@magpie-community/middleware-think-tags

A model's thinking, where its reply's text is, as magpie gateway middleware. It either takes thinking out of the text, or puts it in (New API's thinking_to_content).

Install it from magpie's Plugins › Discover, or with magpie plugin add @magpie-community/middleware-think-tags. Set its options under Plugins › Installed › Options, or with magpie plugin options think-tags '<json>'.

Options

{ "mode": "strip" }

strip (the default). Some models (open reasoning models served by vLLM, Ollama and many relays) write their thinking into the reply as <think>…</think>. This mode takes that block out, so an agent doesn't keep it in its history and send it back every turn.

• Only a block at the start of the text is taken out. A <think> the model writes later, in code for example, stays.
• A tag split across streamed chunks is still found.
• It works on Anthropic, Chat and Responses replies, streamed or whole.

to_content. A Chat reply's reasoning_content (or reasoning) is put into its text as <think>…</think>, for a client that shows only the text.

• This mode works on Chat replies only. An Anthropic thinking block carries a signature, and turned into text it would fail the next turn.
• It changes what the agent keeps and sends back, so it suits a chat client better than a coding agent.

中文：strip 去掉回复开头模型写出的 <think>…</think>；to_content 把 Chat 回复的 reasoning_content 放进正文（New API 的 thinking_to_content）。