Add retry on flood for repeater login

AGENTS.md

@@ -341,7 +341,7 @@ All endpoints are prefixed with `/api` (e.g., `/api/health`).
 | POST | `/api/contacts/{public_key}/routing-override` | Set or clear a forced routing override |
 | POST | `/api/contacts/{public_key}/trace` | Trace route to contact |
 | POST | `/api/contacts/{public_key}/path-discovery` | Discover forward/return paths and persist the learned direct route |
-| POST | `/api/contacts/{public_key}/repeater/login` | Log in to a repeater |
+| POST | `/api/contacts/{public_key}/repeater/login` | Log in to a repeater (escalates to one flood retry if the first attempt draws no reply) |
 | POST | `/api/contacts/{public_key}/repeater/status` | Fetch repeater status telemetry |
 | POST | `/api/contacts/{public_key}/repeater/lpp-telemetry` | Fetch CayenneLPP sensor data |
 | POST | `/api/contacts/{public_key}/repeater/neighbors` | Fetch repeater neighbors |
@@ -354,7 +354,7 @@ All endpoints are prefixed with `/api` (e.g., `/api/health`).
 | GET | `/api/contacts/{public_key}/repeater/telemetry-history` | Stored telemetry history for a repeater (read-only, no radio access) |
 | POST | `/api/contacts/{public_key}/telemetry` | Fetch CayenneLPP telemetry from any contact (single attempt, 10s timeout) |
 | GET | `/api/contacts/{public_key}/telemetry-history` | Stored LPP telemetry history for a contact (read-only, no radio access) |
-| POST | `/api/contacts/{public_key}/room/login` | Log in to a room server |
+| POST | `/api/contacts/{public_key}/room/login` | Log in to a room server (escalates to one flood retry if the first attempt draws no reply) |
 | POST | `/api/contacts/{public_key}/room/status` | Fetch room-server status telemetry |
 | POST | `/api/contacts/{public_key}/room/lpp-telemetry` | Fetch room-server CayenneLPP sensor data |
 | POST | `/api/contacts/{public_key}/room/acl` | Fetch room-server ACL entries |

app/AGENTS.md

@@ -142,6 +142,23 @@ app/
 - ACKs are delivery state, not routing state. Bundled ACKs inside PATH packets still satisfy pending DM sends, but ACK history does not feed contact route learning.
 - DM ACKs are matched from two independent radio emissions, so confirmation does not depend on the radio surfacing a host control frame: (1) the `EventType.ACK`/`SEND_CONFIRMED` host frame via `event_handlers.on_ack`, and (2) the raw RF packet itself via `packet_processor.process_raw_packet`. The packet processor extracts ACK codes both from PATH-return packets (flood replies, ACK embedded in `extra`) and from standalone `PayloadType.ACK` packets (direct replies, 4-byte cleartext payload), feeding both into `apply_dm_ack_code`. This matters for companion firmwares (e.g. pyMC over TCP) that do not reliably emit a separate host ACK frame for direct-routed replies.
 
+### Server login route escalation
+
+`prepare_authenticated_contact_connection` (`routers/server_control.py`, shared by repeater and room login) sends one login over the contact's effective route. If that draws **no reply at all**, it calls `reset_path(...)` and retries exactly once as flood.
+
+This is intentionally *more* than the reference implementations do — do not "correct" it back to single-shot:
+- Firmware `BaseChatMesh::sendLogin` picks flood only when `out_path_len == OUT_PATH_UNKNOWN` and never retries; `CMD_SEND_LOGIN` calls it once.
+- `meshcore_py` has `send_msg_with_retry` (with `flood_after` + `reset_path`) but no login equivalent — `send_login`/`send_login_sync` are single-shot.
+- Firmware only clears a stale path when the *host* asks (`CMD_RESET_PATH`); client-side path learning is otherwise passive via `onContactPathRecv`.
+
+Escalating is still correct because the **server** side treats an inbound flood as its cue to relearn the return path (`simple_repeater`/`simple_room_server`: `if (is_flood) client->out_path_len = OUT_PATH_UNKNOWN`). A flood login is therefore what repairs a broken route in both directions, and login gates the whole repeater dashboard.
+
+Escalation is bounded to one extra attempt and only fires when:
+- the first attempt **timed out**. `LOGIN_FAILED` means the server heard us and refused, so the route is fine and retrying only hammers it with bad credentials; a send error is a local radio problem a different route will not fix.
+- the contact was **not already on flood** (`effective_route_source != "flood"`), since the retry would otherwise be byte-identical.
+
+The retry deliberately does not re-run `_ensure_on_radio` — re-adding the contact would restore the route just cleared. `reset_path` clears the route on the radio only; the stored contact route is untouched, so the next `add_contact` re-stages it. That mirrors the DM retry and keeps one bad login from discarding a route that may be fine.
+
 ### Echo/repeat dedup
 
 - Channel message uniqueness (`idx_messages_dedup_null_safe`): `(type, conversation_key, text, COALESCE(sender_timestamp, 0))` where `type = 'CHAN'`.
@@ -242,7 +259,7 @@ Web Push is a standalone subsystem in `app/push/`, separate from the fanout modu
 - `POST /contacts/{public_key}/routing-override`
 - `POST /contacts/{public_key}/trace`
 - `POST /contacts/{public_key}/path-discovery` — discover forward/return paths, persist the learned direct route, and sync it back to the radio best-effort
-- `POST /contacts/{public_key}/repeater/login`
+- `POST /contacts/{public_key}/repeater/login` — one attempt on the effective route, then one flood retry on timeout
 - `POST /contacts/{public_key}/repeater/status`
 - `POST /contacts/{public_key}/repeater/lpp-telemetry`
 - `POST /contacts/{public_key}/repeater/neighbors`
@@ -255,7 +272,7 @@ Web Push is a standalone subsystem in `app/push/`, separate from the fanout modu
 - `GET /contacts/{public_key}/repeater/telemetry-history` — stored telemetry history for a repeater (read-only, no radio access)
 - `POST /contacts/{public_key}/telemetry` — on-demand CayenneLPP telemetry from any contact (persists in `contact_telemetry_history`)
 - `GET /contacts/{public_key}/telemetry-history` — stored LPP telemetry history for a contact (read-only)
-- `POST /contacts/{public_key}/room/login`
+- `POST /contacts/{public_key}/room/login` — one attempt on the effective route, then one flood retry on timeout
 - `POST /contacts/{public_key}/room/status`
 - `POST /contacts/{public_key}/room/lpp-telemetry`
 - `POST /contacts/{public_key}/room/acl`

app/routers/server_control.py

@@ -80,6 +80,15 @@ def _login_timeout_message(label: str) -> str:
     )
 
 
+def _login_flood_retry_timeout_message(label: str) -> str:
+    return (
+        f"No login confirmation was heard from the {label}, including a retry sent as flood "
+        "in case the stored route was stale. That can mean the password was wrong, the "
+        f"{label} is out of range, or the reply was missed in transit. "
+        "You're free to attempt interaction; try logging in again if authenticated actions fail."
+    )
+
+
 def extract_response_text(event) -> str:
     """Extract text from a CLI response event, stripping the firmware '> ' prefix."""
     text = event.payload.get("text", str(event.payload))
@@ -214,17 +223,20 @@ async def fetch_contact_cli_response(
         subscription.unsubscribe()
 
 
-async def prepare_authenticated_contact_connection(
+async def _attempt_server_login(
     mc,
     contact: Contact,
     password: str,
     *,
-    label: str | None = None,
+    contact_label: str,
-    response_timeout: float = SERVER_LOGIN_RESPONSE_TIMEOUT_SECONDS,
+    response_timeout: float,
 ) -> RepeaterLoginResponse:
-    """Prepare connection to a server-capable contact by adding it to the radio and logging in."""
+    """Send one login and wait for the reply.
+
+    Subscriptions are per-attempt because the resolving future can only be
+    settled once — a retry needs a fresh pair.
+    """
     pubkey_prefix = contact.public_key[:12].lower()
-    contact_label = label or get_server_contact_label(contact)
     loop = asyncio.get_running_loop()
     login_future = loop.create_future()
 
@@ -254,9 +266,6 @@ async def prepare_authenticated_contact_connection(
     )
 
     try:
-        logger.info("Adding %s %s to radio", contact_label, contact.public_key[:12])
-        await _ensure_on_radio(mc, contact)
-
         logger.info("Sending login to %s %s", contact_label, contact.public_key[:12])
         login_result = await mc.commands.send_login(contact.public_key, password)
 
@@ -268,10 +277,7 @@ async def prepare_authenticated_contact_connection(
             )
 
         try:
-            return await asyncio.wait_for(
+            return await asyncio.wait_for(login_future, timeout=response_timeout)
-                login_future,
-                timeout=response_timeout,
-            )
         except TimeoutError:
             logger.warning(
                 "No login response from %s %s within %.1fs",
@@ -284,6 +290,105 @@ async def prepare_authenticated_contact_connection(
                 authenticated=False,
                 message=_login_timeout_message(contact_label),
             )
+    finally:
+        success_subscription.unsubscribe()
+        failed_subscription.unsubscribe()
+
+
+async def prepare_authenticated_contact_connection(
+    mc,
+    contact: Contact,
+    password: str,
+    *,
+    label: str | None = None,
+    response_timeout: float = SERVER_LOGIN_RESPONSE_TIMEOUT_SECONDS,
+) -> RepeaterLoginResponse:
+    """Prepare connection to a server-capable contact by adding it to the radio and logging in.
+
+    A login that draws no reply at all may mean the stored direct route has gone
+    stale, so it escalates to one flood retry — mirroring the DM send path, which
+    already resets the path before its final attempt. This is deliberately more
+    than the reference clients do: firmware ``sendLogin`` and ``send_login_sync``
+    are both single-shot, and firmware only clears a stale path when the *host*
+    asks it to (``CMD_RESET_PATH``). Escalating is still the right call because
+    the server side treats an inbound flood as its cue to relearn the return path
+    (``simple_repeater``/``simple_room_server``: ``if (is_flood)
+    client->out_path_len = OUT_PATH_UNKNOWN``), so a flood login is exactly what
+    repairs a broken route in both directions.
+
+    Escalation is bounded to a single extra attempt and only fires when:
+      - the first attempt timed out. An explicit ``LOGIN_FAILED`` means the
+        server heard us and refused, so the route is fine and retrying would
+        just hammer it with bad credentials. A send error is a local radio
+        problem that a different route will not fix.
+      - the first attempt actually used a route. If the contact was already on
+        flood, the retry would be byte-identical for no gain.
+
+    ``reset_path`` clears the route on the radio only; the stored contact route
+    is left alone, so the next ``add_contact`` re-stages it. That matches the DM
+    retry and keeps a single bad login from discarding a route that may be fine.
+    """
+    contact_label = label or get_server_contact_label(contact)
+
+    try:
+        logger.info("Adding %s %s to radio", contact_label, contact.public_key[:12])
+        await _ensure_on_radio(mc, contact)
+
+        response = await _attempt_server_login(
+            mc,
+            contact,
+            password,
+            contact_label=contact_label,
+            response_timeout=response_timeout,
+        )
+        if response.status != "timeout":
+            return response
+
+        if contact.effective_route_source == "flood":
+            logger.debug(
+                "Login to %s %s timed out on flood; no route to escalate from",
+                contact_label,
+                contact.public_key[:12],
+            )
+            return response
+
+        logger.info(
+            "Login to %s %s timed out on the %s route; resetting path and retrying as flood",
+            contact_label,
+            contact.public_key[:12],
+            contact.effective_route_source,
+        )
+        reset_result = await mc.commands.reset_path(contact.public_key)
+        if reset_result is None:
+            logger.warning(
+                "No response from radio for reset_path to %s before flood login retry",
+                contact.public_key[:12],
+            )
+            return response
+        if reset_result.type == EventType.ERROR:
+            logger.warning(
+                "Failed to reset path before flood login retry to %s: %s",
+                contact.public_key[:12],
+                reset_result.payload,
+            )
+            return response
+
+        # Deliberately no _ensure_on_radio here — re-adding the contact would
+        # restore the very route we just cleared, and the retry would go direct.
+        flood_response = await _attempt_server_login(
+            mc,
+            contact,
+            password,
+            contact_label=contact_label,
+            response_timeout=response_timeout,
+        )
+        if flood_response.status == "timeout":
+            return RepeaterLoginResponse(
+                status="timeout",
+                authenticated=False,
+                message=_login_flood_retry_timeout_message(contact_label),
+            )
+        return flood_response
     except HTTPException as exc:
         logger.warning(
             "%s login setup failed for %s: %s",
@@ -296,9 +401,6 @@ async def prepare_authenticated_contact_connection(
             authenticated=False,
             message=f"{_login_send_failed_message(contact_label)} ({exc.detail})",
         )
-    finally:
-        success_subscription.unsubscribe()
-        failed_subscription.unsubscribe()
 
 
 async def batch_cli_fetch(

frontend/AGENTS.md

@@ -422,7 +422,7 @@ State: `useConversationNavigation` controls open/close via `infoPaneChannelKey`.
 
 For repeater contacts (`type=2`), `ConversationPane.tsx` renders `RepeaterDashboard` instead of the normal chat UI (ChatHeader + MessageList + MessageInput).
 
-**Login**: `RepeaterLogin` component — password or guest login via `POST /api/contacts/{key}/repeater/login`.
+**Login**: `RepeaterLogin` component — password or guest login via `POST /api/contacts/{key}/repeater/login`. The frontend sends exactly one request; the backend internally escalates a timed-out login to one flood retry (see `app/AGENTS.md` § "Server login route escalation"), so a single call may take up to two response windows. Do not add a client-side login retry loop on top — a `LOGIN_FAILED` result means the password was refused, not that the route needs another attempt.
 
 **Dashboard panes** (after login): Telemetry, Node Info, Neighbors, ACL, Radio Settings, Regions, Advert Intervals, Owner Info — each fetched via granular `POST /api/contacts/{key}/repeater/{pane}` endpoints. The Regions pane prefers the admin CLI hierarchy and falls back to the guest anon flood-allowed names, so its payload carries a `source` of `cli` or `anon`. Panes retry up to 3 times client-side. `Neighbors` depends on the smaller `node-info` fetch for repeater GPS, not the heavier radio-settings batch. "Load All" fetches all panes serially (parallel would queue behind the radio lock).
 

tests/test_repeater_routes.py

@@ -818,6 +818,162 @@ class TestPrepareRepeaterConnection:
         assert "No login confirmation was heard from the repeater" in (response.message or "")
 
 
+def _make_routed_contact() -> Contact:
+    """A repeater with a learned direct route, so a login goes out direct."""
+    contact = Contact(
+        public_key=KEY_A,
+        name="Repeater",
+        type=2,
+        direct_path="aabb",
+        direct_path_len=2,
+        direct_path_hash_mode=0,
+    )
+    assert contact.effective_route_source == "direct"
+    return contact
+
+
+class TestRepeaterLoginFloodEscalation:
+    """A login that draws no reply escalates to one flood retry.
+
+    Firmware's own ``sendLogin`` is single-shot, so this is RT going beyond the
+    reference clients. It works because the *server* treats an inbound flood as
+    its cue to relearn the return path.
+    """
+
+    @pytest.mark.asyncio
+    async def test_timeout_on_direct_route_resets_path_and_retries_as_flood(self):
+        mc = _mock_mc()
+        contact = _make_routed_contact()
+        subscriptions: dict[EventType, tuple[object, object]] = {}
+        attempts = 0
+
+        def _subscribe(event_type, callback, attribute_filters=None):
+            subscriptions[event_type] = (callback, attribute_filters)
+            return MagicMock(unsubscribe=MagicMock())
+
+        async def _send_login(*args, **kwargs):
+            nonlocal attempts
+            attempts += 1
+            # First attempt (direct) is never answered; the flood retry lands.
+            if attempts > 1:
+                callback, _filters = subscriptions[EventType.LOGIN_SUCCESS]
+                callback(_radio_result(EventType.LOGIN_SUCCESS, {"pubkey_prefix": KEY_A[:12]}))
+            return _radio_result(EventType.MSG_SENT)
+
+        mc.subscribe = MagicMock(side_effect=_subscribe)
+        mc.commands.send_login = AsyncMock(side_effect=_send_login)
+        mc.commands.reset_path = AsyncMock(return_value=_radio_result(EventType.OK))
+
+        with patch("app.routers.repeaters.REPEATER_LOGIN_RESPONSE_TIMEOUT_SECONDS", 0):
+            response = await prepare_repeater_connection(mc, contact, "secret")
+
+        assert response.status == "ok"
+        assert response.authenticated is True
+        assert attempts == 2
+        mc.commands.reset_path.assert_awaited_once_with(KEY_A)
+        # The contact must not be re-added between attempts — that would restore
+        # the route we just cleared and send the retry direct again.
+        assert mc.commands.add_contact.await_count == 1
+
+    @pytest.mark.asyncio
+    async def test_flood_retry_timeout_reports_that_flood_was_tried(self):
+        mc = _mock_mc()
+        contact = _make_routed_contact()
+        mc.commands.reset_path = AsyncMock(return_value=_radio_result(EventType.OK))
+
+        with patch("app.routers.repeaters.REPEATER_LOGIN_RESPONSE_TIMEOUT_SECONDS", 0):
+            response = await prepare_repeater_connection(mc, contact, "secret")
+
+        assert response.status == "timeout"
+        assert response.authenticated is False
+        assert "retry sent as flood" in (response.message or "")
+        assert mc.commands.send_login.await_count == 2
+
+    @pytest.mark.asyncio
+    async def test_no_escalation_when_contact_is_already_flood(self):
+        mc = _mock_mc()
+        contact = _make_contact()  # no direct route -> already flooding
+        mc.commands.reset_path = AsyncMock(return_value=_radio_result(EventType.OK))
+
+        with patch("app.routers.repeaters.REPEATER_LOGIN_RESPONSE_TIMEOUT_SECONDS", 0):
+            response = await prepare_repeater_connection(mc, contact, "secret")
+
+        assert response.status == "timeout"
+        # A flood retry would be byte-identical, so don't bother the repeater
+        assert mc.commands.send_login.await_count == 1
+        mc.commands.reset_path.assert_not_awaited()
+
+    @pytest.mark.asyncio
+    async def test_rejected_login_does_not_escalate(self):
+        """LOGIN_FAILED means the route works and the password doesn't."""
+        mc = _mock_mc()
+        contact = _make_routed_contact()
+        subscriptions: dict[EventType, tuple[object, object]] = {}
+
+        def _subscribe(event_type, callback, attribute_filters=None):
+            subscriptions[event_type] = (callback, attribute_filters)
+            return MagicMock(unsubscribe=MagicMock())
+
+        async def _send_login(*args, **kwargs):
+            callback, _filters = subscriptions[EventType.LOGIN_FAILED]
+            callback(_radio_result(EventType.LOGIN_FAILED, {"pubkey_prefix": KEY_A[:12]}))
+            return _radio_result(EventType.MSG_SENT)
+
+        mc.subscribe = MagicMock(side_effect=_subscribe)
+        mc.commands.send_login = AsyncMock(side_effect=_send_login)
+        mc.commands.reset_path = AsyncMock(return_value=_radio_result(EventType.OK))
+
+        response = await prepare_repeater_connection(mc, contact, "bad")
+
+        assert response.status == "error"
+        assert mc.commands.send_login.await_count == 1
+        mc.commands.reset_path.assert_not_awaited()
+
+    @pytest.mark.asyncio
+    async def test_send_error_does_not_escalate(self):
+        """A local radio failure is not fixed by picking a different route."""
+        mc = _mock_mc()
+        contact = _make_routed_contact()
+        mc.commands.send_login = AsyncMock(
+            return_value=_radio_result(EventType.ERROR, {"err": "busy"})
+        )
+        mc.commands.reset_path = AsyncMock(return_value=_radio_result(EventType.OK))
+
+        response = await prepare_repeater_connection(mc, contact, "secret")
+
+        assert response.status == "error"
+        assert mc.commands.send_login.await_count == 1
+        mc.commands.reset_path.assert_not_awaited()
+
+    @pytest.mark.asyncio
+    async def test_failed_reset_path_returns_original_timeout(self):
+        mc = _mock_mc()
+        contact = _make_routed_contact()
+        mc.commands.reset_path = AsyncMock(
+            return_value=_radio_result(EventType.ERROR, {"err": "nope"})
+        )
+
+        with patch("app.routers.repeaters.REPEATER_LOGIN_RESPONSE_TIMEOUT_SECONDS", 0):
+            response = await prepare_repeater_connection(mc, contact, "secret")
+
+        assert response.status == "timeout"
+        # Without a cleared path the retry would just go direct again
+        assert mc.commands.send_login.await_count == 1
+        assert "retry sent as flood" not in (response.message or "")
+
+    @pytest.mark.asyncio
+    async def test_no_reset_path_response_returns_original_timeout(self):
+        mc = _mock_mc()
+        contact = _make_routed_contact()
+        mc.commands.reset_path = AsyncMock(return_value=None)
+
+        with patch("app.routers.repeaters.REPEATER_LOGIN_RESPONSE_TIMEOUT_SECONDS", 0):
+            response = await prepare_repeater_connection(mc, contact, "secret")
+
+        assert response.status == "timeout"
+        assert mc.commands.send_login.await_count == 1
+
+
 class TestRepeaterStatus:
     @pytest.mark.asyncio
     async def test_success_with_field_mapping(self, test_db):