Small agent tweaks

package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@ztimson/ai-utils",
-	"version": "1.4.2",
+	"version": "1.4.3",
 	"description": "AI Utility library",
 	"author": "Zak Timson",
 	"license": "MIT",

src/llm.ts

@@ -132,8 +132,8 @@ class LLM {
 			return {
 				name: toolName,
 				description: `${a.delegate ? 'Delegate to ' : ''}Subagent: ${a.description || a.name}`,
-				args: <any>(a.delegate ? {} : {
-					context: {type: 'string', description: 'Summary of related messages, samples, files, etc...', required: true},
+				args: <any>({
+					context: !a.delegate ? {type: 'string', description: 'Summary of related messages, samples, files, etc...', required: true} : undefined,
 					instructions: {type: 'string', description: 'Detailed instructions for subagent to complete', required: true},
 				}),
 				fn: async (args: any, stream: any, ai: any, id?: string) => {
@@ -148,8 +148,9 @@ class LLM {
 					const q = a.delegate ? '' : `${args.instructions}${args.context ? `\n\n<context>${args.context}</context>` : ''}`;
 
 					const request = this.ask(q, {
-						system: `You are a specialized subagent. ${a.delegate ? 'Your output streams directly to the user for the remainder of this turn. You are mid conversation - dispense with greetings.' : 'You are wrapped in a tool call that will be analysis by an LLM - dispense with conversation'}
-As a subagent, focus on executing your task completely using available tools and returning only the final result - no commentary, questions, or dialogue.
+						system: `You are a specialized subagent being called from an orchestrator
+${a.delegate ? 'Your output streams directly to the user for the remainder of this turn. You are mid conversation' : 'You are wrapped in a tool call that will be analysis by an LLM'}
+Dispense with greetings and focus on your instructions using available tools and returning only the final result unless specifically instructed to converse
 
 ${a.system}`,
 						model: a.model || undefined,

src/memory.ts

@@ -186,6 +186,17 @@ export class MemoryManager {
 
 	constructor(private llm: any) {}
 
+	private ghostNodes(memories: Memory[]): string[] {
+		const names = new Set(memories.map(m => m.name));
+		const ghosts = new Set<string>();
+		for (const m of memories) {
+			for (const link of m.links) {
+				if (!names.has(link)) ghosts.add(link);
+			}
+		}
+		return [...ghosts];
+	}
+
 	static normalize(m?: Memory[] | MemoryCache | MemoryOptions) {
 		if(!m) return null;
 		const raw = m instanceof MemoryCache || Array.isArray(m);
@@ -458,6 +469,8 @@ ${currentBody}
 
 	private async factAgent(conversation: string, memories: Memory[], options: LLMRequest, weekKey: string): Promise<FactBucket[]> {
 		const buckets = new Map<string, string[]>();
+		const ghosts = this.ghostNodes(memories);
+
 		await this.llm.ask(conversation, {
 			model: options.model,
 			temperature: 0.2,
@@ -472,14 +485,15 @@ Rules:
 - If nothing worth remembering was said, do not call any tools
 
 When extracting facts, you MUST also decide the exact destination path:
-- Use an existing node name if the facts clearly belong there
+- Reuse node names (including ghost) as much as possible IF the facts belongs there
 - All information primarily about the user should go under "People/User"
 - When required, create a new path following collection/subject format (e.g., People/Sarah, Projects/Oxide) — you are not limited to any fixed list of collections, use whatever fits
 - For journal entries, use "Journal"
 
 Available nodes:
 - Journal
-${this.listNodes(memories).filter(n => !n.name.includes('Journal')).map(n => `- ${n.name}: ${n.description}`).join('\n') || 'None yet.'}`,
+${this.listNodes(memories).filter(n => !n.name.includes('Journal')).map(n => `- ${n.name}: ${n.description}`).join('\n') || 'None yet.'}
+${ghosts.length ? `${ghosts.map(g => `- ${g}: (Ghost)`).join('\n')}` : ''}`,
 			tools: [{
 				name: 'facts_extract',
 				description: 'Submit facts with their destination',