dsh-argp

Third-party context compaction engine for DeepSeek Harness (dsh). It uses a dual-engine design: Stage-1 per-atom compression where the LLM proposes extract/summary/false and deterministic guards decide, and Stage-2 lazy reference-graph pruning with 0 LLM calls. An append-only log is the source of truth, and recall_summary/recall_detail provide paginated recall. npm default mounts only Stage-2; Stage-1 is enabled via a peratom config block.

• Stage-1 per-atom compression: LLM proposes extract/summary/false per atom; deterministic guards reject extract if any high-signal token is missing.
• Stage-2 reference-graph pruning uses 0 LLM calls, reverse topological removal, lifecycle states ACTIVE→COMPLETED→PRUNABLE→PRUNED, and precise compression ratio.
• Stage-2 trigger ladder: turn-start proactive check (step===1), mid-turn 0-LLM pressure prune (midTurnPrune default true), and max-tokens continuation.
• Append-only log is the single source of truth; compressed/pruned content is never destroyed and can be recalled.
• RecallZoom: recall_summary and recall_detail with from/limit pagination; text blocks verbatim, tool-call parameter objects rebuilt with JSON semantic equivalenc
• npm default mounts only Stage-2 (0-LLM graph pruning); Stage-1 pipelines (compressor/declarer/zoom) are enabled by a peratom config block.
• Config knobs: midTurnPrune (default true), midTurnTurnGuard (default 0), continuationNotice (default built-in sentence), reactiveRetries (default 2), midTurnAct
• Install via dsh plugin --profile <name> add dsh-argp; disable stock summarizer with compaction-basic disabled: true in profile cordis.patch.yml.