daftari 1.10.0 → 1.12.0

package/CHANGELOG.md

@@ -7,6 +7,64 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [1.12.0] - 2026-05-26
+
+### Added
+
+- **Cross-platform MCPB packaging (#66).** The `.mcpb` artifact now
+  boots on both macOS (arm64) and Windows (x64). A single universal
+  package bundles platform-tagged native binaries for `better-sqlite3`
+  (under `build/Release-${platform}-${arch}/`) and `sharp`, and a
+  one-line loader patch in `better-sqlite3`'s `lib/database.js`
+  selects the right binary at runtime from `process.platform` /
+  `process.arch`. The manifest's `compatibility.platforms` is back
+  to `["darwin", "win32"]`. `npm run pack:mcpb` (new) builds the
+  universal artifact from a darwin-arm64 host.
+
+- **MCP tool annotations.** All 14 tools now carry a `title` and the
+  appropriate safety hint — `readOnlyHint` for read/search/analysis
+  tools, `destructiveHint` for write and curation tools. MCP clients
+  use these to label tools and to decide when to prompt for
+  confirmation before a call.
+
+### Fixed
+
+- **`vault_write` no longer rejects writes that omit `updated` /
+  `updated_by`.** The server stamps both fields on every write, so requiring
+  callers to also supply them was redundant — and a caller who omitted them
+  (reasonably) got `invalid frontmatter: updated: missing required field;
+  updated_by: missing required field`. The fields are now filled in before
+  built-in schema validation runs, then re-stamped post-validation by
+  `performWrite` as before. Callers that still supply them keep working — the
+  server-side stamp wins, identical to the previous behavior. The MCP input
+  schema description now flags both fields as server-managed.
+
+## [1.11.0] - 2026-05-21
+
+### Added
+
+- **`vault_themes` thematic clustering** (#56). New MCP tool surfaces
+  thematic clusters across the vault. For each document the tool mean-pools
+  its chunk embeddings into one vector, L2-normalises, and clusters the
+  resulting per-document set with hand-rolled k-means (k-means++ init,
+  Lloyd's iterations). Default behaviour sweeps k ∈ {10, 15, 20, 25} and
+  picks the k with the best mean silhouette; an explicit `k` argument
+  skips the sweep. Each theme returns a heuristic label (TF-IDF over
+  titles + tags — no LLM call), a coherence score (mean pairwise cosine
+  inside the cluster — `null` for singleton clusters, where there are no
+  pairs to average), representative documents nearest the centroid, the
+  most frequent tags, and `secondaryDocs`: documents whose primary
+  cluster is elsewhere but whose pooled vector also aligns with this
+  theme's centroid (surfaces cross-cutting documents that the hard
+  one-doc-one-theme partition would otherwise hide). Optional
+  `collection` and `tags` filters scope clustering; RBAC drops documents
+  the caller cannot read. Output is deterministic for the same vault
+  (fixed seed). No new storage — reads the existing `chunks` /
+  `embeddings` tables. v1 is one-doc-one-theme at the partition level
+  (`documentCount` still partitions by primary); true multi-theme
+  membership, HDBSCAN, seeded-search/coverage mode, and LLM labels are
+  deferred.
+
 ## [1.10.0] - 2026-05-21
 
 ### Added