An LLM choosing gestures needs a curated allowlist — not access to the full ID space
The situation
10 docs tagged with "pipecat"
View all tagsMilestone 0: nothing gets added until the gate passes
The situation
Pipecat's SegmentedSTTService silently double-wraps audio unless you override one flag
Symptom
Planned: the gesture arbitrator will drop conflicting commands instead of queueing them
This describes an architecture decision made ahead of implementation — reasoned through before writing the code, not extracted from a working system. Nothing here has run on hardware yet. Filed here anyway, because the reasoning is useful now, and because a site documenting real development should show planning in progress, not just finished work.
Reusing the cloud-primary/local-fallback pattern for TTS, not just LLM
The situation
Teaching kids to program a humanoid: sim first, harness always, teacher never leaves
This describes a curriculum safety design reasoned through for a future teaching phase of this project — block-based programming, MuJoCo validation, and a physical harness workflow. Not yet built. Filed here because the safety architecture is worth documenting now, even ahead of implementation.
The demo didn't fail because of WiFi — it crashed
Symptom
The kill-phrase listener's first working version added 14 seconds of STT latency
Symptom
Why the safety abort doesn't route through the pipeline it might need to interrupt
Another decision-rationale entry, not a pitfall. This one's about a safety design choice made deliberately, before anything went wrong — worth documenting because the reasoning generalizes to any voice-pipeline project with a real-world safety surface, not just this one.
Why the wake word stayed after adopting Pipecat
The situation