yu381792/dsh-superlcm

SuperLcm is an asynchronous context-compaction and lossless recall plugin built specifically for DeepSeek Harness (DSH), inspired by Clint Ehrlich and Theodore Blackman's LCM paper and explicitly scoped to DSH's append-only event log as the canonical source of truth. It runs an out-of-band Worker that uses an independent summarizationProvider and summarizationModel to build a layered summary DAG in the background, while the foreground Agent reply never awaits that compression model and is decoupled via its own AbortController. The plugin exposes six recall tools — lcm_grep, lcm_describe, lcm_expand, lcm_expand_query, lcm_reindex, lcm_doctor — that read exact DSH event seq ranges rather than re-summarizing summaries. Installation uses dsh plugin --profile web|acp add pointing at the GitHub Release tarball SuperLcm-0.3.0-alpha.9.tgz, and operators are instructed to replace the existing compaction provider instead of stacking a second one alongside it.

Typical workflow keeps the active surface small by keeping only the frozen SYSTEM + CHECKPOINT prefix and a recent fresh tail, while older raw history is folded into summary nodes that still carry child_ids and source event seqs. The default policy accepts only mode: rolling with foldTiming: background, retains the last 24 surface nodes and at least 32k fresh tokens, prepares 64k batches off the critical path, and only commits ready summaries under softActiveTokens 160k or hardActiveTokens 220k pressure. Atomic transactions enforce the full compaction/start → compaction/summary → checkpoint user/message → compaction/end lifecycle, and snapshots are pinned before summarization so tail growth or rewrites cause discard-and-rebuild rather than corrupted ranges. It targets DSH power users who want stable Prompt Cache prefixes and bounded attention dilution in long-running sessions.

Hard requirements are Node.js 22.16 or newer and continued DSH exposure of its official compaction, LLM, tools, and session/event interfaces; CI additionally covers Linux on Node.js 24 and Windows/macOS on 22. SQLite lives at ~/.dsh/SuperLcm/lcm.sqlite, overridable via DSH_SUPERLCM_DB and DSH_HOME, holds only derivable indexes (lcm_nodes, lcm_edges, lcm_nodes_fts, lcm_scan_state, lcm_index_state) with WAL, and never stores original transcripts so it can be rebuilt by lcm_reindex. The 160k/220k thresholds are starting points for large-context models and must not be copied verbatim to small-context models; a fallback provider/model pair, if configured, must differ from the primary and is invoked at most once. First-run advice is to validate in an isolated profile, pin both DSH and plugin versions, run npm run validate plus npm pack --dry-run, and avoid automated upgrading into the main workspace. The project is MIT-licensed and explicitly unaffiliated with Voltropy, Martian Engineering, DeepSeek, the official interactive LCM site, or DSH upstream.