diff --git a/proxy/README.md b/proxy/README.md
index 179440d..447e5b6 100644
--- a/proxy/README.md
+++ b/proxy/README.md
@@ -151,7 +151,7 @@ POST /v1/instance/unregister
 GET  /v1/instance/list?include_dead=true
 ```
 
-Instances register static information such as `instance_id`, `host`, `port`, `endpoints`, `tags`, `weight`, and `meta`. Heartbeats refresh `last_seen_at` and can optionally report lightweight load fields such as `inflight`, `qps_1m`, and `gpu_util`.
+Instances register static information such as `instance_id`, `host`, `port`, `endpoints`, `tags`, `weight`, and `meta`. Heartbeats refresh `last_seen_at` and can optionally report lightweight non-lifecycle load fields such as `qps_1m` and `gpu_util`; Proxy request lifecycle hooks maintain `inflight`.
 
 ### Instance resource snapshots
 
@@ -160,8 +160,11 @@ After PR #87, demo Instances can report host resource snapshots to the Proxy con
 ```text
 POST /v1/instance/resource_snapshot
 GET  /debug/instance_resources
+GET  /debug/instance_loads
 ```
 
+Use `/debug/instance_loads` to inspect Proxy-maintained per-Instance `inflight` counters, `qps_1m` when present, and queue-depth hints when the queue snapshot provider is available.
+
 The reporting path is:
 
 ```text
diff --git a/proxy/proxy.py b/proxy/proxy.py
index f19e369..02451c1 100644
--- a/proxy/proxy.py
+++ b/proxy/proxy.py
@@ -393,7 +393,12 @@ def _sse_meta_event(task: ProxyTask) -> bytes:
     ).encode("utf-8")
 
 
-async def _wrap_chat_stream_with_meta(task: ProxyTask, queue_mgr: QueueManager) -> AsyncGenerator[bytes, None]:
+async def _wrap_chat_stream_with_meta(
+    task: ProxyTask,
+    queue_mgr: QueueManager,
+    instance_pool: InstancePool,
+    instance_id: str,
+) -> AsyncGenerator[bytes, None]:
     """
     Forward downstream chat SSE, but delay [DONE] and insert one cacheroute_meta event first.
     """
@@ -423,11 +428,14 @@ async def _wrap_chat_stream_with_meta(task: ProxyTask, queue_mgr: QueueManager)
         task.trace["stream_exception_ms"] = int(time.time() * 1000)
         logger.exception("[Proxy] stream wrapper failed rid=%s", task.request_id)
     finally:
-        if pending:
-            yield pending
-            pending = b""
-        yield _sse_meta_event(task)
-        yield b"data: [DONE]\n\n"
+        try:
+            if pending:
+                yield pending
+                pending = b""
+            yield _sse_meta_event(task)
+            yield b"data: [DONE]\n\n"
+        finally:
+            instance_pool.end_request(instance_id)
 
 
 def select_instance(app: FastAPI, req_obj: SchedulerRequest):
@@ -593,8 +601,15 @@ async def proxy_chat_completions(request: FastAPIRequest):
     # ====================================
     # Send into the queue: enqueue -> manager -> forward
     # ====================================
+    instance_pool: InstancePool = proxy.state.instance_pool  # type: ignore
+    request_counted = instance_pool.begin_request(chosen.instance_id)
+    if not request_counted:
+        return JSONResponse(
+            status_code=503,
+            content={"error": "no_instance", "detail": "selected instance is no longer registered"},
+        )
     try:
-        # 1) Wrap the task (note: chosen comes from RR and has instance_id/host/port fields)
+        # 1) Wrap the task (note: chosen comes from strategy and has instance_id/host/port fields)
         task = ProxyTask(
             request_id=getattr(req_obj, "Request_ID", None),
             req_obj=req_obj,
@@ -611,10 +626,12 @@ async def proxy_chat_completions(request: FastAPIRequest):
 
         await queue_mgr.enqueue_prepare(task)
 
-        stream_gen = _wrap_chat_stream_with_meta(task, queue_mgr)
+        stream_gen = _wrap_chat_stream_with_meta(task, queue_mgr, instance_pool, chosen.instance_id)
         return StreamingResponse(stream_gen, media_type="text/event-stream")
 
     except Exception as e:
+        if request_counted:
+            instance_pool.end_request(chosen.instance_id)
         logger.exception("[Proxy] failed to call Worker(chat)")
         return JSONResponse(
             status_code=502,
@@ -762,6 +779,13 @@ async def proxy_completions(request: FastAPIRequest):
     # ==================================
     # Send into the queue: enqueue -> drain -> forward
     # ==================================
+    instance_pool: InstancePool = proxy.state.instance_pool  # type: ignore
+    request_counted = instance_pool.begin_request(chosen.instance_id)
+    if not request_counted:
+        return JSONResponse(
+            status_code=503,
+            content={"error": "no_instance", "detail": "selected instance is no longer registered"},
+        )
     try:
         task = ProxyTask(
             request_id=getattr(req_obj, "Request_ID", None),
@@ -780,9 +804,14 @@ async def proxy_completions(request: FastAPIRequest):
         await queue_mgr.enqueue_prepare(task)
 
         content_bytes = b""
-        async for chunk in queue_mgr.iter_response(task):
-            if chunk:
-                content_bytes += chunk
+        try:
+            async for chunk in queue_mgr.iter_response(task):
+                if chunk:
+                    content_bytes += chunk
+        finally:
+            if request_counted:
+                instance_pool.end_request(chosen.instance_id)
+                request_counted = False
 
         # completions is non-streaming: the worker should return a one-shot JSON response
         if not content_bytes:
@@ -806,6 +835,8 @@ async def proxy_completions(request: FastAPIRequest):
             )
 
     except Exception as e:
+        if request_counted:
+            instance_pool.end_request(chosen.instance_id)
         logger.exception("[Proxy] failed to call Worker(completions)")
         return JSONResponse(
             status_code=502,
diff --git a/proxy/resource/instance_pool.py b/proxy/resource/instance_pool.py
index 25b8921..38fea8b 100644
--- a/proxy/resource/instance_pool.py
+++ b/proxy/resource/instance_pool.py
@@ -90,6 +90,8 @@ def upsert(
                 it.weight = float(weight)
                 if meta:
                     it.meta.update(meta)
+                if it.load.inflight is None:
+                    it.load.inflight = 0
                 it.last_seen_at = now
                 return it
 
@@ -101,6 +103,7 @@ def upsert(
                 tags=tags or [],
                 weight=float(weight),
                 meta=meta or {},
+                load=InstanceLoad(inflight=0),
                 registered_at=now,
                 last_seen_at=now,
             )
@@ -120,9 +123,8 @@ def heartbeat(
             if not it:
                 return False
             it.last_seen_at = now
-            # Only update when provided to avoid overwriting with 0/None
-            if inflight is not None:
-                it.load.inflight = int(inflight)
+            # Inflight is Proxy-maintained via begin_request/end_request.
+            # Keep heartbeat load updates for non-lifecycle signals only.
             if qps_1m is not None:
                 it.load.qps_1m = float(qps_1m)
             if gpu_util is not None:
@@ -145,6 +147,80 @@ def report_resource_snapshot(
             return True
 
 
+    def begin_request(self, instance_id: str) -> bool:
+        """Increment the Proxy-maintained inflight counter for an Instance."""
+        with self._lock:
+            it = self._items.get(instance_id)
+            if not it:
+                return False
+            current = int(it.load.inflight or 0)
+            it.load.inflight = current + 1
+            return True
+
+    def end_request(self, instance_id: str) -> bool:
+        """Decrement the Proxy-maintained inflight counter without going below zero."""
+        with self._lock:
+            it = self._items.get(instance_id)
+            if not it:
+                return False
+            current = int(it.load.inflight or 0)
+            it.load.inflight = max(0, current - 1)
+            return True
+
+    def snapshot_instance_loads(
+        self,
+        queue_depths: Optional[Dict[str, Any]] = None,
+        include_dead: bool = True,
+    ) -> Dict[str, Any]:
+        """Return per-Instance local load counters and optional queue-depth hints."""
+        now = int(time.time())
+        per_instance_queue = {}
+        if queue_depths:
+            raw = queue_depths.get("per_instance", {})
+            if isinstance(raw, dict):
+                per_instance_queue = raw
+
+        with self._lock:
+            items = list(self._items.values())
+
+        instances: List[Dict[str, Any]] = []
+        for it in items:
+            is_alive = (now - int(it.last_seen_at)) <= self._ttl_s
+            if not include_dead and not is_alive:
+                continue
+            queue_item = per_instance_queue.get(it.instance_id, {})
+            inflight = it.load.inflight
+            qps_1m = it.load.qps_1m
+            instances.append({
+                "instance_id": it.instance_id,
+                "is_alive": is_alive,
+                "inflight": inflight,
+                "qps_1m": qps_1m,
+                "prepare_queue_depth": queue_item.get("prepare_queue_depth"),
+                "ready_queue_depth": queue_item.get("ready_queue_depth"),
+                "active_prepare": queue_item.get("active_prepare"),
+                "active_ready": queue_item.get("active_ready"),
+                "least_load_score": {
+                    "primary": "inflight" if inflight is not None else None,
+                    "primary_value": inflight,
+                    "primary_source": "proxy_lifecycle_counter" if inflight is not None else "unavailable",
+                    "secondary": "qps_1m" if qps_1m is not None else None,
+                    "secondary_value": qps_1m,
+                    "secondary_source": "instance_heartbeat" if qps_1m is not None else "unavailable",
+                },
+            })
+
+        return {
+            "ttl_s": self._ttl_s,
+            "generated_at": time.time(),
+            "metric_source": {
+                "inflight": "proxy_lifecycle_counter",
+                "qps_1m": "instance_heartbeat",
+                "queue_depth": "proxy_queue_manager" if queue_depths is not None else "unavailable",
+            },
+            "instances": instances,
+        }
+
     def build_pool_resource_snapshot(
         self,
         proxy_id: str,
@@ -216,7 +292,7 @@ def build_pool_resource_snapshot(
         metric_source = {
             "instances": "proxy_instance_pool_ttl",
             "resource": "instance_resource_snapshot" if reporting else "unavailable",
-            "inflight_total": "instance_heartbeat" if inflight_values else "unavailable",
+            "inflight_total": "proxy_lifecycle_counter" if inflight_values else "unavailable",
             "qps_1m_total": "instance_heartbeat" if qps_values else "unavailable",
             "load_ratio": "derived_from_inflight_capacity" if load_ratio is not None else "unavailable",
             "capacity": "proxy_config",
diff --git a/proxy/resource/p_control_plane.py b/proxy/resource/p_control_plane.py
index 2610914..d2c99ba 100644
--- a/proxy/resource/p_control_plane.py
+++ b/proxy/resource/p_control_plane.py
@@ -339,6 +339,22 @@ async def debug_instance_resources(include_dead: bool = True) -> Dict[str, Any]:
         })
     return {"instances": resources}
 
+
+@_control_plane.get("/debug/instance_loads")
+async def debug_instance_loads(include_dead: bool = True) -> Dict[str, Any]:
+    pool = get_pool()
+    queue_depths = None
+    if _queue_snapshot_provider is not None:
+        try:
+            queue_depths = _queue_snapshot_provider() or {}
+        except Exception:
+            logger.warning(
+                "[ProxyCP] queue snapshot provider failed; instance load queue metrics unavailable",
+                exc_info=True,
+            )
+    snapshot = pool.snapshot_instance_loads(queue_depths=queue_depths, include_dead=include_dead)
+    return {"ok": True, **snapshot}
+
 @_control_plane.post("/v1/topology/report")
 async def report_topology(req: TopologyReportReq) -> Dict[str, Any]:
     merged = 0