I’m using a locally served DeepSeek coding model through a chat client. Occasionally, a generated function ends halfway through, even though the response is much shorter than the configured output limit.

The client has a custom stop list inherited from another model setup, so I’m wondering whether custom stop sequences could be cutting off otherwise valid output. I haven’t yet established whether the problem is in the client, the serving layer, or the model’s own completion.

1. Which response fields or server logs should I inspect to distinguish a stop condition from a token limit?
2. Can a stop sequence match text inside a code comment or string literal?
3. What should I record about the model and chat template before comparing runs?

I can test the same prompt directly against the server and through the client. What would make that comparison useful for diagnosing premature generation termination, given that two runs may produce different code?