@vellumai/assistant 0.10.4 → 0.10.5-dev.202607022329.d89bf2f

package/src/plugins/defaults/memory/persistence-hooks.ts

@@ -1,9 +1,18 @@
+import { join } from "node:path";
+
 import { getConfig } from "../../../config/loader.js";
+import {
+  clearMessagesLexicalIndex,
+  enqueueDeleteMessageLexical,
+  enqueuePurgeConversationLexical,
+} from "../../../persistence/job-handlers/message-lexical.js";
 import type {
   ConversationForkedEvent,
   MemoryPersistenceHooks,
   MessagePersistedEvent,
 } from "../../../persistence/memory-lifecycle-hooks.js";
+import { getWorkspaceDir } from "../../../util/platform.js";
+import { getMemoryConfig } from "./config.js";
 import { forkGraphMemoryState } from "./graph/graph-memory-state-store.js";
 import { indexMessageNow } from "./indexer.js";
 import { sweepOrphanMemoryRetrospectiveConversations } from "./memory-retrospective-startup-cleanup.js";
@@ -13,6 +22,7 @@ import {
   forkActivationState,
   seedForkActivationState,
 } from "./v2/activation-store.js";
+import { countBufferLines } from "./v2/consolidation-job.js";
 import {
   extractInjectedConceptSlugs,
   readInjectedBlock,
@@ -31,7 +41,7 @@ import {
  */
 export const memoryPersistenceHooks: MemoryPersistenceHooks = {
   async onMessagePersisted(event: MessagePersistedEvent): Promise<void> {
-    await indexMessageNow({ ...event, scopeId: "default" }, getConfig().memory);
+    await indexMessageNow({ ...event, scopeId: "default" }, getMemoryConfig());
   },
 
   onConversationForked(event: ConversationForkedEvent): void {
@@ -110,10 +120,45 @@ export const memoryPersistenceHooks: MemoryPersistenceHooks = {
   },
 
   onConversationWiped(conversationId: string): number {
+    // Cancel pending memory jobs only. The lexical purge is fired from the
+    // shared delete primitive via `onConversationDeleted` (which
+    // `wipeConversation` reaches through its internal `deleteConversation`), so
+    // the purge lands AFTER this cancellation pass and cannot be swept by it.
     return cancelPendingJobsForConversation(conversationId);
   },
 
+  onConversationDeleted(conversationId: string): void {
+    // Purge the conversation's points from the lexical (Qdrant) index. Fired
+    // from the shared delete primitive, so every delete caller — route, wipe,
+    // retrospective cleanup, GC — cleans up. The enqueue helper self-selects:
+    // enqueue a job when memory is enabled, run the delete inline (best-effort,
+    // breaker-wrapped) when it is disabled.
+    enqueuePurgeConversationLexical(conversationId);
+  },
+
+  onMessagesDeleted(messageIds: string[]): void {
+    // Remove each deleted message's point from the lexical index. The enqueue
+    // helper self-selects: enqueue a job when memory is enabled, run the delete
+    // inline (best-effort, breaker-wrapped) when it is disabled.
+    for (const messageId of messageIds) {
+      enqueueDeleteMessageLexical(messageId);
+    }
+  },
+
+  async onAllConversationsCleared(): Promise<void> {
+    // Drop the whole lexical (Qdrant) collection — a "delete all" leaves no ids
+    // to key per-message cleanup on. Awaited so the drop completes before
+    // clear-all returns and writes resume; otherwise a message created right
+    // after clear-all could upsert into the not-yet-dropped collection and then
+    // be erased when the drop lands.
+    await clearMessagesLexicalIndex(getConfig());
+  },
+
   onWorkerStartup(): void {
     sweepOrphanMemoryRetrospectiveConversations();
   },
+
+  countMemoryBufferLines(): number {
+    return countBufferLines(join(getWorkspaceDir(), "memory", "buffer.md"));
+  },
 };

package/src/plugins/defaults/memory/pkb/pkb-index.ts

@@ -15,10 +15,9 @@ import { createHash } from "node:crypto";
 import { readdir, readFile, stat } from "node:fs/promises";
 import { join, relative } from "node:path";
 
-import { getConfig } from "../../../../config/loader.js";
 import { withQdrantBreaker } from "../../../../persistence/embeddings/qdrant-circuit-breaker.js";
 import { getQdrantClient } from "../../../../persistence/embeddings/qdrant-client.js";
-import { embedAndUpsert } from "../../../../persistence/job-utils.js";
+import { embedAndUpsert } from "../embeddings.js";
 import type { PkbIndexEntry } from "./types.js";
 import { PKB_TARGET_TYPE } from "./types.js";
 
@@ -189,8 +188,6 @@ export async function indexPkbFile(
   const relPath = relative(pkbRoot, absPath);
   const chunks = chunkPkbFile(content);
 
-  const config = getConfig();
-
   const qdrant = getQdrantClient();
   const existing = await withQdrantBreaker(() =>
     qdrant.scrollByTargetType(PKB_TARGET_TYPE, {
@@ -209,7 +206,6 @@ export async function indexPkbFile(
     const targetId = `${memoryScopeId}:${relPath}#${chunkIndex}`;
     newTargetIds.add(targetId);
     await embedAndUpsert(
-      config,
       PKB_TARGET_TYPE,
       targetId,
       { type: "text", text: chunk },

package/src/plugins/defaults/memory/pkb/pkb-search.ts

@@ -2,7 +2,6 @@
 // PKB — Qdrant hybrid search for indexed PKB markdown files
 // ---------------------------------------------------------------------------
 
-import { getConfig } from "../../../../config/loader.js";
 import {
   isQdrantBreakerOpen,
   withQdrantBreaker,
@@ -13,6 +12,7 @@ import type {
 } from "../../../../persistence/embeddings/qdrant-client.js";
 import { getQdrantClient } from "../../../../persistence/embeddings/qdrant-client.js";
 import { getLogger } from "../../../../util/logger.js";
+import { getMemoryConfig } from "../config.js";
 import type { PkbSearchResult } from "./types.js";
 import { PKB_TARGET_TYPE } from "./types.js";
 
@@ -44,7 +44,7 @@ export async function searchPkbFiles(
   // v2 owns the read path when enabled; v2 absorbs PKB as a read source,
   // so PKB hint search short-circuits to keep traffic off the v1 collection
   // (avoiding OOM-crash risk from a corrupted sparse segment).
-  if (getConfig().memory.v2.enabled) return [];
+  if (getMemoryConfig().v2.enabled) return [];
 
   if (isQdrantBreakerOpen()) {
     log.warn("Qdrant circuit breaker open, skipping PKB search");

package/src/plugins/defaults/memory/routes/__tests__/memory-v2-simulate-route.test.ts

@@ -22,12 +22,10 @@ import { afterEach, beforeEach, describe, expect, mock, test } from "bun:test";
 import type {
   Message,
   Provider,
-  ProviderResponse,
   SendMessageOptions,
-  ToolUseContent,
 } from "@vellumai/plugin-api";
 
-import type { ToolDefinition } from "../../../../../providers/types.js";
+import type { ToolDefinition } from "../../llm-helpers.js";
 
 // ---------------------------------------------------------------------------
 // Mocks (installed before the route module is imported)
@@ -123,10 +121,10 @@ interface ProviderCall {
   options: SendMessageOptions | undefined;
 }
 const providerCalls: ProviderCall[] = [];
-mock.module("../../../../../providers/provider-send-message.js", () => ({
+// The route imports `getConfiguredProvider` from `@vellumai/plugin-api`; the
+// pure `extractToolUse` helper runs for real from the plugin's `llm-helpers`.
+mock.module("@vellumai/plugin-api", () => ({
   getConfiguredProvider: async () => providerStub,
-  extractToolUse: (response: ProviderResponse) =>
-    response.content.find((b): b is ToolUseContent => b.type === "tool_use"),
 }));
 
 // Platform helpers. `getWorkspaceDir` must return the per-test tmp dir so

package/src/plugins/defaults/memory/routes/memory-item-routes.ts

@@ -27,7 +27,6 @@ import { z } from "zod";
 import { getConfig } from "../../../../config/loader.js";
 import { getDb } from "../../../../persistence/db-connection.js";
 import {
-  embedWithBackend,
   generateSparseEmbedding,
   getMemoryBackendStatus,
 } from "../../../../persistence/embeddings/embedding-backend.js";
@@ -46,6 +45,7 @@ import {
 } from "../../../../runtime/routes/errors.js";
 import type { RouteDefinition } from "../../../../runtime/routes/types.js";
 import { getLogger } from "../../../../util/logger.js";
+import { embedWithBackend } from "../embeddings.js";
 import { createNode, deleteNode, getNode, updateNode } from "../graph/store.js";
 import type {
   Fidelity,

package/src/plugins/defaults/memory/search/semantic.ts

@@ -1,6 +1,5 @@
 import { inArray } from "drizzle-orm";
 
-import { getConfig } from "../../../../config/loader.js";
 import { getDb } from "../../../../persistence/db-connection.js";
 import { withQdrantBreaker } from "../../../../persistence/embeddings/qdrant-circuit-breaker.js";
 import type {
@@ -12,6 +11,7 @@ import {
   memorySegments,
   memorySummaries,
 } from "../../../../persistence/schema/index.js";
+import { getMemoryConfig } from "../config.js";
 import { mapCosineToUnit } from "../validation.js";
 // ── Types (inlined from deleted types.ts) ──────────────────────────
 
@@ -62,7 +62,7 @@ export async function semanticSearch(
   // v2 owns the read path when enabled; the v1 `memory` collection is in
   // active retirement, and routing semantic recall there would re-enter the
   // same corrupted sparse segments that can OOM-crash Qdrant.
-  if (getConfig().memory.v2.enabled) return [];
+  if (getMemoryConfig().v2.enabled) return [];
 
   const qdrant = getQdrantClient();
 

package/src/plugins/defaults/memory/startup.ts

@@ -0,0 +1,244 @@
+/**
+ * Memory plugin — background startup orchestration.
+ *
+ * Boots the Qdrant vector store, reconciles the embedding-identity and v2
+ * concept-page collections, runs the PKB and BM25 reconciles, and seeds the
+ * capability graph. Kicked off fire-and-forget from the memory plugin's `init`
+ * hook during daemon plugin bootstrap (after the runtime HTTP server is up), so
+ * the daemon accepts requests without waiting on Qdrant. Each step contains its
+ * own failure so a memory-subsystem problem never blocks boot.
+ *
+ * Job-handler registration happens in the plugin's `init` hook, synchronously,
+ * before this is kicked off — so the memory handlers are guaranteed to be in the
+ * dispatch table before the jobs worker started near the end of this function
+ * claims its first job.
+ */
+
+import { join } from "node:path";
+
+import type { AssistantConfig } from "../../../config/schema.js";
+import { reconcileEmbeddingIdentity } from "../../../daemon/embedding-reconcile.js";
+import { refreshSkillCapabilityMemories } from "../../../daemon/skill-memory-refresh.js";
+import { selectEmbeddingBackend } from "../../../persistence/embeddings/embedding-backend.js";
+import {
+  initMessagesLexicalIndex,
+  MESSAGES_LEXICAL_COLLECTION,
+} from "../../../persistence/embeddings/messages-lexical-index.js";
+import {
+  clearRebuildSentinel,
+  initQdrantClient,
+} from "../../../persistence/embeddings/qdrant-client.js";
+import { createQdrantManager } from "../../../persistence/embeddings/qdrant-manager.js";
+import {
+  enqueueMemoryJob,
+  isMemoryEnabled,
+} from "../../../persistence/jobs-store.js";
+import { startMemoryJobsWorker } from "../../../persistence/jobs-worker.js";
+import { getLogger } from "../../../util/logger.js";
+import { getWorkspaceDir } from "../../../util/platform.js";
+import { resolveQdrantUrl } from "./embeddings.js";
+import { sweepConceptPageFrontmatter } from "./v2/frontmatter-sweep.js";
+import {
+  maybeRebuildMemoryV2Concepts,
+  rebuildBm25CorpusStatsAndReseedSkills,
+} from "./v2/memory-v2-startup.js";
+
+const log = getLogger("memory-startup");
+
+export async function runMemoryStartup(config: AssistantConfig): Promise<void> {
+  const qdrantUrl = resolveQdrantUrl();
+  log.info({ qdrantUrl }, "Daemon startup: initializing Qdrant");
+  const manager = createQdrantManager({ url: qdrantUrl });
+  const QDRANT_START_MAX_ATTEMPTS = 3;
+  let qdrantStarted = false;
+  for (let attempt = 1; attempt <= QDRANT_START_MAX_ATTEMPTS; attempt++) {
+    try {
+      await manager.start();
+      qdrantStarted = true;
+      break;
+    } catch (err) {
+      if (attempt < QDRANT_START_MAX_ATTEMPTS) {
+        const backoffMs = attempt * 5_000; // 5s, 10s
+        log.warn(
+          {
+            err,
+            attempt,
+            maxAttempts: QDRANT_START_MAX_ATTEMPTS,
+            backoffMs,
+          },
+          "Qdrant startup failed, retrying",
+        );
+        await Bun.sleep(backoffMs);
+      } else {
+        log.warn(
+          { err },
+          "Qdrant failed to start after all attempts — memory features will be unavailable",
+        );
+      }
+    }
+  }
+
+  if (qdrantStarted) {
+    // Skip the v1 Qdrant collection lifecycle when memory v2 is active —
+    // the v1 collection has no writers (handleRemember returns early) or
+    // readers (graph search is bypassed) under v2, so ensuring/migrating
+    // it just maintains a dead-on-arrival collection. Existing on-disk
+    // collections are left intact so flipping v2 off restores v1 cleanly.
+    if (!config.memory.v2.enabled) {
+      try {
+        const embeddingSelection = await selectEmbeddingBackend(config);
+        // Sentinel only encodes the dense provider+model identity; sparse
+        // encoder changes never require collection recreation, so they
+        // intentionally do not contribute to the v1 collection identity.
+        const embeddingModel = embeddingSelection.backend
+          ? `${embeddingSelection.backend.provider}:${embeddingSelection.backend.model}`
+          : undefined;
+        const qdrantClient = initQdrantClient({
+          url: qdrantUrl,
+          collection: config.memory.qdrant.collection,
+          vectorSize: config.memory.qdrant.vectorSize,
+          onDisk: config.memory.qdrant.onDisk,
+          quantization: config.memory.qdrant.quantization,
+          embeddingModel,
+        });
+
+        // Eagerly ensure the collection exists so we detect migrations
+        // (unnamed→named vectors, dimension/model changes) at startup.
+        // If a destructive migration occurred, enqueue a rebuild_index job
+        // to re-embed all memory items from the SQLite cache.
+        const { migrated } = await qdrantClient.ensureCollection();
+        if (migrated && isMemoryEnabled()) {
+          enqueueMemoryJob("rebuild_index", {});
+          // Clear the on-disk sentinel the ensure-path writes before its
+          // destructive delete: now that rebuild_index is queued, the
+          // cross-boot signal can retire. No-op if no sentinel was written.
+          await clearRebuildSentinel();
+          log.info(
+            "Qdrant collection was migrated — enqueued rebuild_index job",
+          );
+        }
+
+        log.info("Qdrant vector store initialized");
+      } catch (err) {
+        log.warn(
+          { err },
+          "Qdrant client initialization failed — memory features will be degraded",
+        );
+      }
+    }
+
+    // Initialize the messages lexical index — a dedicated sparse-only Qdrant
+    // collection that is the lexical (BM25-style) replacement for SQLite FTS5
+    // over message content. Independent of the v1 dense collection lifecycle
+    // (sparse-only, no embedding model), so it is constructed whenever Qdrant
+    // is up regardless of the v2 flag. Lazy collection creation lives in the
+    // client, so this only wires the singleton.
+    try {
+      initMessagesLexicalIndex({
+        url: qdrantUrl,
+        collection: MESSAGES_LEXICAL_COLLECTION,
+        onDisk: config.memory.qdrant.onDisk,
+      });
+    } catch (err) {
+      log.warn(
+        { err },
+        "Messages lexical index initialization failed — lexical search will be degraded",
+      );
+    }
+
+    // Reconcile the committed embedding-collection dimension against a live
+    // backend probe (confirm-before-destroy) before the v2 rebuild and the
+    // worker drain, so `memory.qdrant.vectorSize` is settled first. Its own
+    // try/catch keeps an unreachable backend or reconcile failure from
+    // blocking startup.
+    try {
+      await reconcileEmbeddingIdentity(config);
+    } catch (err) {
+      log.warn(
+        { err },
+        "Embedding-identity reconcile threw — continuing startup",
+      );
+    }
+
+    // Detect schema drift on the v2 concept-page collection (e.g.
+    // pre-#29823 collections lacking summary_dense / summary_sparse) and
+    // recreate + enqueue a reembed when needed. Awaited inline so the
+    // reembed enqueue happens before the memory worker drains its first
+    // batch; the call's own try/catch keeps any v2-side failure from
+    // blocking the v1 PKB reconcile or BM25 build below.
+    try {
+      await maybeRebuildMemoryV2Concepts(config);
+    } catch (err) {
+      log.warn(
+        { err },
+        "Memory v2 collection schema check threw — continuing startup",
+      );
+    }
+
+    // Reconcile the PKB Qdrant index against the on-disk tree. Gated on
+    // !v2 because PKB is the v1 storage layer; under v2 the v1 collection
+    // is not initialized, so calling `getQdrantClient()` here would throw.
+    // Fire-and-forget so enqueued re-index jobs drain in the background
+    // and first-turn latency stays unaffected.
+    if (!config.memory.v2.enabled) {
+      void (async () => {
+        try {
+          const { reconcilePkbIndex } = await import("./pkb/pkb-reconcile.js");
+          const { PKB_WORKSPACE_SCOPE } = await import("./pkb/types.js");
+          const pkbRoot = join(getWorkspaceDir(), "pkb");
+          await reconcilePkbIndex(pkbRoot, PKB_WORKSPACE_SCOPE);
+        } catch (err) {
+          log.warn(
+            { err },
+            "PKB index reconciliation failed — continuing startup",
+          );
+        }
+      })();
+    }
+
+    void rebuildBm25CorpusStatsAndReseedSkills(config);
+
+    try {
+      await sweepConceptPageFrontmatter(config, getWorkspaceDir());
+    } catch (err) {
+      log.warn(
+        { err },
+        "Concept page frontmatter sweep threw — continuing startup",
+      );
+    }
+  }
+
+  // `startMemoryJobsWorker` starts the in-process supervisor (which owns the
+  // synchronous runner and stands down when an out-of-process worker is live)
+  // and spawns the out-of-process worker at boot when `memory.worker.enabled`
+  // is set. Shutdown stops whichever worker is actually running — see
+  // shutdown-handlers.ts. The job handlers were registered synchronously by the
+  // plugin's `init` hook before this function was kicked off, so the dispatch
+  // table is populated before the worker's first claim.
+  log.info("Daemon startup: starting memory worker");
+  startMemoryJobsWorker();
+
+  // Seed capability graph nodes (new memory graph system)
+  try {
+    const { seedCliGraphNodes } = await import("./graph/capability-seed.js");
+    refreshSkillCapabilityMemories(config);
+    await seedCliGraphNodes();
+  } catch (err) {
+    log.warn({ err }, "Graph capability seeding failed — continuing");
+  }
+
+  // Auto-bootstrap: if the graph has no non-procedural nodes but historical
+  // segments exist, enqueue a one-time graph_bootstrap job to populate the
+  // graph from conversation history and journal files.
+  try {
+    const { maybeEnqueueGraphBootstrap, cleanupStaleItemVectors } =
+      await import("./graph/bootstrap.js");
+    maybeEnqueueGraphBootstrap();
+    // Fire-and-forget: clean up orphaned Qdrant vectors from dropped memory_items table
+    void cleanupStaleItemVectors().catch((err) =>
+      log.warn({ err }, "Stale item vector cleanup failed — continuing"),
+    );
+  } catch (err) {
+    log.warn({ err }, "Graph bootstrap check failed — continuing");
+  }
+}

package/src/plugins/defaults/memory/task-memory-cleanup.ts

@@ -35,6 +35,7 @@ function cancelJobsByPayloadIds(
     const chunk = ids.slice(i, i + ID_CHUNK_SIZE);
     const placeholders = chunk.map(() => "?").join(", ");
     affected += rawMemoryRun(
+      "taskMemory:cancelJobsByPayloadIds",
       `UPDATE memory_jobs
           SET status = 'failed',
               last_error = ?,
@@ -62,6 +63,7 @@ export function isConversationFailed(conversationId: string): boolean {
   // which is a monotonically increasing epoch timestamp) has an error status.
   // Note: cron_runs.id is a UUID v4 (random), so we cannot use MAX(id).
   const row = rawGet<{ found: number }>(
+    "taskMemory:isConversationFailed",
     `SELECT 1 AS found
        FROM (
          SELECT 1 FROM task_runs WHERE conversation_id = ? AND status = 'failed'
@@ -98,6 +100,7 @@ export function invalidateAssistantInferredItemsForConversation(
   cancelPendingExtractionJobsForConversation(conversationId);
 
   const affected = rawRun(
+    "taskMemory:invalidateInferredNodes",
     `UPDATE memory_graph_nodes
         SET fidelity = 'gone',
             last_accessed = ?
@@ -164,6 +167,7 @@ export function cancelPendingJobsForConversation(
 
   // Jobs keyed by messageId: embed_attachment
   const messageIds = rawAll<{ id: string }>(
+    "taskMemory:cancelJobs:messageIds",
     `SELECT id FROM messages WHERE conversation_id = ?`,
     conversationId,
   ).map((r) => r.id);
@@ -171,6 +175,7 @@ export function cancelPendingJobsForConversation(
 
   // Jobs keyed by conversationId: graph_extract, build_conversation_summary
   total += rawMemoryRun(
+    "taskMemory:cancelJobs:byConversation",
     `UPDATE memory_jobs
         SET status = 'failed',
             last_error = ?,
@@ -184,6 +189,7 @@ export function cancelPendingJobsForConversation(
 
   // Jobs keyed by segmentId: embed_segment (segments belong to the conversation)
   const segmentIds = rawAll<{ id: string }>(
+    "taskMemory:cancelJobs:segmentIds",
     `SELECT id FROM memory_segments WHERE conversation_id = ?`,
     conversationId,
   ).map((r) => r.id);
@@ -191,6 +197,7 @@ export function cancelPendingJobsForConversation(
 
   // Jobs keyed by nodeId: embed_graph_node (nodes sourced from this conversation)
   const nodeIds = rawAll<{ id: string }>(
+    "taskMemory:cancelJobs:nodeIds",
     `SELECT mgn.id
        FROM memory_graph_nodes mgn, json_each(mgn.source_conversations) jc
       WHERE jc.value = ?`,
@@ -219,6 +226,7 @@ function cancelPendingExtractionJobsForConversation(
 ): number {
   const now = Date.now();
   const cancelled = rawMemoryRun(
+    "taskMemory:cancelExtractionJobs",
     `UPDATE memory_jobs
         SET status = 'failed',
             last_error = 'conversation_failed',

package/src/{daemon → plugins/defaults/memory/v2}/__tests__/memory-v2-startup.test.ts

@@ -1,6 +1,6 @@
 import { describe, expect, mock, test } from "bun:test";
 
-import type { AssistantConfig } from "../../config/schema.js";
+import type { AssistantConfig } from "../../../../../config/schema.js";
 
 function fakeConfig(): AssistantConfig {
   return {
@@ -30,20 +30,20 @@ async function withMocks(opts: { reembedInFlight: boolean }) {
     hasConceptPages: mock(async () => true),
   };
 
-  mock.module("../../persistence/jobs-store.js", () => ({
+  mock.module("../../../../../persistence/jobs-store.js", () => ({
     enqueueMemoryJob: spies.enqueueMemoryJob,
     hasActiveJobOfType: spies.hasActiveJobOfType,
   }));
-  mock.module("../../plugins/defaults/memory/v2/qdrant.js", () => ({
+  mock.module("../qdrant.js", () => ({
     ensureConceptPageCollection: spies.ensureConceptPageCollection,
     countConceptPagePoints: spies.countConceptPagePoints,
     clearReembedSentinel: spies.clearReembedSentinel,
   }));
-  mock.module("../../plugins/defaults/memory/v2/page-store.js", () => ({
+  mock.module("../page-store.js", () => ({
     hasConceptPages: spies.hasConceptPages,
   }));
-  const realPlatform = await import("../../util/platform.js");
-  mock.module("../../util/platform.js", () => ({
+  const realPlatform = await import("../../../../../util/platform.js");
+  mock.module("../../../../../util/platform.js", () => ({
     ...realPlatform,
     getWorkspaceDir: () => "/tmp/workspace",
   }));

package/src/plugins/defaults/memory/v2/__tests__/migration.test.ts

@@ -28,7 +28,6 @@ import { afterEach, beforeEach, describe, expect, mock, test } from "bun:test";
 import type {
   Message,
   Provider,
-  ProviderResponse,
   SendMessageOptions,
 } from "@vellumai/plugin-api";
 
@@ -62,18 +61,8 @@ mock.module("../../../../../persistence/jobs-store.js", () => ({
 // provider, but the safety net mock keeps any accidental real-call from
 // reaching the network — and lets us add a "no provider configured" test.
 let providerStub: Provider | null = null;
-mock.module("../../../../../providers/provider-send-message.js", () => ({
+mock.module("@vellumai/plugin-api", () => ({
   getConfiguredProvider: async () => providerStub,
-  userMessage: (text: string) => ({
-    role: "user" as const,
-    content: [{ type: "text" as const, text }],
-  }),
-  extractText: (response: ProviderResponse) => {
-    const block = response.content.find(
-      (b): b is { type: "text"; text: string } => b.type === "text",
-    );
-    return block?.text?.trim() ?? "";
-  },
 }));
 
 import {
@@ -100,7 +89,7 @@ import * as schema from "../../../../../persistence/schema/index.js";
 import type { Cluster, V1Edge, V1Item } from "../migration.js";
 
 // Dynamic import — runs *after* the mock.module calls above so the migration
-// module's transitive references to logger / jobs-store / provider-send-message
+// module's transitive references to logger / jobs-store / @vellumai/plugin-api
 // resolve to our stubs (matches the pattern used in `qdrant.test.ts`).
 const {
   clusterByTopic,

package/src/plugins/defaults/memory/v2/__tests__/router.test.ts

@@ -30,10 +30,9 @@ import type {
   Provider,
   ProviderResponse,
   SendMessageOptions,
-  ToolUseContent,
 } from "@vellumai/plugin-api";
 
-import type { ToolDefinition } from "../../../../../providers/types.js";
+import type { ToolDefinition } from "../../llm-helpers.js";
 
 // ---------------------------------------------------------------------------
 // Mocks installed BEFORE the router import so the module observes them at
@@ -106,10 +105,10 @@ interface ProviderCall {
 }
 const providerCalls: ProviderCall[] = [];
 
-mock.module("../../../../../providers/provider-send-message.js", () => ({
+// The router imports `getConfiguredProvider` from `@vellumai/plugin-api`; the
+// pure `extractToolUse` helper runs for real from the plugin's `llm-helpers`.
+mock.module("@vellumai/plugin-api", () => ({
   getConfiguredProvider: async () => providerStub,
-  extractToolUse: (response: ProviderResponse) =>
-    response.content.find((b): b is ToolUseContent => b.type === "tool_use"),
 }));
 
 // IDENTITY.md / users/default.md aren't required for these tests — the

package/src/plugins/defaults/memory/v2/__tests__/sweep-job.test.ts

@@ -38,7 +38,6 @@ import type {
   Provider,
   ProviderResponse,
   SendMessageOptions,
-  ToolUseContent,
 } from "@vellumai/plugin-api";
 
 import { makeMockLogger } from "../../../../../__tests__/helpers/mock-logger.js";
@@ -55,14 +54,8 @@ const providerCalls: Array<{
   userText: string;
 }> = [];
 
-mock.module("../../../../../providers/provider-send-message.js", () => ({
+mock.module("@vellumai/plugin-api", () => ({
   getConfiguredProvider: async () => providerStub,
-  userMessage: (text: string) => ({
-    role: "user" as const,
-    content: [{ type: "text" as const, text }],
-  }),
-  extractToolUse: (response: ProviderResponse) =>
-    response.content.find((b): b is ToolUseContent => b.type === "tool_use"),
 }));
 
 // emitNotificationSignal spy — captures every notification the sweep emits

package/src/plugins/defaults/memory/v2/activation.ts

@@ -35,11 +35,9 @@
 //     for the next turn and drop below `epsilon` if no longer relevant.
 
 import type { AssistantConfig } from "../../../../config/types.js";
-import {
-  embedWithBackend,
-  isEmbeddingDimensionAvailable,
-} from "../../../../persistence/embeddings/embedding-backend.js";
+import { isEmbeddingDimensionAvailable } from "../../../../persistence/embeddings/embedding-backend.js";
 import { applyCorrectionIfCalibrated } from "../anisotropy.js";
+import { embedWithBackend } from "../embeddings.js";
 import { clampUnitInterval } from "../validation.js";
 import type { EdgeIndex } from "./edge-index.js";
 import { hybridQueryConceptPages } from "./qdrant.js";

package/src/plugins/defaults/memory/v2/cli-command-store.ts

@@ -21,12 +21,10 @@
 //     summary.
 
 import { getConfig } from "../../../../config/loader.js";
-import {
-  embedWithBackend,
-  generateSparseEmbedding,
-} from "../../../../persistence/embeddings/embedding-backend.js";
+import { generateSparseEmbedding } from "../../../../persistence/embeddings/embedding-backend.js";
 import { getLogger } from "../../../../util/logger.js";
 import { applyCorrectionIfCalibrated } from "../anisotropy.js";
+import { embedWithBackend } from "../embeddings.js";
 import { buildCliCommandContent } from "./cli-command-content.js";
 import { invalidatePageIndex } from "./page-index.js";
 import {