yoza10635/dsh-argp

README 摘要

English 中文 ARGP — 0-LLM Deterministic Context Compaction for DeepSeek Harness ARGP ( A tomic R eference G raph P runing) is a third-party CompactionEngine for DeepSeek Harness (dsh) that compresses conversation context without any LLM calls : instead of rewriting history into a summary, it selectively forgets. - 0 LLM in the compression phase — pure graph rules, deterministic and convergent. - Selective forgetting, not rewriting — pruned content stays in the append-only session log and is retrievable via a built-in recall pruned tool. - Engine-agnostic seam — mounted as a drop-in replacement for compaction-basic through the standard CompactionEngine interface. Status: research/validation stage. The full pipeline (mount → prune → recall, transaction invariants) is validated on dsh 0.1.0-rc.6 with DeepSeek v4-flash (see Reproduce). Declarative production mounting (P4) is in progress. Why Summarizer-based compaction (e.g. compaction-basic ) rewrites history with an LLM at compression time: cost scales with context, information is lossy, and compression ratio is not controllable. ARGP takes the opposite route: dependencies are captured structurally while the conversation happens (a sma…

分类