pi-mega-compact 0.20.53 → 0.20.54

package/README.md

@@ -11,7 +11,7 @@ A local-first context compressor for the [pi coding agent](https://github.com/ea
 - **Per-turn tracking + rewind.** Every turn, checkpoint, and recall hit lands as a row in an isolated `turns.db` — `turns`, `turn_recall`, `conversation_forks`. The dashboard **Turns tab** shows turn-by-turn memory: context pressure, the compact epoch that superseded each turn, and the exact checkpoints recalled into it. A **fork** action branches a conversation at any turn (carrying its recall set); a **rewind** action queues an intent the host consumes at the next `before_agent_start`. The `TurnStore` is contract-first (capability-gated reader/writer/admin views) so the same spine backs the dashboard, the TUI, or an API gateway.
 - **Cross-repo recall** — doors you close in one repo don't reopen when you move to another. A decision stored while hacking repo A is a recall hit the next time you're in repo B.
 - **Durable memory** — on a cadence the store auto-reviews and safe-keeps decisions, facts, and preferences as first-class RAG memories, so long-running projects remember what mattered.
-- **Prompt-cache optimization** — message separation + cache striping (default OFF, opt-in). Targets 82-90% cache hit rate by structuring context around provider cache boundaries. Enables with `MEGACOMPACT_MESSAGE_SEPARATION` and `MEGACOMPACT_CACHE_STRIPING`.
+- **Prompt-cache optimization** — message separation + cache striping (both default ON; disable with MEGACOMPACT_MESSAGE_SEPARATION=0 / MEGACOMPACT_CACHE_STRIPING=0). Targets 82-90% cache hit rate by structuring context around provider cache boundaries.
 - **Context health + KV cache poison validation** (v0.12) — a real-time composite 0-1 health score per turn from five sub-scores (drift, output quality, error rate, cache health, cache poison). Catches garbled/hallucinated output and provider-side KV-cache corruption *before* they waste tokens — the failure mode where a large-context model (e.g. DeepSeek V4 Flash, 1M window) degrades at <1% usage. Tri-layer cache poison validation: prefix hash (L1 FNV-1a), output-quality-by-cache-hit (L2 semantic), error-rate correlation (L3 behavioral). The dashboard **Health tab** shows a gauge, sparkline, sub-score bars, alerts, and per-model breakdown. Auto-mitigation (force compaction on degraded context, prefix break on poisoned cache) is default OFF — toggle it in the Maintenance tab.
 - **RAG suite** — query reformulation (TF-IDF + RRF), tiered routing (L0 cache -> L1 FTS5 -> L2 PGlite), recall-quality metrics (CRAG), memory graph, and HyDE (hypothetical document embeddings). All default ON with graceful fallback — opt out via `MEGACOMPACT_<NAME>_DISABLED=true`. HyDE auto-activates only when an LLM embedder (Ollama/HTTP) is configured — it's a no-op with the default TrigramEmbedder. The dashboard **Metrics tab** (v0.13) shows RAG recall-quality metrics (pass rate, avg lift, telemetry turns, HyDE runs) and the Overview tab has a RAG health card. Toggle any flag from the dashboard Setup tab.
 - **Debug bundle** — Maintenance tab in the dashboard has a **Gather Debug Logs** button that collects events, config, and store state into a shareable archive for bug reports.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-mega-compact",
-  "version": "0.20.53",
+  "version": "0.20.54",
   "description": "Layered, local, vector-backed context compressor for pi — supersede/collapse/cluster compaction with deduped inline recall.",
   "type": "module",
   "license": "BSD-3-Clause",