diff --git a/autoresearch/prefill/supervisor.py b/autoresearch/prefill/supervisor.py
index b6046c6..0d4fb22 100644
--- a/autoresearch/prefill/supervisor.py
+++ b/autoresearch/prefill/supervisor.py
@@ -569,11 +569,20 @@ def _emit(self) -> None:
             ).get("prefill", {})
         except Exception:
             return
-        total = self._delta(current, "remote_job_tokens_total")
-        computed = self._delta(current, "remote_job_tokens_computed")
+        new_remote_jobs = self._delta(current, "remote_jobs")
+        if new_remote_jobs:
+            # These two fields are gauges for the current job, not cumulative
+            # counters. Subtracting the previous same-length job makes a retry
+            # look permanently stuck at 0/0.
+            total = int(current.get("remote_job_tokens_total", 0))
+            computed = int(current.get("remote_job_tokens_computed", 0))
+        else:
+            total = 0
+            computed = 0
         state = (
             computed,
             total,
+            new_remote_jobs,
             self._delta(current, "remote_hits"),
             self._delta(current, "tokens_reused"),
         )
@@ -583,7 +592,7 @@ def _emit(self) -> None:
         percent = min(100.0, 100.0 * computed / total)
         print(
             f"[autoresearch] Strategy Prefill: {computed}/{total} tokens "
-            f"({percent:.1f}%) · remote_hits={state[2]} reused={state[3]}",
+            f"({percent:.1f}%) · remote_hits={state[3]} reused={state[4]}",
             flush=True,
         )
 
diff --git a/tests/inference_engine/bench/test_autoresearch_supervisor.py b/tests/inference_engine/bench/test_autoresearch_supervisor.py
index 1a7df9b..6d17e0a 100644
--- a/tests/inference_engine/bench/test_autoresearch_supervisor.py
+++ b/tests/inference_engine/bench/test_autoresearch_supervisor.py
@@ -397,23 +397,25 @@ def test_runtime_health_check_is_read_only(monkeypatch):
 def test_strategy_prefill_heartbeat_reports_delta(monkeypatch, capsys):
     heartbeat = StrategyPrefillHeartbeat(interval_s=0.01)
     heartbeat._baseline = {
-        "remote_job_tokens_total": 100,
-        "remote_job_tokens_computed": 100,
+        "remote_jobs": 10,
+        "remote_job_tokens_total": 300,
+        "remote_job_tokens_computed": 300,
         "remote_hits": 2,
         "tokens_reused": 20,
     }
     monkeypatch.setattr(
         "autoresearch.prefill.supervisor._json_request",
         lambda _url: {"prefill": {
+            "remote_jobs": 11,
             "remote_job_tokens_total": 300,
-            "remote_job_tokens_computed": 228,
+            "remote_job_tokens_computed": 128,
             "remote_hits": 3,
             "tokens_reused": 84,
         }},
     )
     heartbeat._emit()
     output = capsys.readouterr().out
-    assert "128/200 tokens (64.0%)" in output
+    assert "128/300 tokens (42.7%)" in output
     assert "remote_hits=1 reused=64" in output