@vellumai/assistant 0.8.0 → 0.8.2

package/src/runtime/routes/inbound-stages/guardian-reply-intercept.ts

@@ -101,8 +101,22 @@ export async function handleGuardianReplyIntercept(
   // When delivery-scoped matches exist, union them with any identity-
   // based pending requests so that requests without delivery rows (e.g.
   // tool_approval requests created inline) are not silently excluded.
-  // Pass undefined (not []) when there are zero combined results so the
-  // router's own identity-based fallback stays active.
+  //
+  // On Slack, when no delivery-scoped results exist for the current
+  // chat, pass [] (empty array) instead of undefined. This prevents the
+  // router's identity-based fallback from intercepting unrelated
+  // messages in other channels/threads — a cross-chat hijacking vector
+  // unique to Slack where a single guardian is active in many threaded
+  // contexts. Explicit callbacks (apr:<id>:<action>) and request codes
+  // still work cross-chat because they carry specific request
+  // identifiers and bypass the pendingRequests list.
+  //
+  // Non-Slack channels (Telegram, WhatsApp) keep undefined so the
+  // identity-based fallback stays active. On those channels, delivery
+  // rows are created asynchronously (fire-and-forget .then()) so the
+  // guardian can reply before the row is persisted. Cross-chat
+  // contamination is unlikely there because each chat is a distinct
+  // conversation with no thread concept.
   const deliveryScopedPendingRequests =
     listPendingCanonicalGuardianRequestsByDestinationChat(
       sourceChannel,
@@ -121,6 +135,10 @@ export async function handleGuardianReplyIntercept(
       deliveryIds.add(r.id);
     }
     pendingRequestIds = [...deliveryIds];
+  } else if (sourceChannel === "slack") {
+    // Block identity-based fallback on Slack to prevent cross-chat
+    // NL/free-text interception. See comment above for rationale.
+    pendingRequestIds = [];
   }
 
   const routerResult = await routeGuardianReply({
@@ -168,13 +186,13 @@ export async function handleGuardianReplyIntercept(
     }
 
     return {
-      response: ({
+      response: {
         accepted: true,
         duplicate: false,
         eventId,
         canonicalRouter: routerResult.type,
         requestId: routerResult.requestId,
-      }),
+      },
       skipApprovalInterception: false,
     };
   }

package/src/runtime/routes/index.ts

@@ -14,14 +14,18 @@ import { ROUTES as APP_ROUTES } from "./app-routes.js";
 import { ROUTES as APPROVAL_ROUTES } from "./approval-routes.js";
 import { ROUTES as ATTACHMENT_ROUTES } from "./attachment-routes.js";
 import { ROUTES as AUDIO_ROUTES } from "./audio-routes.js";
+import { ROUTES as AUDIT_ROUTES } from "./audit-routes.js";
+import { ROUTES as AUTH_ROUTES } from "./auth-routes.js";
 import { ROUTES as AVATAR_ROUTES } from "./avatar-routes.js";
 import { ROUTES as BACKGROUND_TOOL_ROUTES } from "./background-tool-routes.js";
 import { ROUTES as BACKUP_ROUTES } from "./backup-routes.js";
+import { ROUTES as BOOKMARK_ROUTES } from "./bookmark-routes.js";
 import { ROUTES as BRAIN_GRAPH_ROUTES } from "./brain-graph-routes.js";
 import { ROUTES as BROWSER_ROUTES } from "./browser-routes.js";
 import { ROUTES as BTW_ROUTES } from "./btw-routes.js";
 import { ROUTES as CACHE_ROUTES } from "./cache-routes.js";
 import { ROUTES as CALL_ROUTES } from "./call-routes.js";
+import { ROUTES as CHANNEL_AVAILABILITY_ROUTES } from "./channel-availability-routes.js";
 import { ROUTES as CHANNEL_READINESS_ROUTES } from "./channel-readiness-routes.js";
 import { CHANNEL_ROUTES } from "./channel-route-definitions.js";
 import { ROUTES as CHANNEL_VERIFICATION_ROUTES } from "./channel-verification-routes.js";
@@ -31,18 +35,23 @@ import { CONTACT_PROMPT_ROUTES } from "./contact-prompt-routes.js";
 import { ROUTES as CONTACT_ROUTES } from "./contact-routes.js";
 import { ROUTES as CONVERSATION_ANALYSIS_ROUTES } from "./conversation-analysis-routes.js";
 import { ROUTES as CONVERSATION_ATTENTION_ROUTES } from "./conversation-attention-routes.js";
+import { ROUTES as CONVERSATION_CLI_ROUTES } from "./conversation-cli-routes.js";
 import { ROUTES as CONVERSATION_LIST_ROUTES } from "./conversation-list-routes.js";
 import { ROUTES as CONVERSATION_MANAGEMENT_ROUTES } from "./conversation-management-routes.js";
 import { ROUTES as CONVERSATION_QUERY_ROUTES } from "./conversation-query-routes.js";
 import { ROUTES as CONVERSATION_MESSAGE_ROUTES } from "./conversation-routes.js";
 import { ROUTES as CONVERSATION_STARTER_ROUTES } from "./conversation-starter-routes.js";
+import { ROUTES as CONVERSATIONS_IMPORT_ROUTES } from "./conversations-import-routes.js";
 import { ROUTES as CREDENTIAL_PROMPT_ROUTES } from "./credential-prompt-routes.js";
+import { ROUTES as CREDENTIAL_ROUTES } from "./credential-routes.js";
 import { ROUTES as DEBUG_BASH_ROUTES } from "./debug-bash-routes.js";
 import { ROUTES as DEBUG_ROUTES } from "./debug-routes.js";
 import { ROUTES as DEFER_ROUTES } from "./defer-routes.js";
 import { ROUTES as DIAGNOSTICS_ROUTES } from "./diagnostics-routes.js";
 import { ROUTES as DISK_PRESSURE_ROUTES } from "./disk-pressure-routes.js";
 import { ROUTES as DOCUMENT_ROUTES } from "./documents-routes.js";
+import { ROUTES as DOMAIN_ROUTES } from "./domain-routes.js";
+import { ROUTES as EMAIL_ROUTES } from "./email-routes.js";
 import { ROUTES as EVENTS_ROUTES } from "./events-routes.js";
 import { ROUTES as FILING_ROUTES } from "./filing-routes.js";
 import { ROUTES as GATEWAY_LOG_ROUTES } from "./gateway-log-routes.js";
@@ -59,6 +68,10 @@ import { ROUTES as HOST_CU_ROUTES } from "./host-cu-routes.js";
 import { ROUTES as HOST_FILE_ROUTES } from "./host-file-routes.js";
 import { ROUTES as HOST_TRANSFER_ROUTES } from "./host-transfer-routes.js";
 import { ROUTES as IDENTITY_ROUTES } from "./identity-routes.js";
+import { ROUTES as IMAGE_GENERATION_ROUTES } from "./image-generation-routes.js";
+import { ROUTES as INFERENCE_PROFILE_SESSION_ROUTES } from "./inference-profile-session-routes.js";
+import { ROUTES as INFERENCE_PROVIDER_CONNECTION_ROUTES } from "./inference-provider-connection-routes.js";
+import { ROUTES as INFERENCE_SEND_ROUTES } from "./inference-send-routes.js";
 import { ROUTES as SLACK_CHANNEL_ROUTES } from "./integrations/slack/channel.js";
 import { ROUTES as SLACK_SHARE_ROUTES } from "./integrations/slack/share.js";
 import { ROUTES as TELEGRAM_ROUTES } from "./integrations/telegram.js";
@@ -76,15 +89,21 @@ import { ROUTES as MIGRATION_ROLLBACK_ROUTES } from "./migration-rollback-routes
 import { ROUTES as MIGRATION_ROUTES } from "./migration-routes.js";
 import { ROUTES as NOTIFICATION_ROUTES } from "./notification-routes.js";
 import { ROUTES as OAUTH_APPS_ROUTES } from "./oauth-apps.js";
+import { ROUTES as OAUTH_COMMANDS_ROUTES } from "./oauth-commands-routes.js";
 import { ROUTES as OAUTH_CONNECT_ROUTES } from "./oauth-connect-routes.js";
+import { ROUTES as OAUTH_LIFECYCLE_ROUTES } from "./oauth-lifecycle-routes.js";
 import { ROUTES as OAUTH_PROVIDERS_ROUTES } from "./oauth-providers.js";
+import { ROUTES as PLATFORM_ROUTES } from "./platform-routes.js";
 import { ROUTES as PLAYGROUND_ROUTES } from "./playground/index.js";
 import { ROUTES as PROFILER_ROUTES } from "./profiler-routes.js";
 import { ROUTES as PS_ROUTES } from "./ps-routes.js";
+import { ROUTES as PUBLISH_ROUTES } from "./publish-routes.js";
+import { ROUTES as QUESTION_ROUTES } from "./question-routes.js";
 import { ROUTES as RECORDING_ROUTES } from "./recording-routes.js";
 import { ROUTES as RENAME_CONVERSATION_ROUTES } from "./rename-conversation-routes.js";
 import { ROUTES as SCHEDULE_ROUTES } from "./schedule-routes.js";
 import { ROUTES as SECRET_ROUTES } from "./secret-routes.js";
+import { ROUTES as SEQUENCE_ROUTES } from "./sequence-routes.js";
 import { ROUTES as SETTINGS_ROUTES } from "./settings-routes.js";
 import { ROUTES as SKILL_ROUTES } from "./skills-routes.js";
 import { ROUTES as STT_ROUTES } from "./stt-routes.js";
@@ -102,8 +121,10 @@ import { ROUTES as UI_REQUEST_ROUTES } from "./ui-request-routes.js";
 import { ROUTES as UPGRADE_BROADCAST_ROUTES } from "./upgrade-broadcast-routes.js";
 import { ROUTES as USAGE_ROUTES } from "./usage-routes.js";
 import { ROUTES as USER_ROUTES } from "./user-routes.js";
+import { ROUTES as USER_ROUTES_CLI } from "./user-routes-cli.js";
 import { ROUTES as WAKE_CONVERSATION_ROUTES } from "./wake-conversation-routes.js";
 import { ROUTES as WATCHER_ROUTES } from "./watcher-routes.js";
+import { ROUTES as WEBHOOK_ROUTES } from "./webhook-routes.js";
 import { ROUTES as WIPE_CONVERSATION_ROUTES } from "./wipe-conversation-routes.js";
 import { ROUTES as WORK_ITEM_ROUTES } from "./work-items-routes.js";
 import { ROUTES as WORKSPACE_COMMIT_ROUTES } from "./workspace-commit-routes.js";
@@ -116,13 +137,17 @@ export const ROUTES: RouteDefinition[] = [
   ...APP_ROUTES,
   ...APPROVAL_ROUTES,
   ...AUDIO_ROUTES,
+  ...AUDIT_ROUTES,
+  ...AUTH_ROUTES,
   ...AVATAR_ROUTES,
   ...BACKGROUND_TOOL_ROUTES,
   ...BACKUP_ROUTES,
+  ...BOOKMARK_ROUTES,
   ...CACHE_ROUTES,
   ...CALL_ROUTES,
   ...CHANNEL_ROUTES,
   ...CHANNEL_VERIFICATION_ROUTES,
+  ...CHANNEL_AVAILABILITY_ROUTES,
   ...CHANNEL_READINESS_ROUTES,
   ...BROWSER_ROUTES,
   ...BTW_ROUTES,
@@ -132,11 +157,14 @@ export const ROUTES: RouteDefinition[] = [
   ...CONTACT_ROUTES,
   ...CONVERSATION_ANALYSIS_ROUTES,
   ...CONVERSATION_ATTENTION_ROUTES,
+  ...CONVERSATION_CLI_ROUTES,
   ...CONVERSATION_LIST_ROUTES,
   ...CONVERSATION_MANAGEMENT_ROUTES,
+  ...CONVERSATIONS_IMPORT_ROUTES,
   ...CONVERSATION_MESSAGE_ROUTES,
   ...CONSOLIDATION_ROUTES,
   ...CREDENTIAL_PROMPT_ROUTES,
+  ...CREDENTIAL_ROUTES,
   ...DEFER_ROUTES,
   ...CONVERSATION_QUERY_ROUTES,
   ...CONVERSATION_STARTER_ROUTES,
@@ -144,7 +172,9 @@ export const ROUTES: RouteDefinition[] = [
   ...DEBUG_ROUTES,
   ...DIAGNOSTICS_ROUTES,
   ...DISK_PRESSURE_ROUTES,
+  ...DOMAIN_ROUTES,
   ...DOCUMENT_ROUTES,
+  ...EMAIL_ROUTES,
   ...EVENTS_ROUTES,
   ...FILING_ROUTES,
   ...GATEWAY_LOG_ROUTES,
@@ -154,6 +184,7 @@ export const ROUTES: RouteDefinition[] = [
   ...HEARTBEAT_ROUTES,
   ...HOME_FEED_ROUTES,
   ...HOME_STATE_ROUTES,
+  ...IMAGE_GENERATION_ROUTES,
   ...HOST_APP_CONTROL_ROUTES,
   ...HOST_BASH_ROUTES,
   ...HOST_BROWSER_ROUTES,
@@ -161,6 +192,9 @@ export const ROUTES: RouteDefinition[] = [
   ...HOST_FILE_ROUTES,
   ...HOST_TRANSFER_ROUTES,
   ...IDENTITY_ROUTES,
+  ...INFERENCE_PROFILE_SESSION_ROUTES,
+  ...INFERENCE_PROVIDER_CONNECTION_ROUTES,
+  ...INFERENCE_SEND_ROUTES,
   ...INTERFACE_ROUTES,
   ...INTERNAL_OAUTH_ROUTES,
   ...MCP_AUTH_ROUTES,
@@ -174,10 +208,15 @@ export const ROUTES: RouteDefinition[] = [
   ...MIGRATION_ROUTES,
   ...NOTIFICATION_ROUTES,
   ...OAUTH_APPS_ROUTES,
+  ...OAUTH_LIFECYCLE_ROUTES,
+  ...OAUTH_COMMANDS_ROUTES,
   ...OAUTH_PROVIDERS_ROUTES,
+  ...PLATFORM_ROUTES,
   ...PLAYGROUND_ROUTES,
   ...PROFILER_ROUTES,
   ...PS_ROUTES,
+  ...PUBLISH_ROUTES,
+  ...QUESTION_ROUTES,
   ...RECORDING_ROUTES,
   ...RENAME_CONVERSATION_ROUTES,
   ...SCHEDULE_ROUTES,
@@ -204,10 +243,13 @@ export const ROUTES: RouteDefinition[] = [
   ...VERCEL_ROUTES,
   ...WORK_ITEM_ROUTES,
   ...WATCHER_ROUTES,
+  ...WEBHOOK_ROUTES,
   ...WIPE_CONVERSATION_ROUTES,
   ...WORKSPACE_COMMIT_ROUTES,
   ...WAKE_CONVERSATION_ROUTES,
   ...WORKSPACE_ROUTES,
+  ...SEQUENCE_ROUTES,
+  ...USER_ROUTES_CLI,
 
   // User-defined routes under /x/* — MUST be last so built-in routes
   // always take priority over the catch-all pattern.

package/src/runtime/routes/inference-profile-session-handler.ts

@@ -0,0 +1,285 @@
+/**
+ * Shared business-logic handler for inference-profile session open/close/list.
+ *
+ * This module is the single source of truth for:
+ *  - setInferenceProfileSession  — open (or replace) a session-backed profile
+ *  - closeInferenceProfileSession — close the active session
+ *  - listInferenceProfileSessionsWithRemaining — list active sessions with
+ *    remaining TTL
+ *
+ * Route handlers in `inference-profile-session-routes.ts` (PR 4) and the IPC
+ * route equivalents delegate to these functions rather than duplicating the
+ * logic.
+ */
+
+import { randomUUID } from "node:crypto";
+
+import { loadConfig } from "../../config/loader.js";
+import {
+  getConversation,
+  listActiveInferenceProfileSessions,
+  setConversationInferenceProfileSession,
+} from "../../memory/conversation-crud.js";
+import { resolveConversationId } from "../../memory/conversation-key-store.js";
+import { publishConversationInferenceProfileChanged } from "../sync/resource-sync-events.js";
+import { BadRequestError, NotFoundError } from "./errors.js";
+
+// ---------------------------------------------------------------------------
+// Types
+// ---------------------------------------------------------------------------
+
+export interface InferenceProfileSessionResult {
+  conversationId: string;
+  profile: string | null;
+  sessionId: string | null;
+  expiresAt: number | null;
+  ttlSeconds: number | null | undefined;
+  replaced: {
+    profile: string | null;
+    sessionId: string | null;
+    expiresAt: number | null;
+  } | null;
+}
+
+// ---------------------------------------------------------------------------
+// setInferenceProfileSession
+// ---------------------------------------------------------------------------
+
+/**
+ * Open (or replace) a session-backed inference-profile override for a
+ * conversation.
+ *
+ * - `profile === null`: clears any existing override (session or non-session).
+ * - `profile` is a string: validates it against `llm.profiles`, then applies
+ *   it.  TTL handling:
+ *   - `ttlSeconds` is a positive number → clamp to
+ *     `[1, llm.profileSession.maxTtlSeconds]`, mint a sessionId, compute
+ *     expiresAt.
+ *   - `ttlSeconds === null` → apply profile without a session (sticky, no
+ *     expiry).  sessionId and expiresAt are both null.
+ *   - `ttlSeconds === undefined` → same as null — sticky, no expiry.
+ *
+ * Returns `replaced` with the prior active session's details when one existed,
+ * or `null` when no active session was replaced.
+ */
+export async function setInferenceProfileSession({
+  conversationId,
+  profile,
+  ttlSeconds,
+  sessionId: callerSessionId,
+}: {
+  conversationId: string;
+  profile: string | null;
+  ttlSeconds?: number | null;
+  sessionId?: string;
+}): Promise<InferenceProfileSessionResult> {
+  const resolvedId = resolveConversationId(conversationId) ?? conversationId;
+  const conversation = getConversation(resolvedId);
+  if (!conversation) {
+    throw new NotFoundError(`Conversation ${conversationId} not found`);
+  }
+
+  // Capture the prior active session (if any) so we can report it as
+  // `replaced` in the response.
+  let replaced: InferenceProfileSessionResult["replaced"] = null;
+  if (
+    conversation.inferenceProfileSessionId != null &&
+    (conversation.inferenceProfileExpiresAt == null ||
+      conversation.inferenceProfileExpiresAt > Date.now())
+  ) {
+    replaced = {
+      profile: conversation.inferenceProfile,
+      sessionId: conversation.inferenceProfileSessionId,
+      expiresAt: conversation.inferenceProfileExpiresAt,
+    };
+  }
+
+  // --- Clear path ---
+  if (profile === null) {
+    // Idempotency: skip the write and event when the row is already fully
+    // cleared (no profile, no session, no expiry). This keeps `updatedAt`
+    // stable and avoids emitting duplicate `conversation_inference_profile_updated`
+    // events for retried/repeated clears, matching the behavior of
+    // `closeInferenceProfileSession` when there is no active session.
+    if (
+      conversation.inferenceProfile == null &&
+      conversation.inferenceProfileSessionId == null &&
+      conversation.inferenceProfileExpiresAt == null
+    ) {
+      return {
+        conversationId: resolvedId,
+        profile: null,
+        sessionId: null,
+        expiresAt: null,
+        ttlSeconds: null,
+        replaced: null,
+      };
+    }
+    setConversationInferenceProfileSession(resolvedId, null, null, null);
+    publishConversationInferenceProfileChanged({
+      conversationId: resolvedId,
+      profile: null,
+      sessionId: null,
+      expiresAt: null,
+    });
+    return {
+      conversationId: resolvedId,
+      profile: null,
+      sessionId: null,
+      expiresAt: null,
+      ttlSeconds: null,
+      replaced,
+    };
+  }
+
+  // --- Validate profile ---
+  const profiles = loadConfig().llm?.profiles ?? {};
+  if (!Object.prototype.hasOwnProperty.call(profiles, profile)) {
+    throw new BadRequestError(
+      `Profile "${profile}" is not defined in llm.profiles`,
+    );
+  }
+
+  // --- Compute session fields based on ttlSeconds ---
+  let newSessionId: string | null;
+  let newExpiresAt: number | null;
+  let clamped: number | null | undefined;
+
+  if (typeof ttlSeconds === "number") {
+    const maxTtl = loadConfig().llm?.profileSession?.maxTtlSeconds ?? 43200;
+    clamped = Math.min(Math.max(1, ttlSeconds), maxTtl);
+    newExpiresAt = Date.now() + clamped * 1000;
+    newSessionId = callerSessionId ?? randomUUID();
+  } else if (ttlSeconds === null) {
+    newExpiresAt = null;
+    newSessionId = null;
+    clamped = null;
+  } else {
+    // undefined
+    newExpiresAt = null;
+    newSessionId = null;
+    clamped = undefined;
+  }
+
+  // Skip write and event when nothing has changed.
+  if (
+    conversation.inferenceProfile === profile &&
+    conversation.inferenceProfileSessionId === (newSessionId ?? null) &&
+    conversation.inferenceProfileExpiresAt === (newExpiresAt ?? null)
+  ) {
+    return {
+      conversationId: resolvedId,
+      profile,
+      sessionId: newSessionId ?? null,
+      expiresAt: newExpiresAt ?? null,
+      ttlSeconds: clamped,
+      replaced,
+    };
+  }
+
+  setConversationInferenceProfileSession(
+    resolvedId,
+    profile,
+    newSessionId ?? null,
+    newExpiresAt ?? null,
+  );
+
+  publishConversationInferenceProfileChanged({
+    conversationId: resolvedId,
+    profile,
+    sessionId: newSessionId ?? null,
+    expiresAt: newExpiresAt ?? null,
+  });
+
+  return {
+    conversationId: resolvedId,
+    profile,
+    sessionId: newSessionId ?? null,
+    expiresAt: newExpiresAt ?? null,
+    ttlSeconds: clamped,
+    replaced,
+  };
+}
+
+// ---------------------------------------------------------------------------
+// closeInferenceProfileSession
+// ---------------------------------------------------------------------------
+
+/**
+ * Close the active session-backed inference-profile override for a
+ * conversation.
+ *
+ * Only closes session-backed overrides (rows with a non-null
+ * `inferenceProfileSessionId`). Sticky overrides set by the composer picker
+ * (profile present, no sessionId) are left untouched — close is not the right
+ * verb for those and callers should use the PUT endpoint to clear them.
+ *
+ * Returns `noop: true` when there was no active session to close.
+ */
+export async function closeInferenceProfileSession(
+  conversationId: string,
+): Promise<{
+  conversationId: string;
+  closed: { profile: string | null; sessionId: string | null } | null;
+  noop: boolean;
+}> {
+  const resolvedId = resolveConversationId(conversationId) ?? conversationId;
+  const conversation = getConversation(resolvedId);
+  if (!conversation) {
+    throw new NotFoundError(`Conversation ${conversationId} not found`);
+  }
+
+  const hasActiveSession =
+    conversation.inferenceProfileSessionId != null &&
+    (conversation.inferenceProfileExpiresAt == null ||
+      conversation.inferenceProfileExpiresAt > Date.now());
+
+  if (!hasActiveSession) {
+    return { conversationId: resolvedId, closed: null, noop: true };
+  }
+
+  const result = await setInferenceProfileSession({
+    conversationId: resolvedId,
+    profile: null,
+  });
+  return {
+    conversationId: result.conversationId,
+    closed:
+      result.replaced !== null
+        ? {
+            profile: result.replaced.profile,
+            sessionId: result.replaced.sessionId,
+          }
+        : null,
+    noop: result.replaced === null,
+  };
+}
+
+// ---------------------------------------------------------------------------
+// listInferenceProfileSessionsWithRemaining
+// ---------------------------------------------------------------------------
+
+/**
+ * List all active (non-expired) session-backed inference-profile overrides,
+ * augmented with `remainingSeconds` relative to `Date.now()`.
+ *
+ * Pass `conversationId` to narrow to a single conversation.
+ */
+export function listInferenceProfileSessionsWithRemaining(
+  conversationId?: string,
+): Array<{
+  conversationId: string;
+  conversationTitle: string | null;
+  profile: string;
+  sessionId: string;
+  expiresAt: number;
+  remainingSeconds: number;
+}> {
+  return listActiveInferenceProfileSessions(conversationId).map((row) => ({
+    ...row,
+    remainingSeconds: Math.max(
+      0,
+      Math.floor((row.expiresAt - Date.now()) / 1000),
+    ),
+  }));
+}

package/src/runtime/routes/inference-profile-session-reaper.ts

@@ -0,0 +1,84 @@
+/**
+ * Inference profile session reaper.
+ *
+ * Periodically scans the `conversations` table for rows whose
+ * session-backed inference profile has expired (`inference_profile_expires_at`
+ * is in the past) and clears them.  For each cleared session a
+ * `conversation_inference_profile_updated` event is published so connected
+ * clients receive an immediate update without polling.
+ *
+ * The reaper runs every 30 seconds and is started alongside the other
+ * background sweeps in `http-server.ts`.
+ */
+
+import { clearExpiredInferenceProfiles } from "../../memory/conversation-crud.js";
+import { getLogger } from "../../util/logger.js";
+import { publishConversationInferenceProfileChanged } from "../sync/resource-sync-events.js";
+
+const log = getLogger("inference-profile-session-reaper");
+
+/** Interval at which the reaper runs (30 seconds). */
+const REAPER_INTERVAL_MS = 30_000;
+
+/** Timer handle for the reaper so it can be stopped in tests and shutdown. */
+let reaperTimer: ReturnType<typeof setInterval> | null = null;
+
+/** Guard against overlapping reaper runs. */
+let reaperInProgress = false;
+
+/**
+ * Tick the inference profile session reaper once.
+ *
+ * Clears all conversations whose session-backed inference profile has expired
+ * and emits a `conversation_inference_profile_updated` event for each one.
+ * Exported for direct use in tests.
+ */
+export function tickInferenceProfileReaper(): void {
+  const now = Date.now();
+  const cleared = clearExpiredInferenceProfiles(now);
+  for (const { conversationId } of cleared) {
+    publishConversationInferenceProfileChanged({
+      conversationId,
+      profile: null,
+      sessionId: null,
+      expiresAt: null,
+    });
+  }
+  if (cleared.length > 0) {
+    log.info(
+      { count: cleared.length },
+      "Inference profile session reaper: cleared expired sessions",
+    );
+  }
+}
+
+/**
+ * Start the periodic inference profile session reaper. Idempotent — calling
+ * it multiple times reuses the same timer.
+ */
+export function startInferenceProfileSessionReaper(): void {
+  if (reaperTimer) return;
+  reaperTimer = setInterval(() => {
+    if (reaperInProgress) return;
+    reaperInProgress = true;
+    try {
+      tickInferenceProfileReaper();
+    } catch (err) {
+      log.error({ err }, "Inference profile session reaper failed");
+    } finally {
+      reaperInProgress = false;
+    }
+  }, REAPER_INTERVAL_MS);
+}
+
+/**
+ * Stop the periodic inference profile session reaper. Used in tests and
+ * shutdown.
+ */
+export function stopInferenceProfileSessionReaper(): void {
+  if (reaperTimer) {
+    clearInterval(reaperTimer);
+    reaperTimer = null;
+  }
+  reaperInProgress = false;
+}