@meridiona/meridian-darwin-arm64 1.66.0 → 1.66.1

package/VERSION

@@ -1 +1 @@
-1.66.0
+1.66.1

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@meridiona/meridian-darwin-arm64",
-  "version": "1.66.0",
+  "version": "1.66.1",
   "description": "Prebuilt Meridian app for macOS arm64 (daemon binary + dashboard + Python services). Installed via @meridiona/meridian.",
   "homepage": "https://github.com/Meridiona/meridian",
   "repository": {

package/services/agents/_state.py

@@ -13,7 +13,7 @@ app_state: dict[str, Any] = {}
 # `received`/`total` are AGGREGATE byte counts summed across every pipeline model
 # (llm + reranker + embedder), so the wizard's progress bar has a single honest
 # denominator. The wire contract the Rust tray decodes is exactly
-# state/received/total/error (tray/src-tauri/src/mlx_server.rs::PrefetchStatus);
+# state/received/total/error/speed (tray/src-tauri/src/mlx_server.rs::PrefetchStatus);
 # `models` is an additive per-model breakdown the tray ignores.
 prefetch_state: dict[str, Any] = {
     "state": "idle",

package/services/agents/routes/prefetch.py

@@ -33,10 +33,9 @@ router = APIRouter()
 # Speed is derived from the growth of on-disk `received` between /prefetch_status
 # polls — NOT from a tqdm hook. hf_xet (the Xet accelerator these models use) does
 # its download in Rust and bypasses the classic tqdm bar, so a tqdm_class never
-# fires; the byte-delta approach works for any backend. These hold the last
-# (received, monotonic-time) sample to diff against.
-_last_recv = 0
-_last_recv_ts = 0.0
+# fires; the byte-delta approach works for any backend. Stored as a dict so both
+# functions can mutate it in-place without `global` declarations.
+_speed_state: dict[str, float | int] = {"recv": 0, "ts": 0.0}
 
 
 def _hf_cache_dir_for(model_id: str) -> Path:
@@ -140,7 +139,7 @@ def _run_prefetch(specs: list[model_registry.ModelSpec]) -> None:
                 prefetch_state["error"] = str(exc)
                 prefetch_state["speed"] = 0.0
             root.set_status(trace.Status(trace.StatusCode.ERROR, str(exc)))
-            log.error("server: model prefetch failed", extra={"error": str(exc)})
+            log.exception("server: model prefetch failed", extra={"error": str(exc)})
 
 
 @router.post("/prefetch_model")
@@ -160,13 +159,13 @@ async def prefetch_model() -> dict:
 
     from fastapi.concurrency import run_in_threadpool
 
-    global _last_recv, _last_recv_ts
     specs = list(model_registry.ALL_SPECS)
 
     with prefetch_lock:
         if prefetch_state["state"] in ("downloading", "done"):
             return dict(prefetch_state)  # idempotent — no duplicate downloads
-        _last_recv, _last_recv_ts = 0, 0.0
+        _speed_state["recv"] = 0
+        _speed_state["ts"] = 0.0
         prefetch_state.update(
             state="downloading",
             received=0,
@@ -189,7 +188,7 @@ async def prefetch_model() -> dict:
         return_exceptions=True,
     )
     totals: list[int] = []
-    for spec, r in zip(specs, probe_results):
+    for spec, r in zip(specs, probe_results, strict=True):
         if isinstance(r, int):
             totals.append(r)
         else:
@@ -223,7 +222,6 @@ async def prefetch_status() -> dict:
     """
     from fastapi.concurrency import run_in_threadpool
 
-    global _last_recv, _last_recv_ts
     with prefetch_lock:
         st = dict(prefetch_state)
         models = list(st.get("models", []))
@@ -252,12 +250,13 @@ async def prefetch_status() -> dict:
     # EMA-smoothed; decays to 0 when bytes stop flowing (delta 0) or the run ends.
     now = time.monotonic()
     with prefetch_lock:
-        if st["state"] == "downloading" and _last_recv_ts and now > _last_recv_ts:
-            inst = max(0, st["received"] - _last_recv) / (now - _last_recv_ts)
+        if st["state"] == "downloading" and _speed_state["ts"] and now > _speed_state["ts"]:
+            inst = max(0, st["received"] - _speed_state["recv"]) / (now - _speed_state["ts"])
             prev = prefetch_state.get("speed", 0.0) or 0.0
             prefetch_state["speed"] = inst if prev == 0.0 else prev * 0.5 + inst * 0.5
         elif st["state"] != "downloading":
             prefetch_state["speed"] = 0.0
-        _last_recv, _last_recv_ts = st["received"], now
+        _speed_state["recv"] = st["received"]
+        _speed_state["ts"] = now
         st["speed"] = prefetch_state["speed"]
     return st

package/services/agents/session_distiller.py

@@ -96,7 +96,7 @@ _ENTITY_RESCUE_CAP = 4
 _embedder = None  # cached (model, tokenizer) tuple
 
 
-def _get_embedder():
+def _get_embedder() -> tuple:
     global _embedder
     if _embedder is None:
         import mlx_embeddings
@@ -135,7 +135,8 @@ def evict_embedder() -> None:
         _embedder = None
         gc.collect()
         try:
-            import mlx.core as mx; mx.clear_cache()
+            import mlx.core as mx
+            mx.clear_cache()
         except Exception:  # noqa: BLE001 — cache flush is best-effort; non-fatal if mlx absent
             pass
         log.info("session_distiller: embedding model evicted from memory")

package/services/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "meridian-agents"
-version = "1.66.0"
+version = "1.66.1"
 description = "Meridian agents — MLX classifier server and Jira worklog synthesis for meridian.db"
 requires-python = ">=3.11"
 authors = [{ name = "Meridiona" }]