@pentatonic-ai/ai-agent-sdk 0.10.27 → 0.10.28

package/dist/index.cjs

@@ -878,7 +878,7 @@ function fireAndForgetEmit(clientConfig, sessionOpts, messages, result, model) {
 }
 
 // src/telemetry.js
-var VERSION = "0.10.27";
+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 = "0.10.27";
+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": "0.10.27",
+  "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/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: