Articles tagged "Model adaptation"

牛马提示词：多模态模型提示词工程专家

2026-10-27 Big Model Skills multimodal model Prompt Engineering AI生成 Model adaptation Prompt optimization

介绍牛马提示词这个多模态模型提示词工程专家 Skill，覆盖全球主流生图/生视频/多模态模型，提供提示词优化、跨模型转换、机制分析能力，帮助开发者减少提示词适配成本。

DeepSeek Harness 插件：nethiel/dsh-reasoning-effort-sniffer

2026-10-26 DeepSeek Harness plugin DeepSeek DSH Plugin Inference Strength Model adaptation

该插件通过实测嗅探每个第三方供应商的端点支持哪些思考强度档位，自动补全档位声明与 wire 标记，使 DSH 模型选择器的强度控件对非官方供应商也生效。支持按供应商开关强度与嗅探，内置限速保护与安全隐私设计。

a42null/dsh-opencode-zen-plugin：DeepSeek Harness 的 OpenCode Zen 模型适配插件

2026-10-20 DeepSeek Harness plugin DeepSeek DSH Plugin OpenCode Zen Model adaptation

`a42null/dsh-opencode-zen-plugin` 是 DeepSeek Harness (DSH) 的一个插件，旨在将 OpenCode Console (Zen) 的模型集成至 DSH 调用链。该插件实现了零配置模型目录获取、自动端点路由（适配 OpenAI、Anthropic、Google 等多种协议）及元数据同步。它支持流式输出、推理等级透传、工具调用及 Vision 模型，并通过缓存和错误优化提升体验。用户通过包管理器安装后，配置 API Key 即可在 DSH 中使用 Zen 模型。

dsh-tool-adapt: Unofficial Model Compatibility and Security Adaptation Layer for DeepSeek Harness Web

2026-10-07 DeepSeek Harness plugin DeepSeek DSH dsh-tool-adapt Plugin Model adaptation

`dsh-tool-adapt` is a compatibility and security adaptation plugin for the unofficial model families of DeepSeek Harness Web. Through core features such as dead-session protection, model-family convention injection, circuit breaking, and gateway compatibility diagnosis and repair, it addresses pain points including gateway compatibility errors, session deadlocks, and disordered permission management. After installation, a process restart is required. It supports hot configuration management via the settings panel and allows displaying status capsules to monitor operational status.

dsh-plugin-judge: DSH Plugin Value Judge

2026-09-28 DeepSeek Harness plugin DeepSeek DSH plugin dsh-plugin-judge Plugin Audit Model adaptation

`dsh-plugin-judge` is a plugin value judgment tool in the DeepSeek Harness ecosystem, designed to address the issue that a plugin's value can fluctuate with model changes. It provides three core capabilities: pre-installation review (static scanning + LLM judgment), post-installation audit (risk scoring), and reminders for re-evaluation when switching models. Through commands or the settings panel, users can assess plugin suitability and manage constraint-related risks, ensuring that plugins do not impair task performance due to rule restrictions or model upgrades. Installation requires environment prerequisites such as Node.js >= 18; judgment results are for reference only, so install third-party code cautiously.