PR: AI User as reviewer causes error (OD-2746)

PR: AI User as reviewer causes error #2746

jbauer opened 2 months ago

AI user uses Mistral.ai with model devstral-2512 and only a very basic system prompt.

Adding the AI user as a reviewer results in the exception below. Before 15.0.1 I got "empty response" comments on the PR instead of the exception:

2026-04-27 18:31:17,333 ERROR [pool-3-thread-99] i.o.s.s.impl.DefaultUserService Error running AI task
dev.langchain4j.exception.InvalidRequestException: {"object":"error","message":"Assistant message must have either content or tool_calls, but not none.","type":"invalid_request_assistant_message","param":null,"code":"3240","raw_status_code":400}
	at dev.langchain4j.internal.ExceptionMapper$DefaultExceptionMapper.mapHttpStatusCode(ExceptionMapper.java:71)
	at dev.langchain4j.internal.ExceptionMapper$DefaultExceptionMapper.mapException(ExceptionMapper.java:44)
	at dev.langchain4j.internal.ExceptionMapper.withExceptionMapper(ExceptionMapper.java:31)
	at dev.langchain4j.internal.RetryUtils.lambda$withRetryMappingExceptions$2(RetryUtils.java:325)
	at dev.langchain4j.internal.RetryUtils$RetryPolicy.withRetry(RetryUtils.java:207)
	at dev.langchain4j.internal.RetryUtils.withRetry(RetryUtils.java:262)
	at dev.langchain4j.internal.RetryUtils.withRetryMappingExceptions(RetryUtils.java:325)
	at dev.langchain4j.internal.RetryUtils.withRetryMappingExceptions(RetryUtils.java:308)
	at dev.langchain4j.model.openai.OpenAiChatModel.doChat(OpenAiChatModel.java:149)
	at dev.langchain4j.model.chat.ChatModel.chat(ChatModel.java:46)
	at io.onedev.server.service.impl.DefaultUserService.lambda$execute$7(DefaultUserService.java:659)
	at java.base/java.util.concurrent.Executors$RunnableAdapter.call(Executors.java:539)
	at java.base/java.util.concurrent.FutureTask.run(FutureTask.java:264)
	at io.onedev.server.security.SecurityUtils.lambda$inheritSubject$1(SecurityUtils.java:648)
	at java.base/java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1136)
	at java.base/java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:635)
	at java.base/java.lang.Thread.run(Thread.java:840)
Caused by: dev.langchain4j.exception.HttpException: {"object":"error","message":"Assistant message must have either content or tool_calls, but not none.","type":"invalid_request_assistant_message","param":null,"code":"3240","raw_status_code":400}
	at dev.langchain4j.http.client.jdk.JdkHttpClient.execute(JdkHttpClient.java:53)
	at dev.langchain4j.model.openai.internal.SyncRequestExecutor.execute(SyncRequestExecutor.java:20)
	at dev.langchain4j.model.openai.internal.RequestExecutor.executeRaw(RequestExecutor.java:44)
	at dev.langchain4j.model.openai.OpenAiChatModel.lambda$doChat$0(OpenAiChatModel.java:150)
	at dev.langchain4j.internal.ExceptionMapper.withExceptionMapper(ExceptionMapper.java:29)
	... 14 common frames omitted

OneDev commented 2 months ago

State changed as build OD-7309 (15.0.3) is successful
OneDev changed state to 'Closed' 2 months ago
Previous Value Current Value
Open

Closed
jbauer commented 2 months ago

@robin Thanks for fixing. I now have "empty reponse" comments again with the mentioned model. I temporarly switched back to OpenAI with some GPT variant to check if the AI user then creates a proper review and well..it does.

I noticed that GPT first creates a reaction with "request for changes" or "approve" and after that the code review comment is added. With Mistral I directly receive a comment with "empty response" text. Could it be that Mistral might not call the tool for "request for changes / approve" correctly or only provides review text without any tool call and OneDev code logic might swallow the review text?

I know prompt writing is super annoying with all the models and their different behaviors and I would be fine if your stance is that you cannot satisfy everyone with their model choice as long as internal prompts are not configurable. But I have a hard time to imagine that Mistral really doesn't respond back anything useful as the model has 123 billion parameters (https://docs.mistral.ai/models/model-cards/devstral-2-25-12)
Robin Shen commented 2 months ago

I tested devstral-2512 via openrouter with some example PRs and it works quite solid.

In 15.0.5, review action is placed under comments like normal user review. Also LLM is allowed to leave comment only if it needs more context for review. And later it can be mentioned again to review the PR.
Login to comment

Previous Value	Current Value
Open	Closed

Type	Bug
Priority	Normal
Assignee	Robin Shen
Affected Versions	15.0.1
Labels	No labels

Issue Votes (0)

Watchers (3)

Reference

OD-2746