vllm 5

Spark Scope's mini window in the DeepSeek Harness sidebar: decode, prefill and every node at a glance

DeepSeek Harness plugin: trims the oldest images from outgoing chat requests and, when the provider rejects the image count with HTTP 400, parses the limit from the error and retries with fewer images. · DSH 插件：裁剪即将发出请求中的历史图片，并在上游因图片数量返回 400 时解析该上限并按更少的图片

Neutralize tool calls with invalid JSON arguments on the wire, so one malformed model generation cannot brick a session against strict OpenAI-compatible servers (vLLM et al).

DSH plugin: read reasoning_efforts from OpenAI-compatible (vLLM) providers' /models listing and sync them into the llm-pi-ai settings section so the Web composer shows a per-model reasoning-effort picker.

DSH plugin: inject per-model sampling parameters (top_p, top_k, min_p, repeat_penalty/repetition_penalty, presence_penalty, frequency_penalty) into every chat-completions request sent to an OpenAI-compatible gateway (llama.cpp, SGLang, vLLM).