Zhuchen00123/dsh-compaction-cacheaware ↗★ 0
dsh-compaction-cacheaware
Reasonix-style cache-aware compaction backend for DeepSeek Harness (DSH). Replaces/enhances compaction-basic with compact_ratio, one structured summary checkpoint, and a stable recent tail.
AI 분석
核心用途是优化长对话的上下文压缩，提高缓存命中率并减少 Token 浪费。适合频繁进行超长对话、关注推理成本的用户。
설치
검증된 bundle이 없거나 호환성 검사에 실패했습니다. 먼저 저장소 설명을 읽어 주세요. 전체 README 읽기 ↗
Configuration
| Key | Default | Meaning |
|---|---|---|
compactRatio | 0.85 | Sole automatic trigger fraction. |
checkpointCeilingRatio | 0.5 | Normal auto-checkpoint acceptance ceiling. |
recentTailRatio | 0.1 | Recent verbatim tail fraction. |
recentTailMinTokens | 32768 | Lower bound for production tail. |
recentTailMaxTokens | 98304 | Upper bound for production tail. |
summaryMaxTokens | 16384 | Summarizer output cap. |
exceptionalMinSavingsRatio | 0.25 | Required savings when fixed prefix exceeds ceiling. |
minRecentKeep | 2 | Minimum recent messages kept. |
minCompactMessages | 2 | Minimum compactable messages. |
maxPinnedFirstUserTokens | 1500 | Pin first user turn if ≤ this. |
pinnedFirstUserWindowFrac | 0.15 | First-user pin window fraction cap. |
protocolReserveTokens | 256 | Framing reserve. |
summarizationProvider / summarizationModel | '' | Optional summary route; defaults to conversation route. |
auto | true | Register automatic pressure/overflow listeners. |