@pentatonic-ai/ai-agent-sdk 0.10.34 → 0.10.36

package/dist/index.cjs

@@ -878,7 +878,7 @@ function fireAndForgetEmit(clientConfig, sessionOpts, messages, result, model) {
 }
 
 // src/telemetry.js
-var VERSION = "0.10.34";
+var VERSION = "0.10.36";
 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 = "0.10.34";
+var VERSION = "0.10.36";
 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": "0.10.34",
+  "version": "0.10.36",
   "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

@@ -507,6 +507,18 @@ ACTIVE_SYSTEM_PROMPT = (
 # guided-JSON traces never mix in a training export).
 SYSTEM_PROMPT_HASH = hashlib.sha256(ACTIVE_SYSTEM_PROMPT.encode()).hexdigest()[:16]
 
+# Sentinel system_prompt_hash for STUDENT traces. The student runs with NO
+# system prompt, so it has no natural hash — and crucially, every trace CONSUMER
+# keys off the teacher's hash: build_retrain_corpus.py filters
+# `system_prompt_hash = <clean teacher hash>` and redistill.py treats a trace at
+# that hash as "already re-distilled clean". Tagging student traces with a
+# distinct sentinel (never a 16-hex teacher hash) keeps them out of BOTH — so
+# student output can never be mistaken for training gold or a clean re-distill.
+# `llm_model` (= STUDENT_MODEL) also distinguishes them; the sentinel is the
+# belt to that suspenders. These rows exist purely as the student side of the
+# student-vs-teacher agreement corpus.
+STUDENT_TRACE_PROMPT_HASH = f"student:{STUDENT_MODEL}"
+
 
 # --------------------------------------------------------------------
 # Content cleaner — strip HTML/CSS so email + doc styling never reaches
@@ -2296,6 +2308,8 @@ def _insert_trace(
     user_prompt: str,
     raw_response: str,
     llm_chunk_ms: float | None,
+    llm_model: str = LLM_MODEL,
+    system_prompt_hash: str = SYSTEM_PROMPT_HASH,
 ) -> None:
     """Append a (user_prompt, raw_response) pair to distillation_traces.
 
@@ -2303,7 +2317,12 @@ def _insert_trace(
     caller wraps this in a try/except and never lets a trace-insert
     failure poison the upsert path. Skip rows with empty raw_response
     (no signal to train on); the model occasionally emits a header
-    with no body."""
+    with no body.
+
+    Defaults tag the row as the teacher's training gold. The student path
+    passes llm_model=STUDENT_MODEL + system_prompt_hash=STUDENT_TRACE_PROMPT_HASH
+    so its rows form the agreement corpus WITHOUT being pulled into the retrain
+    corpus / redistill (both key off the teacher hash)."""
     if not raw_response.strip():
         return
     with conn.cursor() as cur:
@@ -2316,7 +2335,7 @@ def _insert_trace(
             """,
             (
                 event_id, user_prompt, raw_response,
-                LLM_MODEL, SYSTEM_PROMPT_HASH, llm_chunk_ms,
+                llm_model, system_prompt_hash, llm_chunk_ms,
             ),
         )
 
@@ -2429,6 +2448,27 @@ def claim_next_batch(conn: psycopg.Connection) -> list[dict[str, Any]]:
     LLM call on filtered events. They're cheap UPDATE statements scoped
     to the current pending set."""
     with conn.cursor() as cur:
+        # Sweep: retire retry-budget-exhausted rows stuck in 'pending'. The
+        # claim below filters `attempts < MAX_ATTEMPTS`, so a row that reached
+        # MAX_ATTEMPTS while still 'pending' is UNCLAIMABLE — no worker can ever
+        # process it or transition it to 'failed'. Such rows accumulate (legacy
+        # gotcha #11: claiming used to increment attempts; a burst of
+        # deploy-reclaims could push a row to MAX in 'pending') and then pin the
+        # distiller autoscaler awake forever via its flush-on-age trigger (it
+        # counts pending by age) — an idle GPU billed 24/7 for events it can't
+        # touch (62 such rows, oldest 26d, found 2026-07-14). Fail them here so
+        # 'pending' only ever holds claimable work. Terminal by definition:
+        # they already spent the full retry budget.
+        cur.execute(
+            """
+            UPDATE distillation_queue SET
+              status = 'failed',
+              completed_at = NOW(),
+              last_error = COALESCE(last_error, '') || ' | swept: attempts_exhausted_in_pending'
+            WHERE status = 'pending' AND attempts >= %s
+            """,
+            (MAX_ATTEMPTS,),
+        )
         # Pre-filter: agent's-own-output events (by source_kind). Covers any
         # agent producer without enumerating its source label.
         if SKIP_AGENT_SOURCE_KIND:
@@ -2794,6 +2834,30 @@ async def _process_cascade(
         if escalate:
             escalate_items.append(item)
             reason_by_qid[item["id"]] = reason or "escalated"
+            # Capture the student's (discarded) output on escalated events —
+            # the ONLY place the teacher ALSO runs, so the ONLY place a
+            # student-vs-teacher diff is possible. This is the agreement corpus
+            # that makes student quality measurable (esp. the random_sample
+            # subset = an unbiased paired sample of student-passing traffic, and
+            # the high_value_class lane the student now carries). Tagged student
+            # (model + sentinel hash) so no consumer mistakes it for teacher
+            # gold. sresult is None only on student_parse_fail (nothing to
+            # trace). Best-effort — never blocks escalation.
+            if DISTILL_TRACE_ENABLED and event is not None and sresult is not None:
+                try:
+                    _insert_trace(
+                        conn,
+                        event_id=item["event_id"],  # same key the teacher trace uses → rows pair
+                        user_prompt=build_event_block(0, event),
+                        raw_response=sresult.get("raw_slice", ""),
+                        llm_chunk_ms=None,
+                        llm_model=STUDENT_MODEL,
+                        system_prompt_hash=STUDENT_TRACE_PROMPT_HASH,
+                    )
+                except Exception as trace_exc:
+                    log.warning(
+                        f"student trace insert failed event_id={event.get('id')}: {trace_exc}"
+                    )
         else:
             _apply_extraction(
                 conn, item=item, event=event, result=sresult, llm_ms=0.0,