orchestrator: recover from context overflow on fallback

src/backends/native/orchestrator.test.ts

@@ -605,6 +605,94 @@ describe('AgentOrchestrator', () => {
     });
   });
 
+  describe('process()', () => {
+    it('rolls back tool-loop provider errors, hard-trims on context overflow, and retries once', async () => {
+      let callCount = 0;
+      const mockClient: ModelClient = {
+        chat: vi.fn().mockImplementation(async () => {
+          callCount++;
+          if (callCount === 1) {
+            return {
+              content: '',
+              stopReason: 'tool_use',
+              usage: { inputTokens: 10, outputTokens: 5 },
+              toolCalls: [{ id: 'call_1', name: 'test.echo', args: { text: 'hi' } }],
+            } as ChatResponse;
+          }
+          if (callCount === 2) {
+            // Simulate llama.cpp context overflow buried inside an aggregated router error.
+            throw new Error(
+              'llama-server error (400): {"error":{"type":"exceedcontextsizeerror","nprompttokens":9183,"nctx":4096}}',
+            );
+          }
+          return {
+            content: 'ok',
+            stopReason: 'end_turn',
+            usage: { inputTokens: 10, outputTokens: 5 },
+          } as ChatResponse;
+        }),
+      };
+
+      const router = new ModelRouter({
+        default: mockClient,
+        fallbackChain: [],
+      });
+
+      // Minimal Session stub that supports rollback via replaceHistory().
+      const history: any[] = [];
+      const session = {
+        id: 'test',
+        addMessage: vi.fn((m: any) => { history.push(m); }),
+        getHistory: vi.fn(() => [...history]),
+        clear: vi.fn(() => { history.length = 0; }),
+        replaceHistory: vi.fn((msgs: any[]) => {
+          history.length = 0;
+          history.push(...msgs);
+        }),
+        getConfig: vi.fn(() => undefined),
+        setConfig: vi.fn(),
+        deleteConfig: vi.fn(),
+      } as any;
+
+      const registry = new ToolRegistry();
+      registry.register({
+        name: 'test.echo',
+        description: 'echo',
+        inputSchema: { type: 'object', properties: { text: { type: 'string' } }, required: ['text'] },
+        execute: async (args: any) => ({ success: true, output: String(args.text ?? '') }),
+      });
+
+      const hooks = new HookEngine({ confirm: [], log: [], silent: [] });
+      const executor = new ToolExecutor(registry, hooks);
+
+      const orchestrator = new AgentOrchestrator({
+        modelRouter: router,
+        systemPrompt: 'You are helpful.',
+        session,
+        toolRegistry: registry,
+        toolExecutor: executor,
+        primaryTier: 'default',
+        delegation: {
+          compaction: 'fast',
+          memory_extraction: 'default',
+          classification: 'complex',
+          tool_summarisation: 'default',
+          complex_reasoning: 'complex',
+        },
+        maxDelegationDepth: 3,
+      });
+
+      const res = await orchestrator.process('hello');
+      expect(res).toBe('ok');
+
+      // Ensure we didn't persist the low-level error string in history.
+      const textHistory = history
+        .map(m => (typeof m.content === 'string' ? m.content : ''))
+        .join('\n');
+      expect(textHistory).not.toContain('Error in tool loop');
+    });
+  });
+
   describe('setModelTier()', () => {
     it('sets model tier on primary agent', () => {
       const orchestrator = new AgentOrchestrator({