@vellumai/assistant 0.10.2-dev.202606250318.5e7cfb0 → 0.10.2

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

@@ -66,6 +66,14 @@
       "description": "Gate per-turn conversation trace (user/assistant/tool-call/tool-response) collection. The daemon attaches a trace to its turn telemetry only when this flag AND the owner's share_diagnostics consent are both on. Defaults off (fail-closed) — the live value is delivered from LaunchDarkly.",
       "defaultEnabled": false
     },
+    {
+      "id": "workspace-tools-watcher",
+      "scope": "assistant",
+      "key": "workspace-tools-watcher",
+      "label": "Workspace tools file watcher",
+      "description": "Hot-reload workspace tool overrides under `<workspaceDir>/tools/` by watching the directory and re-registering tools as files change. When disabled, workspace tools still load from disk once at daemon startup, but live edits require a restart to take effect.",
+      "defaultEnabled": false
+    },
     {
       "id": "settings-developer-nav",
       "scope": "assistant",
@@ -403,11 +411,11 @@
       "defaultEnabled": false
     },
     {
-      "id": "mcp-add-server",
+      "id": "mcp-settings",
       "scope": "assistant",
-      "key": "mcp-add-server",
-      "label": "MCP Add Server",
-      "description": "Show the Add Server action on the MCP settings page. The MCP page itself remains visible; this flag gates only creating new Model Context Protocol server connections.",
+      "key": "mcp-settings",
+      "label": "MCP Settings",
+      "description": "Show the MCP page in Settings for managing Model Context Protocol server connections: view status, enable/disable, configure, and inspect registered tools per server.",
       "defaultEnabled": false
     },
     {
@@ -415,7 +423,7 @@
       "scope": "assistant",
       "key": "os-beta",
       "label": "OS Beta",
-      "description": "Enable the OS Beta model profile (MiniMax M3 / Together) in the assistant's model profile selection.",
+      "description": "Enable the OS Beta model profile (GLM 5.2 / Fireworks) in the assistant's model profile selection.",
       "defaultEnabled": false
     }
   ]

package/src/config/loader.ts

@@ -14,7 +14,6 @@ import {
   getConfigQuarantineNoticePath,
   getWorkspaceConfigPath,
 } from "../util/platform.js";
-import { pruneSeededCallsiteDefaultsFromConfig } from "./prune-seeded-callsite-defaults.js";
 import { AssistantConfigSchema } from "./schema.js";
 import type { AssistantConfig } from "./types.js";
 
@@ -301,20 +300,14 @@ function validateWithSchema(raw: Record<string, unknown>): AssistantConfig {
   }
 
   // Strip invalid fields by setting them to undefined so Zod defaults apply,
-  // then re-parse. We walk the error paths and delete the offending keys,
-  // pruning any ancestor object the deletion leaves empty. Pruning matters for
-  // nested overrides like `llm.callSites.<id>.profile`: stripping just the
-  // invalid `.profile` leaf would leave `llm.callSites.<id> = {}`, which the
-  // resolver treats as a present (non-default) override and so skips the
-  // shipped call-site default — silently downgrading the call site to the
-  // active profile. Removing the emptied object lets that default apply.
+  // then re-parse. We walk the error paths and delete the offending keys.
   const cleaned = structuredClone(raw);
   for (const issue of result.error.issues) {
     if (issue.path.length === 0) {
       // Top-level error — return full defaults
       return cloneDefaultConfig();
     }
-    deleteNestedKey(cleaned, issue.path as (string | number)[], true);
+    deleteNestedKey(cleaned, issue.path as (string | number)[]);
   }
 
   const retry = AssistantConfigSchema.safeParse(cleaned);
@@ -327,42 +320,17 @@ function validateWithSchema(raw: Record<string, unknown>): AssistantConfig {
   return cloneDefaultConfig();
 }
 
-/**
- * Delete the key at `path` from `obj`. When `pruneEmptyAncestors` is set, also
- * remove any ancestor object the deletion leaves empty, walking up until the
- * first ancestor that still holds other keys. Only empty plain objects are
- * pruned (arrays are left alone), and a still-populated ancestor stops the walk
- * so a container holding other config is never removed.
- */
 function deleteNestedKey(
   obj: Record<string, unknown>,
   path: (string | number)[],
-  pruneEmptyAncestors = false,
 ): void {
-  // Record each (container, key) hop on the way down so we can prune upward
-  // after deleting the leaf.
-  const chain: Array<{ container: Record<string, unknown>; key: string }> = [];
   let current: unknown = obj;
   for (let i = 0; i < path.length - 1; i++) {
     if (current == null || typeof current !== "object") return;
-    const key = String(path[i]);
-    chain.push({ container: current as Record<string, unknown>, key });
-    current = (current as Record<string, unknown>)[key];
+    current = (current as Record<string, unknown>)[String(path[i])];
   }
-  if (current == null || typeof current !== "object") return;
-  delete (current as Record<string, unknown>)[String(path[path.length - 1])];
-
-  if (!pruneEmptyAncestors) return;
-  // Remove ancestors emptied by the deletion, deepest first; stop at the first
-  // that still has keys.
-  for (let i = chain.length - 1; i >= 0; i--) {
-    const { container, key } = chain[i];
-    const child = container[key];
-    if (isPlainObject(child) && Object.keys(child).length === 0) {
-      delete container[key];
-    } else {
-      break;
-    }
+  if (current != null && typeof current === "object") {
+    delete (current as Record<string, unknown>)[String(path[path.length - 1])];
   }
 }
 
@@ -630,7 +598,6 @@ export function mergeDefaultWorkspaceConfig(): DefaultWorkspaceConfigMergeResult
   }
 
   deepMergeOverwrite(existing, defaults as Record<string, unknown>);
-  pruneSeededCallsiteDefaultsFromConfig(existing);
 
   const dir = dirname(configPath);
   if (!existsSync(dir)) {

package/src/config/schemas/__tests__/memory-v3.test.ts

@@ -22,7 +22,6 @@ describe("MemoryV3ConfigSchema", () => {
       needleK: 100,
       denseK: 100,
       replyQueryK: 12,
-      selectorPromptPath: null,
       edge: { hubDegree: 30, seedCount: 18, perSeed: 6, cap: 45 },
     });
   });

package/src/config/schemas/memory-lifecycle.ts

@@ -78,17 +78,6 @@ export const MemoryJobsConfigSchema = MemoryJobsConfigInputSchema.transform(
   },
 ).describe("Memory background job processing configuration");
 
-export const MemoryWorkerConfigSchema = z
-  .object({
-    enabled: z
-      .boolean({ error: "memory.worker.enabled must be a boolean" })
-      .default(false)
-      .describe(
-        "Whether the memory jobs worker runs as a separate OS process spawned at assistant startup (the `assistant memory worker` implementation) instead of on the assistant's main event loop. Only affects startup; shutdown stops whichever worker is actually running.",
-      ),
-  })
-  .describe("Memory jobs worker process configuration");
-
 export const MemoryRetentionConfigSchema = z
   .object({
     keepRawForever: z
@@ -196,7 +185,6 @@ export const MemoryMaintenanceConfigSchema = z
   );
 
 export type MemoryJobsConfig = z.infer<typeof MemoryJobsConfigSchema>;
-export type MemoryWorkerConfig = z.infer<typeof MemoryWorkerConfigSchema>;
 export type MemoryRetentionConfig = z.infer<typeof MemoryRetentionConfigSchema>;
 export type MemoryCleanupConfig = z.infer<typeof MemoryCleanupConfigSchema>;
 export type MemoryMaintenanceConfig = z.infer<

package/src/config/schemas/memory-v3.ts

@@ -259,13 +259,6 @@ export const MemoryV3ConfigSchema = z
       .describe(
         "Per-lane article budget for the reply-query finder pass: needle and dense each re-run over the assistant's previous message as separate queries (never concatenated with the user's message). 0 disables the pass. Deliberately small next to needleK/denseK — the pass adds the assistant-side retrieval signal, not a second full sweep.",
       ),
-    selectorPromptPath: z
-      .string({ error: "memory.v3.selectorPromptPath must be a string" })
-      .nullable()
-      .default(null)
-      .describe(
-        "Optional path to a file whose contents replace the bundled per-turn selector system prompt (the instructions that tell the selector which candidate pages to keep). Absolute paths are used as-is, a leading `~/` is expanded to the home directory, otherwise the path is resolved under the workspace root. The selector prompt takes no placeholders — the candidate pool is supplied separately as the user message — so the file is used verbatim. If the file is missing, unreadable, empty, or over 1 MiB, the bundled prompt is used and a warning is logged.",
-      ),
     edge: MemoryV3EdgeSchema.default(MemoryV3EdgeSchema.parse({})),
   })
   .describe("Memory v3 — section-grain lane retrieval");

package/src/config/schemas/memory.ts

@@ -5,7 +5,6 @@ import {
   MemoryJobsConfigSchema,
   MemoryMaintenanceConfigSchema,
   MemoryRetentionConfigSchema,
-  MemoryWorkerConfigSchema,
 } from "./memory-lifecycle.js";
 import {
   MemoryExtractionConfigSchema,
@@ -40,9 +39,6 @@ export const MemoryConfigSchema = z
       MemorySegmentationConfigSchema.parse({}),
     ),
     jobs: MemoryJobsConfigSchema.default(MemoryJobsConfigSchema.parse({})),
-    worker: MemoryWorkerConfigSchema.default(
-      MemoryWorkerConfigSchema.parse({}),
-    ),
     retention: MemoryRetentionConfigSchema.default(
       MemoryRetentionConfigSchema.parse({}),
     ),

package/src/config/schemas/timeouts.ts

@@ -28,14 +28,6 @@ export const TimeoutConfigSchema = z
       .describe(
         "How long to wait for user permission approval before timing out (seconds)",
       ),
-    questionResponseTimeoutSec: z
-      .number({ error: "timeouts.questionResponseTimeoutSec must be a number" })
-      .finite("timeouts.questionResponseTimeoutSec must be finite")
-      .positive("timeouts.questionResponseTimeoutSec must be a positive number")
-      .default(1800)
-      .describe(
-        "Backstop timeout for an unanswered ask_question prompt (seconds). The primary way an interactive user dismisses a prompt is by moving on — enqueuing another message supersedes it — so this only bounds a prompt left open with no response and no follow-up message.",
-      ),
     toolExecutionTimeoutSec: z
       .number({ error: "timeouts.toolExecutionTimeoutSec must be a number" })
       .finite("timeouts.toolExecutionTimeoutSec must be finite")

package/src/config/seed-inference-profiles.ts

@@ -88,24 +88,15 @@ const MANAGED_PROFILE_TEMPLATES: Record<string, ManagedProfileTemplate> = {
     // profile there's nothing stronger to consult, so the advisor defaults off.
     advisorEnabled: false,
   },
-  // Served by DeepSeek V4 Flash on Fireworks via managed platform inference: a
-  // fast, low-cost open model. `model` is pinned explicitly rather than
-  // resolved via the `latency-optimized` intent (which still maps to Kimi K2.5
-  // on Fireworks and Anthropic Haiku elsewhere).
-  //
-  // `effort: "none"` (not "low") because Fireworks is not thinking-aware: the
-  // disabled `thinking` config is stripped before the request, so a non-"none"
-  // effort would be sent as `reasoning_effort` and make this profile pay for
-  // reasoning despite thinking being off. "none" keeps Speed non-reasoning.
   "cost-optimized": {
-    model: "accounts/fireworks/models/deepseek-v4-flash",
-    provider: "fireworks",
-    connectionName: "fireworks-managed",
+    intent: "latency-optimized",
+    provider: "anthropic",
+    connectionName: "anthropic-managed",
     source: "managed",
     label: "Speed",
-    description: "Fastest responses at lower cost (DeepSeek V4 Flash)",
+    description: "Fastest responses at lower cost",
     maxTokens: 8192,
-    effort: "none",
+    effort: "low",
     thinking: { enabled: false, streamThinking: false },
     contextWindow: { maxInputTokens: DEFAULT_CONTEXT_WINDOW_MAX_INPUT_TOKENS },
   },
@@ -164,20 +155,19 @@ export const OS_BETA_FEATURE_FLAG_KEY = "os-beta";
  * Flag-gated managed profile. NOT in MANAGED_PROFILE_TEMPLATES, so the
  * unconditional boot seed never creates it. Reconciled in/out by
  * the flag-gated profile reconcile based on the `os-beta` feature flag.
- * Balanced defaults, with lower reasoning effort while the profile is in beta.
+ * Balanced-parity defaults; GLM 5.2 pinned explicitly via `model`.
  */
 export const OS_BETA_PROFILE_TEMPLATE: ManagedProfileTemplate = {
-  intent: "balanced",
-  provider: "together",
-  connectionName: "together-managed",
+  model: "accounts/fireworks/models/glm-5p2",
+  provider: "fireworks",
+  connectionName: "fireworks-managed",
   source: "managed",
   label: "OS Beta",
-  description: "Good balance of quality, cost, and speed, in beta",
+  description: "Open-source frontier model (GLM 5.2), in beta",
   maxTokens: 32000,
-  effort: "low",
+  effort: "high",
   thinking: { enabled: true, streamThinking: true },
   contextWindow: { maxInputTokens: DEFAULT_CONTEXT_WINDOW_MAX_INPUT_TOKENS },
-  topP: 0.95,
 };
 
 // Membership here marks a name as managed. The route layer applies managed

package/src/config/skills.ts

@@ -688,9 +688,6 @@ function discoverSkillDirectories(skillsDir: string): string[] {
  * parseable `package.json` whose `name` equals the directory name. This
  * mirrors the external plugin loader's recognition gate, which skips any
  * directory whose `manifest.name` does not match its directory name.
- *
- * The caller is responsible for the missing-`package.json` case (it emits a
- * diagnostic warning); this function only judges a manifest that is present.
  */
 function isRecognizedPluginDir(pluginDir: string, dirName: string): boolean {
   const manifestPath = join(pluginDir, "package.json");
@@ -738,31 +735,12 @@ function discoverPluginResidentSkills(): SkillSummary[] {
   for (const entry of entries) {
     if (!entry.isDirectory() && !entry.isSymbolicLink()) continue;
     const pluginDir = join(pluginsDir, entry.name);
-
-    // A directory under `plugins/` with no `package.json` is not a plugin the
-    // runtime can load, so its skills are never surfaced. This is an easy
-    // footgun — a plugin dropped in without its manifest looks installed but
-    // silently contributes nothing — so warn loudly with the path rather than
-    // skipping in silence, to make the misconfiguration diagnosable.
-    if (!existsSync(join(pluginDir, "package.json"))) {
-      log.warn(
-        { pluginDir },
-        "Plugin directory is missing package.json — skipping; its skills will not be available. Add a package.json whose `name` matches the directory.",
-      );
-      continue;
-    }
-
-    // Honor the `.disabled` sentinel the runtime plugin scan checks
-    // (`plugins/mtime-cache.ts`): a disabled plugin contributes no hooks or
-    // tools, so its resident skills must not be loadable either.
-    if (existsSync(join(pluginDir, ".disabled"))) continue;
-
     // Mirror the plugin loader's recognition gate: a directory is a real
-    // installed plugin only if its `package.json` `name` matches the directory.
-    // This rejects staging dirs and malformed/mismatched clones (e.g. an
-    // un-adapted `caveman-installer`) that the loader itself would skip, so the
-    // catalog never surfaces skills from a directory the runtime would refuse
-    // to load.
+    // installed plugin only if it carries a parseable `package.json` whose
+    // `name` matches the directory. This rejects staging dirs, stray files,
+    // and malformed/mismatched clones (e.g. an un-adapted `caveman-installer`)
+    // that the loader itself would skip, so the catalog never surfaces skills
+    // from a directory the runtime would refuse to load.
     if (!isRecognizedPluginDir(pluginDir, entry.name)) continue;
 
     const skillsDir = join(pluginDir, "skills");

package/src/config/sync-gated-profiles.ts

@@ -23,12 +23,12 @@ const log = getLogger("sync-gated-profiles");
  * Reconcile flag-gated managed profiles against the current feature-flag state.
  *
  * `seedInferenceProfiles()` runs synchronously at boot before feature flags are
- * available, so the OS Beta profile (MiniMax M3 / together-managed) is
- * materialized here once flags have loaded. When the `os-beta` flag is on, the
- * managed profile is created (ordered right after `balanced`); when it is off, a
- * previously managed entry is removed with `profileOrder` / `activeProfile` /
- * `advisorProfile` fallbacks. The reconcile is idempotent and never touches a
- * user-owned profile of the same name.
+ * available, so the OS Beta profile (GLM 5.2 / fireworks-managed) is materialized
+ * here once flags have loaded. When the `os-beta` flag is on, the managed profile
+ * is created (ordered right after `balanced`); when it is off, a previously
+ * managed entry is removed with `profileOrder` / `activeProfile` / `advisorProfile`
+ * fallbacks. The reconcile is idempotent and never touches a user-owned profile of
+ * the same name.
  *
  * Returns whether the on-disk config changed.
  */
@@ -105,22 +105,23 @@ function enableProfile(
     OS_BETA_PROFILE_TEMPLATE.connectionName,
   ) as Record<string, unknown>;
 
-  // BYOK installs seed managed profiles disabled: the managed inference
-  // connection backing this profile isn't usable until the user enables it, so a
-  // fresh OS Beta entry starts disabled to avoid offering an unusable route. A
-  // user's own status override (preserved below) wins on later reconciles.
+  // BYOK installs seed managed profiles disabled: the platform-auth
+  // `fireworks-managed` connection backing this profile isn't usable until the
+  // user enables it, so a fresh OS Beta entry starts disabled to avoid offering
+  // an unusable route. A user's own status override (preserved below) wins on
+  // later reconciles.
   if (isByokMode && !previous) {
     next.status = "disabled";
   }
 
   if (previous) {
-    // Preserve user-owned overrides across reconciles.
+    // The only fields a user may override on a managed profile. Carry `label`
+    // by key-presence so an explicit null (user cleared it) survives too.
     if ("label" in previous) next.label = previous.label;
     if ("status" in previous) next.status = previous.status;
     if ("advisorEnabled" in previous) {
       next.advisorEnabled = previous.advisorEnabled;
     }
-    if ("topP" in previous) next.topP = previous.topP;
   }
 
   let changed = false;

package/src/contacts/contacts-write.ts

@@ -30,9 +30,6 @@ import type {
  * Returns true when a guardian channel was found and revoked, false otherwise.
  */
 export function revokeGuardianBinding(channel: string): boolean {
-  // Local-store read, not the gateway: this read selects the row that the
-  // updateChannelStatus write below mutates, so it must stay transactionally
-  // consistent with that write. Leave for Combo 11 / gateway-bootstrap-binding.
   const guardian = findGuardianForChannel(channel);
   if (!guardian) return false;
 

package/src/daemon/assistant-attachments.ts

@@ -284,19 +284,6 @@ interface VellumLinkExtractResult {
  * markdown links from assistant text and return corresponding directive
  * requests. The text is NOT modified — the links remain as rendered markdown.
  */
-/**
- * Decode a vellum:// path segment, returning null on malformed percent-encoding
- * (e.g. a literal `%` not followed by two hex digits). This prevents a single
- * bad link from throwing URIError and aborting the entire assistant message.
- */
-function safeDecodePath(rawPath: string): string | null {
-  try {
-    return decodeURIComponent(rawPath);
-  } catch {
-    return null;
-  }
-}
-
 export function extractVellumLinks(text: string): VellumLinkExtractResult {
   const directiveRequests: DirectiveRequest[] = [];
   const parseWarnings: string[] = [];
@@ -307,19 +294,9 @@ export function extractVellumLinks(text: string): VellumLinkExtractResult {
     const authority = m[2]!;
     const rawPath = m[3]!;
 
-    const decodedPath = safeDecodePath(rawPath);
-    if (decodedPath === null) {
-      parseWarnings.push(
-        `Ignored vellum://${authority} link "${linkText}": malformed percent-encoding in path.`,
-      );
-      continue;
-    }
-
     if (authority === "workspace") {
       // Strip the leading "/" to get a workspace-relative path
-      const path = decodedPath.startsWith("/")
-        ? decodedPath.slice(1)
-        : decodedPath;
+      const path = rawPath.startsWith("/") ? rawPath.slice(1) : rawPath;
       if (!path) {
         parseWarnings.push(
           `Ignored vellum://workspace link "${linkText}": empty path.`,
@@ -333,8 +310,8 @@ export function extractVellumLinks(text: string): VellumLinkExtractResult {
         mimeType: undefined,
       });
     } else {
-      // host: decodedPath is already absolute (starts with /)
-      if (!decodedPath || decodedPath === "/") {
+      // host: rawPath is already absolute (starts with /)
+      if (!rawPath || rawPath === "/") {
         parseWarnings.push(
           `Ignored vellum://host link "${linkText}": empty path.`,
         );
@@ -342,7 +319,7 @@ export function extractVellumLinks(text: string): VellumLinkExtractResult {
       }
       directiveRequests.push({
         source: "host",
-        path: decodedPath,
+        path: rawPath,
         filename: linkText || undefined,
         mimeType: undefined,
       });

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

@@ -100,10 +100,6 @@ import {
   isUserCancellation,
 } from "./conversation-error.js";
 import { raceWithTimeout } from "./conversation-media-retry.js";
-import {
-  clearConversationNotices,
-  drainConversationNotices,
-} from "./conversation-notices.js";
 import {
   getSlackCompactionWatermarkForPrefix,
   loadSlackChronologicalContext,
@@ -484,11 +480,6 @@ export async function runAgentLoopImpl(
   // resolved once here and threaded into every re-injection — including the
   // post-compaction hook — rather than re-read per assembly call.
   const isNonInteractive = !isInteractiveResolved;
-  // Expose the resolved turn-level interactivity to tool execution so tools
-  // (e.g. ask_question) see whether a human is present to answer, rather than
-  // re-deriving it from live client state that misclassifies a scheduled turn
-  // running on a client-attached conversation.
-  ctx.currentTurnIsNonInteractive = isNonInteractive;
   const diskPressureDecision = classifyDiskPressureTurnPolicy(
     getDiskPressureStatus(),
     {
@@ -1021,15 +1012,6 @@ export async function runAgentLoopImpl(
       );
     }
 
-    const shouldEmitQueuedConversationNotices =
-      !overflowTerminalReason &&
-      !yieldedForHandoff &&
-      !state.providerErrorUserMessage &&
-      !abortController.signal.aborted;
-    if (!shouldEmitQueuedConversationNotices) {
-      clearConversationNotices(ctx.conversationId);
-    }
-
     // Flush remaining tool results. On a normal turn these drain at the next
     // `message_complete`; an aborted or yielded loop exits with them still
     // buffered, so finalize the (possibly already on-arrival-reserved) grouped
@@ -1422,16 +1404,10 @@ export async function runAgentLoopImpl(
             ? { messageId: state.lastAssistantMessageId }
             : {}),
         });
-        if (shouldEmitQueuedConversationNotices) {
-          for (const notice of drainConversationNotices(ctx.conversationId)) {
-            onEvent(notice);
-          }
-        }
         publishLoopMessagesChanged();
       }
     }
   } catch (err) {
-    clearConversationNotices(ctx.conversationId);
     const errorCtx = {
       phase: "agent_loop" as const,
       aborted: abortController.signal.aborted,
@@ -1539,10 +1515,6 @@ export async function runAgentLoopImpl(
     ctx.diskPressureCleanupModeActive = false;
     ctx.preactivatedSkillIds = undefined;
     ctx.currentTurnOverrideProfile = undefined;
-    // Turn-scoped interactivity. Clear it so paths that bypass this loop (e.g.
-    // opportunity wakes calling `agentLoop.run` directly) don't inherit a stale
-    // value and instead fall back to live client state in the tool context.
-    ctx.currentTurnIsNonInteractive = undefined;
     // Channel command intents (e.g. Telegram /start) are single-turn metadata.
     // Clear at turn end so they never leak into subsequent unrelated messages.
     ctx.commandIntent = undefined;

package/src/daemon/conversation-process.ts

@@ -60,21 +60,6 @@ import { resolveVerificationSessionIntent } from "./verification-session-intent.
 
 const log = getLogger("conversation-process");
 
-/**
- * Daemon-injected subagent lifecycle notifications carry `subagentNotification`
- * metadata. They are persisted into the parent conversation so the orchestrator
- * wakes and reads the subagent's result, but they are internal scaffolding — the
- * user sees subagent activity through the inline progress card, not a chat turn.
- * Skip the `user_message_echo` broadcast for these so they never render as a live
- * user bubble; the persisted row is filtered from the rendered transcript on the
- * client.
- */
-function isSubagentNotificationMessage(
-  metadata: Record<string, unknown> | undefined,
-): boolean {
-  return metadata?.subagentNotification != null;
-}
-
 /** Format the result of a forced compaction into a user-facing message. */
 export function formatCompactResult(result: ContextWindowResult): string {
   const fmt = (n: number | undefined) => (n ?? 0).toLocaleString("en-US");
@@ -868,16 +853,14 @@ async function drainSingleMessage(
 
   // Broadcast the user message to all hub subscribers so passive devices
   // see the user turn before the assistant reply starts streaming.
-  if (!isSubagentNotificationMessage(next.metadata)) {
-    next.onEvent({
-      type: "user_message_echo",
-      text: resolvedContent,
-      conversationId: conversation.conversationId,
-      messageId: userMessageId,
-      requestId: next.requestId,
-      clientMessageId: next.clientMessageId,
-    });
-  }
+  next.onEvent({
+    type: "user_message_echo",
+    text: resolvedContent,
+    conversationId: conversation.conversationId,
+    messageId: userMessageId,
+    requestId: next.requestId,
+    clientMessageId: next.clientMessageId,
+  });
   publishConversationMessagesChanged(conversation.conversationId);
 
   // Set the active surface for the dequeued message so runAgentLoop can inject context
@@ -1221,16 +1204,14 @@ async function drainBatch(
 
     // Broadcast the user message to all hub subscribers so passive devices
     // see each batched user turn before the assistant reply starts streaming.
-    if (!isSubagentNotificationMessage(qm.metadata)) {
-      qm.onEvent({
-        type: "user_message_echo",
-        text: qmContent,
-        conversationId: conversation.conversationId,
-        messageId: lastUserMessageId,
-        requestId: qm.requestId,
-        clientMessageId: qm.clientMessageId,
-      });
-    }
+    qm.onEvent({
+      type: "user_message_echo",
+      text: qmContent,
+      conversationId: conversation.conversationId,
+      messageId: lastUserMessageId,
+      requestId: qm.requestId,
+      clientMessageId: qm.clientMessageId,
+    });
     publishConversationMessagesChanged(conversation.conversationId);
 
     // Persist succeeded. Update last-successful markers so a later tail