@vellumai/assistant 0.6.5 → 0.6.6

package/src/daemon/handlers/config-model.ts

@@ -8,6 +8,7 @@ import {
   setServiceField,
 } from "../../config/raw-config-utils.js";
 import { VALID_INFERENCE_PROVIDERS } from "../../config/schemas/services.js";
+import { providerForImageModelPrefix } from "../../media/types.js";
 import type { ProviderCatalogEntry } from "../../providers/model-catalog.js";
 import {
   isModelInCatalog,
@@ -219,6 +220,16 @@ export async function setModel(
 export function setImageGenModel(modelId: string, ctx: ModelSetContext): void {
   const raw = loadRawConfig();
   setServiceField(raw, "image-generation", "model", modelId);
+  // Keep the derived provider in sync with the selected model so downstream
+  // routing never sends a Gemini request to an OpenAI model (or vice versa).
+  // The prefix logic is shared with workspace migration 006-services-config
+  // via providerForImageModelPrefix().
+  setServiceField(
+    raw,
+    "image-generation",
+    "provider",
+    providerForImageModelPrefix(modelId),
+  );
 
   const wasSuppressed = ctx.suppressConfigReload;
   ctx.setSuppressConfigReload(true);

package/src/daemon/handlers/skills.ts

@@ -326,10 +326,14 @@ export function postInstallSkill(
 
 /** Map the old `source` field to the new `kind` axis. */
 function deriveKind(
-  source: "bundled" | "managed" | "workspace" | "extra" | "catalog",
+  source: "bundled" | "managed" | "workspace" | "extra" | "catalog" | "plugin",
 ): SlimSkillResponse["kind"] {
   if (source === "bundled") return "bundled";
   if (source === "catalog") return "catalog";
+  // Plugin-contributed skills are framework-provided like bundled skills —
+  // expose them under the same "bundled" kind so the UI doesn't invent a
+  // new category that existing clients don't know how to render.
+  if (source === "plugin") return "bundled";
   return "installed"; // managed, workspace, extra
 }
 

package/src/daemon/lifecycle.ts

@@ -12,8 +12,6 @@ import { setVoiceBridgeDeps } from "../calls/voice-session-bridge.js";
 import { initFeatureFlagOverrides } from "../config/assistant-feature-flags.js";
 import {
   getPlatformAssistantId,
-  getQdrantHttpPortEnv,
-  getQdrantUrlEnv,
   getRuntimeHttpHost,
   getRuntimeHttpPort,
   setIngressPublicBaseUrl,
@@ -43,8 +41,6 @@ import {
   startFeedScheduler,
 } from "../home/feed-scheduler.js";
 import { backfillRelationshipStateIfMissing } from "../home/relationship-state-writer.js";
-import { getHookManager } from "../hooks/manager.js";
-import { installTemplates } from "../hooks/templates.js";
 import { closeSentry, initSentry, setSentryDeviceId } from "../instrument.js";
 import { getMcpServerManager } from "../mcp/manager.js";
 import * as attachmentsStore from "../memory/attachments-store.js";
@@ -63,7 +59,7 @@ import {
 } from "../memory/embedding-backend.js";
 import { enqueueMemoryJob } from "../memory/jobs-store.js";
 import { startMemoryJobsWorker } from "../memory/jobs-worker.js";
-import { initQdrantClient } from "../memory/qdrant-client.js";
+import { initQdrantClient, resolveQdrantUrl } from "../memory/qdrant-client.js";
 import { QdrantManager } from "../memory/qdrant-manager.js";
 import { rotateToolInvocations } from "../memory/tool-usage-store.js";
 import { deleteOldTraceEvents } from "../memory/trace-event-store.js";
@@ -73,6 +69,7 @@ import {
 } from "../notifications/emit-signal.js";
 import { backfillManualTokenConnections } from "../oauth/manual-token-connection.js";
 import { seedOAuthProviders } from "../oauth/seed-providers.js";
+import { loadUserPlugins } from "../plugins/user-loader.js";
 import { ensurePromptFiles } from "../prompts/system-prompt.js";
 import { resolveManagedProxyContext } from "../providers/managed-proxy/context.js";
 import { buildAssistantEvent } from "../runtime/assistant-event.js";
@@ -107,6 +104,7 @@ import {
   getInterfacesDir,
   getWorkspaceDir,
 } from "../util/platform.js";
+import { APP_VERSION } from "../version.js";
 import {
   listWorkItems,
   updateWorkItem,
@@ -124,6 +122,7 @@ import {
   cleanupPidFileIfOwner,
   writePid,
 } from "./daemon-control.js";
+import { bootstrapPlugins } from "./external-plugins-bootstrap.js";
 import {
   createGuardianActionCopyGenerator,
   createGuardianFollowUpConversationGenerator,
@@ -148,7 +147,6 @@ import {
 import { seedInterfaceFiles } from "./seed-files.js";
 import { DaemonServer } from "./server.js";
 import { installShutdownHandlers } from "./shutdown-handlers.js";
-import { handleWatchObservation } from "./watch-handler.js";
 
 // Re-export public API so existing consumers don't need to change imports
 export type { StopResult } from "./daemon-control.js";
@@ -340,8 +338,7 @@ export async function runDaemon(): Promise<void> {
 
     seedInterfaceFiles();
 
-    log.info("Daemon startup: installing templates and initializing DB");
-    installTemplates();
+    log.info("Daemon startup: initializing DB");
     ensurePromptFiles();
 
     // DB must be initialized before workspace migrations because some
@@ -699,6 +696,33 @@ export async function runDaemon(): Promise<void> {
       }
     }
 
+    // Populate the registry with user plugins from `~/.vellum/plugins/*`
+    // AFTER first-party plugins have already registered via their static
+    // side-effect imports. User plugins may fail to load individually; a
+    // failing user plugin is logged and skipped so one bad install can't
+    // prevent the daemon from starting. Ordering is load-bearing:
+    //   first-party registrations → user registrations → bootstrap (init).
+    // Both groups are fully registered before any `init()` runs so plugins
+    // that depend on each other's registration observably see a stable
+    // registry at init time.
+    await loadUserPlugins();
+
+    // Bootstrap registered plugins. Runs after the plugin registry is
+    // populated (first-party static side-effect imports + user plugins
+    // loaded above) and before the DaemonServer starts handling
+    // conversations. Credential resolution + per-plugin storage directory
+    // creation happen here. Wrapped in try/catch so a failing plugin can't
+    // block daemon startup — bootstrapPlugins internally tears down any
+    // partially-initialized plugins before throwing.
+    try {
+      await bootstrapPlugins({ config, assistantVersion: APP_VERSION });
+    } catch (err) {
+      log.warn(
+        { err },
+        "Plugin bootstrap failed — continuing startup with degraded plugin functionality",
+      );
+    }
+
     // Start the DaemonServer (conversation manager) before Qdrant so HTTP
     // routes can begin accepting requests while Qdrant initializes.
     log.info("Daemon startup: starting DaemonServer");
@@ -741,13 +765,7 @@ export async function runDaemon(): Promise<void> {
     // Initialize Qdrant vector store and memory worker in the background so the
     // RuntimeHttpServer can start accepting requests without waiting for Qdrant.
     async function initializeQdrantAndMemory(): Promise<void> {
-      // Prefer QDRANT_HTTP_PORT (locally-spawned Qdrant on a specific port) over
-      // QDRANT_URL (external Qdrant instance) so the CLI can set the port without
-      // triggering QdrantManager's external mode which skips local process spawn.
-      const qdrantHttpPort = getQdrantHttpPortEnv();
-      const qdrantUrl = qdrantHttpPort
-        ? `http://127.0.0.1:${qdrantHttpPort}`
-        : getQdrantUrlEnv() || config.memory.qdrant.url;
+      const qdrantUrl = resolveQdrantUrl(config);
       log.info({ qdrantUrl }, "Daemon startup: initializing Qdrant");
       const manager = new QdrantManager({ url: qdrantUrl });
       bgRefs.qdrantManager = manager;
@@ -941,29 +959,6 @@ export async function runDaemon(): Promise<void> {
           throwOnError: true,
         });
       },
-      (schedule) => {
-        void emitNotificationSignal({
-          sourceEventName: "schedule.complete",
-          sourceChannel: "scheduler",
-          sourceContextId: schedule.id,
-          attentionHints: {
-            requiresAction: false,
-            urgency: "medium",
-            isAsyncBackground: true,
-            visibleInSourceNow: false,
-          },
-          contextPayload: {
-            scheduleId: schedule.id,
-            name: schedule.name,
-          },
-          conversationMetadata: {
-            groupId: "system:scheduled",
-            scheduleJobId: schedule.id,
-            source: "schedule",
-          },
-          dedupeKey: `schedule:complete:${schedule.id}:${Date.now()}`,
-        });
-      },
       (notification) => {
         void emitNotificationSignal({
           sourceEventName: "watcher.notification",
@@ -1132,28 +1127,6 @@ export async function runDaemon(): Promise<void> {
           );
         },
       },
-      getWatchDeps: () => {
-        const ctx = server.getHandlerContext();
-        return {
-          handleWatchObservation: async (params) => {
-            await handleWatchObservation(
-              {
-                type: "watch_observation",
-                watchId: params.watchId,
-                conversationId: params.conversationId,
-                ocrText: params.ocrText,
-                appName: params.appName,
-                windowTitle: params.windowTitle,
-                bundleIdentifier: params.bundleIdentifier,
-                timestamp: params.timestamp,
-                captureIndex: params.captureIndex,
-                totalExpected: params.totalExpected,
-              },
-              ctx,
-            );
-          },
-        };
-      },
       getRecordingDeps: () => ({
         getHandlerContext: () => server.getHandlerContext(),
       }),
@@ -1390,13 +1363,6 @@ export async function runDaemon(): Promise<void> {
       log.warn({ err }, "Assistant symlink installation failed — continuing");
     }
 
-    const hookManager = getHookManager();
-    hookManager.watch();
-
-    void hookManager.trigger("daemon-start", {
-      pid: process.pid,
-    });
-
     // Download embedding runtime in background (non-blocking).
     // If download fails, local embeddings gracefully fall back to cloud backends.
     void (async () => {
@@ -1510,7 +1476,6 @@ export async function runDaemon(): Promise<void> {
       workspaceHeartbeat,
       heartbeat,
       filing,
-      hookManager,
       runtimeHttp,
       scheduler,
       feedScheduler,

package/src/daemon/message-types/computer-use.ts

@@ -1,4 +1,4 @@
-// Computer use and watch observation types.
+// Computer use and recording types.
 
 import type { CommandIntent, UserMessageAttachment } from "./shared.js";
 
@@ -15,19 +15,6 @@ export interface TaskSubmit {
   commandIntent?: CommandIntent;
 }
 
-export interface WatchObservation {
-  type: "watch_observation";
-  watchId: string;
-  conversationId: string;
-  ocrText: string;
-  appName?: string;
-  windowTitle?: string;
-  bundleIdentifier?: string;
-  timestamp: number;
-  captureIndex: number;
-  totalExpected: number;
-}
-
 // === Recording ===
 
 /** Recording options shared across standalone and CU recording flows. */
@@ -89,30 +76,11 @@ export interface RecordingResume {
   recordingId: string;
 }
 
-export interface WatchStarted {
-  type: "watch_started";
-  conversationId: string;
-  watchId: string;
-  durationSeconds: number;
-  intervalSeconds: number;
-}
-
-export interface WatchCompleteRequest {
-  type: "watch_complete_request";
-  conversationId: string;
-  watchId: string;
-}
-
 // --- Domain-level union aliases (consumed by the barrel file) ---
 
-export type _ComputerUseClientMessages =
-  | TaskSubmit
-  | WatchObservation
-  | RecordingStatus;
+export type _ComputerUseClientMessages = TaskSubmit | RecordingStatus;
 
 export type _ComputerUseServerMessages =
-  | WatchStarted
-  | WatchCompleteRequest
   | RecordingStart
   | RecordingStop
   | RecordingPause

package/src/daemon/message-types/conversations.ts

@@ -421,8 +421,16 @@ export interface UsageResponse {
   model: string;
 }
 
+/**
+ * Emitted after a compaction turn completes (both auto-compaction and
+ * `/compact`). Carries the fresh `estimatedInputTokens` so clients can refresh
+ * the context-window indicator without waiting for the next `usage_update`.
+ *
+ * Scoped per-conversation — see `CompactionCircuitOpen` doc for why.
+ */
 export interface ContextCompacted {
   type: "context_compacted";
+  conversationId: string;
   previousEstimatedInputTokens: number;
   estimatedInputTokens: number;
   maxInputTokens: number;
@@ -432,6 +440,22 @@ export interface ContextCompacted {
   summaryInputTokens: number;
   summaryOutputTokens: number;
   summaryModel: string;
+  /**
+   * Quality signals for the generated summary. Emitted for every
+   * compaction (including truncation-only paths where the summary text
+   * is unchanged from the prior pass). Consumers can use these to detect
+   * regressions without needing to read the summary text itself.
+   *
+   * - `summaryCharCount`: length of the produced summary text.
+   * - `summaryHeaderCount`: number of `## ` section headers in the summary.
+   * - `summaryHadMemoryEcho`: `true` if the summary contains any runtime
+   *   injection tag (e.g. `<memory`, `<turn_context>`, `<workspace>`).
+   *   Should always be `false` — `true` indicates the compaction strip
+   *   logic failed to remove an injected block from the summarizer input.
+   */
+  summaryCharCount?: number;
+  summaryHeaderCount?: number;
+  summaryHadMemoryEcho?: boolean;
 }
 
 /**
@@ -439,14 +463,38 @@ export interface ContextCompacted {
  * summary-LLM failures (with local fallback covering each), auto-compaction is
  * suspended until `openUntil` to avoid repeatedly hammering a broken provider.
  * User-initiated compaction (`/compact`, `force: true`) bypasses the breaker.
+ *
+ * `conversationId` scopes the event so clients can ignore breaker trips from
+ * other conversations — `EventStreamClient` broadcasts every parsed server
+ * message to all subscribers, so without this field a breaker trip in one
+ * conversation would set the "auto-compaction paused" banner on every open
+ * `ChatViewModel`.
  */
 export interface CompactionCircuitOpen {
   type: "compaction_circuit_open";
+  conversationId: string;
   reason: "3_consecutive_failures";
   /** Timestamp (ms since epoch) when the breaker will allow auto-compaction again. */
   openUntil: number;
 }
 
+/**
+ * Emitted when the compaction circuit breaker transitions from open → closed
+ * because a successful compaction reset
+ * `ctx.compactionCircuitOpenUntil`. The Swift client clears its banner state
+ * on receipt so the "auto-compaction paused" indicator dismisses immediately
+ * instead of lingering until the original `openUntil` deadline (up to 1h).
+ *
+ * Only fires on the open→closed transition — successful compactions while
+ * the breaker was already closed would be noise.
+ *
+ * Scoped per-conversation — see `CompactionCircuitOpen` doc for why.
+ */
+export interface CompactionCircuitClosed {
+  type: "compaction_circuit_closed";
+  conversationId: string;
+}
+
 export type ConversationErrorCode =
   | "PROVIDER_NETWORK"
   | "PROVIDER_RATE_LIMIT"
@@ -545,6 +593,7 @@ export type _ConversationsServerMessages =
   | UsageResponse
   | ContextCompacted
   | CompactionCircuitOpen
+  | CompactionCircuitClosed
   | ConversationErrorMessage
   | ConversationInfo
   | ConversationTitleUpdated

package/src/daemon/message-types/messages.ts

@@ -145,6 +145,14 @@ export interface ToolResult {
   imageDataList?: string[];
   /** The tool_use block ID for client-side correlation. */
   toolUseId?: string;
+  /** Risk level from the classifier ("low" | "medium" | "high" | "unknown"). */
+  riskLevel?: string;
+  /** Human-readable reason for the risk classification. */
+  riskReason?: string;
+  /** Whether the daemon is running in a containerized (Docker) environment. */
+  isContainerized?: boolean;
+  /** Scope options ladder for the rule editor modal (narrowest to broadest). */
+  riskScopeOptions?: Array<{ pattern: string; label: string }>;
 }
 
 export interface ConfirmationRequest {
@@ -153,6 +161,10 @@ export interface ConfirmationRequest {
   toolName: string;
   input: Record<string, unknown>;
   riskLevel: string;
+  /** Human-readable reason for the risk classification (e.g. "Modifies remote repository state"). */
+  riskReason?: string;
+  /** Whether the daemon is running in a containerized (Docker) environment. */
+  isContainerized?: boolean;
   executionTarget?: "sandbox" | "host";
   allowlistOptions: Array<{
     label: string;

package/src/daemon/server.ts

@@ -886,6 +886,7 @@ export class DaemonServer {
         trustClass: conversation.trustContext?.trustClass ?? "guardian",
         hostBrowserProxy: conversation.hostBrowserProxy,
         transportInterface: conversation.transportInterface,
+        hostBrowserRegistryRouted: !!conversation.hostBrowserSenderOverride,
       };
     });
 
@@ -1302,9 +1303,10 @@ export class DaemonServer {
     // Chrome-extension host_browser wiring is intentionally not supported
     // through this entry point. `prepareConversationForMessage` constructs
     // host_browser proxies that capture `conversation.getCurrentSender()`
-    // directly, which would route browser frames through the daemon SSE
-    // channel instead of the `ChromeExtensionRegistry`. Chrome-extension
-    // flows reach host_browser exclusively through the
+    // directly, which routes browser frames through the daemon SSE channel.
+    // This is correct for macOS (SSE-based host proxy), but chrome-extension
+    // requires the `ChromeExtensionRegistry` WebSocket transport instead.
+    // Chrome-extension flows reach host_browser exclusively through the
     // `/v1/messages` flow in `conversation-routes.ts`, which wires a
     // registry-aware sender and sets `hostBrowserSenderOverride`.
     //

package/src/daemon/shutdown-handlers.ts

@@ -3,7 +3,6 @@ import * as Sentry from "@sentry/node";
 import type { BackupWorkerHandle } from "../backup/backup-worker.js";
 import type { FilingService } from "../filing/filing-service.js";
 import type { HeartbeatService } from "../heartbeat/heartbeat-service.js";
-import type { HookManager } from "../hooks/manager.js";
 import type { McpServerManager } from "../mcp/manager.js";
 import { getSqlite, resetDb } from "../memory/db.js";
 import type { QdrantManager } from "../memory/qdrant-manager.js";
@@ -23,7 +22,6 @@ export interface ShutdownDeps {
   workspaceHeartbeat: WorkspaceHeartbeatService;
   heartbeat: HeartbeatService;
   filing: FilingService;
-  hookManager: HookManager;
   runtimeHttp: RuntimeHttpServer | null;
   scheduler: { stop(): void };
   feedScheduler: { stop(): void } | null;
@@ -44,11 +42,9 @@ export function installShutdownHandlers(deps: ShutdownDeps): void {
     shuttingDown = true;
     log.info("Shutting down daemon...");
 
-    deps.hookManager.stopWatching();
-
     // Force exit if graceful shutdown takes too long.
-    // Set this BEFORE awaiting heartbeat stop and triggering daemon-stop hooks
-    // so it covers all potentially-blocking async shutdown work.
+    // Set this BEFORE awaiting heartbeat stop so it covers all
+    // potentially-blocking async shutdown work.
     //
     // 20s budget: 15s reserved for Meet session teardown
     // (`MeetSessionManager.shutdownAll`), plus ~5s for the remaining
@@ -67,12 +63,6 @@ export function installShutdownHandlers(deps: ShutdownDeps): void {
     await deps.heartbeat.stop();
     await deps.filing.stop();
 
-    try {
-      await deps.hookManager.trigger("daemon-stop", { pid: process.pid });
-    } catch {
-      // Don't let hook failures block shutdown
-    }
-
     // Run registered skill shutdown hooks (e.g. meet-join session teardown)
     // before stopping the server so any HTTP round-trips and SSE emissions
     // still have live transports.

package/src/daemon/tool-side-effects.ts

@@ -7,9 +7,7 @@
  * registry entry instead of another if/else branch.
  */
 
-import { compileApp } from "../bundler/app-compiler.js";
 import { generateAppIcon } from "../media/app-icon-generator.js";
-import { getApp, getAppDirPath, isMultifileApp } from "../memory/app-store.js";
 import { findActiveSession } from "../runtime/channel-verification-service.js";
 import { deliverVerificationSlack } from "../runtime/verification-outbound-actions.js";
 import { updatePublishedAppDeployment } from "../services/published-app-updater.js";
@@ -39,41 +37,22 @@ export type PostExecutionHook = (
 
 // ── Helpers ──────────────────────────────────────────────────────────
 
-/** Shared logic for refreshing app surfaces, broadcasting changes, and triggering auto-deploy. */
-function handleAppChange(
+/**
+ * Propagate an app change to connected clients and the publish pipeline.
+ *
+ * Compilation is the responsibility of the tool executor (see
+ * `executeAppCreate`, `executeAppRefresh`): executors own the source→dist
+ * transform and surface `compile_errors` in their result when it fails.
+ * Post-execution hooks only observe the outcome and notify — they must
+ * not re-run a compile because `compileApp()` begins with `rm -rf dist/`
+ * and would race with the executor's own output (LUM-1153).
+ */
+function notifyAppChanged(
   ctx: ToolSetupContext,
   appId: string,
   broadcastToAllClients: ((msg: ServerMessage) => void) | undefined,
   opts?: { fileChange?: boolean; status?: string },
 ): void {
-  const app = getApp(appId);
-
-  // Multifile apps need a recompile before refreshing surfaces so the
-  // WebView picks up the latest compiled output.
-  if (app && isMultifileApp(app)) {
-    const appDir = getAppDirPath(appId);
-    void compileApp(appDir)
-      .then((result) => {
-        if (!result.ok) {
-          log.warn(
-            { appId, errors: result.errors },
-            "Recompile failed on app change, serving stale dist/",
-          );
-        }
-        refreshSurfacesForApp(ctx, appId, opts);
-        broadcastToAllClients?.({ type: "app_files_changed", appId });
-        void updatePublishedAppDeployment(appId);
-      })
-      .catch((err) => {
-        log.warn({ appId, err }, "Recompile threw on app change");
-        // Still refresh surfaces with stale output
-        refreshSurfacesForApp(ctx, appId, opts);
-        broadcastToAllClients?.({ type: "app_files_changed", appId });
-        void updatePublishedAppDeployment(appId);
-      });
-    return;
-  }
-
   refreshSurfacesForApp(ctx, appId, opts);
   broadcastToAllClients?.({ type: "app_files_changed", appId });
   void updatePublishedAppDeployment(appId);
@@ -115,7 +94,7 @@ registerHook(
         // trigger live reload.
         ensureAppSourceWatcher();
 
-        handleAppChange(ctx, parsed.id, broadcastToAllClients);
+        notifyAppChanged(ctx, parsed.id, broadcastToAllClients);
 
         // Fire-and-forget: generate an app icon in the background.
         // When complete, broadcast again so clients pick up the new icon.
@@ -165,33 +144,12 @@ registerHook(
 );
 
 // Trigger surface refresh + broadcast when an app is refreshed.
-// If the executor already compiled (multifile path), skip the redundant
-// recompile and just refresh surfaces / broadcast / deploy directly.
 registerHook(
   "app_refresh",
-  (_name, input, result, { ctx, broadcastToAllClients }) => {
+  (_name, input, _result, { ctx, broadcastToAllClients }) => {
     const appId = input.app_id as string | undefined;
     if (!appId) return;
-
-    // executeAppRefresh already compiled multifile apps and included a
-    // "compiled" field in the result. Skip the expensive recompile and
-    // go straight to surface refresh + broadcast + deploy.
-    let alreadyCompiled = false;
-    try {
-      const parsed = JSON.parse(result.content) as { compiled?: boolean };
-      alreadyCompiled = parsed.compiled !== undefined;
-    } catch {
-      // Result wasn't valid JSON — fall through to handleAppChange.
-    }
-
-    if (alreadyCompiled) {
-      const opts = { fileChange: true };
-      refreshSurfacesForApp(ctx, appId, opts);
-      broadcastToAllClients?.({ type: "app_files_changed", appId });
-      void updatePublishedAppDeployment(appId);
-    } else {
-      handleAppChange(ctx, appId, broadcastToAllClients, { fileChange: true });
-    }
+    notifyAppChanged(ctx, appId, broadcastToAllClients, { fileChange: true });
   },
 );