[eric] health: boot-time bundled-CLI integrity check rides the login-health pill with repair guidance

backend/apps/agents/agents.py

@@ -531,15 +531,18 @@ async def subscriptions_health():
     so the frontend can retry once instead of reading 'all healthy' off a cold boot."""
     from backend.apps.nine_router import is_running, get_providers
     from backend.apps.nine_router.subscription_health import probe_subscription_health
+    from backend.apps.agents.core.bundled_cli_missing import bundled_cli_missing
+    # Independent of the router: the bundled-CLI integrity check rides the same boot fetch so an AV-quarantined runtime surfaces as a pill instead of dead turns.
+    p_cli_missing = bundled_cli_missing() is not None
     if not is_running():
-        return {"dead": [], "skipped": True}
+        return {"dead": [], "skipped": True, "cli_missing": p_cli_missing}
     try:
         connections = await get_providers()
         dead = await probe_subscription_health(connections)
-        return {"dead": dead, "skipped": False}
+        return {"dead": dead, "skipped": False, "cli_missing": p_cli_missing}
     except Exception as e:
         logger.debug(f"subscription health probe failed: {e}")
-        return {"dead": [], "skipped": True}
+        return {"dead": [], "skipped": True, "cli_missing": p_cli_missing}
 
 
 @agents.router.get("/subscriptions/models")

backend/apps/agents/core/bundled_cli_missing.py

@@ -0,0 +1,26 @@
+"""Detect the AV-quarantine failure class at boot: the claude-agent-sdk's bundled
+CLI binary deleted out from under an installed app (Windows-only in field data;
+22 of 25 affected installs never produced another agent reply). Returns the
+expected path when the binary is gone so the health pill can tell the user how
+to repair, None when it's present or there is nothing to verify."""
+import platform
+from pathlib import Path
+from typing import Optional
+from typeguard import typechecked
+
+
+@typechecked
+def bundled_cli_missing() -> Optional[str]:
+    try:
+        import claude_agent_sdk
+        p_pkg = Path(claude_agent_sdk.__file__).parent
+    except Exception:
+        return None
+    p_name = "claude.exe" if platform.system() == "Windows" else "claude"
+    p_bundled = p_pkg / "_bundled" / p_name
+    # No _bundled dir at all = a source install running the PATH CLI; nothing to verify (quarantine removes the file, not the dir).
+    if not p_bundled.parent.is_dir():
+        return None
+    if p_bundled.is_file():
+        return None
+    return str(p_bundled)

backend/tests/test_bundled_cli_missing.py

@@ -0,0 +1,40 @@
+"""Pins the boot-time AV-quarantine detector: binary present = quiet, binary
+deleted out of an intact _bundled dir = flagged with the expected path, no
+_bundled dir at all (source installs) = quiet, so it can never cry wolf on dev."""
+import platform
+import sys
+import types
+
+from backend.apps.agents.core.bundled_cli_missing import bundled_cli_missing
+
+P_CLI_NAME = "claude.exe" if platform.system() == "Windows" else "claude"
+
+
+def p_fake_sdk(monkeypatch, tmp_path, with_dir=True, with_binary=True):
+    pkg = tmp_path / "claude_agent_sdk"
+    pkg.mkdir()
+    (pkg / "__init__.py").write_text("")
+    if with_dir:
+        (pkg / "_bundled").mkdir()
+        if with_binary:
+            (pkg / "_bundled" / P_CLI_NAME).write_text("stub")
+    mod = types.ModuleType("claude_agent_sdk")
+    mod.__file__ = str(pkg / "__init__.py")
+    monkeypatch.setitem(sys.modules, "claude_agent_sdk", mod)
+
+
+def test_present_binary_is_quiet(monkeypatch, tmp_path):
+    p_fake_sdk(monkeypatch, tmp_path)
+    assert bundled_cli_missing() is None
+
+
+def test_deleted_binary_is_flagged_with_path(monkeypatch, tmp_path):
+    p_fake_sdk(monkeypatch, tmp_path, with_binary=False)
+    result = bundled_cli_missing()
+    assert result is not None
+    assert result.endswith(P_CLI_NAME)
+
+
+def test_no_bundled_dir_is_quiet(monkeypatch, tmp_path):
+    p_fake_sdk(monkeypatch, tmp_path, with_dir=False)
+    assert bundled_cli_missing() is None

frontend/src/app/components/overlays/ProviderHealthToast.tsx

@@ -16,6 +16,7 @@ export default function ProviderHealthToast() {
   const dispatch = useAppDispatch();
   const open = useAppSelector((s) => s.subscriptions.healthToastOpen);
   const dead = useAppSelector((s) => s.subscriptions.healthDead);
+  const cliMissing = useAppSelector((s) => s.subscriptions.healthCliMissing);
 
   const onReconnect = React.useCallback(() => {
     dispatch(openSettingsModal('models'));
@@ -26,7 +27,7 @@ export default function ProviderHealthToast() {
 
   return (
     <Snackbar
-      open={open && dead.length > 0}
+      open={open && (dead.length > 0 || cliMissing)}
       autoHideDuration={null}
       // Clickaway would kill the pill on the user's first canvas click, before they read it; only the X or Reconnect dismisses.
       onClose={(event, reason) => { if (reason !== 'clickaway') dispatch(hideProviderHealthToast()); }}
@@ -43,9 +44,11 @@ export default function ProviderHealthToast() {
         }}
         action={
           <>
-            <Button size="small" onClick={onReconnect} sx={{ color: c.accent.primary, fontWeight: 700 }}>
-              Reconnect
-            </Button>
+            {dead.length > 0 && (
+              <Button size="small" onClick={onReconnect} sx={{ color: c.accent.primary, fontWeight: 700 }}>
+                Reconnect
+              </Button>
+            )}
             <IconButton
               size="small"
               aria-label="Dismiss"
@@ -57,7 +60,9 @@ export default function ProviderHealthToast() {
           </>
         }
       >
-        Your {labels} login{dead.length > 1 ? 's have' : ' has'} expired; chats on {dead.length > 1 ? 'them' : 'it'} will fail until you reconnect.
+        {cliMissing
+          ? 'A core OpenSwarm component is missing, usually antivirus quarantine. Restore it from quarantine and add an exclusion for OpenSwarm, or reinstall from openswarm.com; agents cannot run until then.'
+          : `Your ${labels} login${dead.length > 1 ? 's have' : ' has'} expired; chats on ${dead.length > 1 ? 'them' : 'it'} will fail until you reconnect.`}
       </Alert>
     </Snackbar>
   );

frontend/src/shared/state/subscriptionsSlice.ts

@@ -25,6 +25,7 @@ export interface DeadProvider {
 export interface SubscriptionsState {
   status: SubscriptionStatus | null;
   healthDead: DeadProvider[];
+  healthCliMissing: boolean;
   healthToastOpen: boolean;
 }
 
@@ -34,6 +35,7 @@ type WithSubscriptions = { subscriptions: SubscriptionsState };
 const initialState: SubscriptionsState = {
   status: null,
   healthDead: [],
+  healthCliMissing: false,
   healthToastOpen: false,
 };
 
@@ -56,9 +58,9 @@ export const fetchSubscriptionStatus = createAsyncThunk(
 /** Boot-time login-health check; `skipped` means the router wasn't up yet, caller may retry once. */
 export const fetchProviderHealth = createAsyncThunk(
   'subscriptions/fetchHealth',
-  async (): Promise<{ dead: DeadProvider[]; skipped: boolean }> => {
+  async (): Promise<{ dead: DeadProvider[]; skipped: boolean; cli_missing?: boolean }> => {
     const r = await fetch(`${API_BASE}/agents/subscriptions/health`);
-    return (await r.json()) as { dead: DeadProvider[]; skipped: boolean };
+    return (await r.json()) as { dead: DeadProvider[]; skipped: boolean; cli_missing?: boolean };
   },
 );
 
@@ -99,9 +101,14 @@ const subscriptionsSlice = createSlice({
       state.status = action.payload;
     });
     builder.addCase(fetchProviderHealth.fulfilled, (state, action) => {
-      if (action.payload.skipped) return;
+      // cli_missing is filesystem truth, valid even when the router probe was skipped.
+      state.healthCliMissing = action.payload.cli_missing ?? false;
+      if (action.payload.skipped) {
+        state.healthToastOpen = state.healthToastOpen || state.healthCliMissing;
+        return;
+      }
       state.healthDead = action.payload.dead ?? [];
-      state.healthToastOpen = state.healthDead.length > 0;
+      state.healthToastOpen = state.healthDead.length > 0 || state.healthCliMissing;
     });
   },
 });