Staying on-topic after answering a sub, by generating two responses and swapping them around.

[edgy remnant]

Discussion of problem: #neurotic-neurons message
(Nomenclature: comments for chat, messages/responses for Neuro)
TL;DR:
When a sub comes in, generate an on-topic message and store it somewhere. Produce a response to the sub. Send the stored message to TTS (and context). The next message will be on topic too (assumption).

ASSUMPTIONS:
Neuro's context is a list of previous Neuro messages.
Neuro will stay on-topic if the previous message is on-topic.
Each message cycle, Neuro is presented with a comment pool.
Subs are comments.
Neuro thanks a sub if there's a sub in the comment pool presented to her.

PROCESS:
When an incoming sub is detected via regex, generate an on-topic message[1]. Store this response somewhere, don't put it in the context.
Then produce a sub response[2], TTS it and stick it in the context.
Send the stored response to TTS and stick it in context.
The next message should be on-topic because the last message was on-topic.

CRINGINESS:
The cached message can't refer to the sub response (little cringe), but messages after that can see the sub response (no cringe).
Hardcoded regex is used to detect if the chat comment pool contains a sub. (some cringe).
[1] You'll need to remove the sub from the comment pool to make the on-topic message, and then return it to make a sub response. (some cringe)
Creates large gaps where Neuro doesn't react to chat and the cached message will be outdated. (cringe)
Increased latency jitter if the LLM is the slow part of the pipeline. (some cringe)
Overall manipulation of chatting pipeline (cringe)

ISSUES:
[2] If Neuro thanks a sub with 90% probability, not 100%, then that generated sub response won't be an actual sub response. You'll instead end up with two messages that sort of exist in the same time and are probably similar. If you discard one, then you basically tanked that latency for nothing.
Can't cope with multiple consecutive sub responses.
If the on-topic message is a lengthy rant, then you might have a lot of deadair. If this is gap is indeed large, maybe somehow force this message to be short, or only cache the first 200 tokens. After responding to the sub, complete the rest of this 200token seed string, then TTS it. Hopefully the seed string is enough to keep the model on-topic.