dsh-compaction-instant
VCC-style instant, near-lossless deterministic compaction engine for the DeepSeek Harness — a drop-in replacement for @deepseek-ai/dsh-compaction-basic
43 results
VCC-style instant, near-lossless deterministic compaction engine for the DeepSeek Harness — a drop-in replacement for @deepseek-ai/dsh-compaction-basic
Headroom-inspired automatic context compression plugin for DeepSeek Harness (dsh): compresses tool outputs before they reach the model and keeps every lossy compression reversible via CCR retrieval tools.
DeepSeek Harness plugin that uses configured model providers for image analysis and context compaction.
Context compression tool (context_compact): the agent writes the replacement checkpoint itself and hands it to the host compaction engine, which skips the LLM summarizer call; automatic compaction stays on the official engine.
Cut token cost without cutting model intelligence — a DeepSeek Harness (dsh) bundle plugin: structure-aware compression + TOON-style lossless encoding first + reversible spill-to-disk + real-time dashboard.
Context Assembler DSH V0.99 — Context Assembler plugin for DeepSeek Harness (dsh): context compaction, cache-friendly topic-block management, water-pressure topic splitting, tool trace/rewrite, handoff planning and reality recall injection.
DSH ↔ agentmemory session-memory bridge (backend-only). Mirrors session lifecycle into agentmemory with standard hookTypes (compression-friendly, dedup-safe), exposes memory_recall / memory_remember tools, and injects memory context via agent/pre-step. Configuration comes from the cordis.yml row.
Guarded context compaction for DeepSeek Harness (dsh): the LLM proposes, deterministic guards dispose — eager per-atom shrink (extract/summary/false under verbatim guards) + lazy reference-graph eviction (0-LLM) + byte-exact recall from an append-only log. 压缩率精确兑现,历史永不销毁。
dsh网页端网络优化:通过缓存与压缩技术降低传输,从而大幅提升网页加载速度;同时提供网络断连指示与自动断网重连功能。非常适合追求极致性能或网络不稳定用户使用。
输出 token 压缩:让 agent 用紧凑的穴居人语回复(省 ~65% 输出 token),代码/命令/错误保持字节精确;token 下沉点审计(caveman learn:safe fix/offload/habit/load-bearing 分类)。受 JuliusBrussee/caveman(86k★ MIT)启发。
Generic stateless dialogue compression that mimics human memory — forgetting-curve bounded context for LLM conversations, with a ready-to-use adapter module for DeepSeek Harness (DSH).
Content-aware tool output compression for DeepSeek Harness.
Adaptive Reversible Context (ARC) for DeepSeek Harness — provider-aware, cache-conscious context governance with reversible recovery.
DeepSeek Harness plugin: gzip /api responses — fixes history loading failures over slow links
Live token/cost-savings dashboard for the Headroom compression proxy inside the DeepSeek Harness (DSH) Web UI: a settings-page dashboard plus a persistent stats line under the composer. 在 DSH 内实时展示 Headroom 压节省统计:设置页仪表盘 + 输入区常驻统计行。
Live llmtrim savings dashboard inside the DeepSeek Harness (DSH) Web UI: a settings dashboard plus a rotating carousel stats strip under the composer.
Context compression skill for DeepSeek Harness -- intelligent conversation summarization
Pure-incremental directive-driven compaction for DeepSeek Harness: /compact-directive <requirement> summarizes the session middle per your requirement, /trim-directive <requirement> trims the whole conversation per your requirement
AgentFrame compaction backend for DeepSeek Harness: semantic + physical KV compression (28.4x) replacing LLM summarization
Custom wallpaper engine for the dsh web GUI: upload an image (client-side WebP/JPEG compression), frosted-glass blur, pane opacity, and auto font-color linkage.
DSH plugin: auto-downscale oversized images (and a portable Agent Skill). Powered by sharp. Install from GitHub — no npm publish.
把长文本渲染成图片发送来压缩 LLM 输入 token(DeepSeek 视觉模型、每图 384 token 封顶)— text-as-image token compression for DeepSeek Harness
附件自动降级:DSH 图片准入(单边 ≤2000px / ≤3.5MB / ≤4000万像素)超限图片不再拒绝,用 sharp 自动缩小/重编码后入库,用户上传手机原图不再踩线。
Context compression for small-context models on the DeepSeek Harness Web GUI: one click condenses the whole conversation into a few sentences, opens a fresh session in the same workspace, and re-injects the summary as the first message so a small model keeps working with a clean, compact context.