@vellumai/assistant 0.5.5 → 0.5.6

package/src/config/env.ts

@@ -51,9 +51,15 @@ export function getGatewayPort(): number {
   return int("GATEWAY_PORT", DEFAULT_GATEWAY_PORT);
 }
 
-/** Resolve the gateway base URL for internal service-to-service calls. */
+/**
+ * Resolve the gateway base URL for internal service-to-service calls.
+ *
+ * In containerized deployments the gateway runs in a separate container,
+ * reachable via `GATEWAY_INTERNAL_URL` (e.g. `http://gateway:7822`).
+ * Falls back to `http://127.0.0.1:<GATEWAY_PORT>` for local deployments.
+ */
 export function getGatewayInternalBaseUrl(): string {
-  return `http://127.0.0.1:${getGatewayPort()}`;
+  return str("GATEWAY_INTERNAL_URL") ?? `http://127.0.0.1:${getGatewayPort()}`;
 }
 
 // ── Ingress ──────────────────────────────────────────────────────────────────

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

@@ -25,14 +25,6 @@
       "description": "Show the Contacts tab in Settings for viewing and managing contacts",
       "defaultEnabled": true
     },
-    {
-      "id": "custom-inference-provider",
-      "scope": "macos",
-      "key": "custom_inference_provider_enabled",
-      "label": "Custom Inference Provider",
-      "description": "Allow selecting a specific LLM provider and model for inference in Your Own mode",
-      "defaultEnabled": false
-    },
     {
       "id": "email-channel",
       "scope": "assistant",

package/src/config/schema.ts

@@ -106,18 +106,6 @@ export {
   MemoryDynamicBudgetConfigSchema,
   MemoryRetrievalConfigSchema,
 } from "./schemas/memory-retrieval.js";
-export type {
-  MemorySimplifiedArchiveRecallConfig,
-  MemorySimplifiedBriefConfig,
-  MemorySimplifiedConfig,
-  MemorySimplifiedReducerConfig,
-} from "./schemas/memory-simplified.js";
-export {
-  MemorySimplifiedArchiveRecallConfigSchema,
-  MemorySimplifiedBriefConfigSchema,
-  MemorySimplifiedConfigSchema,
-  MemorySimplifiedReducerConfigSchema,
-} from "./schemas/memory-simplified.js";
 export type {
   MemoryEmbeddingsConfig,
   MemorySegmentationConfig,

package/src/config/schemas/memory.ts

@@ -10,7 +10,6 @@ import {
   MemorySummarizationConfigSchema,
 } from "./memory-processing.js";
 import { MemoryRetrievalConfigSchema } from "./memory-retrieval.js";
-import { MemorySimplifiedConfigSchema } from "./memory-simplified.js";
 import {
   MemoryEmbeddingsConfigSchema,
   MemorySegmentationConfigSchema,
@@ -46,9 +45,6 @@ export const MemoryConfigSchema = z
     summarization: MemorySummarizationConfigSchema.default(
       MemorySummarizationConfigSchema.parse({}),
     ),
-    simplified: MemorySimplifiedConfigSchema.default(
-      MemorySimplifiedConfigSchema.parse({}),
-    ),
   })
   .describe(
     "Long-term memory system — stores, retrieves, and manages persistent knowledge across conversations",

package/src/config/schemas/platform.ts

@@ -43,7 +43,7 @@ export const DaemonConfigSchema = z
       .number({ error: "daemon.titleGenerationMaxTokens must be a number" })
       .int("daemon.titleGenerationMaxTokens must be an integer")
       .positive("daemon.titleGenerationMaxTokens must be a positive integer")
-      .default(30)
+      .default(50)
       .describe(
         "Maximum number of tokens for auto-generated conversation titles",
       ),

package/src/config/schemas/security.ts

@@ -77,6 +77,10 @@ export const PermissionsConfigSchema = z
       .describe(
         "Permission mode — 'strict' requires explicit approval for all operations, 'workspace' allows operations within the workspace",
       ),
+    dangerouslySkipPermissions: z
+      .boolean({ error: "permissions.dangerouslySkipPermissions must be a boolean" })
+      .default(false)
+      .describe("Auto-accept all permission prompts without asking"),
   })
   .describe("Permission enforcement mode for tool operations");
 

package/src/context/window-manager.ts

@@ -538,12 +538,12 @@ export class ContextWindowManager {
     }
 
     const keepTurns = lo;
-    const keepFromIndex =
+    const rawKeepFromIndex =
       keepTurns === 0
         ? messages.length
         : (userTurnStarts[userTurnStarts.length - keepTurns] ??
           messages.length);
-
+    const keepFromIndex = adjustForToolPairs(messages, rawKeepFromIndex);
     return { keepFromIndex, keepTurns };
   }
 
@@ -703,6 +703,57 @@ function isToolResultOnly(message: Message): boolean {
   );
 }
 
+/**
+ * Walk the keep boundary backward to ensure tool_use/tool_result pairs are
+ * never split across the compaction boundary. If the first kept message is
+ * a user message containing tool_result blocks whose matching tool_use blocks
+ * live in the preceding (compacted-away) assistant message, include that
+ * assistant message in the kept set.
+ */
+function adjustForToolPairs(
+  messages: Message[],
+  keepFromIndex: number,
+): number {
+  let idx = keepFromIndex;
+  while (idx > 0) {
+    const msg = messages[idx];
+    if (!msg || msg.role !== "user") break;
+
+    // Collect tool_use_ids referenced by tool_results in this user message
+    const referencedIds = new Set<string>();
+    for (const block of msg.content) {
+      if ((block.type === "tool_result" || block.type === "web_search_tool_result") && "tool_use_id" in block) {
+        referencedIds.add((block as { tool_use_id: string }).tool_use_id);
+      }
+    }
+    if (referencedIds.size === 0) break;
+
+    // Check if the preceding assistant message contains matching tool_uses
+    const prev = messages[idx - 1];
+    if (!prev || prev.role !== "assistant") break;
+
+    const hasOrphanedPair = prev.content.some(
+      (block) =>
+        (block.type === "tool_use" || block.type === "server_tool_use") &&
+        "id" in block &&
+        referencedIds.has((block as { id: string }).id),
+    );
+    if (!hasOrphanedPair) break;
+
+    // Include the assistant message
+    idx--;
+
+    // The assistant message may itself be preceded by a tool_result user
+    // message that pairs with an even earlier assistant — continue the check
+    if (idx > 0 && messages[idx - 1]?.role === "user") {
+      idx--;
+    } else {
+      break;
+    }
+  }
+  return idx;
+}
+
 export function getSummaryFromContextMessage(
   message: Message | undefined,
 ): string | null {

package/src/daemon/config-watcher.ts

@@ -12,7 +12,6 @@ import {
 } from "node:fs";
 import { join } from "node:path";
 
-import { getIsContainerized } from "../config/env-registry.js";
 import { getConfig, invalidateConfigCache } from "../config/loader.js";
 import { clearEmbeddingBackendCache } from "../memory/embedding-backend.js";
 import { clearCache as clearTrustCache } from "../permissions/trust-store.js";
@@ -210,9 +209,7 @@ export class ConfigWatcher {
       );
     }
 
-    if (!getIsContainerized()) {
-      this.startSignalsWatcher();
-    }
+    this.startSignalsWatcher();
     this.startSkillsWatchers(onConversationEvict);
   }
 

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

@@ -33,7 +33,6 @@ import {
 } from "../instrument.js";
 import { commitAppTurnChanges } from "../memory/app-git-service.js";
 import { getApp, listAppFiles, resolveAppDir } from "../memory/app-store.js";
-import { insertCompactionEpisode } from "../memory/archive-store.js";
 import {
   addMessage,
   deleteMessageById,
@@ -514,12 +513,6 @@ export async function runAgentLoopImpl(
         compacted.summaryText,
         ctx.contextCompactedMessageCount,
       );
-      dualWriteCompactionEpisode(
-        ctx.conversationId,
-        ctx.memoryPolicy.scopeId,
-        compacted.summaryText,
-        compacted.summaryOutputTokens,
-      );
       onEvent({
         type: "context_compacted",
         previousEstimatedInputTokens: compacted.previousEstimatedInputTokens,
@@ -787,12 +780,6 @@ export async function runAgentLoopImpl(
             step.compactionResult.summaryText,
             ctx.contextCompactedMessageCount,
           );
-          dualWriteCompactionEpisode(
-            ctx.conversationId,
-            ctx.memoryPolicy.scopeId,
-            step.compactionResult.summaryText,
-            step.compactionResult.summaryOutputTokens,
-          );
           onEvent({
             type: "context_compacted",
             previousEstimatedInputTokens:
@@ -977,12 +964,6 @@ export async function runAgentLoopImpl(
           midLoopCompact.summaryText,
           ctx.contextCompactedMessageCount,
         );
-        dualWriteCompactionEpisode(
-          ctx.conversationId,
-          ctx.memoryPolicy.scopeId,
-          midLoopCompact.summaryText,
-          midLoopCompact.summaryOutputTokens,
-        );
         onEvent({
           type: "context_compacted",
           previousEstimatedInputTokens:
@@ -1179,12 +1160,6 @@ export async function runAgentLoopImpl(
             step.compactionResult.summaryText,
             ctx.contextCompactedMessageCount,
           );
-          dualWriteCompactionEpisode(
-            ctx.conversationId,
-            ctx.memoryPolicy.scopeId,
-            step.compactionResult.summaryText,
-            step.compactionResult.summaryOutputTokens,
-          );
           onEvent({
             type: "context_compacted",
             previousEstimatedInputTokens:
@@ -1292,12 +1267,6 @@ export async function runAgentLoopImpl(
                 emergencyCompact.summaryText,
                 ctx.contextCompactedMessageCount,
               );
-              dualWriteCompactionEpisode(
-                ctx.conversationId,
-                ctx.memoryPolicy.scopeId,
-                emergencyCompact.summaryText,
-                emergencyCompact.summaryOutputTokens,
-              );
               onEvent({
                 type: "context_compacted",
                 previousEstimatedInputTokens:
@@ -1402,12 +1371,6 @@ export async function runAgentLoopImpl(
               emergencyCompact.summaryText,
               ctx.contextCompactedMessageCount,
             );
-            dualWriteCompactionEpisode(
-              ctx.conversationId,
-              ctx.memoryPolicy.scopeId,
-              emergencyCompact.summaryText,
-              emergencyCompact.summaryOutputTokens,
-            );
             onEvent({
               type: "context_compacted",
               previousEstimatedInputTokens:
@@ -1873,26 +1836,3 @@ function collapseRawResponses(rawResponses?: unknown[]): unknown | undefined {
   if (!rawResponses || rawResponses.length === 0) return undefined;
   return rawResponses.length === 1 ? rawResponses[0] : rawResponses;
 }
-
-/**
- * Dual-write a compaction summary as an archive episode so it becomes
- * searchable via vector recall. Called after each successful compaction
- * that produces a new summary.
- */
-function dualWriteCompactionEpisode(
-  conversationId: string,
-  scopeId: string,
-  summaryText: string,
-  summaryOutputTokens: number,
-): void {
-  const now = Date.now();
-  insertCompactionEpisode({
-    conversationId,
-    scopeId,
-    title: truncate(summaryText, 120, ""),
-    summary: summaryText,
-    tokenEstimate: summaryOutputTokens,
-    startAt: now,
-    endAt: now,
-  });
-}

package/src/daemon/conversation-memory.ts

@@ -1,8 +1,5 @@
 import { getConfig } from "../config/loader.js";
 import { estimatePromptTokens } from "../context/token-estimator.js";
-import { buildArchiveRecall } from "../memory/archive-recall.js";
-import { compileMemoryBrief } from "../memory/brief.js";
-import { getDb } from "../memory/db.js";
 import { buildMemoryQuery } from "../memory/query-builder.js";
 import { computeRecallBudget } from "../memory/retrieval-budget.js";
 import {
@@ -12,11 +9,8 @@ import {
 import type { ScopePolicyOverride } from "../memory/search/types.js";
 import type { Message } from "../providers/types.js";
 import type { Provider } from "../providers/types.js";
-import { getLogger } from "../util/logger.js";
 import type { ServerMessage } from "./message-protocol.js";
 
-const log = getLogger("conversation-memory");
-
 export interface MemoryRecallResult {
   runMessages: Message[];
   recall: Awaited<ReturnType<typeof buildMemoryRecall>>;
@@ -121,14 +115,6 @@ export async function prepareMemoryContext(
 
   const runtimeConfig = getConfig();
 
-  // ── Simplified memory path ──────────────────────────────────────────
-  // When `memory.simplified.enabled` is true, inject the brief and
-  // optional archive recall instead of the legacy hybrid pipeline.
-  if (runtimeConfig.memory?.simplified?.enabled) {
-    return prepareSimplifiedMemoryContext(ctx, content, userMessageId, onEvent);
-  }
-
-  // ── Legacy memory path (fallback) ──────────────────────────────────
   // Memory recall via the V2 hybrid pipeline
   const recallQuery = buildMemoryQuery(content, ctx.messages);
   const dynamicBudgetConfig = runtimeConfig.memory?.retrieval?.dynamicBudget;
@@ -221,106 +207,3 @@ export async function prepareMemoryContext(
     recall,
   };
 }
-
-// ── Simplified memory injection ─────────────────────────────────────────
-
-/**
- * Build simplified memory context for a turn: compiles the `<memory_brief>`
- * block and conditionally appends `<supporting_recall>` from the archive.
- *
- * Non-empty blocks are injected as text content blocks prepended to the
- * last user message, following the same injection pattern as the legacy
- * pipeline. Stripping is handled by `RUNTIME_INJECTION_PREFIXES` which
- * already includes `<memory_brief>`.
- */
-function prepareSimplifiedMemoryContext(
-  ctx: MemoryPrepareContext,
-  content: string,
-  userMessageId: string,
-  onEvent: (msg: ServerMessage) => void,
-): MemoryRecallResult {
-  const start = Date.now();
-
-  // Build a no-op recall result matching the legacy shape.
-  const noopRecall = (): Awaited<ReturnType<typeof buildMemoryRecall>> =>
-    ({
-      enabled: true,
-      degraded: false,
-      injectedText: "",
-      semanticHits: 0,
-      recencyHits: 0,
-      mergedCount: 0,
-      selectedCount: 0,
-      injectedTokens: 0,
-      latencyMs: 0,
-      topCandidates: [],
-      tier1Count: 0,
-      tier2Count: 0,
-    }) as Awaited<ReturnType<typeof buildMemoryRecall>>;
-
-  try {
-    const db = getDb();
-
-    // Step 1: Build the memory brief
-    const briefResult = compileMemoryBrief(db, ctx.scopeId, userMessageId);
-
-    // Step 2: Conditionally build supporting recall from the archive
-    const archiveResult = buildArchiveRecall(ctx.scopeId, content);
-
-    // Step 3: Assemble the injection blocks (non-empty only)
-    const blocks: string[] = [];
-    if (briefResult.text.length > 0) {
-      blocks.push(briefResult.text);
-    }
-    if (archiveResult.text.length > 0) {
-      blocks.push(archiveResult.text);
-    }
-
-    const latencyMs = Date.now() - start;
-
-    // Emit memory status for the simplified path
-    onEvent({
-      type: "memory_status",
-      enabled: true,
-      degraded: false,
-    });
-
-    // Inject non-empty blocks into the last user message
-    let runMessages = ctx.messages;
-    if (blocks.length > 0) {
-      const injectedText = blocks.join("\n\n");
-      const userTail = ctx.messages[ctx.messages.length - 1];
-      if (userTail && userTail.role === "user") {
-        runMessages = injectMemoryRecallAsUserBlock(ctx.messages, injectedText);
-      }
-
-      log.debug(
-        {
-          briefLength: briefResult.text.length,
-          recallTrigger: archiveResult.trigger,
-          recallBullets: archiveResult.bullets.length,
-          latencyMs,
-        },
-        "Simplified memory injection completed",
-      );
-    }
-
-    return {
-      runMessages,
-      recall: {
-        ...noopRecall(),
-        injectedText: blocks.length > 0 ? blocks.join("\n\n") : "",
-        latencyMs,
-      },
-    };
-  } catch (err) {
-    log.warn({ err }, "Simplified memory injection failed, returning no-op");
-    return {
-      runMessages: ctx.messages,
-      recall: {
-        ...noopRecall(),
-        latencyMs: Date.now() - start,
-      },
-    };
-  }
-}

package/src/daemon/conversation-runtime-assembly.ts

@@ -961,8 +961,6 @@ const RUNTIME_INJECTION_PREFIXES = [
   "<inbound_actor_context>",
   "<interface_turn_context>",
   "<turn_context>",
-  "<memory_brief>",
-  "<supporting_recall>",
   "<memory_context __injected>",
   "<memory_context>", // backward-compat: strip legacy blocks from pre-__injected history
   "<voice_call_control>",

package/src/daemon/handlers/conversations.ts

@@ -23,7 +23,6 @@ import {
   queueGenerateConversationTitle,
   UNTITLED_FALLBACK,
 } from "../../memory/conversation-title-service.js";
-import { reduceBeforeSwitch } from "../../memory/reducer-scheduler.js";
 import * as pendingInteractions from "../../runtime/pending-interactions.js";
 import { getSubagentManager } from "../../subagent/index.js";
 import { truncate } from "../../util/truncate.js";
@@ -234,12 +233,6 @@ export async function handleConversationCreate(
     conversationType: normalizeConversationType(conversation.conversationType),
   });
 
-  // Reduce the previous dirty conversation before processing the initial
-  // message so its memory is fresh for the next read.
-  if (msg.initialMessage) {
-    await reduceBeforeSwitch(conversation.id);
-  }
-
   // Auto-send the initial message if provided, kick-starting the skill.
   if (msg.initialMessage) {
     // Queue title generation eagerly — some processMessage paths (guardian
@@ -350,10 +343,6 @@ export async function switchConversation(
     return null;
   }
 
-  // Reduce the previous dirty conversation before switching so its memory
-  // is fresh for the next read.
-  await reduceBeforeSwitch(conversationId);
-
   // If the target conversation is headless-locked (actively executing a task run),
   // skip rebinding so tool confirmations stay suppressed.
   const existingConversation = ctx.conversations.get(conversationId);

package/src/daemon/lifecycle.ts

@@ -30,7 +30,6 @@ import {
   getConversationType,
   getMessages,
   purgePrivateConversations,
-  sweepStaleReducerJobs,
 } from "../memory/conversation-crud.js";
 import { resolveConversationId } from "../memory/conversation-key-store.js";
 import { initializeDb } from "../memory/db.js";
@@ -57,8 +56,8 @@ import { assistantEventHub } from "../runtime/assistant-event-hub.js";
 import { DAEMON_INTERNAL_ASSISTANT_ID } from "../runtime/assistant-scope.js";
 import {
   initAuthSigningKey,
-  loadOrCreateSigningKey,
   mintPairingBearerToken,
+  resolveSigningKey,
 } from "../runtime/auth/token-service.js";
 import { ensureVellumGuardianBinding } from "../runtime/guardian-vellum-migration.js";
 import { RuntimeHttpServer } from "../runtime/http-server.js";
@@ -148,7 +147,7 @@ export async function runDaemon(): Promise<void> {
     // Load (or generate + persist) the auth signing key so tokens survive
     // daemon restarts. Must happen after ensureDataDir() creates the
     // protected directory.
-    const signingKey = loadOrCreateSigningKey();
+    const signingKey = await resolveSigningKey();
     initAuthSigningKey(signingKey);
 
     seedInterfaceFiles();
@@ -213,40 +212,16 @@ export async function runDaemon(): Promise<void> {
           targetId: summaryId,
         });
       }
-      for (const obsId of deletedMemory.deletedObservationIds) {
-        enqueueMemoryJob("delete_qdrant_vectors", {
-          targetType: "observation",
-          targetId: obsId,
-        });
-      }
-      for (const chunkId of deletedMemory.deletedChunkIds) {
-        enqueueMemoryJob("delete_qdrant_vectors", {
-          targetType: "chunk",
-          targetId: chunkId,
-        });
-      }
-      for (const episodeId of deletedMemory.deletedEpisodeIds) {
-        enqueueMemoryJob("delete_qdrant_vectors", {
-          targetType: "episode",
-          targetId: episodeId,
-        });
-      }
       if (
         deletedMemory.segmentIds.length > 0 ||
         deletedMemory.orphanedItemIds.length > 0 ||
-        deletedMemory.deletedSummaryIds.length > 0 ||
-        deletedMemory.deletedObservationIds.length > 0 ||
-        deletedMemory.deletedChunkIds.length > 0 ||
-        deletedMemory.deletedEpisodeIds.length > 0
+        deletedMemory.deletedSummaryIds.length > 0
       ) {
         log.info(
           {
             segments: deletedMemory.segmentIds.length,
             orphanedItems: deletedMemory.orphanedItemIds.length,
             deletedSummaries: deletedMemory.deletedSummaryIds.length,
-            deletedObservations: deletedMemory.deletedObservationIds.length,
-            deletedChunks: deletedMemory.deletedChunkIds.length,
-            deletedEpisodes: deletedMemory.deletedEpisodeIds.length,
           },
           "Enqueued Qdrant vector cleanup jobs for purged private conversations",
         );
@@ -271,24 +246,6 @@ export async function runDaemon(): Promise<void> {
       );
     }
 
-    // Sweep dirty conversations whose tail messages are already past the
-    // idle delay — they should have been reduced while the daemon was down.
-    // Enqueue immediate reducer jobs so the memory worker picks them up.
-    try {
-      const sweepCount = sweepStaleReducerJobs();
-      if (sweepCount > 0) {
-        log.info(
-          { sweepCount },
-          `Enqueued reducer jobs for ${sweepCount} stale dirty conversation(s)`,
-        );
-      }
-    } catch (err) {
-      log.warn(
-        { err },
-        "Startup sweep for stale reducer jobs failed — continuing startup",
-      );
-    }
-
     // Ensure a vellum guardian binding exists so the identity system works
     // without requiring a manual bootstrap step.
     try {

package/src/followups/followup-store.ts

@@ -1,4 +1,4 @@
-import { and, desc, eq, lte, or } from "drizzle-orm";
+import { and, asc, eq, lte, or, sql } from "drizzle-orm";
 import { v4 as uuid } from "uuid";
 
 import { getDb } from "../memory/db.js";
@@ -217,7 +217,10 @@ export function getPendingAndOverdueFollowUps(): BriefFollowUp[] {
         eq(followups.status, "nudged"),
       ),
     )
-    .orderBy(desc(followups.expectedResponseBy))
+    .orderBy(
+      sql`CASE WHEN ${followups.expectedResponseBy} IS NULL THEN 1 ELSE 0 END`,
+      asc(followups.expectedResponseBy),
+    )
     .all();
 
   return rows as BriefFollowUp[];