Diagnostics: surface upstream WS close reason + break the null-spin loop

The Pi journal showed 170 000 'upstream silent NxN' messages within one
second — i.e. ReceiveJsonAsync was returning null in zero time, so the
upstream WS to OpenAI was closed. The relay was spinning until idle
watchdog (30 s) fired. Add IRealtimeUpstream.IsClosed + CloseReason,
populated by OpenAIRealtimeUpstream from the close status / WS exception,
and have PumpAsync stop on close and forward
  error code=upstream_closed message=close status=<X> desc=<Y>
to the device. Closes the spin AND gives us the OpenAI close reason.

README.md

@@ -4,8 +4,12 @@ Voice assistant on a Raspberry Pi talking to the OpenAI Realtime API via an
 ASP.NET backend deployed on Coolify.
 
 See `docs/superpowers/specs/2026-06-11-smart-assistant-design.md` for the full
-design and `docs/superpowers/plans/2026-06-11-smart-assistant-backend-foundation.md`
-for the Plan 1 implementation steps that produced this code.
+design. Implementation plans:
+
+- Plan 1 — `docs/superpowers/plans/2026-06-11-smart-assistant-backend-foundation.md`
+  (Identity, devices, pairing, install endpoints).
+- Plan 2 — `docs/superpowers/plans/2026-06-11-smart-assistant-device-hub-realtime-tools.md`
+  (device WebSocket hub, OpenAI Realtime relay, tools registry).
 
 ## Repo layout
 
@@ -84,11 +88,126 @@ curl -sf -X POST "$COOLIFY_URL/api/v1/deploy?uuid=$APP" \
 Fire-and-forget; the API returns a `deployment_uuid` immediately. Don't poll
 (per the global Coolify rule in `CLAUDE.md`).
 
+## Plan 2: device hub + Realtime relay
+
+- `wss://<domain>/device` — bearer-token WebSocket. Pair a device first to get a token.
+- Handshake: client sends `{type:"hello", device_id, client_version}`, backend
+  replies `hello_ack` with per-device config (voice, model, system prompt, idle
+  timeout, enabled tools).
+- Heartbeat: client sends `{type:"ping"}` every 15 s, backend replies `pong`
+  and updates `Devices.LastSeenAt`.
+- Wake: `{type:"wake"}` opens a Realtime session against OpenAI; backend emits
+  `session_started` (with `conversation_id`), forwards `response.audio.delta`
+  as binary frames to the device, persists `response.audio_transcript.done` as
+  an assistant turn, persists `conversation.item.input_audio_transcription.completed`
+  as a user turn, and emits `assistant_done` after each `response.done`.
+- Tools shipped (per-device toggle via `DeviceConfig.EnabledToolsJson`):
+  - `get_current_time` (server-side, returns `{"now":"<ISO-UTC>"}`),
+  - `end_session` (closes the session after the current `response.done`
+    completes; emits `session_ended(reason="tool")`),
+  - `set_volume` (round-trips a `tool_call`/`tool_result` envelope to the Pi;
+    the Pi side will run `amixer` in Plan 3).
+- Idle timeout: server-side. With no `input_audio_buffer.speech_started` upstream
+  event for `IdleTimeoutSeconds`, the relay emits `session_ended(reason="idle")`.
+
+The Python Pi client (Plan 3) is the production consumer of this surface. A
+quick handshake-only smoke against the live deploy:
+
+```sh
+DOMAIN="https://assistant.volcanic.tes.gd"
+# Reuse the curl recipe above to log in, generate a pair-code, and pair
+# a "smoke" device — capture $TOKEN from the /api/pair response.
+python3 - <<PY
+import asyncio, json, websockets
+URL, TOKEN = "wss://assistant.volcanic.tes.gd/device", "$TOKEN"
+async def main():
+    async with websockets.connect(
+        URL, additional_headers={"Authorization": f"Bearer {TOKEN}"}
+    ) as ws:
+        await ws.send(json.dumps({
+            "type":"hello","device_id":"00000000-0000-0000-0000-000000000000",
+            "client_version":"0.1"}))
+        print("ack:", json.loads(await ws.recv())["type"])
+        await ws.send(json.dumps({"type":"ping"}))
+        print("pong:", json.loads(await ws.recv())["type"])
+asyncio.run(main())
+PY
+```
+
+Don't send `wake` from the smoke — it would open a real OpenAI session and
+burn budget. Plan 3 exercises that end-to-end.
+
+## Plan 3: Pi client
+
+The Python client lives in `client/`. It is shipped to the Pi inside
+`client.tar.gz` (built into the backend Docker image) and installed by
+`install.sh`. Modules:
+
+- `main.py` — entry point. Sets `PA_ALSA_PLUGHW=1` before importing
+  sounddevice. Wires audio, state, wakeword, session, and playback.
+- `state.py` — `IDLE → WAKE_PENDING → LISTENING ↔ ASSISTANT_SPEAKING →
+  IDLE_PENDING → IDLE`. `wakeword_enabled` is true only in `IDLE`;
+  `uplink_enabled` is true only in `LISTENING`.
+- `audio.py` — USB device discovery, persistent 24 kHz/mono/int16 InputStream
+  and OutputStream, and the 24 → 16 kHz resample for the wakeword feed.
+- `wakeword.py` — openwakeword `alexa` model, threshold + cooldown.
+- `playback.py` — single worker thread that owns the OutputStream and
+  implements `set_volume` as a software gain (the USB Speaker Phone has no
+  hardware playback volume control; only a `PCM Playback Switch`).
+- `session.py` — backend WS client, dispatch, tool round-trip,
+  exponential-backoff reconnect.
+- `config.py` — `~/assistant/state/config.json` (mode 0600).
+- `pair.py` — `python -m client.pair --backend <url> [--code <code>]`.
+
+### Updating the Pi
+
+For a fast iteration loop:
+
+```sh
+sshpass -p 'assistant' rsync -az --delete --exclude __pycache__ --exclude tests \
+  client/ pi@192.168.50.115:assistant/code/client/
+sshpass -p 'assistant' ssh pi@192.168.50.115 'systemctl --user restart assistant'
+sshpass -p 'assistant' ssh pi@192.168.50.115 \
+  'journalctl --user -u assistant -n 50 --no-pager'
+```
+
+For a production update (after a backend deploy):
+
+```sh
+sshpass -p 'assistant' ssh pi@192.168.50.115 \
+  'curl -fsSL https://assistant.volcanic.tes.gd/install.sh | bash'
+```
+
+### Local tests
+
+```sh
+python3 -m venv .venv
+.venv/bin/pip install -r requirements-dev.txt requests websockets numpy scipy
+.venv/bin/python -m pytest client/tests
+```
+
+Hardware-touching modules (`audio.py` streams, `wakeword.py`, `playback.py`,
+`session.run_forever`) are smoke-tested via the live deploy. Deterministic
+modules (`state.py`, `config.py`, `pair.py`, the resample helper, the session
+dispatcher) are covered by the pytest suite.
+
+### End-to-end smoke
+
+With the unit running on the Pi:
+
+```sh
+sshpass -p 'assistant' ssh pi@192.168.50.115 'journalctl --user -u assistant -f'
+```
+
+…then say `alexa` near the Speaker Phone. Expected log sequence: `wakeword
+fired`, `state IDLE -> WAKE_PENDING`, `session_started conversation_id=…`,
+`state WAKE_PENDING -> LISTENING`. Speak a question; the assistant replies
+through the speaker (`state LISTENING -> ASSISTANT_SPEAKING`, then
+`assistant_done`). Say "bye" → `session_ended reason=tool`, then `state
+LISTENING -> IDLE_PENDING -> IDLE`. The conversation row + turns are visible
+via the deployed backend's DB (admin UI in Plan 4).
+
 ## What's next
 
-- **Plan 2** — backend device hub + OpenAI Realtime relay + tools registry
-  (the `end_session`, `get_current_time`, `set_volume` tools the spec describes).
-- **Plan 3** — full Python client on the Pi (wakeword, VAD, audio plumbing,
-  state machine, install + pair CLIs).
-- **Plan 4** — React management UI for the admin dashboard and conversation
-  history viewer.
+- **Plan 4** — React management UI for the admin dashboard, per-device
+  config editor, and conversation history viewer.

backend.tests/ApiFactory.cs

@@ -1,7 +1,11 @@
+using backend.Realtime;
 using Microsoft.AspNetCore.Hosting;
 using Microsoft.AspNetCore.Mvc.Testing;
+using Microsoft.AspNetCore.TestHost;
 using Microsoft.Data.Sqlite;
 using Microsoft.Extensions.Configuration;
+using Microsoft.Extensions.DependencyInjection;
+using Microsoft.Extensions.Logging;
 
 namespace backend.tests;
 
@@ -17,8 +21,14 @@ public class ApiFactory : WebApplicationFactory<Program>
             cfg.AddInMemoryCollection(new Dictionary<string, string?>
             {
                 ["ConnectionStrings:Default"] = $"Data Source={_dbPath}",
+                ["OPENAI_API_KEY"] = "test-key",
             });
         });
+
+        builder.ConfigureTestServices(services =>
+        {
+            services.AddScoped<RealtimeSessionFactory, FakeRealtimeSessionFactory>();
+        });
     }
 
     protected override void Dispose(bool disposing)
@@ -32,3 +42,24 @@ public class ApiFactory : WebApplicationFactory<Program>
         }
     }
 }
+
+public class FakeRealtimeSessionFactory(
+    OpenAIKeyProvider keys,
+    backend.Tools.ToolRegistry registry,
+    backend.Conversations.ConversationLog log,
+    ILoggerFactory loggerFactory)
+    : RealtimeSessionFactory(keys, registry, log, loggerFactory)
+{
+    public static ScriptedRealtimeUpstream Upstream { get; set; } = new();
+
+    public override Task<RealtimeSession> CreateAsync(
+        Guid deviceId, RealtimeSettings cfg, ISessionOutput output,
+        backend.Tools.IDeviceChannel channel, CancellationToken ct)
+    {
+        Upstream.Push(new System.Text.Json.Nodes.JsonObject { ["type"] = "session.updated" });
+        var s = new RealtimeSession(
+            deviceId, Upstream, output, log, registry, cfg, channel,
+            loggerFactory.CreateLogger<RealtimeSession>());
+        return Task.FromResult(s);
+    }
+}

backend.tests/ConversationLogTests.cs

@@ -0,0 +1,71 @@
+using backend.Conversations;
+using backend.Data;
+using FluentAssertions;
+using Microsoft.EntityFrameworkCore;
+using Microsoft.Extensions.DependencyInjection;
+using Xunit;
+
+namespace backend.tests;
+
+public class ConversationLogTests(ApiFactory factory) : IClassFixture<ApiFactory>
+{
+    private readonly ApiFactory _factory = factory;
+
+    private async Task<(AppDbContext db, ConversationLog log, Guid deviceId)> ArrangeAsync()
+    {
+        var scope = _factory.Services.CreateScope();
+        var db = scope.ServiceProvider.GetRequiredService<AppDbContext>();
+        var log = scope.ServiceProvider.GetRequiredService<ConversationLog>();
+        var deviceId = Guid.NewGuid();
+        return (db, log, deviceId);
+    }
+
+    [Fact]
+    public async Task Start_creates_conversation_row()
+    {
+        var (db, log, deviceId) = await ArrangeAsync();
+
+        var convo = await log.StartAsync(deviceId, default);
+
+        var row = await db.Conversations.FirstAsync(c => c.Id == convo.Id);
+        row.DeviceId.Should().Be(deviceId);
+        row.EndedAt.Should().BeNull();
+        row.EndReason.Should().Be(EndReason.Unknown);
+    }
+
+    [Fact]
+    public async Task Turns_are_persisted_with_monotonic_index_per_conversation()
+    {
+        var (db, log, deviceId) = await ArrangeAsync();
+        var convo = await log.StartAsync(deviceId, default);
+
+        await log.AppendUserAsync(convo.Id, "hi", default);
+        await log.AppendAssistantAsync(convo.Id, "hello there", default);
+        await log.AppendToolAsync(convo.Id, "get_current_time", "{}", "{\"now\":\"2026-06-11T00:00:00Z\"}", default);
+
+        var turns = await db.Turns
+            .Where(t => t.ConversationId == convo.Id)
+            .OrderBy(t => t.Index)
+            .ToListAsync();
+
+        turns.Should().HaveCount(3);
+        turns.Select(t => t.Index).Should().Equal(0, 1, 2);
+        turns.Select(t => t.Role).Should().Equal(TurnRole.User, TurnRole.Assistant, TurnRole.Tool);
+        turns[0].Text.Should().Be("hi");
+        turns[1].Text.Should().Be("hello there");
+        turns[2].ToolName.Should().Be("get_current_time");
+    }
+
+    [Fact]
+    public async Task End_sets_ended_at_and_reason()
+    {
+        var (db, log, deviceId) = await ArrangeAsync();
+        var convo = await log.StartAsync(deviceId, default);
+
+        await log.EndAsync(convo.Id, EndReason.Idle, default);
+
+        var row = await db.Conversations.FirstAsync(c => c.Id == convo.Id);
+        row.EndedAt.Should().NotBeNull();
+        row.EndReason.Should().Be(EndReason.Idle);
+    }
+}

backend.tests/DeviceHubHandshakeTests.cs

@@ -0,0 +1,91 @@
+using System.Net.Http.Json;
+using System.Net.WebSockets;
+using System.Text;
+using System.Text.Json;
+using FluentAssertions;
+using Xunit;
+
+namespace backend.tests;
+
+public class DeviceHubHandshakeTests(ApiFactory factory) : IClassFixture<ApiFactory>
+{
+    private readonly ApiFactory _factory = factory;
+
+    private async Task<string> PairAndGetTokenAsync(string email)
+    {
+        var http = _factory.CreateClient();
+        await http.PostAsJsonAsync("/api/auth/register",
+            new { email, password = "Passw0rd!" });
+        await http.PostAsJsonAsync("/api/auth/login",
+            new { email, password = "Passw0rd!" });
+        var codeRes = await http.PostAsJsonAsync("/api/pair-code", new { name = "ws-test" });
+        var code = (await codeRes.Content.ReadFromJsonAsync<Dictionary<string, object>>())!["code"].ToString()!;
+
+        var pairRes = await _factory.CreateClient().PostAsJsonAsync("/api/pair",
+            new { code, hostname = "smoke", client_version = "0.1" });
+        var pairBody = await pairRes.Content.ReadFromJsonAsync<Dictionary<string, object>>();
+        return pairBody!["device_token"].ToString()!;
+    }
+
+    [Fact]
+    public async Task Connect_without_token_fails_upgrade()
+    {
+        var client = _factory.Server.CreateWebSocketClient();
+        var uri = new Uri(_factory.Server.BaseAddress, "device");
+        var act = async () => await client.ConnectAsync(uri, CancellationToken.None);
+        await act.Should().ThrowAsync<Exception>();
+    }
+
+    [Fact]
+    public async Task Hello_ack_includes_config()
+    {
+        var token = await PairAndGetTokenAsync("ws1@example.com");
+        var client = _factory.Server.CreateWebSocketClient();
+        client.ConfigureRequest = req => req.Headers["Authorization"] = $"Bearer {token}";
+
+        var uri = new Uri(_factory.Server.BaseAddress, "device");
+        var ws = await client.ConnectAsync(uri, CancellationToken.None);
+
+        await SendJson(ws, new { type = "hello", device_id = Guid.NewGuid(), client_version = "0.1" });
+        var ack = await ReceiveJson(ws);
+        ack.GetProperty("type").GetString().Should().Be("hello_ack");
+        ack.GetProperty("config").GetProperty("voice").GetString().Should().NotBeNullOrEmpty();
+        ack.GetProperty("config").GetProperty("model").GetString().Should().NotBeNullOrEmpty();
+        ack.GetProperty("config").GetProperty("enabled_tools").GetArrayLength().Should().BeGreaterThan(0);
+    }
+
+    [Fact]
+    public async Task Ping_replies_with_pong()
+    {
+        var token = await PairAndGetTokenAsync("ws2@example.com");
+        var client = _factory.Server.CreateWebSocketClient();
+        client.ConfigureRequest = req => req.Headers["Authorization"] = $"Bearer {token}";
+        var uri = new Uri(_factory.Server.BaseAddress, "device");
+        var ws = await client.ConnectAsync(uri, CancellationToken.None);
+
+        await SendJson(ws, new { type = "hello", device_id = Guid.NewGuid(), client_version = "0.1" });
+        await ReceiveJson(ws);
+        await SendJson(ws, new { type = "ping" });
+        var pong = await ReceiveJson(ws);
+        pong.GetProperty("type").GetString().Should().Be("pong");
+    }
+
+    private static async Task SendJson(WebSocket ws, object payload)
+    {
+        var bytes = Encoding.UTF8.GetBytes(JsonSerializer.Serialize(payload));
+        await ws.SendAsync(bytes, WebSocketMessageType.Text, true, CancellationToken.None);
+    }
+
+    private static async Task<JsonElement> ReceiveJson(WebSocket ws)
+    {
+        var buf = new byte[8192];
+        using var stream = new MemoryStream();
+        while (true)
+        {
+            var r = await ws.ReceiveAsync(buf, CancellationToken.None);
+            stream.Write(buf, 0, r.Count);
+            if (r.EndOfMessage) break;
+        }
+        return JsonDocument.Parse(stream.ToArray()).RootElement.Clone();
+    }
+}

backend.tests/DeviceHubWakeFlowTests.cs

@@ -0,0 +1,129 @@
+using System.Net.Http.Json;
+using System.Net.WebSockets;
+using System.Text;
+using System.Text.Json;
+using System.Text.Json.Nodes;
+using FluentAssertions;
+using Xunit;
+
+namespace backend.tests;
+
+[Collection("WakeFlow")] // tests touch FakeRealtimeSessionFactory.Upstream static — serialize
+public class DeviceHubWakeFlowTests(ApiFactory factory) : IClassFixture<ApiFactory>
+{
+    private readonly ApiFactory _factory = factory;
+
+    private async Task<string> PairAndGetTokenAsync(string email)
+    {
+        var http = _factory.CreateClient();
+        await http.PostAsJsonAsync("/api/auth/register",
+            new { email, password = "Passw0rd!" });
+        await http.PostAsJsonAsync("/api/auth/login",
+            new { email, password = "Passw0rd!" });
+        var codeRes = await http.PostAsJsonAsync("/api/pair-code", new { name = "wake-test" });
+        var code = (await codeRes.Content.ReadFromJsonAsync<Dictionary<string, object>>())!["code"].ToString()!;
+        var pairRes = await _factory.CreateClient().PostAsJsonAsync("/api/pair",
+            new { code, hostname = "smoke", client_version = "0.1" });
+        var pairBody = await pairRes.Content.ReadFromJsonAsync<Dictionary<string, object>>();
+        return pairBody!["device_token"].ToString()!;
+    }
+
+    [Fact]
+    public async Task Wake_opens_session_started_then_assistant_done_then_session_ended_on_cancel()
+    {
+        FakeRealtimeSessionFactory.Upstream = new ScriptedRealtimeUpstream();
+        var token = await PairAndGetTokenAsync("wake1@example.com");
+        var client = _factory.Server.CreateWebSocketClient();
+        client.ConfigureRequest = req => req.Headers["Authorization"] = $"Bearer {token}";
+        var uri = new Uri(_factory.Server.BaseAddress, "device");
+        var ws = await client.ConnectAsync(uri, CancellationToken.None);
+
+        await SendJson(ws, new { type = "hello", device_id = Guid.NewGuid(), client_version = "0.1" });
+        await ReceiveJson(ws); // ack
+
+        await SendJson(ws, new { type = "wake", at = 0L });
+        var started = await ReceiveJson(ws);
+        started.GetProperty("type").GetString().Should().Be("session_started");
+
+        FakeRealtimeSessionFactory.Upstream.Push(new JsonObject
+        {
+            ["type"] = "response.audio_transcript.done",
+            ["transcript"] = "hello!",
+        });
+        FakeRealtimeSessionFactory.Upstream.Push(new JsonObject { ["type"] = "response.done" });
+
+        var done = await ReceiveUntil(ws, "assistant_done");
+        done.GetProperty("type").GetString().Should().Be("assistant_done");
+
+        await SendJson(ws, new { type = "cancel" });
+        var ended = await ReceiveUntil(ws, "session_ended");
+        ended.GetProperty("type").GetString().Should().Be("session_ended");
+    }
+
+    [Fact]
+    public async Task Binary_uplink_after_wake_reaches_upstream_as_input_audio_buffer_append()
+    {
+        FakeRealtimeSessionFactory.Upstream = new ScriptedRealtimeUpstream();
+        var token = await PairAndGetTokenAsync("wake2@example.com");
+        var client = _factory.Server.CreateWebSocketClient();
+        client.ConfigureRequest = req => req.Headers["Authorization"] = $"Bearer {token}";
+        var uri = new Uri(_factory.Server.BaseAddress, "device");
+        var ws = await client.ConnectAsync(uri, CancellationToken.None);
+
+        await SendJson(ws, new { type = "hello", device_id = Guid.NewGuid(), client_version = "0.1" });
+        await ReceiveJson(ws);
+        await SendJson(ws, new { type = "wake", at = 0L });
+        await ReceiveJson(ws); // session_started
+
+        var frame = new byte[3840];
+        for (int i = 0; i < frame.Length; i++) frame[i] = 0x55;
+        await ws.SendAsync(frame, WebSocketMessageType.Binary, true, CancellationToken.None);
+
+        // Allow time for the pump's 250 ms poll + send.
+        var deadline = DateTime.UtcNow + TimeSpan.FromSeconds(2);
+        while (DateTime.UtcNow < deadline)
+        {
+            if (FakeRealtimeSessionFactory.Upstream.Sent.Any(e =>
+                (string?)e["type"] == "input_audio_buffer.append")) break;
+            await Task.Delay(50);
+        }
+        FakeRealtimeSessionFactory.Upstream.Sent
+            .Any(e => (string?)e["type"] == "input_audio_buffer.append")
+            .Should().BeTrue();
+
+        await SendJson(ws, new { type = "cancel" });
+        await ReceiveJson(ws);
+    }
+
+    private static async Task SendJson(WebSocket ws, object payload)
+    {
+        var bytes = Encoding.UTF8.GetBytes(JsonSerializer.Serialize(payload));
+        await ws.SendAsync(bytes, WebSocketMessageType.Text, true, CancellationToken.None);
+    }
+
+    private static async Task<JsonElement> ReceiveJson(WebSocket ws)
+    {
+        var buf = new byte[16384];
+        using var stream = new MemoryStream();
+        while (true)
+        {
+            var r = await ws.ReceiveAsync(buf, CancellationToken.None);
+            if (r.MessageType == WebSocketMessageType.Binary) { stream.Write(buf, 0, r.Count); continue; }
+            stream.Write(buf, 0, r.Count);
+            if (r.EndOfMessage) break;
+        }
+        return JsonDocument.Parse(stream.ToArray()).RootElement.Clone();
+    }
+
+    // The relay forwards an upstream-event "debug" envelope for every event
+    // it sees, so envelopes the test cares about are interleaved with debugs.
+    private static async Task<JsonElement> ReceiveUntil(WebSocket ws, string type)
+    {
+        for (int i = 0; i < 50; i++)
+        {
+            var env = await ReceiveJson(ws);
+            if (env.GetProperty("type").GetString() == type) return env;
+        }
+        throw new TimeoutException($"never saw envelope type={type}");
+    }
+}

backend.tests/DeviceRegistryTests.cs

@@ -0,0 +1,25 @@
+using System.Net.WebSockets;
+using backend.DeviceHub;
+using FluentAssertions;
+using Xunit;
+
+namespace backend.tests;
+
+public class DeviceRegistryTests
+{
+    [Fact]
+    public void Register_unregister_round_trip()
+    {
+        var reg = new DeviceRegistry();
+        var id = Guid.NewGuid();
+        var dev = new ActiveDevice(id, new ClientWebSocket());
+        reg.TryRegister(dev).Should().BeTrue();
+        reg.TryRegister(dev).Should().BeFalse();
+        reg.IsOnline(id).Should().BeTrue();
+        reg.OnlineCount.Should().Be(1);
+
+        reg.Unregister(id).Should().BeTrue();
+        reg.IsOnline(id).Should().BeFalse();
+        reg.OnlineCount.Should().Be(0);
+    }
+}

backend.tests/EndSessionToolTests.cs

@@ -0,0 +1,35 @@
+using System.Text.Json;
+using backend.Tools;
+using FluentAssertions;
+using Xunit;
+
+namespace backend.tests;
+
+public class EndSessionToolTests
+{
+    [Fact]
+    public async Task Returns_ok_true()
+    {
+        var tool = new EndSessionTool();
+        var ctx = new DeviceContext(Guid.NewGuid(), Guid.NewGuid(), new NoopChannel());
+
+        var res = await tool.ExecuteAsync(
+            new ToolInvocation("call_1", JsonDocument.Parse("{}").RootElement), ctx, default);
+
+        res.Output.GetProperty("ok").GetBoolean().Should().BeTrue();
+    }
+
+    [Fact]
+    public void Name_and_runs_during_response_match_spec()
+    {
+        var tool = new EndSessionTool();
+        tool.Name.Should().Be("end_session");
+        tool.RunsDuringResponse.Should().BeFalse();
+    }
+
+    private class NoopChannel : IDeviceChannel
+    {
+        public Task<JsonElement> CallPiToolAsync(string n, JsonElement a, CancellationToken ct)
+            => throw new NotSupportedException();
+    }
+}

backend.tests/FakeDeviceChannel.cs

@@ -0,0 +1,32 @@
+using System.Collections.Concurrent;
+using System.Text.Json;
+using backend.DeviceHub;
+using backend.Tools;
+
+namespace backend.tests;
+
+public class FakeDeviceChannel : IDeviceChannel, IDeviceClient
+{
+    public List<HubEnvelope> SentEnvelopes { get; } = new();
+    public List<byte[]> SentBinary { get; } = new();
+    public ConcurrentQueue<JsonElement> PiToolReplies { get; } = new();
+
+    public Task SendEnvelopeAsync<T>(T envelope, CancellationToken ct) where T : HubEnvelope
+    {
+        SentEnvelopes.Add(envelope);
+        return Task.CompletedTask;
+    }
+
+    public Task SendBinaryAsync(ReadOnlyMemory<byte> bytes, CancellationToken ct)
+    {
+        SentBinary.Add(bytes.ToArray());
+        return Task.CompletedTask;
+    }
+
+    public Task<JsonElement> CallPiToolAsync(string name, JsonElement args, CancellationToken ct)
+    {
+        if (PiToolReplies.TryDequeue(out var reply))
+            return Task.FromResult(reply);
+        return Task.FromResult(JsonDocument.Parse("""{"ok":true}""").RootElement);
+    }
+}

backend.tests/GetCurrentTimeToolTests.cs

@@ -0,0 +1,29 @@
+using System.Text.Json;
+using backend.Tools;
+using FluentAssertions;
+using Xunit;
+
+namespace backend.tests;
+
+public class GetCurrentTimeToolTests
+{
+    [Fact]
+    public async Task Returns_now_field_with_iso_utc_timestamp()
+    {
+        var tool = new GetCurrentTimeTool();
+        var args = JsonDocument.Parse("{}").RootElement;
+        var ctx = new DeviceContext(Guid.NewGuid(), Guid.NewGuid(), new NoopChannel());
+
+        var res = await tool.ExecuteAsync(new ToolInvocation("call_1", args), ctx, default);
+
+        res.Output.TryGetProperty("now", out var now).Should().BeTrue();
+        now.GetString().Should().MatchRegex(@"^\d{4}-\d{2}-\d{2}T\d{2}:\d{2}:\d{2}");
+        now.GetString()!.Should().EndWith("Z");
+    }
+
+    private class NoopChannel : IDeviceChannel
+    {
+        public Task<JsonElement> CallPiToolAsync(string name, JsonElement args, CancellationToken ct)
+            => throw new NotSupportedException();
+    }
+}

backend.tests/RealtimeSessionAudioTests.cs

@@ -0,0 +1,81 @@
+using System.Text.Json.Nodes;
+using backend.Conversations;
+using backend.Realtime;
+using backend.Tools;
+using FluentAssertions;
+using Microsoft.Extensions.DependencyInjection;
+using Xunit;
+
+namespace backend.tests;
+
+public class RealtimeSessionAudioTests(ApiFactory factory) : IClassFixture<ApiFactory>
+{
+    private readonly ApiFactory _factory = factory;
+
+    [Fact]
+    public async Task Uplink_bytes_become_input_audio_buffer_append()
+    {
+        using var scope = _factory.Services.CreateScope();
+        var log = scope.ServiceProvider.GetRequiredService<ConversationLog>();
+        var reg = new ToolRegistry(Array.Empty<ITool>());
+        var upstream = new ScriptedRealtimeUpstream();
+        var sink = new RecordingSink();
+        var cfg = new RealtimeSettings("m", "v", "p", 30, new HashSet<string>());
+        var session = new RealtimeSession(Guid.NewGuid(), upstream, sink, log, reg, cfg,
+            new FakeDeviceChannel(), NullLogger.Instance);
+
+        upstream.Push(new JsonObject { ["type"] = "session.updated" });
+        var cts = new CancellationTokenSource();
+        var run = session.RunAsync(cts.Token);
+        await sink.WaitForAsync("session_started", TimeSpan.FromSeconds(2));
+
+        var frame = new byte[3840];
+        for (int i = 0; i < frame.Length; i++) frame[i] = (byte)(i % 256);
+        await session.WriteUplinkAsync(frame, default);
+
+        await Task.Delay(100);
+
+        var appended = upstream.Sent
+            .FirstOrDefault(e => (string?)e["type"] == "input_audio_buffer.append");
+        appended.Should().NotBeNull();
+        var b64 = (string?)appended!["audio"];
+        Convert.FromBase64String(b64!).Length.Should().Be(3840);
+
+        cts.Cancel();
+        await run;
+    }
+
+    [Fact]
+    public async Task Audio_delta_from_upstream_becomes_binary_to_device()
+    {
+        using var scope = _factory.Services.CreateScope();
+        var log = scope.ServiceProvider.GetRequiredService<ConversationLog>();
+        var reg = new ToolRegistry(Array.Empty<ITool>());
+        var upstream = new ScriptedRealtimeUpstream();
+        var sink = new RecordingSink();
+        var cfg = new RealtimeSettings("m", "v", "p", 30, new HashSet<string>());
+        var session = new RealtimeSession(Guid.NewGuid(), upstream, sink, log, reg, cfg,
+            new FakeDeviceChannel(), NullLogger.Instance);
+
+        upstream.Push(new JsonObject { ["type"] = "session.updated" });
+        var cts = new CancellationTokenSource();
+        var run = session.RunAsync(cts.Token);
+        await sink.WaitForAsync("session_started", TimeSpan.FromSeconds(2));
+
+        var payload = new byte[3840];
+        for (int i = 0; i < payload.Length; i++) payload[i] = (byte)((i + 1) % 256);
+        upstream.Push(new JsonObject
+        {
+            ["type"] = "response.audio.delta",
+            ["delta"] = Convert.ToBase64String(payload),
+        });
+
+        await Task.Delay(100);
+        sink.Binary.Should().NotBeEmpty();
+        sink.Binary[0].Length.Should().Be(3840);
+        sink.Binary[0][0].Should().Be(1);
+
+        cts.Cancel();
+        await run;
+    }
+}

backend.tests/RealtimeSessionEndSessionTests.cs

@@ -0,0 +1,54 @@
+using System.Text.Json.Nodes;
+using backend.Conversations;
+using backend.Realtime;
+using backend.Tools;
+using FluentAssertions;
+using Microsoft.Extensions.DependencyInjection;
+using Xunit;
+
+namespace backend.tests;
+
+public class RealtimeSessionEndSessionTests(ApiFactory factory) : IClassFixture<ApiFactory>
+{
+    private readonly ApiFactory _factory = factory;
+
+    [Fact]
+    public async Task End_session_call_closes_session_with_reason_tool_and_skips_response_create()
+    {
+        using var scope = _factory.Services.CreateScope();
+        var log = scope.ServiceProvider.GetRequiredService<ConversationLog>();
+        var reg = new ToolRegistry(new ITool[] { new EndSessionTool() });
+        var upstream = new ScriptedRealtimeUpstream();
+        var sink = new RecordingSink();
+        var cfg = new RealtimeSettings("m", "v", "p", 30, new HashSet<string> { "end_session" });
+        var session = new RealtimeSession(Guid.NewGuid(), upstream, sink, log, reg, cfg,
+            new FakeDeviceChannel(), NullLogger.Instance);
+
+        upstream.Push(new JsonObject { ["type"] = "session.updated" });
+        var run = session.RunAsync(default);
+        await sink.WaitForAsync("session_started", TimeSpan.FromSeconds(2));
+
+        upstream.Push(new JsonObject
+        {
+            ["type"] = "response.audio.delta",
+            ["delta"] = Convert.ToBase64String(new byte[3840]),
+        });
+        upstream.Push(new JsonObject
+        {
+            ["type"] = "response.function_call_arguments.done",
+            ["call_id"] = "c1",
+            ["name"] = "end_session",
+            ["arguments"] = "{}",
+        });
+        upstream.Push(new JsonObject { ["type"] = "response.done" });
+
+        await run;
+
+        sink.Binary.Should().NotBeEmpty();
+
+        var ended = sink.Envelopes.Last(e => (string?)e["type"] == "session_ended");
+        ((string?)ended["reason"]).Should().Be("tool");
+
+        upstream.Sent.Should().NotContain(e => (string?)e["type"] == "response.create");
+    }
+}

backend.tests/RealtimeSessionIdleTimeoutTests.cs

@@ -0,0 +1,63 @@
+using System.Text.Json.Nodes;
+using backend.Conversations;
+using backend.Realtime;
+using backend.Tools;
+using FluentAssertions;
+using Microsoft.Extensions.DependencyInjection;
+using Xunit;
+
+namespace backend.tests;
+
+public class RealtimeSessionIdleTimeoutTests(ApiFactory factory) : IClassFixture<ApiFactory>
+{
+    private readonly ApiFactory _factory = factory;
+
+    [Fact]
+    public async Task Session_ends_with_reason_idle_when_no_speech_for_timeout()
+    {
+        using var scope = _factory.Services.CreateScope();
+        var log = scope.ServiceProvider.GetRequiredService<ConversationLog>();
+        var reg = new ToolRegistry(Array.Empty<ITool>());
+        var upstream = new ScriptedRealtimeUpstream();
+        var sink = new RecordingSink();
+        var cfg = new RealtimeSettings("m", "v", "p", IdleTimeoutSeconds: 1, new HashSet<string>());
+        var session = new RealtimeSession(Guid.NewGuid(), upstream, sink, log, reg, cfg,
+            new FakeDeviceChannel(), NullLogger.Instance);
+
+        upstream.Push(new JsonObject { ["type"] = "session.updated" });
+        var run = session.RunAsync(default);
+        await sink.WaitForAsync("session_started", TimeSpan.FromSeconds(2));
+
+        await run.WaitAsync(TimeSpan.FromSeconds(4));
+
+        var ended = sink.Envelopes.Last(e => (string?)e["type"] == "session_ended");
+        ((string?)ended["reason"]).Should().Be("idle");
+    }
+
+    [Fact]
+    public async Task Speech_started_resets_the_idle_timer()
+    {
+        using var scope = _factory.Services.CreateScope();
+        var log = scope.ServiceProvider.GetRequiredService<ConversationLog>();
+        var reg = new ToolRegistry(Array.Empty<ITool>());
+        var upstream = new ScriptedRealtimeUpstream();
+        var sink = new RecordingSink();
+        var cfg = new RealtimeSettings("m", "v", "p", IdleTimeoutSeconds: 2, new HashSet<string>());
+        var session = new RealtimeSession(Guid.NewGuid(), upstream, sink, log, reg, cfg,
+            new FakeDeviceChannel(), NullLogger.Instance);
+
+        upstream.Push(new JsonObject { ["type"] = "session.updated" });
+        var run = session.RunAsync(default);
+        await sink.WaitForAsync("session_started", TimeSpan.FromSeconds(2));
+
+        await Task.Delay(1000);
+        upstream.Push(new JsonObject { ["type"] = "input_audio_buffer.speech_started" });
+        await Task.Delay(1500);
+
+        sink.Envelopes.Any(e => (string?)e["type"] == "session_ended").Should().BeFalse();
+
+        await run.WaitAsync(TimeSpan.FromSeconds(4));
+        var ended = sink.Envelopes.Last(e => (string?)e["type"] == "session_ended");
+        ((string?)ended["reason"]).Should().Be("idle");
+    }
+}

backend.tests/RealtimeSessionLifecycleTests.cs

@@ -0,0 +1,92 @@
+using System.Text.Json.Nodes;
+using backend.Conversations;
+using backend.Data;
+using backend.Realtime;
+using backend.Tools;
+using FluentAssertions;
+using Microsoft.Extensions.DependencyInjection;
+using Xunit;
+
+namespace backend.tests;
+
+public class RealtimeSessionLifecycleTests(ApiFactory factory) : IClassFixture<ApiFactory>
+{
+    private readonly ApiFactory _factory = factory;
+
+    [Fact]
+    public async Task Open_emits_session_update_then_session_started_then_session_ended_on_cancel()
+    {
+        using var scope = _factory.Services.CreateScope();
+        var log = scope.ServiceProvider.GetRequiredService<ConversationLog>();
+        var registry = new ToolRegistry(new ITool[] { new GetCurrentTimeTool() });
+
+        var upstream = new ScriptedRealtimeUpstream();
+        var sink = new RecordingSink();
+        var cfg = new RealtimeSettings(
+            Model: "gpt-4o-realtime-preview",
+            Voice: "alloy",
+            SystemPrompt: "test",
+            IdleTimeoutSeconds: 30,
+            EnabledTools: new HashSet<string> { "get_current_time" });
+
+        var session = new RealtimeSession(
+            deviceId: Guid.NewGuid(),
+            upstream, sink, log, registry, cfg,
+            channel: new FakeDeviceChannel(),
+            logger: NullLogger.Instance);
+
+        upstream.Push(new JsonObject { ["type"] = "session.updated" });
+
+        var cts = new CancellationTokenSource();
+        var runTask = session.RunAsync(cts.Token);
+
+        await sink.WaitForAsync("session_started", TimeSpan.FromSeconds(2));
+
+        cts.Cancel();
+        await runTask;
+
+        upstream.Sent.Select(e => (string?)e["type"]).Should().StartWith("session.update");
+        sink.Envelopes.Select(e => (string?)e["type"]).Should().Contain("session_started");
+        sink.Envelopes.Select(e => (string?)e["type"]).Should().Contain("session_ended");
+    }
+}
+
+internal class NullLogger : Microsoft.Extensions.Logging.ILogger
+{
+    public static readonly NullLogger Instance = new();
+    public IDisposable? BeginScope<TState>(TState state) where TState : notnull => null;
+    public bool IsEnabled(Microsoft.Extensions.Logging.LogLevel logLevel) => false;
+    public void Log<TState>(Microsoft.Extensions.Logging.LogLevel l, Microsoft.Extensions.Logging.EventId e,
+        TState s, Exception? ex, Func<TState, Exception?, string> f) { }
+}
+
+internal class RecordingSink : ISessionOutput
+{
+    public List<JsonObject> Envelopes { get; } = new();
+    public List<byte[]> Binary { get; } = new();
+    public Task WriteEnvelopeAsync(JsonObject e, CancellationToken ct)
+    {
+        lock (Envelopes) Envelopes.Add(e);
+        _gate.Release();
+        return Task.CompletedTask;
+    }
+    public Task WriteBinaryAsync(ReadOnlyMemory<byte> b, CancellationToken ct)
+    {
+        Binary.Add(b.ToArray());
+        return Task.CompletedTask;
+    }
+    private readonly SemaphoreSlim _gate = new(0, int.MaxValue);
+    public async Task WaitForAsync(string type, TimeSpan timeout)
+    {
+        var deadline = DateTime.UtcNow + timeout;
+        while (DateTime.UtcNow < deadline)
+        {
+            lock (Envelopes)
+            {
+                if (Envelopes.Any(e => (string?)e["type"] == type)) return;
+            }
+            await _gate.WaitAsync(timeout);
+        }
+        throw new TimeoutException($"never saw {type}");
+    }
+}

backend.tests/RealtimeSessionToolsTests.cs

@@ -0,0 +1,67 @@
+using System.Text.Json.Nodes;
+using backend.Conversations;
+using backend.Data;
+using backend.Realtime;
+using backend.Tools;
+using FluentAssertions;
+using Microsoft.EntityFrameworkCore;
+using Microsoft.Extensions.DependencyInjection;
+using Xunit;
+
+namespace backend.tests;
+
+public class RealtimeSessionToolsTests(ApiFactory factory) : IClassFixture<ApiFactory>
+{
+    private readonly ApiFactory _factory = factory;
+
+    [Fact]
+    public async Task Get_current_time_call_returns_function_call_output_and_persists_tool_turn()
+    {
+        using var scope = _factory.Services.CreateScope();
+        var db = scope.ServiceProvider.GetRequiredService<AppDbContext>();
+        var log = scope.ServiceProvider.GetRequiredService<ConversationLog>();
+        var reg = new ToolRegistry(new ITool[] { new GetCurrentTimeTool() });
+        var upstream = new ScriptedRealtimeUpstream();
+        var sink = new RecordingSink();
+        var cfg = new RealtimeSettings("m", "v", "p", 30, new HashSet<string> { "get_current_time" });
+        var session = new RealtimeSession(Guid.NewGuid(), upstream, sink, log, reg, cfg,
+            new FakeDeviceChannel(), NullLogger.Instance);
+
+        upstream.Push(new JsonObject { ["type"] = "session.updated" });
+        var cts = new CancellationTokenSource();
+        var run = session.RunAsync(cts.Token);
+        await sink.WaitForAsync("session_started", TimeSpan.FromSeconds(2));
+
+        upstream.Push(new JsonObject
+        {
+            ["type"] = "response.function_call_arguments.done",
+            ["call_id"] = "call_abc",
+            ["name"] = "get_current_time",
+            ["arguments"] = "{}",
+        });
+        upstream.Push(new JsonObject { ["type"] = "response.done" });
+
+        var deadline = DateTime.UtcNow + TimeSpan.FromSeconds(2);
+        while (DateTime.UtcNow < deadline && !upstream.Sent.Any(e =>
+            (string?)e["type"] == "conversation.item.create"))
+        {
+            await Task.Delay(20);
+        }
+
+        var item = upstream.Sent.First(e => (string?)e["type"] == "conversation.item.create");
+        ((string?)item["item"]!["type"]).Should().Be("function_call_output");
+        ((string?)item["item"]!["call_id"]).Should().Be("call_abc");
+        ((string?)item["item"]!["output"]).Should().Contain("\"now\"");
+
+        upstream.Sent.Should().Contain(e => (string?)e["type"] == "response.create");
+
+        var convoId = session.ConversationId!.Value;
+        var toolTurn = await db.Turns.FirstAsync(t =>
+            t.ConversationId == convoId && t.Role == TurnRole.Tool);
+        toolTurn.ToolName.Should().Be("get_current_time");
+        toolTurn.ToolResultJson.Should().Contain("\"now\"");
+
+        cts.Cancel();
+        await run;
+    }
+}

backend.tests/RealtimeSessionTranscriptTests.cs

@@ -0,0 +1,60 @@
+using System.Text.Json.Nodes;
+using backend.Conversations;
+using backend.Data;
+using backend.Realtime;
+using backend.Tools;
+using FluentAssertions;
+using Microsoft.EntityFrameworkCore;
+using Microsoft.Extensions.DependencyInjection;
+using Xunit;
+
+namespace backend.tests;
+
+public class RealtimeSessionTranscriptTests(ApiFactory factory) : IClassFixture<ApiFactory>
+{
+    private readonly ApiFactory _factory = factory;
+
+    [Fact]
+    public async Task User_and_assistant_transcripts_are_persisted_and_response_done_emits_assistant_done()
+    {
+        using var scope = _factory.Services.CreateScope();
+        var db = scope.ServiceProvider.GetRequiredService<AppDbContext>();
+        var log = scope.ServiceProvider.GetRequiredService<ConversationLog>();
+        var reg = new ToolRegistry(Array.Empty<ITool>());
+        var upstream = new ScriptedRealtimeUpstream();
+        var sink = new RecordingSink();
+        var cfg = new RealtimeSettings("m", "v", "p", 30, new HashSet<string>());
+        var session = new RealtimeSession(Guid.NewGuid(), upstream, sink, log, reg, cfg,
+            new FakeDeviceChannel(), NullLogger.Instance);
+
+        upstream.Push(new JsonObject { ["type"] = "session.updated" });
+        var cts = new CancellationTokenSource();
+        var run = session.RunAsync(cts.Token);
+        await sink.WaitForAsync("session_started", TimeSpan.FromSeconds(2));
+
+        upstream.Push(new JsonObject
+        {
+            ["type"] = "conversation.item.input_audio_transcription.completed",
+            ["transcript"] = "hello",
+        });
+        upstream.Push(new JsonObject
+        {
+            ["type"] = "response.audio_transcript.done",
+            ["transcript"] = "hi there",
+        });
+        upstream.Push(new JsonObject { ["type"] = "response.done" });
+
+        await sink.WaitForAsync("assistant_done", TimeSpan.FromSeconds(2));
+
+        var convoId = session.ConversationId!.Value;
+        var turns = await db.Turns
+            .Where(t => t.ConversationId == convoId)
+            .OrderBy(t => t.Index).ToListAsync();
+        turns.Select(t => t.Role).Should().Equal(TurnRole.User, TurnRole.Assistant);
+        turns[0].Text.Should().Be("hello");
+        turns[1].Text.Should().Be("hi there");
+
+        cts.Cancel();
+        await run;
+    }
+}

backend.tests/ScriptedRealtimeUpstream.cs

@@ -0,0 +1,55 @@
+using System.Text.Json.Nodes;
+using System.Threading.Channels;
+using backend.Realtime;
+
+namespace backend.tests;
+
+public class ScriptedRealtimeUpstream : IRealtimeUpstream
+{
+    private readonly Channel<JsonObject> _toRelay =
+        Channel.CreateUnbounded<JsonObject>(new UnboundedChannelOptions { SingleReader = true });
+
+    private readonly Channel<JsonObject> _fromRelay =
+        Channel.CreateUnbounded<JsonObject>(new UnboundedChannelOptions { SingleWriter = true });
+
+    public IList<JsonObject> Sent { get; } = new List<JsonObject>();
+    public bool IsClosed => _toRelay.Reader.Completion.IsCompleted;
+    public string? CloseReason => IsClosed ? "scripted upstream closed" : null;
+
+    public Task SendJsonAsync(JsonObject envelope, CancellationToken ct)
+    {
+        Sent.Add(envelope);
+        return _fromRelay.Writer.WriteAsync(envelope, ct).AsTask();
+    }
+
+    public async Task<JsonObject?> ReceiveJsonAsync(CancellationToken ct)
+    {
+        try
+        {
+            return await _toRelay.Reader.ReadAsync(ct);
+        }
+        catch (ChannelClosedException)
+        {
+            return null;
+        }
+    }
+
+    public void Push(JsonObject evt) => _toRelay.Writer.TryWrite(evt);
+    public void CloseUpstream() => _toRelay.Writer.TryComplete();
+
+    public async Task<JsonObject> WaitForSentAsync(string type, CancellationToken ct)
+    {
+        await foreach (var item in _fromRelay.Reader.ReadAllAsync(ct))
+        {
+            if ((string?)item["type"] == type) return item;
+        }
+        throw new TimeoutException($"never saw outbound {type}");
+    }
+
+    public ValueTask DisposeAsync()
+    {
+        _toRelay.Writer.TryComplete();
+        _fromRelay.Writer.TryComplete();
+        return ValueTask.CompletedTask;
+    }
+}

backend.tests/SetVolumeToolTests.cs

@@ -0,0 +1,52 @@
+using System.Text.Json;
+using backend.Tools;
+using FluentAssertions;
+using Xunit;
+
+namespace backend.tests;
+
+public class SetVolumeToolTests
+{
+    [Fact]
+    public async Task Forwards_call_to_pi_and_returns_its_result()
+    {
+        var fake = new RecordingChannel(
+            JsonDocument.Parse("""{"ok":true}""").RootElement);
+        var tool = new SetVolumeTool();
+        var ctx = new DeviceContext(Guid.NewGuid(), Guid.NewGuid(), fake);
+
+        var args = JsonDocument.Parse("""{"level":42}""").RootElement;
+        var res = await tool.ExecuteAsync(new ToolInvocation("c1", args), ctx, default);
+
+        fake.LastName.Should().Be("set_volume");
+        fake.LastArgs.GetProperty("level").GetInt32().Should().Be(42);
+        res.Output.GetProperty("ok").GetBoolean().Should().BeTrue();
+    }
+
+    [Fact]
+    public async Task Rejects_out_of_range_level_without_calling_pi()
+    {
+        var fake = new RecordingChannel(JsonDocument.Parse("""{"ok":true}""").RootElement);
+        var tool = new SetVolumeTool();
+        var ctx = new DeviceContext(Guid.NewGuid(), Guid.NewGuid(), fake);
+
+        var args = JsonDocument.Parse("""{"level":150}""").RootElement;
+        var res = await tool.ExecuteAsync(new ToolInvocation("c1", args), ctx, default);
+
+        fake.LastName.Should().BeNull();
+        res.Output.GetProperty("ok").GetBoolean().Should().BeFalse();
+        res.Output.GetProperty("error").GetString().Should().Contain("0..100");
+    }
+
+    private class RecordingChannel(JsonElement reply) : IDeviceChannel
+    {
+        public string? LastName { get; private set; }
+        public JsonElement LastArgs { get; private set; }
+        public Task<JsonElement> CallPiToolAsync(string name, JsonElement args, CancellationToken ct)
+        {
+            LastName = name;
+            LastArgs = args;
+            return Task.FromResult(reply);
+        }
+    }
+}

backend.tests/ToolRegistryTests.cs

@@ -0,0 +1,29 @@
+using backend.Tools;
+using FluentAssertions;
+using Xunit;
+
+namespace backend.tests;
+
+public class ToolRegistryTests
+{
+    private readonly ITool[] _all = new ITool[]
+    {
+        new GetCurrentTimeTool(),
+    };
+
+    [Fact]
+    public void Get_returns_tool_by_exact_name()
+    {
+        var reg = new ToolRegistry(_all);
+        reg.Get("get_current_time").Should().NotBeNull();
+        reg.Get("does_not_exist").Should().BeNull();
+    }
+
+    [Fact]
+    public void Filter_returns_only_enabled_tools_preserving_order()
+    {
+        var reg = new ToolRegistry(_all);
+        var enabled = new HashSet<string> { "get_current_time", "ghost" };
+        reg.EnabledFor(enabled).Select(t => t.Name).Should().Equal("get_current_time");
+    }
+}

backend/Conversations/Conversation.cs

@@ -0,0 +1,19 @@
+namespace backend.Conversations;
+
+public enum EndReason
+{
+    Unknown = 0,
+    Tool = 1,
+    Idle = 2,
+    Error = 3,
+}
+
+public class Conversation
+{
+    public Guid Id { get; set; } = Guid.NewGuid();
+    public Guid DeviceId { get; set; }
+    public DateTimeOffset StartedAt { get; set; }
+    public DateTimeOffset? EndedAt { get; set; }
+    public EndReason EndReason { get; set; } = EndReason.Unknown;
+    public List<Turn> Turns { get; set; } = new();
+}

backend/Conversations/ConversationLog.cs

@@ -0,0 +1,63 @@
+using backend.Data;
+using Microsoft.EntityFrameworkCore;
+
+namespace backend.Conversations;
+
+public class ConversationLog(AppDbContext db)
+{
+    public async Task<Conversation> StartAsync(Guid deviceId, CancellationToken ct)
+    {
+        var c = new Conversation
+        {
+            DeviceId = deviceId,
+            StartedAt = DateTimeOffset.UtcNow,
+        };
+        db.Conversations.Add(c);
+        await db.SaveChangesAsync(ct);
+        return c;
+    }
+
+    public Task AppendUserAsync(Guid conversationId, string text, CancellationToken ct)
+        => AppendAsync(conversationId, TurnRole.User, text: text, ct: ct);
+
+    public Task AppendAssistantAsync(Guid conversationId, string text, CancellationToken ct)
+        => AppendAsync(conversationId, TurnRole.Assistant, text: text, ct: ct);
+
+    public Task AppendToolAsync(
+        Guid conversationId, string toolName, string argsJson, string resultJson, CancellationToken ct)
+        => AppendAsync(conversationId, TurnRole.Tool,
+            toolName: toolName, argsJson: argsJson, resultJson: resultJson, ct: ct);
+
+    public async Task EndAsync(Guid conversationId, EndReason reason, CancellationToken ct)
+    {
+        var row = await db.Conversations.FirstAsync(c => c.Id == conversationId, ct);
+        row.EndedAt = DateTimeOffset.UtcNow;
+        row.EndReason = reason;
+        await db.SaveChangesAsync(ct);
+    }
+
+    private async Task AppendAsync(
+        Guid conversationId, TurnRole role,
+        string? text = null, string? toolName = null,
+        string? argsJson = null, string? resultJson = null,
+        CancellationToken ct = default)
+    {
+        var next = await db.Turns
+            .Where(t => t.ConversationId == conversationId)
+            .Select(t => (int?)t.Index)
+            .MaxAsync(ct) ?? -1;
+
+        db.Turns.Add(new Turn
+        {
+            ConversationId = conversationId,
+            Index = next + 1,
+            Role = role,
+            Text = text,
+            ToolName = toolName,
+            ToolArgsJson = argsJson,
+            ToolResultJson = resultJson,
+            CreatedAt = DateTimeOffset.UtcNow,
+        });
+        await db.SaveChangesAsync(ct);
+    }
+}

backend/Conversations/Turn.cs

@@ -0,0 +1,21 @@
+namespace backend.Conversations;
+
+public enum TurnRole
+{
+    User = 0,
+    Assistant = 1,
+    Tool = 2,
+}
+
+public class Turn
+{
+    public Guid Id { get; set; } = Guid.NewGuid();
+    public Guid ConversationId { get; set; }
+    public int Index { get; set; }
+    public TurnRole Role { get; set; }
+    public string? Text { get; set; }
+    public string? ToolName { get; set; }
+    public string? ToolArgsJson { get; set; }
+    public string? ToolResultJson { get; set; }
+    public DateTimeOffset CreatedAt { get; set; }
+}

backend/Data/AppDbContext.cs

@@ -1,4 +1,5 @@
 using backend.Auth;
+using backend.Conversations;
 using backend.Devices;
 using backend.Pairing;
 using backend.Settings;
@@ -15,6 +16,8 @@ public class AppDbContext(DbContextOptions<AppDbContext> options)
     public DbSet<DeviceConfig> DeviceConfigs => Set<DeviceConfig>();
     public DbSet<SystemSettings> SystemSettings => Set<SystemSettings>();
     public DbSet<PairingCode> PairingCodes => Set<PairingCode>();
+    public DbSet<Conversation> Conversations => Set<Conversation>();
+    public DbSet<Turn> Turns => Set<Turn>();
 
     protected override void OnModelCreating(ModelBuilder b)
     {
@@ -36,5 +39,13 @@ public class AppDbContext(DbContextOptions<AppDbContext> options)
         b.Entity<SystemSettings>().Property(s => s.Id).ValueGeneratedNever();
         b.Entity<PairingCode>().HasKey(p => p.Code);
         b.Entity<PairingCode>().HasIndex(p => p.ExpiresAt);
+
+        b.Entity<Conversation>().HasIndex(c => c.DeviceId);
+        b.Entity<Conversation>()
+            .HasMany(c => c.Turns)
+            .WithOne()
+            .HasForeignKey(t => t.ConversationId)
+            .OnDelete(DeleteBehavior.Cascade);
+        b.Entity<Turn>().HasIndex(t => new { t.ConversationId, t.Index }).IsUnique();
     }
 }

backend/DeviceHub/ActiveDevice.cs

@@ -0,0 +1,112 @@
+using System.Collections.Concurrent;
+using System.Net.WebSockets;
+using System.Text;
+using System.Text.Json;
+using System.Text.Json.Nodes;
+using backend.Realtime;
+using backend.Tools;
+
+namespace backend.DeviceHub;
+
+public interface IDeviceClient
+{
+    Task SendEnvelopeAsync<T>(T envelope, CancellationToken ct) where T : HubEnvelope;
+    Task SendBinaryAsync(ReadOnlyMemory<byte> bytes, CancellationToken ct);
+}
+
+public class ActiveDevice : IDeviceChannel, IDeviceClient, ISessionOutput, IAsyncDisposable
+{
+    public Guid DeviceId { get; }
+    public WebSocket Ws => _ws;
+    public RealtimeSession? CurrentSession { get; set; }
+    public CancellationTokenSource? SessionCts { get; set; }
+
+    private readonly WebSocket _ws;
+    private readonly SemaphoreSlim _sendLock = new(1, 1);
+    private readonly ConcurrentDictionary<string, TaskCompletionSource<JsonElement>> _pendingTools = new();
+
+    public ActiveDevice(Guid deviceId, WebSocket ws)
+    {
+        DeviceId = deviceId;
+        _ws = ws;
+    }
+
+    public async Task SendEnvelopeAsync<T>(T envelope, CancellationToken ct) where T : HubEnvelope
+    {
+        var json = JsonSerializer.Serialize(envelope, envelope.GetType());
+        await _sendLock.WaitAsync(ct);
+        try
+        {
+            await _ws.SendAsync(Encoding.UTF8.GetBytes(json),
+                WebSocketMessageType.Text, endOfMessage: true, ct);
+        }
+        finally { _sendLock.Release(); }
+    }
+
+    public async Task SendBinaryAsync(ReadOnlyMemory<byte> bytes, CancellationToken ct)
+    {
+        await _sendLock.WaitAsync(ct);
+        try
+        {
+            await _ws.SendAsync(bytes, WebSocketMessageType.Binary, endOfMessage: true, ct);
+        }
+        finally { _sendLock.Release(); }
+    }
+
+    public Task WriteEnvelopeAsync(JsonObject envelope, CancellationToken ct)
+    {
+        var json = envelope.ToJsonString();
+        return WriteRawTextAsync(json, ct);
+    }
+
+    private async Task WriteRawTextAsync(string json, CancellationToken ct)
+    {
+        await _sendLock.WaitAsync(ct);
+        try
+        {
+            await _ws.SendAsync(Encoding.UTF8.GetBytes(json),
+                WebSocketMessageType.Text, endOfMessage: true, ct);
+        }
+        finally { _sendLock.Release(); }
+    }
+
+    public Task WriteBinaryAsync(ReadOnlyMemory<byte> bytes, CancellationToken ct)
+        => SendBinaryAsync(bytes, ct);
+
+    public Task<JsonElement> CallPiToolAsync(string name, JsonElement args, CancellationToken ct)
+    {
+        var callId = "srv_" + Guid.NewGuid().ToString("N")[..12];
+        var tcs = new TaskCompletionSource<JsonElement>(TaskCreationOptions.RunContinuationsAsynchronously);
+        _pendingTools[callId] = tcs;
+
+        var envelope = new ToolCallEnvelope("tool_call", callId, name, args);
+        _ = SendEnvelopeAsync(envelope, ct);
+
+        ct.Register(() => tcs.TrySetCanceled());
+        return tcs.Task;
+    }
+
+    public void CompleteToolResult(ToolResultEnvelope env)
+    {
+        if (!_pendingTools.TryRemove(env.CallId, out var tcs)) return;
+        if (env.Ok && env.Result is JsonElement r) tcs.TrySetResult(r);
+        else
+        {
+            var msg = (env.Error ?? "unknown").Replace("\\", "\\\\").Replace("\"", "\\\"");
+            var doc = JsonDocument.Parse($$"""{"ok":false,"error":"{{msg}}"}""");
+            tcs.TrySetResult(doc.RootElement.Clone());
+        }
+    }
+
+    public async ValueTask DisposeAsync()
+    {
+        try
+        {
+            if (_ws.State == WebSocketState.Open)
+                await _ws.CloseAsync(WebSocketCloseStatus.NormalClosure, "bye", CancellationToken.None);
+        }
+        catch { }
+        _ws.Dispose();
+        _sendLock.Dispose();
+    }
+}

backend/DeviceHub/DeviceAuth.cs

@@ -0,0 +1,21 @@
+using backend.Data;
+using backend.Devices;
+using Microsoft.EntityFrameworkCore;
+
+namespace backend.DeviceHub;
+
+public class DeviceAuth(AppDbContext db, DeviceTokenService tokens)
+{
+    public async Task<Device?> AuthenticateAsync(string? authHeader, CancellationToken ct)
+    {
+        if (string.IsNullOrEmpty(authHeader)) return null;
+        const string prefix = "Bearer ";
+        if (!authHeader.StartsWith(prefix, StringComparison.Ordinal)) return null;
+        var token = authHeader[prefix.Length..].Trim();
+        if (string.IsNullOrEmpty(token)) return null;
+        var hash = tokens.Hash(token);
+        var device = await db.Devices.Include(d => d.Config)
+            .FirstOrDefaultAsync(d => d.TokenHash == hash && !d.IsRevoked, ct);
+        return device;
+    }
+}

backend/DeviceHub/DeviceHubEndpoint.cs

@@ -0,0 +1,207 @@
+using System.Net.WebSockets;
+using System.Text;
+using System.Text.Json;
+using System.Text.Json.Nodes;
+using backend.Conversations;
+using backend.Data;
+using backend.Devices;
+using backend.Realtime;
+using backend.Tools;
+using Microsoft.EntityFrameworkCore;
+using Microsoft.Extensions.DependencyInjection;
+
+namespace backend.DeviceHub;
+
+public static class DeviceHubEndpoint
+{
+    public static IEndpointRouteBuilder MapDeviceHub(this IEndpointRouteBuilder app)
+    {
+        app.Map("/device", async (
+            HttpContext ctx,
+            DeviceAuth auth,
+            DeviceRegistry reg,
+            AppDbContext db,
+            RealtimeSessionFactory sessions,
+            IServiceScopeFactory scopes) =>
+        {
+            if (!ctx.WebSockets.IsWebSocketRequest)
+            {
+                ctx.Response.StatusCode = StatusCodes.Status400BadRequest;
+                return;
+            }
+
+            var device = await auth.AuthenticateAsync(
+                ctx.Request.Headers.Authorization.ToString(), ctx.RequestAborted);
+            if (device is null)
+            {
+                ctx.Response.StatusCode = StatusCodes.Status401Unauthorized;
+                return;
+            }
+
+            using var ws = await ctx.WebSockets.AcceptWebSocketAsync();
+            await using var active = new ActiveDevice(device.Id, ws);
+            if (!reg.TryRegister(active))
+            {
+                await ws.CloseAsync(WebSocketCloseStatus.PolicyViolation,
+                    "device already connected", ctx.RequestAborted);
+                return;
+            }
+
+            try
+            {
+                await HandleAsync(active, device, db, sessions, scopes, ctx.RequestAborted);
+            }
+            finally
+            {
+                reg.Unregister(device.Id);
+            }
+        });
+        return app;
+    }
+
+    private static async Task HandleAsync(
+        ActiveDevice active, Device device, AppDbContext db,
+        RealtimeSessionFactory sessions, IServiceScopeFactory scopes,
+        CancellationToken ct)
+    {
+        var buffer = new byte[8192];
+        var ws = active.Ws;
+
+        while (ws.State == WebSocketState.Open && !ct.IsCancellationRequested)
+        {
+            using var msg = new MemoryStream();
+            WebSocketReceiveResult res;
+            while (true)
+            {
+                res = await ws.ReceiveAsync(buffer, ct);
+                if (res.MessageType == WebSocketMessageType.Close) return;
+                msg.Write(buffer, 0, res.Count);
+                if (res.EndOfMessage) break;
+            }
+
+            if (res.MessageType == WebSocketMessageType.Binary)
+            {
+                if (msg.Length == 3840 && active.CurrentSession is not null)
+                {
+                    await active.CurrentSession.WriteUplinkAsync(msg.ToArray(), ct);
+                }
+                continue;
+            }
+
+            var text = Encoding.UTF8.GetString(msg.ToArray());
+            using var doc = JsonDocument.Parse(text);
+            var type = doc.RootElement.GetProperty("type").GetString();
+            switch (type)
+            {
+                case "hello":
+                    await SendHelloAckAsync(active, device, ct);
+                    break;
+
+                case "ping":
+                    device.LastSeenAt = DateTimeOffset.UtcNow;
+                    await db.SaveChangesAsync(ct);
+                    await active.SendEnvelopeAsync(new PongEnvelope("pong"), ct);
+                    break;
+
+                case "wake":
+                    if (active.CurrentSession is null)
+                        await StartSessionAsync(active, device, sessions, scopes, ct);
+                    break;
+
+                case "cancel":
+                    active.SessionCts?.Cancel();
+                    break;
+
+                case "playback_done":
+                    break; // v1: client coordination only
+
+                case "tool_result":
+                    var env = JsonSerializer.Deserialize<ToolResultEnvelope>(text)!;
+                    active.CompleteToolResult(env);
+                    break;
+            }
+        }
+    }
+
+    private static async Task StartSessionAsync(
+        ActiveDevice active, Device device,
+        RealtimeSessionFactory sessions, IServiceScopeFactory scopes,
+        CancellationToken parentCt)
+    {
+        var enabledTools = JsonSerializer.Deserialize<string[]>(device.Config?.EnabledToolsJson ?? "[]")
+            ?? Array.Empty<string>();
+        // Promote any deprecated Beta-era model name to the GA name. Plan 4's
+        // admin UI will let users set a specific value; until then, anything
+        // that mentions the old preview slug routes to gpt-realtime.
+        var model = device.Config?.Model ?? "gpt-realtime";
+        if (model.StartsWith("gpt-4o-realtime-preview", StringComparison.Ordinal))
+            model = "gpt-realtime";
+        var cfg = new RealtimeSettings(
+            Model: model,
+            Voice: device.Config?.Voice ?? "alloy",
+            SystemPrompt: device.Config?.SystemPrompt ?? "",
+            IdleTimeoutSeconds: device.Config?.IdleTimeoutSeconds ?? 30,
+            EnabledTools: new HashSet<string>(enabledTools));
+
+        active.SessionCts = CancellationTokenSource.CreateLinkedTokenSource(parentCt);
+        var ct = active.SessionCts.Token;
+
+        var sessionScope = scopes.CreateAsyncScope();
+        var scopedSessions = sessionScope.ServiceProvider.GetRequiredService<RealtimeSessionFactory>();
+
+        RealtimeSession session;
+        try
+        {
+            session = await scopedSessions.CreateAsync(
+                device.Id, cfg, active, active, ct);
+        }
+        catch (Exception ex)
+        {
+            await active.SendEnvelopeAsync(
+                new ErrorEnvelope("error", "upstream", ex.Message, true), parentCt);
+            await sessionScope.DisposeAsync();
+            return;
+        }
+
+        active.CurrentSession = session;
+        _ = Task.Run(async () =>
+        {
+            try { await session.RunAsync(ct); }
+            finally
+            {
+                active.CurrentSession = null;
+                active.SessionCts?.Dispose();
+                active.SessionCts = null;
+                await sessionScope.DisposeAsync();
+            }
+        }, ct);
+    }
+
+    // Bump on every backend change so we can verify a Coolify deploy landed
+    // by looking at the hello_ack config.server_version field in the Pi journal.
+    public const string ServerVersion = "plan3-debug-2026-06-12-09-closeprobe";
+
+    private static async Task SendHelloAckAsync(ActiveDevice active, Device device, CancellationToken ct)
+    {
+        var enabled = device.Config?.EnabledToolsJson ?? "[]";
+        var config = JsonNode.Parse($$"""
+        {
+          "voice": "{{Escape(device.Config?.Voice)}}",
+          "model": "{{Escape(device.Config?.Model)}}",
+          "system_prompt": "{{Escape(device.Config?.SystemPrompt)}}",
+          "idle_timeout_seconds": {{device.Config?.IdleTimeoutSeconds ?? 30}},
+          "enabled_tools": {{enabled}},
+          "server_version": "{{ServerVersion}}"
+        }
+        """)!.AsObject();
+
+        await active.WriteEnvelopeAsync(new JsonObject
+        {
+            ["type"] = "hello_ack",
+            ["config"] = config,
+        }, ct);
+    }
+
+    private static string Escape(string? s) =>
+        (s ?? "").Replace("\\", "\\\\").Replace("\"", "\\\"");
+}

backend/DeviceHub/DeviceRegistry.cs

@@ -0,0 +1,21 @@
+using System.Collections.Concurrent;
+
+namespace backend.DeviceHub;
+
+public class DeviceRegistry
+{
+    private readonly ConcurrentDictionary<Guid, ActiveDevice> _online = new();
+
+    public bool TryRegister(ActiveDevice dev)
+        => _online.TryAdd(dev.DeviceId, dev);
+
+    public bool Unregister(Guid deviceId)
+        => _online.TryRemove(deviceId, out _);
+
+    public ActiveDevice? Get(Guid deviceId)
+        => _online.TryGetValue(deviceId, out var d) ? d : null;
+
+    public bool IsOnline(Guid deviceId) => _online.ContainsKey(deviceId);
+
+    public int OnlineCount => _online.Count;
+}

backend/DeviceHub/HubEnvelopes.cs

@@ -0,0 +1,42 @@
+using System.Text.Json;
+using System.Text.Json.Serialization;
+
+namespace backend.DeviceHub;
+
+public record HubEnvelope([property: JsonPropertyName("type")] string Type);
+
+public record HelloEnvelope(string Type,
+    [property: JsonPropertyName("device_id")] Guid DeviceId,
+    [property: JsonPropertyName("client_version")] string ClientVersion) : HubEnvelope(Type);
+
+public record HelloAckEnvelope(string Type,
+    [property: JsonPropertyName("config")] JsonElement Config) : HubEnvelope(Type);
+
+public record PongEnvelope(string Type) : HubEnvelope(Type);
+
+public record WakeEnvelope(string Type,
+    [property: JsonPropertyName("at")] long? At) : HubEnvelope(Type);
+
+public record SessionStartedEnvelope(string Type,
+    [property: JsonPropertyName("conversation_id")] Guid ConversationId) : HubEnvelope(Type);
+
+public record SessionEndedEnvelope(string Type,
+    [property: JsonPropertyName("reason")] string Reason) : HubEnvelope(Type);
+
+public record AssistantDoneEnvelope(string Type) : HubEnvelope(Type);
+
+public record ToolCallEnvelope(string Type,
+    [property: JsonPropertyName("call_id")] string CallId,
+    [property: JsonPropertyName("name")] string Name,
+    [property: JsonPropertyName("arguments")] JsonElement Arguments) : HubEnvelope(Type);
+
+public record ToolResultEnvelope(string Type,
+    [property: JsonPropertyName("call_id")] string CallId,
+    [property: JsonPropertyName("ok")] bool Ok,
+    [property: JsonPropertyName("result")] JsonElement? Result,
+    [property: JsonPropertyName("error")] string? Error) : HubEnvelope(Type);
+
+public record ErrorEnvelope(string Type,
+    [property: JsonPropertyName("code")] string Code,
+    [property: JsonPropertyName("message")] string Message,
+    [property: JsonPropertyName("fatal")] bool? Fatal) : HubEnvelope(Type);

backend/Migrations/20260611210852_AddConversationsAndTurns.Designer.cs

@@ -0,0 +1,482 @@
+// <auto-generated />
+using System;
+using Microsoft.EntityFrameworkCore;
+using Microsoft.EntityFrameworkCore.Infrastructure;
+using Microsoft.EntityFrameworkCore.Migrations;
+using Microsoft.EntityFrameworkCore.Storage.ValueConversion;
+using backend.Data;
+
+#nullable disable
+
+namespace backend.Migrations
+{
+    [DbContext(typeof(AppDbContext))]
+    [Migration("20260611210852_AddConversationsAndTurns")]
+    partial class AddConversationsAndTurns
+    {
+        /// <inheritdoc />
+        protected override void BuildTargetModel(ModelBuilder modelBuilder)
+        {
+#pragma warning disable 612, 618
+            modelBuilder.HasAnnotation("ProductVersion", "9.0.17");
+
+            modelBuilder.Entity("Microsoft.AspNetCore.Identity.IdentityRole<System.Guid>", b =>
+                {
+                    b.Property<Guid>("Id")
+                        .ValueGeneratedOnAdd()
+                        .HasColumnType("TEXT");
+
+                    b.Property<string>("ConcurrencyStamp")
+                        .IsConcurrencyToken()
+                        .HasColumnType("TEXT");
+
+                    b.Property<string>("Name")
+                        .HasMaxLength(256)
+                        .HasColumnType("TEXT");
+
+                    b.Property<string>("NormalizedName")
+                        .HasMaxLength(256)
+                        .HasColumnType("TEXT");
+
+                    b.HasKey("Id");
+
+                    b.HasIndex("NormalizedName")
+                        .IsUnique()
+                        .HasDatabaseName("RoleNameIndex");
+
+                    b.ToTable("AspNetRoles", (string)null);
+                });
+
+            modelBuilder.Entity("Microsoft.AspNetCore.Identity.IdentityRoleClaim<System.Guid>", b =>
+                {
+                    b.Property<int>("Id")
+                        .ValueGeneratedOnAdd()
+                        .HasColumnType("INTEGER");
+
+                    b.Property<string>("ClaimType")
+                        .HasColumnType("TEXT");
+
+                    b.Property<string>("ClaimValue")
+                        .HasColumnType("TEXT");
+
+                    b.Property<Guid>("RoleId")
+                        .HasColumnType("TEXT");
+
+                    b.HasKey("Id");
+
+                    b.HasIndex("RoleId");
+
+                    b.ToTable("AspNetRoleClaims", (string)null);
+                });
+
+            modelBuilder.Entity("Microsoft.AspNetCore.Identity.IdentityUserClaim<System.Guid>", b =>
+                {
+                    b.Property<int>("Id")
+                        .ValueGeneratedOnAdd()
+                        .HasColumnType("INTEGER");
+
+                    b.Property<string>("ClaimType")
+                        .HasColumnType("TEXT");
+
+                    b.Property<string>("ClaimValue")
+                        .HasColumnType("TEXT");
+
+                    b.Property<Guid>("UserId")
+                        .HasColumnType("TEXT");
+
+                    b.HasKey("Id");
+
+                    b.HasIndex("UserId");
+
+                    b.ToTable("AspNetUserClaims", (string)null);
+                });
+
+            modelBuilder.Entity("Microsoft.AspNetCore.Identity.IdentityUserLogin<System.Guid>", b =>
+                {
+                    b.Property<string>("LoginProvider")
+                        .HasColumnType("TEXT");
+
+                    b.Property<string>("ProviderKey")
+                        .HasColumnType("TEXT");
+
+                    b.Property<string>("ProviderDisplayName")
+                        .HasColumnType("TEXT");
+
+                    b.Property<Guid>("UserId")
+                        .HasColumnType("TEXT");
+
+                    b.HasKey("LoginProvider", "ProviderKey");
+
+                    b.HasIndex("UserId");
+
+                    b.ToTable("AspNetUserLogins", (string)null);
+                });
+
+            modelBuilder.Entity("Microsoft.AspNetCore.Identity.IdentityUserRole<System.Guid>", b =>
+                {
+                    b.Property<Guid>("UserId")
+                        .HasColumnType("TEXT");
+
+                    b.Property<Guid>("RoleId")
+                        .HasColumnType("TEXT");
+
+                    b.HasKey("UserId", "RoleId");
+
+                    b.HasIndex("RoleId");
+
+                    b.ToTable("AspNetUserRoles", (string)null);
+                });
+
+            modelBuilder.Entity("Microsoft.AspNetCore.Identity.IdentityUserToken<System.Guid>", b =>
+                {
+                    b.Property<Guid>("UserId")
+                        .HasColumnType("TEXT");
+
+                    b.Property<string>("LoginProvider")
+                        .HasColumnType("TEXT");
+
+                    b.Property<string>("Name")
+                        .HasColumnType("TEXT");
+
+                    b.Property<string>("Value")
+                        .HasColumnType("TEXT");
+
+                    b.HasKey("UserId", "LoginProvider", "Name");
+
+                    b.ToTable("AspNetUserTokens", (string)null);
+                });
+
+            modelBuilder.Entity("backend.Auth.ApplicationUser", b =>
+                {
+                    b.Property<Guid>("Id")
+                        .ValueGeneratedOnAdd()
+                        .HasColumnType("TEXT");
+
+                    b.Property<int>("AccessFailedCount")
+                        .HasColumnType("INTEGER");
+
+                    b.Property<string>("ConcurrencyStamp")
+                        .IsConcurrencyToken()
+                        .HasColumnType("TEXT");
+
+                    b.Property<string>("Email")
+                        .HasMaxLength(256)
+                        .HasColumnType("TEXT");
+
+                    b.Property<bool>("EmailConfirmed")
+                        .HasColumnType("INTEGER");
+
+                    b.Property<bool>("LockoutEnabled")
+                        .HasColumnType("INTEGER");
+
+                    b.Property<DateTimeOffset?>("LockoutEnd")
+                        .HasColumnType("TEXT");
+
+                    b.Property<string>("NormalizedEmail")
+                        .HasMaxLength(256)
+                        .HasColumnType("TEXT");
+
+                    b.Property<string>("NormalizedUserName")
+                        .HasMaxLength(256)
+                        .HasColumnType("TEXT");
+
+                    b.Property<string>("PasswordHash")
+                        .HasColumnType("TEXT");
+
+                    b.Property<string>("PhoneNumber")
+                        .HasColumnType("TEXT");
+
+                    b.Property<bool>("PhoneNumberConfirmed")
+                        .HasColumnType("INTEGER");
+
+                    b.Property<string>("SecurityStamp")
+                        .HasColumnType("TEXT");
+
+                    b.Property<bool>("TwoFactorEnabled")
+                        .HasColumnType("INTEGER");
+
+                    b.Property<string>("UserName")
+                        .HasMaxLength(256)
+                        .HasColumnType("TEXT");
+
+                    b.HasKey("Id");
+
+                    b.HasIndex("NormalizedEmail")
+                        .HasDatabaseName("EmailIndex");
+
+                    b.HasIndex("NormalizedUserName")
+                        .IsUnique()
+                        .HasDatabaseName("UserNameIndex");
+
+                    b.ToTable("AspNetUsers", (string)null);
+                });
+
+            modelBuilder.Entity("backend.Conversations.Conversation", b =>
+                {
+                    b.Property<Guid>("Id")
+                        .ValueGeneratedOnAdd()
+                        .HasColumnType("TEXT");
+
+                    b.Property<Guid>("DeviceId")
+                        .HasColumnType("TEXT");
+
+                    b.Property<int>("EndReason")
+                        .HasColumnType("INTEGER");
+
+                    b.Property<DateTimeOffset?>("EndedAt")
+                        .HasColumnType("TEXT");
+
+                    b.Property<DateTimeOffset>("StartedAt")
+                        .HasColumnType("TEXT");
+
+                    b.HasKey("Id");
+
+                    b.HasIndex("DeviceId");
+
+                    b.ToTable("Conversations");
+                });
+
+            modelBuilder.Entity("backend.Conversations.Turn", b =>
+                {
+                    b.Property<Guid>("Id")
+                        .ValueGeneratedOnAdd()
+                        .HasColumnType("TEXT");
+
+                    b.Property<Guid>("ConversationId")
+                        .HasColumnType("TEXT");
+
+                    b.Property<DateTimeOffset>("CreatedAt")
+                        .HasColumnType("TEXT");
+
+                    b.Property<int>("Index")
+                        .HasColumnType("INTEGER");
+
+                    b.Property<int>("Role")
+                        .HasColumnType("INTEGER");
+
+                    b.Property<string>("Text")
+                        .HasColumnType("TEXT");
+
+                    b.Property<string>("ToolArgsJson")
+                        .HasColumnType("TEXT");
+
+                    b.Property<string>("ToolName")
+                        .HasColumnType("TEXT");
+
+                    b.Property<string>("ToolResultJson")
+                        .HasColumnType("TEXT");
+
+                    b.HasKey("Id");
+
+                    b.HasIndex("ConversationId", "Index")
+                        .IsUnique();
+
+                    b.ToTable("Turns");
+                });
+
+            modelBuilder.Entity("backend.Devices.Device", b =>
+                {
+                    b.Property<Guid>("Id")
+                        .ValueGeneratedOnAdd()
+                        .HasColumnType("TEXT");
+
+                    b.Property<bool>("IsRevoked")
+                        .HasColumnType("INTEGER");
+
+                    b.Property<DateTimeOffset?>("LastSeenAt")
+                        .HasColumnType("TEXT");
+
+                    b.Property<string>("Name")
+                        .IsRequired()
+                        .HasColumnType("TEXT");
+
+                    b.Property<Guid>("OwnerUserId")
+                        .HasColumnType("TEXT");
+
+                    b.Property<DateTimeOffset>("PairedAt")
+                        .HasColumnType("TEXT");
+
+                    b.Property<string>("TokenHash")
+                        .IsRequired()
+                        .HasColumnType("TEXT");
+
+                    b.HasKey("Id");
+
+                    b.HasIndex("OwnerUserId");
+
+                    b.HasIndex("TokenHash")
+                        .IsUnique();
+
+                    b.ToTable("Devices");
+                });
+
+            modelBuilder.Entity("backend.Devices.DeviceConfig", b =>
+                {
+                    b.Property<Guid>("DeviceId")
+                        .HasColumnType("TEXT");
+
+                    b.Property<string>("EnabledToolsJson")
+                        .IsRequired()
+                        .HasColumnType("TEXT");
+
+                    b.Property<int>("IdleTimeoutSeconds")
+                        .HasColumnType("INTEGER");
+
+                    b.Property<string>("Model")
+                        .IsRequired()
+                        .HasColumnType("TEXT");
+
+                    b.Property<string>("SystemPrompt")
+                        .IsRequired()
+                        .HasColumnType("TEXT");
+
+                    b.Property<string>("Voice")
+                        .IsRequired()
+                        .HasColumnType("TEXT");
+
+                    b.HasKey("DeviceId");
+
+                    b.ToTable("DeviceConfigs");
+                });
+
+            modelBuilder.Entity("backend.Pairing.PairingCode", b =>
+                {
+                    b.Property<string>("Code")
+                        .HasColumnType("TEXT");
+
+                    b.Property<DateTimeOffset?>("ConsumedAt")
+                        .HasColumnType("TEXT");
+
+                    b.Property<string>("DeviceName")
+                        .IsRequired()
+                        .HasColumnType("TEXT");
+
+                    b.Property<DateTimeOffset>("ExpiresAt")
+                        .HasColumnType("TEXT");
+
+                    b.Property<Guid>("OwnerUserId")
+                        .HasColumnType("TEXT");
+
+                    b.HasKey("Code");
+
+                    b.HasIndex("ExpiresAt");
+
+                    b.ToTable("PairingCodes");
+                });
+
+            modelBuilder.Entity("backend.Settings.SystemSettings", b =>
+                {
+                    b.Property<int>("Id")
+                        .HasColumnType("INTEGER");
+
+                    b.Property<int>("DefaultIdleTimeoutSeconds")
+                        .HasColumnType("INTEGER");
+
+                    b.Property<string>("DefaultModel")
+                        .IsRequired()
+                        .HasColumnType("TEXT");
+
+                    b.Property<string>("DefaultSystemPrompt")
+                        .IsRequired()
+                        .HasColumnType("TEXT");
+
+                    b.Property<string>("DefaultVoice")
+                        .IsRequired()
+                        .HasColumnType("TEXT");
+
+                    b.HasKey("Id");
+
+                    b.ToTable("SystemSettings");
+                });
+
+            modelBuilder.Entity("Microsoft.AspNetCore.Identity.IdentityRoleClaim<System.Guid>", b =>
+                {
+                    b.HasOne("Microsoft.AspNetCore.Identity.IdentityRole<System.Guid>", null)
+                        .WithMany()
+                        .HasForeignKey("RoleId")
+                        .OnDelete(DeleteBehavior.Cascade)
+                        .IsRequired();
+                });
+
+            modelBuilder.Entity("Microsoft.AspNetCore.Identity.IdentityUserClaim<System.Guid>", b =>
+                {
+                    b.HasOne("backend.Auth.ApplicationUser", null)
+                        .WithMany()
+                        .HasForeignKey("UserId")
+                        .OnDelete(DeleteBehavior.Cascade)
+                        .IsRequired();
+                });
+
+            modelBuilder.Entity("Microsoft.AspNetCore.Identity.IdentityUserLogin<System.Guid>", b =>
+                {
+                    b.HasOne("backend.Auth.ApplicationUser", null)
+                        .WithMany()
+                        .HasForeignKey("UserId")
+                        .OnDelete(DeleteBehavior.Cascade)
+                        .IsRequired();
+                });
+
+            modelBuilder.Entity("Microsoft.AspNetCore.Identity.IdentityUserRole<System.Guid>", b =>
+                {
+                    b.HasOne("Microsoft.AspNetCore.Identity.IdentityRole<System.Guid>", null)
+                        .WithMany()
+                        .HasForeignKey("RoleId")
+                        .OnDelete(DeleteBehavior.Cascade)
+                        .IsRequired();
+
+                    b.HasOne("backend.Auth.ApplicationUser", null)
+                        .WithMany()
+                        .HasForeignKey("UserId")
+                        .OnDelete(DeleteBehavior.Cascade)
+                        .IsRequired();
+                });
+
+            modelBuilder.Entity("Microsoft.AspNetCore.Identity.IdentityUserToken<System.Guid>", b =>
+                {
+                    b.HasOne("backend.Auth.ApplicationUser", null)
+                        .WithMany()
+                        .HasForeignKey("UserId")
+                        .OnDelete(DeleteBehavior.Cascade)
+                        .IsRequired();
+                });
+
+            modelBuilder.Entity("backend.Conversations.Turn", b =>
+                {
+                    b.HasOne("backend.Conversations.Conversation", null)
+                        .WithMany("Turns")
+                        .HasForeignKey("ConversationId")
+                        .OnDelete(DeleteBehavior.Cascade)
+                        .IsRequired();
+                });
+
+            modelBuilder.Entity("backend.Devices.Device", b =>
+                {
+                    b.HasOne("backend.Auth.ApplicationUser", null)
+                        .WithMany()
+                        .HasForeignKey("OwnerUserId")
+                        .OnDelete(DeleteBehavior.Cascade)
+                        .IsRequired();
+                });
+
+            modelBuilder.Entity("backend.Devices.DeviceConfig", b =>
+                {
+                    b.HasOne("backend.Devices.Device", null)
+                        .WithOne("Config")
+                        .HasForeignKey("backend.Devices.DeviceConfig", "DeviceId")
+                        .OnDelete(DeleteBehavior.Cascade)
+                        .IsRequired();
+                });
+
+            modelBuilder.Entity("backend.Conversations.Conversation", b =>
+                {
+                    b.Navigation("Turns");
+                });
+
+            modelBuilder.Entity("backend.Devices.Device", b =>
+                {
+                    b.Navigation("Config");
+                });
+#pragma warning restore 612, 618
+        }
+    }
+}

backend/Migrations/20260611210852_AddConversationsAndTurns.cs

@@ -0,0 +1,76 @@
+using System;
+using Microsoft.EntityFrameworkCore.Migrations;
+
+#nullable disable
+
+namespace backend.Migrations
+{
+    /// <inheritdoc />
+    public partial class AddConversationsAndTurns : Migration
+    {
+        /// <inheritdoc />
+        protected override void Up(MigrationBuilder migrationBuilder)
+        {
+            migrationBuilder.CreateTable(
+                name: "Conversations",
+                columns: table => new
+                {
+                    Id = table.Column<Guid>(type: "TEXT", nullable: false),
+                    DeviceId = table.Column<Guid>(type: "TEXT", nullable: false),
+                    StartedAt = table.Column<DateTimeOffset>(type: "TEXT", nullable: false),
+                    EndedAt = table.Column<DateTimeOffset>(type: "TEXT", nullable: true),
+                    EndReason = table.Column<int>(type: "INTEGER", nullable: false)
+                },
+                constraints: table =>
+                {
+                    table.PrimaryKey("PK_Conversations", x => x.Id);
+                });
+
+            migrationBuilder.CreateTable(
+                name: "Turns",
+                columns: table => new
+                {
+                    Id = table.Column<Guid>(type: "TEXT", nullable: false),
+                    ConversationId = table.Column<Guid>(type: "TEXT", nullable: false),
+                    Index = table.Column<int>(type: "INTEGER", nullable: false),
+                    Role = table.Column<int>(type: "INTEGER", nullable: false),
+                    Text = table.Column<string>(type: "TEXT", nullable: true),
+                    ToolName = table.Column<string>(type: "TEXT", nullable: true),
+                    ToolArgsJson = table.Column<string>(type: "TEXT", nullable: true),
+                    ToolResultJson = table.Column<string>(type: "TEXT", nullable: true),
+                    CreatedAt = table.Column<DateTimeOffset>(type: "TEXT", nullable: false)
+                },
+                constraints: table =>
+                {
+                    table.PrimaryKey("PK_Turns", x => x.Id);
+                    table.ForeignKey(
+                        name: "FK_Turns_Conversations_ConversationId",
+                        column: x => x.ConversationId,
+                        principalTable: "Conversations",
+                        principalColumn: "Id",
+                        onDelete: ReferentialAction.Cascade);
+                });
+
+            migrationBuilder.CreateIndex(
+                name: "IX_Conversations_DeviceId",
+                table: "Conversations",
+                column: "DeviceId");
+
+            migrationBuilder.CreateIndex(
+                name: "IX_Turns_ConversationId_Index",
+                table: "Turns",
+                columns: new[] { "ConversationId", "Index" },
+                unique: true);
+        }
+
+        /// <inheritdoc />
+        protected override void Down(MigrationBuilder migrationBuilder)
+        {
+            migrationBuilder.DropTable(
+                name: "Turns");
+
+            migrationBuilder.DropTable(
+                name: "Conversations");
+        }
+    }
+}

backend/Migrations/AppDbContextModelSnapshot.cs

@@ -208,6 +208,69 @@ namespace backend.Migrations
                     b.ToTable("AspNetUsers", (string)null);
                 });
 
+            modelBuilder.Entity("backend.Conversations.Conversation", b =>
+                {
+                    b.Property<Guid>("Id")
+                        .ValueGeneratedOnAdd()
+                        .HasColumnType("TEXT");
+
+                    b.Property<Guid>("DeviceId")
+                        .HasColumnType("TEXT");
+
+                    b.Property<int>("EndReason")
+                        .HasColumnType("INTEGER");
+
+                    b.Property<DateTimeOffset?>("EndedAt")
+                        .HasColumnType("TEXT");
+
+                    b.Property<DateTimeOffset>("StartedAt")
+                        .HasColumnType("TEXT");
+
+                    b.HasKey("Id");
+
+                    b.HasIndex("DeviceId");
+
+                    b.ToTable("Conversations");
+                });
+
+            modelBuilder.Entity("backend.Conversations.Turn", b =>
+                {
+                    b.Property<Guid>("Id")
+                        .ValueGeneratedOnAdd()
+                        .HasColumnType("TEXT");
+
+                    b.Property<Guid>("ConversationId")
+                        .HasColumnType("TEXT");
+
+                    b.Property<DateTimeOffset>("CreatedAt")
+                        .HasColumnType("TEXT");
+
+                    b.Property<int>("Index")
+                        .HasColumnType("INTEGER");
+
+                    b.Property<int>("Role")
+                        .HasColumnType("INTEGER");
+
+                    b.Property<string>("Text")
+                        .HasColumnType("TEXT");
+
+                    b.Property<string>("ToolArgsJson")
+                        .HasColumnType("TEXT");
+
+                    b.Property<string>("ToolName")
+                        .HasColumnType("TEXT");
+
+                    b.Property<string>("ToolResultJson")
+                        .HasColumnType("TEXT");
+
+                    b.HasKey("Id");
+
+                    b.HasIndex("ConversationId", "Index")
+                        .IsUnique();
+
+                    b.ToTable("Turns");
+                });
+
             modelBuilder.Entity("backend.Devices.Device", b =>
                 {
                     b.Property<Guid>("Id")
@@ -374,6 +437,15 @@ namespace backend.Migrations
                         .IsRequired();
                 });
 
+            modelBuilder.Entity("backend.Conversations.Turn", b =>
+                {
+                    b.HasOne("backend.Conversations.Conversation", null)
+                        .WithMany("Turns")
+                        .HasForeignKey("ConversationId")
+                        .OnDelete(DeleteBehavior.Cascade)
+                        .IsRequired();
+                });
+
             modelBuilder.Entity("backend.Devices.Device", b =>
                 {
                     b.HasOne("backend.Auth.ApplicationUser", null)
@@ -392,6 +464,11 @@ namespace backend.Migrations
                         .IsRequired();
                 });
 
+            modelBuilder.Entity("backend.Conversations.Conversation", b =>
+                {
+                    b.Navigation("Turns");
+                });
+
             modelBuilder.Entity("backend.Devices.Device", b =>
                 {
                     b.Navigation("Config");

backend/Program.cs

@@ -1,5 +1,6 @@
 using backend.Auth;
 using backend.Data;
+using backend.DeviceHub;
 using backend.Devices;
 using backend.Install;
 using backend.Pairing;
@@ -33,6 +34,15 @@ builder.Services.AddAuthorization();
 builder.Services.AddScoped<backend.Devices.DeviceTokenService>();
 builder.Services.AddScoped<backend.Pairing.PairingService>();
 builder.Services.AddSingleton<backend.Install.InstallScriptBuilder>();
+builder.Services.AddScoped<backend.Conversations.ConversationLog>();
+builder.Services.AddSingleton<backend.Realtime.OpenAIKeyProvider>();
+builder.Services.AddSingleton<backend.Tools.ITool, backend.Tools.GetCurrentTimeTool>();
+builder.Services.AddSingleton<backend.Tools.ITool, backend.Tools.EndSessionTool>();
+builder.Services.AddSingleton<backend.Tools.ITool, backend.Tools.SetVolumeTool>();
+builder.Services.AddSingleton<backend.Tools.ToolRegistry>();
+builder.Services.AddScoped<backend.Realtime.RealtimeSessionFactory>();
+builder.Services.AddSingleton<backend.DeviceHub.DeviceRegistry>();
+builder.Services.AddScoped<backend.DeviceHub.DeviceAuth>();
 
 var app = builder.Build();
 
@@ -61,6 +71,9 @@ app.MapInstallEndpoints();
 
 app.MapGet("/health", () => Results.Ok(new { ok = true }));
 
+app.UseWebSockets();
+app.MapDeviceHub();
+
 app.Run();
 
 public partial class Program { }

backend/Realtime/IRealtimeUpstream.cs

@@ -0,0 +1,13 @@
+using System.Text.Json.Nodes;
+
+namespace backend.Realtime;
+
+public interface IRealtimeUpstream : IAsyncDisposable
+{
+    Task SendJsonAsync(JsonObject envelope, CancellationToken ct);
+    Task<JsonObject?> ReceiveJsonAsync(CancellationToken ct);
+    // Once a receive returns null because the WebSocket transitioned out of
+    // the Open state, these surface why. Both stay null until that happens.
+    bool IsClosed { get; }
+    string? CloseReason { get; }
+}

backend/Realtime/OpenAIKeyProvider.cs

@@ -0,0 +1,21 @@
+namespace backend.Realtime;
+
+public class OpenAIKeyProvider
+{
+    private readonly string? _key;
+
+    public OpenAIKeyProvider(IConfiguration cfg)
+    {
+        _key = cfg["OPENAI_API_KEY"] ?? cfg["OpenAI:ApiKey"];
+    }
+
+    public bool HasKey => !string.IsNullOrWhiteSpace(_key);
+
+    public string GetRequired()
+    {
+        if (string.IsNullOrWhiteSpace(_key))
+            throw new InvalidOperationException(
+                "OPENAI_API_KEY is not configured. Set the env var on the Coolify app.");
+        return _key!;
+    }
+}

backend/Realtime/OpenAIRealtimeUpstream.cs

@@ -0,0 +1,77 @@
+using System.Net.WebSockets;
+using System.Text;
+using System.Text.Json.Nodes;
+
+namespace backend.Realtime;
+
+public class OpenAIRealtimeUpstream : IRealtimeUpstream
+{
+    private readonly ClientWebSocket _ws = new();
+    private string? _closeReason;
+
+    public bool IsClosed => _ws.State != WebSocketState.Open && _ws.State != WebSocketState.Connecting;
+    public string? CloseReason => _closeReason;
+
+    public async Task ConnectAsync(string apiKey, string model, CancellationToken ct)
+    {
+        _ws.Options.SetRequestHeader("Authorization", $"Bearer {apiKey}");
+        // The GA Realtime API does NOT take the OpenAI-Beta: realtime=v1 header.
+        // Sending it returns upstream error code=beta_api_shape_disabled.
+        var uri = new Uri($"wss://api.openai.com/v1/realtime?model={Uri.EscapeDataString(model)}");
+        await _ws.ConnectAsync(uri, ct);
+    }
+
+    public async Task SendJsonAsync(JsonObject envelope, CancellationToken ct)
+    {
+        var json = envelope.ToJsonString();
+        var bytes = Encoding.UTF8.GetBytes(json);
+        await _ws.SendAsync(bytes, WebSocketMessageType.Text, endOfMessage: true, ct);
+    }
+
+    public async Task<JsonObject?> ReceiveJsonAsync(CancellationToken ct)
+    {
+        var buffer = new byte[8192];
+        using var stream = new MemoryStream();
+        while (true)
+        {
+            WebSocketReceiveResult result;
+            try
+            {
+                result = await _ws.ReceiveAsync(buffer, ct);
+            }
+            catch (WebSocketException exc)
+            {
+                _closeReason ??= $"WebSocketException:{exc.WebSocketErrorCode}:{exc.Message}";
+                return null;
+            }
+            catch (OperationCanceledException)
+            {
+                return null;
+            }
+
+            if (result.MessageType == WebSocketMessageType.Close)
+            {
+                var status = _ws.CloseStatus?.ToString() ?? "<none>";
+                var desc = _ws.CloseStatusDescription ?? "<no description>";
+                _closeReason ??= $"close status={status} desc={desc}";
+                return null;
+            }
+            stream.Write(buffer, 0, result.Count);
+            if (result.EndOfMessage)
+            {
+                var json = Encoding.UTF8.GetString(stream.ToArray());
+                return JsonNode.Parse(json)?.AsObject();
+            }
+        }
+    }
+
+    public async ValueTask DisposeAsync()
+    {
+        if (_ws.State == WebSocketState.Open)
+        {
+            try { await _ws.CloseAsync(WebSocketCloseStatus.NormalClosure, "bye", CancellationToken.None); }
+            catch { }
+        }
+        _ws.Dispose();
+    }
+}

backend/Realtime/RealtimeEvents.cs

@@ -0,0 +1,92 @@
+using System.Text.Json;
+using System.Text.Json.Nodes;
+using backend.Tools;
+
+namespace backend.Realtime;
+
+public static class RealtimeEvents
+{
+    // GA Realtime API shape (the Beta API shape was retired in 2026 with
+    // upstream error "beta_api_shape_disabled"). Key differences:
+    //   - session.type = "realtime" is required.
+    //   - voice / audio formats / transcription / turn_detection now nested
+    //     under session.audio.{input,output}.
+    //   - No OpenAI-Beta header on the upgrade (see OpenAIRealtimeUpstream).
+    public static JsonObject SessionUpdate(RealtimeSettings cfg, IEnumerable<ITool> enabledTools)
+    {
+        var tools = new JsonArray();
+        foreach (var t in enabledTools)
+        {
+            tools.Add(new JsonObject
+            {
+                ["type"] = "function",
+                ["name"] = t.Name,
+                ["description"] = t.Description,
+                ["parameters"] = JsonNode.Parse(t.ParameterSchema.GetRawText())!,
+            });
+        }
+
+        // Minimal GA session.update. We drop explicit audio.input/output.format
+        // objects (the GA shape accepts them but OpenAI defaults to 24 kHz PCM16
+        // mono on both directions, which is what we use). Removing the explicit
+        // format objects matches the documented minimal session config and
+        // eliminates a known source of OpenAI silently ignoring the audio.
+        return new JsonObject
+        {
+            ["type"] = "session.update",
+            ["session"] = new JsonObject
+            {
+                ["type"] = "realtime",
+                ["model"] = cfg.Model,
+                ["instructions"] = cfg.SystemPrompt,
+                ["audio"] = new JsonObject
+                {
+                    ["input"] = new JsonObject
+                    {
+                        ["transcription"] = new JsonObject { ["model"] = "whisper-1" },
+                        ["turn_detection"] = new JsonObject
+                        {
+                            ["type"] = "server_vad",
+                            ["threshold"] = 0.5,
+                            ["prefix_padding_ms"] = 300,
+                            ["silence_duration_ms"] = 500,
+                            ["create_response"] = true,
+                            ["interrupt_response"] = true,
+                        },
+                    },
+                    ["output"] = new JsonObject
+                    {
+                        ["voice"] = cfg.Voice,
+                    },
+                },
+                ["tools"] = tools,
+                ["tool_choice"] = "auto",
+            },
+        };
+    }
+
+    public static JsonObject InputAudioBufferAppend(ReadOnlySpan<byte> pcm16)
+    {
+        return new JsonObject
+        {
+            ["type"] = "input_audio_buffer.append",
+            ["audio"] = Convert.ToBase64String(pcm16),
+        };
+    }
+
+    public static JsonObject FunctionCallOutput(string callId, string outputJson)
+    {
+        return new JsonObject
+        {
+            ["type"] = "conversation.item.create",
+            ["item"] = new JsonObject
+            {
+                ["type"] = "function_call_output",
+                ["call_id"] = callId,
+                ["output"] = outputJson,
+            },
+        };
+    }
+
+    public static JsonObject ResponseCreate() => new() { ["type"] = "response.create" };
+}

backend/Realtime/RealtimeSession.cs

@@ -0,0 +1,358 @@
+using System.Text.Json.Nodes;
+using backend.Conversations;
+using backend.Tools;
+using Microsoft.Extensions.Logging;
+
+namespace backend.Realtime;
+
+public interface ISessionOutput
+{
+    Task WriteEnvelopeAsync(JsonObject envelope, CancellationToken ct);
+    Task WriteBinaryAsync(ReadOnlyMemory<byte> bytes, CancellationToken ct);
+}
+
+public class RealtimeSession
+{
+    private readonly Guid _deviceId;
+    private readonly IRealtimeUpstream _upstream;
+    private readonly ISessionOutput _output;
+    private readonly ConversationLog _log;
+    private readonly ToolRegistry _registry;
+    private readonly RealtimeSettings _cfg;
+    private readonly IDeviceChannel _channel;
+    private readonly ILogger _logger;
+
+    private Conversation? _conversation;
+    private EndReason _endReason = EndReason.Unknown;
+    private bool _endRequested;
+    private DateTime _lastSpeechAt = DateTime.UtcNow;
+    private readonly List<(string CallId, string Name, string ArgsJson)> _pendingCalls = new();
+
+    private readonly System.Threading.Channels.Channel<byte[]> _uplink =
+        System.Threading.Channels.Channel.CreateBounded<byte[]>(
+            new System.Threading.Channels.BoundedChannelOptions(64)
+            {
+                FullMode = System.Threading.Channels.BoundedChannelFullMode.DropOldest,
+            });
+
+    public ValueTask WriteUplinkAsync(byte[] frame, CancellationToken ct)
+        => _uplink.Writer.WriteAsync(frame, ct);
+
+    public RealtimeSession(
+        Guid deviceId,
+        IRealtimeUpstream upstream,
+        ISessionOutput output,
+        ConversationLog log,
+        ToolRegistry registry,
+        RealtimeSettings cfg,
+        IDeviceChannel channel,
+        ILogger logger)
+    {
+        _deviceId = deviceId;
+        _upstream = upstream;
+        _output = output;
+        _log = log;
+        _registry = registry;
+        _cfg = cfg;
+        _channel = channel;
+        _logger = logger;
+    }
+
+    public Guid? ConversationId => _conversation?.Id;
+
+    public async Task RunAsync(CancellationToken ct)
+    {
+        try
+        {
+            var enabled = _registry.EnabledFor(_cfg.EnabledTools).ToList();
+            _logger.LogInformation(
+                "relay opening session: model={Model} voice={Voice} tools=[{Tools}]",
+                _cfg.Model, _cfg.Voice, string.Join(",", enabled.Select(t => t.Name)));
+            await _upstream.SendJsonAsync(RealtimeEvents.SessionUpdate(_cfg, enabled), ct);
+
+            var ack = await WaitForUpstreamTypeAsync("session.updated", ct);
+            if (ack is null)
+            {
+                _logger.LogWarning("upstream closed before session.updated");
+                _endReason = EndReason.Error;
+                return;
+            }
+
+            _conversation = await _log.StartAsync(_deviceId, ct);
+            await _output.WriteEnvelopeAsync(new JsonObject
+            {
+                ["type"] = "session_started",
+                ["conversation_id"] = _conversation.Id.ToString(),
+            }, ct);
+            _lastSpeechAt = DateTime.UtcNow;
+
+            var uplinkTask = Task.Run(() => PumpUplinkAsync(ct), ct);
+            var idleTask = Task.Run(() => WatchdogAsync(ct), ct);
+            await PumpAsync(ct);
+            _uplink.Writer.TryComplete();
+            await uplinkTask;
+            await idleTask;
+        }
+        catch (OperationCanceledException)
+        {
+            if (_endReason == EndReason.Unknown) _endReason = EndReason.Error;
+        }
+        catch (Exception ex)
+        {
+            _logger.LogError(ex, "RealtimeSession crashed");
+            _endReason = EndReason.Error;
+        }
+        finally
+        {
+            await _upstream.DisposeAsync();
+            if (_conversation is not null)
+            {
+                await _log.EndAsync(_conversation.Id, _endReason, CancellationToken.None);
+            }
+            var endEnv = new JsonObject
+            {
+                ["type"] = "session_ended",
+                ["reason"] = _endReason.ToString().ToLowerInvariant(),
+            };
+            await _output.WriteEnvelopeAsync(endEnv, CancellationToken.None);
+        }
+    }
+
+    private async Task PumpAsync(CancellationToken ct)
+    {
+        int nullPolls = 0;
+        while (!ct.IsCancellationRequested && !_endRequested)
+        {
+            var evt = await ReceiveOrIdleAsync(ct);
+            if (evt is null && _endRequested) return;
+            if (evt is null && ct.IsCancellationRequested) return;
+            if (evt is null)
+            {
+                // If the upstream WS is closed, stop spinning — surface the
+                // close reason to the device and bail with EndReason=Error.
+                if (_upstream.IsClosed)
+                {
+                    var reason = _upstream.CloseReason ?? "upstream closed";
+                    _logger.LogWarning("upstream closed mid-session: {Reason}", reason);
+                    await _output.WriteEnvelopeAsync(new JsonObject
+                    {
+                        ["type"] = "error",
+                        ["code"] = "upstream_closed",
+                        ["message"] = reason,
+                        ["fatal"] = false,
+                    }, ct);
+                    _endReason = EndReason.Error;
+                    return;
+                }
+                // Surface "alive but silent" every ~2s so we can tell that
+                // from "the upstream WS was closed". Coolify logs aren't
+                // reachable from the dev sandbox.
+                if (++nullPolls % 8 == 0)
+                {
+                    _logger.LogInformation("upstream silent ({Polls} x 250ms polls)", nullPolls);
+                    await _output.WriteEnvelopeAsync(new JsonObject
+                    {
+                        ["type"] = "debug",
+                        ["message"] = "upstream silent " + nullPolls + "x250ms",
+                    }, ct);
+                }
+                continue;
+            }
+            nullPolls = 0;
+            var type = (string?)evt["type"];
+            _logger.LogInformation("relay evt={EvtType}", type);
+            // Forward the event type to the device so it appears in the Pi
+            // journal — Coolify logs aren't accessible from the dev sandbox.
+            await _output.WriteEnvelopeAsync(new JsonObject
+            {
+                ["type"] = "debug",
+                ["message"] = "relay evt=" + (type ?? "<null>"),
+            }, ct);
+            switch (type)
+            {
+                case "error":
+                {
+                    var err = evt["error"] as JsonObject;
+                    var code = (string?)err?["code"] ?? "unknown";
+                    var msg = (string?)err?["message"] ?? evt.ToJsonString();
+                    _logger.LogWarning("upstream mid-session error code={Code} message={Message}", code, msg);
+                    await _output.WriteEnvelopeAsync(new JsonObject
+                    {
+                        ["type"] = "error",
+                        ["code"] = "upstream_" + code,
+                        ["message"] = msg,
+                        ["fatal"] = false,
+                    }, ct);
+                    break;
+                }
+                case "input_audio_buffer.speech_started":
+                    _lastSpeechAt = DateTime.UtcNow;
+                    break;
+                case "response.audio.delta":
+                    await ForwardAudioDeltaAsync(evt, ct);
+                    break;
+                case "conversation.item.input_audio_transcription.completed":
+                {
+                    var text = (string?)evt["transcript"];
+                    if (_conversation is not null && !string.IsNullOrEmpty(text))
+                        await _log.AppendUserAsync(_conversation.Id, text!, ct);
+                    break;
+                }
+                case "response.audio_transcript.done":
+                {
+                    var text = (string?)evt["transcript"];
+                    if (_conversation is not null && !string.IsNullOrEmpty(text))
+                        await _log.AppendAssistantAsync(_conversation.Id, text!, ct);
+                    break;
+                }
+                case "response.function_call_arguments.done":
+                {
+                    var callId = (string?)evt["call_id"] ?? "";
+                    var name = (string?)evt["name"] ?? "";
+                    var args = (string?)evt["arguments"] ?? "{}";
+                    _pendingCalls.Add((callId, name, args));
+                    break;
+                }
+                case "response.done":
+                {
+                    await _output.WriteEnvelopeAsync(
+                        new JsonObject { ["type"] = "assistant_done" }, ct);
+
+                    if (_pendingCalls.Count == 0) break;
+                    var calls = _pendingCalls.ToList();
+                    _pendingCalls.Clear();
+
+                    if (calls.Any(c => c.Name == "end_session"))
+                    {
+                        _endReason = EndReason.Tool;
+                        _endRequested = true;
+                        break;
+                    }
+
+                    foreach (var c in calls)
+                    {
+                        await ExecuteToolAsync(c.CallId, c.Name, c.ArgsJson, ct);
+                    }
+                    break;
+                }
+            }
+        }
+    }
+
+    private async Task ExecuteToolAsync(string callId, string name, string argsJson, CancellationToken ct)
+    {
+        var tool = _registry.Get(name);
+        string outputJson;
+        if (tool is null)
+        {
+            outputJson = $$"""{"ok":false,"error":"unknown tool: {{name}}"}""";
+        }
+        else
+        {
+            try
+            {
+                using var argsDoc = System.Text.Json.JsonDocument.Parse(argsJson);
+                var ctx = new DeviceContext(_deviceId, _conversation!.Id, _channel);
+                var res = await tool.ExecuteAsync(
+                    new ToolInvocation(callId, argsDoc.RootElement.Clone()), ctx, ct);
+                outputJson = res.Output.GetRawText();
+            }
+            catch (Exception ex)
+            {
+                outputJson = System.Text.Json.JsonSerializer.Serialize(new
+                {
+                    ok = false,
+                    error = ex.Message,
+                });
+            }
+        }
+
+        if (_conversation is not null)
+        {
+            await _log.AppendToolAsync(_conversation.Id, name, argsJson, outputJson, ct);
+        }
+
+        await _upstream.SendJsonAsync(RealtimeEvents.FunctionCallOutput(callId, outputJson), ct);
+        await _upstream.SendJsonAsync(RealtimeEvents.ResponseCreate(), ct);
+    }
+
+    private async Task ForwardAudioDeltaAsync(JsonObject evt, CancellationToken ct)
+    {
+        var b64 = (string?)evt["delta"];
+        if (string.IsNullOrEmpty(b64)) return;
+        var bytes = Convert.FromBase64String(b64);
+        await _output.WriteBinaryAsync(bytes, ct);
+    }
+
+    private async Task WatchdogAsync(CancellationToken ct)
+    {
+        try
+        {
+            var timeout = TimeSpan.FromSeconds(_cfg.IdleTimeoutSeconds);
+            while (!ct.IsCancellationRequested && !_endRequested)
+            {
+                await Task.Delay(250, ct);
+                if (DateTime.UtcNow - _lastSpeechAt > timeout)
+                {
+                    _endReason = EndReason.Idle;
+                    _endRequested = true;
+                    return;
+                }
+            }
+        }
+        catch (OperationCanceledException) { }
+    }
+
+    private async Task<JsonObject?> ReceiveOrIdleAsync(CancellationToken ct)
+    {
+        using var timeout = CancellationTokenSource.CreateLinkedTokenSource(ct);
+        var receive = _upstream.ReceiveJsonAsync(timeout.Token);
+        var delay = Task.Delay(250, timeout.Token);
+        var done = await Task.WhenAny(receive, delay);
+        if (done == receive) return await receive;
+        timeout.Cancel();
+        try { await receive; } catch { /* ignore */ }
+        return null;
+    }
+
+    private async Task PumpUplinkAsync(CancellationToken ct)
+    {
+        try
+        {
+            await foreach (var frame in _uplink.Reader.ReadAllAsync(ct))
+            {
+                await _upstream.SendJsonAsync(RealtimeEvents.InputAudioBufferAppend(frame), ct);
+            }
+        }
+        catch (OperationCanceledException) { }
+        catch (Exception ex) { _logger.LogWarning(ex, "uplink pump errored"); }
+    }
+
+    private async Task<JsonObject?> WaitForUpstreamTypeAsync(string type, CancellationToken ct)
+    {
+        while (true)
+        {
+            var evt = await _upstream.ReceiveJsonAsync(ct);
+            if (evt is null) return null;
+            var evtType = (string?)evt["type"];
+            _logger.LogInformation("upstream evt={EvtType}", evtType);
+            if (evtType == "error")
+            {
+                var err = evt["error"] as JsonObject;
+                var code = (string?)err?["code"] ?? "unknown";
+                var msg = (string?)err?["message"] ?? evt.ToJsonString();
+                _logger.LogWarning("upstream error code={Code} message={Message}", code, msg);
+                await _output.WriteEnvelopeAsync(new JsonObject
+                {
+                    ["type"] = "error",
+                    ["code"] = "upstream_" + code,
+                    ["message"] = msg,
+                    ["fatal"] = false,
+                }, ct);
+                _endReason = EndReason.Error;
+                return null;
+            }
+            if (evtType == type) return evt;
+        }
+    }
+}

backend/Realtime/RealtimeSessionFactory.cs

@@ -0,0 +1,37 @@
+using backend.Conversations;
+using backend.Tools;
+using Microsoft.Extensions.Logging;
+
+namespace backend.Realtime;
+
+public class RealtimeSessionFactory
+{
+    protected readonly OpenAIKeyProvider keys;
+    protected readonly ToolRegistry registry;
+    protected readonly ConversationLog log;
+    protected readonly ILoggerFactory loggerFactory;
+
+    public RealtimeSessionFactory(
+        OpenAIKeyProvider keys, ToolRegistry registry,
+        ConversationLog log, ILoggerFactory loggerFactory)
+    {
+        this.keys = keys;
+        this.registry = registry;
+        this.log = log;
+        this.loggerFactory = loggerFactory;
+    }
+
+    public virtual async Task<RealtimeSession> CreateAsync(
+        Guid deviceId,
+        RealtimeSettings cfg,
+        ISessionOutput output,
+        IDeviceChannel channel,
+        CancellationToken ct)
+    {
+        var upstream = new OpenAIRealtimeUpstream();
+        await upstream.ConnectAsync(keys.GetRequired(), cfg.Model, ct);
+        return new RealtimeSession(
+            deviceId, upstream, output, log, registry, cfg, channel,
+            loggerFactory.CreateLogger<RealtimeSession>());
+    }
+}

backend/Realtime/RealtimeSettings.cs

@@ -0,0 +1,8 @@
+namespace backend.Realtime;
+
+public record RealtimeSettings(
+    string Model,
+    string Voice,
+    string SystemPrompt,
+    int IdleTimeoutSeconds,
+    IReadOnlySet<string> EnabledTools);

backend/Settings/SystemSettings.cs

@@ -6,6 +6,6 @@ public class SystemSettings
     public string DefaultSystemPrompt { get; set; } =
         "You are a helpful voice assistant. Keep responses concise.";
     public string DefaultVoice { get; set; } = "alloy";
-    public string DefaultModel { get; set; } = "gpt-4o-realtime-preview";
+    public string DefaultModel { get; set; } = "gpt-realtime";
     public int DefaultIdleTimeoutSeconds { get; set; } = 30;
 }

backend/Tools/EndSessionTool.cs

@@ -0,0 +1,20 @@
+using System.Text.Json;
+
+namespace backend.Tools;
+
+public class EndSessionTool : ITool
+{
+    public string Name => "end_session";
+    public string Description =>
+        "Call this when the user has said goodbye or otherwise indicates the conversation is over.";
+    public bool RunsDuringResponse => false;
+
+    public JsonElement ParameterSchema { get; } =
+        JsonDocument.Parse("""{"type":"object","properties":{},"required":[]}""").RootElement;
+
+    public Task<ToolResult> ExecuteAsync(ToolInvocation invocation, DeviceContext device, CancellationToken ct)
+    {
+        var output = JsonDocument.Parse("""{"ok":true}""").RootElement;
+        return Task.FromResult(new ToolResult(output));
+    }
+}

backend/Tools/GetCurrentTimeTool.cs

@@ -0,0 +1,20 @@
+using System.Text.Json;
+
+namespace backend.Tools;
+
+public class GetCurrentTimeTool : ITool
+{
+    public string Name => "get_current_time";
+    public string Description => "Returns the current UTC time as an ISO-8601 string.";
+    public bool RunsDuringResponse => false;
+
+    public JsonElement ParameterSchema { get; } =
+        JsonDocument.Parse("""{"type":"object","properties":{},"required":[]}""").RootElement;
+
+    public Task<ToolResult> ExecuteAsync(ToolInvocation invocation, DeviceContext device, CancellationToken ct)
+    {
+        var now = DateTimeOffset.UtcNow.ToString("yyyy-MM-ddTHH:mm:ssZ");
+        var output = JsonDocument.Parse($$"""{"now":"{{now}}"}""").RootElement;
+        return Task.FromResult(new ToolResult(output));
+    }
+}

backend/Tools/IDeviceChannel.cs

@@ -0,0 +1,8 @@
+using System.Text.Json;
+
+namespace backend.Tools;
+
+public interface IDeviceChannel
+{
+    Task<JsonElement> CallPiToolAsync(string name, JsonElement args, CancellationToken ct);
+}

backend/Tools/ITool.cs

@@ -0,0 +1,19 @@
+using System.Text.Json;
+
+namespace backend.Tools;
+
+public record DeviceContext(Guid DeviceId, Guid ConversationId, IDeviceChannel Channel);
+public record ToolInvocation(string CallId, JsonElement Arguments);
+public record ToolResult(JsonElement Output);
+
+public interface ITool
+{
+    string Name { get; }
+    string Description { get; }
+    JsonElement ParameterSchema { get; }
+    bool RunsDuringResponse { get; }
+    Task<ToolResult> ExecuteAsync(
+        ToolInvocation invocation,
+        DeviceContext device,
+        CancellationToken ct);
+}

backend/Tools/SetVolumeTool.cs

@@ -0,0 +1,34 @@
+using System.Text.Json;
+
+namespace backend.Tools;
+
+public class SetVolumeTool : ITool
+{
+    public string Name => "set_volume";
+    public string Description =>
+        "Sets the assistant's audio output volume on the device. Level is 0..100.";
+    public bool RunsDuringResponse => true;
+
+    public JsonElement ParameterSchema { get; } = JsonDocument.Parse(
+        """
+        {
+          "type":"object",
+          "properties":{"level":{"type":"integer","minimum":0,"maximum":100}},
+          "required":["level"]
+        }
+        """).RootElement;
+
+    public async Task<ToolResult> ExecuteAsync(ToolInvocation invocation, DeviceContext device, CancellationToken ct)
+    {
+        if (!invocation.Arguments.TryGetProperty("level", out var lvlEl)
+            || !lvlEl.TryGetInt32(out var level)
+            || level < 0 || level > 100)
+        {
+            var err = JsonDocument.Parse("""{"ok":false,"error":"level must be an integer in 0..100"}""").RootElement;
+            return new ToolResult(err);
+        }
+
+        var piReply = await device.Channel.CallPiToolAsync(Name, invocation.Arguments, ct);
+        return new ToolResult(piReply);
+    }
+}

backend/Tools/ToolRegistry.cs

@@ -0,0 +1,20 @@
+namespace backend.Tools;
+
+public class ToolRegistry
+{
+    private readonly Dictionary<string, ITool> _byName;
+    private readonly List<ITool> _all;
+
+    public ToolRegistry(IEnumerable<ITool> tools)
+    {
+        _all = tools.ToList();
+        _byName = _all.ToDictionary(t => t.Name);
+    }
+
+    public IReadOnlyList<ITool> All => _all;
+
+    public ITool? Get(string name) => _byName.GetValueOrDefault(name);
+
+    public IEnumerable<ITool> EnabledFor(IReadOnlySet<string> enabled)
+        => _all.Where(t => enabled.Contains(t.Name));
+}

client/audio.py

@@ -0,0 +1,139 @@
+"""Audio helpers for the Pi client.
+
+This module is split into two layers:
+
+* Deterministic helpers (no hardware) — `resample_24k_to_16k`. Tested.
+* `AudioStreams` — owns the persistent PortAudio InputStream / OutputStream.
+  Smoke-tested by `main.py` running on the Pi.
+"""
+import queue
+import sys
+from dataclasses import dataclass, field
+from typing import Optional
+
+import numpy as np
+from scipy.signal import resample_poly
+
+from client.log import get_logger
+
+_log = get_logger("audio")
+
+SAMPLE_RATE = 24_000
+FRAME_SAMPLES = 1920            # 80 ms at 24 kHz
+CHANNELS = 1
+DTYPE = "int16"
+BYTES_PER_FRAME = FRAME_SAMPLES * 2  # int16 -> 2 bytes
+
+
+def resample_24k_to_16k(frame: np.ndarray) -> np.ndarray:
+    """Downsample a 24 kHz int16 mono frame to 16 kHz int16 mono (2:3 poly).
+
+    Input length is preserved as `len(frame) * 2 // 3`. For the 1920-sample
+    frames produced by our 80 ms blocksize at 24 kHz, this yields 1280 samples
+    (80 ms at 16 kHz), which is exactly what openwakeword expects.
+    """
+    if frame.dtype != np.int16:
+        raise TypeError(f"expected int16, got {frame.dtype}")
+    if frame.size == 0:
+        raise ValueError("empty input")
+    floats = frame.astype(np.float32)
+    resampled = resample_poly(floats, up=2, down=3)
+    return np.clip(resampled, -32768, 32767).astype(np.int16)
+
+
+def find_usb_device(sd) -> int:
+    """Return the PortAudio device index of the USB Speaker Phone.
+
+    `sd` is the imported `sounddevice` module — passed in so the deferred
+    import of sounddevice (which requires PA_ALSA_PLUGHW already set in env)
+    stays in `main.py`.
+    """
+    devices = sd.query_devices()
+    for idx, dev in enumerate(devices):
+        name = dev["name"].lower()
+        if "usb" in name and dev["max_input_channels"] >= 1:
+            return idx
+    print("USB audio device not found. Devices:", file=sys.stderr)
+    print(devices, file=sys.stderr)
+    raise SystemExit("no USB audio device found")
+
+
+@dataclass
+class AudioStreams:
+    """Wrapper around persistent 24 kHz/mono/int16 InputStream + OutputStream.
+
+    The InputStream is callback-driven; each 1920-sample frame is pushed onto
+    `mic_queue` (a `queue.Queue[bytes]`). The bridge thread drains it. The
+    OutputStream is opened but writes are driven by `PlaybackWorker`.
+    """
+    sd: object                              # sounddevice module
+    device_index: int
+    mic_queue: queue.Queue
+    overflow_counter: int = 0
+    _in_stream: Optional[object] = field(default=None, repr=False)
+    _out_stream: Optional[object] = field(default=None, repr=False)
+
+    @classmethod
+    def open(cls, sd, mic_queue: queue.Queue) -> "AudioStreams":
+        device = find_usb_device(sd)
+        _log.info("opening audio on device %d: %r", device, sd.query_devices(device)["name"])
+
+        streams = cls(sd=sd, device_index=device, mic_queue=mic_queue)
+
+        def _on_input(indata, frames, time_info, status):  # PortAudio thread
+            if status:
+                streams.overflow_counter += 1
+                _log.warning("input status: %s", status)
+            mono = indata[:, 0].tobytes()
+            try:
+                mic_queue.put_nowait(mono)
+            except queue.Full:
+                try:
+                    mic_queue.get_nowait()
+                except queue.Empty:
+                    pass
+                try:
+                    mic_queue.put_nowait(mono)
+                except queue.Full:
+                    pass
+                _log.warning("mic_queue full; dropped oldest frame")
+
+        streams._in_stream = sd.InputStream(
+            samplerate=SAMPLE_RATE,
+            channels=CHANNELS,
+            dtype=DTYPE,
+            device=device,
+            blocksize=FRAME_SAMPLES,
+            callback=_on_input,
+        )
+        streams._out_stream = sd.OutputStream(
+            samplerate=SAMPLE_RATE,
+            channels=CHANNELS,
+            dtype=DTYPE,
+            device=device,
+            blocksize=FRAME_SAMPLES,
+        )
+        return streams
+
+    def start(self) -> None:
+        self._in_stream.start()
+        self._out_stream.start()
+
+    def write(self, pcm16_bytes: bytes) -> None:
+        """Blocking write to the OutputStream (from the playback worker thread)."""
+        if not pcm16_bytes:
+            return
+        buf = np.frombuffer(pcm16_bytes, dtype=np.int16)
+        self._out_stream.write(buf)
+
+    def close(self) -> None:
+        for s in (self._in_stream, self._out_stream):
+            if s is not None:
+                try:
+                    s.stop()
+                except Exception:
+                    pass
+                try:
+                    s.close()
+                except Exception:
+                    pass

client/config.py

@@ -0,0 +1,49 @@
+"""Read / write the device config file at ~/assistant/state/config.json."""
+import json
+import os
+from dataclasses import asdict, dataclass
+from pathlib import Path
+
+
+class ConfigMissingError(RuntimeError):
+    """Raised when config.json is missing or malformed."""
+
+
+@dataclass(frozen=True)
+class Config:
+    device_id: str
+    device_token: str
+    backend_ws: str
+
+
+def config_path() -> Path:
+    return Path(os.environ["HOME"]) / "assistant" / "state" / "config.json"
+
+
+def load_config() -> Config:
+    path = config_path()
+    if not path.exists():
+        raise ConfigMissingError(f"{path} not found; run `python -m client.pair` first")
+    try:
+        raw = json.loads(path.read_text())
+    except json.JSONDecodeError as exc:
+        raise ConfigMissingError(f"{path} is not valid JSON: {exc}") from exc
+    try:
+        return Config(
+            device_id=str(raw["device_id"]),
+            device_token=str(raw["device_token"]),
+            backend_ws=str(raw["backend_ws"]),
+        )
+    except KeyError as exc:
+        raise ConfigMissingError(f"{path} missing field: {exc}") from exc
+
+
+def save_config(cfg: Config) -> None:
+    path = config_path()
+    path.parent.mkdir(parents=True, exist_ok=True)
+    # Write to a temp file then atomically rename, so a partial write never
+    # leaves a half-baked config on disk.
+    tmp = path.with_suffix(".json.tmp")
+    tmp.write_text(json.dumps(asdict(cfg), indent=2) + "\n")
+    os.chmod(tmp, 0o600)
+    os.replace(tmp, path)

client/log.py

@@ -0,0 +1,28 @@
+"""Structured stdout logging suitable for journald capture."""
+import logging
+import os
+import sys
+
+_CONFIGURED = False
+
+
+def configure(level: str | int | None = None) -> None:
+    """Configure root logging once. Safe to call multiple times."""
+    global _CONFIGURED
+    if _CONFIGURED:
+        return
+    if level is None:
+        level = os.environ.get("ASSISTANT_LOG_LEVEL", "INFO").upper()
+    logging.basicConfig(
+        level=level,
+        format="%(asctime)s %(levelname)-5s %(name)s %(message)s",
+        datefmt="%H:%M:%S",
+        stream=sys.stdout,
+    )
+    _CONFIGURED = True
+
+
+def get_logger(name: str) -> logging.Logger:
+    """Return a logger named `name`; auto-configure on first call."""
+    configure()
+    return logging.getLogger(name)

client/main.py

@@ -1,8 +1,191 @@
-"""Smart Assistant client entry point. Implemented in Plan 3."""
+"""Smart Assistant Pi client entry point.
+
+CRITICAL: set PA_ALSA_PLUGHW BEFORE importing sounddevice (see findings.md §1).
+"""
+import os
+os.environ.setdefault("PA_ALSA_PLUGHW", "1")
+
+import asyncio
+import queue
+import signal
+import sys
+import threading
+from pathlib import Path
+
+import numpy as np
+import sounddevice as sd  # safe to import now
+
+from client.audio import (
+    AudioStreams,
+    FRAME_SAMPLES,
+    resample_24k_to_16k,
+)
+from client.config import ConfigMissingError, load_config
+from client.log import get_logger
+from client.playback import PlaybackWorker
+from client.session import Session
+from client.state import StateMachine
+from client.wakeword import WakewordDetector
+
+_log = get_logger("main")
+
+ASSISTANT_STATE_DIR = Path(os.environ["HOME"]) / "assistant" / "state"
+WAKE_WAV = ASSISTANT_STATE_DIR / "wake.wav"
+SLEEP_WAV = ASSISTANT_STATE_DIR / "sleep.wav"
+
+
+def _build_bridge_thread(
+    mic_queue: queue.Queue,
+    sm: StateMachine,
+    wakeword: WakewordDetector,
+    session: Session,
+    playback: PlaybackWorker,
+    stop_event: threading.Event,
+) -> threading.Thread:
+    def run():
+        frame_counter = 0
+        recent_rms_max = 0.0
+        while not stop_event.is_set():
+            try:
+                pcm16_bytes = mic_queue.get(timeout=0.25)
+            except queue.Empty:
+                continue
+            frame_counter += 1
+            raw = np.frombuffer(pcm16_bytes, dtype=np.int16)
+            rms = float(np.sqrt(np.mean(raw.astype(np.float32) ** 2)))
+            if rms > recent_rms_max:
+                recent_rms_max = rms
+            # Every ~5 s emit a heartbeat so we can see if the InputStream is
+            # actually producing frames and what state we're routing to.
+            if frame_counter % 62 == 0:
+                _log.info(
+                    "bridge frames=%d state=%s wake_en=%s up_en=%s qsize=%d rms_max_5s=%.0f",
+                    frame_counter, sm.state.name,
+                    sm.wakeword_enabled, sm.uplink_enabled, mic_queue.qsize(),
+                    recent_rms_max,
+                )
+                recent_rms_max = 0.0
+            if sm.wakeword_enabled:
+                frame = np.frombuffer(pcm16_bytes, dtype=np.int16)
+                if frame.size != FRAME_SAMPLES:
+                    continue
+                feed = resample_24k_to_16k(frame)
+                if wakeword.predict(feed):
+                    if sm.wake_fired():
+                        if WAKE_WAV.exists():
+                            playback.enqueue_wav(WAKE_WAV, "wake")
+                        else:
+                            # Even with no wav, fire the done marker so the
+                            # playback_done(for=wake) envelope is sent.
+                            playback.enqueue_done_marker("wake")
+                        session.call_soon_threadsafe_send_wake()
+            elif sm.uplink_enabled:
+                session.call_soon_threadsafe_send_binary(pcm16_bytes)
+            # else: ASSISTANT_SPEAKING / WAKE_PENDING / IDLE_PENDING -> drop
+    t = threading.Thread(target=run, name="bridge", daemon=True)
+    return t
+
+
+def _make_callbacks(sm: StateMachine, playback: PlaybackWorker):
+    class _CB:
+        def on_session_started(self, conversation_id: str) -> None:
+            _log.info("session_started conversation_id=%s", conversation_id)
+            sm.session_started()
+
+        def on_assistant_audio(self, pcm16: bytes) -> None:
+            sm.assistant_audio_arrived()
+            playback.enqueue_audio(pcm16)
+
+        def on_assistant_done(self) -> None:
+            _log.info("assistant_done")
+            sm.assistant_done()
+
+        def on_session_ended(self, reason: str) -> None:
+            _log.info("session_ended reason=%s", reason)
+            sm.session_ended()
+            if SLEEP_WAV.exists():
+                playback.enqueue_wav(SLEEP_WAV, "sleep")
+            else:
+                playback.enqueue_done_marker("sleep")
+
+        def on_disconnected(self) -> None:
+            # The conversation is gone with the WS — reset to IDLE so the
+            # wakeword detector can re-arm. Any pending assistant audio left
+            # in the playback queue still drains, but we drop a "sleep" done
+            # marker only if we were mid-session.
+            _log.info("on_disconnected: forcing state to IDLE")
+            sm.force_idle()
+    return _CB()
 
 
 def main() -> None:
-    raise SystemExit("client.main: placeholder; full implementation comes in Plan 3.")
+    try:
+        cfg = load_config()
+    except ConfigMissingError as exc:
+        raise SystemExit(str(exc)) from exc
+
+    _log.info("device_id=%s backend=%s", cfg.device_id, cfg.backend_ws)
+
+    mic_queue: queue.Queue = queue.Queue(maxsize=64)
+    streams = AudioStreams.open(sd, mic_queue=mic_queue)
+    streams.start()
+
+    sm = StateMachine()
+
+    session_ref: list[Session] = []
+
+    def on_playback_done(label: str) -> None:
+        _log.info("playback_done for=%s", label)
+        if not session_ref:
+            return
+        session = session_ref[0]
+        session.call_soon_threadsafe_send_playback_done(label)
+        if label == "sleep":
+            sm.sleep_played()
+
+    playback = PlaybackWorker(streams, on_done=on_playback_done)
+    playback.start()
+
+    wakeword = WakewordDetector()
+
+    callbacks = _make_callbacks(sm, playback)
+    session = Session(
+        backend_ws=cfg.backend_ws,
+        device_token=cfg.device_token,
+        device_id=cfg.device_id,
+        callbacks=callbacks,
+    )
+    session_ref.append(session)
+
+    def set_volume_handler(args: dict) -> dict:
+        level = int(args.get("level", -1))
+        playback.set_volume(level)
+        return {"ok": True, "level": level}
+
+    session.register_tool("set_volume", set_volume_handler)
+
+    stop_event = threading.Event()
+    bridge = _build_bridge_thread(mic_queue, sm, wakeword, session, playback, stop_event)
+    bridge.start()
+
+    def _shutdown(signum, _frame):
+        _log.info("signal %d received; shutting down", signum)
+        stop_event.set()
+        try:
+            playback.stop()
+        finally:
+            streams.close()
+        sys.exit(0)
+
+    signal.signal(signal.SIGTERM, _shutdown)
+    signal.signal(signal.SIGINT, _shutdown)
+
+    try:
+        asyncio.run(session.run_forever())
+    finally:
+        stop_event.set()
+        playback.stop()
+        streams.close()
 
 
 if __name__ == "__main__":

client/pair.py

@@ -1,8 +1,67 @@
-"""Smart Assistant pairing CLI. Implemented in Plan 3."""
+"""Pair this Pi with the backend.
+
+Usage:
+    python -m client.pair --backend https://backend.example.com [--code XYZ12345]
+
+Called from install.sh after it prompts the user for the code.
+"""
+import argparse
+import socket
+
+import requests
+
+from client.config import Config, save_config
+from client.log import get_logger
+
+_log = get_logger("pair")
+
+CLIENT_VERSION = "0.1.0"
+
+
+def pair(*, backend: str, code: str | None, hostname: str) -> None:
+    backend = backend.rstrip("/")
+    if not code:
+        code = input("Pairing code: ").strip()
+    if not code:
+        raise SystemExit("no pairing code supplied")
+
+    url = f"{backend}/api/pair"
+    payload = {"code": code, "hostname": hostname, "client_version": CLIENT_VERSION}
+    _log.info("POST %s code=%s hostname=%s", url, code, hostname)
+
+    try:
+        resp = requests.post(url, json=payload, timeout=30)
+    except requests.RequestException as exc:
+        raise SystemExit(f"pair request failed: {exc}") from exc
+
+    if resp.status_code != 200:
+        try:
+            err = resp.json().get("error", resp.text)
+        except ValueError:
+            err = resp.text
+        raise SystemExit(f"/api/pair returned {resp.status_code}: {err}")
+
+    body = resp.json()
+    cfg = Config(
+        device_id=str(body["device_id"]),
+        device_token=str(body["device_token"]),
+        backend_ws=str(body["backend_ws"]),
+    )
+    save_config(cfg)
+    _log.info("paired device %s; config saved", cfg.device_id)
 
 
 def main() -> None:
-    raise SystemExit("client.pair: placeholder; full implementation comes in Plan 3.")
+    parser = argparse.ArgumentParser(description="Pair this Pi with the backend.")
+    parser.add_argument("--backend", required=True, help="Backend base URL (https://...)")
+    parser.add_argument("--code", default=None, help="Pairing code (prompted if omitted)")
+    parser.add_argument(
+        "--hostname",
+        default=socket.gethostname(),
+        help="Hostname reported to the backend (defaults to socket.gethostname())",
+    )
+    args = parser.parse_args()
+    pair(backend=args.backend, code=args.code, hostname=args.hostname)
 
 
 if __name__ == "__main__":

client/playback.py

@@ -0,0 +1,117 @@
+"""Playback worker thread + wake/sleep cue + software volume.
+
+Items pushed to the queue:
+  ("audio", bytes)     - PCM16 LE 24 kHz mono data to write
+  ("wav", (Path, label)) - WAV file to read and play; triggers on_done(label) after
+  ("done_marker", str) - synthetic marker: after this is popped, call on_done(label)
+  ("stop", None)       - shut the worker down
+
+Volume is applied as a software gain in [0..1.0] on int16 buffers because the
+USB Speaker Phone has no hardware playback volume control (only `PCM Playback
+Switch`, a mute).
+"""
+import queue
+import threading
+import wave
+from pathlib import Path
+from typing import Callable, Tuple
+
+import numpy as np
+
+from client.audio import SAMPLE_RATE
+from client.log import get_logger
+
+_log = get_logger("playback")
+
+
+class PlaybackWorker:
+    def __init__(self, audio_streams, on_done: Callable[[str], None]):
+        """
+        `audio_streams.write(pcm16_bytes)` is the blocking write into the
+        persistent OutputStream. `on_done(label)` is invoked when a tagged
+        cue (e.g. "wake", "sleep") finishes playing.
+        """
+        self._streams = audio_streams
+        self._on_done = on_done
+        self._q: "queue.Queue[Tuple[str, object]]" = queue.Queue()
+        self._volume = 1.0
+        self._thread = threading.Thread(target=self._run, name="playback", daemon=True)
+        self._running = False
+
+    def start(self) -> None:
+        self._running = True
+        self._thread.start()
+
+    def stop(self) -> None:
+        self._running = False
+        self._q.put(("stop", None))
+        self._thread.join(timeout=2.0)
+
+    def set_volume(self, level: int) -> None:
+        """Set software gain. `level` is 0..100."""
+        if not 0 <= level <= 100:
+            raise ValueError("level must be 0..100")
+        self._volume = level / 100.0
+        _log.info("volume set to %d%%", level)
+
+    def enqueue_audio(self, pcm16_bytes: bytes) -> None:
+        self._q.put(("audio", pcm16_bytes))
+
+    def enqueue_wav(self, path: Path, label: str) -> None:
+        """Queue a WAV cue. When fully drained, on_done(label) fires."""
+        self._q.put(("wav", (path, label)))
+
+    def enqueue_done_marker(self, label: str) -> None:
+        """Fire on_done(label) once this marker is processed (after all prior audio)."""
+        self._q.put(("done_marker", label))
+
+    def _apply_gain(self, pcm16: bytes) -> bytes:
+        if self._volume == 1.0:
+            return pcm16
+        arr = np.frombuffer(pcm16, dtype=np.int16).astype(np.int32)
+        arr = (arr * int(self._volume * 1024)) >> 10
+        arr = np.clip(arr, -32768, 32767).astype(np.int16)
+        return arr.tobytes()
+
+    def _play_wav(self, path: Path, label: str) -> None:
+        if not path.exists():
+            _log.info("wav %s not present at %s; skipping", label, path)
+            self._on_done(label)
+            return
+        try:
+            with wave.open(str(path), "rb") as w:
+                if w.getframerate() != SAMPLE_RATE or w.getnchannels() != 1 or w.getsampwidth() != 2:
+                    _log.warning(
+                        "wav %s is %dHz/%dch/%dB; expected %dHz/mono/16-bit",
+                        path, w.getframerate(), w.getnchannels(), w.getsampwidth() * 8, SAMPLE_RATE,
+                    )
+                    self._on_done(label)
+                    return
+                data = w.readframes(w.getnframes())
+        except wave.Error as exc:
+            _log.warning("could not read wav %s: %s", path, exc)
+            self._on_done(label)
+            return
+        chunk = 3840
+        for i in range(0, len(data), chunk):
+            self._streams.write(self._apply_gain(data[i:i + chunk]))
+        self._on_done(label)
+
+    def _run(self) -> None:
+        while self._running:
+            try:
+                kind, payload = self._q.get(timeout=0.5)
+            except queue.Empty:
+                continue
+            try:
+                if kind == "stop":
+                    return
+                if kind == "audio":
+                    self._streams.write(self._apply_gain(payload))
+                elif kind == "wav":
+                    path, label = payload
+                    self._play_wav(path, label)
+                elif kind == "done_marker":
+                    self._on_done(payload)
+            except Exception as exc:  # pragma: no cover - hardware path
+                _log.exception("playback worker error: %s", exc)

client/session.py

@@ -0,0 +1,235 @@
+"""Backend WebSocket client (asyncio)."""
+import asyncio
+import json
+import random
+from typing import Callable, Protocol
+
+from client.log import get_logger
+
+_log = get_logger("session")
+
+CLIENT_VERSION = "0.1.0"
+PING_INTERVAL_S = 15.0
+RECONNECT_MIN_S = 1.0
+RECONNECT_MAX_S = 30.0
+
+
+class Callbacks(Protocol):
+    def on_session_started(self, conversation_id: str) -> None: ...
+    def on_assistant_audio(self, pcm16: bytes) -> None: ...
+    def on_assistant_done(self) -> None: ...
+    def on_session_ended(self, reason: str) -> None: ...
+    # Called once per backend WS lifecycle when the connection drops, so the
+    # owner can reset state (the conversation is dead).
+    def on_disconnected(self) -> None: ...
+
+
+class Session:
+    def __init__(
+        self,
+        backend_ws: str,
+        device_token: str,
+        device_id: str,
+        callbacks: Callbacks,
+    ) -> None:
+        self._backend_ws = backend_ws
+        self._device_token = device_token
+        self._device_id = device_id
+        self._cb = callbacks
+        self._tools: dict[str, Callable[[dict], dict]] = {}
+        self._ws = None
+        self._loop: asyncio.AbstractEventLoop | None = None
+
+    def register_tool(self, name: str, handler: Callable[[dict], dict]) -> None:
+        self._tools[name] = handler
+
+    async def run_forever(self) -> None:
+        """Connect, pump, reconnect with exponential backoff. Never returns
+        cleanly except on cancellation."""
+        # Lazy import so the module is importable without `websockets`.
+        import websockets
+
+        self._loop = asyncio.get_running_loop()
+        backoff = RECONNECT_MIN_S
+        while True:
+            try:
+                _log.info("connecting to %s", self._backend_ws)
+                async with websockets.connect(
+                    self._backend_ws,
+                    additional_headers={"Authorization": f"Bearer {self._device_token}"},
+                    max_size=4 * 1024 * 1024,
+                ) as ws:
+                    self._ws = ws
+                    backoff = RECONNECT_MIN_S
+                    ping_task = asyncio.create_task(self._ping_loop(ws))
+                    try:
+                        await self._pump(ws)
+                    finally:
+                        ping_task.cancel()
+                        self._ws = None
+                        self._notify_disconnected()
+            except asyncio.CancelledError:
+                raise
+            except Exception as exc:  # noqa: BLE001
+                _log.warning("session error: %s; reconnecting in %.1fs", exc, backoff)
+                self._notify_disconnected()
+            jitter = random.uniform(0, backoff * 0.1)
+            await asyncio.sleep(backoff + jitter)
+            backoff = min(backoff * 2.0, RECONNECT_MAX_S)
+
+    async def _pump(self, ws) -> None:
+        # 1. Send hello.
+        await ws.send(json.dumps({
+            "type": "hello",
+            "device_id": self._device_id,
+            "client_version": CLIENT_VERSION,
+        }))
+        # Track ws so the threadsafe shims work when _pump is exercised directly
+        # (i.e. without going through run_forever, as in unit tests).
+        self._ws = ws
+        try:
+            self._loop = asyncio.get_running_loop()
+        except RuntimeError:
+            pass
+        # 2. Pump.
+        while True:
+            frame = await ws.recv()
+            if isinstance(frame, (bytes, bytearray, memoryview)):
+                self._cb.on_assistant_audio(bytes(frame))
+                continue
+            try:
+                env = json.loads(frame)
+            except json.JSONDecodeError:
+                _log.warning("non-JSON text frame; dropping")
+                continue
+            await self._dispatch(env, ws)
+
+    async def _dispatch(self, env: dict, ws) -> None:
+        t = env.get("type")
+        if t == "hello_ack":
+            _log.info("hello_ack config=%s", env.get("config"))
+        elif t == "pong":
+            pass
+        elif t == "session_started":
+            self._cb.on_session_started(str(env.get("conversation_id", "")))
+        elif t == "assistant_done":
+            self._cb.on_assistant_done()
+        elif t == "session_ended":
+            self._cb.on_session_ended(str(env.get("reason", "unknown")))
+        elif t == "tool_call":
+            await self._handle_tool_call(env, ws)
+        elif t == "error":
+            _log.warning(
+                "backend error code=%s message=%s fatal=%s",
+                env.get("code"), env.get("message"), env.get("fatal"),
+            )
+        elif t == "config_updated":
+            _log.info("config_updated: %s", env.get("config"))
+        elif t == "debug":
+            _log.info("backend %s", env.get("message"))
+        else:
+            _log.info("unhandled envelope type=%s", t)
+
+    async def _handle_tool_call(self, env: dict, ws) -> None:
+        call_id = env.get("call_id", "")
+        name = env.get("name", "")
+        args = env.get("arguments") or {}
+        if isinstance(args, str):
+            try:
+                args = json.loads(args)
+            except json.JSONDecodeError:
+                args = {}
+        handler = self._tools.get(name)
+        if handler is None:
+            await ws.send(json.dumps({
+                "type": "tool_result",
+                "call_id": call_id,
+                "ok": False,
+                "error": f"unknown tool: {name}",
+            }))
+            return
+        try:
+            result = handler(args)
+        except Exception as exc:  # noqa: BLE001
+            await ws.send(json.dumps({
+                "type": "tool_result",
+                "call_id": call_id,
+                "ok": False,
+                "error": str(exc),
+            }))
+            return
+        await ws.send(json.dumps({
+            "type": "tool_result",
+            "call_id": call_id,
+            "ok": True,
+            "result": result,
+        }))
+
+    def _notify_disconnected(self) -> None:
+        # Tolerant: Callbacks Protocol marks on_disconnected as part of the
+        # surface, but real callers (and tests) may omit it.
+        fn = getattr(self._cb, "on_disconnected", None)
+        if fn is None:
+            return
+        try:
+            fn()
+        except Exception as exc:  # noqa: BLE001
+            _log.warning("on_disconnected raised: %s", exc)
+
+    async def _ping_loop(self, ws) -> None:
+        try:
+            while True:
+                await asyncio.sleep(PING_INTERVAL_S)
+                await ws.send(json.dumps({"type": "ping"}))
+        except asyncio.CancelledError:
+            pass
+
+    # ---- Public helpers used by main.py from the asyncio loop ----
+    #
+    # All of these are called fire-and-forget via call_soon_threadsafe, so they
+    # must NEVER raise — the asyncio loop has nowhere to surface the exception
+    # except as an unhandled-task-exception traceback in the journal.
+
+    async def _safe_send(self, payload) -> None:
+        if self._ws is None:
+            return
+        try:
+            await self._ws.send(payload)
+        except Exception as exc:  # noqa: BLE001
+            # ConnectionClosed / ConnectionClosedOK / ConnectionClosedError all
+            # land here. The reconnect loop already handles WS lifecycle.
+            _log.debug("send dropped: %s", exc)
+
+    async def send_wake(self) -> None:
+        await self._safe_send(json.dumps({"type": "wake"}))
+
+    async def send_binary(self, pcm16: bytes) -> None:
+        await self._safe_send(pcm16)
+
+    async def send_playback_done(self, label: str) -> None:
+        await self._safe_send(json.dumps({"type": "playback_done", "for": label}))
+
+    async def send_cancel(self) -> None:
+        await self._safe_send(json.dumps({"type": "cancel"}))
+
+    def call_soon_threadsafe_send_binary(self, pcm16: bytes) -> None:
+        """Thread-safe shim used by the bridge thread to queue an uplink send."""
+        if self._loop is None or self._ws is None:
+            return
+        self._loop.call_soon_threadsafe(
+            lambda: asyncio.create_task(self.send_binary(pcm16))
+        )
+
+    def call_soon_threadsafe_send_wake(self) -> None:
+        if self._loop is None or self._ws is None:
+            return
+        self._loop.call_soon_threadsafe(
+            lambda: asyncio.create_task(self.send_wake())
+        )
+
+    def call_soon_threadsafe_send_playback_done(self, label: str) -> None:
+        if self._loop is None or self._ws is None:
+            return
+        self._loop.call_soon_threadsafe(
+            lambda: asyncio.create_task(self.send_playback_done(label))
+        )

client/state.py

@@ -0,0 +1,77 @@
+"""Assistant client state machine.
+
+States flow: IDLE -> WAKE_PENDING -> LISTENING <-> ASSISTANT_SPEAKING
+                                       \\         /
+                                        session_ended
+                                                 \\
+                                                  IDLE_PENDING -> IDLE
+"""
+from enum import Enum, auto
+
+from client.log import get_logger
+
+_log = get_logger("state")
+
+
+class State(Enum):
+    IDLE = auto()
+    WAKE_PENDING = auto()
+    LISTENING = auto()
+    ASSISTANT_SPEAKING = auto()
+    IDLE_PENDING = auto()
+
+
+class StateMachine:
+    def __init__(self) -> None:
+        self._state = State.IDLE
+
+    @property
+    def state(self) -> State:
+        return self._state
+
+    @property
+    def wakeword_enabled(self) -> bool:
+        return self._state is State.IDLE
+
+    @property
+    def uplink_enabled(self) -> bool:
+        return self._state is State.LISTENING
+
+    def _transition(self, target: State) -> None:
+        if target is self._state:
+            return
+        _log.info("state %s -> %s", self._state.name, target.name)
+        self._state = target
+
+    def wake_fired(self) -> bool:
+        if self._state is not State.IDLE:
+            return False
+        self._transition(State.WAKE_PENDING)
+        return True
+
+    def session_started(self) -> None:
+        if self._state in (State.WAKE_PENDING, State.IDLE):
+            self._transition(State.LISTENING)
+
+    def assistant_audio_arrived(self) -> None:
+        if self._state is State.LISTENING:
+            self._transition(State.ASSISTANT_SPEAKING)
+
+    def assistant_done(self) -> None:
+        if self._state is State.ASSISTANT_SPEAKING:
+            self._transition(State.LISTENING)
+
+    def session_ended(self) -> None:
+        if self._state in (State.LISTENING, State.ASSISTANT_SPEAKING, State.WAKE_PENDING):
+            self._transition(State.IDLE_PENDING)
+
+    def sleep_played(self) -> None:
+        if self._state is State.IDLE_PENDING:
+            self._transition(State.IDLE)
+
+    def force_idle(self) -> None:
+        """Unconditional reset to IDLE — used when the backend WS drops and
+        the conversation is dead regardless of what state we thought we were in.
+        Re-enables the wakeword detector."""
+        if self._state is not State.IDLE:
+            self._transition(State.IDLE)

client/tests/conftest.py

@@ -0,0 +1,8 @@
+"""Shared pytest fixtures for the client test suite."""
+import sys
+from pathlib import Path
+
+# Make the repo root importable so `import client.<module>` works without install.
+REPO_ROOT = Path(__file__).resolve().parents[2]
+if str(REPO_ROOT) not in sys.path:
+    sys.path.insert(0, str(REPO_ROOT))

client/tests/pytest.ini

@@ -0,0 +1,2 @@
+[pytest]
+asyncio_mode = auto

client/tests/test_audio_resample.py

@@ -0,0 +1,47 @@
+import numpy as np
+import pytest
+
+scipy_signal = pytest.importorskip("scipy.signal")
+
+from client.audio import resample_24k_to_16k
+
+
+def test_length_and_dtype():
+    frame = np.zeros(1920, dtype=np.int16)
+    out = resample_24k_to_16k(frame)
+    assert out.shape == (1280,)
+    assert out.dtype == np.int16
+
+
+def test_constant_signal_stays_constant():
+    frame = np.full(1920, 1000, dtype=np.int16)
+    out = resample_24k_to_16k(frame)
+    # poly resample of a constant is the same constant up to filter ringing
+    assert np.abs(out.astype(int).mean() - 1000) < 5
+
+
+def test_silence_in_silence_out():
+    frame = np.zeros(1920, dtype=np.int16)
+    out = resample_24k_to_16k(frame)
+    assert int(np.max(np.abs(out))) == 0
+
+
+def test_sine_peak_frequency_preserved():
+    # 1 kHz sine at 24 kHz: 80 ms = 80 cycles.
+    t = np.arange(1920) / 24_000.0
+    sig = (8000 * np.sin(2 * np.pi * 1000 * t)).astype(np.int16)
+    out = resample_24k_to_16k(sig)
+    # 1 kHz in a 1280-sample 16 kHz window is bin (1000 / 16000) * 1280 = 80.
+    spec = np.abs(np.fft.rfft(out.astype(np.float32)))
+    peak = int(np.argmax(spec))
+    assert abs(peak - 80) <= 1
+
+
+def test_zero_length_input_raises():
+    with pytest.raises(ValueError):
+        resample_24k_to_16k(np.zeros(0, dtype=np.int16))
+
+
+def test_wrong_dtype_rejected():
+    with pytest.raises(TypeError):
+        resample_24k_to_16k(np.zeros(1920, dtype=np.float32))

client/tests/test_config.py

@@ -0,0 +1,59 @@
+import json
+import os
+import stat
+
+import pytest
+
+from client.config import (
+    Config,
+    ConfigMissingError,
+    config_path,
+    load_config,
+    save_config,
+)
+
+
+def test_config_path_under_assistant_state(tmp_path, monkeypatch):
+    monkeypatch.setenv("HOME", str(tmp_path))
+    assert config_path() == tmp_path / "assistant" / "state" / "config.json"
+
+
+def test_load_missing_raises(tmp_path, monkeypatch):
+    monkeypatch.setenv("HOME", str(tmp_path))
+    with pytest.raises(ConfigMissingError):
+        load_config()
+
+
+def test_save_then_load_round_trip(tmp_path, monkeypatch):
+    monkeypatch.setenv("HOME", str(tmp_path))
+    cfg = Config(
+        device_id="11111111-1111-1111-1111-111111111111",
+        device_token="t0k3n",
+        backend_ws="wss://example/device",
+    )
+    save_config(cfg)
+    loaded = load_config()
+    assert loaded == cfg
+
+
+def test_save_creates_parent_dirs(tmp_path, monkeypatch):
+    monkeypatch.setenv("HOME", str(tmp_path))
+    cfg = Config(device_id="x", device_token="y", backend_ws="wss://z")
+    save_config(cfg)
+    assert (tmp_path / "assistant" / "state").is_dir()
+
+
+def test_save_sets_mode_0600(tmp_path, monkeypatch):
+    monkeypatch.setenv("HOME", str(tmp_path))
+    cfg = Config(device_id="x", device_token="y", backend_ws="wss://z")
+    save_config(cfg)
+    mode = stat.S_IMODE(os.stat(config_path()).st_mode)
+    assert mode == 0o600
+
+
+def test_load_rejects_missing_fields(tmp_path, monkeypatch):
+    monkeypatch.setenv("HOME", str(tmp_path))
+    config_path().parent.mkdir(parents=True, exist_ok=True)
+    config_path().write_text(json.dumps({"device_id": "x"}))
+    with pytest.raises(ConfigMissingError):
+        load_config()

client/tests/test_pair.py

@@ -0,0 +1,86 @@
+import io
+
+import pytest
+import responses
+
+from client.config import load_config
+from client.pair import pair
+
+
+@responses.activate
+def test_pair_posts_to_api_pair_and_writes_config(tmp_path, monkeypatch):
+    monkeypatch.setenv("HOME", str(tmp_path))
+    responses.post(
+        "https://backend.test/api/pair",
+        json={
+            "device_id": "00000000-0000-0000-0000-000000000001",
+            "device_token": "T0K3N",
+            "backend_ws": "wss://backend.test/device",
+        },
+        status=200,
+    )
+
+    pair(backend="https://backend.test", code="ABC12345", hostname="testhost")
+
+    cfg = load_config()
+    assert cfg.device_id == "00000000-0000-0000-0000-000000000001"
+    assert cfg.device_token == "T0K3N"
+    assert cfg.backend_ws == "wss://backend.test/device"
+
+    body = responses.calls[0].request.body
+    assert b"ABC12345" in body
+    assert b"testhost" in body
+
+
+@responses.activate
+def test_pair_strips_trailing_slash_from_backend(tmp_path, monkeypatch):
+    monkeypatch.setenv("HOME", str(tmp_path))
+    responses.post(
+        "https://backend.test/api/pair",
+        json={"device_id": "x", "device_token": "y", "backend_ws": "wss://z"},
+        status=200,
+    )
+    pair(backend="https://backend.test/", code="ABC12345", hostname="h")
+    assert responses.calls[0].request.url == "https://backend.test/api/pair"
+
+
+@responses.activate
+def test_pair_raises_on_404(tmp_path, monkeypatch):
+    monkeypatch.setenv("HOME", str(tmp_path))
+    responses.post(
+        "https://backend.test/api/pair",
+        json={"error": "code invalid or expired"},
+        status=404,
+    )
+    with pytest.raises(SystemExit) as exc:
+        pair(backend="https://backend.test", code="WRONGCOD", hostname="h")
+    assert "code invalid or expired" in str(exc.value)
+
+
+def test_prompt_used_when_code_not_provided(tmp_path, monkeypatch):
+    monkeypatch.setenv("HOME", str(tmp_path))
+    monkeypatch.setattr("sys.stdin", io.StringIO("FROMSTDIN\n"))
+
+    called_with = {}
+
+    def fake_post(url, json, timeout):
+        called_with["json"] = json
+
+        class R:
+            status_code = 200
+
+            def json(self_):
+                return {
+                    "device_id": "x",
+                    "device_token": "y",
+                    "backend_ws": "wss://z",
+                }
+
+            def raise_for_status(self_):
+                pass
+
+        return R()
+
+    monkeypatch.setattr("client.pair.requests.post", fake_post)
+    pair(backend="https://backend.test", code=None, hostname="h")
+    assert called_with["json"]["code"] == "FROMSTDIN"

client/tests/test_session.py

@@ -0,0 +1,186 @@
+import asyncio
+import json
+
+import pytest
+
+from client.session import Session
+
+
+class FakeWebSocket:
+    """Minimal stand-in for a `websockets` client connection."""
+    def __init__(self, inbox: list):
+        # inbox is a list of either str (text frame) or bytes (binary frame).
+        self._inbox = list(inbox)
+        self.sent: list = []
+
+    async def send(self, frame):
+        self.sent.append(frame)
+
+    async def recv(self):
+        if not self._inbox:
+            await asyncio.sleep(3600)  # block forever; tests cancel
+        return self._inbox.pop(0)
+
+    async def close(self):
+        pass
+
+
+class _FakeCallbacks:
+    """Implements client.session.Callbacks via duck-typing."""
+    def __init__(self):
+        self.started: list = []
+        self.audio: list = []
+        self.assistant_done = 0
+        self.session_ended: list = []
+
+    def on_session_started(self, conversation_id: str) -> None:
+        self.started.append(conversation_id)
+
+    def on_assistant_audio(self, pcm16: bytes) -> None:
+        self.audio.append(pcm16)
+
+    def on_assistant_done(self) -> None:
+        self.assistant_done += 1
+
+    def on_session_ended(self, reason: str) -> None:
+        self.session_ended.append(reason)
+
+
+async def _run_pump_briefly(session, ws, settle: float = 0.05):
+    task = asyncio.create_task(session._pump(ws))
+    await asyncio.sleep(settle)
+    task.cancel()
+    try:
+        await task
+    except (asyncio.CancelledError, BaseException):
+        pass
+
+
+async def test_hello_is_sent_on_connect_and_hello_ack_logged():
+    ws = FakeWebSocket(inbox=[
+        json.dumps({"type": "hello_ack", "config": {"voice": "alloy"}}),
+    ])
+    session = Session(
+        backend_ws="wss://x/device",
+        device_token="T",
+        device_id="00000000-0000-0000-0000-000000000001",
+        callbacks=_FakeCallbacks(),
+    )
+    await _run_pump_briefly(session, ws)
+
+    first = json.loads(ws.sent[0])
+    assert first["type"] == "hello"
+    assert first["device_id"] == "00000000-0000-0000-0000-000000000001"
+    assert first["client_version"]
+
+
+async def test_session_started_calls_callback_with_conversation_id():
+    cb = _FakeCallbacks()
+    ws = FakeWebSocket(inbox=[
+        json.dumps({"type": "hello_ack", "config": {}}),
+        json.dumps({
+            "type": "session_started",
+            "conversation_id": "cccccccc-cccc-cccc-cccc-cccccccccccc",
+        }),
+    ])
+    session = Session("wss://x/device", "T", "did", cb)
+    await _run_pump_briefly(session, ws)
+    assert cb.started == ["cccccccc-cccc-cccc-cccc-cccccccccccc"]
+
+
+async def test_binary_frame_goes_to_audio_callback():
+    cb = _FakeCallbacks()
+    payload = bytes(range(256)) * 15  # 3840 bytes
+    ws = FakeWebSocket(inbox=[
+        json.dumps({"type": "hello_ack", "config": {}}),
+        payload,
+    ])
+    session = Session("wss://x/device", "T", "did", cb)
+    await _run_pump_briefly(session, ws)
+    assert cb.audio == [payload]
+
+
+async def test_assistant_done_and_session_ended_callbacks():
+    cb = _FakeCallbacks()
+    ws = FakeWebSocket(inbox=[
+        json.dumps({"type": "hello_ack", "config": {}}),
+        json.dumps({"type": "assistant_done"}),
+        json.dumps({"type": "session_ended", "reason": "idle"}),
+    ])
+    session = Session("wss://x/device", "T", "did", cb)
+    await _run_pump_briefly(session, ws)
+    assert cb.assistant_done == 1
+    assert cb.session_ended == ["idle"]
+
+
+async def test_tool_call_runs_handler_and_sends_tool_result():
+    cb = _FakeCallbacks()
+    ws = FakeWebSocket(inbox=[
+        json.dumps({"type": "hello_ack", "config": {}}),
+        json.dumps({
+            "type": "tool_call",
+            "call_id": "c1",
+            "name": "set_volume",
+            "arguments": {"level": 42},
+        }),
+    ])
+    session = Session("wss://x/device", "T", "did", cb)
+    captured = {}
+
+    def handler(args):
+        captured["args"] = args
+        return {"ok": True, "level": args["level"]}
+
+    session.register_tool("set_volume", handler)
+    await _run_pump_briefly(session, ws)
+
+    assert captured["args"] == {"level": 42}
+    reply = json.loads(ws.sent[-1])
+    assert reply["type"] == "tool_result"
+    assert reply["call_id"] == "c1"
+    assert reply["ok"] is True
+    assert reply["result"] == {"ok": True, "level": 42}
+
+
+async def test_tool_call_with_exception_sends_ok_false():
+    cb = _FakeCallbacks()
+    ws = FakeWebSocket(inbox=[
+        json.dumps({"type": "hello_ack", "config": {}}),
+        json.dumps({
+            "type": "tool_call",
+            "call_id": "c2",
+            "name": "broken",
+            "arguments": {},
+        }),
+    ])
+    session = Session("wss://x/device", "T", "did", cb)
+
+    def handler(args):
+        raise RuntimeError("kaboom")
+
+    session.register_tool("broken", handler)
+    await _run_pump_briefly(session, ws)
+
+    reply = json.loads(ws.sent[-1])
+    assert reply["type"] == "tool_result"
+    assert reply["call_id"] == "c2"
+    assert reply["ok"] is False
+    assert "kaboom" in reply["error"]
+
+
+async def test_unknown_tool_sends_ok_false():
+    cb = _FakeCallbacks()
+    ws = FakeWebSocket(inbox=[
+        json.dumps({"type": "hello_ack", "config": {}}),
+        json.dumps({
+            "type": "tool_call",
+            "call_id": "c3",
+            "name": "ghost",
+            "arguments": {},
+        }),
+    ])
+    session = Session("wss://x/device", "T", "did", cb)
+    await _run_pump_briefly(session, ws)
+    reply = json.loads(ws.sent[-1])
+    assert reply["ok"] is False
+    assert "ghost" in reply["error"]

client/tests/test_state.py

@@ -0,0 +1,127 @@
+from client.state import State, StateMachine
+
+
+def test_starts_idle_with_wakeword_enabled_and_uplink_disabled():
+    sm = StateMachine()
+    assert sm.state is State.IDLE
+    assert sm.wakeword_enabled
+    assert not sm.uplink_enabled
+
+
+def test_wake_fired_goes_idle_to_wake_pending_and_disables_wakeword():
+    sm = StateMachine()
+    assert sm.wake_fired() is True
+    assert sm.state is State.WAKE_PENDING
+    assert not sm.wakeword_enabled
+    assert not sm.uplink_enabled
+
+
+def test_wake_fired_is_idempotent_when_not_idle():
+    sm = StateMachine()
+    sm.wake_fired()
+    assert sm.wake_fired() is False
+    assert sm.state is State.WAKE_PENDING
+
+
+def test_session_started_goes_wake_pending_to_listening_and_unmutes_uplink():
+    sm = StateMachine()
+    sm.wake_fired()
+    sm.session_started()
+    assert sm.state is State.LISTENING
+    assert sm.uplink_enabled
+    assert not sm.wakeword_enabled
+
+
+def test_assistant_audio_arrived_mutes_uplink():
+    sm = StateMachine()
+    sm.wake_fired()
+    sm.session_started()
+    sm.assistant_audio_arrived()
+    assert sm.state is State.ASSISTANT_SPEAKING
+    assert not sm.uplink_enabled
+
+
+def test_assistant_audio_arrived_is_noop_outside_listening():
+    sm = StateMachine()
+    sm.assistant_audio_arrived()
+    assert sm.state is State.IDLE  # ignored from IDLE
+    sm.wake_fired()
+    sm.assistant_audio_arrived()
+    assert sm.state is State.WAKE_PENDING  # ignored from WAKE_PENDING
+
+
+def test_assistant_done_returns_to_listening_from_speaking():
+    sm = StateMachine()
+    sm.wake_fired()
+    sm.session_started()
+    sm.assistant_audio_arrived()
+    sm.assistant_done()
+    assert sm.state is State.LISTENING
+    assert sm.uplink_enabled
+
+
+def test_assistant_done_is_noop_when_not_speaking():
+    sm = StateMachine()
+    sm.wake_fired()
+    sm.session_started()  # LISTENING; never went to ASSISTANT_SPEAKING
+    sm.assistant_done()
+    assert sm.state is State.LISTENING
+
+
+def test_session_ended_routes_through_idle_pending_then_idle():
+    sm = StateMachine()
+    sm.wake_fired()
+    sm.session_started()
+    sm.session_ended()
+    assert sm.state is State.IDLE_PENDING
+    assert not sm.wakeword_enabled
+    assert not sm.uplink_enabled
+    sm.sleep_played()
+    assert sm.state is State.IDLE
+    assert sm.wakeword_enabled
+
+
+def test_session_ended_works_from_assistant_speaking():
+    sm = StateMachine()
+    sm.wake_fired()
+    sm.session_started()
+    sm.assistant_audio_arrived()
+    sm.session_ended()
+    assert sm.state is State.IDLE_PENDING
+
+
+def test_session_ended_from_idle_is_noop():
+    sm = StateMachine()
+    sm.session_ended()
+    assert sm.state is State.IDLE
+
+
+def test_sleep_played_outside_idle_pending_is_noop():
+    sm = StateMachine()
+    sm.sleep_played()
+    assert sm.state is State.IDLE
+
+
+def test_force_idle_from_listening_resets_state_and_reenables_wakeword():
+    sm = StateMachine()
+    sm.wake_fired()
+    sm.session_started()
+    assert sm.state is State.LISTENING
+    sm.force_idle()
+    assert sm.state is State.IDLE
+    assert sm.wakeword_enabled
+
+
+def test_force_idle_from_assistant_speaking_resets():
+    sm = StateMachine()
+    sm.wake_fired()
+    sm.session_started()
+    sm.assistant_audio_arrived()
+    sm.force_idle()
+    assert sm.state is State.IDLE
+
+
+def test_force_idle_on_idle_is_noop():
+    sm = StateMachine()
+    sm.force_idle()
+    assert sm.state is State.IDLE

client/wakeword.py

@@ -0,0 +1,52 @@
+"""openwakeword wrapper for the Pi client.
+
+Loads the stock `alexa` model. Each call to `predict()` takes a 1280-sample
+16 kHz mono int16 frame (produced by `client.audio.resample_24k_to_16k`).
+Returns True at most once per cooldown window.
+"""
+import time
+
+import numpy as np
+
+from client.log import get_logger
+
+_log = get_logger("wakeword")
+
+WAKEWORD = "alexa"
+DEFAULT_THRESHOLD = 0.5
+DEFAULT_COOLDOWN_S = 1.5
+
+
+class WakewordDetector:
+    def __init__(self, *, threshold: float = DEFAULT_THRESHOLD,
+                 cooldown_s: float = DEFAULT_COOLDOWN_S) -> None:
+        # Lazy-import openwakeword to keep the test suite importable without it.
+        import openwakeword.utils
+        from openwakeword.model import Model
+
+        _log.info("downloading openwakeword models (idempotent)")
+        # No-arg form dodges the None-handling crash documented in findings.md §4.
+        openwakeword.utils.download_models()
+        _log.info("loading openwakeword model %r", WAKEWORD)
+        t0 = time.monotonic()
+        self._model = Model(wakeword_models=[WAKEWORD], inference_framework="onnx")
+        _log.info("openwakeword model loaded in %.2fs", time.monotonic() - t0)
+
+        self._threshold = threshold
+        self._cooldown_s = cooldown_s
+        self._last_fired = 0.0
+
+    def predict(self, frame_16k_int16: np.ndarray) -> bool:
+        scores = self._model.predict(frame_16k_int16)
+        score = float(scores.get(WAKEWORD, 0.0))
+        rms = float(np.sqrt(np.mean(frame_16k_int16.astype(np.float32) ** 2)))
+        now = time.monotonic()
+        # Diagnostic: log any non-trivial score so we can see what's getting
+        # close to firing.
+        if score >= 0.1:
+            _log.info("score=%.3f rms=%.0f", score, rms)
+        if score >= self._threshold and (now - self._last_fired) >= self._cooldown_s:
+            self._last_fired = now
+            _log.info("wakeword fired score=%.3f rms=%.0f", score, rms)
+            return True
+        return False

requirements-dev.txt

@@ -0,0 +1,3 @@
+pytest>=8.0
+pytest-asyncio>=0.23
+responses>=0.25