Text + Voice Demo Baseline physical validation — 2026-10-04, Asia/Shanghai

DEMO BASELINE: PHYSICAL FAIL
UI / Wearable Shell 0.6a gate: NOT PASSED
Speech output: FROZEN. No Read Screen, AUTO_REPLY, AUTO_SUMMARY, buffered playback,
or streaming playback was invoked on hardware. Speech playback start count: 0.

Baseline branch: codex/demo-text-voice-baseline-0.5.1
Baseline / Gateway source: 0f4f47ccbe79142a1b8cc324c62b856ef6433ff6
Firmware behavior source: 6bffe571b7c1b61cfdb168b0f7a03737bad394f1
Tracked firmware / boards diff against 6bffe57: EMPTY
Firmware VERSION: 0.5.1
Firmware provisioning: original baseline profile, user-confirmed powellli hotspot.
An unsupported legacy CONFIG_DEVICE_LAB_SPEECH_TRACE line was removed from the
private provisioning file. No tracked firmware/server source was modified.
This is a newly rebuilt candidate from physically proven source, not a
byte-identical copy of the historical physical-pass binary.

merged-flash.bin: 2,123,712 bytes
SHA256: 099ac2b8d25bf1c04fb5181344d59466fe34b509eb50377486821d8e30facf0b
firmware.bin: 2,058,176 bytes
SHA256: dff6d7cf4a4bacf86f8806f68407a53c4d066167a7b88fe5eb676e5e82482dbb
Mac SHA256: MATCHES LIGHTHOUSE
Mac artifact: /Users/lili/workspace/esp32-dev/.cache/demo-baseline-validation-20261004/firmware-hotspot
Serial port: /dev/cu.usbmodem1101
Chip: ESP32-S3 revision v0.2, 16 MB Flash, 8 MB PSRAM
Flash: PASS, esptool write hash verified

Gateway PID: 419467
Gateway WorkingDirectory: /home/lighthouse/esp32-dev/.cache/demo-text-voice-baseline-0.5.1
Gateway ExecStart: same worktree/scripts/run-server.sh
Gateway status: active/running
Gateway start: 2026-10-04 20:56:13 CST
Gateway NRestarts: 0
Gateway health: PASS
Speech Mode for restored target: OFF
Restored Pi target: agent:pi:01a0fcd2-7b96-77fb-a57a-40bbfb9fe5da
Maintenance restart count: 1. No subsequent Gateway restart.
buffered-experiment.conf: saved and moved outside the active drop-in directory.
All Speech experiment branches and worktrees: preserved. No merge into main.
Remote evidence: /home/lighthouse/esp32-dev/.cache/demo-baseline-validation-20261004-PSHmES

Software validation, rerun on this baseline:
Python: 138 PASS, one existing deprecation warning
Node: 18 PASS
Firmware host helpers / render: PASS
HTTP / WebSocket smoke: PASS
Pi / Codex authenticated TLS WSS Voice smoke: PASS, FAKE STT
ESP-IDF v5.5.4 build: PASS
git diff --check: PASS
Restored server health/Admin/JS/server/status pages: HTTP 200
Existing Pi control settings endpoint: HTTP 200

Boot:
Wi-Fi connected: PASS
Cloud connected / hello acknowledged: PASS
Device online / original Agent target restored: PASS
Speech OFF: PASS
PI / IDLE screen: user confirmed normal
Basic monitor pages: user confirmed normal, returned to PI

Idle 5 min: PASS
Window: 21:05:02–21:10:02, 300.1 seconds, 61 samples
Unexpected offline / WSS reconnect / reboot during this window: 0 / 0 / 0
Gateway active throughout; PID unchanged; NRestarts=0

Text Agent: PASS for displayed and delivered behavior
RUNNING: PASS
Elapsed timer: PASS
Chinese activity: PASS
Tool activity: PASS
A APPROVE: PASS, physical A click while approval=1
B REJECT: PASS, separate harmless sleep 1 request, physical B click while approval=1
DONE: PASS
final_text: exact backend match “检查完成”, 12 UTF-8 bytes; user confirmed display
Chinese rendering: PASS, user confirmed
Reject task final_text: exact backend match “已取消”, 9 UTF-8 bytes; user confirmed
Notification vibration: completed; minimum free stack 2016–2028 bytes

Voice 1: complete single-voice physical path PASS, after an earlier WSS reconnect
Voice ID: 58c10ac51284da79
VOICE_RX start: device serial voice_start sent=1; Gateway INFO marker NOT EMITTED
VOICE_RX end: device serial voice_end sent=1; Gateway INFO marker NOT EMITTED
Upload: 46 chunks, 162,288 PCM bytes; capture failure=none
Maximum send time: 282 ms
STT: successful transcript was delivered; Gateway STT INFO marker NOT EMITTED
STT provider duration: UNKNOWN, not present in journal
Transcript: displayed; user confirmed correct; no transcript content saved here
A SEND: PASS, voice_decision queued and voice_state=SENT
Pi received: PASS, user explicitly confirmed receipt and response
Agent reply: DONE, final_bytes=307; user confirmed device reply display

Voice repeated:
1: NOT RUN
2: NOT RUN
3: NOT RUN
Post-Voice idle 3 min: NOT RUN
Stopped additional physical tests after the earlier unexpected disconnect was
discovered. The first disconnect was noticed retrospectively while examining
the single-voice trace; the prior “still no disconnect” progress statement was
explicitly corrected. No further recording was requested.

Unexpected offline: 1
WSS reconnect: 1
Wi-Fi disconnect: 0
Reboot: 0
Stack overflow: 0
Watchdog / heap failure: 0 observed
These counts cover the final hotspot baseline serial evidence, not the earlier
wrong-network boot/preflight or planned flash resets.

Exact failure boundary:
WSS/TLS receive failed during idle after the successful Text task and before
Voice recording. This is not an STT failure. Exact initiating root cause and
Gateway WebSocket close code/reason remain UNKNOWN.
First connection lasted about 504 seconds. DONE preceded the failure by 60.8 s.
Last successful application heartbeat was 27.8 s before the receive error;
configured application heartbeat interval is 30 s. No explicit no-PONG failure
was logged. RSSI was -38 dBm, PS=1, internal_free=97935, dma_largest=31744.
No Wi-Fi disconnect, heap failure, or Gateway process restart was observed.
The current evidence does not distinguish hotspot/public route stalls from
ESP32 TLS/WebSocket receive behavior, or identify which peer initiated closure.

First relevant serial log:
E (511373) transport_base: esp_tls_conn_read error, errno=No more processes
W (511373) transport_ws: Requested to read 2, actually read 0 bytes
E (511373) transport_ws: Error read data(-1)
E (511373) websocket_client: esp_transport_read() failed with -1, transport_error=ESP_OK, tls_error_code=0, tls_flags=0, errno=11
I (511373) CLOUD: [CLOUD] link reason=receive-error bytes=728 parts=1 elapsed_ms=10004 rssi=-38 ps=1 internal_free=97935 dma_largest=31744
W (511403) CLOUD: [CLOUD] disconnected
I (519603) CLOUD: [CLOUD] connected

First relevant Gateway log:
2026-10-04T21:12:55+08:00 device-gateway[419467]: INFO: device esp32s3-28848543a200 offline
2026-10-04T21:12:55+08:00 device-gateway[419467]: INFO: connection closed
2026-10-04T21:13:03+08:00 device-gateway[419467]: INFO: device esp32s3-28848543a200 online board=m5stack-stopwatch firmware=0.5.1
No close code/reason or corresponding Cloudflare service error was logged in
the inspected 21:12:40–21:13:10 window.

Additional diagnostic gap:
voice.py uses logger device_gateway.voice, outside the configured uvicorn.error
logger hierarchy. With this launcher its effective level is WARNING; INFO is
disabled. Therefore existing [VOICE_RX]/[VOICE_STT] lifecycle markers were
suppressed. The failure is in diagnostic visibility, not proof of failed STT.

Next smallest diagnostic action:
Enable the existing content-free Voice INFO diagnostics and record Gateway
WebSocket close code/reason, then repeat only a same-firmware Text-to-idle
connection observation. Correlate closure with serial receive-error/heartbeats.
Do not change PS, TCP buffers, TX locking, timeout values, STT, or Speech output.
No such code/deployment change was made during this run.

Preservation:
Mac tracked source unchanged; config/wifi-secondary.local.json preserved.
Historical physical-pass merged SHA still present unchanged:
673ba20d8d010a562e441c7f136ec48afcb0853f45d684e40c489465c0009cbf
Factory backup SHA before/after matches:
910416566f113205d7eddc9739c4ea9d7cf0e5dec7dab20c73e4a269f7a6b33a
Serial monitor and Gateway journal capture remain active for follow-up.