@miller-tech/uap 1.20.45 → 1.20.47

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@miller-tech/uap",
-  "version": "1.20.45",
+  "version": "1.20.47",
   "description": "Autonomous AI agent memory system with CLAUDE.md protocol enforcement",
   "type": "module",
   "main": "dist/index.js",

package/tools/agents/scripts/anthropic_proxy.py

@@ -812,7 +812,12 @@ class SessionMonitor:
         turns_str = f"~{turns} turns remaining" if turns is not None else "unknown"
 
         if warning == "CRITICAL":
-            logger.error(
+            # WARNING, not ERROR: critical context utilization is a *handled*
+            # condition — the proxy force-prunes and the session continues.
+            # Logging it at ERROR floods the error stream (100+/2h for a few
+            # context-saturated agentic sessions) and drowns genuine failures.
+            # CONTEXT HIGH below is already WARNING; this keeps parity.
+            logger.warning(
                 "CONTEXT CRITICAL: %d/%d tokens (%.1f%%), %s, pruned=%d, overflows=%d",
                 self.last_input_tokens,
                 self.context_window,
@@ -1523,6 +1528,14 @@ PROXY_SLOT_SAVE_DIR = os.environ.get(
 PROXY_SLOT_CACHE_MAX_FILES = int(os.environ.get("PROXY_SLOT_CACHE_MAX_FILES", "12"))
 # llama-server slot id — always 0 under --parallel 1.
 PROXY_SLOT_ID = int(os.environ.get("PROXY_SLOT_ID", "0"))
+# HTTP timeouts for the /slots save|restore calls. A large session's KV
+# state (131k ctx) is ~1 GiB; serializing it to / loading it from disk on
+# a slower model (e.g. Qwen3.6-35B-A3B MoE) can exceed the original
+# hardcoded 60s/120s, surfacing as `SLOT SAVE/RESTORE error` with an empty
+# httpx-timeout exception. Restore is given more headroom than save since
+# it also waits on the disk read + KV reload.
+PROXY_SLOT_SAVE_TIMEOUT = float(os.environ.get("PROXY_SLOT_SAVE_TIMEOUT", "180"))
+PROXY_SLOT_RESTORE_TIMEOUT = float(os.environ.get("PROXY_SLOT_RESTORE_TIMEOUT", "300"))
 
 # Module state. Mutated only inside the upstream_semaphore-held section
 # (_post_with_retry), so no extra lock is needed.
@@ -1556,7 +1569,9 @@ async def _save_slot(client: httpx.AsyncClient, session_id: str) -> bool:
     fn = _slot_filename(session_id)
     url = f"{_slot_endpoint_base()}/slots/{PROXY_SLOT_ID}?action=save"
     try:
-        resp = await client.post(url, json={"filename": fn}, timeout=60.0)
+        resp = await client.post(
+            url, json={"filename": fn}, timeout=PROXY_SLOT_SAVE_TIMEOUT
+        )
         if resp.status_code == 200:
             logger.info("SLOT SAVE: session=%s -> %s", session_id, fn)
             return True
@@ -1565,7 +1580,12 @@ async def _save_slot(client: httpx.AsyncClient, session_id: str) -> bool:
             session_id, resp.status_code, resp.text[:200],
         )
     except Exception as exc:
-        logger.warning("SLOT SAVE error: session=%s %s", session_id, exc)
+        # Include the exception TYPE — httpx timeout exceptions stringify
+        # to "" and an empty message log line is undiagnosable.
+        logger.warning(
+            "SLOT SAVE error: session=%s %s: %s",
+            session_id, type(exc).__name__, exc,
+        )
     return False
 
 
@@ -1581,7 +1601,9 @@ async def _restore_slot(client: httpx.AsyncClient, session_id: str) -> bool:
         return False
     url = f"{_slot_endpoint_base()}/slots/{PROXY_SLOT_ID}?action=restore"
     try:
-        resp = await client.post(url, json={"filename": fn}, timeout=120.0)
+        resp = await client.post(
+            url, json={"filename": fn}, timeout=PROXY_SLOT_RESTORE_TIMEOUT
+        )
         if resp.status_code == 200:
             logger.info("SLOT RESTORE: session=%s <- %s", session_id, fn)
             return True
@@ -1590,7 +1612,12 @@ async def _restore_slot(client: httpx.AsyncClient, session_id: str) -> bool:
             session_id, resp.status_code, resp.text[:200],
         )
     except Exception as exc:
-        logger.warning("SLOT RESTORE error: session=%s %s", session_id, exc)
+        # Include the exception TYPE — httpx timeout exceptions stringify
+        # to "" and an empty message log line is undiagnosable.
+        logger.warning(
+            "SLOT RESTORE error: session=%s %s: %s",
+            session_id, type(exc).__name__, exc,
+        )
     return False
 
 

package/tools/agents/tests/test_anthropic_proxy_streaming.py

@@ -5212,14 +5212,14 @@ class TestThinkingBlockExtraction(unittest.TestCase):
 
 
 class _SlotFakeClient:
-    """Records POST calls for slot save/restore tests."""
+    """Records POST calls (incl. the timeout kwarg) for slot tests."""
 
     def __init__(self, status_code=200):
         self.calls = []
         self._status = status_code
 
     async def post(self, url, json=None, timeout=None):  # noqa: A002
-        self.calls.append({"url": url, "json": json})
+        self.calls.append({"url": url, "json": json, "timeout": timeout})
         return _FakeResponse({}, status_code=self._status)
 
 
@@ -5240,6 +5240,8 @@ class TestSlotSaveRestore(unittest.TestCase):
                 "PROXY_SLOT_SAVE_RESTORE",
                 "PROXY_SLOT_CACHE_MAX_FILES",
                 "PROXY_SLOT_ID",
+                "PROXY_SLOT_SAVE_TIMEOUT",
+                "PROXY_SLOT_RESTORE_TIMEOUT",
                 "_slot_owner_session",
             )
         }
@@ -5326,6 +5328,29 @@ class TestSlotSaveRestore(unittest.TestCase):
         self.assertIn("fp:aaaa", proxy._slot_lru)
         self.assertIn("fp:bbbb", proxy._slot_lru)
 
+    def test_slot_timeout_defaults_are_sane(self):
+        """Slot save/restore HTTP timeouts must be configurable and large
+        enough for a slow model's ~1 GiB KV serialization. Restore gets more
+        headroom than save (it also waits on disk read + KV reload)."""
+        self.assertIsInstance(proxy.PROXY_SLOT_SAVE_TIMEOUT, float)
+        self.assertIsInstance(proxy.PROXY_SLOT_RESTORE_TIMEOUT, float)
+        # Both above the original hardcoded 60s/120s that were too tight
+        # for the 35B-A3B (surfaced as empty-message SLOT SAVE/RESTORE errors).
+        self.assertGreaterEqual(proxy.PROXY_SLOT_SAVE_TIMEOUT, 120.0)
+        self.assertGreaterEqual(proxy.PROXY_SLOT_RESTORE_TIMEOUT, 180.0)
+        self.assertGreaterEqual(
+            proxy.PROXY_SLOT_RESTORE_TIMEOUT, proxy.PROXY_SLOT_SAVE_TIMEOUT
+        )
+
+    def test_save_slot_passes_configured_timeout(self):
+        """_save_slot must hand its httpx POST the configured
+        PROXY_SLOT_SAVE_TIMEOUT, not a hardcoded value."""
+        proxy.PROXY_SLOT_SAVE_TIMEOUT = 222.0
+        client = _SlotFakeClient(status_code=200)
+        asyncio.run(proxy._save_slot(client, "fp:timeoutcheck"))
+        self.assertEqual(len(client.calls), 1)
+        self.assertEqual(client.calls[0]["timeout"], 222.0)
+
     def test_evict_slot_files_respects_lru_cap_and_owner(self):
         """LRU eviction removes oldest entries beyond the cap but never the
         session currently owning the slot."""