BaseChatTemplateParser strips every TokenConfig string from message content to block control-token injection. For Qwen that set included tool_start_token ("\n<tool_call>\n") and tool_end_token ("\n</tool_call>"), which are ordinary text in tool-use system prompts and in assistant tool-call turns. Stripping them removed the <tool_call> wrapper from the function-calling format example that the Qwen3.5 chat template writes into the system prompt, so the model emitted bare <function=...> blocks and often failed to end its turn. In agentic RL on DeepSWE with Qwen3.5-35B-A3B, about a fifth of rollouts degenerated into repeated </function> until they hit the response-length cap, which more than doubled step time.

[copybara-service]

BaseChatTemplateParser strips every TokenConfig string from message content to block control-token injection. For Qwen that set included tool_start_token ("\n<tool_call>\n") and tool_end_token ("\n</tool_call>"), which are ordinary text in tool-use system prompts and in assistant tool-call turns. Stripping them removed the <tool_call> wrapper from the function-calling format example that the Qwen3.5 chat template writes into the system prompt, so the model emitted bare <function=...> blocks and often failed to end its turn. In agentic RL on DeepSWE with Qwen3.5-35B-A3B, about a fifth of rollouts degenerated into repeated </function> until they hit the response-length cap, which more than doubled step time.

Exclude tool_start_token and tool_end_token from sanitization. The exclusion is in the base class, so Gemma4 now also keeps its <|tool_call> / <tool_call|> markup and Llama no longer strips its tool header from content; turn terminators such as <|im_end|> and <|eot_id|> are still stripped.

Also render runs of consecutive tool messages the way the Qwen chat template does, as a single user turn with one <tool_response> block per message, which changes the token layout for turns with several tool calls. QwenChatTemplateParser.preprocess_messages marks such runs so that tokenize_and_generate_masks, which renders one message at a time, produces the same text as parse.