Handle Ollama responses without message wrappers

Treat Ollama chat messages as optional so qwen3/template fallback responses do not crash when the API returns a bare completion payload.

Add regression coverage for bare non-stream responses and bare streaming done chunks.

Co-authored-by: Autohand Evolve <code-noreply@autohand.ai>

Author: igorcosta

src/providers/OllamaProvider.ts

@@ -42,7 +42,7 @@ interface OllamaRequestToolCall {
 }
 
 interface OllamaChatResponse {
-    message: {
+    message?: {
         role: string;
         content: string;
         tool_calls?: OllamaToolCall[];
@@ -247,11 +247,12 @@ export class OllamaProvider implements LLMProvider {
         }
 
         const data = await response.json() as OllamaChatResponse;
+        const message = data.message ?? { role: 'assistant', content: '' };
 
         // Parse tool calls if present (Ollama returns arguments as object, not string)
         let toolCalls: LLMToolCall[] | undefined;
-        if (data.message.tool_calls && Array.isArray(data.message.tool_calls)) {
-            toolCalls = data.message.tool_calls.map((tc: OllamaToolCall, index: number) => {
+        if (message.tool_calls && Array.isArray(message.tool_calls)) {
+            toolCalls = message.tool_calls.map((tc: OllamaToolCall, index: number) => {
                 let argumentsStr: string;
                 try {
                     // Ollama returns arguments as object, convert to JSON string for consistency
@@ -281,7 +282,7 @@ export class OllamaProvider implements LLMProvider {
         return {
             id: `ollama-${Date.now()}`,
             created: Math.floor(new Date(data.created_at).getTime() / 1000),
-            content: data.message.content,
+            content: message.content,
             toolCalls,
             finishReason: toolCalls?.length ? 'tool_calls' : 'stop',
             usage,
@@ -550,7 +551,7 @@ export class OllamaProvider implements LLMProvider {
                 for (const line of lines) {
                     try {
                         const data: OllamaChatResponse = JSON.parse(line);
-                        fullContent += data.message.content;
+                        fullContent += data.message?.content ?? '';
                         lastData = data;
                         // Ollama signals completion via the JSON "done" field
                         if (data.done) {

tests/providers/OllamaProvider.test.ts

@@ -155,6 +155,25 @@ describe('OllamaProvider', () => {
             );
         });
 
+        it('handles bare Ollama chat responses without a message wrapper', async () => {
+            const p = new OllamaProvider(config, { maxRetries: 0 });
+            global.fetch = vi.fn().mockResolvedValue({
+                ok: true,
+                json: async () => ({
+                    created_at: '2024-11-21T10:30:00Z',
+                    done: true
+                })
+            });
+
+            const response = await p.complete({
+                messages: [{ role: 'user', content: 'Hello' }]
+            });
+
+            expect(response.content).toBe('');
+            expect(response.toolCalls).toBeUndefined();
+            expect(response.finishReason).toBe('stop');
+        });
+
         it('should handle streaming responses', async () => {
             const mockStream = new ReadableStream({
                 start(controller) {
@@ -181,6 +200,30 @@ describe('OllamaProvider', () => {
             expect(response.content).toContain('Hello');
         });
 
+        it('honors bare Ollama stream chunks without a message wrapper', async () => {
+            const mockStream = new ReadableStream({
+                start(controller) {
+                    controller.enqueue(new TextEncoder().encode(
+                        '{"created_at":"2024-11-21T10:30:00Z","done":true}\n'
+                    ));
+                    controller.close();
+                }
+            });
+
+            global.fetch = vi.fn().mockResolvedValue({
+                ok: true,
+                body: mockStream
+            });
+
+            const response = await provider.complete({
+                messages: [{ role: 'user', content: 'Hello' }],
+                stream: true
+            });
+
+            expect(response.content).toBe('');
+            expect(response.finishReason).toBe('stop');
+        });
+
         // -----------------------------------------------------------------------
         // Error handling tests (TDD — these fail before the fix is implemented)
         // -----------------------------------------------------------------------