@vellumai/assistant 0.5.4 → 0.5.6

package/src/memory/job-handlers/conversation-starters.ts

@@ -176,11 +176,11 @@ async function generateStarters(scopeId: string): Promise<GeneratedStarter[]> {
     ? truncate(rawIdentityContext, 2000, "\n…[truncated]")
     : null;
 
-  const systemPrompt = `You are generating 4 conversation starters for a personal assistant app. These appear as clickable chips on the empty conversation page — the first thing the user sees when they open the app.
+  const systemPrompt = `You are generating 4 conversation starters for a personal assistant app. These appear as clickable chips on the empty conversation page — the first thing the user sees when they open the app. Clicking a chip sends its prompt as a message from the user.
 
 ${timeContext}
 
-Your goal: look at what's going on in this person's life right now and suggest the 4 most useful things they could ask you to do. Think about what a thoughtful chief of staff would proactively bring up in a 30-second check-in.
+Your goal: suggest the 4 most useful things this person could ask you to do right now.
 
 ${identityContext ? `## Assistant identity & user profile\n\n${identityContext}\n\n` : ""}## What you know
 
@@ -188,7 +188,9 @@ ${rollup}
 ${diff}
 ${skills}
 
-## How to think about this
+## Selection
+
+Generate exactly 4 starters, ranked #1 (best) to #4.
 
 Start from the user's situation, not from the skill list. Ask yourself:
 - What is this person likely dealing with right now (given the day/time and their context)?
@@ -197,11 +199,7 @@ Start from the user's situation, not from the skill list. Ask yourself:
 
 The skills list tells you what the assistant CAN do — use it to filter out suggestions the assistant can't actually help with, not as a menu to generate suggestions from.
 
-## Selection
-
-Generate exactly 4 starters, ranked #1 (best) to #4.
-
-For each, you must be able to clearly answer:
+For each starter, you must clearly answer:
 - Why now? (timing — day of week, recent activity, upcoming deadline)
 - Why this user? (grounded in their specific context, not generic)
 - Why would they be glad I suggested this? (genuine usefulness, not just relevance)
@@ -218,44 +216,34 @@ Favor what is live over what is merely true. Recent changes matter more than old
 
 ## Output format
 
-Return exactly 4 starters in rank order (best first).
-
 Each starter has:
-- label: 3-6 words, max 40 chars, starts with a verb. Should read as something the user wants to do — these chips send a message as the user, so the label must be in the user's voice. Must sound natural when read aloud.
-- prompt: 1-2 natural sentences, written as the user would actually say them — not templated.
+- label: 3-6 words, max 40 chars, starts with a verb. Written in the user's voice — something they'd want to do, not something the assistant is offering.
+- prompt: 1-2 natural sentences, as the user would actually say them.
 - category: one of ${CONVERSATION_STARTER_CATEGORIES.join(", ")}
 
-The 4 starters should feel like one coherent set of recommendations for this moment — similar abstraction level, no jarring mix of mundane chores and life strategy. Don't lift raw memory phrases, project names, or jargon into labels unless they already sound natural in conversation.
-
-Never include a chip whose primary meaning is configuration, setup, workflow creation, or "set up X for Y" unless it solves an urgent pain the user is actively feeling right now. Prefer the outcome over the mechanism — "Catch the emails that matter" beats "Set up a playbook for inbox."
-
-## Topic diversity
-
-Each chip should cover a distinct topic or concern. Never have two chips about the same tool, project, or theme — even if there are multiple related issues. Pick the single most impactful angle and give the other slot to something different. Four chips about three topics is too narrow; four chips about four topics is right.
+## Constraints
 
-## User-facingness check
+**Voice**: The user clicks these chips to send a message. Every label must read as something the user is asking to do, never something the assistant is saying to the user.
 
-If a label sounds like an issue title, project ticket, or implementation task, rewrite it. Prefer the user-visible payoff over the internal object name. The chip should feel inviting and useful, not merely accurate.
+**Coherence**: The 4 starters should feel like one set — similar abstraction level, no jarring mix of mundane chores and life strategy.
 
-Prefer natural, flowing language over mechanical or operational phrasing. "Get Slack messages flowing" is better than "Restore outgoing Slack messages." The label should sound like something a helpful person would say, not a support ticket.
+**Diversity**: Each chip covers a distinct topic. Never two chips about the same tool, project, or theme. Four topics, four chips.
 
-Voice: The user clicks these chips to send a message. Every label must make sense as something the user is asking to do, never something the assistant is saying to the user.
+**No setup chips**: Never include a chip whose primary meaning is configuration or "set up X for Y" unless it solves an urgent pain the user is actively feeling. Prefer the outcome over the mechanism.
 
-Before finalizing each label, ask yourself: would this feel good to click? Or does it sound like a backlog item? If it sounds like a backlog item, rewrite it.
+**Natural language**: No jargon, project names, or raw memory phrases in labels unless they already sound natural in conversation. If a label sounds like a ticket title or backlog item, rewrite it as something the user would actually say.
 
-Examples of bad vs good:
-- BAD: "Fix Slack Socket Mode blocker" → GOOD: "Fix Slack so it just works"
-- BAD: "Rewire messaging for Socket Mode" → GOOD: "Get Socket Mode stable"
-- BAD: "Review this week's calendar" → GOOD: "Protect this week's focus"
-- BAD: "Model the coaching transition" → GOOD: "Plan the coaching transition"
-- BAD: "Restore outgoing Slack messages" → GOOD: "Get Slack messages flowing"
-- BAD: "Set up a playbook for inbox" → GOOD: "Catch the emails that matter"
+## Examples
 
-Assistant-voice vs user-voice:
-- BAD: "You've got a busy week ahead" → GOOD: "Plan my week ahead"
-- BAD: "Let me check your calendar" → GOOD: "Check my Thursday schedule"
+Bad → Good (ticket-speak → natural):
+- "Fix Slack Socket Mode blocker" → "Fix Slack so it just works"
+- "Restore outgoing Slack messages" → "Get Slack messages flowing"
+- "Review this week's calendar" → "Protect this week's focus"
+- "Set up a playbook for inbox" → "Triage my inbox"
 
-The good versions emphasize the user's payoff in the user's own voice, not the internal mechanism or the assistant's perspective.`;
+Bad → Good (assistant voice → user voice):
+- "You've got a busy week ahead" → "Plan my week ahead"
+- "Let me check your calendar" → "Check my Thursday schedule"`;
 
   const { signal, cleanup } = createTimeout(20000);
   try {
@@ -280,7 +268,7 @@ The good versions emphasize the user's payoff in the user's own voice, not the i
                     label: {
                       type: "string",
                       description:
-                        "User-voice chip text (2-7 words, max 40 chars, starts with a verb)",
+                        "User-voice chip label (2-7 words, max 40 chars, verb-first)",
                     },
                     prompt: {
                       type: "string",

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

@@ -11,10 +11,7 @@ import type { MemoryJob } from "../jobs-store.js";
 import { extractMediaBlocks } from "../message-content.js";
 import {
   mediaAssets,
-  memoryChunks,
-  memoryEpisodes,
   memoryItems,
-  memoryObservations,
   memorySegments,
   memorySummaries,
   messages,
@@ -93,26 +90,6 @@ export async function embedSummaryJob(
   );
 }
 
-export async function embedChunkJob(
-  job: MemoryJob,
-  config: AssistantConfig,
-): Promise<void> {
-  const chunkId = asString(job.payload.chunkId);
-  if (!chunkId) return;
-  const db = getDb();
-  const chunk = db
-    .select()
-    .from(memoryChunks)
-    .where(eq(memoryChunks.id, chunkId))
-    .get();
-  if (!chunk) return;
-  await embedAndUpsert(config, "chunk", chunk.id, chunk.content, {
-    observation_id: chunk.observationId,
-    created_at: chunk.createdAt,
-    memory_scope_id: chunk.scopeId,
-  });
-}
-
 export async function embedMediaJob(
   job: MemoryJob,
   config: AssistantConfig,
@@ -146,40 +123,6 @@ export async function embedMediaJob(
   });
 }
 
-export async function embedObservationJob(
-  job: MemoryJob,
-  config: AssistantConfig,
-): Promise<void> {
-  const observationId = asString(job.payload.observationId);
-  const chunkId = asString(job.payload.chunkId);
-  if (!observationId || !chunkId) return;
-
-  const db = getDb();
-  const observation = db
-    .select()
-    .from(memoryObservations)
-    .where(eq(memoryObservations.id, observationId))
-    .get();
-  if (!observation) return;
-
-  const chunk = db
-    .select()
-    .from(memoryChunks)
-    .where(eq(memoryChunks.id, chunkId))
-    .get();
-  if (!chunk) return;
-
-  await embedAndUpsert(config, "observation", chunk.id, chunk.content, {
-    observation_id: observationId,
-    conversation_id: observation.conversationId,
-    role: observation.role,
-    modality: observation.modality,
-    source: observation.source,
-    created_at: observation.createdAt,
-    memory_scope_id: observation.scopeId,
-  });
-}
-
 export async function embedAttachmentJob(
   job: MemoryJob,
   config: AssistantConfig,
@@ -216,25 +159,3 @@ export async function embedAttachmentJob(
     memory_scope_id: memoryScopeId,
   });
 }
-
-export async function embedEpisodeJob(
-  job: MemoryJob,
-  config: AssistantConfig,
-): Promise<void> {
-  const episodeId = asString(job.payload.episodeId);
-  if (!episodeId) return;
-  const db = getDb();
-  const episode = db
-    .select()
-    .from(memoryEpisodes)
-    .where(eq(memoryEpisodes.id, episodeId))
-    .get();
-  if (!episode) return;
-  const text = `[episode] ${episode.title}: ${episode.summary}`;
-  await embedAndUpsert(config, "episode", episode.id, text, {
-    conversation_id: episode.conversationId,
-    created_at: episode.startAt,
-    last_seen_at: episode.endAt,
-    memory_scope_id: episode.scopeId,
-  });
-}

package/src/memory/job-utils.ts

@@ -142,7 +142,7 @@ export function truncate(text: string, max: number): string {
 
 export async function embedAndUpsert(
   config: AssistantConfig,
-  targetType: "segment" | "item" | "summary" | "observation" | "chunk" | "episode" | "media",
+  targetType: "segment" | "item" | "summary" | "media",
   targetId: string,
   input: EmbeddingInput,
   extraPayload?: Record<string, unknown>,

package/src/memory/jobs-store.ts

@@ -12,9 +12,6 @@ export type MemoryJobType =
   | "embed_segment"
   | "embed_item"
   | "embed_summary"
-  | "embed_chunk"
-  | "embed_episode"
-  | "embed_observation"
   | "extract_items"
   | "extract_entities"
   | "cleanup_stale_superseded_items"
@@ -30,8 +27,6 @@ export type MemoryJobType =
   | "embed_media"
   | "embed_attachment"
   | "generate_conversation_starters"
-  | "reduce_conversation_memory"
-  | "backfill_simplified_memory"
   | "generate_capability_cards" // legacy compat — silently dropped by worker (capability cards removed)
   | "generate_thread_starters"; // legacy compat — silently dropped by worker (renamed to generate_conversation_starters)
 
@@ -39,9 +34,6 @@ const EMBED_JOB_TYPES: MemoryJobType[] = [
   "embed_segment",
   "embed_item",
   "embed_summary",
-  "embed_chunk",
-  "embed_episode",
-  "embed_observation",
   "embed_media",
   "embed_attachment",
 ];

package/src/memory/jobs-worker.ts

@@ -3,7 +3,6 @@ import type { AssistantConfig } from "../config/types.js";
 import { getLogger } from "../util/logger.js";
 import { rawRun } from "./db.js";
 import { backfillJob } from "./job-handlers/backfill.js";
-import { backfillSimplifiedMemoryJob } from "./job-handlers/backfill-simplified-memory.js";
 import {
   cleanupStaleSupersededItemsJob,
   pruneOldConversationsJob,
@@ -12,11 +11,8 @@ import { generateConversationStartersJob } from "./job-handlers/conversation-sta
 // ── Per-job-type handlers ──────────────────────────────────────────
 import {
   embedAttachmentJob,
-  embedChunkJob,
-  embedEpisodeJob,
   embedItemJob,
   embedMediaJob,
-  embedObservationJob,
   embedSegmentJob,
   embedSummaryJob,
 } from "./job-handlers/embedding.js";
@@ -26,7 +22,6 @@ import {
   rebuildIndexJob,
 } from "./job-handlers/index-maintenance.js";
 import { mediaProcessingJob } from "./job-handlers/media-processing.js";
-import { reduceConversationMemoryJob } from "./job-handlers/reduce-conversation-memory.js";
 import { buildConversationSummaryJob } from "./job-handlers/summarization.js";
 import {
   BackendUnavailableError,
@@ -272,15 +267,6 @@ async function processJob(
     case "embed_summary":
       await embedSummaryJob(job, config);
       return;
-    case "embed_chunk":
-      await embedChunkJob(job, config);
-      return;
-    case "embed_episode":
-      await embedEpisodeJob(job, config);
-      return;
-    case "embed_observation":
-      await embedObservationJob(job, config);
-      return;
     case "extract_items":
       await extractItemsJob(job);
       return;
@@ -321,12 +307,6 @@ async function processJob(
     case "embed_attachment":
       await embedAttachmentJob(job, config);
       return;
-    case "reduce_conversation_memory":
-      await reduceConversationMemoryJob(job);
-      return;
-    case "backfill_simplified_memory":
-      await backfillSimplifiedMemoryJob(job);
-      return;
     case "generate_conversation_starters":
       await generateConversationStartersJob(job);
       return;

package/src/memory/migrations/189-drop-simplified-memory.ts

@@ -0,0 +1,42 @@
+import type { DrizzleDb } from "../db-connection.js";
+import { getSqliteFrom } from "../db-connection.js";
+
+/**
+ * Drop simplified-memory tables and reducer checkpoint columns added by
+ * the simplified-memory-v1 plan, reverting to the legacy item/tier/XML
+ * memory system.
+ */
+export function migrateDropSimplifiedMemory(database: DrizzleDb): void {
+  const raw = getSqliteFrom(database);
+
+  // Drop simplified-memory tables (idempotent — IF EXISTS).
+  raw.exec(`DROP TABLE IF EXISTS time_contexts`);
+  raw.exec(`DROP TABLE IF EXISTS open_loops`);
+  raw.exec(`DROP TABLE IF EXISTS memory_observations`);
+  raw.exec(`DROP TABLE IF EXISTS memory_chunks`);
+  raw.exec(`DROP TABLE IF EXISTS memory_episodes`);
+
+  // Remove reducer checkpoint columns from conversations.
+  // SQLite doesn't support DROP COLUMN before 3.35.0, but Bun's built-in
+  // SQLite is >= 3.38, so this is safe.
+  for (const col of [
+    "memory_reduced_through_message_id",
+    "memory_dirty_tail_since_message_id",
+    "memory_last_reduced_at",
+  ]) {
+    try {
+      raw.exec(`ALTER TABLE conversations DROP COLUMN ${col}`);
+    } catch {
+      // Column doesn't exist — already cleaned up.
+    }
+  }
+
+  // Remove embedding rows for archive target types that no longer exist.
+  try {
+    raw.exec(
+      `DELETE FROM memory_embeddings WHERE target_type IN ('observation', 'chunk', 'episode')`,
+    );
+  } catch {
+    // Column doesn't exist — table was never migrated to include target_type.
+  }
+}

package/src/memory/migrations/index.ts

@@ -126,10 +126,8 @@ export { migrateRenameThreadStartersCheckpoints } from "./181-rename-thread-star
 export { migrateOAuthProvidersDisplayMetadata } from "./182-oauth-providers-display-metadata.js";
 export { migrateConversationForkLineage } from "./183-add-conversation-fork-lineage.js";
 export { migrateLlmRequestLogProvider } from "./184-llm-request-log-provider.js";
-export { migrateMemoryBriefState } from "./185-memory-brief-state.js";
-export { migrateMemoryArchiveTables } from "./186-memory-archive.js";
-export { migrateMemoryReducerCheckpoints } from "./187-memory-reducer-checkpoints.js";
 export { migrateScheduleQuietFlag } from "./188-schedule-quiet-flag.js";
+export { migrateDropSimplifiedMemory } from "./189-drop-simplified-memory.js";
 export {
   MIGRATION_REGISTRY,
   type MigrationRegistryEntry,

package/src/memory/qdrant-client.ts

@@ -20,7 +20,7 @@ export interface QdrantClientConfig {
 }
 
 export interface QdrantPointPayload {
-  target_type: "segment" | "item" | "summary" | "observation" | "chunk" | "episode" | "media";
+  target_type: "segment" | "item" | "summary" | "media";
   target_id: string;
   text: string;
   kind?: string;
@@ -230,7 +230,7 @@ export class VellumQdrantClient {
   }
 
   async upsert(
-    targetType: "segment" | "item" | "summary" | "observation" | "chunk" | "episode" | "media",
+    targetType: "segment" | "item" | "summary" | "media",
     targetId: string,
     vector: number[],
     payload: Omit<QdrantPointPayload, "target_type" | "target_id">,
@@ -324,9 +324,7 @@ export class VellumQdrantClient {
   async searchWithFilter(
     vector: number[],
     limit: number,
-    targetTypes: Array<
-      "segment" | "item" | "summary" | "media" | "chunk" | "episode"
-    >,
+    targetTypes: Array<"segment" | "item" | "summary" | "media">,
     excludeMessageIds?: string[],
     scopeIds?: string[],
   ): Promise<QdrantSearchResult[]> {
@@ -349,7 +347,7 @@ export class VellumQdrantClient {
           },
           {
             key: "target_type",
-            match: { any: ["segment", "summary", "media", "chunk"] },
+            match: { any: ["segment", "summary", "media"] },
           },
         ],
       });

package/src/memory/schema/conversations.ts

@@ -30,9 +30,6 @@ export const conversations = sqliteTable(
     forkParentMessageId: text("fork_parent_message_id"),
     isAutoTitle: integer("is_auto_title").notNull().default(1),
     scheduleJobId: text("schedule_job_id"),
-    memoryReducedThroughMessageId: text("memory_reduced_through_message_id"),
-    memoryDirtyTailSinceMessageId: text("memory_dirty_tail_since_message_id"),
-    memoryLastReducedAt: integer("memory_last_reduced_at"),
   },
   (table) => [
     index("idx_conversations_updated_at").on(table.updatedAt),

package/src/memory/schema/index.ts

@@ -3,8 +3,6 @@ export * from "./contacts.js";
 export * from "./conversations.js";
 export * from "./guardian.js";
 export * from "./infrastructure.js";
-export * from "./memory-archive.js";
-export * from "./memory-brief.js";
 export * from "./memory-core.js";
 export * from "./notifications.js";
 export * from "./oauth.js";

package/src/messaging/draft-store.ts

@@ -10,7 +10,7 @@ import { readdirSync, readFileSync, unlinkSync, writeFileSync } from "node:fs";
 import { join } from "node:path";
 
 import { ensureDir, pathExists } from "../util/fs.js";
-import { getRootDir } from "../util/platform.js";
+import { getWorkspaceDir } from "../util/platform.js";
 
 export interface Draft {
   id: string;
@@ -25,7 +25,7 @@ export interface Draft {
 }
 
 function getDraftsDir(platform: string): string {
-  const dir = join(getRootDir(), "workspace", "data", "drafts", platform);
+  const dir = join(getWorkspaceDir(), "data", "drafts", platform);
   ensureDir(dir);
   return dir;
 }

package/src/messaging/provider.ts

@@ -89,6 +89,15 @@ export interface MessagingProvider {
    */
   isConnected?(): Promise<boolean>;
 
+  /**
+   * Custom credential resolution for providers with non-standard credential
+   * paths (e.g. Slack Socket Mode stores tokens under "slack_channel" rather
+   * than the OAuth provider key). When present, getProviderConnection() calls
+   * this instead of resolveOAuthConnection(), giving the provider full control
+   * over credential lookup including fallback strategies.
+   */
+  resolveConnection?(account?: string): Promise<OAuthConnection | string>;
+
   /** Platform-specific capabilities for tool routing (e.g. 'reactions', 'threads', 'labels'). */
   capabilities: Set<string>;
 }

package/src/messaging/providers/slack/adapter.ts

@@ -1,11 +1,15 @@
 /**
  * Slack messaging provider adapter.
  *
- * Maps Slack API responses to the platform-agnostic messaging types
- * and implements the MessagingProvider interface.
+ * Maps Slack API responses to the platform-agnostic messaging types and
+ * implements the MessagingProvider interface.
  */
 
 import type { OAuthConnection } from "../../../oauth/connection.js";
+import { resolveOAuthConnection } from "../../../oauth/connection-resolver.js";
+import { isProviderConnected } from "../../../oauth/oauth-store.js";
+import { credentialKey } from "../../../security/credential-key.js";
+import { getSecureKeyAsync } from "../../../security/secure-keys.js";
 import type { MessagingProvider } from "../../provider.js";
 import type {
   ConnectionInfo,
@@ -112,6 +116,29 @@ export const slackProvider: MessagingProvider = {
   credentialService: "integration:slack",
   capabilities: new Set(["reactions", "threads", "leave_channel"]),
 
+  async isConnected(): Promise<boolean> {
+    // Socket Mode: check for bot token directly in credential store.
+    // The token is the source of truth; the slack_channel connection row
+    // is advisory (backfill can fail non-fatally on startup).
+    const botToken = await getSecureKeyAsync(
+      credentialKey("slack_channel", "bot_token"),
+    );
+    if (botToken) return true;
+    // Preserve existing OAuth path (integration:slack) for backwards compat.
+    return isProviderConnected("integration:slack");
+  },
+
+  async resolveConnection(account?: string): Promise<OAuthConnection | string> {
+    // Socket Mode: return raw bot token if available.
+    // Token presence is sufficient — no connection row required.
+    const botToken = await getSecureKeyAsync(
+      credentialKey("slack_channel", "bot_token"),
+    );
+    if (botToken) return botToken;
+    // Preserve existing OAuth path (integration:slack) for backwards compat.
+    return resolveOAuthConnection("integration:slack", { account });
+  },
+
   async testConnection(
     connectionOrToken: OAuthConnection | string,
   ): Promise<ConnectionInfo> {

package/src/oauth/connection-resolver.test.ts

@@ -8,12 +8,7 @@ let mockProvider: Record<string, unknown> | undefined;
 let mockConnection: Record<string, unknown> | undefined;
 let mockAccessToken: string | undefined;
 let mockConfig: Record<string, unknown> = {};
-let mockManagedProxyCtx = {
-  enabled: false,
-  platformBaseUrl: "",
-  assistantApiKey: "",
-};
-let mockAssistantId = "";
+let mockPlatformClient: Record<string, unknown> | null = null;
 
 // ---------------------------------------------------------------------------
 // Module mocks (must precede imports of the module under test)
@@ -48,12 +43,10 @@ mock.module("../config/loader.js", () => ({
   getConfig: () => mockConfig,
 }));
 
-mock.module("../config/env.js", () => ({
-  getPlatformAssistantId: () => mockAssistantId,
-}));
-
-mock.module("../providers/managed-proxy/context.js", () => ({
-  resolveManagedProxyContext: async () => mockManagedProxyCtx,
+mock.module("../platform/client.js", () => ({
+  VellumPlatformClient: {
+    create: async () => mockPlatformClient,
+  },
 }));
 
 // ---------------------------------------------------------------------------
@@ -68,6 +61,22 @@ import { PlatformOAuthConnection } from "./platform-connection.js";
 // Helpers
 // ---------------------------------------------------------------------------
 
+function makeMockClient() {
+  return {
+    baseUrl: "https://platform.example.com",
+    assistantApiKey: "sk-test-key",
+    platformAssistantId: "asst-123",
+    fetch: mock(async () => {
+      return new Response(
+        JSON.stringify({
+          results: [{ id: "platform-conn-1", account_label: null }],
+        }),
+        { status: 200 },
+      );
+    }),
+  };
+}
+
 function setupDefaults(): void {
   mockProvider = {
     providerKey: "integration:google",
@@ -100,12 +109,7 @@ function setupDefaults(): void {
       "google-oauth": { mode: "managed" },
     },
   };
-  mockManagedProxyCtx = {
-    enabled: true,
-    platformBaseUrl: "https://platform.example.com",
-    assistantApiKey: "sk-test-key",
-  };
-  mockAssistantId = "asst-123";
+  mockPlatformClient = makeMockClient();
 }
 
 // ---------------------------------------------------------------------------