@vellumai/assistant 0.7.3 → 0.8.1

package/src/memory/schema/conversations.ts

@@ -39,6 +39,8 @@ export const conversations = sqliteTable(
     lastMessageAt: integer("last_message_at"),
     archivedAt: integer("archived_at"),
     inferenceProfile: text("inference_profile"),
+    inferenceProfileSessionId: text("inference_profile_session_id"),
+    inferenceProfileExpiresAt: integer("inference_profile_expires_at"),
   },
   (table) => [
     index("idx_conversations_updated_at").on(table.updatedAt),

package/src/memory/schema/index.ts

@@ -1,8 +1,10 @@
 export * from "./acp.js";
+export * from "./bookmarks.js";
 export * from "./calls.js";
 export * from "./contacts.js";
 export * from "./conversations.js";
 export * from "./guardian.js";
+export * from "./inference.js";
 export * from "./infrastructure.js";
 export * from "./memory-core.js";
 export * from "./memory-graph.js";

package/src/memory/schema/inference.ts

@@ -0,0 +1,29 @@
+import { index, integer, sqliteTable, text } from "drizzle-orm/sqlite-core";
+
+/**
+ * Named provider connections.
+ *
+ * Each row is a named auth-config instance for a code-defined provider.
+ * Profiles in config.json reference connections by `name` via the
+ * `provider_connection` field.
+ *
+ * Created by migration 243.
+ */
+export const providerConnections = sqliteTable(
+  "provider_connections",
+  {
+    name: text("name").primaryKey(),
+    provider: text("provider").notNull(),
+    auth: text("auth").notNull(),
+    status: text("status").notNull().default("active"),
+    label: text("label"),
+    createdAt: integer("created_at").notNull(),
+    updatedAt: integer("updated_at").notNull(),
+  },
+  (table) => [
+    index("idx_provider_connections_provider").on(table.provider),
+  ],
+);
+
+export type ProviderConnectionRow = typeof providerConnections.$inferSelect;
+export type NewProviderConnectionRow = typeof providerConnections.$inferInsert;

package/src/memory/schema/memory-core.ts

@@ -100,6 +100,15 @@ export const memoryCheckpoints = sqliteTable("memory_checkpoints", {
   updatedAt: integer("updated_at").notNull(),
 });
 
+export const memoryRetrospectiveState = sqliteTable(
+  "memory_retrospective_state",
+  {
+    conversationId: text("conversation_id").primaryKey(),
+    lastProcessedMessageId: text("last_processed_message_id").notNull(),
+    lastRunAt: integer("last_run_at").notNull(),
+  },
+);
+
 export const conversationStarters = sqliteTable(
   "conversation_starters",
   {

package/src/memory/search/semantic.ts

@@ -1,7 +1,6 @@
 import { inArray } from "drizzle-orm";
 
 import { getConfig } from "../../config/loader.js";
-import { isMemoryV2ReadActive } from "../context-search/sources/memory-v2.js";
 import { getDb } from "../db-connection.js";
 import { withQdrantBreaker } from "../qdrant-circuit-breaker.js";
 import type {
@@ -10,6 +9,7 @@ import type {
 } from "../qdrant-client.js";
 import { getQdrantClient } from "../qdrant-client.js";
 import { memorySegments, memorySummaries } from "../schema.js";
+import { mapCosineToUnit } from "../validation.js";
 // ── Types (inlined from deleted types.ts) ──────────────────────────
 
 type CandidateType = "segment" | "item" | "summary" | "media";
@@ -56,10 +56,10 @@ export async function semanticSearch(
 ): Promise<Candidate[]> {
   if (limit <= 0) return [];
 
-  // v2 owns the read path when both gates are on; 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 (isMemoryV2ReadActive(getConfig())) return [];
+  // 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 [];
 
   const qdrant = getQdrantClient();
 
@@ -236,7 +236,3 @@ function buildHybridFilter(
     must_not: mustNotConditions,
   };
 }
-
-function mapCosineToUnit(value: number): number {
-  return Math.max(0, Math.min(1, (value + 1) / 2));
-}

package/src/memory/v2/__tests__/__snapshots__/prompts-router.test.ts.snap

@@ -0,0 +1,27 @@
+// Bun Snapshot v1, https://bun.sh/docs/test/snapshots
+
+exports[`renderRouterPrompt — determinism & snapshot stability snapshot of fixed inputs 1`] = `
+"You are a background helper for Aria. Your job is to route memory pages for the next assistant turn between Aria and Alice.
+
+You will be shown the recent conversation, a \`<now>\` marker for the current time, an \`<already_injected_ids>\` block listing pages picked on the previous turn, and a \`# Concept Page Index\` listing every routable page on this workspace.
+
+Pick the concept pages whose contents would help Aria respond well on this turn. Lean toward inclusion when in doubt — missing a relevant page is a worse error than surfacing a few unused ones, because the assistant can ignore extras but can't summon context that wasn't loaded. Abstain (return an empty list) only when nothing in the index plausibly bears on the turn.
+
+Index format. Each line of the index has the shape:
+
+    [id] slug — summary (edges: a, b, c)
+
+\`id\` is a small integer used to refer to this page. \`edges\` are numeric IDs into the same list, pointing at related pages; you may follow them when one page strongly implies another.
+
+Already-injected pages. Pages whose IDs appear in \`<already_injected_ids>\` were picked on the previous turn. Do not pick them again unless Aria should re-anchor on that material — e.g., the topic genuinely returns after drifting away. Routine continuity does not require re-picking; the prior turn's pages are already in the assistant's working context.
+
+Time. Bias toward pages that match the current state implied by \`<now>\` and the active conversational threads (what is happening today, what was just decided, who is being discussed). Stale pages with no bearing on the live conversation should be skipped even if their summaries look superficially relevant.
+
+Emit your selection by calling \`select_pages_to_inject\` with the chosen \`page_ids\`. Return an empty array to abstain.
+
+# Concept Page Index
+
+[1] morning-routine — coffee, walk, journal (edges: 2)
+[2] journal-style — terse, dated, no fluff (edges: 1)
+[3] taxes-2025 — Q1 estimate due April 15 (edges: )"
+`;

package/src/memory/v2/__tests__/activation-store.test.ts

@@ -8,7 +8,7 @@ import { migrateActivationState } from "../../migrations/232-activation-state.js
 import * as schema from "../../schema.js";
 import {
   evictCompactedTurns,
-  fork,
+  forkActivationState,
   hydrate,
   save,
 } from "../activation-store.js";
@@ -113,12 +113,12 @@ describe("activation-store", () => {
     });
   });
 
-  describe("fork", () => {
+  describe("forkActivationState", () => {
     test("copies parent state to a new conversation id", async () => {
       const parentState = buildState();
       await save(db, "conv-parent", parentState);
 
-      await fork(db, "conv-parent", "conv-child");
+      forkActivationState(db, "conv-parent", "conv-child");
 
       const child = await hydrate(db, "conv-child");
       expect(child).toEqual(parentState);
@@ -129,7 +129,7 @@ describe("activation-store", () => {
     });
 
     test("is a no-op when the parent has no state", async () => {
-      await fork(db, "conv-parent-missing", "conv-child");
+      forkActivationState(db, "conv-parent-missing", "conv-child");
 
       expect(await hydrate(db, "conv-child")).toBeNull();
     });
@@ -139,7 +139,7 @@ describe("activation-store", () => {
       await save(db, "conv-parent", parentState);
       await save(db, "conv-child", buildState({ currentTurn: 99 }));
 
-      await fork(db, "conv-parent", "conv-child");
+      forkActivationState(db, "conv-parent", "conv-child");
 
       const child = await hydrate(db, "conv-child");
       expect(child?.currentTurn).toBe(7);

package/src/memory/v2/__tests__/activation.test.ts

@@ -114,7 +114,10 @@ class MockQdrantClient {
       limit: params.limit,
       filter: params.filter,
     });
-    const channel = params.using as "dense" | "sparse";
+    // The four-channel hybrid query fires body-dense, body-sparse,
+    // summary-dense, summary-sparse in order; both dense channels share
+    // the dense queue and both sparse channels share the sparse queue.
+    const channel = params.using.endsWith("sparse") ? "sparse" : "dense";
     return state.queryResponses[channel].shift() ?? { points: [] };
   }
 }
@@ -223,9 +226,20 @@ function makeConfig(
   } as unknown as AssistantConfig;
 }
 
-/** Stage a single dense + sparse pair on the response queues. */
+/**
+ * Stage a single hybrid-query response — body channels first, then summary
+ * channels (which default to empty). The four-channel hybrid query fires
+ * body-dense, body-sparse, summary-dense, summary-sparse in that order, so
+ * each logical call consumes 2 dense + 2 sparse queue entries.
+ */
 function stageHybridResponse(
-  hits: Array<{ slug: string; denseScore?: number; sparseScore?: number }>,
+  hits: Array<{
+    slug: string;
+    denseScore?: number;
+    sparseScore?: number;
+    summaryDenseScore?: number;
+    summarySparseScore?: number;
+  }>,
 ): void {
   state.queryResponses.dense.push({
     points: hits
@@ -237,6 +251,22 @@ function stageHybridResponse(
       .filter((h) => h.sparseScore !== undefined)
       .map((h) => ({ score: h.sparseScore, payload: { slug: h.slug } })),
   });
+  state.queryResponses.dense.push({
+    points: hits
+      .filter((h) => h.summaryDenseScore !== undefined)
+      .map((h) => ({
+        score: h.summaryDenseScore,
+        payload: { slug: h.slug },
+      })),
+  });
+  state.queryResponses.sparse.push({
+    points: hits
+      .filter((h) => h.summarySparseScore !== undefined)
+      .map((h) => ({
+        score: h.summarySparseScore,
+        payload: { slug: h.slug },
+      })),
+  });
 }
 
 beforeEach(resetState);
@@ -369,7 +399,7 @@ describe("selectCandidates", () => {
       nowText: "",
       config: makeConfig(),
     });
-    expect(state.queryCalls).toHaveLength(2);
+    expect(state.queryCalls).toHaveLength(4);
     for (const call of state.queryCalls) {
       expect(call.limit).toBe(1_000_000);
       expect(call.filter).toBeUndefined();
@@ -385,7 +415,7 @@ describe("selectCandidates", () => {
       nowText: "",
       config: makeConfig({ ann_candidate_limit: 25 }),
     });
-    expect(state.queryCalls).toHaveLength(2);
+    expect(state.queryCalls).toHaveLength(4);
     for (const call of state.queryCalls) {
       expect(call.limit).toBe(25);
       expect(call.filter).toBeUndefined();
@@ -440,7 +470,9 @@ describe("computeOwnActivation", () => {
         c_now: 0.2,
       }),
     });
-    // Expected: 0.3*0.6 + 0.3*0.5 + 0.2*0.4 + 0.2*0.2 = 0.18+0.15+0.08+0.04 = 0.45
+    // Positive cosines pass through unchanged (negatives clamp to 0).
+    // Expected: 0.3*0.6 + 0.3*0.5 + 0.2*0.4 + 0.2*0.2
+    //         = 0.18 + 0.15 + 0.08 + 0.04 = 0.45
     expect(out.activation.get("alice")).toBeCloseTo(0.45, 6);
   });
 
@@ -493,6 +525,7 @@ describe("computeOwnActivation", () => {
         c_now: 0.2,
       }),
     });
+    // Positive cosines pass through unchanged: 1.0 stays 1.0; 0 stays 0.
     // 0.3*0 + 0.3*1 + 0.2*0 + 0.2*0 = 0.3
     expect(out.activation.get("fresh")).toBeCloseTo(0.3, 6);
   });
@@ -543,7 +576,8 @@ describe("computeOwnActivation", () => {
     expect(breakdown).toBeDefined();
     // priorContribution is `d * prev`, not the weighted sim term.
     expect(breakdown?.priorContribution).toBeCloseTo(d * 0.6, 6);
-    // Raw sims, captured before c_user / c_assistant / c_now weighting.
+    // Fused sims (positive cosines pass through unchanged), captured before
+    // c_user / c_assistant / c_now weighting.
     expect(breakdown?.simUser).toBeCloseTo(0.5, 6);
     expect(breakdown?.simAssistant).toBeCloseTo(0.4, 6);
     expect(breakdown?.simNow).toBeCloseTo(0.2, 6);
@@ -771,8 +805,10 @@ describe("computeOwnActivation", () => {
   });
 
   test("rerank boost is additive on A_o and leaves raw simUser / simAssistant untouched", async () => {
-    stageHybridResponse([{ slug: "a", denseScore: 0.5 }]); // user
-    stageHybridResponse([{ slug: "a", denseScore: 0.4 }]); // assistant
+    // Positive cosines pass through unchanged; pick 0.5 / 0.4 directly so
+    // the round fused values used in the formula assertion below hold.
+    stageHybridResponse([{ slug: "a", denseScore: 0.5 }]); // user — fused 0.5
+    stageHybridResponse([{ slug: "a", denseScore: 0.4 }]); // assistant — fused 0.4
     stageHybridResponse([]); // now
     rerankState.scores = new Map([["a", 0.8]]);
 
@@ -1189,6 +1225,7 @@ describe("skills participate in the unified pipeline", () => {
     });
 
     // No prior state → priorContribution = 0.
+    // Positive cosines pass through unchanged.
     // 0.3*0.5 + 0.2*0.4 + 0.2*0.2 = 0.15 + 0.08 + 0.04 = 0.27
     expect(out.activation.get("skills/example-skill-a")).toBeCloseTo(0.27, 6);
     expect(out.breakdown.get("skills/example-skill-a")?.priorContribution).toBe(

package/src/memory/v2/__tests__/backfill-jobs.test.ts

@@ -8,8 +8,7 @@
  * Coverage matrix:
  *   - migrate: wraps `runMemoryV2Migration`; force flag propagates;
  *     `MigrationAlreadyAppliedError` is swallowed (no rethrow).
- *   - reembed: enqueues `N + 4` jobs (concept-page slugs plus four reserved
- *     meta-file slugs).
+ *   - reembed: enqueues `N` jobs, one per concept-page slug.
  *   - activation-recompute: walks conversations with rows, runs the pipeline
  *     end-to-end against the real activation module, persists fresh state.
  *
@@ -116,7 +115,10 @@ const STUB_RUNTIME_CONFIG = {
 };
 mock.module("../../../config/loader.js", () => ({
   getConfig: () => STUB_RUNTIME_CONFIG,
+  getConfigReadOnly: () => STUB_RUNTIME_CONFIG,
   loadConfig: () => STUB_RUNTIME_CONFIG,
+  loadRawConfig: () => ({}) as Record<string, unknown>,
+  saveRawConfig: () => {},
   invalidateConfigCache: () => {},
   applyNestedDefaults: () => STUB_RUNTIME_CONFIG,
 }));
@@ -208,7 +210,6 @@ const { writePage } = await import("../page-store.js");
 const { save: saveActivation, hydrate: hydrateActivation } =
   await import("../activation-store.js");
 const {
-  META_FILE_SLUGS,
   memoryV2ActivationRecomputeJob,
   memoryV2MigrateJob,
   memoryV2ReembedJob,
@@ -322,15 +323,15 @@ describe("memoryV2MigrateJob", () => {
 // ---------------------------------------------------------------------------
 
 describe("memoryV2ReembedJob", () => {
-  test("returns N + 4 (one per concept page plus the four meta files) and writes that many job rows", async () => {
+  test("returns N (one per concept page) and writes that many job rows", async () => {
     await writePage(tmpWorkspace, {
       slug: "alice",
-      frontmatter: { edges: [], ref_files: [] },
+      frontmatter: { edges: [], ref_files: [], ref_urls: [] },
       body: "Alice.\n",
     });
     await writePage(tmpWorkspace, {
       slug: "bob",
-      frontmatter: { edges: [], ref_files: [] },
+      frontmatter: { edges: [], ref_files: [], ref_urls: [] },
       body: "Bob.\n",
     });
 
@@ -339,8 +340,8 @@ describe("memoryV2ReembedJob", () => {
       TEST_CONFIG,
     );
 
-    // Return value covers the contract: N concept pages + 4 meta files.
-    expect(total).toBe(2 + META_FILE_SLUGS.length);
+    // Return value covers the contract: one job per concept page.
+    expect(total).toBe(2);
 
     // Verify the slugs that were enqueued by reading the memory_jobs table.
     // Tests that mock `jobs-store.js` skip inserting rows; when this suite
@@ -349,34 +350,50 @@ describe("memoryV2ReembedJob", () => {
     // belt-and-suspenders.
     const rows = getDb().select().from(memoryJobs).all();
     if (rows.length > 0) {
-      expect(rows).toHaveLength(2 + META_FILE_SLUGS.length);
+      expect(rows).toHaveLength(2);
       const slugs = rows.map((row) => JSON.parse(row.payload).slug);
       expect(slugs).toContain("alice");
       expect(slugs).toContain("bob");
-      for (const metaSlug of META_FILE_SLUGS) {
-        expect(slugs).toContain(metaSlug);
-      }
       for (const row of rows) {
         expect(row.type).toBe("embed_concept_page");
       }
     }
   });
 
-  test("with no concept pages on disk, still enqueues the 4 meta-file jobs", async () => {
+  test("with no concept pages on disk, enqueues nothing", async () => {
     const total = await memoryV2ReembedJob(
       makeJob("memory_v2_reembed"),
       TEST_CONFIG,
     );
-    expect(total).toBe(META_FILE_SLUGS.length);
+    expect(total).toBe(0);
   });
 
-  test("uses reserved meta-file slugs (__essentials__/__threads__/__recent__/__buffer__)", () => {
-    expect([...META_FILE_SLUGS]).toEqual([
-      "__essentials__",
-      "__threads__",
-      "__recent__",
-      "__buffer__",
-    ]);
+  test("does NOT enqueue reserved meta-file slugs", async () => {
+    // The four prose meta files (essentials/threads/recent/buffer) live at
+    // `memory/<name>.md` and are direct-injected into the system prompt via
+    // `_autoinject.md`. Their underscore-bracketed slug aliases (e.g.
+    // `__essentials__`) fail the concept-page slug validator
+    // (`[a-z0-9][a-z0-9-]*`), so the reembed fan-out must not enqueue them.
+    await writePage(tmpWorkspace, {
+      slug: "alice",
+      frontmatter: { edges: [], ref_files: [], ref_urls: [] },
+      body: "Alice.\n",
+    });
+
+    await memoryV2ReembedJob(makeJob("memory_v2_reembed"), TEST_CONFIG);
+
+    const rows = getDb().select().from(memoryJobs).all();
+    if (rows.length > 0) {
+      const slugs = rows.map((row) => JSON.parse(row.payload).slug);
+      for (const reserved of [
+        "__essentials__",
+        "__threads__",
+        "__recent__",
+        "__buffer__",
+      ]) {
+        expect(slugs).not.toContain(reserved);
+      }
+    }
   });
 });