Release 3.6.0 · TarsLab/obsidian-tars

🆕 3.6.0 New providers

Model lists now come from the provider. Claude, DeepSeek, Qwen, OpenAI, Doubao, Ollama, Gemini and Zhipu now ask their own API which models exist, rather than offering a list baked into the plugin that goes stale. A model missing from the list can still be set under "Override input parameters". (#119, #121)
Thinking appears in a callout for more models, however they send it. A provider reports reasoning either in its own field or wrapped in <think> tags, and some model families use each — Zhipu uses both. Both are handled now, and the blank lines models leave around their thinking are trimmed. (#116)
Adding an assistant opens its own settings page, where its tag and API key are.
Azure asks for a deployment name, in a text field. It was a dropdown of model names, which can never be right: an Azure deployment is named by whoever created it.

ModelScope and other gateways no longer fail CORS. The bundled OpenAI SDK stamps X-Stainless-* telemetry headers onto every request, and a gateway whose allowlist does not name them rejects the request entirely. They are no longer sent. (#108)
Kimi and Grok no longer stop an answer part-way. A streamed frame that arrived split across two network reads was parsed in halves, which ended the stream mid-sentence and reported nothing. It only happened when the packets landed that way, which is why it looked random.
OpenRouter no longer reports a rejected request as a success. An HTTP error produced no text and no error, and the editor still said the text had been generated.
Image generation works on mobile. It used a Node API that mobile does not have, and threw every time.
"Override input parameters" reaches Claude, Gemini and image generation. Every key beyond a provider's own settings was being dropped there in silence.
A provider's own settings no longer leak into its request body — only declared parameters are sent.
Adding a second assistant of the same provider now names it (Qwen2), instead of saving it with no tag — which left it impossible to trigger, while still showing an entry reading # : in the command palette.
Clearing an assistant's tag now says so instead of silently keeping the previous one while the field looks empty.
Azure says which field is missing when its endpoint is blank, rather than failing obscurely.

模型列表改为向服务商获取。 Claude、DeepSeek、Qwen、OpenAI、Doubao、Ollama、Gemini、Zhipu 现在调用官方接口获取模型列表，不再使用插件里写死、会过期的列表。列表里没有的模型，仍然可以在「覆盖输入参数」里设置。(#119, #121)
更多模型的推理内容会显示为 callout，无论它用哪种方式传出来。 服务商要么用单独的字段返回推理内容，要么用 <think> 标签包裹，有些系列两种都用——智谱就是。现在两种都支持，模型在推理内容前后留下的多余空行也会被去掉。(#116)
新增 AI 助手后直接进入它自己的设置页，标签和 API key 都在那里。
Azure 的「模型」改为填部署名称的输入框。 原来是模型名下拉框，那不可能对：Azure 的部署名是创建者自己起的。

ModelScope 等网关不再被 CORS 拦截。 内置的 OpenAI SDK 会给每个请求打上 X-Stainless-* 遥测请求头，而允许列表里没有列出它们的网关会直接拒绝整个请求。现在不再发送这些头。(#108)
Kimi 和 Grok 的回答不再中途断掉。 一个流式数据帧如果被拆到两次网络读取里，就会被拆成两半去解析，导致回答从中间中断且不报任何错。它只在数据包恰好那样落下时发生，所以看起来像是随机的。
OpenRouter 不再把被拒绝的请求报告成成功。 HTTP 错误既不产生文本也不产生报错，而编辑器仍然提示「文本生成成功」。
图像生成在手机端可用了。 它用了移动端没有的 Node API，此前每次都会抛错。
「覆盖输入参数」对 Claude、Gemini 和图像生成生效了。 此前这三处会静默丢弃服务商自身设置以外的所有键。
服务商自身的设置不再混进请求体——只发送声明过的参数。
给同一个服务商添加第二个助手时会自动命名（如 Qwen2），不再存成空标签——空标签的助手永远无法触发，却仍会在命令面板留下一条 # : 的条目。
清空助手标签时会给出提示，而不是在输入框看起来是空的情况下静默保留原值。
Azure 端点为空时会指明缺哪一项，不再以晦涩的方式失败。