversationId,
@@ -1431,9 +1329,6 @@ export interface DeletedMemoryIds {
   segmentIds: string[];
   orphanedItemIds: string[];
   deletedSummaryIds: string[];
-  deletedObservationIds: string[];
-  deletedChunkIds: string[];
-  deletedEpisodeIds: string[];
 }
 
 export interface WipeConversationResult extends DeletedMemoryIds {
@@ -1507,9 +1402,6 @@ export function deleteMessageById(messageId: string): DeletedMemoryIds {
     segmentIds: [],
     orphanedItemIds: [],
     deletedSummaryIds: [],
-    deletedObservationIds: [],
-    deletedChunkIds: [],
-    deletedEpisodeIds: [],
   };
 
   // Collect attachment IDs linked to this message before cascade-delete
@@ -1598,134 +1490,6 @@ export function deleteMessageById(messageId: string): DeletedMemoryIds {
   return result;
 }
 
-/**
- * Mark a conversation as having unreduced messages starting from the given
- * message. Sets `memoryDirtyTailSinceMessageId` only when it is currently
- * null so the earliest unreduced boundary is preserved across multiple
- * messages — later messages must not clobber the original dirty marker.
- *
- * Also upserts a pending `reduce_conversation_memory` job scheduled at
- * `now + idleDelayMs`. If a pending job for this conversation already exists,
- * its `runAfter` is pushed forward (rescheduled) so the reducer waits for
- * the full idle window after the *latest* message — avoiding premature runs
- * while the user is still actively typing.
- */
-export function markConversationMemoryDirty(
-  conversationId: string,
-  messageId: string,
-): void {
-  const db = getDb();
-  db.update(conversations)
-    .set({ memoryDirtyTailSinceMessageId: messageId })
-    .where(
-      and(
-        eq(conversations.id, conversationId),
-        isNull(conversations.memoryDirtyTailSinceMessageId),
-      ),
-    )
-    .run();
-
-  // Schedule (or reschedule) a deferred reducer job for this conversation.
-  scheduleReducerJob(conversationId);
-}
-
-/**
- * Upsert a pending `reduce_conversation_memory` job for the given
- * conversation, scheduled `idleDelayMs` from now. If one already exists in
- * pending state, its `runAfter` is pushed forward to restart the idle timer.
- * This ensures exactly one pending reducer job per conversation — new
- * messages reschedule rather than duplicate.
- */
-export function scheduleReducerJob(
-  conversationId: string,
-  runAfter?: number,
-): void {
-  const idleDelayMs = getReducerIdleDelayMs();
-  const scheduledAt = runAfter ?? Date.now() + idleDelayMs;
-
-  const existing = rawGet<{ id: string; status: string }>(
-    `SELECT id, status FROM memory_jobs
-     WHERE type = 'reduce_conversation_memory'
-       AND json_extract(payload, '$.conversationId') = ?
-       AND status = 'pending'
-     LIMIT 1`,
-    conversationId,
-  );
-
-  if (existing) {
-    // Reschedule: push runAfter forward so the idle timer resets.
-    rawRun(
-      `UPDATE memory_jobs SET run_after = ?, updated_at = ? WHERE id = ?`,
-      scheduledAt,
-      Date.now(),
-      existing.id,
-    );
-  } else {
-    enqueueMemoryJob(
-      "reduce_conversation_memory",
-      { conversationId },
-      scheduledAt,
-    );
-  }
-}
-
-/**
- * Startup sweep: find conversations that are marked dirty and whose tail
- * message is already older than the idle delay. For these conversations the
- * reducer should have run but didn't (daemon was down). Enqueue immediate
- * reducer jobs for each so they are processed on the next worker tick.
- *
- * Conversations whose tail is still within the idle window are skipped —
- * the normal `markConversationMemoryDirty` path will schedule them when
- * new messages arrive (or on the next conversation interaction).
- *
- * Returns the number of jobs enqueued.
- */
-export function sweepStaleReducerJobs(): number {
-  const idleDelayMs = getReducerIdleDelayMs();
-  const cutoff = Date.now() - idleDelayMs;
-
-  // Find dirty conversations whose latest message is older than the idle
-  // window AND that don't already have a pending reducer job.
-  const stale = rawAll<{ conversationId: string }>(
-    `SELECT c.id AS conversationId
-     FROM conversations c
-     WHERE c.memory_dirty_tail_since_message_id IS NOT NULL
-       AND NOT EXISTS (
-         SELECT 1 FROM memory_jobs mj
-         WHERE mj.type = 'reduce_conversation_memory'
-           AND json_extract(mj.payload, '$.conversationId') = c.id
-           AND mj.status IN ('pending', 'running')
-       )
-       AND (
-         SELECT MAX(m.created_at) FROM messages m
-         WHERE m.conversation_id = c.id
-       ) <= ?`,
-    cutoff,
-  );
-
-  for (const { conversationId } of stale) {
-    enqueueMemoryJob("reduce_conversation_memory", { conversationId });
-  }
-
-  return stale.length;
-}
-
-function getReducerIdleDelayMs(): number {
-  // Some test suites mock getConfig() with partial objects; fall back to the
-  // schema default so reducer scheduling stays stable outside full config load.
-  const config = getConfig() as {
-    memory?: {
-      simplified?: {
-        reducer?: {
-          idleDelayMs?: number;
-        };
-      };
-    };
-  };
-  return config.memory?.simplified?.reducer?.idleDelayMs ?? 30_000;
-}
-
 export function setConversationOriginChannelIfUnset(
   conversationId: string,
   channel: ChannelId,

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

@@ -133,6 +133,8 @@ export async function generateAndPersistConversationTitle(
   const result = await runBtwSidechain({
     content: prompt,
     provider,
+    systemPrompt: buildTitleSystemPrompt(),
+    tools: [],
     maxTokens: config.daemon.titleGenerationMaxTokens,
     modelIntent: "latency-optimized",
     signal,
@@ -236,6 +238,8 @@ export async function regenerateConversationTitle(
   const result = await runBtwSidechain({
     content: prompt,
     provider,
+    systemPrompt: buildTitleSystemPrompt(),
+    tools: [],
     maxTokens: config.daemon.titleGenerationMaxTokens,
     modelIntent: "latency-optimized",
     signal,
@@ -277,14 +281,30 @@ export function queueRegenerateConversationTitle(
 
 // ── Internal helpers ─────────────────────────────────────────────────
 
+/**
+ * Dedicated system prompt for title generation. Replaces the default
+ * assistant system prompt that btw-sidechain would otherwise inject,
+ * which caused the model to respond to the conversation content instead
+ * of titling it.
+ */
+function buildTitleSystemPrompt(): string {
+  return [
+    "You generate short conversation titles. Output ONLY the title text — no explanation, no quotes, no markdown, no preamble.",
+    "",
+    "Rules:",
+    "- Maximum 5 words and 40 characters",
+    "- Summarize the TOPIC the user is asking about",
+    "- Do NOT respond to the conversation content",
+    "- Do NOT assess feasibility or comment on capabilities",
+  ].join("\n");
+}
+
 function buildTitlePrompt(
   context?: TitleContext,
   userMessage?: string,
   assistantResponse?: string,
 ): string {
-  const parts: string[] = [
-    "Generate a very short title summarizing the TOPIC of this conversation. Rules: at most 5 words, at most 40 characters, no quotes, no markdown formatting. IMPORTANT: Summarize what the user is asking about — do NOT respond to the message, do NOT assess feasibility, and do NOT comment on your own capabilities.",
-  ];
+  const parts: string[] = [];
 
   if (context) {
     const hints: string[] = [];
@@ -295,12 +315,12 @@ function buildTitlePrompt(
     if (context.metadataHints?.length)
       hints.push(`Hints: ${context.metadataHints.join(", ")}`);
     if (hints.length > 0) {
-      parts.push("", "Metadata:", ...hints);
+      parts.push("Metadata:", ...hints, "");
     }
   }
 
   if (userMessage) {
-    parts.push("", `User: ${truncate(userMessage, 200, "")}`);
+    parts.push(`User: ${truncate(userMessage, 200, "")}`);
   }
   if (assistantResponse) {
     parts.push(`Assistant: ${truncate(assistantResponse, 200, "")}`);
@@ -339,11 +359,7 @@ function deriveFallbackTitle(context?: TitleContext): string | null {
 }
 
 function buildRegenerationPrompt(recentMessages: MessageRow[]): string {
-  const parts: string[] = [
-    "Generate a very short title summarizing the TOPIC of this conversation based on the recent messages below. Rules: at most 5 words, at most 40 characters, no quotes, no markdown formatting. IMPORTANT: Summarize what the user is asking about — do NOT respond to the messages, do NOT assess feasibility, and do NOT comment on your own capabilities.",
-    "",
-    "Recent messages:",
-  ];
+  const parts: string[] = ["Recent messages:"];
 
   for (const msg of recentMessages) {
     const role = msg.role === "user" ? "User" : "Assistant";

package/src/memory/db-init.ts

@@ -66,6 +66,7 @@ import {
   migrateDropMemorySegmentFts,
   migrateDropOrphanedMediaTables,
   migrateDropRemindersTable,
+  migrateDropSimplifiedMemory,
   migrateDropUsageCompositeIndexes,
   migrateFkCascadeRebuilds,
   migrateGuardianActionFollowup,
@@ -82,10 +83,7 @@ import {
   migrateInviteContactId,
   migrateLlmRequestLogMessageId,
   migrateLlmRequestLogProvider,
-  migrateMemoryArchiveTables,
-  migrateMemoryBriefState,
   migrateMemoryItemSupersession,
-  migrateMemoryReducerCheckpoints,
   migrateMessagesFtsBackfill,
   migrateNormalizePhoneIdentities,
   migrateNotificationDeliveryThreadDecision,
@@ -488,18 +486,12 @@ export function initializeDb(): void {
   // 84. Add nullable conversation fork lineage columns and parent lookup index
   migrateConversationForkLineage(database);
 
-  // 85. Memory brief state tables (time_contexts, open_loops) for simplified memory system
-  migrateMemoryBriefState(database);
-
-  // 86. Memory archive tables (observations, chunks, episodes) for simplified memory v1
-  migrateMemoryArchiveTables(database);
-
-  // 87. Add memory reducer checkpoint columns to conversations
-  migrateMemoryReducerCheckpoints(database);
-
-  // 88. Add quiet flag to schedule jobs
+  // 85. Add quiet flag to schedule jobs
   migrateScheduleQuietFlag(database);
 
+  // 86. Drop simplified-memory tables and reducer checkpoint columns
+  migrateDropSimplifiedMemory(database);
+
   validateMigrationState(database);
 
   if (process.env.BUN_TEST === "1") {

package/src/memory/embedding-local.ts

@@ -1,6 +1,7 @@
 import { existsSync, unlinkSync, writeFileSync } from "node:fs";
 import { join } from "node:path";
 
+import { getIsContainerized } from "../config/env-registry.js";
 import { getLogger } from "../util/logger.js";
 import { getEmbeddingModelsDir, getRootDir } from "../util/platform.js";
 import { PromiseGuard } from "../util/promise-guard.js";
@@ -353,12 +354,17 @@ export class LocalEmbeddingBackend implements EmbeddingBackend {
 
   private static readonly PID_FILENAME = "embed-worker.pid";
 
+  /** PID files are process-local state — store in /tmp when containerized to keep shared volumes clean. */
+  private getPidFilePath(): string {
+    if (getIsContainerized()) {
+      return join("/tmp", LocalEmbeddingBackend.PID_FILENAME);
+    }
+    return join(getRootDir(), LocalEmbeddingBackend.PID_FILENAME);
+  }
+
   private writePidFile(pid: number): void {
     try {
-      writeFileSync(
-        join(getRootDir(), LocalEmbeddingBackend.PID_FILENAME),
-        String(pid),
-      );
+      writeFileSync(this.getPidFilePath(), String(pid));
     } catch {
       // Best-effort — doesn't affect functionality
     }
@@ -366,7 +372,7 @@ export class LocalEmbeddingBackend implements EmbeddingBackend {
 
   private removePidFile(): void {
     try {
-      unlinkSync(join(getRootDir(), LocalEmbeddingBackend.PID_FILENAME));
+      unlinkSync(this.getPidFilePath());
     } catch {
       // Best-effort
     }

package/src/memory/indexer.ts

@@ -5,7 +5,6 @@ import { getConfig } from "../config/loader.js";
 import type { MemoryConfig } from "../config/types.js";
 import type { TrustClass } from "../runtime/actor-trust-resolver.js";
 import { getLogger } from "../util/logger.js";
-import { computeChunkContentHash } from "./archive-store.js";
 import { getDb } from "./db.js";
 import { selectedBackendSupportsMultimodal } from "./embedding-backend.js";
 import { enqueueMemoryJob } from "./jobs-store.js";
@@ -13,7 +12,7 @@ import {
   extractMediaBlockMeta,
   extractTextFromStoredMessageContent,
 } from "./message-content.js";
-import { memoryChunks, memoryObservations, memorySegments } from "./schema.js";
+import { memorySegments } from "./schema.js";
 import { segmentText } from "./segmenter.js";
 
 const log = getLogger("memory-indexer");
@@ -54,12 +53,7 @@ export async function indexMessageNow(
     input.provenanceTrustClass === undefined;
 
   const text = extractTextFromStoredMessageContent(input.content);
-  const hasText = text.length > 0;
-  const candidateMediaMeta = extractMediaBlockMeta(input.content).filter(
-    (b) => b.type === "image",
-  );
-  const hasMedia = candidateMediaMeta.length > 0;
-  if (!hasText && !hasMedia) {
+  if (text.length === 0) {
     enqueueMemoryJob("build_conversation_summary", {
       conversationId: input.conversationId,
     });
@@ -68,13 +62,11 @@ export async function indexMessageNow(
 
   const db = getDb();
   const now = Date.now();
-  const segments = hasText
-    ? segmentText(
-        text,
-        config.segmentation.targetTokens,
-        config.segmentation.overlapTokens,
-      )
-    : [];
+  const segments = segmentText(
+    text,
+    config.segmentation.targetTokens,
+    config.segmentation.overlapTokens,
+  );
   const shouldExtract =
     input.role === "user" ||
     (input.role === "assistant" && config.extraction.extractFromAssistant);
@@ -84,6 +76,9 @@ export async function indexMessageNow(
   // overhead for messages on non-multimodal backends.
   // selectedBackendSupportsMultimodal requires async key resolution, so we
   // skip it entirely for text-only messages.
+  const candidateMediaMeta = extractMediaBlockMeta(input.content).filter(
+    (b) => b.type === "image",
+  );
   const mediaBlocks =
     candidateMediaMeta.length > 0 &&
     (await selectedBackendSupportsMultimodal(getConfig()))
@@ -93,10 +88,7 @@ export async function indexMessageNow(
   // Wrap all segment inserts and job enqueues in a single transaction so they
   // either all succeed or all roll back, preventing partial/orphaned state.
   let skippedEmbedJobs = 0;
-  let skippedChunkEmbedJobs = 0;
-  const scopeId = input.scopeId ?? "default";
   db.transaction((tx) => {
-    // ── Legacy segment path (kept intact for parallel validation) ───
     for (const segment of segments) {
       const segmentId = buildSegmentId(input.messageId, segment.segmentIndex);
       const hash = createHash("sha256").update(segment.text).digest("hex");
@@ -117,7 +109,7 @@ export async function indexMessageNow(
           segmentIndex: segment.segmentIndex,
           text: segment.text,
           tokenEstimate: segment.tokenEstimate,
-          scopeId,
+          scopeId: input.scopeId ?? "default",
           contentHash: hash,
           createdAt: input.createdAt,
           updatedAt: now,
@@ -127,7 +119,7 @@ export async function indexMessageNow(
           set: {
             text: segment.text,
             tokenEstimate: segment.tokenEstimate,
-            scopeId,
+            scopeId: input.scopeId ?? "default",
             contentHash: hash,
             updatedAt: now,
           },
@@ -141,65 +133,6 @@ export async function indexMessageNow(
       }
     }
 
-    // ── Archive chunk dual-write (mirrors segment boundaries) ──────
-    // Create a single observation per message, then create one chunk per
-    // segment using the same segmentation boundaries. Chunks are
-    // deduplicated by (scopeId, contentHash) via onConflictDoNothing so
-    // unchanged content does not enqueue duplicate embed_chunk jobs.
-    const observationId = buildObservationId(input.messageId);
-    tx.insert(memoryObservations)
-      .values({
-        id: observationId,
-        scopeId,
-        conversationId: input.conversationId,
-        messageId: input.messageId,
-        role: input.role,
-        content: hasText ? text : input.content,
-        modality: hasMedia ? "multimodal" : "text",
-        source: null,
-        createdAt: input.createdAt,
-      })
-      .onConflictDoNothing({ target: memoryObservations.id })
-      .run();
-
-    for (const segment of segments) {
-      const chunkId = buildChunkId(input.messageId, segment.segmentIndex);
-      const chunkHash = computeChunkContentHash(scopeId, segment.text);
-
-      // Check if this chunk already exists with the same content hash
-      const existingChunk = tx
-        .select({ contentHash: memoryChunks.contentHash })
-        .from(memoryChunks)
-        .where(eq(memoryChunks.id, chunkId))
-        .get();
-
-      tx.insert(memoryChunks)
-        .values({
-          id: chunkId,
-          scopeId,
-          observationId,
-          content: segment.text,
-          tokenEstimate: segment.tokenEstimate,
-          contentHash: chunkHash,
-          createdAt: input.createdAt,
-        })
-        .onConflictDoUpdate({
-          target: memoryChunks.id,
-          set: {
-            content: segment.text,
-            tokenEstimate: segment.tokenEstimate,
-            contentHash: chunkHash,
-          },
-        })
-        .run();
-
-      if (existingChunk?.contentHash === chunkHash) {
-        skippedChunkEmbedJobs++;
-      } else {
-        enqueueMemoryJob("embed_chunk", { chunkId, scopeId }, Date.now(), tx);
-      }
-    }
-
     // Enqueue embed_attachment jobs for image content blocks when the
     // embedding provider supports multimodal (Gemini only).
     for (const block of mediaBlocks) {
@@ -214,7 +147,7 @@ export async function indexMessageNow(
     if (shouldExtract && isTrustedActor && !input.automated) {
       enqueueMemoryJob(
         "extract_items",
-        { messageId: input.messageId, scopeId },
+        { messageId: input.messageId, scopeId: input.scopeId ?? "default" },
         Date.now(),
         tx,
       );
@@ -233,12 +166,6 @@ export async function indexMessageNow(
     );
   }
 
-  if (skippedChunkEmbedJobs > 0) {
-    log.debug(
-      `Skipped ${skippedChunkEmbedJobs}/${segments.length} embed_chunk jobs (content unchanged)`,
-    );
-  }
-
   if (!isTrustedActor && shouldExtract) {
     log.info(
       `Skipping extraction jobs for untrusted actor (trustClass=${input.provenanceTrustClass})`,
@@ -250,11 +177,9 @@ export async function indexMessageNow(
   }
 
   const extractionGated = !isTrustedActor || !!input.automated;
-  const segmentEmbedJobs = segments.length - skippedEmbedJobs;
-  const chunkEmbedJobs = segments.length - skippedChunkEmbedJobs;
   const enqueuedJobs =
-    segmentEmbedJobs +
-    chunkEmbedJobs +
+    segments.length -
+    skippedEmbedJobs +
     mediaBlocks.length +
     (shouldExtract && !extractionGated ? 2 : 1);
   return {
@@ -288,19 +213,3 @@ export function getRecentSegmentsForConversation(
 function buildSegmentId(messageId: string, segmentIndex: number): string {
   return `${messageId}:${segmentIndex}`;
 }
-
-/**
- * Deterministic observation ID derived from the messageId so repeated
- * indexer runs for the same message converge on the same observation row.
- */
-function buildObservationId(messageId: string): string {
-  return `obs:${messageId}`;
-}
-
-/**
- * Deterministic chunk ID derived from the messageId and segment index so
- * the dual-write path mirrors the legacy segment identity scheme exactly.
- */
-function buildChunkId(messageId: string, segmentIndex: number): string {
-  return `chunk:${messageId}:${segmentIndex}`;
-}