Fixed delegate agent history... again

src/llm.ts

@@ -34,6 +34,10 @@ export type LLMMessage = {
 	content: string | any;
 	/** Timestamp */
 	timestamp?: number;
+	/** Response duration in ms */
+	duration?: number;
+	/** Tokens per second */
+	tps?: number;
 } | {
 	/** Tool call */
 	role: 'tool';
@@ -135,12 +139,15 @@ class LLM {
 				fn: async (args: any, stream: any, ai: any, id?: string) => {
 					if(depth >= MAX_AGENT_DEPTH) return 'Max agent delegation depth exceeded';
 
-					// Opt-in only, self always excluded regardless of whitelist
 					const nested = (a.agents || [])
 						.map(name => allAgents.find(x => x.name === name))
 						.filter((x): x is Agent => !!x && x.name !== a.name);
 
-					const request = this.ask(a.delegate ? '' : `${args.instructions}${args.context ? `\n\n<context>${args.context}</context>` : ''}`, {
+					// Delegate continues the SAME live conversation - no new user turn needed,
+					// `history` is always current (shared, mutated in place) by the time this runs
+					const q = a.delegate ? '' : `${args.instructions}${args.context ? `\n\n<context>${args.context}</context>` : ''}`;
+
+					const request = this.ask(q, {
 						system: `You are a specialized subagent. ${a.delegate ? 'Your output streams directly to the user for the remainder of this turn. You are mid conversation - dispense with greetings.' : 'You are wrapped in a tool call that will be analysis by an LLM - dispense with conversation'}
 As a subagent, focus on executing your task completely using available tools and returning only the final result - no commentary, questions, or dialogue.
 
@@ -265,7 +272,10 @@ ${a.system}`,
 		promise = (async () => {
 			let tools: AiTool[] = options.tools || this.ai.options.llm?.tools || [];
 			const prompts: string[] = [];
+			// `history` is the single source of truth from here on - mutated in place by
+			// this call AND by any nested/delegated agent calls sharing the same array
 			let history = options.history || [];
+			if(message) history.push({role: 'user', content: message, timestamp: Date.now()});
 
 			// MCP
 			const mcp = options.mcp || this.ai.options?.llm?.mcp;
@@ -334,7 +344,9 @@ Also relevant but not preloaded (use \`memory_recall\`): ${listed.map(r => r.nam
 			if(aborted) throw Object.assign(new Error('Aborted'), {name: 'AbortError'});
 
 			prompts.unshift(options.system || this.ai.options.llm?.system || '');
-			request = this.models[m].ask(message, {...options, tools, system: prompts.filter(Boolean).join('\n\n')});
+			// Message already appended to shared `history` above - pass '' so the provider
+			// doesn't push a duplicate user turn
+			request = this.models[m].ask('', {...options, tools, system: prompts.filter(Boolean).join('\n\n')});
 			let resp = await request;
 
 			// Capture meta (duration / tps)