@tom are you using Jev to filter unworthy requests?
This is what we use with our own bot:
const REASONS = {
direct_ask:
"Asks Lead (or the swarm, the bot, 'you' meaning the AI) a question or for something, explicitly or clearly implied.",
followup_to_lead:
"Answers, confirms, corrects or continues something Lead or another swarm agent ([Agent]) just said or asked, so Lead needs it to proceed.",
instruction_to_lead:
"Gives a task, instruction, request or decision for the AI to act on, even without naming it (for example 'can you create X', 'ship it', 'do it').",
reaction_only:
"Only a reaction: emoji, laughter (lol, haha, jaja), 'boom!', 'nice', 'thanks', '+1', with no request and nothing new for Lead.",
banter_between_others:
"Chat between the humans, or between other people and other bots, that is not aimed at Lead.",
not_addressed_to_lead:
"Addressed to a specific other person or another assistant, or an FYI that needs no answer from Lead.",
other: "None of the above, or unclear.",
};
questions: {
reason: {
type: "choice",
instructions:
"Which label best describes new_message from Lead's point of view? Use recent_thread_messages to tell who it is talking to. Lines starting with [Agent] are swarm bots.",
criteria: REASONS,
},
},commit ef638ec