fix(agents): accept exact remote KV hits in gate

Treat a complete allens cache restore as a valid warmup without requiring redundant remote recomputation before the Primary hot-hit inference pass.

Co-authored-by: Cursor <cursoragent@cursor.com>

Author: 2 people

scripts/agent_gan_inference_demo.py

@@ -18,8 +18,11 @@
 
 def _agent_cache_gate(warm_delta: dict, actual_delta: dict) -> bool:
     return (
-        warm_delta["remote_jobs"] >= 1
-        and warm_delta["remote_hits"] >= 1
+        warm_delta["remote_hits"] >= 1
+        and warm_delta["tokens_reused"] >= 1
+        and warm_delta["tokens_computed"] == 0
+        and warm_delta["fallbacks"] == 0
+        and warm_delta["remote_job_failures"] == 0
         and actual_delta["local_hits"] >= 1
         and actual_delta["remote_jobs"] == 0
         and actual_delta["tokens_computed"] == 0

tests/inference_engine/bridge/test_agent_gan_demo.py

@@ -6,16 +6,27 @@
 )
 
 
-def test_agent_gate_requires_remote_warmup_and_primary_hot_inference():
-    warm = {"remote_jobs": 1, "remote_hits": 1}
+def test_agent_gate_accepts_remote_compute_or_exact_remote_cache_hit():
+    warm = {
+        "remote_jobs": 1,
+        "remote_hits": 1,
+        "tokens_reused": 10,
+        "tokens_computed": 0,
+        "fallbacks": 0,
+        "remote_job_failures": 0,
+    }
     actual = {
         "local_hits": 1,
         "remote_jobs": 0,
         "tokens_computed": 0,
         "fallbacks": 0,
     }
     assert _agent_cache_gate(warm, actual)
-    assert not _agent_cache_gate({**warm, "remote_jobs": 0}, actual)
+    assert _agent_cache_gate({**warm, "remote_jobs": 0}, actual)
+    assert not _agent_cache_gate({**warm, "remote_hits": 0}, actual)
+    assert not _agent_cache_gate({**warm, "tokens_reused": 0}, actual)
+    assert not _agent_cache_gate({**warm, "fallbacks": 1}, actual)
+    assert not _agent_cache_gate({**warm, "remote_job_failures": 1}, actual)
     assert not _agent_cache_gate(warm, {**actual, "local_hits": 0})
     assert not _agent_cache_gate(warm, {**actual, "fallbacks": 1})
 

tests/inference_engine/bridge/test_agent_gan_repl.py

@@ -34,6 +34,9 @@ def test_repl_stage_is_redacted_and_passes_cache_gate():
             "remote_jobs": 1,
             "remote_hits": 1,
             "tokens_reused": 10,
+            "tokens_computed": 0,
+            "fallbacks": 0,
+            "remote_job_failures": 0,
         },
     }
     actual = {
@@ -101,7 +104,14 @@ def test_stage_includes_full_context_metrics():
     warm = {
         "prefix_tokens": 10,
         "e2e_s": 1,
-        "delta": {"remote_jobs": 1, "remote_hits": 1, "tokens_reused": 10},
+        "delta": {
+            "remote_jobs": 1,
+            "remote_hits": 1,
+            "tokens_reused": 10,
+            "tokens_computed": 0,
+            "fallbacks": 0,
+            "remote_job_failures": 0,
+        },
     }
     actual = {
         "prefix_tokens": 10,