Gateway

• Added runtime-adjustable beat-mode onset sensitivity for venue tuning; the
  default maps to the real-device verified 0.004 RMS floor, and metadata
  snapshots report both the selected sensitivity and effective floor. (#301)
• Added gateway-only beat mode MCP tools for continuous ambient audio capture
  through the existing listen wire path, dependency-free BPM estimation,
  beat-synced head sway/base-ring LED flashes, polling metadata snapshots, and
  WAV clip export from a bounded rolling buffer. (#301)
• Beat mode now requests the raw listen capture profile so ambient music
  reaches the gateway without device-side speech AFE suppression. (#349)
• Lowered the beat tracker's minimum onset RMS floor (0.025 → 0.004 full
  scale) to match real-device microphone levels; onset detection now works
  against actual ambient music (verified on device). The adaptive threshold
  continues to prevent false onsets in quiet rooms. (#301)

Highlights

• Raw microphone capture profile (#349, #351): the inbound listen
  wire message accepts an optional "profile": "voice" | "raw" field.
  The default voice profile behaves exactly as before (speech-tuned
  AFE path for STT). The new raw profile streams pre-AFE microphone
  PCM through the existing Opus audio path, so ambient music reaches
  the gateway at natural levels — this is what enables the gateway-side
  beat mode (gateway v0.17.0) to detect BPM from live music and drive
  beat-synced head sway and LED expression.

Pair release with gateway v0.17.0
(pipx install --force stackchan-mcp or uv tool upgrade stackchan-mcp),
which adds the beat-mode MCP tools that use this profile.

Flashing

• xiaozhi.bin — app-only update at offset 0x20000 (preserves NVS:
  Wi-Fi credentials, gateway URL settings, avatar assets)
• merged-binary.bin — full clean install at 0x0 (resets NVS)
• v2.2.6_stackchan.zip — full artifact set

License

Distributed under MIT (canonical build, MIT end-to-end).

Notes

• Unknown or missing profile values fall back to voice; older
  gateways that never send the field see unchanged behavior.
• Verified on device: beat mode over the raw profile detected
  76.7 BPM at 0.77 confidence from ambient music at ordinary indoor
  volume, driving motion and LED commands.

Gateway

• Added a color_order option (grb default, rgb) for Port B/C WS2812
  gateway tools and stackchan_follow_led_stream, allowing RGB-wired LEDs to
  render correct colors by swapping R/G in the gateway before relay. (#343)
• Added stackchan_follow_led_stream, a gateway-side WebSocket LED-frame
  subscriber for driving the base ring or a Port B WS2812 strip from external
  event / continuous color frames. (#335)
• Log ESP32 WebSocket disconnect close codes, reasons, close class, last-frame
  age, and connection lifetime, and make the gateway keepalive policy explicit.
  (#338)
• Exposed the Port C WS2812 device tools and added port_c as a
  stackchan_follow_led_stream target. (#340)

Highlights

• Port C (GPIO 17) WS2812 support — the second rear Grove port can now drive a WS2812-compatible LED strip via five new MCP tools self.port_c.ws2812.{init,set_pixel,set_strip,refresh,clear}, mirroring the existing Port B set (#340, #341). Both ports can run strips simultaneously on separate RMT channels.
• Fully opt-in: nothing touches GPIO 17 until self.port_c.ws2812.init is explicitly called; existing setups using Port C for other purposes are unaffected.

Pairs with gateway v0.16.0, which exposes the Port C tools, adds a port_c target to stackchan_follow_led_stream, and adds a color_order option for RGB-wired NeoPixels (e.g. 8 mm through-hole units).

Files

• merged-binary.bin — clean install (flashes everything, resets NVS)
• xiaozhi.bin — app-only update to 0x20000 (preserves NVS / Wi-Fi settings)
• v2.2.6_stackchan.zip — full artifact set

License

Distributed under MIT (canonical build path; the optional GPL-3.0 SCServo fallback remains opt-in only, see CONTRIBUTING.md).

Gateway

• Re-dispatch the emoji-selected avatar face after successful speech playback,
  so emoji+text say calls keep the expression visible after lip-sync stops.
  (#296)

Pairing

Pairs with firmware-v1.14.0 (released the same day), which adds an on-screen
listening indicator on the stack-chan LCD during STT capture (#332). The two
changes are independent — this gateway version also works with earlier firmware
unchanged.

Highlights

• On-screen listening indicator (#332): the stack-chan LCD now shows a small pulsing red dot while STT capture is active, so it is visible at a glance that the device is listening. Previously a listen window gave no visual cue on the device itself. The avatar render path (face / blink / lip-sync) is untouched. Verified on hardware (details on PR #333).

Pairing

Pairs with Gateway v0.15.0 (released the same day), which re-dispatches the emoji-selected avatar face after successful speech playback so emoji+text say calls keep the expression visible after lip-sync stops (#296). The two changes are independent — earlier gateway versions continue to work with this firmware unchanged.

License

Firmware binaries in this release are distributed under MIT (canonical build; the optional GPL servo-library fallback remains opt-in and is not part of these artifacts).

Assets

• merged-binary.bin — clean install (flash at 0x0, resets NVS / WiFi settings)
• xiaozhi.bin — app-only update (flash at 0x20000, preserves NVS / WiFi settings)
• v2.2.6_stackchan.zip — full build output

Highlights

• Server-driven listen() no longer drops the connection (#328): requesting listening mode on an already connected WebSocket now arms the logical audio session in place instead of tearing down and rebuilding the socket. Previously, any gateway-side listen() call issued while the device sat idle caused an immediate disconnect/reconnect and returned zero audio frames; the same rebuild also affected touch and wake-word activations from the idle state. Verified on hardware: repeated listen windows over a single uninterrupted connection, with audio frames flowing end to end (details on PR #330).

Pairing

No paired gateway release — this is a firmware-only patch. Recent gateway versions work unchanged (verified against gateway v0.14.0).

License

Firmware binaries in this release are distributed under MIT (canonical build; the optional GPL servo-library fallback remains opt-in and is not part of these artifacts).

Assets

• merged-binary.bin — clean install (flash at 0x0, resets NVS / WiFi settings)
• xiaozhi.bin — app-only update (flash at 0x20000, preserves NVS / WiFi settings)
• v2.2.6_stackchan.zip — full build output

Gateway

• Auto-render the idle avatar after a new ESP32 device session finishes
  initialization and tool discovery, unless set_avatar was already sent
  on that connection. (#77)
• Fail in-flight load_avatar_set calls with disconnected immediately
  when the ESP32 connection drops, instead of waiting for the avatar
  load timeout. (#228)
• Scoped the gateway ownership lock per WS port (owner-<ws_port>.lock
  instead of the machine-global owner.lock), so one gateway per device
  can run on separate WS ports on the same host without the second
  gateway being rejected at startup. Single-instance use is unaffected
  (the default lock is simply named owner-8765.lock) and a duplicate
  start on the same port is still rejected; --check reads the per-port
  lock. (#320)
• Fixed /capture rejecting photo uploads with HTTP 413 on aiohttp >= 3.14.
  The capture app raised the per-request body cap with client_max_size=0
  (so /pcm can stream long PCM), but aiohttp's multipart reader treats 0
  as a zero-byte limit — unlike request.read()/.post() — so every upload
  carrying a non-empty question field was rejected with 413. Use a large
  finite client_max_size instead; /capture's real limit stays the explicit
  per-route byte cap.
• Hardened /capture to tolerate a non-UTF-8 question field (decode with
  errors="replace") instead of failing the upload with HTTP 500.

BREAKING: stackchan_follow_pose_stream.downsample_hz schema maximum is now 20 (was 60). Callers that previously set values above 20 will be rejected with "downsample_hz must be a number in (0, 20]". The schema default (20) is unchanged, so calls that omit the argument behave the same. Rationale in the entry below. (#315)

Gateway

• Added optional user-local user-defaults.toml support for gateway-side
  MCP argument defaults, starting with stackchan_follow_pose_stream. Explicit
  MCP call arguments still take precedence, while absent, empty, or invalid
  config files fall back to schema defaults. (#311)
• Added set_touch_sensor_enabled and get_touch_sensor_enabled MCP
  wrappers for the firmware head-touch enable flag. (#312)
• stackchan_follow_pose_stream now exposes smoothing_window as an
  MCP tool argument (integer, default 5, range 1..20; 1 = passthrough).
  Callers whose upstream pose source already applies smoothing can
  disable the redundant gateway-side moving average. Omitting the
  argument preserves the previous behaviour. (#309)
• Lowered stackchan_follow_pose_stream.downsample_hz schema maximum
  from 60 to 20 to match the SCS0009 servo's observed sustained
  WritePos rate. Continuous command rates above ~20 Hz triggered
  UART hangs during real-device dogfood. BREAKING: callers that
  previously set values above 20 will now be rejected with
  "downsample_hz must be a number in (0, 20]". The schema default
  (20) is unchanged, so calls that omit the argument behave the
  same. (#315)
• Added edge-tts TTS engine — a subprocess-based engine using
  Microsoft's Edge TTS CLI and ffmpeg, registered as edge-tts.
  Provides natural English and multilingual voices (voicevox is
  Japanese-only by default). Setup: install edge-tts CLI on PATH
  (e.g. pip install edge-tts) plus ffmpeg. Default voice is
  en-GB-SoniaNeural, overridable per-call via the new
  say(speaker_name=...) argument (added to the say() schema as a
  string field, distinct from speaker_id integer and the voice
  engine selector) or globally via the STACKCHAN_EDGE_TTS_DEFAULT_VOICE
  env var. (#317)
• Exposed the optional Port A I2C scl_speed_hz argument in the gateway
  schemas for i2c_read, i2c_write, and i2c_write_read, matching the
  firmware-side 100000..1000000 Hz range so schema-driven MCP clients can
  discover slower per-transaction I2C clocks. Behaviour is unchanged when
  omitted. (#321)

Pairs with firmware-v1.13.0 (released the same day) for the firmware-side touch sensor disable (#312) and Port A I2C scl_speed_hz (#319).

Install / upgrade: pipx install --force stackchan-mcp

Highlights

• User-configurable touch sensor disable (#312): new NVS-backed touch.enabled field with self.robot.set_touch_sensor_enabled / self.robot.get_touch_sensor_enabled MCP tools. Disabling stops both local touch reactions and stackchan/event emission, and the setting persists across reboots.
• Per-transaction I2C clock control (#319): the Port A I2C tools (self.i2c.read / self.i2c.write / self.i2c.write_read) accept an optional scl_speed_hz property (default 400000, range 100000–1000000), so slow Units — e.g. the RCWL-9620 ultrasonic ranger, which fails transfers with ESP_ERR_INVALID_STATE at 400 kHz — can be driven without recompiling the firmware. Behaviour is unchanged when the property is omitted.

Pairing

Pairs with Gateway v0.13.0 (released the same day), which exposes the set_touch_sensor_enabled / get_touch_sensor_enabled wrappers and declares scl_speed_hz on the gateway-side I2C tool schemas so schema-driven MCP clients can discover it. Earlier gateway versions continue to work; the I2C clock argument still passes through, it just is not advertised in their schemas.

License

Firmware binaries in this release are distributed under MIT (canonical build; the optional GPL servo-library fallback remains opt-in and is not part of these artifacts).

Assets

• merged-binary.bin — clean install (flash at 0x0, resets NVS / WiFi settings)
• xiaozhi.bin — app-only update (flash at 0x20000, preserves NVS / WiFi settings)
• v2.2.6_stackchan.zip — full build output