dsh-coldread-cache：DSH 长会话解析优化插件
DeepSeek Harness (DSH) 插件解决长会话冷读开销问题。原缓存机制因文件变更导致频繁整份解析，引发卡顿。该插件通过在进程生命周期内保留已解析事件日志，实现长会话仅解析一次。当文件未变时直接返回，变长时仅读取新增字节并追加解析。插件严格校验条件以避免错误，并配合 `freezeStoredEvents` 处理。配置项包含内存预算与淘汰策略，适用于长会话场景。
Read MoreDeepSeek Harness Plugin: dsh-prompt-injector (General Per-Turn Context Injection)
`dsh-prompt-injector` is a plugin for DeepSeek Harness (DSH) designed to address the issue of rules being forgotten in long sessions. It leverages the memory mechanism, allows prompt lists to be managed via a UI, and injects context in the form of a “Context injection” notification line for each conversation (or after context compaction), ensuring that the model always “sees” critical rules. The plugin’s core features include per-turn injection, selectable trigger modes, abbreviation filtering, a zero-intrusion design, and UI-based management. It is responsible only for injecting text into the context and does not execute any actions, making it suitable for repeatedly emphasizing specific constraints in coding or complex Q&A scenarios.
Read MoreDSH Question Navigator: User Question Navigation Plugin in Long Sessions
`Cocowwy/dsh-plugin-question-navigator` is a long-session question navigation plugin designed for the DeepSeek Harness (DSH) Web client. It aims to address the pain points of quickly locating user historical questions and reviewing context in long conversations without requiring modifications to the DSH source code. Integrated via the `shell.overlay` slot, this plugin supports displaying controls alongside DSH's native "Back to bottom" button. It enables one-click navigation to the most recent user message and allows sequential traversal of earlier questions, supporting both rendered and paginated history messages. The plugin adapts to light/dark/custom themes, displaying text or icons depending on the device type (desktop or mobile), and switches between Chinese and English interfaces according to the browser language. It requires DSH Web 0.1.0-rc.6 or later.
Read Moredsh-argp: DSH's third-party context compression engine
dsh-argp is a third-party context compression engine for DeepSeek Harness, designed to address the issue of increasing context window pressure in long sessions while retaining historical original text. Its core mechanism comprises two stages: Stage-1 involves the model proposing extraction, summarization, or ignore actions for atomic units, which are implemented after adjudication by a deterministic guard, ensuring the LLM only proposes candidates rather than directly rewriting; Stage-2 prunes atomic units based on reference graph rules when thresholds are exceeded, a process that does not require LLM calls. Compressed or pruned content is stored in an append-only log, supporting retrieval at both summary and detail levels. During installation, the native summarizer must be disabled in the profile and the plugin added to avoid duplicate loader configurations. Users can customize the model backend, with default support for OpenAI-compatible interfaces. The repository provides multiple verification commands to test performance. This tool is suitable for scenarios requiring control over long-session context and preservation...
Read More