openai

/

gpt-oss-20b

like 4.87k

Follow

OpenAI 38.7k

Summary

The chat template raises TypeError on an assistant message that has content: null and tool_calls. That is the exact shape the OpenAI Chat Completions API returns for a turn that only calls tools, so appending an assistant response to the conversation and rendering it back, the normal agent loop crashes.

Reproduction

from transformers import AutoTokenizer

tok = AutoTokenizer.from_pretrained("openai/gpt-oss-120b")

tools = [{"type": "function", "function": {
    "name": "get_weather", "description": "Get weather.",
    "parameters": {"type": "object", "properties": {"city": {"type": "string"}}}}}]

messages = [
    {"role": "user", "content": "Weather in Zurich?"},
    # exactly what the API returns for a tool-only turn
    {"role": "assistant", "content": None, "tool_calls": [
        {"id": "c1", "type": "function",
         "function": {"name": "get_weather", "arguments": {"city": "Zurich"}}}]},
    {"role": "tool", "name": "get_weather", "tool_call_id": "c1", "content": "18C"},
]

tok.apply_chat_template(messages, tools=tools, add_generation_prompt=True)
# TypeError: argument of type 'NoneType' is not iterable

Cause

{%- if "content" in message %}
    {%- if "<|channel|>analysis<|message|>" in message.content or ... %}

The guard tests whether the key is present. The body then runs a substring match against its value. When content is present and null, the guard passes and "..." in None raises.

The check is a validation guard: it exists to reject <|channel|> tags smuggled in through content. A null content cannot contain a tag, so it should be skipped, not inspected.

The identical pattern appears two lines below for thinking, and fails the same way on thinking: null.

Fix

Test the value rather than the key:

-        {%- if "content" in message %}
+        {%- if message.content %}
             {%- if "<|channel|>analysis<|message|>" in message.content or "<|channel|>final<|message|>" in message.content %}

-        {%- if "thinking" in message %}
+        {%- if message.thinking %}
             {%- if "<|channel|>analysis<|message|>" in message.thinking or "<|channel|>final<|message|>" in message.thinking %}

Both are validation guards, so skipping them for empty or null values changes nothing about what gets rendered for any value that previously worked.

Verification

I rendered every snapshot of 49 conversation shapes through the template before and after the change. Every conversation that rendered before still renders, and produces a byte-identical prompt. The only difference is that the shapes which previously raised now render.

Omitting the content key entirely, or passing "", already worked before this patch. Only the spec-compliant null failed, which is why this has been easy to miss locally and hard to avoid in production.

Related, not included

Line 305 has a third instance of the same root cause:

{%- if "thinking" in message %}
    {{- "<|start|>assistant<|channel|>analysis<|message|>" + message.thinking + "<|end|>" }}

str + None also raises. I left it out because it is on the loop.last and not add_generation_prompt path that the comment describes as training-only, and because changing it alters behaviour for thinking: "" (currently emits an empty analysis block, would emit nothing). Happy to include it if you would prefer the complete fix.

Found with ChatLint, which renders chat templates against a corpus of conversation shapes and checks that they do not crash, drop content, or rewrite history.