Fix open-ai early termination & memory improvements

src/open-ai.ts

@@ -36,8 +36,10 @@ export class OpenAi extends LLMProvider {
 	private toWire(history: LLMMessage[], system?: string): any[] {
 		const wire: any[] = [];
 		if(system) wire.push({role: 'system', content: system});
+
 		for(let i = 0; i < history.length; i++) {
 			const h = history[i];
+
 			if(h.role !== 'tool') {
 				wire.push({role: h.role, content: this.toWireContent(h.content)});
 				continue;
@@ -45,24 +47,34 @@ export class OpenAi extends LLMProvider {
 
 			const calls: any[] = [];
 			const results: any[] = [];
+
 			while(i < history.length && history[i].role === 'tool') {
-				const tool = <any>history[i];
+				const tool: any = history[i];
+
 				calls.push({
 					id: tool.id,
 					type: 'function',
 					function: {
 						name: tool.name,
-						arguments: JSON.stringify(tool.args)
+						arguments: JSON.stringify(tool.args || {})
 					}
 				});
+
 				results.push({
 					role: 'tool',
 					tool_call_id: tool.id,
 					content: tool.error || tool.content || ''
 				});
+
 				i++;
 			}
-			wire.push({role: 'assistant', content: null, tool_calls: calls});
+
+			wire.push({
+				role: 'assistant',
+				content: null,
+				tool_calls: calls
+			});
+
 			wire.push(...results);
 			i--;
 		}
@@ -76,13 +88,12 @@ export class OpenAi extends LLMProvider {
 			if(!options.history) options.history = [];
 			const history = options.history;
 			if(message) history.push({role: 'user', content: message, timestamp: Date.now()});
-
 			const tools = options.tools || this.ai.options.llm?.tools || [];
 			const requestParams: any = {
 				model: options.model || this.model,
 				stream: !!options.stream,
-				max_completion_tokens: options.maxTokens || this.ai.options.llm?.maxTokens || undefined,
-				temperature: options.temperature || this.ai.options.llm?.temperature || undefined,
+				max_completion_tokens: options.maxTokens ?? this.ai.options.llm?.maxTokens,
+				temperature: options.temperature ?? this.ai.options.llm?.temperature,
 				tools: tools.map(t => ({
 					type: 'function',
 					function: {
@@ -112,7 +123,10 @@ export class OpenAi extends LLMProvider {
 
 			try {
 				let terminal = false;
+				let iteration = 0;
+
 				do {
+					iteration++;
 					requestParams.messages = this.toWire(history.filter(h => h.role !== 'system'), options.system);
 
 					const callStart = Date.now();
@@ -126,36 +140,46 @@ export class OpenAi extends LLMProvider {
 					let usage: any;
 					let finishReason: string | undefined;
 					let msg: any = {content: '', tool_calls: []};
+					let streamedChars = 0;
+
 					if(options.stream) {
 						let streamCompleted = false;
 						try {
 							for await (const chunk of resp) {
 								if(controller.signal.aborted) break;
 								if(chunk.usage) usage = chunk.usage;
+
 								const choice = chunk.choices?.[0];
 								if(choice?.finish_reason) finishReason = choice.finish_reason;
+
 								if(choice?.delta?.content) {
 									msg.content += choice.delta.content;
+									streamedChars += choice.delta.content.length;
 									options.stream({text: choice.delta.content});
 								}
+
 								if(choice?.delta?.tool_calls) {
 									for(const deltaTC of choice.delta.tool_calls) {
 										const index = deltaTC.index ?? msg.tool_calls.length;
 										let existing = msg.tool_calls.find((tc: any) => tc.index === index);
+
 										if(!existing) {
 											existing = {index, id: '', function: {name: '', arguments: ''}};
 											msg.tool_calls.push(existing);
 										}
+
 										if(deltaTC.id) existing.id = deltaTC.id;
 										if(deltaTC.function?.name) existing.function.name = deltaTC.function.name;
 										if(deltaTC.function?.arguments) existing.function.arguments += deltaTC.function.arguments;
 									}
 								}
 							}
+
 							streamCompleted = true;
 						} catch(err) {
 							if(!controller.signal.aborted) throw err;
 						}
+
 						if(streamCompleted && !finishReason) finishReason = msg.tool_calls.length ? 'tool_calls' : 'stop';
 					} else {
 						usage = resp.usage;
@@ -164,9 +188,7 @@ export class OpenAi extends LLMProvider {
 					}
 
 					const duration = Date.now() - callStart;
-					const tps = usage?.completion_tokens && duration > 0
-						? usage.completion_tokens / (duration / 1000)
-						: 0;
+					const tps = usage?.completion_tokens && duration > 0 ? usage.completion_tokens / (duration / 1000) : 0;
 
 					if(finishReason === 'length' && !controller.signal.aborted) {
 						if(msg.content?.trim()) history.push({role: 'assistant', content: msg.content.trim(), timestamp: Date.now(), duration, tps});
@@ -178,10 +200,20 @@ export class OpenAi extends LLMProvider {
 					}
 
 					const toolCalls = msg.tool_calls || [];
+
 					if(toolCalls.length && !controller.signal.aborted) {
 						if(msg.content?.trim()) history.push({role: 'assistant', content: msg.content.trim(), timestamp: Date.now(), duration, tps});
+
 						const entries = toolCalls.map((tc: any) => {
-							const entry: any = {role: 'tool', id: tc.id, name: tc.function.name, args: JSONAttemptParse(tc.function.arguments, {}), content: undefined, timestamp: Date.now()};
+							const entry: any = {
+								role: 'tool',
+								id: tc.id,
+								name: tc.function.name,
+								args: JSONAttemptParse(tc.function.arguments, {}),
+								content: undefined,
+								timestamp: Date.now()
+							};
+
 							history.push(entry);
 							return {tc, entry};
 						});
@@ -195,6 +227,7 @@ export class OpenAi extends LLMProvider {
 									if(chunk.done) return;
 									options.stream!(chunk);
 								});
+
 								const result = await tool.fn(entry.args, toolStream, this.ai, tc.id);
 								entry.content = typeof result === 'object' ? JSONSanitize(result) : result;
 							} catch(err: any) {