diff --git a/frontend/src/pages/ChatView.tsx b/frontend/src/pages/ChatView.tsx
index 3d191e42..7907e004 100644
--- a/frontend/src/pages/ChatView.tsx
+++ b/frontend/src/pages/ChatView.tsx
@@ -226,7 +226,9 @@ export function ChatView() {
           <option value="claude-opus-4-7">Opus 4.7</option>
           <option value="claude-opus-4-7:max">Opus 4.7 Max</option>
           <option value="claude-opus-4-6">Opus 4.6</option>
+          <option value="claude-opus-4-5">Opus 4.5</option>
           <option value="claude-sonnet-4-6">Sonnet 4.6</option>
+          <option value="claude-sonnet-4-5">Sonnet 4.5</option>
           <option value="claude-haiku-4-5">Haiku 4.5</option>
         </select>
         {!keyboardOpen && (
diff --git a/packages/harness/__tests__/auto-rename.test.ts b/packages/harness/__tests__/auto-rename.test.ts
index d3fa8572..a277124e 100644
--- a/packages/harness/__tests__/auto-rename.test.ts
+++ b/packages/harness/__tests__/auto-rename.test.ts
@@ -4,6 +4,7 @@ import {
   extractRecentPrompts,
   generateSessionName,
   generateSessionNameFallback,
+  sanitizeSessionName,
   setClientFactory,
   resetClientFactory,
   createAnthropicClient,
@@ -140,16 +141,58 @@ describe('generateSessionName', () => {
 
     expect(result).toBe('Auth Bug Fix Session');
     expect(mockCreate).toHaveBeenCalledOnce();
-    expect(mockCreate).toHaveBeenCalledWith(
-      {
-        model: AUTO_RENAME_MODEL,
-        max_tokens: 20,
-        system:
-          'Generate a 3-6 word title for this chat session. Be specific and descriptive. Return only the title, nothing else.',
-        messages: [{ role: 'user', content: 'Fix the auth bug\nUpdate login page' }],
-      },
-      { timeout: 5000 },
-    );
+    const call = mockCreate.mock.calls[0];
+    expect(call[0].model).toBe(AUTO_RENAME_MODEL);
+    expect(call[0].max_tokens).toBe(20);
+    expect(call[0].messages[0].content).toBe('Fix the auth bug\n---\nUpdate login page');
+    expect(call[1]).toEqual({ timeout: 5000 });
+  });
+
+  it('sanitizes Haiku output with embedded quotes', async () => {
+    const mockCreate = vi.fn().mockResolvedValue({
+      content: [{ type: 'text', text: '"Fix Auth Module"' }],
+    });
+
+    setClientFactory(() => ({ messages: { create: mockCreate } }) as never);
+
+    const result = await generateSessionName(['Fix the auth bug']);
+    expect(result).toBe('Fix Auth Module');
+  });
+
+  it('falls back when Haiku generates conversational response', async () => {
+    const mockCreate = vi.fn().mockResolvedValue({
+      content: [
+        {
+          type: 'text',
+          text: 'I apologize, but I cannot access the specific Jira ticket list',
+        },
+      ],
+    });
+
+    setClientFactory(() => ({ messages: { create: mockCreate } }) as never);
+
+    const prompts = ['check my jira tickets'];
+    const result = await generateSessionName(prompts);
+    const fallback = generateSessionNameFallback(prompts);
+    expect(result).toBe(fallback);
+  });
+
+  it('truncates Haiku output at newlines', async () => {
+    const mockCreate = vi.fn().mockResolvedValue({
+      content: [
+        {
+          type: 'text',
+          text: '"OpenShift Agent Integration"\n\nWould you like me to',
+        },
+      ],
+    });
+
+    setClientFactory(() => ({ messages: { create: mockCreate } }) as never);
+
+    const result = await generateSessionName(['integrate openshift agent']);
+    expect(result).toBe('OpenShift Agent Integration');
+    expect(result).not.toContain('\n');
+    expect(result).not.toContain('Would');
   });
 
   it('falls back to keyword extraction when API call fails', async () => {
@@ -177,6 +220,56 @@ describe('generateSessionName', () => {
   });
 });
 
+describe('sanitizeSessionName', () => {
+  it('passes through clean titles', () => {
+    expect(sanitizeSessionName('Fix PR Shepherd CI Failures')).toBe('Fix PR Shepherd CI Failures');
+  });
+
+  it('strips surrounding double quotes', () => {
+    expect(sanitizeSessionName('"Decoupling AI Architecture"')).toBe('Decoupling AI Architecture');
+  });
+
+  it('strips surrounding single quotes', () => {
+    expect(sanitizeSessionName("'Some Title'")).toBe('Some Title');
+  });
+
+  it('truncates at first newline', () => {
+    expect(sanitizeSessionName('Good Title\n\nWould you like me to')).toBe('Good Title');
+  });
+
+  it('rejects apology responses', () => {
+    expect(sanitizeSessionName('I apologize, but I cannot access')).toBe('');
+    expect(sanitizeSessionName("I can't determine the topic")).toBe('');
+    expect(sanitizeSessionName("I don't have enough context")).toBe('');
+    expect(sanitizeSessionName('I am sorry but')).toBe('');
+  });
+
+  it('rejects generic conversation starters', () => {
+    expect(sanitizeSessionName('Start a Conversation')).toBe('');
+    expect(sanitizeSessionName('Start A Conversation')).toBe('');
+    expect(sanitizeSessionName('Begin New Chat Session')).toBe('');
+    expect(sanitizeSessionName('New Conversation')).toBe('');
+  });
+
+  it('rejects conversational filler', () => {
+    expect(sanitizeSessionName('Sure, here is the title')).toBe('');
+    expect(sanitizeSessionName('Certainly! Fix the bug')).toBe('');
+    expect(sanitizeSessionName('Here is the session name')).toBe('');
+  });
+
+  it('enforces max length', () => {
+    const long = 'A'.repeat(80);
+    const result = sanitizeSessionName(long);
+    expect(result.length).toBeLessThanOrEqual(60);
+  });
+
+  it('handles combined issues: quotes + newline + trailing text', () => {
+    expect(sanitizeSessionName('"OpenShift Integration"\n\nWould you like me to elaborate?')).toBe(
+      'OpenShift Integration',
+    );
+  });
+});
+
 describe('AUTO_RENAME_INTERVAL', () => {
   it('is 2', () => {
     expect(AUTO_RENAME_INTERVAL).toBe(2);
diff --git a/packages/harness/src/auto-rename.ts b/packages/harness/src/auto-rename.ts
index c2cd6882..6101ecbc 100644
--- a/packages/harness/src/auto-rename.ts
+++ b/packages/harness/src/auto-rename.ts
@@ -12,7 +12,7 @@ export const AUTO_RENAME_INTERVAL = 2;
 export const AUTO_RENAME_MODEL = 'claude-haiku-4-5-20251001';
 
 /** Model name on Vertex AI (uses different naming convention). */
-const VERTEX_MODEL = 'claude-3-5-haiku@20241022';
+const VERTEX_MODEL = 'claude-haiku-4-5@20251001';
 
 /** Max total characters of concatenated prompts sent to the LLM. */
 const MAX_PROMPT_INPUT_CHARS = 2000;
@@ -23,6 +23,16 @@ const MAX_RECENT_PROMPTS = 8;
 /** Maximum length of the generated session name. */
 const MAX_NAME_LENGTH = 60;
 
+/** Patterns that indicate Haiku went off-script — fall back to keyword extraction. */
+const REJECTED_PATTERNS = [
+  /^I\s+(apologize|apologise|cannot|can't|don't|am\s+sorry)/i,
+  /^(start|begin|new)\s+(a\s+|new\s+)?(conversation|chat|session)/i,
+  /^(hello|hi|hey|greetings)/i,
+  /^(sure|certainly|of\s+course)/i,
+  /^(here|let\s+me)/i,
+  /^would you like/i,
+];
+
 /** Common stop words to filter out when extracting key terms. */
 const STOP_WORDS = new Set([
   'a',
@@ -220,6 +230,53 @@ export function generateSessionNameFallback(prompts: string[]): string {
   return name;
 }
 
+/**
+ * Sanitize a Haiku-generated session name.
+ * Strips quotes, truncates at newlines, and rejects conversational responses.
+ * Returns empty string if the name is rejected.
+ */
+export function sanitizeSessionName(raw: string): string {
+  // Truncate at first newline — Haiku sometimes generates multi-line output
+  let name = raw.split('\n')[0].trim();
+
+  // Strip surrounding quotes (single, double, smart quotes)
+  name = name.replace(/^[\s"'\u201C\u201D\u2018\u2019]+|[\s"'\u201C\u201D\u2018\u2019]+$/g, '');
+
+  // Reject conversational responses — Haiku went off-script
+  for (const pattern of REJECTED_PATTERNS) {
+    if (pattern.test(name)) {
+      log.warn('rejected Haiku title (conversational)', { raw: raw.slice(0, 80) });
+      return '';
+    }
+  }
+
+  // Enforce length limit
+  if (name.length > MAX_NAME_LENGTH) {
+    name = name.slice(0, MAX_NAME_LENGTH).trim();
+  }
+
+  return name;
+}
+
+/** System prompt for LLM-based session naming. */
+const NAMING_SYSTEM_PROMPT = `Extract a 3-6 word title from the user messages below. The title should capture the primary topic or goal.
+
+Rules:
+- Output ONLY the title, no quotes, no explanation, no punctuation except hyphens
+- Be specific: "Fix PR Shepherd CI Failures" not "Fix Bug"
+- Never apologize, ask questions, or add commentary
+- If messages discuss multiple topics, title the dominant one
+
+Examples:
+Messages: "can you check why the tests are failing on the pr shepherd branch"
+Title: Debug PR Shepherd Test Failures
+
+Messages: "I need to update the morning briefing to include slack data"
+Title: Add Slack Data to Briefing
+
+Messages: "what's the status of my open PRs across all repos"
+Title: Cross-Repo PR Status Check`;
+
 /**
  * Generate a short session name using Claude Haiku.
  * Falls back to keyword extraction if the API call fails.
@@ -230,7 +287,7 @@ export async function generateSessionName(prompts: string[]): Promise<string> {
   try {
     const client = clientFactory();
     const isVertex = client instanceof AnthropicVertex;
-    let input = prompts.join('\n');
+    let input = prompts.join('\n---\n');
     if (input.length > MAX_PROMPT_INPUT_CHARS) {
       input = input.slice(0, MAX_PROMPT_INPUT_CHARS);
     }
@@ -238,8 +295,7 @@ export async function generateSessionName(prompts: string[]): Promise<string> {
       {
         model: isVertex ? VERTEX_MODEL : AUTO_RENAME_MODEL,
         max_tokens: 20,
-        system:
-          'What is the user trying to accomplish? Generate a 3-6 word title capturing their intent or goal. Be action-oriented and specific. Return only the title, nothing else.',
+        system: NAMING_SYSTEM_PROMPT,
         messages: [
           {
             role: 'user',
@@ -251,14 +307,16 @@ export async function generateSessionName(prompts: string[]): Promise<string> {
     );
 
     const textBlock = response.content.find((b) => b.type === 'text');
-    const name = textBlock?.text?.trim() ?? '';
+    const rawName = textBlock?.text?.trim() ?? '';
+    const name = sanitizeSessionName(rawName);
 
-    if (name && name.length <= MAX_NAME_LENGTH) {
-      return name;
-    }
     if (name) {
-      return name.slice(0, MAX_NAME_LENGTH).trim();
+      return name;
     }
+
+    log.info('Haiku returned empty or rejected title, using fallback', {
+      raw: rawName.slice(0, 80),
+    });
   } catch (err: unknown) {
     log.warn('Haiku auto-rename failed, falling back to keyword extraction', {
       error: err instanceof Error ? err.message : 'unknown',
diff --git a/packages/harness/src/providers/anthropic-vertex.ts b/packages/harness/src/providers/anthropic-vertex.ts
index 17066abe..2911cd12 100644
--- a/packages/harness/src/providers/anthropic-vertex.ts
+++ b/packages/harness/src/providers/anthropic-vertex.ts
@@ -13,13 +13,6 @@ import { calculateCost } from './types.js';
 
 const log = createLogger('provider:anthropic');
 
-/** Vertex AI model name mapping (Vertex uses different naming). */
-const VERTEX_MODEL_MAP: Record<string, string> = {
-  'claude-opus-4-6': 'claude-opus-4-6@20250514',
-  'claude-sonnet-4-6': 'claude-sonnet-4-6@20250514',
-  'claude-haiku-4-5': 'claude-3-5-haiku@20241022',
-};
-
 export interface AnthropicVertexProviderOptions {
   /** GCP project ID. Falls back to ANTHROPIC_VERTEX_PROJECT_ID env var. */
   projectId?: string;
@@ -58,7 +51,7 @@ export class AnthropicVertexModelProvider implements ModelProvider {
     const systemMessages = messages.filter((m) => m.role === 'system');
     const conversationMessages = messages.filter((m) => m.role !== 'system');
 
-    const apiModel = this.isVertex ? (VERTEX_MODEL_MAP[this.model] ?? this.model) : this.model;
+    const apiModel = this.model;
 
     const response = await this.client.messages.create({
       model: apiModel,
diff --git a/server/__tests__/auto-rename.test.ts b/server/__tests__/auto-rename.test.ts
index d47c7b34..03a09272 100644
--- a/server/__tests__/auto-rename.test.ts
+++ b/server/__tests__/auto-rename.test.ts
@@ -156,21 +156,12 @@ describe('generateSessionName', () => {
 
     expect(result).toBe('Auth Bug Fix Session');
     expect(mockCreate).toHaveBeenCalledOnce();
-    expect(mockCreate).toHaveBeenCalledWith(
-      {
-        model: AUTO_RENAME_MODEL,
-        max_tokens: 20,
-        system:
-          'What is the user trying to accomplish? Generate a 3-6 word title capturing their intent or goal. Be action-oriented and specific. Return only the title, nothing else.',
-        messages: [
-          {
-            role: 'user',
-            content: 'Fix the auth bug\nUpdate login page',
-          },
-        ],
-      },
-      { timeout: 5000 },
-    );
+    const call = mockCreate.mock.calls[0];
+    expect(call[0].model).toBe(AUTO_RENAME_MODEL);
+    expect(call[0].max_tokens).toBe(20);
+    expect(call[0].messages[0].content).toContain('Fix the auth bug');
+    expect(call[0].messages[0].content).toContain('Update login page');
+    expect(call[1]).toEqual({ timeout: 5000 });
   });
 
   it('falls back to keyword extraction when API call fails', async () => {
diff --git a/server/chat.ts b/server/chat.ts
index 50579b77..6442db70 100644
--- a/server/chat.ts
+++ b/server/chat.ts
@@ -369,7 +369,9 @@ export const AVAILABLE_MODELS = [
   { id: 'claude-opus-4-7', label: 'Opus 4.7', desc: 'Adaptive thinking' },
   { id: 'claude-opus-4-7:max', label: 'Opus 4.7 Max', desc: 'Max thinking (128k)' },
   { id: 'claude-opus-4-6', label: 'Opus 4.6', desc: 'Previous Opus' },
+  { id: 'claude-opus-4-5', label: 'Opus 4.5', desc: 'Legacy Opus' },
   { id: 'claude-sonnet-4-6', label: 'Sonnet 4.6', desc: 'Balanced' },
+  { id: 'claude-sonnet-4-5', label: 'Sonnet 4.5', desc: 'Previous Sonnet' },
   { id: 'claude-haiku-4-5', label: 'Haiku 4.5', desc: 'Fastest' },
 ];