🤖 Re-enable thinking for Anthropic models in compaction (#324)

Now that tools are disabled during compaction, Anthropic models can use
thinking without conflicts. Also includes continue message in compaction
prompt to help focus the summarization.

## Changes

- Remove Anthropic-specific thinking restriction in compactionOptions.ts
- Add continue message context to compaction prompt when provided
- Update tests to reflect new behavior

## Testing

- ✅ Unit tests pass (compactionOptions.test.ts)
- ✅ Slash command tests pass (compact.test.ts)
- ✅ Type checking passes

_Generated with `cmux`_

Author: ammar-agent

src/utils/chatCommands.ts

@@ -135,7 +135,12 @@ export function prepareCompactionMessage(options: CompactionOptions): {
 } {
   const targetWords = options.maxOutputTokens ? Math.round(options.maxOutputTokens / 1.3) : 2000;
 
-  const messageText = `Summarize this conversation into a compact form for a new Assistant to continue helping the user. Use approximately ${targetWords} words.`;
+  // Build compaction message with optional continue context
+  let messageText = `Summarize this conversation into a compact form for a new Assistant to continue helping the user. Use approximately ${targetWords} words.`;
+
+  if (options.continueMessage) {
+    messageText += `\n\nThe user wants to continue with: ${options.continueMessage}`;
+  }
 
   // Handle model preference (sticky globally)
   const effectiveModel = resolveCompactionModel(options.model);

src/utils/messages/compactionOptions.test.ts

@@ -31,22 +31,20 @@ describe("applyCompactionOverrides", () => {
     expect(result.model).toBe("anthropic:claude-haiku-4-5");
   });
 
-  it("sets thinking to off for Anthropic models", () => {
-    const compactData: CompactionRequestData = {
+  it("preserves workspace thinking level for all models", () => {
+    // Test Anthropic model
+    const anthropicData: CompactionRequestData = {
       model: "anthropic:claude-haiku-4-5",
     };
-    const result = applyCompactionOverrides(baseOptions, compactData);
-
-    expect(result.thinkingLevel).toBe("off");
-  });
+    const anthropicResult = applyCompactionOverrides(baseOptions, anthropicData);
+    expect(anthropicResult.thinkingLevel).toBe("medium");
 
-  it("preserves workspace thinking level for non-Anthropic models", () => {
-    const compactData: CompactionRequestData = {
+    // Test OpenAI model
+    const openaiData: CompactionRequestData = {
       model: "openai:gpt-5-pro",
     };
-    const result = applyCompactionOverrides(baseOptions, compactData);
-
-    expect(result.thinkingLevel).toBe("medium");
+    const openaiResult = applyCompactionOverrides(baseOptions, openaiData);
+    expect(openaiResult.thinkingLevel).toBe("medium");
   });
 
   it("applies maxOutputTokens override", () => {

src/utils/messages/compactionOptions.ts

@@ -26,14 +26,11 @@ export function applyCompactionOverrides(
   // Use custom model if specified, otherwise use workspace default
   const compactionModel = compactData.model ?? baseOptions.model;
 
-  // Anthropic models don't support thinking, always use "off"
-  // Non-Anthropic models keep workspace default (backend will enforce policy)
-  const isAnthropic = compactionModel.startsWith("anthropic:");
-
   return {
     ...baseOptions,
     model: compactionModel,
-    thinkingLevel: isAnthropic ? "off" : baseOptions.thinkingLevel,
+    // Keep workspace default thinking level - all models support thinking now that tools are disabled
+    thinkingLevel: baseOptions.thinkingLevel,
     maxOutputTokens: compactData.maxOutputTokens,
     mode: "compact" as const,
     toolPolicy: [], // Disable all tools during compaction