@pentatonic-ai/ai-agent-sdk 0.10.26 → 0.10.28

package/dist/index.cjs

@@ -878,7 +878,7 @@ function fireAndForgetEmit(clientConfig, sessionOpts, messages, result, model) {
 }
 
 // src/telemetry.js
-var VERSION = "0.10.26";
+var VERSION = "0.10.28";
 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 = "0.10.26";
+var VERSION = "0.10.28";
 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": "0.10.26",
+  "version": "0.10.28",
   "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/compat/server.py

@@ -460,7 +460,7 @@ class StoreRequest(BaseModel):
 
 class StoreBatchRequest(BaseModel):
     records: list[dict[str, Any]] = Field(default_factory=list)
-    arena: str | None = "general"
+    arena: str | None = None  # no "general" default — fail closed if absent
     # v1's optional pre-computed embeddings — passed through but we
     # re-embed regardless. The shared-embed optimisation lives at the
     # SDK level now (PR #58 retry-with-jitter); compat trusts the
@@ -611,7 +611,14 @@ async def _extract(arena: str, clientId: str, userId: str | None,
     return r.json()["event_id"]
 
 
-def _arena_of(meta: dict[str, Any] | None, fallback: str = "general") -> str:
+def _arena_of(meta: dict[str, Any] | None, fallback: str | None = None) -> str | None:
+    # No "general" fallback. An arena-less write used to silently land in a
+    # shared "general" bucket that belongs to no tenant — a cross-tenant
+    # hazard (and the one thing that could co-mingle entities across tenants,
+    # since entity ids embed the arena). Callers MUST supply an arena; the
+    # /store + /store-batch handlers reject when this returns None. The SDK
+    # (engineStore/composeArena) always stamps one, so only a raw direct
+    # caller can hit the rejection — which is the point (fail closed).
     if not meta:
         return fallback
     if isinstance(meta.get("arena"), str) and meta["arena"]:
@@ -690,6 +697,10 @@ async def store(req: StoreRequest):
     { id, content, layerId, engine }."""
     meta = req.metadata or {}
     arena = _arena_of(meta)
+    if not arena:
+        # Fail closed: no arena = no tenant. Mirrors /search. (Was a silent
+        # "general" bucket — a cross-tenant hazard.)
+        raise HTTPException(400, "arena required")
     clientId = meta.get("clientId") or arena.split(":")[0]
     userId = meta.get("user_id") or (arena.split(":", 1)[1] if ":" in arena else None)
     source_kind = _source_kind_of(meta)
@@ -778,7 +789,7 @@ async def store_batch(req: StoreBatchRequest):
     if not req.records:
         return {"status": "ok", "inserted": 0, "ids": [], "engine": {}}
 
-    arena_default = req.arena or "general"
+    arena_default = req.arena  # may be None; per-record arena can still supply it
     texts = [r["content"] for r in req.records]
     embeddings = await _embed_batch(texts)
     if len(embeddings) != len(texts):
@@ -809,6 +820,10 @@ async def store_batch(req: StoreBatchRequest):
     for r in req.records:
         meta = r.get("metadata") or {}
         arena = _arena_of(meta, fallback=arena_default)
+        if not arena:
+            # Fail closed: every record needs an arena (batch-level or its own).
+            # No "general" bucket — an arena-less write belongs to no tenant.
+            raise HTTPException(400, "arena required (batch-level or per-record)")
         clientId = meta.get("clientId") or arena.split(":")[0]
         userId = meta.get("user_id") or (arena.split(":", 1)[1] if ":" in arena else None)
         source_kind = _source_kind_of(meta)

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

@@ -121,6 +121,16 @@ STUDENT_API_KEY = os.environ.get("STUDENT_API_KEY", "")
 STUDENT_MODEL = os.environ.get("STUDENT_MODEL", "nuextract-2.0-4b-ft")
 STUDENT_TIMEOUT_SEC = float(os.environ.get("STUDENT_TIMEOUT_SEC", "60"))
 STUDENT_MAX_TOKENS = int(os.environ.get("STUDENT_MAX_TOKENS", "768"))
+# Student-path concurrency — SEPARATE from CONCURRENT_LLM_CALLS. That knob is
+# tuned for the teacher (N concurrent 27B *batch* calls of EVENTS_PER_LLM_CALL
+# events each); the cascade reused it to bound student calls, throttling a 4B
+# model — which batches 30–60 single-event calls at <5% KV cache on an L4 — to
+# just 6/worker. Under a bursty backlog that left the student GPU-pinned but
+# starved (Running:1, Waiting:0, KV:0.1%; 25k-backlog incident 2026-07-08). Each
+# student call is ONE event (tiny KV), so it parallelises far higher. Tune per
+# the box's KV-cache headroom. Do NOT raise CONCURRENT_LLM_CALLS to fix student
+# throughput — that also raises teacher 27B concurrency (OOM risk).
+STUDENT_CONCURRENT_CALLS = int(os.environ.get("STUDENT_CONCURRENT_CALLS", "48"))
 # Fraction of student-passing events ALSO sent to the teacher — NOT a quality
 # lever, this is the monitoring + active-learning sample (drift detection and
 # fresh teacher-gold on live traffic). Default conservative-ish 5%.
@@ -2649,7 +2659,10 @@ async def _process_cascade(
     gates: a pass writes the student's extraction (producer='student'); a fail/
     sample escalates to the teacher. Disjoint — the student writes XOR the event
     is escalated, so there is nothing to supersede."""
-    sem = asyncio.Semaphore(CONCURRENT_LLM_CALLS)
+    # STUDENT_CONCURRENT_CALLS, not CONCURRENT_LLM_CALLS: the student is a cheap
+    # single-event 4B call that batches far higher than the teacher's 27B
+    # batch-calls (see the config note). Reusing the teacher knob starved it.
+    sem = asyncio.Semaphore(STUDENT_CONCURRENT_CALLS)
 
     async def _student(item):
         async with sem: