游戏开发专家：在 AI 对话里落地 Unity / Unreal 系统与性能优化
昆仑增长在 SkillHub 发布的「游戏开发专家」（@kunlungrowth/game-developer，v1.0.0）面向 Unity 与 Unreal 的游戏系统实现与性能优化，基于 SKILL.md，无需 API Key。文章说明其定位、可核实能力边界，以及通过 skillhub CLI 或官方提示词安装到 Cursor、Claude Code、Codex 等 skills 目录的方式，并提醒以安装后的 SKILL.md 与引擎实测为准。
Read More高级开发工程师：面向 Web 全栈与性能目标的 Agent Skill
SkillHub 社区 Skill「高级开发工程师」（@user_ad872d55/senior-developer，v1.1.0）面向 Premium Web 全栈：Laravel/Livewire/FluxUI、Three.js 与高级 CSS，强调技术决策框架、调试实战，以及 60fps 与首屏约 1.5s 目标。无需 API Key。可通过 SkillHub 安装文档提示词或 skillhub CLI 安装到 Cursor 等 skills 目录。适合架构设计、性能优化、重构与选型类对话。
Read MoreDeepSeek Harness Plugin: dsh-history-fold
The DeepSeek Harness plugin dsh-history-fold addresses the increased token consumption and reduced speed caused by growing conversation turns. It works by folding the oldest history into a single message and writing it to a local lossless archive file. The plugin operates in a lazy execution mode, does not log by default, and DSH upgrades do not break it. Its core advantages are lossless archiving, recoverability, and hierarchical path indexing, enabling complete restoration of historical content while maintaining decoding speed.
Read Moredsh-call-shrink: DeepSeek Harness Call Compression Plugin
`dsh-call-shrink` is DeepSeek Harness's ultimate call compression plugin, designed to merge N LLM calls into a single call. It significantly reduces costs through caching, deduplication, and batch consolidation within a 10-second window. The plugin features stable fingerprinting, protocolized merging, a failure circuit breaker, and a real-time monitoring dashboard. It is easy to install, supports RPC operations, and automatically falls back when parsing fails without affecting the final output.
Read Moredsh-chat-offscreen: Keeping Long Conversations Responsive
The DeepSeek Harness (DSH) Web chat session does not virtualize chat history, so the excessive number of DOM nodes causes UI lag during long conversations. The `dsh-chat-offscreen` plugin injects a stylesheet and uses the CSS `content-visibility` property to skip layout and painting for offscreen areas, providing a low-overhead approximation to virtualization, while using `contain-intrinsic-size` to keep the scroll range stable. This solution is compatible with Chrome/Edge 85+, Firefox 125+, and Safari 18+, and can be used together with the retired P0 rule as a temporary workaround until official DSH virtualization support is available.
Read Moremomojie-s/dsh-archive-retention: Periodic Cleanup of Archived Sessions
The DSH plugin `dsh-archive-retention` automatically cleans up expired archived data—including physical archive stacks, page archive sessions, and orphaned child agents—using Cron scheduling to improve system performance. The plugin provides a web UI for visual configuration, supporting the setting of retention periods, cleanup schedules, and cleanup scope. It strictly adheres to safety boundaries and only removes archived and expired data, ensuring a lightweight system environment.
Read Moredsh-title-index: DeepSeek Harness Web Session Title Index Acceleration Plugin
`dsh-title-index` is a plugin developed for the DeepSeek Harness Web client, designed to address the performance bottleneck caused by the official engine having to decompress large amounts of logs when querying conversation titles. By maintaining a disk-based index cache, it reduces the Top-50 menu title read time from approximately 9.7 seconds to 0.3 seconds. The plugin is implemented by extending and overriding the official engine's `readTitleSnapshots` method, with support for automatic invalidation and re-collapsing, as well as a fallback strategy. Although this feature has now been integrated into the official version and the plugin has been deprecated, it remains in the community directory as a reference implementation.
Read Moresatan9394/dsh-performance: Performance Optimization and Budget Management
`satan9394/dsh-performance` is a performance optimization plugin for DeepSeek Harness (DSH). It is designed to address the challenge of blind optimization in agent development by providing a complete closed-loop toolchain for performance tuning through mechanisms such as measurement-first, bottleneck identification, lazy cache loading, and performance budgeting. Developers must obtain the source code from GitHub and deploy it independently. When using the plugin, they should pay attention to permission security and the unpublished license agreement. It effectively helps developers quantify metrics, manage resources, and prevent performance degradation.
Read Moredsh-algo-compare: Algorithm Execution Time Comparison
The DSH plugin `dsh-algo-compare` focuses on comparing algorithm runtime in model inference. Implemented in pure Node.js with no network dependencies, it supports precise parameter-based calculations. Developers can install it using `dsh plugin add github:uckkk/dsh-algo-compare` to quantify performance, making it suitable for developers who need to optimize inference speed.
Read Moredsh-fork-perf: DSH Conversation Fork Performance Optimization Plugin
The `dsh-fork-perf` plugin for DeepSeek Harness (DSH) is designed to address performance bottlenecks when forking large sessions, reducing the time taken by approximately 1/18. It reuses references to frozen JSON trees through a "zero-copy fork" and optimizes the initialization process of `PersistenceCoordinator.initFor`. The plugin provides an HTTP API for monitoring and configuration, along with robust fallback mechanisms and patch restoration capabilities. Note that it depends on DSH version 0.1.0-rc.6, and is limited by the fact that copies inside the `enqueue` closure cannot be eliminated.
Read Moredsh-large-proj-perf: Eliminating blocking caused by fork, history loading, and disk spilling in DSH super large sessions
**Preface** The design philosophy of DSH (DeepSeek Harness) is "everything is a plugin." Several performance-oriented plugins already exist in the community: @linxin666/dsh-perf handles observability and frontend rendering load reduction, while dsh-pref-kit performs streaming incremental merging at the source. However, there is a class of issues regarding large existing sessions that these plugins cannot address: once the number of events in a single session accumulates to several hundred thousand, forking sub-sessions, reopening historical sessions, and persisting fork results to disk all become synchronous heavy tasks on the event loop—causing anything from a few hundred milliseconds of lag to a single core running at full capacity for 20 minutes, or even a direct RangeError. This article introduces dsh-large-proj-perf (v1.2.0, MIT license, author orangeofc).
Read Moredsh-session-repacker: Repacks DSH session logs into the official two-frame format, opening old sessions about 6 times faster.
# Preface If you have used DSH for long sessions, you have likely encountered this issue: opening an old session takes several seconds to respond. The problem lies in the storage layer. DSH's JSONL session backend compresses each batch write into a standalone zstd frame and appends it to the file's tail. A long session can easily contain tens of thousands of frames; reading requires decompressing frame by frame. Furthermore, the backend does not support seek-based suffix reading, meaning the entire file must be decoded whether you need just the tail or the whole thing. The number of frames is determined by historical writes and cannot be shrunk afterward. `kkishapppy/dsh-session-repacker` addresses this issue by repacking scattered thousands of log frames into the official native format. Measured results show the maximum time to open a session reduced from approximately 661ms to approximately 126ms.
Read Moredsh-pref-kit: Merging Adjacent Text Increments at the Source of DSH Streaming Output
`dsh-pref-kit` is a DSH performance optimization plugin designed to address high resource consumption and stuttering caused by the massive number of small events generated by streaming output under long conversations and multi-agent concurrency. The plugin merges adjacent `text`/`reasoning` increments within the same text block at the data stream source before writing them to logs and pushing, thereby significantly reducing the volume of session events while remaining fully compatible with existing interfaces and protocols without altering the underlying structure. Core features include streaming merging enabled by default (controlled by the `windowMs` parameter, default 30ms), as well as a configurable "full passthrough" mode (when `windowMs = 0`, it passes through token-by-token to ensure minimal latency). Additionally, the plugin provides a settings card for users to adjust configurations in real time, and includes several experimental features disabled by default, such as replacing the trace view.
Read MoreInstall a RN Performance Optimization Manual for AI Assistant with react-native-best-practices
Callstack split the RN performance e-book into Agent Skill react-native-best-practices, which covers FPS, TTI, package volume, memory and native modules. This article is verified against the official SKILL.md and the repository README: 29 references cover how to triage, how to install the skills CLI, as well as list virtualization, React DevTools measurement and the usage of Android R8. The core is to measure before making changes.
Read More