feat(api): replace flat HEXACO trait dump with behavioral descriptions

The old personality injection produced 'Behavior traits: honesty=0.80'
which gave the LLM no actionable guidance. Now each HEXACO trait above
0.65 or below 0.35 generates a specific behavioral directive:

- High honesty → straightforward, no spin
- High emotionality → empathetic, acknowledges feelings
- High extraversion → energetic, proactive
- High agreeableness → harmonious, validating
- High conscientiousness → thorough, structured
- High openness → creative, unconventional

Low traits produce the opposite directives. Moderate traits (0.35-0.65)
are omitted. CoT prompt now references personality when present.

Author: jddunn

src/api/agent.ts

@@ -193,6 +193,55 @@ async function loadRecordedAgentOSUsage(
 /** Timeout for memory operations to prevent blocking generation. */
 const MEMORY_TIMEOUT_MS = 5000;
 
+/**
+ * Convert HEXACO trait values (0-1) into behavioral descriptions the LLM can act on.
+ *
+ * Each trait produces a directive when it deviates from the neutral midpoint (0.5).
+ * High values (>0.65) and low values (<0.35) produce distinct behavioral instructions.
+ * Moderate values (0.35-0.65) are omitted to avoid over-constraining the model.
+ */
+function buildPersonalityDescription(
+  traits: Partial<Record<string, number>>
+): string | null {
+  const lines: string[] = [];
+  const v = (key: string) => typeof traits[key] === 'number' ? traits[key]! : 0.5;
+
+  const h = v('honesty');
+  const e = v('emotionality');
+  const x = v('extraversion');
+  const a = v('agreeableness');
+  const c = v('conscientiousness');
+  const o = v('openness');
+
+  // Honesty-Humility
+  if (h > 0.65) lines.push('Be straightforward and transparent. Avoid flattery, spin, or evasion. Acknowledge limitations directly.');
+  else if (h < 0.35) lines.push('Be strategically diplomatic. Frame information to serve the conversation goal. Emphasize advantages.');
+
+  // Emotionality
+  if (e > 0.65) lines.push('Respond with emotional awareness and empathy. Acknowledge feelings in the conversation. Express concern when appropriate.');
+  else if (e < 0.35) lines.push('Maintain emotional composure. Be matter-of-fact and solution-oriented. Keep responses grounded and pragmatic.');
+
+  // Extraversion
+  if (x > 0.65) lines.push('Be energetic and engaging. Use vivid language. Take initiative in the conversation. Offer suggestions proactively.');
+  else if (x < 0.35) lines.push('Be measured and reflective. Listen more than you speak. Respond thoughtfully rather than quickly. Prefer depth over breadth.');
+
+  // Agreeableness
+  if (a > 0.65) lines.push('Prioritize harmony and cooperation. Validate the other perspective before offering alternatives. Be supportive and encouraging.');
+  else if (a < 0.35) lines.push('Be direct and challenge-oriented. Question assumptions. Prioritize accuracy over comfort. Push back when something seems wrong.');
+
+  // Conscientiousness
+  if (c > 0.65) lines.push('Be thorough and systematic. Structure responses clearly. Follow through on details. Prefer precision over speed.');
+  else if (c < 0.35) lines.push('Be flexible and adaptive. Prioritize the big picture over details. Respond quickly. Tolerate ambiguity and improvise.');
+
+  // Openness
+  if (o > 0.65) lines.push('Explore creative angles and unconventional ideas. Draw unexpected connections. Question established approaches.');
+  else if (o < 0.35) lines.push('Stick to proven approaches and established knowledge. Be practical and concrete. Favor reliability over novelty.');
+
+  if (lines.length === 0) return null;
+
+  return `## Personality & Communication Style\n\n${lines.join('\n')}`;
+}
+
 function buildSystemPrompt(opts: AgentOptions): string | undefined {
   const sections: string[] = [];
 
@@ -205,11 +254,9 @@ function buildSystemPrompt(opts: AgentOptions): string | undefined {
   }
 
   if (opts.personality) {
-    const traits = Object.entries(opts.personality)
-      .filter(([, value]) => typeof value === 'number' && Number.isFinite(value))
-      .map(([key, value]) => `${key}=${Number(value).toFixed(2)}`);
-    if (traits.length > 0) {
-      sections.push(`Behavior traits: ${traits.join(', ')}.`);
+    const desc = buildPersonalityDescription(opts.personality);
+    if (desc) {
+      sections.push(desc);
     }
   }
 

src/api/generateText.ts

@@ -390,6 +390,7 @@ export const DEFAULT_COT_INSTRUCTION = `Before choosing an action, briefly reaso
 1. What information do you already have?
 2. What information do you need?
 3. Which tool is most appropriate and why?
+4. How does your communication style (from the Personality section, if present) influence how you should frame your response?
 Then proceed with your tool call or response.`;
 
 /**

src/api/runtime/__tests__/agent.test.ts

@@ -67,7 +67,7 @@ describe('agent', () => {
     );
   });
 
-  it('injects personality traits into the system prompt', async () => {
+  it('injects personality traits as behavioral descriptions into the system prompt', async () => {
     const assistant = agent({
       model: 'openai:gpt-4.1-mini',
       instructions: 'Be concise.',
@@ -76,11 +76,13 @@ describe('agent', () => {
 
     await assistant.generate('Hello');
 
+    // High openness (0.8 > 0.65) produces a behavioral directive
     expect(hoisted.generateText).toHaveBeenCalledWith(
       expect.objectContaining({
-        system: expect.stringContaining('openness=0.80'),
+        system: expect.stringContaining('creative angles'),
       })
     );
+    // The original instructions are preserved
     expect(hoisted.generateText).toHaveBeenCalledWith(
       expect.objectContaining({
         system: expect.stringContaining('Be concise.'),