Qwen3.5/Qwen3.6 35B-A3B 多轮工具调用时,Agent RL 训练中出现工具调用格式异常,导致崩溃

I am confused about the following points: In the SFT and RL stages, the tool call format of Qwen3.5/Qwen3.6 has changed from the previous JSON-based format to the XML-style format. However, during the RL training in verl, the pending stage of the `agent_loop` still uses the original YAML tool parameter definition to concatenate the tool schema, and the tool schema itself is still in JSON format. In other words, the final tool call required by the model is in the XML-style format: