Repository navigation
v0.35.0
·
62 commits
to main
since this release
Immutable
release. Only release title and notes can be modified.
Highlights
- Native omp extension and bundled codebase-search skill, with host-compatible schemas and preserved system-prompt blocks.
- Better conceptual retrieval admission and relevance-aware diversity; faster cross-file deduplication.
- JavaScript/TypeScript generator indexing and call attribution, with incremental migration of existing indexes.
- Preserve outgoing call-graph edges during symbol refreshes; incrementally repair legacy catalogs without deleting indexes or requesting new embeddings.
- Correct mixed-evidence evaluation grading and Pi integration evidence targets (agent-context dataset 1.0.1).
- Correct nested-directory CLI/MCP checkout discovery, Codex installation docs, and host-provided TypeBox dependency declarations.
Release checks
- Full delta since v0.34.0 reconciled, including #397, #399, #400, #401, and #403.
- Package, lockfile, Claude/Codex manifests, and security support policy synchronized at 0.35.0.
- Refreshed local build/typecheck/lint/full suite passed: 2,582 tests passed, 6 skipped. Release PR and merged-commit CI passed at
559b7ea. - All six native targets built successfully, including Windows ARM64. Trusted npm publishing workflow passed.
- Both npm identities publish
0.35.0aslatest, with SLSA provenance. Clean installs passed ESM imports, CJS requires, native loading, all four CLI names across the identities, and MCP initialize/tools-list (19 tools).cbiandocbidry runs parsed one TypeScript file into one chunk without embedding requests.npm audit --omit=devreported zero vulnerabilities for each install.
Known non-blockers
- Dependabot #72: xxhash-rust 0.8.15 has a custom-secret API issue. This project's native callers use xxh3_64 and Xxh3::new; no custom-secret calls are present. The dependency alert remains open.
- Windows ARM64 builds use a preview runner and are best-effort; the other five targets are required.
- Retrieval improvements do not imply uniform ranking improvement for every query. Evaluation scores using agent-context 1.0.1 are not directly comparable to earlier dataset fingerprints.