dsh-context-compression-improved
DSH 插件 dsh-context-compression-improved
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add dsh-context-compression-improved@0.1.1
DSH 插件 dsh-context-compression-improved
Context Compression Improved 以 dsh-context-compression-improved 发布，当前版本为 0.1.1。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
The installable Product Bundle for the unofficial community DeepSeek Harness context-compression selector.
0.1.0 highlights: DeepSeek V4 Flash Vision's official tokenizer is included; users can choose the model-driven Auto Compact threshold; standard profile watermarks and compression parameters follow that choice.
dsh plugin --profile web add dsh-context-compression-improved@latest
dsh --profile web --dump-config
This one command installs the Bundle and its pinned @huggingface/tokenizers dependency; there is no separate runtime package. The Bundle contributes Host settings, the Web UI, and a reversible preset overlay. All presets except the exact built-in minimal id receive the compression stack; switching non-Minimal presets preserves the saved selector settings. Minimal pauses plugin compression without deleting the setting.
It provides deterministic Fresh, Aggregate, History, Native tool-result, and Custom TailTrim compression; a plugin-owned context_compression_retrieve recovery tool; structured audit records; and a pinned offline DeepSeek V4 tokenizer. It uses only public DeepSeek Harness APIs and supports both 0.1.1-rc.2 and 0.1.2-alpha.5; it does not patch Harness core.
Lossy rewrites require exact same-revision counts before and after replacement. Verified model ids are deepseek-v4-flash, deepseek-v4-pro, and deepseek-v4-flash-vision-exp. Vision text counting uses the separately pinned deepseek-ai/DeepSeek-V4-Flash-Vision-Exp tokenizer. Vision images use the official image-processor arithmetic validated against official golden fixtures and are reported as tokenizer-estimate: the four alignment residues over valid intrinsic dimensions are collapsed to their midpoint, with 384 tokens retained as the per-image upper bound; malformed or unevaluable dimensions use a fixed 256-token fallback. The estimate is not exact because the absolute prompt position and the adapter's final image projection are not publicly observable. Image-bearing tool-result candidates remain exact-ineligible and intact; unknown models, unavailable tokenizer assets, unsafe tool groups, and incomplete text measurements still fail open.
Audit log records use the prefix context-compression audit and distinguish policy snapshots, skipped/disabled components, committed rewrites, fail-open errors, and observed core compaction/summary events. No audit record contains prompt or tool-result content.
The first policy-frozen record carries the complete settings/deployment snapshot. It is emitted through the Harness logger, so its retention follows the configured log sink; standard prune/replacement Session events remain the durable proof of committed rewrites. This package does not hard-code a private path below ~/.dsh.
Verified compatibility: DeepSeek Harness dsh-v0.1.1-rc.2 and dsh-v0.1.2-alpha.5, Node ^22.19.0 || >=24. No Harness core patch is required.
See the full README for profiles, defaults, audit evidence, Adaptive/cache limitations, upgrade/removal, and security guidance. Tokenizer provenance and checksums are in assets/deepseek-v4/manifest.json and assets/deepseek-v4-vision-exp/manifest.json, summarized in THIRD_PARTY_NOTICES.md. This package is not affiliated with or endorsed by DeepSeek.