fix: handle thinking model reasoning-only responses as valid content

Thinking models (Kimi K2.5, DeepSeek-R1, QwQ) may produce only
reasoning_content with no regular text content or tool calls. Previously
this was treated as an empty/failed response, triggering the "language
model did not provide any assistant messages" error and unnecessary
retries.

Changes:
- Task.ts: Include reasoningMessage in the content check so
  reasoning-only responses are recognized as valid model output
  instead of triggering the empty response error path
- openai.ts: Handle both "reasoning_content" (standard) and
  "reasoning" (Ollama /v1/) field names in streaming responses
- openai.ts: Yield reasoning content from non-streaming responses

Fixes #10603
Related: #9959, #10064, #9551

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

src/api/providers/openai.ts

@@ -199,10 +199,16 @@ export class OpenAiHandler extends BaseProvider implements SingleCompletionHandl
 					}
 				}
 
-				if ("reasoning_content" in delta && delta.reasoning_content) {
+				// Handle reasoning/thinking tokens from various providers:
+				// - reasoning_content: DeepSeek, OpenAI (standard)
+				// - reasoning: Ollama /v1/ (non-standard, used by Kimi K2.5, etc.)
+				const reasoningText =
+					("reasoning_content" in delta && (delta as any).reasoning_content) ||
+					("reasoning" in delta && (delta as any).reasoning)
+				if (reasoningText) {
 					yield {
 						type: "reasoning",
-						text: (delta.reasoning_content as string | undefined) || "",
+						text: String(reasoningText),
 					}
 				}
 
@@ -260,6 +266,18 @@ export class OpenAiHandler extends BaseProvider implements SingleCompletionHandl
 				}
 			}
 
+			// Yield reasoning content from non-streaming responses
+			// Thinking models (Kimi K2.5, DeepSeek-R1) may return reasoning
+			// in the message alongside (or instead of) regular content
+			const messageAny = message as any
+			const nonStreamReasoning = messageAny?.reasoning_content || messageAny?.reasoning
+			if (nonStreamReasoning) {
+				yield {
+					type: "reasoning",
+					text: String(nonStreamReasoning),
+				}
+			}
+
 			yield {
 				type: "text",
 				text: message?.content || "",

src/core/task/Task.ts

@@ -3401,14 +3401,19 @@ export class Task extends EventEmitter<TaskEvents> implements TaskLike {
 				// the assistant message is already in history. Otherwise, tool_result blocks would appear
 				// BEFORE their corresponding tool_use blocks, causing API errors.
 
-				// Check if we have any content to process (text or tool uses)
+				// Check if we have any content to process (text, reasoning, or tool uses)
+				// Thinking models (e.g., Kimi K2, DeepSeek-R1, QwQ) may produce only
+				// reasoning_content with no regular text content. This should not be
+				// treated as an empty/failed response — the model did respond, just
+				// entirely in reasoning tokens.
 				const hasTextContent = assistantMessage.length > 0
+				const hasReasoningContent = reasoningMessage.length > 0
 
 				const hasToolUses = this.assistantMessageContent.some(
 					(block) => block.type === "tool_use" || block.type === "mcp_tool_use",
 				)
 
-				if (hasTextContent || hasToolUses) {
+				if (hasTextContent || hasToolUses || hasReasoningContent) {
 					// Reset counter when we get a successful response with content
 					this.consecutiveNoAssistantMessagesCount = 0
 					// Display grounding sources to the user if they exist