Release Notes

A month of code-intelligence upgrades across Gortex v0.19 to v0.39: smarter retrieval, wider language ingest, deeper analysis, better agent ergonomics, and a rebuilt storage layer.

A plain-English tour of the May–June 2026 Gortex updates for the engineers who use it: what each change means in practice, plus the exact commands and settings to get the benefit — no graph theory required.

Gortex tracks each git worktree as an independent repository instance with its own graph, so usages, callers, and resolution stay correctly isolated per checkout.

A rebuilt storage layer — a pure-Go SQLite backend by default (no CGO, FTS5 search), sidecars that keep agent knowledge off the hot path, warm-restart reconcile in seconds, and graph walks pushed into indexed Store capabilities.

Make the model a configuration choice, not a code dependency: one provider interface, nine interchangeable LLM backends, and graph-aware routing by task complexity.

Two memory layers, a multi-agent coordination registry, and a preview-and-verify edit gate make Gortex a safe, productive teammate for autonomous coding agents over MCP.

Push hard codebase questions into the graph: an interprocedural bottleneck analyzer for hidden hot paths that compound across calls, and first-class capability edges that make a least-privilege audit a one-hop query.

Gortex now ingests infrastructure-as-code, legacy systems, docs, images, PDFs, and live database schemas into one graph — plus two config-driven SPIs to add your own formats without a fork.

A code graph is only as good as its edges. A framework dynamic-dispatch synthesizer engine recovers the cross-language and cross-framework call edges that static parsing alone misses.

smart_context packs the smallest correct working set for a task: dependency-closure selection, a budget that scales with the graph, fidelity tiers with compress_bodies, and delta packing that re-sends only what changed.

How Gortex turned code retrieval from tuned-by-feel into measured-and-ranked — graph-centrality reranking, per-query lexical/semantic blending, query understanding, and a real P@K/R@K/MRR evaluation harness.