@vellumai/assistant 0.6.5 → 0.6.6

package/src/daemon/conversation-history.ts

@@ -191,10 +191,7 @@ export function consolidateAssistantMessages(
 
     // Clean up Qdrant vectors (fire-and-forget)
     if (allSegmentIds.length > 0) {
-      cleanupQdrantVectors(
-        conversationId,
-        allSegmentIds,
-      ).catch((err) => {
+      cleanupQdrantVectors(conversationId, allSegmentIds).catch((err) => {
         log.warn(
           { err, conversationId },
           "Qdrant cleanup after consolidation failed (non-fatal)",
@@ -326,10 +323,7 @@ export function consolidateAssistantMessages(
 
   // Clean up Qdrant vectors (fire-and-forget)
   if (allSegmentIds.length > 0) {
-    cleanupQdrantVectors(
-      conversationId,
-      allSegmentIds,
-    ).catch((err) => {
+    cleanupQdrantVectors(conversationId, allSegmentIds).catch((err) => {
       log.warn(
         { err, conversationId },
         "Qdrant cleanup after consolidation failed (non-fatal)",
@@ -365,7 +359,6 @@ export interface HistoryConversationContext {
     userMessageId: string,
     onEvent: (msg: ServerMessage) => void,
     options?: {
-      skipPreMessageRollback?: boolean;
       isUserMessage?: boolean;
       titleText?: string;
     },
@@ -534,15 +527,14 @@ export async function regenerate(
   }
 
   // Clean up Qdrant vectors (fire-and-forget).
-  cleanupQdrantVectors(
-    conversation.conversationId,
-    allSegmentIds,
-  ).catch((err) => {
-    log.warn(
-      { err, conversationId: conversation.conversationId },
-      "Qdrant cleanup after regenerate failed (non-fatal)",
-    );
-  });
+  cleanupQdrantVectors(conversation.conversationId, allSegmentIds).catch(
+    (err) => {
+      log.warn(
+        { err, conversationId: conversation.conversationId },
+        "Qdrant cleanup after regenerate failed (non-fatal)",
+      );
+    },
+  );
 
   // Re-extract the user message content for the agent loop.
   // Use all content blocks (text, image, file) so attachments are
@@ -581,7 +573,6 @@ export async function regenerate(
   // observability contract is preserved on those paths too.
   void conversation
     .runAgentLoop(content, existingUserMessageId, onEvent, {
-      skipPreMessageRollback: true,
       isUserMessage: true,
     })
     .catch((err) => {

package/src/daemon/conversation-lifecycle.ts

@@ -8,7 +8,6 @@ import { createContextSummaryMessage } from "../context/window-manager.js";
 import type { EventBus } from "../events/bus.js";
 import type { AssistantDomainEvents } from "../events/domain-events.js";
 import type { ToolProfiler } from "../events/tool-profiling-listener.js";
-import { getHookManager } from "../hooks/manager.js";
 import { enqueueAutoAnalysisIfEnabled } from "../memory/auto-analysis-enqueue.js";
 import { isAutoAnalysisConversation } from "../memory/auto-analysis-guard.js";
 import {
@@ -27,10 +26,7 @@ import {
 import { unregisterConversationSender } from "../tools/browser/browser-screencast.js";
 import { type AbortReason, createAbortReason } from "../util/abort-reasons.js";
 import { getLogger } from "../util/logger.js";
-import {
-  unregisterCallNotifiers,
-  unregisterWatchNotifiers,
-} from "./conversation-notifiers.js";
+import { unregisterCallNotifiers } from "./conversation-notifiers.js";
 import type { MessageQueue } from "./conversation-queue-manager.js";
 import { resetSkillToolProjection } from "./conversation-skill-tools.js";
 import { repairHistory } from "./history-repair.js";
@@ -208,21 +204,34 @@ export async function loadFromDb(ctx: LoadFromDbContext): Promise<void> {
 
       content = reinjectImageSourcePaths(content, role, m.metadata);
 
-      // Re-inject persisted memory block from metadata so it survives
+      // Re-inject persisted injection blocks from metadata so it survives
       // conversation reloads (eviction, restart, fork).
       if (role === "user" && m.metadata) {
         try {
           const meta = JSON.parse(m.metadata);
           const isTail = index === arr.length - 1;
 
-          // turn_context and system_reminder rehydrate for historical rows
-          // only; the tail gets fresh blocks via applyRuntimeInjections on
-          // the next turn. All three rehydration steps are prepends — the
-          // ordering below (system_reminder first, memory second,
-          // turn_context last) builds the documented shape right-to-left,
-          // since each prepend shifts previously-prepended blocks one slot
-          // right:
-          //   [<turn_context>, <memory __injected>, <system_reminder>, ...original]
+          // All rehydration steps are prepends — the ordering below
+          // (innermost block first, outermost last) builds the documented
+          // shape right-to-left, since each prepend shifts previously-
+          // prepended blocks one slot right:
+          //   [<workspace>, <turn_context>, <NOW.md>, <memory __injected>,
+          //    <system_reminder>, <knowledge_base>, ...original]
+          //
+          // Persisted non-tail rows rehydrate the full set so Anthropic's
+          // prefix cache keeps matching msg[0] across daemon restarts and
+          // conversation eviction. The tail row gets fresh blocks via
+          // applyRuntimeInjections on the next turn, so rehydration for the
+          // tail is limited to blocks that the next turn's injection pipeline
+          // cannot reconstruct (currently just `memoryInjectedBlock`, which
+          // is not always re-injected on the next turn).
+          if (!isTail && typeof meta.pkbContextBlock === "string") {
+            content = [
+              { type: "text" as const, text: meta.pkbContextBlock },
+              ...content,
+            ];
+          }
+
           if (!isTail && typeof meta.pkbSystemReminderBlock === "string") {
             content = [
               { type: "text" as const, text: meta.pkbSystemReminderBlock },
@@ -241,12 +250,26 @@ export async function loadFromDb(ctx: LoadFromDbContext): Promise<void> {
             ];
           }
 
+          if (!isTail && typeof meta.nowScratchpadBlock === "string") {
+            content = [
+              { type: "text" as const, text: meta.nowScratchpadBlock },
+              ...content,
+            ];
+          }
+
           if (!isTail && typeof meta.turnContextBlock === "string") {
             content = [
               { type: "text" as const, text: meta.turnContextBlock },
               ...content,
             ];
           }
+
+          if (!isTail && typeof meta.workspaceBlock === "string") {
+            content = [
+              { type: "text" as const, text: meta.workspaceBlock },
+              ...content,
+            ];
+          }
         } catch {
           /* ignore parse errors — metadata may be malformed */
         }
@@ -314,7 +337,6 @@ export function abortConversation(
     ctx.surfaceActionRequestIds.clear();
     ctx.surfaceState.clear();
     ctx.accumulatedSurfaceState.clear();
-    unregisterWatchNotifiers(ctx.conversationId);
     for (const queued of ctx.queue) {
       queued.onEvent({
         type: "generation_cancelled",
@@ -328,10 +350,6 @@ export function abortConversation(
 // ── dispose ──────────────────────────────────────────────────────────
 
 export function disposeConversation(ctx: DisposeContext): void {
-  void getHookManager().trigger("conversation-end", {
-    conversationId: ctx.conversationId,
-  });
-
   // Trigger graph extraction for end-of-conversation sweep.
   // Only extract from guardian conversations to preserve the memory trust
   // boundary — untrusted content must not influence future memory retrieval.

package/src/daemon/conversation-notifiers.ts

@@ -1,5 +1,5 @@
 /**
- * Watch and call notifier registration/unregistration, extracted from
+ * Call notifier registration/unregistration, extracted from
  * the Conversation constructor and dispose/abort methods.
  *
  * Notifier callbacks read from the provided context object at invocation
@@ -22,24 +22,9 @@ import {
   addMessage,
   provenanceFromTrustContext,
 } from "../memory/conversation-crud.js";
-import { setMessageIdOnLogs } from "../memory/llm-request-log-store.js";
 import type { Message } from "../providers/types.js";
-import type { WatchSession } from "../tools/watch/watch-state.js";
-import {
-  pruneWatchSessions,
-  registerWatchCommentaryNotifier,
-  registerWatchCompletionNotifier,
-  registerWatchStartNotifier,
-  unregisterWatchCommentaryNotifier,
-  unregisterWatchCompletionNotifier,
-  unregisterWatchStartNotifier,
-} from "../tools/watch/watch-state.js";
 import type { TrustContext } from "./conversation-runtime-assembly.js";
 import type { ServerMessage } from "./message-protocol.js";
-import {
-  lastCommentaryByConversation,
-  lastSummaryByConversation,
-} from "./watch-handler.js";
 
 /**
  * Subset of Conversation state that notifier callbacks need to read at
@@ -52,7 +37,7 @@ export interface NotifierConversationContext {
 }
 
 /**
- * Register watch and call notifiers for a conversation. Call once during
+ * Register call notifiers for a conversation. Call once during
  * construction; the notifier callbacks close over `ctx` so they see
  * live sendToClient/messages values.
  */
@@ -60,88 +45,6 @@ export function registerConversationNotifiers(
   conversationId: string,
   ctx: NotifierConversationContext,
 ): void {
-  registerWatchStartNotifier(conversationId, (session: WatchSession) => {
-    ctx.sendToClient({
-      type: "watch_started",
-      conversationId: conversationId,
-      watchId: session.watchId,
-      durationSeconds: session.durationSeconds,
-      intervalSeconds: session.intervalSeconds,
-    });
-  });
-
-  registerWatchCommentaryNotifier(
-    conversationId,
-    async (_session: WatchSession) => {
-      const result = lastCommentaryByConversation.get(conversationId);
-      if (result) {
-        lastCommentaryByConversation.delete(conversationId);
-
-        const msg = await addMessage(
-          conversationId,
-          "assistant",
-          JSON.stringify([{ type: "text", text: result.text }]),
-          provenanceFromTrustContext(ctx.trustContext),
-        );
-        ctx.messages.push(createAssistantMessage(result.text));
-
-        try {
-          setMessageIdOnLogs(result.logIds, msg.id);
-        } catch {
-          // non-fatal — message is persisted even if log linkage fails
-        }
-
-        ctx.sendToClient({
-          type: "assistant_text_delta",
-          text: result.text,
-          conversationId: conversationId,
-        });
-        ctx.sendToClient({
-          type: "message_complete",
-          conversationId: conversationId,
-          messageId: msg.id,
-          source: "aux",
-        });
-      }
-    },
-  );
-
-  registerWatchCompletionNotifier(
-    conversationId,
-    async (_session: WatchSession) => {
-      const result = lastSummaryByConversation.get(conversationId);
-      if (result) {
-        lastSummaryByConversation.delete(conversationId);
-
-        const msg = await addMessage(
-          conversationId,
-          "assistant",
-          JSON.stringify([{ type: "text", text: result.text }]),
-          provenanceFromTrustContext(ctx.trustContext),
-        );
-        ctx.messages.push(createAssistantMessage(result.text));
-
-        try {
-          setMessageIdOnLogs(result.logIds, msg.id);
-        } catch {
-          // non-fatal — message is persisted even if log linkage fails
-        }
-
-        ctx.sendToClient({
-          type: "assistant_text_delta",
-          text: result.text,
-          conversationId: conversationId,
-        });
-        ctx.sendToClient({
-          type: "message_complete",
-          conversationId: conversationId,
-          messageId: msg.id,
-          source: "aux",
-        });
-      }
-    },
-  );
-
   registerCallQuestionNotifier(
     conversationId,
     async (callSessionId: string, question: string) => {
@@ -213,17 +116,6 @@ export function registerConversationNotifiers(
   });
 }
 
-/**
- * Unregister watch notifiers and prune watch sessions. Called during
- * abort when the conversation is actively processing.
- */
-export function unregisterWatchNotifiers(conversationId: string): void {
-  unregisterWatchStartNotifier(conversationId);
-  unregisterWatchCommentaryNotifier(conversationId);
-  unregisterWatchCompletionNotifier(conversationId);
-  pruneWatchSessions(conversationId);
-}
-
 /**
  * Unregister call notifiers. Called during dispose regardless of
  * processing state (notifiers are registered in the constructor).

package/src/daemon/conversation-process.ts

@@ -137,7 +137,6 @@ export interface ProcessConversationContext {
     userMessageId: string,
     onEvent: (msg: ServerMessage) => void,
     options?: {
-      skipPreMessageRollback?: boolean;
       isInteractive?: boolean;
       isUserMessage?: boolean;
       titleText?: string;
@@ -157,7 +156,7 @@ export interface ProcessConversationContext {
    * separately with the correct registry-routed sender.
    */
   restoreProxyAvailability(options?: { skipBrowser?: boolean }): void;
-  /** Restore only the host browser proxy (used by chrome-extension and macOS+extension drains). */
+  /** Restore only the host browser proxy (used by chrome-extension and macOS drains). */
   restoreBrowserProxyAvailability(): void;
   /**
    * Registry-routed sender override for the host browser proxy. When set,
@@ -477,6 +476,10 @@ async function drainSingleMessage(
       // When hostBrowserSenderOverride is set, skip the browser proxy here
       // — restoreBrowserProxyAvailability() below will handle it with the
       // correct registry-routed sender instead of the SSE hub emitter.
+      // When no override is set (macOS without extension), restoring the
+      // browser proxy via restoreProxyAvailability is correct — it will
+      // use the SSE sender so host_browser_request frames reach the
+      // desktop client directly.
       conversation.restoreProxyAvailability(
         conversation.hostBrowserSenderOverride
           ? { skipBrowser: true }
@@ -485,11 +488,11 @@ async function drainSingleMessage(
       conversation.addPreactivatedSkillId("computer-use");
     }
     // Tear down a stale hostBrowserProxy inherited from a prior turn on a
-    // different interface (e.g. chrome-extension installed one, then a CLI
-    // turn drains). Without this, restoreProxyAvailability() above would
-    // re-enable the proxy and getCdpClient() would route browser tools
-    // through host_browser_request and hang waiting for a client that this
-    // turn's interface can't service.
+    // different interface (e.g. chrome-extension or macOS installed one,
+    // then a CLI turn drains). Without this, restoreProxyAvailability()
+    // above would re-enable the proxy and getCdpClient() would route
+    // browser tools through host_browser_request and hang waiting for a
+    // client that this turn's interface can't service.
     //
     // Skip teardown only when BOTH conditions hold:
     //   1. `hostBrowserSenderOverride` is set (live registry-routed sender)
@@ -498,6 +501,10 @@ async function drainSingleMessage(
     // Without the interface check, queued turns from CLI/iOS/Vellum would
     // inherit a stale override left by a prior extension-connected turn
     // and keep the proxy alive, causing cross-interface misrouting.
+    //
+    // macOS and chrome-extension both support host_browser natively, so
+    // this teardown only fires for interfaces that do not support
+    // host_browser at all (CLI, iOS, Vellum, channels).
     const currentTurnCanServiceBrowser =
       !!sourceInterface && canServiceRegistryBrowser(sourceInterface);
     if (