Bootstrap AI Layer (ai-layer Claude Code plugin)

A README for ai-layer, a free MIT Claude Code plugin from Kris Codes LLC that gives a session persistent memory and code intelligence: 88 skills and 25 agents across three packs (core, marketing, infra), built on two MCP servers — duckdb-brain (a DuckDB "brain" that remembers across restarts) and typescript-lsp (symbol/call-site lookup so the agent doesn't read whole files to find who calls a function).

The interesting engineering is the three-mode brain resolver: gateway (a commercial Koinessa Brain Gateway — not included), local (single-machine DuckDB file), or setup (deps not yet installed). Precedence is deliberate and not a fallback: a configured-but-unreachable gateway stays in gateway mode and fails open with a named error rather than silently downgrading to a different brain. Setup mode is equally explicit — one status tool per server that points at /ai-layer:doctor --fix, so nothing degrades quietly. One command (doctor --fix) does the whole install because the native DuckDB binding and the spawned language server can't be bundled into a git-cloned plugin.

The lifecycle is the other half: init → create-prd → plan-review → quality-gate → ci-local → fix-ci → ship → security-review, plus session memory commands (recall/remember/resume-work) and a retrieval-first SOP that routes symbol lookups through the LSP index instead of grep. The README even ships a test (tests/docs/readme.test.mjs) that checks the architecture diagram against the resolver, so a fourth mode fails CI rather than shipping a stale doc. Privacy is a stated design constraint: local mode makes no network connections; gateway mode talks to exactly one host you configure; no telemetry in any mode. Distribution is a GitLab git URL — releases are one-commit snapshots of a private monorepo, with merge requests re-applied upstream rather than merged.