@pentatonic-ai/ai-agent-sdk 0.10.14 → 0.10.15

package/dist/index.cjs

@@ -878,7 +878,7 @@ function fireAndForgetEmit(clientConfig, sessionOpts, messages, result, model) {
 }
 
 // src/telemetry.js
-var VERSION = "0.10.14";
+var VERSION = "0.10.15";
 var TELEMETRY_URL = "https://sdk-telemetry.philip-134.workers.dev";
 function machineId() {
   const raw = typeof process !== "undefined" ? `${process.env?.USER || process.env?.USERNAME || "u"}:${process.platform || "x"}:${process.arch || "x"}` : "browser";

package/dist/index.js

@@ -847,7 +847,7 @@ function fireAndForgetEmit(clientConfig, sessionOpts, messages, result, model) {
 }
 
 // src/telemetry.js
-var VERSION = "0.10.14";
+var VERSION = "0.10.15";
 var TELEMETRY_URL = "https://sdk-telemetry.philip-134.workers.dev";
 function machineId() {
   const raw = typeof process !== "undefined" ? `${process.env?.USER || process.env?.USERNAME || "u"}:${process.platform || "x"}:${process.arch || "x"}` : "browser";

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pentatonic-ai/ai-agent-sdk",
-  "version": "0.10.14",
+  "version": "0.10.15",
   "description": "TES SDK — LLM observability and lifecycle tracking via Pentatonic Thing Event System. Track token usage, tool calls, and conversations. Manage things through event-sourced lifecycle stages with AI enrichment and vector search.",
   "type": "module",
   "main": "./dist/index.cjs",

package/packages/memory-engine-v2/extractor-async/extraction_schema.py

@@ -35,9 +35,26 @@ from typing import Any
 # Allowed-value enums. Moved here from worker.py (which now imports
 # them) so the schema pins to the SAME constants the KV prompt and
 # downstream normalisation use — change them in one place only.
+#
+# 2026-06-16 — ONTOLOGY ALIGNMENT (entity-ontology-the-spine.md). This enum is
+# specifically the set of types the LLM extracts FROM PROSE as named entities.
+# Removed the NLP byproducts that polluted ~28% of the graph and are not
+# business entities:
+#   - `place`, `date` → ATTRIBUTES of real entities (a meeting's location/time),
+#     never standalone entities. The guided enum no longer admits them, so the
+#     model stops minting bare place/date nodes (the info still lands in facts).
+#   - `concept`     → folds into `topic` (the model now emits `topic`).
+# NOT added here (deliberately): meeting / document / thread / task / decision.
+# Those are NOT LLM-prose entities — they are created by structured-event paths
+# (meetings/actions/thread module projections; the sync extractor already emits
+# `document`) or modelled as facts (`decision` category). Forcing the LLM to
+# mint them from prose would create spurious nodes. They join the ontology via
+# their own paths, not this enum.
+# Forward-only: existing place/date/concept rows are untouched and demoted at
+# READ time by the ontology ENGINE_TYPE_MAP (concept→topic, place/date→attribute,
+# other→unresolved). No re-distill required for this change to take effect.
 ALLOWED_ENT_TYPES = {
-    "person", "org", "product", "place", "project",
-    "concept", "topic", "date", "other",
+    "person", "org", "product", "project", "topic", "other",
 }
 ALLOWED_FCT_CATEGORIES = {
     "decision", "commitment", "state", "mention",

package/packages/memory-engine-v2/extractor-async/test_guided_json_parser.py

@@ -61,6 +61,20 @@ def test_schema_enums_pin_to_shared_constants() -> None:
     assert fct_enum == sorted(fct_enum)
 
 
+def test_entity_type_enum_is_ontology_aligned() -> None:
+    """Ontology alignment (entity-ontology-the-spine.md): the LLM-extracted
+    entity types are the genuine named-entity work types — NOT the NLP
+    byproducts. Pins the decision so a future edit can't silently re-admit
+    place/date/concept (which polluted ~28% of the graph). meeting/document/
+    thread/task/decision are deliberately NOT here — they come from
+    structured-event paths / are facts, not LLM prose."""
+    assert xs.ALLOWED_ENT_TYPES == {
+        "person", "org", "product", "project", "topic", "other",
+    }
+    for byproduct in ("place", "date", "concept"):
+        assert byproduct not in xs.ALLOWED_ENT_TYPES
+
+
 def test_schema_caps_mirror_prompt_hard_caps() -> None:
     """8 ENT / 6 FCT / 6 REL per event, statement <= 140 — what
     BATCH_SYSTEM_PROMPT requests, the schema enforces."""

package/packages/memory-engine-v2/extractor-async/worker.py

@@ -222,7 +222,10 @@ RULES:
 matching the input index). NEVER skip an event — if an event has \
 nothing to extract, emit ONLY the header.
 - ENT lines have 3 or 4 fields: literal `ENT`, type, name, [email].
-  type ∈ {person, org, product, place, project, concept, topic, date, other}
+  type ∈ {person, org, product, project, topic, other}
+  Do NOT emit a bare date or place as an entity — those are attributes of
+  other entities (a meeting's time/location), not entities themselves. An
+  abstract idea or theme is a `topic`. Use `other` only when nothing fits.
   email (OPTIONAL, person only): when the event body or attributes
   show an email address that unambiguously identifies the person,
   append it as the 4th field. This pairs the name+email forms so a
@@ -277,8 +280,10 @@ Each per-event object has:
 RULES:
 - NEVER skip an event — if an event has nothing to extract, emit its \
 object with "index" set and empty arrays.
-- entities: type ∈ {person, org, product, place, project, concept, \
-topic, date, other}.
+- entities: type ∈ {person, org, product, project, topic, other}. \
+Do NOT emit a bare date or place as an entity (those are attributes of other \
+entities, not entities); an abstract idea or theme is a `topic`; use `other` \
+only when nothing else fits.
   email (OPTIONAL, person only): when the event body or attributes
   show an email address that unambiguously identifies the person,
   include it. This pairs the name+email forms so a later event seeing