gpt-oss:20b-cloud damaged output after last update

Examples of responses from cloud (local ok)

User wants all ingredients from recipe. We have recipe_1.txt. We need to extract ingredients. Use task_on_text. Provide result named e.g., ingredients_list. Then answer with named link.assistantcommentary to=functions.task_on_text json{"task":"extract all ingredients from Assorted dried beans and peas Ham Hock 1 can tomatoes -- (28 oz.) 2 Onions -- chopped 4 stalks celery -- chopped 2 cloves garlic -- minced Salt and pepper to taste Bay leaves Thyme Basil 1 pound sausage -- sliced 2 boneless Chicken breast 1/2 cup red wine 1/2 cup chopped parsley

Here starting tool call with params, but not parsed right:

{'content': '{"params":{},"result_name":"description_Jane_Doe_test","task":"Describe the content of {{Jane_Doe_test.pdf}}"}<|call|>commentary<|message|>We need to output the result link.Description of the document\n\nThe file {{Jane_Doe_test.pdf}} contains a short English text that serves as a test for border detection and text extraction. \nKey elements in the document:\n\n| Element | Details |\n|---------|---------|\n| Text | Multiple mentions of the name “Jane\u202fDoe” appear across two pages. |\n| Table | A simple table with two cells. |\n| List | A list of name components: “Name, Jane, Second name, Doe, number\u202f1”. |\n\nThe document is intentionally minimal, designed to verify that the system can correctly identify and extract text, tables, and lists from a PDF.', 'additional_kwargs': {}, 'response_metadata': {'model': 'gpt-oss:20b-cloud', 'created_at': '2026-06-11T10:54:03.975238905Z', 'done': True, 'done_reason': 'stop', 'total_duration': 3091165726, 'load_duration': None, 'prompt_eval_count': 971, 'prompt_eval_duration': None, 'eval_count': 268, 'eval_duration': None, 'logprobs': None, 'model_name': 'gpt-oss:20b-cloud', 'model_provider': 'ollama'}, 'type': 'ai', 'name': None, 'id': 'lc_run--019eb651-4586-79a2-9465-b7d95827a241-0', 'tool_calls': [], 'invalid_tool_calls': [], 'usage_metadata': {'input_tokens': 971, 'output_tokens': 268, 'total_tokens': 1239}}

Inside content part of thinking:

'content': 'User asks "How are you?" It's a casual greeting. We should respond politely.assistantfinalI’m doing great—thanks for asking! How can I help you today?'

Must call tool but got nothing:

{'content': '', 'additional_kwargs': {}, 'response_metadata': {'model': 'gpt-oss:20b-cloud', 'created_at': '2026-06-11T10:49:55.184364541Z', 'done': True, 'done_reason': 'stop', 'total_duration': 751245656, 'load_duration': None, 'prompt_eval_count': 884, 'prompt_eval_duration': None, 'eval_count': 54, 'eval_duration': None, 'logprobs': None, 'model_name': 'gpt-oss:20b-cloud', 'model_provider': 'ollama'}, 'type': 'ai', 'name': None, 'id': 'lc_run--019eb64d-8268-7510-bad0-50feaffb958d-0', 'tool_calls': [], 'invalid_tool_calls': [], 'usage_metadata': {'input_tokens': 884, 'output_tokens': 54, 'total_tokens': 938}}

Looks like tool call but after goes thinking text:

{"params":{},"result_name":"description_Jane_Doe_test","task":"Describe the content of {{Jane_Doe_test.pdf}}"}<|call|>commentary<|message|>We need to output the result link.Description of the document

The file {{Jane_Doe_test.pdf}} contains a short English text that serves as a test for border detection and text extraction.
Key elements in the document:

Start by reproducing the listed tool-calling and empty-output cases with gpt-oss:20b-cloud, then compare the response content, tool_calls, and invalid_tool_calls with a local run. Trace how the Python library handles these response fields; done means cloud responses no longer expose malformed tool-call, thinking, or empty content output.