Updated OpenAI models for chat, using proper context prompt instead of prefixing message.

config/default.js

@@ -69,11 +69,11 @@ module.exports = {
 	chat: {
 		apiKey: null, // OpenAI
 		validModels: [
-			'gpt-3.5-turbo',
+			'gpt-6-luna',
-			'text-davinci-003',
+			'gpt-6.1-sol',
-			'gpt-4',
 		],
-		model: 'gpt-3.5-turbo',
+		model: 'gpt-6-luna',
+		reasoningEffort: 'none', // none, low, medium, high; reasoning tokens count towards the user token limit
 		userTokenLimit: 20000, // daily, roughly 100+ messages or $0.04 per user
 		userTokenPeriod: 24, // hours
 		replyTokenLimit: 1000,

src/games/chat.js

@@ -2,6 +2,7 @@
 
 const config = require('config');
 const bhttp = require('bhttp');
+const { format } = require('date-fns');
 
 const style = require('../utils/style');
 const knex = require('../knex');
@@ -145,6 +146,20 @@ function setRule(rule, context) {
 	context.sendMessage(`Chat rule must be at least 3 characters long, ${context.user.prefixedUsername}`, context.room?.id, { label: false }, context.user.username);
 }
 
+function getInstructions(context) {
+	return {
+		role: 'developer',
+		content: [
+			`You are ${config.user.username}, ${settings.rule}, chatting in a group chat.`,
+			`You are talking to ${context.user.username}. Today is ${format(new Date(), 'EEEE d MMMM yyyy')}.`,
+			config.platform === 'irc'
+				? 'Reply in plain text on a single line, IRC does not support any formatting.'
+				: 'Reply in a single paragraph without headings or lists. You may use **bold** and *italic*.',
+			settings.replyWordLimit && `Use ${settings.replyWordLimit} words or fewer.`,
+		].filter(Boolean).join(' '),
+	};
+}
+
 async function getTokens(username) {
 	const { used_tokens: usedTokens } = await knex('chat_tokens')
 		.sum('tokens as used_tokens')
@@ -229,21 +244,17 @@ async function onCommand(args, context, isConversation) {
 			return;
 		}
 
-		const message = {
+		const message = { role: 'user', content: prompt };
-			role: 'user',
-			content: settings.replyWordLimit
-				? `Using ${settings.replyWordLimit} words or fewer, answer as if you're ${settings.rule}. ${prompt}`
-				: `Answer as if you're ${settings.rule}. ${prompt}`,
-		};
-
 		const userHistory = (history.get(context.user.username) || []).concat(message);
 
-		context.logger.info(`${context.user.username} asked OpenAI '${config.chat.model}' (${userHistory.length}): ${message.content}`);
+		context.logger.info(`${context.user.username} asked OpenAI '${settings.model}' (${userHistory.length}) to be ${settings.rule}: ${prompt}`);
 
 		const res = await bhttp.post('https://api.openai.com/v1/chat/completions', JSON.stringify({
 			model: settings.model,
-			max_tokens: config.chat.replyTokenLimit,
+			max_completion_tokens: config.chat.replyTokenLimit,
-			messages: userHistory,
+			reasoning_effort: config.chat.reasoningEffort,
+			temperature: settings.temperature,
+			messages: [getInstructions(context), ...userHistory], // instructions are rebuilt for every request and not stored in history, so rule changes apply immediately
 			user: `${context.user.id}:${context.user.username}`,
 		}), {
 			headers: {
@@ -252,11 +263,20 @@ async function onCommand(args, context, isConversation) {
 			},
 		});
 
-		const reply = res.body.choices?.[0].message;
+		if (res.statusCode !== 200) {
+			throw new Error(`OpenAI responded with ${res.statusCode}: ${res.body.error?.message || JSON.stringify(res.body)}`);
+		}
+
+		const reply = res.body.choices?.[0]?.message;
+
+		if (!reply?.content) {
+			throw new Error(`OpenAI returned no content (finish reason '${res.body.choices?.[0]?.finish_reason}', ${res.body.usage?.total_tokens} tokens)`);
+		}
+
 		const target = prompt.match(/@\s*[\w-]+\s*$/)?.[0].slice(1).trim();
 		const curatedContent = reply.content.replace(/\n+/g, '. ');
 
-		context.logger.info(`OpenAI ${config.chat.model} replied to ${target || context.user.username} (${res.body.usage.total_tokens} tokens, ${(usedTokens || 0) + res.body.usage.total_tokens}/${config.chat.userTokenLimit} used): ${curatedContent}`);
+		context.logger.info(`OpenAI ${settings.model} replied to ${target || context.user.username} (${res.body.usage.total_tokens} tokens, ${(usedTokens || 0) + res.body.usage.total_tokens}/${config.chat.userTokenLimit} used): ${curatedContent}`);
 
 		if (isConversation) {
 			context.sendMessage(`${curatedContent}`, context.room?.id, { label: false }, context.user.username);
@@ -270,16 +290,17 @@ async function onCommand(args, context, isConversation) {
 			created: knex.raw("datetime('now')"),
 		});
 
-		if (config.chat.history > 0) {
+		if (settings.history > 0) {
-			history.set(context.user.username, userHistory.concat(reply).slice(-settings.history));
+			// only keep the fields the API accepts as input, newer models return extras like refusal and annotations
+			history.set(context.user.username, userHistory.concat({ role: reply.role, content: reply.content }).slice(-settings.history));
 		}
 	} catch (error) {
-		context.logger.error(error.response ? `${error.response.status}: ${JSON.stringify(error.response.data)}` : error.message);
+		context.logger.error(error.message);
 		context.sendMessage('Sorry, I can\'t think right now', context.room?.id, { label: false }, context.user.username);
-	}
+	} finally {
-
 		processing.delete(context.user.username);
 	}
+}
 
 function onMessage(message, context) {
 	const isConversation = message.recipient === config.user.username && message.type !== 'whisper' && !context.containsCommand;