@kuanfu0430/dsh-compaction-tune
Composer control for DeepSeek Harness auto-compaction: set a percent (30–80) or absolute token threshold that persists in settings.yaml.
430 results
Composer control for DeepSeek Harness auto-compaction: set a percent (30–80) or absolute token threshold that persists in settings.yaml.
在每次会话开头固定注入一段用户配置的文本上下文(设置页输入框可编辑),作为模型可见的 instructions 注入第一轮请求。 | Injects a user-configured text block as model-visible instructions context at the start of every DSH session (editable from the settings page).
DeepSeek Harness plugin: aggressive token optimization without affecting task quality. Trims tool results, manages context window pressure, optional LLM-summary compaction, stable prompt-cache layout.
Probe local model servers (llama.cpp / Ollama) for the real runtime context window and keep llm-pi-ai's contextWindow in sync automatically — tighter before overflow, wider when the server grows. Zero-config profile plugin for DeepSeek Harness.
Persistent task state and context-aware compaction for DeepSeek Harness coding sessions.
Inject Context for DeepSeek Harness (DSH): persist your own context entries and inject them as a standalone 'Remember:' row — right after your message — per turn or once per session (per-entry 'Every turn' switch). Simple/Advanced datasets, native Settings editor, standalone manager page.
会话注入门控:给 DeepSeek Harness (dsh) 加一个会话级上下文裁剪闸门,按会话取消/恢复插件贡献的 skill、tool、系统提示词段落——3 条影子通道、5 条本地路由、零依赖零配置。Per-session injection gate for DeepSeek Harness (dsh): trim skills, tools and prompt sections per session. Zero-config, zero dependencies.
DSH plugin: Codex-style context management — token-budget notice, model-facing new_context, and handoff reset instead of summary compaction.
DeepSeek Harness plugin: plants a canary string in model-visible context and denies any tool argument or outbound URL that echoes it back
Keep native vision, drop historical images before they hit relay providers. Global DeepSeek Harness bundle for every agent.
Hindsight long-term memory for DeepSeek Harness: auto-recall context injection before each agent step (agent/pre-step) and auto-retain of each finished turn (session/event).
LinkDigest for DeepSeek Harness: mounts the remote LinkDigest MCP (https://linkdigest.dev/mcp) via the in-box @deepseek-ai/dsh-mcp-client bridge — turn a Xiaohongshu, Douyin, TikTok, YouTube or X link into text an agent can read.
Fix the DeepSeek Harness Web GUI over plain HTTP on a LAN: injects a crypto.randomUUID polyfill (getRandomValues-backed) before dsh bundle code runs. No official source changes needed.
DSH plugin: adds a platform-labelled 'reveal in file manager' row to the Workspace context menu (Windows/WSL Explorer, macOS Finder)
dsh plugin: compresses verbose tool error payloads (drops stack frames, caps message lines) on the tools/post-execute waterfall before they reach model context.
DSH plugin with an in-process JavaScript port of ContextPrism (Local Context Engine)
DSH 对抗多智能体三层失效的防御插件:读取台账去重度量(read-ledger)、免疫压缩规则区+完成机械门禁(immutable-core+completion-gate)、低成本审查lane(role-router)。零外部依赖,注入即用。
一个基于Model Context Protocol (MCP)的WHOIS域名查询服务器,支持877+顶级域名和169个国家代码顶级域名的解析,提供全面的域名注册信息查询功能。
DeepSeek Harness web plugin: a Model Capabilities settings page for declaring model input modalities, context window, output cap and image budgets without hand-editing settings.yaml.
Summary-free context window rollover, model-authored notes, and scoped raw history lookup plugin for DeepSeek Harness (DSH)
组装式上下文(Assembled Context): a context tree over the dsh session surface with per-node assemble modes (full / key / off), agent-authored presets, and a browser panel that decides what the model actually sees.
DSH 增强型 MCP 客户端插件:在设置界面管理多台 MCP 服务器,自动发现并注册工具给 Agent
DeepSeek Harness 圆盘快捷菜单:按轮次分叉、打断生成、复制路径与会话 ID、打开目录和导出日志。
Session image payload governor for DeepSeek Harness: audit which images a session still ships, pick the ones to keep, and move the rest out of the model context.