Every figure in the latency budget was stale, in both directions. TTS was
listed at ~1.9 s per sentence but measures ~0.24 s warm for 4.5 s of audio;
the full Tatlock flow was listed at 11-25 s but measures ~10-13 s for simple
turns. Both sets of numbers predate the current model.
The VRAM section now carries real figures and the reason they matter: on
2026-08-07 Tatlock ran against a 9.3 GB model, leaving 7 MiB free, and every
transcription failed with CUDA out of memory while the Speaches container
still reported healthy. The budget is the constraint, not slack.
Also replaces the retired tatlock.schweitz.internal hostname in the topology
diagram with the docker container name.
Co-Authored-By: Claude <noreply@anthropic.com>
The homelab is retiring the *.schweitz.internal domain; in-network
machine-to-machine traffic uses docker container names on the
docker-dataplane network. The deployed tatlock-ui stack already overrides
DESKLOCK_TATLOCK_BASE_URL (Tatlock runs on the host), so only the
fallback default changes.
Docs follow: AGENTS.md M2M guidance now points at container names with
*.schweitz.net reserved for browsers, architecture.md drops the retired
domain (the registry name now matches what CI actually pushes since
c477019), and the README diagram loses the stale hostname.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Recognize simple device commands in the transcript and act on them without
a Tatlock round-trip. commands.match() maps volume up/down, mute/unmute,
"set volume to N", and "goes to eleven"/max to a "command" message sent
straight to the device; the utterance never reaches the LLM. Matching is
deliberately precise so real requests ("set an alarm for a quarter to
eleven") are not hijacked. Adds the "command" message to the device
protocol in docs/architecture.md.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LKPbR6DY2JygHbyLjxm7Uu
The device now runs the complete face contract on glass: seven states
with glyph expressions (custom 140px face font), matrix rain as pooled
label streams (22px DejaVu+Noto katakana font), rage kaomoji frames
(64px), orbit arc + elapsed counter wait cues, blink/talk/thought
animations, idle clock (SNTP, Europe/Amsterdam), and the power ladder
(active/ambient 35%/dormant 5% with rain parked).
Voice path: ES7210 mic capture task streams 16k PCM over
esp_websocket_client to the gateway; reply PCM buffers to PSRAM and
plays via the audio task; touch-to-talk (tap to speak, tap to send).
Protocol mapping per docs: thinking->effort, audio->speaking,
error event->rage (auto-composes after 2 loops), WS loss->error face.
Bring-up fixes: 8MB factory partition (fonts overflowed 1.5M),
bsp_display_lock(0) is try-lock in this adapter (use UINT32_MAX),
gong synth yields to feed IDLE0, wifi scan diagnostics (found SSID
case mismatch), esp_websocket_client pinned ~1.3 (1.4 needs IDF>5.5).
Verified on hardware: boots, connects to Wi-Fi and holds an open
WebSocket to the gateway. Sauron (iMac TTY endpoint + ops console)
planned in architecture.md.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
User verdict: audible and butler-non-intrusive on the bare 2W speaker.
Recorded as the house sound signature in architecture.md.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- firmware boots to a rendered LVGL face in 1.57s (eyes + smile + tag)
- sdkconfig: CONFIG_IDF_EXPERIMENTAL_FEATURES=y unlocks SPIRAM 200MHz
(silently degraded to 20MHz before -> MIPI-DSI underrun -> LVGL lock
starvation -> task watchdog); mirrors official 08_lvgl_demo_v9 config
- architecture.md: power management section (user prime concern) —
active/ambient/dormant/night ladder, levers, hard edges, <=300ms wake
- AGENTS/CLAUDE: replace stale bring-up warnings with verified commands
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The gateway service now lives in system-admin-toj/containers/stacks/
tatlock-ui.yml (shared docker-dataplane network: speaches by service
name, host-run Tatlock by LAN IP). Drop deploy/desklock-gateway.yml.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Settles the Python version question: floor >=3.11, no ceiling.
Container moves to python:3.13-slim, CI tests on 3.13. The embedded
Piper fallback resamples with numpy (already present via [speech]).
Verified with a live round trip at 16 kHz.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- Face: new 'rage' state — 3-frame kaomoji loop (stare, flip the table,
put it back) for in-flight request failures; 'error' stays the quiet
persistent face for a dead link. Sim + artifact + design doc updated.
- Gateway: stt.py/tts.py are now pluggable backends. Default 'speaches'
talks OpenAI-format HTTP to the live container on :8601
(faster-whisper-small STT, Kokoro bm_george TTS with 24->16 kHz
audioop resample); 'embedded' fallback kept behind the [speech] extra.
Verified with a live TTS->STT round trip (warm: STT 0.27s, TTS 1.9s).
Docker image is now slim (no CUDA/ML deps). Python pinned to 3.12
(system 3.8 too old, audioop gone in 3.13).
- CI: .gitea/workflows/build.yml — lint+test on main pushes; on v* tags
test, build gateway image, push to registry, release, and trigger
Watchtower (tatlock pattern; needs REGISTRY_USER/REGISTRY_PASSWORD/
WATCHTOWER_TOKEN secrets). Runtime stack in deploy/desklock-gateway.yml.
- architecture.md: measured speech latencies, deployed-Speaches status,
CI & deployment section.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- sim/face/index.html: browser simulator of the 800x800 round panel with
six states (idle/listening/pensive/effort/speaking/error), state-driven
rain density, blink/talk/thought animations, idle clock
- effort state gets hard-required wait cues: orbiting bezel arc, elapsed
counter, max rain (Tatlock turns run 10-25s)
- rain driven by setInterval, not rAF: renders under the screenshot
tool's --virtual-time-budget and mirrors LVGL lv_timer
- architecture.md: Face design contract (state table, protocol mapping,
LVGL port notes) + latency table updated to GPU-era benchmarks
(Steward ~6s warm, full flow 11-25s; old CPU-era figures retired)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- STT/TTS move to a shared Speaches container (OpenAI-format, GPU, port
8601 proposed); gateway becomes a thin orchestrator with pluggable
speech backends (speaches default, embedded fallback)
- Record the on-device ceiling: WakeNet wake word, VAD, ES7210 AEC,
optional MultiNet fixed commands; open-vocabulary STT permanently out
- Record the real latency bottleneck (Tatlock ~2 min full local flow):
gateway must stream chat tokens and synthesize sentence-by-sentence
- Plan reply_delta + barge-in protocol additions
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
DeskLock gives the Tatlock butler a face and voice on a Waveshare
ESP32-P4-WIFI6-Touch-LCD-3.4C round display in the living room.
- firmware/: ESP-IDF project targeting esp32p4 with the Waveshare XC BSP
- gateway/: FastAPI voice bridge (faster-whisper STT, Tatlock chat, Piper TTS)
- docs/architecture.md: component design and device<->gateway WS protocol
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
jpmschweitzerandClaude