thinking budget doesn't work correctly in long prompts

curl --location --request POST 'http://localhost:8111/v1/chat/completions' \
--header 'Content-Type: application/json' \
--data-raw '{
"model": "seed_oss",
"stream": false,
"messages": [
{
"role": "user",
"content": prompt
}
],
"chat_template_kwargs": {
"thinking_budget": 4096
}
}'
When the prompt is relatively long, it seems that the thinking_budget is not working correctly. The beginning of the output is:

“content”:"[seed:think](https://panda.baidu-int.com/llm?appId=gpt-4o)\n[seed:cot_budget_reflect](https://panda.baidu-int.com/llm?appId=gpt-4o)I have used 516 tokens, and there are 3580 tokens remaining for use.\n

why is there no thinking?

Start by reproducing the POST request to /v1/chat/completions with a long prompt and chat_template_kwargs set to a thinking_budget of 4096. Trace how the seed_oss chat template handles the budget, then verify that the response's thinking behavior matches the requested budget for long prompts.

Written by the indexing model from the issue text.