diff --git a/app/main.py b/app/main.py
index 73eade58..3c693e28 100644
--- a/app/main.py
+++ b/app/main.py
@@ -1399,9 +1399,34 @@ async def _retry_device_links() -> None:
         )
 
 
-def _unlinked_devices_warning(result: Any) -> str:
-    """The notice for deployed devices the provider says are not on the account."""
-    ids = sorted(getattr(result, "unlinked_devices", None) or [])
+async def _retrying_device_ids() -> dict[str, set[str]]:
+    """Per slug, the device ids _retry_device_links is still adding to the account.
+
+    Only inside the retry window: a row the retry job never got to mark as given
+    up (the job failing every time) must not hide the notice for ever. A failed
+    read answers {}: the plain notice then shows, as it did before.
+    """
+    try:
+        rows = await database.list_pending_device_links()
+    except Exception as exc:
+        logger.warning("Could not read the device registrations being retried: %s", exc)
+        return {}
+    window_start = _db_time(datetime.now(UTC) - DEVICE_LINK_GIVE_UP)
+    retrying: dict[str, set[str]] = {}
+    for row in rows:
+        if not row["gave_up"] and str(row.get("first_at") or "") > window_start:
+            retrying.setdefault(row["slug"], set()).add(row["device_id"])
+    return retrying
+
+
+def _unlinked_devices_warning(result: Any, retrying: set[str] | frozenset[str] = frozenset()) -> str:
+    """The notice for deployed devices the provider says are not on the account.
+
+    A device in ``retrying`` is left out: the app is still adding it, so asking
+    the user to add it by hand is wrong, and the notice would take the day's
+    push from the give-up message that follows if the retries fail.
+    """
+    ids = sorted(set(getattr(result, "unlinked_devices", None) or []) - set(retrying))
     if not ids:
         return ""
     svc = catalog.get_service(result.platform) or {}
@@ -1459,6 +1484,7 @@ async def _run_collection() -> None:
                 if c.device_ids:
                     device_checked.add(slug)
             results = await asyncio.gather(*(_collect_bounded(c) for c in collectors), return_exceptions=True)
+            retrying = await _retrying_device_ids()
             alerts: list[dict[str, str]] = []
             # Platforms that were already failing before this run: used to detect a
             # recovery, so a service that breaks again later notifies again rather
@@ -1531,7 +1557,7 @@ async def _run_collection() -> None:
                     # A caveat on a SUCCESSFUL reading. It is shown, because the
                     # figure is partial and the user should know, but it is not
                     # a failure: the balance below is stored exactly as any other.
-                    if unlinked := _unlinked_devices_warning(result):
+                    if unlinked := _unlinked_devices_warning(result, retrying.get(result.platform, set())):
                         result.warning = f"{result.warning} {unlinked}" if result.warning else unlinked
                     if getattr(result, "warning", None):
                         safe_warning = notify.redact(result.warning)
diff --git a/tests/test_retrying_device_notice.py b/tests/test_retrying_device_notice.py
new file mode 100644
index 00000000..02192a63
--- /dev/null
+++ b/tests/test_retrying_device_notice.py
@@ -0,0 +1,63 @@
+"""A device the app is still adding to the account gets no "add it by hand" notice.
+
+A new ProxyRack UUID is unknown for about five minutes, so the hourly device
+check finds it "invalid" while _retry_device_links is still adding it. The
+notice then told the user to add it by hand, and it took the day's notice push
+for that service, so the give-up message a day later was never pushed. While a
+registration is being retried the device is left out of the notice; once the
+retries give up it is reported as before.
+"""
+
+from __future__ import annotations
+
+from datetime import UTC, datetime, timedelta
+from unittest.mock import AsyncMock, patch
+
+from app.collectors.base import EarningsResult
+from tests.test_unlinked_devices import UNLINKED, _run_loop
+
+
+def _ago(hours: float) -> str:
+    return (datetime.now(UTC) - timedelta(hours=hours)).strftime("%Y-%m-%d %H:%M:%S")
+
+
+NOW = _ago(0.1)
+
+
+def _loop_with_pending(rows):
+    with patch("app.main.database.list_pending_device_links", new_callable=AsyncMock, **rows):
+        alerts, _, _ = _run_loop(EarningsResult(platform="earnapp", balance=1.0, unlinked_devices=[UNLINKED]))
+    return [a for a in alerts if a.get("kind") == "notice" and a["platform"] == "earnapp"]
+
+
+def test_a_device_being_retried_gets_no_notice():
+    pending = [{"slug": "earnapp", "device_id": UNLINKED, "gave_up": 0, "first_at": NOW}]
+    assert _loop_with_pending({"return_value": pending}) == []
+
+
+def test_a_device_whose_retries_gave_up_is_reported_again():
+    pending = [{"slug": "earnapp", "device_id": UNLINKED, "gave_up": 1, "first_at": _ago(25)}]
+    (notice,) = _loop_with_pending({"return_value": pending})
+    assert UNLINKED in notice["error"]
+
+
+def test_a_retry_of_another_service_or_device_changes_nothing():
+    pending = [
+        {"slug": "proxyrack", "device_id": UNLINKED, "gave_up": 0, "first_at": NOW},
+        {"slug": "earnapp", "device_id": "sdk-node-" + "c" * 32, "gave_up": 0, "first_at": NOW},
+    ]
+    (notice,) = _loop_with_pending({"return_value": pending})
+    assert UNLINKED in notice["error"]
+
+
+def test_an_unreadable_retry_table_keeps_the_plain_notice():
+    (notice,) = _loop_with_pending({"side_effect": RuntimeError("db locked")})
+    assert UNLINKED in notice["error"]
+
+
+def test_a_retry_left_past_its_window_no_longer_hides_the_notice():
+    # The retry job never marked it given up (it failed every time); after the
+    # day the retries were allowed, the device is reported again.
+    pending = [{"slug": "earnapp", "device_id": UNLINKED, "gave_up": 0, "first_at": _ago(25)}]
+    (notice,) = _loop_with_pending({"return_value": pending})
+    assert UNLINKED in notice["error"]