@vellumai/assistant 0.4.56 → 0.4.57

package/src/daemon/conversation-agent-loop-handlers.ts

@@ -168,12 +168,12 @@ export function emitLlmCallStartedIfNeeded(
 }
 
 // ── Client Payload Size Caps ─────────────────────────────────────────
-// The client truncates tool results anyway (2 000 chars in ChatViewModel),
+// The client truncates tool results anyway (20 000 chars in ChatViewModel),
 // but the full string can be megabytes (file_read, bash output). Capping
 // here avoids sending oversized payloads which get decoded on the
 // client's main thread.
 
-const TOOL_RESULT_MAX_CHARS = 2_000;
+const TOOL_RESULT_MAX_CHARS = 20_000;
 const TOOL_RESULT_TRUNCATION_SUFFIX = "...[truncated]";
 
 // tool_input_delta streams accumulated JSON as tools run. For non-app

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

@@ -49,7 +49,6 @@ import {
   queueRegenerateConversationTitle,
   UNTITLED_FALLBACK,
 } from "../memory/conversation-title-service.js";
-import { stripMemoryRecallMessages } from "../memory/retriever.js";
 import type { PermissionPrompter } from "../permissions/prompter.js";
 import type { ContentBlock, Message } from "../providers/types.js";
 import type { Provider } from "../providers/types.js";
@@ -915,14 +914,7 @@ export async function runAgentLoopImpl(
 
       // Strip injected context from updated history before compacting,
       // so we compact the "raw" persistent messages.
-      const rawHistory = stripInjectedContext(updatedHistory, {
-        stripRecall: (msgs) =>
-          stripMemoryRecallMessages(
-            msgs,
-            recall.injectedText,
-            "separate_context_message",
-          ),
-      });
+      const rawHistory = stripInjectedContext(updatedHistory);
       ctx.messages = rawHistory;
 
       ctx.emitActivityState(
@@ -1055,14 +1047,7 @@ export async function runAgentLoopImpl(
     // convergence loop operates on the full (larger) history.
     if (state.contextTooLargeDetected) {
       if (updatedHistory.length > preRunHistoryLength) {
-        ctx.messages = stripInjectedContext(updatedHistory, {
-          stripRecall: (msgs) =>
-            stripMemoryRecallMessages(
-              msgs,
-              recall.injectedText,
-              "separate_context_message",
-            ),
-        });
+        ctx.messages = stripInjectedContext(updatedHistory);
         preRepairMessages = updatedHistory;
         preRunHistoryLength = updatedHistory.length;
       }
@@ -1509,22 +1494,14 @@ export async function runAgentLoopImpl(
         errChannelMeta,
       );
       newMessages.push(errorAssistantMessage);
-      onEvent({
-        type: "assistant_text_delta",
-        text: state.providerErrorUserMessage,
-        conversationId: ctx.conversationId,
-      });
+      // Do NOT send assistant_text_delta here — handleProviderError already
+      // emitted a conversation_error event for this same error text, and the
+      // client renders it as an InlineChatErrorAlert. Sending a text delta
+      // would create a duplicate plain-text bubble below the alert card.
     }
 
     const restoredHistory = [...preRepairMessages, ...newMessages];
-    ctx.messages = stripInjectedContext(restoredHistory, {
-      stripRecall: (msgs) =>
-        stripMemoryRecallMessages(
-          msgs,
-          recall.injectedText,
-          "separate_context_message",
-        ),
-    });
+    ctx.messages = stripInjectedContext(restoredHistory);
 
     emitUsage(
       ctx,

package/src/daemon/conversation-error.ts

@@ -23,6 +23,7 @@ const NETWORK_PATTERNS = [
   /ETIMEDOUT/i,
   /ENOTFOUND/i,
   /socket hang up/i,
+  /socket.*closed unexpectedly/i,
   /network.*error/i,
   /fetch failed/i,
   /connection.*refused/i,
@@ -85,6 +86,14 @@ const WEB_SEARCH_ORDERING_PATTERNS = [
   /web_search.*tool_result/i,
 ];
 
+// Streaming corruption patterns (Anthropic SDK throws non-HTTP errors for SSE issues)
+const STREAMING_ERROR_PATTERNS = [
+  /unexpected event order/i,
+  /stream ended without producing/i,
+  /request ended without sending any chunks/i,
+  /stream has ended.*this shouldn't happen/i,
+];
+
 // User-initiated cancellation patterns — these should NOT produce conversation_error
 const CANCEL_PATTERNS = [/abort/i, /cancel/i];
 
@@ -288,6 +297,11 @@ export function isOrderingError(message: string): boolean {
   return ORDERING_ERROR_PATTERNS.some((p) => p.test(message));
 }
 
+/** Check whether an error message indicates an Anthropic SDK streaming corruption. */
+export function isStreamingError(message: string): boolean {
+  return STREAMING_ERROR_PATTERNS.some((p) => p.test(message));
+}
+
 function classifyByMessage(
   message: string,
 ): Omit<ClassifiedConversationError, "debugDetails"> {
@@ -372,6 +386,16 @@ function classifyByMessage(
     }
   }
 
+  // Streaming corruption errors (Anthropic SDK SSE issues — transient, retryable)
+  if (isStreamingError(message)) {
+    return {
+      code: "PROVIDER_API",
+      userMessage: "The AI provider's response was interrupted. Please try again.",
+      retryable: true,
+      errorCategory: "stream_corruption",
+    };
+  }
+
   // Non-user abort/failure (e.g. AbortError from internal logic, not user cancel)
   for (const pattern of CANCEL_PATTERNS) {
     if (pattern.test(message)) {

package/src/daemon/conversation-memory.ts

@@ -4,7 +4,7 @@ import { buildMemoryQuery } from "../memory/query-builder.js";
 import { computeRecallBudget } from "../memory/retrieval-budget.js";
 import {
   buildMemoryRecall,
-  injectMemoryRecallAsSeparateMessage,
+  injectMemoryRecallAsUserBlock,
 } from "../memory/retriever.js";
 import type { ScopePolicyOverride } from "../memory/search/types.js";
 import type { Message } from "../providers/types.js";
@@ -68,9 +68,10 @@ export function needsMemory(messages: Message[], content: string): boolean {
  * pipeline. Returns the augmented run messages and metadata for
  * downstream event emission.
  *
- * The V2 pipeline always uses `separate_context_message` injection
- * strategy (user + assistant ack pair). When injection text is empty,
- * no synthetic messages are added.
+ * Memory context is injected as a text content block prepended to the
+ * last user message (same pattern as workspace/temporal injections).
+ * Stripping is handled by `stripUserTextBlocksByPrefix` matching the
+ * `<memory_context __injected>` prefix in `RUNTIME_INJECTION_PREFIXES`.
  */
 export async function prepareMemoryContext(
   ctx: MemoryPrepareContext,
@@ -165,13 +166,13 @@ export async function prepareMemoryContext(
     model: recall.model,
   });
 
-  // Inject recall into messages using separate_context_message strategy.
-  // When injection text is empty, skip injection entirely (no synthetic messages).
+  // Inject recall as a text block prepended to the last user message.
+  // When injection text is empty, skip injection entirely.
   let runMessages = ctx.messages;
   if (recall.injectedText.length > 0) {
     const userTail = ctx.messages[ctx.messages.length - 1];
     if (userTail && userTail.role === "user") {
-      runMessages = injectMemoryRecallAsSeparateMessage(
+      runMessages = injectMemoryRecallAsUserBlock(
         ctx.messages,
         recall.injectedText,
       );