dsh-turn-eraser：在 DSH 里按轮删除对话且可追溯
dsh-turn-eraser 是 Nagiko0739 维护的 DeepSeek Harness 插件，通过墓碑事件遮蔽模型可见上下文中的单轮或多轮对话，原始会话日志不改写，可用 trace_deletions.py 读回。支持删除本轮、删除此处及以后、轮尾「第 N 轮」编号及空回轮次。安装包名 dsh-turn-eraser，MIT 0.2.1。需注意无撤销、子 Agent 快照、非防篡改等限制。
Read Moredsh-plugin-clm：把 DSH 上下文窗口镜像到磁盘并可编辑
dsh-plugin-clm 为 DeepSeek Harness 提供 Context Language Models 式上下文管理：将会话镜像到 .dsh/clm/LIVE_CONTEXT.md，支持裁剪压缩与 budgetTokens/reserveTokens 防护，并提供 /clm 斜杠命令与 UI 开关。本文基于官方 README 介绍安装（profile package.json 或 dsh plugin add dsh-plugin-clm）、用法与适用场景，维护者 maxeone，MIT 许可证。
Read Moredsh-context-management：DSH 换窗压缩、会话笔记与分页历史
dsh-context-management 是 overact 维护的 DSH 记忆类插件（v0.3.6），在官方 BasicCompactionEngine 扩展点上实现换窗压缩：自动阈值、new_context、/compact 与 overflow 四类入口统一生成长度受限的交接文档，并保留 goal、todo 与近期用户指令。配套 notes_* 工具在会话内持久化笔记，history_* 工具分页检索完整会话日志。通过 /ctx 与 Web 设置开关，需本地 link 安装、preflight 与 sync-presets 挂载换窗引擎。借鉴 Codex 多窗口思路，非 1:1 移植且不保证语义无损恢复。
Read Moretoken节省大师：面向 OpenClaw 链路的 Token 优化 Skill
token节省大师是 SkillHub 社区 Skill（@user_65cfe060/token-saver-king，v1.0.0），面向 OpenClaw 链路，宣称通过智能压缩、上下文管理、自动摘要与代码优化等多引擎降低 Token 消耗。本文依据 SkillHub 页面与安装文档说明定位、能力与 skillhub CLI 安装步骤（含 Cursor 等 skills 目录），并提醒节省比例需自行验证、摘要可能影响原文精度。无需 API Key（requires_api_key=false）。
Read MoreToken节省王：在 Agent 全流程中压缩 token 消耗的 Skill
Token节省王（@user_65cfe060/burgess-123）是 SkillHub 社区 ai-agent 类 Skill，面向 OpenClaw 问题解决全流程，通过输入压缩、思考精炼、输出瘦身与上下文管理降低 token 消耗，支持 minimal、balanced、detailed 自动切换。官方页面宣称整体约可省 20%–40%，无需 API Key。本文基于 SkillHub 资料说明四引擎、安装命令（skillhub install 与 Cursor/Codex 等 skills 目录）及长对话与代码生成等适用场景与使用注意。
Read MoreDeepSeek Harness 插件：orangeofcarl0-sys/dsh-epistemic-fold
`dsh-epistemic-fold` 是一款 DeepSeek Harness 插件，旨在解决长周期智能体历史记录积累导致的上下文窗口溢出及信息失真问题。它通过热上下文折叠、单调折叠前沿及精确归档机制，将有损摘要降级为辅助信息，确保智能体可随时检索原始历史。该插件提供经济、平衡、质量三种模式，支持运行时监控及上下文检索工具，确保智能体在受限窗口下仍能保持行为可追溯与状态一致。
Read Moreattention-scheduler：注意力调度与记忆检索规则
`attention-scheduler` 是 DeepSeek Harness (DSH) 的一个插件，旨在解决上下文幻觉问题。它通过显式规则管理记忆的检索、召回、保留及丢弃，确保记忆管理的一致性与可审计性。 核心功能包括检索四级触发、注意力预算、成本归属（L0/L1）及冲突写回策略。插件提供标准插件与文件技能两种安装方式，需配合记忆插件使用。它将模糊的记忆管理转化为可配置的流程，适用于优化长会话成本及减少幻觉。
Read More记忆管家：多 Agent 上下文与记忆文件的监控与复盘
记忆管家是 SkillHub 社区 ai-agent 类 Skill（v1.0.3，无需 API Key），面向多 Agent 长期运行：监控会话上下文与记忆文件磁盘占用并在超阈值时提示压缩，配合 crontab 每日增量扫描记忆、提炼偏好与待办并更新 MEMORY.md，安装后引导初始化目录与定时任务。本文基于官方页面与 SkillHub 安装文档说明安装坐标、CLI 命令及 Cursor/Claude Code/Codex 的 skills 路径，并归纳适用场景与限制。
Read More记忆管理：用文件与规则流程稳住 AI 协作上下文
记忆管理（@user_7d16a562/memory-v1，v1.0.1）是 SkillHub 社区 ai-agent 类 Skill，通过文件系统与规则化流程实现持久记忆，核心为任务开始判断、读取记忆、执行、保存，缓解多任务协作中「今天讲明天忘」。本文基于 SkillHub 公开信息介绍定位、能力与安装方式（skillhub CLI 或按 skillhub.md 由 Agent 安装），并说明适用场景与使用注意，不编造未公开的目录与配置细节。
Read More玄奘.skill：在用户施压或反复失败时切换 Agent 工作姿态
玄奘.skill 是 SkillHub ai-agent 分类下的提示词 Skill，在用户明确请求 pua 模式、反复失败两次及以上、抱怨质量或发送 /pua 等触发词时切换 Agent 策略，正常首次编程与查询则不触发。本文基于 SkillHub 公开简介与安装文档，说明触发边界、skillhub CLI 与 Cursor 等 skills 目录安装方式，以及典型对话用法与误触发注意。官方入口：https://api.skillhub.cn/user_741dc82b/xuanzang-skill。
Read More记忆管家 V3.9.0：管理 `.workbuddy/memory/` 的 Token 节流与相关性加载
记忆管家 V3.9.0 是 SkillHub 上的 AI 记忆加速器与 Token 节流阀，面向 .workbuddy/memory/ 目录。核心包括 tiktoken 分词计量、--query/--tags 相关性加载（CJK bigram，可选 embedding）、硬预算闸与 throttle 压实、三色告警与近 7 日均值、--force 与全命令 --json、self-test 安全自检，以及并行摘要与 rank 缓存等性能优化。文章说明安装 skillhub CLI、--dir 指向各 Agent skills 目录的方式，以及 Token 检查、智能加载、清理备份等典型触发场景与使用注意。
Read Moredsh-shake: DeepSeek Harness Manual Context Reduction Plugin
dsh-shake is a third-party plugin for DeepSeek Harness (DSH) designed to reduce excessive token consumption caused by overly long conversation history. It frees up context space by replacing long tool outputs and large code/XML blocks with short placeholders, while enabling on-demand restoration of the original text through the `shake_read` tool. The plugin includes an intelligent protection mechanism that, by default, preserves the most recent 4,000 tokens of history and only cleans up content that meets the eligibility criteria. Users can manually trigger cleanup by sending the `/shake` command, with flexible configuration to effectively balance context management and content preservation.
Read Moredsh-compact-selected-model: Use the currently selected model when compacting conversation summaries
The DeepSeek Harness (DSH) plugin `dsh-compact-selected-model` is designed to address the issue where default compaction uses the model from the previous turn. After switching models in the Web UI, the plugin allows session compaction to be performed directly with the currently selected model. Installation requires a DSH version ranging from 0.1.7-rc.2 to 0.2.x. The plugin works by listening for events to intercept requests and supports both the `/compact` command and automatic compaction, though it has limitations in headless mode and automatic compaction threshold calculation.
Read Moredsh-compact-saviour: DeepSeek Harness Model Compression Rescue and Manual Compression Plugin
dsh-compact-saviour is a plugin for DeepSeek Harness designed to address context overflow in long conversations. It provides a dedicated compression model, chunk-based compression strategies, and a failure rescue mechanism. Users can trigger compression via long-press or manual action, and the plugin automatically switches to rescue mode after two consecutive native compression failures. Supporting both desktop and web installations without source-code dependencies, it aims to improve inference stability in long-conversation scenarios.
Read MoreContext Guardian: A Practical Solution for Local Model Context Management
Context Guardian is an open-source plugin maintained by LuminariSoftwares, designed to resolve session interruptions caused by context window overflow in local models. It offers two modes—an OpenAI-compatible proxy and a DeepSeek Harness native engine—with support for deterministic compression. Its key features include fail-open behavior, disk persistence of raw data, and retrieval of archived context. Users can enable this functionality by configuring the proxy or integrating DSH, ensuring stable sessions and preserving historical details while keeping context length within limits.
Read Moredsh-better-composer: Markdown Demo and Context Enhancement for Web Composer
`dsh-better-composer` is an enhancement plugin for DSH Web Composer, designed to improve Markdown rendering and context management capabilities through public interfaces. Core features include Markdown visual enhancement, local deterministic completion, an interactive context map, and a long-paste clipboard that supports file materialization. Installation requires DSH 0.1.7-alpha.2 and Node.js 20+, and uses pnpm. The plugin leverages DSH public interfaces without taking over source code control, and provides an experimental context doctor feature along with detailed statistical data analysis, making it suitable for developers seeking a refined web interaction experience.
Read Moredsh-switchman: DeepSeek Harness Context Management and Scheduling Plugin
dsh-switchman is a plugin for DeepSeek Harness designed to address context bloat in long sessions. It manages context through watermark control (soft, hard, and forced thresholds) and six scheduling pools (economy, mechanical, main model, among others), transforming the main model from an executor into a scheduler that allocates tasks effectively to reduce costs. The plugin supports agent-team mode, hierarchical validation, and multilingual interfaces, significantly improving the stability and efficiency of autonomous coding workflows.
Read Moredsh-context-truth: Let the Model Know the Actual Context Usage
The DSH plugin `dsh-context-truth` resolves model non-convergence in long sessions caused by misjudged context state, using measured-usage backfill and compaction instruction rule patching. It displays accurate used/remaining token counts and the compaction threshold to the model, and prevents the model from writing its own statements about context exhaustion into summaries. It also provides idle-boundary folding to prevent the model from avoiding critical operations when nearing the threshold. The plugin depends on the host `tokenMeter` and `compaction`, requires no build, and is version 0.2.0.
Read Moredsh-context-window: Replicating Codex's Context Window Management
`dsh-context-window` is a DSH plugin designed to address information loss caused by traditional summarization compression when the context window fills up. It replicates Codex’s context-window swapping mechanism, using context swapping instead of frequent compression and preserving details through notes and handoff checkpoints. Core functionality includes two modes: the `handoff` mode without summarization (recommended, directly writing to a handoff checkpoint) and the `seam-region` mode with traditional compression. The plugin provides `new_context` and `notes` tools for the model to invoke, and generates handoff information containing the objective, tasks, notes, and the last request when swapping context windows. Note that `handoff` mode does not emit a UI-visible compression event, and swapping is only allowed at step boundaries.
Read Morelaodonge/col-dsh-plugin: Context Organization Layer for DeepSeek Harness
laodonge/col-dsh-plugin is a plugin designed for DeepSeek Harness (DSH), intended to address the issues of tight coupling between context management and execution logic and complex permission control. It introduces persistent context slots and provides core features such as fail-closed permission checks, two-stage auditing, permission pruning, and zero runtime dependencies. Developers should be aware of the security risk of running with the current process's permissions and are advised to review the source code before installation.
Read More@rayfalling/dsh-tool-visibility: Controlling Tool Injection in the DeepSeek Harness Model Context
`@rayfalling/dsh-tool-visibility` is a static plugin for DeepSeek Harness (DSH), designed to address context overflow caused by limited model context length and an excessive number of tools. It allows developers to statically control which tool schemas are injected into the context through a browser settings page or a command-line tool. Core features include filtering by name prefix, searching, and batch operations; leveraging DSH's native `tools.restrict` mechanism to filter tools; and persistently saving hidden states. Before model invocation, the plugin removes unauthorized tools by assembling the system prompt, thereby optimizing context space and preventing context overflow. After installation, developers can easily manage tool visibility, making it suitable for tool-only scenarios.
Read Moredsh-codex-context: Lossless Retrieval Context Manager for DeepSeek Harness
`dsh-codex-context` is a plugin for DeepSeek Harness (DSH) designed to address context window exhaustion and information loss caused by the default LLM summarization backend. It employs mechanisms such as token-budget-based sliding windows, zero-orphan boundaries, and model-agnostic checkpoints to ensure information integrity in long sessions. The plugin provides tools such as persistent working notes and cold history search, and supports installation via Git or locally. When installing, be aware of potential tool name conflicts and regular expression query limitations.
Read Morehw-cola/dsh-dynamic-agents: DSH Web Dynamic Context Plugin
DSH Web provides an agent context rule management plugin that adds a docked control above the chat box. The plugin supports two levels of rule libraries, global and workspace, allowing users to select, add, or delete rules and inject them into the system prompt in real time. Once a session starts, the control is locked and must be unlocked to make changes. Installation requires the `dsh` and `pnpm` environments, and it has been tested to work with DSH `0.1.7-rc.1`.
Read Moredsh-context-mode: DeepSeek Harness Context Window Management Plugin
`dsh-context-mode` is a DeepSeek Harness (DSH) plugin designed to solve context overflow issues caused by large files and web content through the “everything is a plugin” philosophy. It primarily features three capabilities: local knowledge base indexing (based on SQLite FTS5, with support for web scraping and chunked retrieval), routing interception (hard-blocking large-file tools and soft-guiding precise reading), and sandboxed execution (running code in an isolated environment and returning only the conclusions). By optimizing data inputs, indexing, and isolated execution, it effectively reduces context window pressure and improves the efficiency of context utilization.
Read MoreDeepSeek Harness Workspace Memory Plugin: Approval-Based Instructions and Project Memory Maintenance
The `dsh-workspace-memory` plugin addresses cross-session context sharing by maintaining two local files: `AGENTS.md` (behavioral rules) and `.dsh-memory.md` (project facts). It uses an interactive approval mechanism, writing changes only after user confirmation to keep the context stable and auditable. The plugin relies on standard Markdown storage, makes no network requests, and supports sharing across concurrent sessions. Installation requires specific DSH and Node.js versions. Its flexible configuration makes it suitable for users who need to manage long-term instructions and facts within a DeepSeek Harness workspace.
Read Moredsh-context-budget: A Plugin for Keeping Local Model Context Within a Controllable GPU Range
The `dsh-context-budget` plugin is designed to address the sharp performance degradation of local models under long contexts. It measures prefill speed and enforces hard limits and checks for token count, first-token latency, and cold prefill time. When a limit is exceeded, it can print a warning or perform compaction, keeping the context within the range the GPU can process. The plugin coexists with `compaction-basic` and provides a `/context-budget` command to monitor status, helping users optimize the experience of using local large language models.
Read Morecontext-budget: Managing the Context Window as a Budget
This plugin aims to address long-task failures that often occur around the 80% mark due to re-reading and re-derivation by managing the context budget. Its core features include a hierarchical reading strategy that first reads file headers and then loads content on demand; immediate writing to persistent “working notes” once facts are established, to retain key information; and “checkpoint recaps” at task-stage transitions to reconstruct context. It is well suited for handling complex long-running tasks and significantly improves the stability of large language models when processing lengthy workflows.
Read MoreAssembled Context: DeepSeek Harness Assembled Context Plugin
The default context handling in DeepSeek Harness (DSH) is limited and can easily allow junk data to occupy the context window. The `catsenior507/dsh-context-assembler` plugin provides an inspectable context tree solution, allowing users to precisely control what the model receives through a browser panel or proxy tools. The plugin supports three node modes—`full`, `key` (summary), and `off`—and is compatible with KV caching. It shifts context management from passive compression to active decision-making, effectively optimizing the token budget and performance.
Read Moredsh-trilogy: Project Memory Plugin for DeepSeek Harness
`dsh-trilogy` is a Host plugin for DeepSeek Harness designed to address the lack of persistent memory in models. It enforces the maintenance of three Markdown files for each workspace: `PROJECT.md` (current state), `DECISIONS.md` (decisions), and `SESSIONS.md` (logs), with automatic creation, loading, and recording. The plugin provides tools such as `memory_checkpoint` for categorized logging, supports BM25 retrieval, and includes a graphical interface for memory management. Its core lies in an intelligent injection strategy: it controls priority via a budget, automatically injects context at session start, and weights categories based on activity. All routing is limited to local access, uninstallation does not delete data, and it is suitable for development scenarios that require long-term maintenance of project context.
Read Moredsh-context-management: Windowed and Reversible Context Management in DeepSeek Harness
`dsh-context-management` is a plugin for DeepSeek Harness, designed to address the context window bottleneck in long conversations or complex tasks. It provides windowed and reversible context management, with core features including automatic takeover of native compression, history retrieval, adaptive fidelity, and graceful degradation. Installation requires DSH 0.1.7 from the prerelease series and Node.js ≥ 22.12; on desktop, install it via the in-app settings, while in the CLI you can run the installation command directly. This plugin is especially suitable for development scenarios that require long-term memory and precise retrieval.
Read Moredsh-context-rollover: A Model-Driven Context Window Rollback Approach
`dsh-context-rollover` is a DeepSeek Harness plugin designed to address the challenges of managing long-conversation context. It takes over the compression engine and provides a model-driven context rollover mechanism, enabling context extension without relying on LLM summarization. The plugin supports locally persisted notes, preserves raw data and the verbatim tail, and offers both manual and automatic rollover capabilities. Users must select a specific preset to enable it and should be mindful of dependency resolution issues on Windows. It is intended for advanced users managing long-term local conversations.
Read MoreDeepSeek Harness Memory Tree Bridge Plugin
`dsh-memory-bridge` is a memory bridge plugin for DeepSeek Harness (DSH), designed to address the limitations of local model context windows (4K–32K). By distilling conversations into a visualized memory tree and providing persistent storage and retrieval, it enables agents to have long-term memory. The plugin’s core capabilities include automatic memory card accumulation, LLM-free retrieval (BM25 + RRF), memory governance (forgetting curves and auditing), and profile distillation. Its architecture follows a Host JS + Python Sidecar pattern, supports plaintext Markdown storage, and ensures data controllability and fault isolation. Through a dual-channel write strategy, it effectively resolves the pain points of agent “forgetfulness” and “context overflow.”
Read Moredsh-dcp: DeepSeek Harness Deterministic Compression Backend
dsh-dcp is a deterministic context pruning and compression plugin for DeepSeek Harness (DSH). It does not use an LLM to summarize content; it only retains factual details from the conversation, such as commands, error messages, and original wording, ensuring stable output with zero additional token consumption. The plugin inherits DSH's security mechanisms, supports management via the `/dcp` command, and optimizes pricing for Chinese-language scenarios. It is easy to install and well suited for low-cost, stable management of Chinese conversational context.
Read Moredsh-token-usage: Add a Token Usage Analysis Page to DSH Web GUI
dsh-token-usage is a Token usage analysis plugin for DSH Web GUI (npm package @samecorner/dsh-client-ui-token-usage, v0.1.3, MIT licensed). It adds a "Token Usage" tab at the top of the session, rendering projection data such as `tokenUsage` and `contextPressure` from the host as visual panels. Features include per-round KPI cards, context pressure progress bars (warning at 80%), context composition, and per-round bar charts.
Read Moredsh-plugin-context-manager: A DSH plugin that automatically records conversations, injects memory, and folds real context.
# Introduction Running long sessions in DeepSeek Harness (DSH) often presents two common challenges: First, the limited context window means that once early conclusions are squeezed out, the model "forgets" them; Second, when switching to a different session, previously recorded key points cannot be carried over. The common solution is to inject summaries into the system prompt, but this only affects auxiliary prompt blocks, while the real dialogue history continues to grow, leading to increasing token consumption. The `dsh-plugin-context-manager` introduced below addresses these two issues: it automatically records every round of dialogue, injects memory and prioritized cross-session notes into each model request, and directly collapses the real dialogue history visible to the model to reduce token usage. # What is dsh-plugin-context-manager
Read Moredsh-agent-context: Injection, Metering, and Pruning of DSH Session Context
dsh-agent-context is a DSH plugin maintained by jonah791, licensed under the MIT license, and is positioned as a session context injector.
Read Moreflowctx-dsh: Local First Context Engine for DeepSeek Harness
flowctx-dsh is a local-first context management plugin for DeepSeek Harness (DSH) designed to address context blurring, information loss, and repetitive search issues caused by long sessions in agent programming. As an additive extension of `dsh-compaction-basic`, it introduces an engineering memory model, with core capabilities including: 1. **Engineer Handover Note Summaries**: Enabled by default, converts summaries into structured handover notes containing failure paths and key identifiers, surpassing general narrative summaries. 2. **Hierarchical DAG Summaries**: Enabled by default, collapses history into hierarchical summary nodes to prevent information generalization in long sessions. 3. **Reversible Tool Result Projection**: Enabled by default, structurally compresses tool results exceeding a threshold and stores the original hash, supporting on-demand retrieval of the original content. 4
Read Moredsh-operating-context: Unified Management of Model Work Windows in DeepSeek Harness
dsh-operating-context is a DeepSeek Harness plugin designed to address the issues of scattered context window sizes and frequent overflow in multi-model service configuration. It resolves this by introducing a "Work Window" settings page that restricts each model service to a manageable range, preventing excessive settings. Maintained by AIMFllyYS, this plugin is licensed under the MIT license and is currently at version v0.1.0. It requires Node.js and pnpm environments and cannot run independently. Its core features include safe configuration writing (writing to different fields based on route shapes), reading model directory limits (without guessing when unknown), idempotent application (removing invalid overrides to restore native capacity), and preserving original route metadata. Once installed, users can adjust the window size and apply settings via the Web UI, with configurations being persisted to "set".
Read Moredsh-plugin-focus: Fixing Targets, Constraints, and Decision-Making in the DSH Agent
dsh-plugin-focus is an MIT-licensed plugin for DSH agents designed to address context compression and cross-session information loss. It uses a persistent focus board to record goals, constraints, and decisions, ensuring that key information is stably retained within the model's context. Core features include a focus tool supporting set, get, append, and clear actions, as well as a mechanism that automatically injects focus content into the model context at the start of each round or when changes occur. When clear is executed, old content is automatically archived to a backup file. The plugin provides configurable prompt guide sections, session projection, and an experimental read-only Web panel. After installation, DSH must be restarted. Configuration options cover file paths, archiving strategies, and injection length limits (maxChars), with all file access restricted.
Read Moredsh-fusion: Merging multiple DeepSeek Harness sessions into one resumable session
dsh-fusion is an MIT-licensed plugin for DeepSeek Harness that merges 2 to 32 scattered DSH sessions into a new, persistent, and continuous target session context. It generates a snapshot at the current point in time rather than a real-time connection, leaving the source sessions unchanged. Core functions include ordering messages by time or source, deduplicating shared prefixes via stable IDs, retaining only user and assistant text content, and inheriting the caller's environment configuration. The plugin offers four ways to invoke it: a UI dialog, a /fuse text command, a conversation_fusion model tool, and a Service API. Users can configure the maximum number of messages, byte limits, and whether to enable the Smart Planning Agent. The Smart Planning Agent is a tool-less temporary A...
Read Moredsh-pin-recall: A plugin to pin and recall context for DeepSeek Harness sessions
dsh-pin-recall is an MIT-licensed DeepSeek Harness (DSH) plugin maintained by kerwin2046, aimed at addressing the pain point of referencing historical content in sessions. This plugin allows users to pin assistant replies, tool results, or free text as persistent events and inject this content into the next round of model context via the `/recall` command or Web UI cards, avoiding manual copy-pasting. Core features include pinning by latest result, specific ID, or text, and providing commands such as `/pin`, `/pins`, `/unpin`, and `/recall`. `/recall` can wake up the agent to process pinned content, while `/recall quiet` only queues it without waking up.
Read More