Merge pull request #1069 from sgaluza/feat/opus-4-7-adaptive-thinking

fix(cliproxy): use adaptive thinking for Opus 4.7 + add 'max' level

tests/unit/cliproxy/composite-thinking.test.ts

@@ -538,6 +538,20 @@ describe('applyThinkingConfig - composite variant integration', () => {
     expect(result.ANTHROPIC_DEFAULT_OPUS_MODEL).toBe('claude-opus-4-6-thinking(32768)');
   });
 
+  it('rewrites legacy budget suffixes for claude level-based models', () => {
+    const envVars: NodeJS.ProcessEnv = {
+      ANTHROPIC_MODEL: 'claude-opus-4-7(32768)',
+      ANTHROPIC_DEFAULT_OPUS_MODEL: 'claude-opus-4-7(32768)',
+      ANTHROPIC_DEFAULT_SONNET_MODEL: 'claude-sonnet-4-6',
+      ANTHROPIC_DEFAULT_HAIKU_MODEL: 'claude-haiku-4-5-20251001',
+    };
+
+    const result = applyThinkingConfig(envVars, 'claude' as CLIProxyProvider, 'max');
+
+    expect(result.ANTHROPIC_MODEL).toBe('claude-opus-4-7(max)');
+    expect(result.ANTHROPIC_DEFAULT_OPUS_MODEL).toBe('claude-opus-4-7(max)');
+  });
+
   it('should use codex effort suffix style when provider is codex', () => {
     const envVars: NodeJS.ProcessEnv = {
       ANTHROPIC_MODEL: 'gpt-5.3-codex',

tests/unit/cliproxy/model-catalog-compat.test.ts

@@ -58,4 +58,24 @@ describe('model-catalog compatibility lookups', () => {
 
     expect(catalog?.models.map((model) => model.id)).toEqual(['gemini-2.5-pro']);
   });
+
+  it('preserves static maxLevel when live thinking metadata omits it', () => {
+    const catalog = mergeCatalog('claude', [
+      {
+        id: 'claude-opus-4-7',
+        display_name: 'Claude Opus 4.7',
+        thinking: {
+          levels: ['low', 'medium', 'high', 'xhigh', 'max'],
+          dynamic_allowed: true,
+        },
+      },
+    ]);
+
+    expect(catalog?.models[0]?.thinking).toMatchObject({
+      type: 'levels',
+      levels: ['low', 'medium', 'high', 'xhigh', 'max'],
+      maxLevel: 'max',
+      dynamicAllowed: true,
+    });
+  });
 });

tests/unit/cliproxy/model-catalog.test.js

@@ -147,9 +147,11 @@ describe('Model Catalog', () => {
       const opus47 = MODEL_CATALOG.claude.models.find((m) => m.id === 'claude-opus-4-7');
       assert(opus47, 'Should include Claude Opus 4.7');
       assert.strictEqual(opus47.name, 'Claude Opus 4.7');
-      // Claude provider differs from AGY: zero thinking budget is disallowed,
-      // and extended (1M) context is available on the Anthropic API.
-      assert.strictEqual(opus47.thinking.zeroAllowed, false);
+      // Opus 4.7 requires adaptive thinking (type: 'levels'); manual budget_tokens
+      // is rejected by the Anthropic API with 400. Extended (1M) context is available.
+      assert.strictEqual(opus47.thinking.type, 'levels');
+      assert.deepStrictEqual(opus47.thinking.levels, ['low', 'medium', 'high', 'xhigh', 'max']);
+      assert.strictEqual(opus47.thinking.maxLevel, 'max');
       assert.strictEqual(opus47.extendedContext, true);
     });
 

tests/unit/cliproxy/thinking-validator.test.ts

@@ -29,6 +29,7 @@ describe('Thinking Validator', () => {
       expect(VALID_THINKING_LEVELS).toContain('medium');
       expect(VALID_THINKING_LEVELS).toContain('high');
       expect(VALID_THINKING_LEVELS).toContain('xhigh');
+      expect(VALID_THINKING_LEVELS).toContain('max');
       expect(VALID_THINKING_LEVELS).toContain('auto');
     });
 
@@ -38,6 +39,24 @@ describe('Thinking Validator', () => {
       expect(THINKING_LEVEL_BUDGETS.medium).toBe(8192);
       expect(THINKING_LEVEL_BUDGETS.high).toBe(24576);
       expect(THINKING_LEVEL_BUDGETS.xhigh).toBe(32768);
+      // `max` sits above xhigh — unconstrained thinking on Opus 4.7 / Mythos
+      expect(THINKING_LEVEL_BUDGETS.max).toBeGreaterThan(THINKING_LEVEL_BUDGETS.xhigh);
+    });
+
+    it('should treat max as a distinct top tier on models that list it (Opus 4.7)', () => {
+      // Opus 4.7 exposes both xhigh and max; max must not collapse into xhigh.
+      const result = validateThinking('claude', 'claude-opus-4-7', 'max');
+      expect(result.valid).toBe(true);
+      expect(result.value).toBe('max');
+      expect(result.warning).toBeUndefined();
+    });
+
+    it('should still alias max -> xhigh for models without a max level (backcompat)', () => {
+      // Codex catalog uses ['low','medium','high','xhigh'] with maxLevel 'xhigh'.
+      // User input "max" should map down to xhigh rather than be rejected.
+      const result = validateThinking('codex', 'gpt-5.4', 'max');
+      expect(result.valid).toBe(true);
+      expect(result.value).toBe('xhigh');
     });
 
     it('should export off values', () => {

tests/unit/commands/config-thinking-command.test.ts

@@ -35,6 +35,7 @@ describe('config thinking override normalization', () => {
 
   it('accepts valid levels', () => {
     expect(parseThinkingOverrideInput('High')).toEqual({ value: 'high' });
+    expect(parseThinkingOverrideInput('Max')).toEqual({ value: 'max' });
   });
 
   it('validates numeric bounds', () => {

tests/unit/cursor/cursor-anthropic-translator.test.ts

@@ -90,6 +90,16 @@ describe('translateAnthropicRequest', () => {
     ]);
   });
 
+  it('maps adaptive anthropic thinking into Cursor reasoning effort', () => {
+    const translated = translateAnthropicRequest({
+      thinking: { type: 'adaptive' },
+      output_config: { effort: 'xhigh' },
+      messages: [{ role: 'user', content: 'hello' }],
+    });
+
+    expect(translated.reasoning_effort).toBe('high');
+  });
+
   it('preserves mixed user text around tool_result blocks in order', () => {
     const translated = translateAnthropicRequest({
       messages: [

tests/unit/proxy/transformers/request-transformer-regressions.test.ts

@@ -30,6 +30,17 @@ describe('ProxyRequestTransformer regressions', () => {
     expect(result.reasoning).toEqual({ enabled: true, effort: 'high' });
   });
 
+  it('explicitly normalizes anthropic xhigh adaptive effort for OpenAI-compatible upstreams', () => {
+    const result = new ProxyRequestTransformer().transform({
+      messages: [{ role: 'user', content: 'hello' }],
+      thinking: { type: 'adaptive' },
+      output_config: { effort: 'xhigh' },
+    });
+
+    expect(result.reasoning_effort).toBe('high');
+    expect(result.reasoning).toEqual({ enabled: true, effort: 'high' });
+  });
+
   it('rejects unsupported thinking types instead of silently dropping them', () => {
     expect(() =>
       new ProxyRequestTransformer().transform({

tests/unit/targets/droid-config-manager.test.ts

@@ -143,6 +143,23 @@ describe('droid-config-manager', () => {
       expect(settings.customModels[0].extraArgs?.thinking?.budget_tokens).toBe(40960);
     });
 
+    it('writes adaptive anthropic thinking for claude-opus-4-7 overrides', async () => {
+      await upsertCcsModel('claude', {
+        model: 'claude-opus-4-7',
+        displayName: 'CCS claude',
+        baseUrl: 'https://api.anthropic.com',
+        apiKey: 'anthropic-key',
+        provider: 'anthropic',
+        reasoningOverride: 'max',
+      });
+
+      const settingsPath = path.join(tmpDir, '.factory', 'settings.json');
+      const settings = JSON.parse(fs.readFileSync(settingsPath, 'utf8'));
+      expect(settings.customModels[0].extraArgs?.thinking?.type).toBe('adaptive');
+      expect(settings.customModels[0].extraArgs?.thinking?.budget_tokens).toBeUndefined();
+      expect(settings.customModels[0].extraArgs?.output_config?.effort).toBe('max');
+    });
+
     it('should clear prior reasoning config when override disables thinking', async () => {
       await upsertCcsModel('glm', {
         model: 'glm-4.7',