@geohar/pi-cribsheet 0.22.0-dev2 → 0.22.0

package/CHANGELOG.md

@@ -9,6 +9,56 @@ This file ships in the Pi package, where Pi shows the entries newer than the
 version you last ran. `scripts/bump-version.sh` refuses a release whose
 version has no heading here.
 
+## [0.22.0] - 2026-11-09
+
+### Fixed
+
+- **Shared store recovery.** One failed reconnect no longer leaves the
+  shared (Quack) client unusable until the daemon restarts; threads that
+  lose the same session relaunch the server once between them; and with
+  `[duckdb] launch = false` a reconnect waits for the supervised server.
+- **`tags=` lookups return k results.** The tag filter now runs inside the
+  store query instead of after the top results were picked.
+- **The embedding cache can't be poisoned** by the hash-embedder fallback:
+  it keys on the embedder actually built and never caches hash vectors.
+- A zero vector scores 0 instead of NaN, which sorted above every match.
+- Hybrid retrieval scores keyword-only candidates on their stored vectors
+  (heading breadcrumb plus body, as the dense search ranks them) instead of
+  re-embedding the bare body.
+
+### Changed
+
+- `code_lookup` and summary-alias scoring use numpy (about 140 ms to 0.1 ms
+  per code query at 16k symbols). numpy is now a direct dependency.
+- Shared mode sends each vector as one literal: writes about 20x faster.
+- A note write is one transaction; metadata refreshes are one statement
+  per batch.
+- A `crib_meta` table records the schema version, record-cache format and
+  embedder; a change drops or rebuilds what depended on it, and switching
+  embedders at the same vector size now recreates the store.
+- The generation ledger forgets files that no longer need a description;
+  the embedding cache drops other embedders' vectors.
+- `crib status` shows `store_stats` (rows, versions, sizes, reconnects) and
+  embedding-cache hits; the store checkpoints after a startup reconcile.
+
+### Added
+
+- `crib store install-extensions`: fetch the shared-mode DuckDB extensions
+  ahead of time, for containers and offline machines.
+
+### Removed
+
+- The `json` store backend (`[store] backend = "json"` now warns and uses
+  DuckDB; the old `~/.cache/crib/store.json` can be deleted), and leftover
+  chroma-era code paths.
+
+### Upgrading from a 0.22.0 dev release
+
+- Chunks whose tags were stored as a comma-joined string are converted to the
+  new list form when the store opens, so `tags=` lookups work at once.
+- Embedding-cache rows keyed by the old configured model name are dropped
+  once; the vectors in the store itself are untouched.
+
 ## [0.22.0-dev2] - 2026-11-09
 
 ### Added
@@ -33,8 +83,8 @@ version has no heading here.
 
 - **chroma.** DuckDB is the only vector store. Gone with it: `ChromaStore`,
   the shared chroma server, the `[chroma]` config table (now ignored with a
-  one-time warning), the chroma CVE guard, `crib store migrate`, and the
-  `chromadb` dependency. Upgrading: [docs/migrating-to-duckdb.md](docs/migrating-to-duckdb.md).
+  one-time warning), the chroma CVE guard, and the `chromadb` dependency.
+  Upgrading: [docs/migrating-to-duckdb.md](https://github.com/georgeharker/cribsheet/blob/main/docs/migrating-to-duckdb.md).
 
 ### Added
 

package/package.json

@@ -1,6 +1,6 @@
 {
     "name": "@geohar/pi-cribsheet",
-    "version": "0.22.0-dev2",
+    "version": "0.22.0",
     "description": "Pi extension: run the crib (cribsheet) memory + code-index MCP server via sharedserver, inject the reach-for-crib directive, and ship the /crib recall command. Pairs with pi-mcp-adapter.",
     "type": "module",
     "main": "dist/index.js",