@melaya/runner 1.0.112 → 1.0.115

package/dist/assistantHost.py

@@ -79,10 +79,29 @@ def _emit_usage(agent, turn_id: str) -> None:
     spans (Overview dashboard); the live in/out counter stays cloud-only for now."""
     try:
         mem = _agent_memory(agent)
-        budget = int(getattr(mem, "max_tokens", 0) or 0) if mem is not None else 0
-        if budget > 0:
-            _emit(turn_id, "usage", turnInTok=0, turnOutTok=0,
-                  ctxUsed=_memory_watermark(agent), ctxBudget=budget)
+        if mem is None:
+            return
+        budget = int(getattr(mem, "max_tokens", 0) or 0)
+        # Real context fullness = current retained-content tokens (NOT the rehydration
+        # watermark, which reads ~0). Falls back to 0 on an old shared bundle without
+        # current_tokens() — the gauge just won't move until the bundle updates.
+        used = int(mem.current_tokens()) if hasattr(mem, "current_tokens") else 0
+        # Live in/out tokens for THIS turn: force-flush the pending trace spans, then
+        # drain the exporter's accumulator (the SAME tokens that land in agents.spans).
+        # The BatchSpanProcessor exports asynchronously, so flush first or the turn's
+        # spans may not have been counted yet. Best-effort; 0 on an old shared bundle.
+        in_tok = out_tok = 0
+        try:
+            from opentelemetry import trace as _ot
+            tp = _ot.get_tracer_provider()
+            if hasattr(tp, "force_flush"):
+                tp.force_flush()
+            from shared.runtime.tracing_exporter import drain_token_totals
+            in_tok, out_tok = drain_token_totals()
+        except Exception:
+            pass
+        if budget > 0 or in_tok or out_tok:
+            _emit(turn_id, "usage", turnInTok=int(in_tok), turnOutTok=int(out_tok), ctxUsed=used, ctxBudget=budget)
     except Exception:
         pass
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@melaya/runner",
-  "version": "1.0.112",
+  "version": "1.0.115",
   "description": "Run Melaya AI pipelines locally with your own LM Studio or Ollama models",
   "license": "UNLICENSED",
   "private": false,