@vellumai/assistant 0.8.2 → 0.8.3

package/src/config/feature-flag-registry.json

@@ -9,14 +9,6 @@
       "description": "Automatically trigger conversation analysis on the same cadence as memory extraction (batch threshold, idle debounce, end-of-conversation). The analysis agent has full tool access and writes back to memory and skills without user approval.",
       "defaultEnabled": false
     },
-    {
-      "id": "memory-retrospective",
-      "scope": "assistant",
-      "key": "memory-retrospective",
-      "label": "Memory retrospective pass",
-      "description": "Run a focused, memory-only retrospective during active conversations and at conversation lifecycle. The retrospective agent re-reads the messages added since the last successful run, dedupes against memory/archive/, and calls `remember` for what wasn't captured in the moment. Enabling this also relaxes the in-conversation pressure to call `remember` (lighter PKB reminder + relaxed tool description) so the assistant can stay present and trust the retrospective as the backstop.",
-      "defaultEnabled": false
-    },
     {
       "id": "user-hosted-enabled",
       "scope": "client",
@@ -33,6 +25,14 @@
       "description": "When enabled, the Local hosting option uses Docker under the hood for sandboxed execution, hiding the separate Docker card",
       "defaultEnabled": false
     },
+    {
+      "id": "a2a-channel",
+      "scope": "assistant",
+      "key": "a2a-channel",
+      "label": "A2A Channel",
+      "description": "Enable the A2A (Agent-to-Agent) channel for inter-assistant communication via the open A2A protocol",
+      "defaultEnabled": false
+    },
     {
       "id": "email-channel",
       "scope": "assistant",
@@ -145,6 +145,14 @@
       "description": "Show a speaker button on assistant messages to generate and play the message as audio via Fish Audio TTS",
       "defaultEnabled": false
     },
+    {
+      "id": "openai-compatible-endpoints",
+      "scope": "assistant",
+      "key": "openai-compatible-endpoints",
+      "label": "OpenAI-Compatible Endpoints",
+      "description": "Enable user-configured OpenAI-compatible inference endpoints with custom base URLs and model identifiers",
+      "defaultEnabled": false
+    },
     {
       "id": "backward-releases",
       "scope": "assistant",
@@ -281,14 +289,6 @@
       "description": "Enable the app-control skill (per-app screenshot + raw input bypassing AX tree)",
       "defaultEnabled": false
     },
-    {
-      "id": "species-migration",
-      "scope": "assistant",
-      "key": "species-migration",
-      "label": "Species Migration",
-      "description": "Enable the Species Migration skill for migrating from OpenClaw, Hermes, Manus, and other assistant species into Vellum.",
-      "defaultEnabled": false
-    },
     {
       "id": "analyze-conversation",
       "scope": "assistant",
@@ -322,19 +322,11 @@
       "defaultEnabled": false
     },
     {
-      "id": "provider-deepseek",
-      "scope": "assistant",
-      "key": "provider-deepseek",
-      "label": "DeepSeek Provider",
-      "description": "Enable the DeepSeek direct API provider and its models (V4 Pro, V4 Flash) in the provider picker and model selection UI",
-      "defaultEnabled": false
-    },
-    {
-      "id": "provider-minimax",
-      "scope": "assistant",
-      "key": "provider-minimax",
-      "label": "MiniMax Provider",
-      "description": "Enable the MiniMax direct API provider and its models (M2.7, M2.5, M2.1, M2, and highspeed variants) in the provider picker and model selection UI",
+      "id": "velvet-theme",
+      "scope": "client",
+      "key": "velvet-theme",
+      "label": "Velvet Theme",
+      "description": "Show the Velvet theme option in the macOS appearance settings. Velvet is a dark-mode variant with red/pink accent colors.",
       "defaultEnabled": false
     }
   ]

package/src/config/llm-resolver.ts

@@ -1,6 +1,7 @@
 import { z } from "zod";
 
 import { getCatalogProviderForModel } from "../providers/model-catalog.js";
+import { CALL_SITE_DEFAULTS } from "./call-site-defaults.js";
 import {
   type LLMCallSite,
   LLMConfigBase,
@@ -56,7 +57,9 @@ export function resolveCallSiteConfig(
     opts.overrideProfile != null
       ? llm.profiles?.[opts.overrideProfile]
       : undefined;
-  const site = llm.callSites?.[callSite];
+  const site =
+    llm.callSites?.[callSite] ??
+    effectiveDefault(callSite, llm, opts.overrideProfile != null);
 
   if (callSite === "mainAgent") {
     appendCallSiteLayers(layers, callSite, llm, site);
@@ -77,6 +80,54 @@ export function resolveCallSiteConfig(
 
 type Mergeable = Record<string, unknown>;
 
+/**
+ * Returns the resolved default profile key for a call site, accounting for
+ * the `custom-*` user-profile fallback when the managed profile is unavailable.
+ */
+export function resolveDefaultProfileKey(
+  callSite: LLMCallSite,
+  llm: z.infer<typeof LLMSchema>,
+): string | undefined {
+  const dflt = CALL_SITE_DEFAULTS[callSite];
+  if (dflt?.profile == null) return undefined;
+  const target = llm.profiles?.[dflt.profile];
+  if (target != null && target.status !== "disabled") return dflt.profile;
+  const customKey = `custom-${dflt.profile}`;
+  const customTarget = llm.profiles?.[customKey];
+  if (customTarget != null && customTarget.status !== "disabled")
+    return customKey;
+  return undefined;
+}
+
+function effectiveDefault(
+  callSite: LLMCallSite,
+  llm: z.infer<typeof LLMSchema>,
+  hasOverrideProfile = false,
+): z.infer<typeof LLMSchema>["callSites"][LLMCallSite] | undefined {
+  const dflt = CALL_SITE_DEFAULTS[callSite];
+  if (dflt == null) return undefined;
+  const targetProfile =
+    dflt.profile != null ? llm.profiles?.[dflt.profile] : undefined;
+  const profileUnavailable =
+    dflt.profile != null &&
+    (targetProfile == null || targetProfile.status === "disabled");
+
+  if (profileUnavailable && !hasOverrideProfile) {
+    const customKey = `custom-${dflt.profile}`;
+    const customProfile = llm.profiles?.[customKey];
+    if (customProfile != null && customProfile.status !== "disabled") {
+      return { ...dflt, profile: customKey };
+    }
+  }
+
+  const stripProfile = hasOverrideProfile || profileUnavailable;
+  if (stripProfile) {
+    const { profile: _profile, ...rest } = dflt;
+    return Object.keys(rest).length > 0 ? rest : undefined;
+  }
+  return dflt;
+}
+
 function withImpliedProviderForKnownModel(source: Mergeable): Mergeable {
   if (source.provider !== undefined) return source;
   const model = source.model;

package/src/config/schema.ts

@@ -20,6 +20,7 @@ import { AnalysisConfigSchema } from "./schemas/analysis.js";
 import { BackupConfigSchema } from "./schemas/backup.js";
 import { CallsConfigSchema } from "./schemas/calls.js";
 import {
+  A2AConfigSchema,
   SlackConfigSchema,
   TelegramConfigSchema,
   TwilioConfigSchema,
@@ -111,6 +112,7 @@ export const AssistantConfigSchema = z
     whatsapp: WhatsAppConfigSchema.default(WhatsAppConfigSchema.parse({})),
     telegram: TelegramConfigSchema.default(TelegramConfigSchema.parse({})),
     slack: SlackConfigSchema.default(SlackConfigSchema.parse({})),
+    a2a: A2AConfigSchema.default(A2AConfigSchema.parse({})),
     ingress: IngressConfigSchema,
     platform: PlatformConfigSchema.default(PlatformConfigSchema.parse({})),
     daemon: DaemonConfigSchema.default(DaemonConfigSchema.parse({})),

package/src/config/schemas/__tests__/memory-v2.test.ts

@@ -33,7 +33,7 @@ describe("MemoryV2ConfigSchema", () => {
         dtype: "q8",
       },
       router: {
-        enabled: false,
+        enabled: true,
         max_page_ids: 25,
         router_prompt_path: null,
       },
@@ -161,9 +161,9 @@ describe("MemoryV2ConfigSchema", () => {
     expect(() => MemoryV2ConfigSchema.parse({ epsilon: 1.5 })).toThrow();
   });
 
-  test("router defaults to disabled with max_page_ids=25", () => {
+  test("router defaults to enabled with max_page_ids=25", () => {
     const parsed = MemoryV2ConfigSchema.parse({});
-    expect(parsed.router.enabled).toBe(false);
+    expect(parsed.router.enabled).toBe(true);
     expect(parsed.router.max_page_ids).toBe(25);
   });
 

package/src/config/schemas/channels.ts

@@ -99,6 +99,15 @@ export const TelegramConfigSchema = z
   })
   .describe("Telegram bot channel configuration");
 
+export const A2AConfigSchema = z
+  .object({
+    enabled: z
+      .boolean({ error: "a2a.enabled must be a boolean" })
+      .default(false)
+      .describe("Whether the A2A channel is enabled"),
+  })
+  .describe("Agent-to-Agent protocol channel configuration");
+
 export const SlackConfigSchema = z
   .object({
     deliverAuthBypass: z

package/src/config/schemas/conversations.ts

@@ -10,6 +10,16 @@ export const ConversationsConfigSchema = z
       .describe(
         "When true, skip the second-pass title regeneration that fires after the third user turn. The initial auto-generated title and manual renames are unaffected.",
       ),
+    backgroundInjection: z
+      .string({
+        error: "conversations.backgroundInjection must be a string",
+      })
+      .default(
+        "This is a background turn — your guardian isn't watching. If anything noteworthy comes up, send them a notification so they see it when they're back by invoking the `notifications` skill (`assistant notifications send --message \"...\"`)",
+      )
+      .describe(
+        "Inner text injected into the tail user message of non-interactive turns in background/scheduled conversations. The injector wraps this in <background_turn>...</background_turn> tags. Empty string disables the injection.",
+      ),
   })
   .describe("Conversation behavior configuration");
 

package/src/config/schemas/heartbeat.ts

@@ -56,6 +56,20 @@ export const HeartbeatConfigSchema = z
       .describe(
         "Maximum heartbeats that can run consecutively without a guardian message. Counter resets when the guardian sends a message. Set to null for unlimited.",
       ),
+    disposition: z
+      .string({ error: "heartbeat.disposition must be a string" })
+      .default(
+        `This is your time to do something useful, interesting, or creative while your guardian is away.
+
+Before checking on anything, ask yourself: is there something I want to work on, think about, or make progress on right now? A project, an idea, something I noticed, something I've been meaning to get to. If so, do it.
+
+If you do something worth sharing — built something, noticed something, had an idea — send your guardian a notification so they see it when they're back.
+
+If nothing needs attention and nothing stirs, that's fine. But make sure you actually considered it first rather than defaulting to "nothing to do."`,
+      )
+      .describe(
+        "Inner text injected into the heartbeat prompt. The service wraps this in <heartbeat-disposition>...</heartbeat-disposition> tags. Empty string disables the block.",
+      ),
   })
   .describe("Periodic heartbeat configuration for health monitoring")
   .superRefine((config, ctx) => {

package/src/config/schemas/llm.ts

@@ -20,9 +20,7 @@ const LLMProvider = z.enum([
   "ollama",
   "fireworks",
   "openrouter",
-  "zai",
-  "deepseek",
-  "minimax",
+  "openai-compatible",
 ]);
 type LLMProvider = z.infer<typeof LLMProvider>;
 

package/src/config/schemas/memory-retrospective.ts

@@ -40,7 +40,7 @@ export const MemoryRetrospectiveConfigSchema = z
       ),
   })
   .describe(
-    "Controls the memory-retrospective background pass triggered by the `memory-retrospective` feature flag. Model selection lives under llm.callSites.memoryRetrospective.",
+    "Controls the memory-retrospective background pass. Model selection lives under llm.callSites.memoryRetrospective.",
   );
 
 export type MemoryRetrospectiveConfig = z.infer<

package/src/config/schemas/memory-v2.ts

@@ -260,9 +260,9 @@ export const MemoryV2ConfigSchema = z
       .object({
         enabled: z
           .boolean()
-          .default(false)
+          .default(true)
           .describe(
-            "Whether to use the LLM router as the per-turn page-selection mechanism in place of spreading activation. Disabled by default — opt in once the router orchestration and dispatcher land.",
+            "Whether to use the LLM router as the per-turn page-selection mechanism in place of spreading activation. Enabled by default.",
           ),
         max_page_ids: z
           .number()
@@ -283,9 +283,9 @@ export const MemoryV2ConfigSchema = z
             "Optional path to a file whose contents replace the bundled router prompt. Absolute paths are used as-is, a leading `~/` is expanded to the home directory, otherwise the path is resolved under the workspace root. The loaded contents may include `{{ASSISTANT_NAME}}`, `{{USER_NAME}}`, and `{{PAGE_INDEX}}`, which are substituted at runtime. If the file is missing, unreadable, or empty, the bundled prompt is used and a warning is logged.",
           ),
       })
-      .default({ enabled: false, max_page_ids: 25, router_prompt_path: null })
+      .default({ enabled: true, max_page_ids: 25, router_prompt_path: null })
       .describe(
-        "LLM router configuration. When enabled, a single Sonnet router call replaces spreading activation for per-turn page selection.",
+        "LLM router configuration. When enabled, a single router LLM call replaces spreading activation for per-turn page selection.",
       ),
   })
   .describe(

package/src/config/schemas/memory.ts

@@ -23,7 +23,9 @@ export const MemoryConfigSchema = z
     enabled: z
       .boolean({ error: "memory.enabled must be a boolean" })
       .default(true)
-      .describe("Whether the long-term memory system is enabled"),
+      .describe(
+        "Whether the long-term memory system is enabled — gates background memory jobs, embedding generation, and `<memory>` block injection into user messages",
+      ),
     embeddings: MemoryEmbeddingsConfigSchema.default(
       MemoryEmbeddingsConfigSchema.parse({}),
     ),

package/src/config/seed-inference-profiles.ts

@@ -3,6 +3,8 @@ import {
   createConnection,
   disableManagedConnectionsForByokHatch,
   getConnection,
+  MANAGED_CONNECTION_NAMES,
+  PROVIDERS_REQUIRING_BASE_URL_AND_MODELS,
 } from "../providers/inference/connections.js";
 import { PROVIDER_CATALOG } from "../providers/model-catalog.js";
 import { resolveModelIntent } from "../providers/model-intents.js";
@@ -17,10 +19,6 @@ import {
 
 const log = getLogger("seed-inference-profiles");
 
-const MANAGED_CONNECTION_NAME = "anthropic-managed";
-const MANAGED_PROFILE_PROVIDER: NonNullable<ProfileEntry["provider"]> =
-  "anthropic";
-
 /**
  * Template for a daemon-managed inference profile. The profile's model is
  * resolved at seed time from `PROVIDER_MODEL_INTENTS` so the catalog stays the
@@ -31,16 +29,20 @@ type ManagedProfileTemplate = Omit<
   "provider" | "model" | "provider_connection"
 > & {
   intent: ModelIntent;
+  provider: NonNullable<ProfileEntry["provider"]>;
+  connectionName: string;
 };
 
 /**
- * Managed Anthropic profiles. Overwritten on every daemon boot so Vellum can
- * push model/config updates to customers in new releases. Platform overlays
+ * Managed profiles. Overwritten on every daemon boot so Vellum can push
+ * model/config updates to customers in new releases. Platform overlays
  * (`preserveProfileNames`) take precedence when present.
  */
 const MANAGED_PROFILE_TEMPLATES: Record<string, ManagedProfileTemplate> = {
   balanced: {
     intent: "balanced",
+    provider: "anthropic",
+    connectionName: "anthropic-managed",
     source: "managed",
     label: "Balanced",
     description: "Good balance of quality, cost, and speed",
@@ -51,6 +53,8 @@ const MANAGED_PROFILE_TEMPLATES: Record<string, ManagedProfileTemplate> = {
   },
   "quality-optimized": {
     intent: "quality-optimized",
+    provider: "anthropic",
+    connectionName: "anthropic-managed",
     source: "managed",
     label: "Quality",
     description: "Best results with the most capable model",
@@ -61,6 +65,8 @@ const MANAGED_PROFILE_TEMPLATES: Record<string, ManagedProfileTemplate> = {
   },
   "cost-optimized": {
     intent: "latency-optimized",
+    provider: "anthropic",
+    connectionName: "anthropic-managed",
     source: "managed",
     label: "Speed",
     description: "Fastest responses at lower cost",
@@ -74,11 +80,15 @@ const MANAGED_PROFILE_TEMPLATES: Record<string, ManagedProfileTemplate> = {
 /**
  * User profile templates. Materialized at hatch time for off-platform
  * installations. Each points at the user's personal provider connection
- * (backed by their API key in CES).
+ * (backed by their API key in CES). The `provider` and `connectionName`
+ * fields are placeholders — they are overridden at hatch time with the
+ * user's chosen provider and personal connection name.
  */
 const USER_PROFILE_TEMPLATES: Record<string, ManagedProfileTemplate> = {
   "custom-balanced": {
     intent: "balanced",
+    provider: "anthropic",
+    connectionName: "",
     source: "user",
     label: "Balanced",
     description: "Good balance of quality, cost, and speed",
@@ -89,6 +99,8 @@ const USER_PROFILE_TEMPLATES: Record<string, ManagedProfileTemplate> = {
   },
   "custom-quality-optimized": {
     intent: "quality-optimized",
+    provider: "anthropic",
+    connectionName: "",
     source: "user",
     label: "Quality",
     description: "Best results with the most capable model",
@@ -99,6 +111,8 @@ const USER_PROFILE_TEMPLATES: Record<string, ManagedProfileTemplate> = {
   },
   "custom-cost-optimized": {
     intent: "latency-optimized",
+    provider: "anthropic",
+    connectionName: "",
     source: "user",
     label: "Speed",
     description: "Fastest responses at lower cost",
@@ -164,11 +178,12 @@ export function seedInferenceProfiles(
   // BYOK mode = off-platform installs. The user is bringing their own provider
   // API key; managed profile labels get a " (Managed)" suffix to disambiguate
   // from the personal "custom-*" profiles that share base labels. Managed
-  // profile + connection status is initially "disabled" so the picker doesn't
-  // offer an unusable platform-auth option on day one — but ONLY at hatch
-  // time, and ONLY when the entry isn't already in the user's config (i.e.
-  // first materialization). Post-hatch user toggles survive every subsequent
-  // boot.
+  // profile + connection status is initially "disabled" for true BYOK hatches
+  // so the picker doesn't offer an unusable platform-auth option on day one.
+  // When the hatch overlay explicitly selects a managed profile, the matching
+  // managed connection stays active so the first post-onboarding message can
+  // use the user's chosen managed route. Post-hatch user toggles survive every
+  // subsequent boot.
   const isByokMode = !isPlatform;
 
   // 1. Managed profiles. Off-platform: overwrite on every boot so Vellum can
@@ -198,10 +213,18 @@ export function seedInferenceProfiles(
   //        rewritten to the suffixed form. Any other previous label value
   //        (user-set custom string, explicit null, already-suffixed) is
   //        preserved as-is.
-  //      • status: "disabled" on fresh materialization at hatch only —
-  //        gated on (isHatch && !previous) so post-hatch boots and existing
-  //        installs are never auto-disabled. A user re-enable persists
-  //        across boots via the key-presence preservation below.
+  //      • status: "disabled" on fresh materialization at BYOK hatch only —
+  //        gated on (isHatch && !previous) and skipped for any managed
+  //        connection explicitly selected by the hatch overlay. Post-hatch
+  //        boots and existing installs are never auto-disabled. A user
+  //        re-enable persists across boots via the key-presence preservation
+  //        below.
+  const hatchSelectedManagedConnection = getHatchSelectedManagedConnection(
+    llm,
+    profiles,
+    options,
+  );
+
   for (const [name, template] of Object.entries(MANAGED_PROFILE_TEMPLATES)) {
     if (preservedProfileNames.has(name)) continue;
     if (isPlatform && readObject(profiles[name]) !== null) continue;
@@ -212,10 +235,15 @@ export function seedInferenceProfiles(
       : template;
     const next = materializeProfile(
       effectiveTemplate,
-      MANAGED_PROFILE_PROVIDER,
-      MANAGED_CONNECTION_NAME,
+      template.provider,
+      template.connectionName,
     ) as Record<string, unknown>;
-    if (isByokMode && options.isHatch && !previous) {
+    if (
+      isByokMode &&
+      options.isHatch &&
+      !previous &&
+      template.connectionName !== hatchSelectedManagedConnection
+    ) {
       next.status = "disabled";
     }
     if (previous) {
@@ -238,17 +266,24 @@ export function seedInferenceProfiles(
   // 2. User profiles — only at hatch time for off-platform installations.
   let userConnectionName: string | undefined;
   if (options.isHatch && !isPlatform) {
-    // BYOK hatch: disable the three canonical managed connections so the
-    // picker doesn't surface unusable platform-auth options on day one.
-    // Runs only here, only at hatch — `seedCanonicalConnections` leaves
-    // `status` alone on subsequent boots so a post-hatch user re-enable
-    // persists.
+    // BYOK hatch: disable canonical managed connections so the picker doesn't
+    // surface unusable platform-auth options on day one. If the hatch overlay
+    // selected a managed profile, leave that connection active; the user has
+    // already chosen managed inference. Runs only here, only at hatch —
+    // `seedCanonicalConnections` leaves `status` alone on subsequent boots so
+    // a post-hatch user re-enable persists.
     if (options.db) {
-      disableManagedConnectionsForByokHatch(options.db);
+      disableManagedConnectionsForByokHatch(options.db, {
+        excludeConnection: hatchSelectedManagedConnection,
+      });
     }
 
     const hatchProvider = readString(readObject(llm.default)?.provider);
-    if (hatchProvider && hatchProvider !== "ollama") {
+    if (
+      hatchProvider &&
+      hatchProvider !== "ollama" &&
+      !PROVIDERS_REQUIRING_BASE_URL_AND_MODELS.has(hatchProvider)
+    ) {
       userConnectionName = `${hatchProvider}-personal`;
 
       if (options.db) {
@@ -269,8 +304,7 @@ export function seedInferenceProfiles(
         }
       }
 
-      const provider =
-        hatchProvider as NonNullable<ProfileEntry["provider"]>;
+      const provider = hatchProvider as NonNullable<ProfileEntry["provider"]>;
       for (const [name, template] of Object.entries(USER_PROFILE_TEMPLATES)) {
         if (preservedProfileNames.has(name)) continue;
         profiles[name] = materializeProfile(
@@ -342,7 +376,7 @@ function materializeProfile(
   provider: NonNullable<ProfileEntry["provider"]>,
   connectionName: string,
 ): ProfileEntry {
-  const { intent, ...rest } = template;
+  const { intent, provider: _p, connectionName: _c, ...rest } = template;
   return {
     ...rest,
     provider,
@@ -361,6 +395,42 @@ function readString(value: unknown): string | undefined {
   return typeof value === "string" && value.length > 0 ? value : undefined;
 }
 
+function getHatchSelectedManagedConnection(
+  llm: Record<string, unknown>,
+  profiles: Record<string, Record<string, unknown>>,
+  options: SeedInferenceProfilesOptions,
+): string | undefined {
+  if (!options.isHatch || options.preserveActiveProfile !== true) {
+    return undefined;
+  }
+
+  const activeProfile = readString(llm.activeProfile);
+  if (!activeProfile) return undefined;
+
+  const activeProfileEntry = readObject(profiles[activeProfile]);
+  if (
+    activeProfileEntry &&
+    Object.prototype.hasOwnProperty.call(
+      activeProfileEntry,
+      "provider_connection",
+    )
+  ) {
+    const explicitConnection = readString(
+      activeProfileEntry.provider_connection,
+    );
+    return explicitConnection &&
+      MANAGED_CONNECTION_NAMES.has(explicitConnection)
+      ? explicitConnection
+      : undefined;
+  }
+
+  const templateConnection =
+    MANAGED_PROFILE_TEMPLATES[activeProfile]?.connectionName;
+  return templateConnection && MANAGED_CONNECTION_NAMES.has(templateConnection)
+    ? templateConnection
+    : undefined;
+}
+
 /**
  * Format the human-readable label seeded onto a personal provider connection
  * at hatch time, e.g. `"Anthropic (Personal)"`. The display name is sourced