@unified-product-graph/core 0.9.22 → 0.9.23

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/CHANGELOG.md CHANGED
@@ -7,6 +7,18 @@ This project uses [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
7
7
 
8
8
  ---
9
9
 
10
+ ## [0.9.23] - 2026-06-11
11
+
12
+ **Closes the duplicate-delivery hole that 0.9.22 only half-fixed, and stops a cross-product write from duplicating on retry.** 0.9.22 deduped a re-delivered mutating call by its JSON-RPC request id, but the real re-delivery carries a *fresh* request id (a client-level re-issue, invisible to a request-id ledger): each create handler re-executed and minted a second copy with new ids. Because the replay lands on the *next* mutating call and never on a read, a "write then recount" check passed falsely — which is why 0.9.22 looked fixed.
13
+
14
+ ### Fixed
15
+ - **Content-level idempotency for mutating MCP calls (`@unified-product-graph/mcp-server`).** The local server adds a second dedup layer keyed on the call's payload (tool + active product + normalised arguments), independent of the request id. A re-delivered mutating call whose payload matches a recent one replays the original result instead of writing a duplicate. The window is bounded by COUNT (the most recent 64 distinct mutating payloads), not wall-clock: the replay lands on the next mutating call, so a sliding count window catches it without timing guesswork. Only successful results are recorded, so a transient error stays retryable. Pass `allow_duplicate: true` on a mutating call to opt out for a deliberate identical re-create. The store was never the source: it dedupes by id and never mints ids, so a fresh-id duplicate can only mean the create path re-ran.
16
+ - **Cross-product edge writes are idempotent and no longer self-corrupt on a tmp-rename race (`@unified-product-graph/sdk`).** `create_cross_product_edge` could throw `ENOENT: rename .../portfolio.upg.tmp -> .../portfolio.upg` even though the edge had already persisted (a debounced save racing an explicit flush on a SHARED tmp path), and a naive retry then appended a second identical edge. The portfolio writer now uses a per-write unique tmp path (cleaned up on any failure), and `addCrossEdge` collapses an identical `(source, target, type)` re-create onto the existing edge, so a retry is a safe no-op.
17
+
18
+ No entity, domain, region, edge, or tool-count change (entities 315, edges 980, local tools 123). The cloud server (Postgres-backed, separate write path, no file watcher) is unaffected.
19
+
20
+ ---
21
+
10
22
  ## [0.9.22] - 2026-06-11
11
23
 
12
24
  **Stops silent data duplication from re-delivered MCP writes, and stops dedupe from destroying structure.** Three fixes for a HIGH-severity report where a mutating call was applied twice (the duplicate landing a few calls later, past any immediate recount).
package/dist/index.cjs CHANGED
@@ -28431,7 +28431,7 @@ function isCanonical(text) {
28431
28431
  }
28432
28432
 
28433
28433
  // src/index.ts
28434
- var UPG_VERSION = "0.9.22";
28434
+ var UPG_VERSION = "0.9.23";
28435
28435
  var UPG_FORMAT_VERSION = "0.4.0";
28436
28436
  var MARKDOWN_FORMAT_VERSION = "0.1";
28437
28437
  var UPG_TYPES = getTypes();