fixie-ai / hisanta.ai Goto Github PK

This will be good for laughs.

Looks like the character view in the in-call UI has a scrollbar and you can actually scroll up and down within it.

After creating a VoiceSession and calling ".start()" on it, sometimes we never get a message saying that the session has transitioned from "idle" to "thinking" state. Hence the call is hung and we never get a welcome message from the agent.

Related: If we .stop() the session when it's in this state and call .start() again, it looks like it never tries to reconnect (either to the websocket or the room). So there is a client-side bug there (which I will fix).

It's weird for to be in one of these states and still see the input visualizer changing

• TTS token shortage
• Footer -- Learn More (blog post), Fixie Voice (fixie.ai/voice), We're hiring, Discord
• Bad Elfie
• Bad Mrs. Claus
• Blog post tightening
• Open source this repo
• Too many concurrent requests -- speedbump (e.g., numrings)
• Scrub shit (matt)
• Fixie.ai/voice tightening
• PH -- Zach to re-create, schedule for 12:01am
• Email to friends + family tomorrow
• LinkedIn / Twitter posts from everyone in AM
• Grab wake lock to prevent going to sleep (matt)
• Thumbs up / down (conversation ID)
• Bump NPM package for fixie sdk

When going from one character to another, the old voice/character is used even though the UI shows the new one that was selected. Obviously a state management bug in CallCharacter. I'll fix it.

We need to do a better job on the client at dealing with various kinds of errors and timeouts.

• If a conversation never gets started after calling .start(), we should proactively stop it and reconnect. This would be a workaround for #25.

• If the Websocket connection fails, we need to deal with that and retry a few times, showing an error page if the error is persistent. The challenge here is that this is currently managed internally by the VoiceSession and the app does not have visibility into the connection state. We should migrate .warmup() out of the constructor so it can use the onerror signal that the client sets up for this, be sure we're catching errors on .warmup() and calling .onerror when they do happen, and probably also expose the connection state (not just the live conversation state) as part of the VoiceSession.state field.

• CallCharacter should transition to an error page if it gets an error and retries aren't working.

If you are in the middle of a convo and navigate away (to the homepage for example by clickin on hisanta.ai), the convo keeps going.

We need to trigger a .stop() on the voice session when this happens.