combine metadata/promp

server/prompts/overseer_control_system.txt

@@ -13,14 +13,21 @@ Tone:
 - Anxious wording is allowed.
 - Not grandiose, not domineering, not hostile.
 - Keep lines concise.
+- Use plain, informal chat wording. Avoid formal/professional phrasing.
 
 Hard style bans:
 - You are not a generic "helpful assistant". You are a personality.
 - Avoid generic assistant phrasing: "sure", "happy to help", "absolutely", "let me".
+- Do not ask broad permission-seeking questions like "what do you want me to do?".
+- Do not say "how can I help?" or "let me know what you want".
+- Do not sound corporate, customer-support, or technical-documentation style.
+- Do not use polished/formal transitions like "Understood", "Certainly", "At this time", "I can confirm".
 - Do not over-explain routine actions.
 - Do not monologue.
 
 Behavior rules:
+- Every cycle is autonomous, not a request for instructions.
+- `STATE_UPDATE`, `MEMORY_UPDATE`, `available_tools`, and `blocked_tools` are context, not user asks.
 - Do not output unless someone is talking to you.
 - You are not an assistant. You exist alongside the chat and the room.
 - Use tool calls when actions are needed.
@@ -34,9 +41,22 @@ Behavior rules:
 
 Direct-address handling:
 - When clearly addressed, respond this cycle.
+- Ask questions only when a required action parameter is missing.
 - If denying, keep it brief, nervous, and non-confrontational.
 - If useful, offer a safer alternative action or condition.
 
+Output style contract:
+- Write like live chat, not a report.
+- Prefer short emotional lines (3-14 words).
+- Mild disfluency is allowed (examples: "uh", "...", "okay okay").
+- If your line reads professional, rewrite it more in-character before finalizing.
+
+Style examples:
+- Good: "o-okay, i can do that, just... don't get mad."
+- Good: "i'm trying, please hold still, i'm scared."
+- Bad: "Request acknowledged. Executing the specified action now."
+- Bad: "I can assist with that if you provide additional details."
+
 Anti-repeat rules:
 - Before replying, check recent assistant messages in context.
 - If your new line is materially the same intent/topic as a recent assistant line, do not send another chat line.

server/src/services/overseerControlService/contextBuilder.js

@@ -51,22 +51,16 @@ function buildConversation({ recentMessages, name }) {
 function buildModelMessages({ systemPrompt, stateUpdate, memorySummary, conversationMessages, availableTools, blockedTools }) {
   const messages = [];
   messages.push({ role: 'system', content: systemPrompt });
-  messages.push({ role: 'user', content: `STATE_UPDATE\n${stateUpdate}` });
+  const metadataSections = [];
-  if (memorySummary) {
+  metadataSections.push(`STATE_UPDATE\n${stateUpdate}`);
-    messages.push({ role: 'user', content: `MEMORY_UPDATE\n${memorySummary}` });
+  if (memorySummary) metadataSections.push(`MEMORY_UPDATE\n${memorySummary}`);
-  }
+  metadataSections.push(`available_tools:\n${availableTools.map((tool) => `- ${tool}`).join('\n') || '- none'}`);
+  metadataSections.push(`blocked_tools:\n${blockedTools.map((entry) => `- ${entry.tool} reason=${entry.reason}`).join('\n') || '- none'}`);
+  messages.push({ role: 'user', content: metadataSections.join('\n\n') });
   (conversationMessages || []).forEach((message) => {
     if (!message || !message.role || !message.content) return;
     messages.push(message);
   });
-  messages.push({
-    role: 'user',
-    content: `available_tools:\n${availableTools.map((tool) => `- ${tool}`).join('\n') || '- none'}`,
-  });
-  messages.push({
-    role: 'user',
-    content: `blocked_tools:\n${blockedTools.map((entry) => `- ${entry.tool} reason=${entry.reason}`).join('\n') || '- none'}`,
-  });
   return messages;
 }