Concepts
Explanation pages — the why behind senkani. Reference tells you what each tool does; concepts tell you why the system works. Start with What is senkani? if this is your first visit.
Compression layer
Three places Senkani compresses: input (before the LLM sees it), redundancy (before the tool runs), output (before responses grow). Why all three.
Hook relay
How the senkani-hook binary intercepts your agent's Read/Bash/Grep, routes to senkani equivalents, and denies with cached results when appropriate.
MCP intelligence
What "MCP tool" means and why senkani has 19 of them. How tree-sitter + FTS5 + MiniLM compose into sub-5 ms symbol search.
Three-layer stack
Layer 1 tools, Layer 2 hooks, Layer 3 smart denials. Why denial-with-cached-result is the biggest win on pathological sessions.
Compound learning
Four artifact types — filter rules, context docs, instruction patches, workflow playbooks — mined from your own sessions. The .recurring → .staged → .applied flow.
Knowledge base
Project entities + decisions + links. Markdown as source-of-truth, SQLite as rebuilt index. How it stays coherent when you hand-edit.
Security posture
v0.2.0 defaults: prompt-injection guard on, SSRF hardening on, secret redaction on, socket auth opt-in. Why senkani is a trust boundary and what that means concretely.