@vellumai/assistant 0.4.44 → 0.4.45

package/src/daemon/session-agent-loop.ts

@@ -139,6 +139,7 @@ const TOOL_FRIENDLY_LABEL: Record<string, string> = {
   app_create: "Create App",
   app_update: "Update App",
   skill_load: "Load Skill",
+  skill_execute: "Run Skill Tool",
   app_file_edit: "Edit App File",
   app_file_write: "Write App File",
 };
@@ -226,6 +227,7 @@ export interface AgentLoopSessionContext {
       | "thinking_delta"
       | "first_text_delta"
       | "tool_use_start"
+      | "preview_start"
       | "tool_result_received"
       | "confirmation_requested"
       | "confirmation_resolved"
@@ -1301,60 +1303,10 @@ export async function runAgentLoopImpl(
       sessionId: ctx.conversationId,
     });
 
-    // Resolve attachments
-    const attachmentResult = await resolveAssistantAttachments(
-      state.accumulatedDirectives,
-      state.accumulatedToolContentBlocks,
-      state.directiveWarnings,
-      ctx.workingDir,
-      async (filePath) =>
-        approveHostAttachmentRead(
-          filePath,
-          ctx.workingDir,
-          ctx.prompter,
-          ctx.conversationId,
-          ctx.hasNoClient,
-        ),
-      state.lastAssistantMessageId,
-      state.toolContentBlockToolNames,
-    );
-    const { assistantAttachments, emittedAttachments } = attachmentResult;
-
-    ctx.lastAssistantAttachments = assistantAttachments;
-    ctx.lastAttachmentWarnings = attachmentResult.directiveWarnings;
-
-    const warningText = formatAttachmentWarnings(
-      attachmentResult.directiveWarnings,
-    );
-    if (warningText) {
-      onEvent({
-        type: "assistant_text_delta",
-        text: warningText,
-        sessionId: ctx.conversationId,
-      });
-    }
-
-    // Emit completion event
-    if (yieldedForHandoff) {
-      ctx.traceEmitter.emit(
-        "generation_handoff",
-        "Handing off to next queued message",
-        {
-          requestId: reqId,
-          status: "info",
-          attributes: { queuedCount: ctx.getQueueDepth() },
-        },
-      );
-      onEvent({
-        type: "generation_handoff",
-        sessionId: ctx.conversationId,
-        requestId: reqId,
-        queuedCount: ctx.getQueueDepth(),
-        ...(emittedAttachments.length > 0
-          ? { attachments: emittedAttachments }
-          : {}),
-      });
-    } else if (abortController.signal.aborted) {
+    // Fast-path: when the user cancelled, skip expensive post-loop work
+    // (attachment resolution) and emit the cancellation event immediately
+    // so the client can re-enable the UI without delay.
+    if (abortController.signal.aborted) {
       ctx.emitActivityState("idle", "generation_cancelled", "global", reqId);
       ctx.traceEmitter.emit(
         "generation_cancelled",
@@ -1366,18 +1318,97 @@ export async function runAgentLoopImpl(
       );
       onEvent({ type: "generation_cancelled", sessionId: ctx.conversationId });
     } else {
-      ctx.emitActivityState("idle", "message_complete", "global", reqId);
-      ctx.traceEmitter.emit("message_complete", "Message processing complete", {
-        requestId: reqId,
-        status: "success",
-      });
-      onEvent({
-        type: "message_complete",
-        sessionId: ctx.conversationId,
-        ...(emittedAttachments.length > 0
-          ? { attachments: emittedAttachments }
-          : {}),
-      });
+      // Resolve attachments (only when not cancelled — this is expensive async I/O)
+      const attachmentResult = await resolveAssistantAttachments(
+        state.accumulatedDirectives,
+        state.accumulatedToolContentBlocks,
+        state.directiveWarnings,
+        ctx.workingDir,
+        async (filePath) =>
+          approveHostAttachmentRead(
+            filePath,
+            ctx.workingDir,
+            ctx.prompter,
+            ctx.conversationId,
+            ctx.hasNoClient,
+          ),
+        state.lastAssistantMessageId,
+        state.toolContentBlockToolNames,
+      );
+      const { assistantAttachments, emittedAttachments } = attachmentResult;
+
+      ctx.lastAssistantAttachments = assistantAttachments;
+      ctx.lastAttachmentWarnings = attachmentResult.directiveWarnings;
+
+      const warningText = formatAttachmentWarnings(
+        attachmentResult.directiveWarnings,
+      );
+      if (warningText) {
+        onEvent({
+          type: "assistant_text_delta",
+          text: warningText,
+          sessionId: ctx.conversationId,
+        });
+      }
+
+      // Re-check: the user may have cancelled during attachment resolution
+      if (abortController.signal.aborted) {
+        ctx.emitActivityState("idle", "generation_cancelled", "global", reqId);
+        ctx.traceEmitter.emit(
+          "generation_cancelled",
+          "Generation cancelled by user",
+          {
+            requestId: reqId,
+            status: "warning",
+          },
+        );
+        onEvent({
+          type: "generation_cancelled",
+          sessionId: ctx.conversationId,
+        });
+      } else if (yieldedForHandoff) {
+        ctx.traceEmitter.emit(
+          "generation_handoff",
+          "Handing off to next queued message",
+          {
+            requestId: reqId,
+            status: "info",
+            attributes: { queuedCount: ctx.getQueueDepth() },
+          },
+        );
+        onEvent({
+          type: "generation_handoff",
+          sessionId: ctx.conversationId,
+          requestId: reqId,
+          queuedCount: ctx.getQueueDepth(),
+          ...(emittedAttachments.length > 0
+            ? { attachments: emittedAttachments }
+            : {}),
+          ...(state.lastAssistantMessageId
+            ? { messageId: state.lastAssistantMessageId }
+            : {}),
+        });
+      } else {
+        ctx.emitActivityState("idle", "message_complete", "global", reqId);
+        ctx.traceEmitter.emit(
+          "message_complete",
+          "Message processing complete",
+          {
+            requestId: reqId,
+            status: "success",
+          },
+        );
+        onEvent({
+          type: "message_complete",
+          sessionId: ctx.conversationId,
+          ...(emittedAttachments.length > 0
+            ? { attachments: emittedAttachments }
+            : {}),
+          ...(state.lastAssistantMessageId
+            ? { messageId: state.lastAssistantMessageId }
+            : {}),
+        });
+      }
     }
 
     // Second title pass: after 3 completed turns, re-generate the title

package/src/daemon/session-attachments.ts

@@ -199,7 +199,7 @@ export async function resolveAssistantAttachments(
   }
 
   // Persist resolved attachments and link to the last assistant message.
-  // Large video attachments are omitted from the IPC payload and lazy-loaded
+  // Large video attachments are omitted from the event payload and lazy-loaded
   // by the client via the HTTP endpoint (same pattern as history_response).
   const MAX_INLINE_B64_SIZE = 512 * 1024;
 

package/src/daemon/session-error.ts

@@ -1,8 +1,11 @@
 import { ProviderError } from "../util/errors.js";
-import type { SessionErrorCode, SessionErrorMessage } from "./message-protocol.js";
+import type {
+  SessionErrorCode,
+  SessionErrorMessage,
+} from "./message-protocol.js";
 
 /**
- * Classified session error ready for IPC emission.
+ * Classified session error ready for client emission.
  */
 export interface ClassifiedSessionError {
   code: SessionErrorCode;
@@ -73,7 +76,7 @@ const CANCEL_PATTERNS = [/abort/i, /cancel/i];
  */
 export interface ErrorContext {
   /** Where in the processing pipeline the error occurred. */
-  phase: "agent_loop" | "queue" | "regenerate" | "handler" | "persist";
+  phase: "agent_loop" | "regenerate" | "handler" | "persist";
   /** Whether the abort signal was active when the error occurred. */
   aborted?: boolean;
 }
@@ -90,11 +93,11 @@ export function isUserCancellation(error: unknown, ctx: ErrorContext): boolean {
   return false;
 }
 
-/** Maximum length for debugDetails to prevent unbounded IPC payloads. */
+/** Maximum length for debugDetails to prevent unbounded event payloads. */
 const MAX_DEBUG_DETAIL_LENGTH = 4000;
 
 /**
- * Truncate debug details to a reasonable size for IPC transport.
+ * Truncate debug details to a reasonable size for transport.
  */
 function truncateDebugDetails(details: string): string {
   if (details.length <= MAX_DEBUG_DETAIL_LENGTH) return details;
@@ -121,15 +124,6 @@ export function classifySessionError(
   const debugDetails = truncateDebugDetails(rawDetails);
 
   // Phase-specific overrides
-  if (ctx.phase === "queue") {
-    return {
-      code: "QUEUE_FULL",
-      userMessage: "Message queue is full (10 messages pending).",
-      retryable: true,
-      debugDetails: truncateDebugDetails(message),
-    };
-  }
-
   if (ctx.phase === "regenerate") {
     const base = classifyCore(error, message);
     return {
@@ -165,6 +159,13 @@ function classifyCore(
         retryable: false,
       };
     }
+    if (error.statusCode === 401) {
+      return {
+        code: "PROVIDER_BILLING",
+        userMessage: "Your API key is invalid or expired.",
+        retryable: false,
+      };
+    }
     if (error.statusCode === 429) {
       return {
         code: "PROVIDER_RATE_LIMIT",
@@ -209,7 +210,7 @@ function classifyCore(
       return {
         code: "PROVIDER_API",
         userMessage: "The AI provider rejected the request.",
-        retryable: false,
+        retryable: true,
       };
     }
   }
@@ -311,7 +312,7 @@ function classifyByMessage(
 }
 
 /**
- * Build a `session_error` IPC message from a classified error.
+ * Build a `session_error` server message from a classified error.
  */
 export function buildSessionErrorMessage(
   sessionId: string,

package/src/daemon/session-lifecycle.ts

@@ -24,7 +24,11 @@ import {
 import { unregisterSessionSender } from "../tools/browser/browser-screencast.js";
 import { getLogger } from "../util/logger.js";
 import { repairHistory } from "./history-repair.js";
-import type { SurfaceData, SurfaceType, UsageStats } from "./message-protocol.js";
+import type {
+  SurfaceData,
+  SurfaceType,
+  UsageStats,
+} from "./message-protocol.js";
 import {
   unregisterCallNotifiers,
   unregisterWatchNotifiers,
@@ -74,6 +78,7 @@ export interface LoadFromDbContext {
   contextCompactedAt: number | null;
   trustContext?: { trustClass: TrustClass };
   loadedHistoryTrustClass?: TrustClass;
+  hasAttachments?: boolean;
 }
 
 export interface AbortContext {
@@ -177,6 +182,20 @@ export async function loadFromDb(ctx: LoadFromDbContext): Promise<void> {
 
   ctx.loadedHistoryTrustClass = trustClass;
 
+  // Scan ALL db messages (including compacted ones) for attachments so that
+  // asset tools remain available after context compaction.
+  if (
+    ctx.contextCompactedMessageCount > 0 &&
+    dbMessages.some(
+      (m) =>
+        m.role === "user" &&
+        (m.content.includes('"type":"image"') ||
+          m.content.includes('"type":"file"')),
+    )
+  ) {
+    ctx.hasAttachments = true;
+  }
+
   log.info(
     { conversationId: ctx.conversationId, count: ctx.messages.length },
     "Loaded messages from DB",

package/src/daemon/session-media-retry.ts

@@ -19,29 +19,15 @@ export function stripMediaPayloadsForRetry(messages: Message[]): {
   latestUserIndex: number | null;
 } {
   let latestUserIndex: number | null = null;
-  let lastSummaryUserIndex: number | null = null;
   for (let i = messages.length - 1; i >= 0; i--) {
     const msg = messages[i];
     if (msg.role !== "user") continue;
     if (isToolResultOnlyMessage(msg)) continue;
-    if (getSummaryFromContextMessage(msg) != null) {
-      // Track the last summary message as a fallback — after aggressive
-      // compaction (minKeepRecentUserTurns: 0), the summary may be the only
-      // user message left and it can contain preserved image blocks that
-      // should not be stripped.
-      if (lastSummaryUserIndex == null) lastSummaryUserIndex = i;
-      continue;
-    }
+    if (getSummaryFromContextMessage(msg) != null) continue;
     latestUserIndex = i;
     break;
   }
 
-  // Fall back to the summary message when compaction consumed all user turns,
-  // so images preserved by compaction are not unconditionally stripped.
-  if (latestUserIndex == null && lastSummaryUserIndex != null) {
-    latestUserIndex = lastSummaryUserIndex;
-  }
-
   let modified = false;
   let replacedBlocks = 0;
   let keptLatestMediaBlocks = 0;

package/src/daemon/session-messaging.ts

@@ -28,7 +28,10 @@ import {
 import type { SecretPrompter } from "../permissions/secret-prompter.js";
 import type { Message } from "../providers/types.js";
 import { getLogger } from "../util/logger.js";
-import type { ServerMessage, UserMessageAttachment } from "./message-protocol.js";
+import type {
+  ServerMessage,
+  UserMessageAttachment,
+} from "./message-protocol.js";
 import type { MessageQueue } from "./session-queue-manager.js";
 import type { TrustContext } from "./session-runtime-assembly.js";
 
@@ -212,7 +215,7 @@ export function enqueueMessage(
   metadata?: Record<string, unknown>,
   options?: { isInteractive?: boolean },
   displayContent?: string,
-): { queued: boolean; rejected?: boolean; requestId: string } {
+): { queued: boolean; requestId: string; rejected?: boolean } {
   if (!ctx.processing) {
     return { queued: false, requestId };
   }
@@ -225,7 +228,7 @@ export function enqueueMessage(
     extractTurnInterfaceContext(metadata) ??
     ctx.getTurnInterfaceContext() ??
     undefined;
-  const pushed = ctx.queue.push({
+  const accepted = ctx.queue.push({
     content,
     attachments,
     requestId,
@@ -236,11 +239,16 @@ export function enqueueMessage(
     turnChannelContext,
     turnInterfaceContext,
     isInteractive: options?.isInteractive,
-    queuedAt: Date.now(),
     displayContent,
   });
-  if (!pushed) {
-    return { queued: false, rejected: true, requestId };
+  if (!accepted) {
+    onEvent({
+      type: "error",
+      message:
+        "The assistant is busy and cannot accept more messages right now. Please try again shortly.",
+      category: "queue_full",
+    });
+    return { queued: false, requestId, rejected: true };
   }
   return { queued: true, requestId };
 }

package/src/daemon/session-process.ts

@@ -36,14 +36,18 @@ import type {
 import type { MessageQueue } from "./session-queue-manager.js";
 import type { QueueDrainReason } from "./session-queue-manager.js";
 import type { TrustContext } from "./session-runtime-assembly.js";
-import { resolveSlash, type SlashContext } from "./session-slash.js";
+import {
+  isProviderShortcut,
+  resolveSlash,
+  type SlashContext,
+} from "./session-slash.js";
 import type { TraceEmitter } from "./trace-emitter.js";
 import { resolveVerificationSessionIntent } from "./verification-session-intent.js";
 
 const log = getLogger("session-process");
 
 /** Build a model_info event with fresh config data. */
-function buildModelInfoEvent(): ServerMessage {
+export function buildModelInfoEvent(): ServerMessage {
   const config = getConfig();
   const configured = Object.keys(config.apiKeys).filter(
     (k) => !!config.apiKeys[k],
@@ -58,7 +62,7 @@ function buildModelInfoEvent(): ServerMessage {
 }
 
 /** True when the trimmed content is a /model or /models slash command. */
-function isModelSlashCommand(content: string): boolean {
+export function isModelSlashCommand(content: string): boolean {
   const trimmed = content.trim();
   return (
     trimmed === "/model" ||
@@ -114,6 +118,10 @@ export interface ProcessSessionContext {
   setTurnChannelContext(ctx: TurnChannelContext): void;
   getTurnInterfaceContext(): TurnInterfaceContext | null;
   setTurnInterfaceContext(ctx: TurnInterfaceContext): void;
+  /** Mark host proxies as unavailable so tool execution uses local fallback. */
+  clearProxyAvailability(): void;
+  /** Restore host proxy availability based on whether a real client is connected. */
+  restoreProxyAvailability(): void;
   emitActivityState(
     phase:
       | "idle"
@@ -131,7 +139,8 @@ export interface ProcessSessionContext {
       | "confirmation_resolved"
       | "message_complete"
       | "generation_cancelled"
-      | "error_terminal",
+      | "error_terminal"
+      | "preview_start",
     anchor?: "assistant_turn" | "user_turn" | "global",
     requestId?: string,
     statusText?: string,
@@ -260,6 +269,23 @@ export async function drainQueue(
     session.setTurnInterfaceContext(queuedInterfaceCtx);
   }
 
+  // Non-interactive queued messages (channel requests) must not execute tools
+  // via the desktop host proxy. Clear proxy availability so isAvailable()
+  // returns false and tool execution falls back to local.
+  if (next.isInteractive === false) {
+    session.clearProxyAvailability();
+  } else {
+    // Restore proxy availability only for desktop-originating turns (macos/ios)
+    // in case a prior non-interactive drain disabled it. Non-desktop interactive
+    // interfaces (CLI, Vellum) should not re-enable desktop host proxies.
+    const interfaceCtx =
+      queuedInterfaceCtx ?? session.getTurnInterfaceContext();
+    const sourceInterface = interfaceCtx?.userMessageInterface;
+    if (sourceInterface === "macos" || sourceInterface === "ios") {
+      session.restoreProxyAvailability();
+    }
+  }
+
   // Resolve slash commands for queued messages
   const slashResult = resolveSlash(next.content, buildSlashContext(session));
 
@@ -326,7 +352,10 @@ export async function drainQueue(
 
       // Emit fresh model info before the text delta so the client has
       // up-to-date configuredProviders when rendering /model or /models UI.
-      if (isModelSlashCommand(next.content)) {
+      if (
+        isModelSlashCommand(next.content) ||
+        isProviderShortcut(next.content)
+      ) {
         next.onEvent(buildModelInfoEvent());
       }
       next.onEvent({ type: "assistant_text_delta", text: slashResult.message });
@@ -510,7 +539,7 @@ export async function drainQueue(
 
 /**
  * Convenience function that persists a user message and runs the agent loop
- * in a single call. Used by the IPC path where blocking is expected.
+ * in a single call. Used by the message-handler path where blocking is expected.
  */
 export async function processMessage(
   session: ProcessSessionContext,
@@ -676,7 +705,7 @@ export async function processMessage(
 
     // Emit fresh model info before the text delta so the client has
     // up-to-date configuredProviders when rendering /model or /models UI.
-    if (isModelSlashCommand(content)) {
+    if (isModelSlashCommand(content) || isProviderShortcut(content)) {
       onEvent(buildModelInfoEvent());
     }
     onEvent({ type: "assistant_text_delta", text: slashResult.message });