v1.3.0

[1.3.0] — 2026-10-16

The typed-calls release: TypeScript receivers resolve through declared types, interfaces dispatch to implementations, imports resolve through project metadata, and the graph reports how completely it understood each file.

• Explicit query seeds. astria query --seed <symbol> (repeatable; MCP query_graph gains a seeds argument) anchors traversals at known symbols instead of keyword ranking. Seeds resolve by id, label, bare name, or unique source path; unresolved seeds are reported, and all-unresolved returns an explicit miss — never a silent keyword fallback.
• Type-aware receiver resolution (TypeScript). this.service.save() resolves through constructor parameter properties, typed class fields, typed locals, and new initializers, following inherited members through extends. Bindings stay name/type-derived (RESOLVED), never compiler-proven; ambiguous receivers stay unresolved.
• Interface dispatch. Interfaces are extracted with their members and each member links to every implementing class via INFERRED dispatch edges — possible implementations for DI-style flows, never presented as proven runtime calls.
• Typed cross-repository member calls. Member calls whose receiver type is known but whose member lives in another repository retain receiver::type::<T>::<m> evidence; the global resolver links them on a unique foreign name+member match.
• Project-aware import resolution. tsconfig paths aliases with baseUrl (longest prefix), workspace package entries (authored TypeScript only — build outputs never map), and scoped subpaths resolve to real source files. Config changes refresh every affected binding on the next run.
• Extraction completeness reporting. Every file carries complete / partial / unsupported / skipped / suspicious-empty status with parser-error lines (schema v11), surfaced in astria stats, astria diagnose, and MCP graph_stats — missing relationships are now distinguishable from incompletely-understood files.
• Evidence-weighted community detection. Clustering weights edges by stored confidence (EXTRACTED > RESOLVED > INFERRED > SEMANTIC) through every pass, so many weak links can no longer outvote strong structure; query-usage learned edges never shape module boundaries.
• Actionable memory. astria reflect produces preferred sources, corrections, dead ends, contested evidence, and tentative findings, discarding citations of deleted nodes; recorded verdicts gate learned-edge promotion so repeated unsuccessful exploration never becomes authoritative (schema v12).
• Grammar-aware extraction caching. Cache keys embed per-language tree-sitter versions derived from Cargo.lock, so a grammar upgrade re-extracts only the affected language — one AST-cache invalidation on upgrade, per-language afterwards.
• Project language overrides. .astria.toml [languages] remaps ambiguous suffixes (.inc = PHP) consistently across discovery, parsing, watch, and cache identity.
• Stable community labels. LLM labels survive refactors by majority continuity (strict majority of both memberships), splits and merges assign one identity unambiguously, and summaries refresh separately from labels.
• Post-merge git hook refreshes the graph after true merges; the editor hook-guard now answers searches and reads with matching graph symbols, locations, and a ready-to-run command instead of a generic nudge (bounded, once per key per 30 minutes, fail-open).
• Measured before release (RESULTS-2026-10-16.md): ten-suite paired eval vs Graphify won or tied on all ten (file recall@5 72.9–100% vs 33.3–100%); token parity 2,875x vs 2,575x with ~11% fewer tokens per answer; self quality gate passed (hit@5 0.800, 0 budget violations); LoCoMo full set recall@5 83.6%; version A/B vs 1.2.0 shows no regressions with reads 9–37% faster.

What's Changed

• test(query): pin identifier-style questions to their definitions by @nodesify-technology in #136
• release: promote develop to main (typed calls + project-aware imports + completeness) by @erictong0602 in #151

Full Changelog: v1.2.0...v1.3.0