psycheros-v0.8.21

Fixed

• Voice: stop the "sent" cue tone firing mid-speaking in vanilla mode. TTS audio
  leaking back into the mic triggered the browser VAD during the entity's turn,
  which forced the pipeline state to recording mid-response and let the next
  user_silence run a new turn on top of the in-flight one. pushAudio now
  drops frames while the entity is processing/speaking, and the
  user_speech_start handler keeps the userSpeaking flag (Pulse draining) but
  no longer transitions state out of a mid-response turn.
• Voice (playback): TTS chunk scheduling moved off the JS event loop. Each chunk
  now starts at a tracked nextStartTime (sample-accurate via
  source.start(nextStartTime)) instead of waiting for the previous source's
  onended, which fires 1–5ms late and produced a click at every chunk boundary
  — audible as "crackling fire" on Bluetooth headsets and small-chunk providers
  like ElevenLabs.

---

How to install

The canonical artifact for the harness is the container image:

docker pull ghcr.io/psycherosai/psycheros:<version>
# or `:latest` for the most recent release

See the README's Docker block
for the recommended docker run invocation, env-var reference, and volume
layout.

The archives prefixed Psycheros- below are GitHub's auto-attached
full-monorepo source — useful for building the image yourself, but the published
image at GHCR is the supported install path.