dsh-bite（咬钩）：压缩后自动找回钉住的关键回合
dsh-bite 是 DeepSeek Harness 客户端插件，将用户选定的提问与助手回复整轮钉住；上下文压缩冲掉原文后，在下一次模型请求前自动打捞并重注入，并标记为历史背景。支持 Web UI 钉子按钮、肥鱼掠食动画，以及 /pin、/pins、/unpin 斜杠命令。默认完整保留原文，每会话最多 12 条，存储于 $DSH_HOME/storages/dsh-bite/。安装可用 codeload 直装 v0.1.1，无需必填配置，可选 maxPins、maxPinChars、storageDir。
Read Moredsh-compaction-fidelity：为 DSH 长会话补上上下文压缩保真层
dsh-compaction-fidelity 是 TLNing260310 维护的 DSH 记忆类插件（v0.3.1，MIT），面向 0.2.0-rc.2 运行时提供双语 compaction 后端：默认 350K 压缩线、跨语言保真指纹与补偿、.dsh/compaction-fidelity 项目索引与架构锚点回查。文章基于 SkillHub 目录与同步 README 说明安装命令、核心命令与适用边界；插件以 dsh 进程权限运行，安装前需审阅源码与 bundle patch。
Read Moredsh-plugin-local-prompt-bridge：本地 token 精算与上下文溢出压缩重试
社区 DSH 插件 local-prompt-bridge 在不修改 @deepseek-ai 包的前提下，将 context 溢出 400 改写为 CONTEXT_WINDOW_EXCEEDED 以触发官方 compaction-basic；配置 routes 后对本机 OpenAI-compat 服务做 /tokenize 步进前精算，超阈值则主动压缩再试。本文基于 README 说明 web profile 的 pnpm 安装、bundles 注册及 cordis.patch.yml 配置，并给出自测与适用注意。
Read Moredsh-context-control：DSH 上下文用量环与可恢复压缩
dsh-context-control 是面向 DSH 0.2.0-rc.2 的社区记忆类插件，在模型选择器旁显示上下文用量环，支持 1 万至 500 万字符可配置自动压缩阈值、手动压缩与安全边界排队，并保留可分页回查的原始会话与结构化摘要。安装需使用 Releases 中的 .tgz 与官方 dsh plugin 命令；与 billion-context 等插件并存时需独立 profile 验证。摘要有损且产生 API 费用，插件以 DSH 进程权限运行。
Read Moredsh-lossless-context：可溯源的 DSH 上下文压缩与检索
Kerberos255 的 dsh-lossless-context 为 DeepSeek Harness 提供分层上下文压缩与检索后端：沿用原生 compaction 事务，不删除源会话日志；将受界事件组整理为可追溯摘要 DAG 与本地 SQLite 索引，并通过会话工具 lcm_recall 支持 search、describe、expand。安装后于 Settings → Plugins → Lossless Context 启用，并在 preset 选用 dsh-lossless-context/agent。含 idle/压力压缩、模型自适应预算、可选 Dream & Memory 后 DM 轮换与 MIT 许可。
Read Moredsh-context-manager：可审查的对话上下文压缩
dsh-context-manager（SkanderBog，v0.2.0）为 DeepSeek Harness 0.2.0-rc.2+ 提供 Context 面板与 /context-manager 命令，用于查看 local context 与用量构成，对历史分组选择 Keep、Summarize 或 Omit，支持决策笔记与 preview 后再 Apply。本文基于官方 README 介绍安装（Release tgz 与 dsh plugin add）、ACP 状态读取、压缩历史与权限注意事项，便于长会话开发者人工把关上下文。
Read Moredsh-context-pilot：长会话上下文用量可见，压缩与思考档位交给模型
dsh-context-pilot 是 Shuffle-1992 维护的 DSH 宿主插件（面板名「上下文智能压缩」），面向长 Agent 会话的窗口压力：每轮注入上下文用量，模型通过工具在智能压缩线自主决定是否压缩并自动续跑，强制压缩线在 pre-step 无条件执行；压缩在下一步前落地，本轮不中断、不伪造用户消息。可选智能思考向 Agent 暴露思考强度档位并允许工具换档。安装命令为 dsh plugin --profile web add github:Shuffle-1992/dsh-context-pilot，需重启 DSH，无 npm 包。当前 v0.1.61，MIT 许可。
Read Moredsh-handoff-compaction：DSH 上下文压缩与任务移交插件
`dsh-handoff-compaction` 是 DSH 插件，用于解决小上下文模型因历史过长导致的溢出问题。它将旧对话转化为结构化的 `# Context Handoff`，保留近期消息，并利用 SQLite 历史工具检索完整记录。该插件支持 Web 界面配置、缓存友好压缩，并能通过 `/compact` 命令手动触发。安装后会全局替换相关逻辑，确保数据可检索且支持 DSH 0.1.x 和 0.2.x 版本。
Read Moredsh-super-memory：跨压缩记忆插件
dsh-super-memory 是一个 DeepSeek Harness 跨压缩记忆插件。它解决超长会话压缩后模型遗忘旧结论的问题。核心功能包括压缩时自动入库、通过 BM25 检索相关内容并注入上下文，以及支持本地检索与可选大模型辅助。该插件仅限单会话使用，不依赖外部服务，支持 MIT 协议开源。
Read MoreHeadroom Token压缩助手：为 Agent 接入 HeadRoom 上下文压缩
Headroom Token压缩助手是 SkillHub 社区 Skill（@user_922b1001/headroom-token-compress-wrap，1.0.1，ai-agent，无需 Skill 级 API Key），封装 headroom-wrap，引导 Agent 使用 headroom-ai 压缩对话、工具输出与 RAG 上下文。文章说明 Skill 定位、SDK/Proxy/MCP 能力、skillhub install 与 Cursor 等 skills 目录、pip/npm 安装底层依赖，并给出 compress 与 proxy 示例及本地运行与实测注意点。
Read MoreDeepSeek Harness 插件：dsh-headroom
`dsh-headroom` 是一个专为 DeepSeek Harness（DSH）设计的本地代理插件，旨在解决 2026 年智能体开发中上下文窗口过大的 token 消耗问题。该插件通过管理本地 Python 服务，自动压缩历史对话和大工具输出，并支持在压缩后按需取回原文。 安装需通过 npm 安装包并在配置文件中注册，首次运行会自动下载模型文件。核心功能包括历史压缩、工具输出压缩、取回工具及配置管理。插件支持通过对话命令或 YAML 文件进行灵活配置，提供无损及有损压缩模式，能显著降低 token 消耗，同时保证不会比不装更差。
Read Moredsh-plan-compact-execute：在 plan 模式下压缩上下文并执行计划
`dsh-plan-compact-execute` 插件旨在解决 DSH plan 模式下的上下文窗口占用问题。该插件在审批卡片增加“压缩后执行”选项，允许在批准前压缩历史上下文为摘要，同时逐字保留计划正文和未完成的工具批次。安装需满足 Node.js 24 及 DSH 版本要求，点击按钮即可在压缩历史后，带着保留的计划内容进入执行阶段。
Read Moretoken省流助手：在 Agent 对话前压缩上下文，节省 Token
SkillHub 社区 Skill「token省流助手」（@user_29dc410e/token-saver-qclaw，v1.0.1，ai-agent）面向长对话与大段代码粘贴场景，自动压缩用户输入、代码片段与对话内容，官方称可节省约 20%–70% Token，无需 API Key。文章说明其定位、已核实功能，以及通过 skillhub CLI 或 AI 提示词安装到 Cursor、Claude Code 等 skills 目录的步骤与使用注意。
Read Morecontext-compression: Context Compression and Token Optimization for Long-Session Agents
context-compression, from Agent-Skills-for-Context-Engineering, focuses on context compression and handoff summaries for ultra-long conversations. The article sets the optimization target as tokens-per-task, introduces three paths—anchored iterative summarization, opaque compression, and full-summary regeneration—emphasizes artifact trails and mandatory sectioned summaries, and provides trigger strategies, probe-based evaluation methods, and common enablement pathways such as Claude Code and npx skills.
Read Moredsh-quilt-compact: DeepSeek Harness Context Compression Plugin
The DeepSeek Harness plugin `dsh-quilt-compact` is designed to compress conversations using “chunk-wise summarization followed by merging” with a specified model, in order to reduce Token costs or handle short context windows. It takes over the default `ctx.compaction` backend, while its workflow includes preprocessing, chunking, summarization, merging, and writing checkpoints. During installation, `compaction-basic` must be disabled. The configuration supports model pools, merge windows, chunk ratios, and preprocessing options. The plugin supports DSH versions between 0.1.7-alpha.2 and 0.3.0-rc.1 and provides an optional execution logging feature. If the model pool fails, a fallback mechanism is enabled.
Read Moredsh-context-compressor: DeepSeek Harness Context Compression Control
`dsh-context-compressor` is an external plugin for DeepSeek Harness that provides flexible control over the context window. It supports setting an absolute token limit or triggering automatic compression based on a context window ratio, and includes a dedicated settings page plus tool result pruning. Installation requires meeting the environment requirements; it can be installed via the official command or a local link (local installation requires `pnpm install`), and a refresh is needed after installation to take effect. During use, the status can be checked via the API. Note that token counts are estimates only, ratio mode relies on the model's declared values, and modifying the source code requires restarting the application.
Read Moredsh-plugin-jev-compaction: Context Compaction Based on Relevance Scoring
By default, DeepSeek Harness manages context in chronological order, which often causes critical constraints or error messages to be mistakenly removed. The `dsh-plugin-jev-compaction` plugin addresses this issue by introducing a Jev relevance scoring mechanism that prioritizes high-value messages and supports protecting specific information via regex patterns. It provides configurable byte budgets, graceful degradation, credential logging, and CLI/programmatic interfaces, ensuring that important data is accurately preserved during context overflow. This makes it well suited for long conversations or complex reasoning tasks.
Read Morehighlightpin: DeepSeek Harness Level-3 Source Text Pinning Plugin
The DSH plugin highlightpin allows users to pin key text in long conversations to prevent information loss caused by context compression. It provides a three-level pinning mechanism (pin, resident, and highlight), supports code and mathematical formulas, and can be managed via the CLI or the interface. After installation, DSH must be restarted, and a character limit is configured to ensure that important content is prioritized and retained in the conversation.
Read MoreDeepSeek Harness Plugin: dsh-hypercompact
`dsh-hypercompact` is an open-source plugin for DeepSeek Harness (DSH) designed to address context-limit or transmission-failure issues in long sessions caused by retransmitting the full history with every request. It provides deterministic, zero-LLM-cost, byte-level context compression by locally compressing older history in milliseconds and supports exact verbatim restoration, ensuring a reasonable request body size without information loss. This article details its core features, installation steps (adding the package, creating a preset, and restarting to enable it), and compatibility requirements (DSH `>=0.1.5-rc.2` and `<0.3.0`, Node.js `^22.19.0` or `>=24.0.0`).
Read Moredsh-compact-manager: Hierarchical Compaction Threshold Manager for DeepSeek Harness
dsh-compact-manager is a DeepSeek Harness plugin designed to address premature triggering caused by an overly low default compression threshold. It provides a three-level override strategy spanning global, context-length tier, and model scopes, allowing users to customize the trigger formula (such as a ratio, output reserve, or fixed value). Through the sidebar, the plugin enables real-time viewing and editing of configurations and supports applying presets, primarily for finely controlling the timing of context compression and preventing premature loss of historical information.
Read MoreJev-LCM Compaction Plugin for DeepSeek Harness
Jev-LCM Compaction is an experimental plugin for DeepSeek Harness (DSH), designed to address state budget saturation and text floor growth caused by standalone Jev compression. The plugin introduces a Jev sorting layer before native compression, uses local SQLite archives to losslessly preserve raw evidence, and optimizes retention rates through rolling retention-rate percentile calibration. It supports multiple providers, overflow markers, and rich session debugging tools. Although it offers functional completeness and fallback mechanisms, it is developer preview software only and does not guarantee production-grade parity.
Read Moredsh-clear-mind: Model-Autonomous Context Compression and Checkpoint Management
`dsh-clear-mind` is a memory management plugin for DeepSeek Harness (DSH) designed to address context-window saturation in long conversations. By providing the `mind_map` and `clear_mind` tools, it enables the root agent to autonomously compress conversation history, replacing redundant information with auditable checkpoints. The plugin also supports proactive threshold alerts, generates GUI tombstone records, and runs entirely locally without requiring network access. It can effectively reduce inference costs while preserving session traceability.
Read Moredsh-jev-prune: A DeepSeek Harness Context Compression Plugin Based on Jev Structured Output
The `dsh-jev-prune` plugin provides a two-layer context compression mechanism for DeepSeek Harness (DSH) based on Jev structured output. It replaces the existing size-based truncation and model summarization logic via an interception point, using deterministic code to generate receipts instead of model inference and ensuring that the decision process follows the principle that “the model will not generate what it should not generate.” It includes two functional layers—result pruning and receipt compression—and uses strict gating mechanisms (such as a tool allowlist, freshness checks, and balance checks) to determine compression behavior, thereby freeing context while retaining key results, avoiding hallucinations, and restoring the original text.
Read Moredsh-ctxopt-compaction: DeepSeek Harness Context Compression Plugin
`dsh-ctxopt-compaction` is a context compaction plugin for DeepSeek Harness, designed to resolve long conversation overflow issues. It offers benefit evaluation, summary caching, explicit memory control, and diagnostic information retention, and requires host-side dependency support. Its configuration enables fine-grained control over the compaction window, cache size, and tool output trimming strategy, while allowing users to manage memories and to-do items via commands.
Read Moredsh-plugin-om: DSH Plugin
The DSH plugin `dsh-plugin-om` is designed to solve context-window overflow issues in long conversations. Maintained by FanetheDivine, it supports automatic compression of the main session history, summary generation, and preservation of the source text for tools and skills. It provides recall by message index (`recall`) and local semantic retrieval (`recall-semantic`), and logs details of the compression process. The plugin requires DSH 0.2.0-rc.2 or higher. If installation fails with pnpm 11, build permissions must be configured. In the configuration, `observeThresholdTokens` controls the compression frequency, while `tailMessageCount` determines the number of messages to retain. The plugin supports hot updates and allows compression results and errors to be viewed in the browser, making it suitable for scenarios that require efficient management of long conversations.
Read Moredsh-auto-compact: Enhanced Auto-Compact Defaults for DeepSeek Harness Agent Presets
`dsh-auto-compact` is a “workflow” category plugin for DeepSeek Harness (DSH), designed to optimize the default context-window auto-compaction configuration when agent presets are running. Maintained by developer `songoao25`, it improves context resource management efficiency by adjusting compaction strategies. The official documentation does not provide a ready-made installation command, so installation guidance must be obtained from the repository. When using the plugin, run it with the current process permissions and review the source code and license to ensure compliance.
Read Moredsh-idle-compactor: An Idle-Triggered Context Compression Plugin
The `dsh-idle-compactor` plugin adds a time- and state-based context compaction mechanism to DeepSeek Harness. It triggers compaction when a session has remained idle for an extended period and exceeds the token threshold, avoiding costs for stale context tails. The plugin supports configurable idle time, context size, and scan interval, and handles background tasks and archived sessions, effectively addressing context bloat in long conversations.
Read Moredsh-compactor: A Context Compression Plugin for DeepSeek Harness
The DeepSeek Harness plugin `dsh-compactor` is designed to address context bloat and infinite-loop issues in long conversations. It employs a three-tier processing mechanism (real-time pruning, batch summarization, and archive restoration) and offers two modes: API-based summarization and local rule-based compression. The plugin supports semantic understanding, data rollback, and infinite-loop detection, and requires user confirmation before compression is executed to balance context length with information retention.
Read Moredsh-compaction-micro: The Micro-Cleanup Layer for DeepSeek Harness
`dsh-compaction-micro` is a standalone plugin for DeepSeek Harness (DSH), designed to quickly free up token space before the official summarization layer by replacing rerunnable tool results on the whitelist with placeholders, thereby avoiding unnecessary and costly model calls. It follows the observer pattern and triggers cleanup only when context pressure exceeds a threshold or an overflow occurs. It supports configuring the cleanup tool whitelist, the number of recent results to retain, and the trigger ratio. After installation, it must be mounted before `dsh-compaction-basic` and applies only to safe, read-only tools.
Read Moredsh-squeeze-command: On-Demand Context Compression for DeepSeek Harness
`dsh-squeeze-command` is a plugin for DeepSeek Harness (DSH) designed to address the soaring input token costs caused by excessively long conversations. It uses a manually triggered mechanism, allowing the conversational model to select the compression scope and employing a cheap, high-speed model to generate summaries, thereby effectively controlling the token budget while preserving the original logs. The tool features no discarded content, resistance to post-compression perplexity, and refusal to run during agent turns. During installation, the token budget and a flash-tier summarization model must be configured. Through commands such as `/squeeze`, users can flexibly adjust the context size or check the compression status, making it a practical tool for long-running, cost-sensitive scenarios.
Read Moredsh-context-compression-improved: Context Compression and Code Skeleton Compression Plugin for DeepSeek Harness
The `dsh-context-compression-improved` plugin in DeepSeek Harness addresses context bloat in agent tasks through mechanisms such as pre-compression, aggregate replacement, and history compaction. Its core new feature, the Code Skeleton Gate, substantially reduces oversized source-code tool results by retaining only declarations and key lines, significantly improving inference efficiency. The plugin supports multiple strategies, including Fresh and Aggregate, is compatible only with DeepSeek models, and requires manual enabling of code skeleton compression in the settings.
Read MoreDeepSeek Harness Plugin: Context Compression Fix for qwen3.8-27b
The DeepSeek Harness (DSH) plugin fixes an issue where the qwen3.8-27b model running on the NInfer engine exhausts its output tokens due to the default high reasoning level, causing context compression and title generation to be truncated. By disabling the thinking mode, applying recommended sampling parameters, increasing the output token limit, and constraining the reasoning intensity, the plugin ensures that the compression feature operates properly. After installation, restart the `dsh web` process and exactly match the model ID in the configuration file for the plugin to take effect.
Read Moredsh-brain-compaction — Context Compaction Coordination Plugin
`dsh-brain-compaction` is a composite plugin for DeepSeek Harness (DSH), designed to address the large context usage and high token consumption in long-running tasks. By integrating multiple community plugins such as `dsh-argp`, `dsh-memory-vault`, and `dsh-mcp-lens`, it provides unified configuration, intelligent routing (We-Need), and cross-session memory management. Users need to install it in the DSH 0.1.5 environment using the `file:` or `github:` protocol, and be aware that `dsh-compaction-instant` is currently unavailable and some local dependency requirements apply. This plugin offers developers a fine-grained token-usage control solution for handling complex long-running tasks.
Read MoreDeepSeek Harness Context Compression Plugin: dsh-plugin-context-compressor
`dsh-plugin-context-compressor` is a context management plugin for DeepSeek Harness, designed to address high token consumption and truncation issues in long conversations. It intelligently filters and retains key decisions, errors, and preferences, removes redundant or failed records, and summarizes the remaining content in a structured format. Users can trigger compression with specific commands, reducing costs while improving the agent's reasoning efficiency.
Read Moredsh-compact: DeepSeek Harness Context Compression and Overflow Recovery Plugin
`dsh-compact` is a context management plugin for DeepSeek Harness, designed to address context overflow caused by long conversations or frequent tool calls. It automatically compresses historical messages, provides manual compression controls, and includes an overflow recovery mechanism: when context usage reaches 75%, it automatically generates summaries to replace older content, and it attempts retries when an overflow error occurs. The plugin supports basic mounting and Agent-level mounting, offering flexible configuration strategies to effectively maintain coherence in long conversations.
Read MoreDeepSeek Harness Compaction UI: Visual Configuration Plugin for Context Compaction
`deepseek-harness-compaction-ui` is a visual configuration plugin for DeepSeek Harness, designed to simplify context window management. It replaces complex YAML editing with a Web interface, supports configuring compaction trigger thresholds and recent-content retention based on token count or percentage, and features automatic conversion, hot reloading, and independent policy management. The plugin does not handle core compaction logic; it is responsible only for displaying, converting, and validating configurations. Users should note that it does not resolve issues where a single message exceeds limits or system prompts consume too much context, and that critical state should be stored externally.
Read Moredsh-plugin-anti-rot: DeepSeek Harness plugin for compressing tool error payloads
The DeepSeek Harness (DSH) plugin mitigates “context rot” in long sessions by compressing verbose tool error stack traces. In the `tools/post-execute` phase, it identifies failure signals (`isError` or a non-zero exit code), truncates error text to a specified number of lines and removes indented stack lines, while preserving a truncation marker. After installation, it must be enabled in the configuration file with a positive integer `maxLines` parameter. The tool effectively reduces noise and saves context budget, leaving successful calls unchanged.
Read Moredsh-compaction-pro: DeepSeek Harness High-Fidelity Context Compression Backend
The default compaction mechanism in DeepSeek Harness (DSH), `dsh-compaction-basic`, suffers from issues such as losing exact values and forcing English output. `dsh-compaction-pro` is an upgraded plugin that rewrites the summarization hook to achieve high-fidelity, bilingual, and recursive compaction. It enforces the preservation of numeric values, paths, and commands, supports Chinese summaries, and requires no changes to the original policy. Installation requires disabling the built-in backend and registering the plugin in the configuration file. This plugin is suitable for long sessions and can effectively address information loss and language fragmentation issues.
Read Moredsh-premise-guard-cn: A Context-Premise Loss Prevention Plugin for Chinese Novel Writing
`dsh-premise-guard-cn` is a Chinese-optimized version of the `ICCuse/dsh-premise-guard` plugin, designed to address the loss of critical premises caused by long-context compression in DeepSeek Harness. It includes built-in automatic anchor-extraction rules for Chinese (such as Chinese quotation marks, terminology chains, and paths), supports manually maintaining an anchor list, and provides the `premise_anchor` tool and `/anchor` command for management. After context compression, the plugin automatically detects and injects warnings, prompting the model to verify any lost settings and effectively preventing logical collapse in long-form Chinese creative writing.
Read Moredsh-arc-context: DeepSeek Harness Reversible Context Compression Plugin
# Introduction For developers of agents running long sessions, a critical issue arises when the context approaches the window limit: compression is mandatory. The Basic compression built into dsh relies on an irreversible model summarization approach, which is lossy and irreversible. In scenarios such as code engineering that depend on precise facts, losing a single constraint can cause subsequent changes to diverge. dsh's philosophy is that "everything is a plugin," meaning compression strategies are replaceable. dsh-arc-context takes a different approach: before compression, it first stores the original text completely in a local append-only log, allowing compressed blocks within the context to be restored character-by-character at any time. Below is an introduction to its implementation, installation, and real-world test data. # What is it dsh-arc-context (ARC = Adaptive Revers...
Read Moredsh-asc: A DeepSeek Harness Plugin That Lets the Model Decide Context Compression
# Introduction Long conversations are an inevitable challenge in agent development: with multi-round tool calls and a large accumulation of intermediate results in the context, the window quickly fills up after a few rounds. The common approach is to set a threshold; upon triggering, the system performs automatic summarization—the model itself has no say in what is discarded or retained, which is entirely determined by external rules. dsh-asc follows a different approach: it allows the model to decide when and what to compress. Every compression is executed as a replacement of events with persistent session logs, making the process replayable, searchable, and reversible. DSH's philosophy is "everything is a plugin," and compression is no exception. Below, we introduce the positioning, operation, installation steps, and considerations for this plugin. # What is it The full name of dsh-asc is DeepSeek Harness Agentic Surface...
Read Moredsh-auto-compact: DSH Session-level Auto Context Compression Plugin
`dsh-auto-compact` is a session-level automatic context compression plugin for DeepSeek Harness (DSH), maintained by JohnathonYe and released under the MIT license. It aims to address the excessive context pressure encountered in long sessions by monitoring session context usage. When usage exceeds a user-defined independent threshold, it automatically performs `compact` compression and injects a summary into the context to seamlessly continue the session, eliminating the need to send a "continue" message throughout the process. The plugin supports threshold adjustment ranging from 1% to 90% (default: 50%) and persists settings based on `sessionId`. It coexists with the built-in DSH `compaction-basic` (default threshold: 80%) without interference. Installation requires Node.js ≥ 20 and pnpm ≥ 10.
Read Moredsh-compaction-instant: An Instant Context Compression Plugin for DeepSeek Harness
dsh-compaction-instant is an instant context compression plugin for the DeepSeek Harness environment, designed to replace traditional LLM summarization with deterministic text processing. This plugin avoids model calls or network requests, eliminating additional delays and rewriting risks. By preserving the original tokens and generating seq, result, or checkpoint pointers that link to persistent logs, it achieves near-lossless compression. As a drop-in replacement for dsh-compaction-basic, it maintains the same interface and semantics, offering recall and search tools along with the /recall command to support on-demand restoration of omitted content or searching of persistent logs. This solution is ideal for developers seeking to reduce model call overhead and ensure data recoverability, requiring No
Read MoreBillion-Context-DSH: Empowering Models to Decide When to Compress Context
Billion-context-dsh provides an Active Context Pruning backend for DeepSeek Harness, where the model determines when to compress and restore context through the compress, decompress, search_context, and acp_status functions. Ported from billion-context-pi and reusing acp-kernel, it can be installed via the bundle command: dsh plugin --profile web add billion-context-dsh. Version v0.2.13 (beta) is suitable for long conversation and long task scenarios.
Read MoreUsing dsh-compaction-instant to transform the context compression of DeepSeek Harness into near-lossless compilation
`dsh-compaction-instant` is a DeepSeek Harness session plugin maintained by KitDoesIt, which replaces the built-in summary engine `compaction-basic` with a deterministic compilation method that does not rely on LLMs. This article checks against the catalog page and GitHub README, covering near-lossless checkpoints, seq readback, three installation methods, as well as the compression ratio and precautions on tool-intensive sessions.
Read MoreUse billion-context-dsh to let the model handle the context compression of DeepSeek Harness by itself
### 正式翻译： billion-context-dsh is a DSH session plugin maintained by Tyan66666, which implements Active Context Pruning as a CompactionEngine backend: the model calls `compress` to generate its own summaries, and the automatic strategy only nudges. The original text remains in the session log, and decompression and search are both supported. This article verifies the installation commands, mounting methods and differences from the built-in automatic compression according to the directory page, GitHub and npm v0.2.2.
Read More