Status

Current release: v0.2.0. 1,510 unit tests passing. All features below are live + tested + reproducible on your hardware.

Shipped this release (v0.2.0)

• Hook relay + filter pipeline — 24 command-specific rules; Layer 3 smart denials (re-read suppression, command replay, trivial routing, symbol upgrade, redundant validation); 80.37× fixture-suite reduction.
• Symbol indexer — 25 tree-sitter grammars, FTS5 search, dependency graph, container hierarchy, incremental FSEvents updates; cold < 5 ms / cached < 1 ms.
• Session DB + analytics — SQLite + FTS5; per-call token tracking; budget enforcement at hook layer; BM25+RRF search fusion; @-mention context pinning; session continuity; JSONL data export.
• Local ML — MiniLM embeddings (384-dim) + Gemma vision via MLX on Neural Engine; fully offline after download; L0/L1/L2 tiered context pre-warming.
• Compound learning — recurring→staged→applied with Laplace-smoothed confidence; four artifact types; Schneier gate on instruction-patch auto-apply.
• Knowledge base — markdown source-of-truth + rebuilt SQLite index; enrichment validator, rollback, timeline; KB↔learning bridge.
• Security hardening — SSRF guard with DNS pre-resolution + redirect re-validation + subresource blocklist; prompt-injection guard (multilingual + homoglyph NFKC-folded); 13 secret-detector families; socket-auth handshake; opt-in structured JSON logs with sink redaction.
• JS-rendered web fetch — senkani_web returns AXTree markdown via WKWebView; version negotiation via senkani_version.
• New website — this rebuild. Multi-page wiki + landing page per spec/website_rebuild.md.

Links

• Full changelog
• GitHub repo
• About + license