Reasoning models: thinking tags not parsed correctly when opening tag is in prompt/template
Reasoning models: thinking tags not parsed correctly when opening tag is in prompt/template: a task in LegoFlow-SWE (Harbor dataset). When using reasoning models (like Qwen-3, GLM) that output <thinking ...</thinking or 🙇..<／thinking blocks, the reasoning content is not correctly separated into…
The task
When using reasoning models (like Qwen-3, GLM) that output `<thinking>...</thinking>` or `🙇..<／thinking>` blocks, the reasoning content is not correctly separated into the `reasoning_content` field of the OpenAI‑compatible API response. Instead, the entire block remains in the `content` field.
Part of Lego-X/LegoFlow-SWE.