家居维修预约：用 Agent Skill 整理渠道、价格与防坑要点
家居维修预约是 SkillHub 社区 Skill（@user_70c2f807/home-repair-booking，v1.0.1，life-service），通过 SKILL.md 在用户提及找师傅、上门维修等场景时，按三步收集故障信息并输出官方/平台/社区分级渠道、价格区间与防坑清单。技能声明不能直接预约、价格仅供参考且无需 API Key。本文介绍其能力边界、SkillHub CLI 与 Zip 安装方式（含 Cursor 等 skills 目录），以及典型对话示例与使用注意。
Read Moredouyin downloader：在 AI 对话里下载抖音视频与图文
douyin downloader 是 SkillHub 社区 Skill（@user_09701b7c/douyin-zh，v1.0.6），在用户分享抖音链接或提出下载需求时，帮助 Agent 下载抖音视频与图文笔记；标注免登录、免注册其他 API、无需 API Key。本文介绍其定位、触发词、通过官方提示词或 skillhub CLI 的安装方式，以及典型对话用法与合规注意。
Read Morecheng_rec_llm_skills：推荐广告与大模型技术情报推送
cheng_rec_llm_skills 是 SkillHub 社区技能（@user_d28b18a4/chengrecllmskill，v1.0.0，ai-agent），通过 SKILL.md 定义推荐广告与大模型两条线的每日情报流程：微信与知乎深度长文检索、GitHub 开源趋势扫描，以及日更、月总结与历史回溯输出格式。标签 requires_api_key=false。本文说明其监控目标池、任务 A/B 检索语法与安装命令（skillhub install --dir），并提示对 OpenClaw、/data/monitoring_log.md 的依赖与落地注意。
Read Morectf ai ml：面向 CTF 的 AI / 机器学习攻防 Skill
ctf ai ml（@user_257ad2a8/ctf-ai-ml，v1.0.0）是 SkillHub 社区 it-ops-security 类 Skill，面向 CTF 中的对抗样本、模型提取、提示词注入、成员推断、投毒、LoRA 与 LLM 越狱等场景。本文基于官网简介与 SkillHub 安装文档，说明定位、能力范围、skillhub install 命令及 Cursor/Claude Code 等 skills 目录，并交代适用人群与使用注意。requires_api_key=false。
Read MoreDeepSeek Harness (DSH) Real-Time Output Repetition Guard
`dsh-dupguard` is a DeepSeek Harness (DSH) plugin that detects and halts consecutive repeated output in real time during the model inference stream. By default, it triggers an interruption after 10 or more consecutive repetitions, with relaxed handling for code blocks and similar contexts. It allows configuring thresholds, allowlists, and global notifications via the settings page. Installation requires DSH ≥ 0.1.2-rc.1 and Node.js ≥ 20, and DSH must be restarted for the plugin to take effect.
Read Moredsh-plugin-task-runner: Task Decomposition Mode
The `dsh-plugin-task-runner` plugin in DeepSeek Harness addresses the limited context windows of local large models through a “task decomposition mode.” The main agent decomposes tasks, subagents execute them in parallel within isolated, short contexts, and the main agent aggregates the results. The plugin supports concurrency control, external memory persistence, budget control, and cloud fallback, making it suitable for handling complex, long-chain tasks.
Read Morezhangyh-x/dsh-model-pricing - LLM Pricing and Policy Quick Reference
The `zhangyh-x/dsh-model-pricing` plugin for DeepSeek Harness (DSH) is an “online tool” designed to provide a quick-reference table for querying large model pricing and policies. It helps developers stay informed about frequent pricing changes and policy updates. Users can obtain and install it from its GitHub repository or plugin catalog page. It is suitable for developers who need to frequently check large model pricing or the latest policies. Note that the plugin runs with the permissions of the current DSH process; it is recommended to review the source code before installation to ensure compliance with usage guidelines.
Read Moredsh-steer: Guiding and Interrupting Large Models During Their Thinking Process
dsh-steer is a DSH plugin maintained by CAI-MH that allows real-time injection of corrective information or interruption of an ongoing conversation while the model is running, without waiting for the turn to end. It provides two core modes, “send guidance” and “interrupt,” and supports visual controls as well as same-origin HTTP interfaces. By injecting messages into the next-step boundary or directly cancelling a session, developers can finely control the model state. Installation requires placing files in node_modules and registering them in package.json; the interface is restricted to local access by default.
Read Moredsh-inline-images: Inline Image/Video Support in DeepSeek Harness Conversations
The `dsh-inline-images` plugin addresses the inability to preview local image paths in DeepSeek Harness (DSH). It converts local paths and Markdown syntax into renderable URLs, supports inline image and video display, and offers click-to-enlarge, size adjustment, and automatic refresh of expired URLs. Through dual-channel rewriting in both streaming and persistence pipelines, it ensures consistency between real-time display and saved messages. Installation requires attention to dependency injection and reverse proxy configuration.
Read Moredsh-cortex: A Low-Cost Multi-Model Orchestration and Quality Gating Plugin
dsh-cortex is a plugin for DeepSeek Harness, designed to address the trade-off between the high processing costs of large models and the difficulty of guaranteeing the quality of smaller models. It adopts a “main model as Supervisor, small models as Executors” architecture and provides capabilities such as recursive task decomposition, dynamic routing, four-level quality gating, budget control, and organizational pattern governance. This plugin depends on Node.js 20 and related libraries, supports DAG structures and soft-defect correction, and achieves cost reduction and quality improvement through intelligent scheduling.
Read MoreGoogle Unveils Gemini 3.6 Flash: 17% Token Efficiency Improvement, Battle of Cost-Performance for Default Agent Models
On July 21, 2026, Google released Gemini 3.6 Flash. Its output token count decreased by 17% compared to 3.5 Flash (up to 65% on the DeepSWE benchmark), with an output pricing of $7.5 per million tokens, and the DeepSWE score rose from 37% to 49%. Also released in the same batch were 3.5 Flash-Lite and 3.5 Flash Cyber. On July 28, Gemini API Managed Agents was switched to 3.6 Flash by default, with new features including Environment Hooks, budget control and a free tier. This article sorts out the model capabilities, benchmark data and developer onboarding essentials based on the official blog.
Read MoreDeploying Baidu Wenxin 4.5 Open-Source Model for Android Device Calls
In the previous article "Usage and Deployment of the ERNIE 4.5 Open-Source Large Model", we introduced how to use FastDeploy to deploy the ERNIE 4.5 open-source large model and briefly called its interface. This article will describe how Android can call this deployed interface and implement conversations.
Read MoreUsage and Deployment of ERNIE 4.5 Open-Source Large Model
The ERNIE 4.5 series open-source models consist of a total of 10 models, covering Mixture-of-Experts (MoE) models with activation parameter scales of 47B and 3B (with the largest model having a total parameter count of 424B), as well as dense parameter models with 0.3B parameters. Below, we will introduce how to quickly use ERNIE 4.5 models for inference and deploy the interface for client-side calls on platforms such as Android and WeChat Mini Programs. Note that only text-type models are accepted here; in reality, ERNIE 4.5 also has multimodal models.
Read More