fix(agents): remove default response cap

deploy/launchd/ai.kakeya.grpc-runtime-prefill.plist

@@ -38,7 +38,7 @@
     <string>--cache-link-mbps</string><string>10000</string>
     <string>--cache-default-rtt-ms</string><string>0.55</string>
     <string>--remote-prefill-min-tokens</string><string>0</string>
-    <string>--prefill-worker-timeout-s</string><string>300</string>
+    <string>--prefill-worker-timeout-s</string><string>900</string>
     <string>--prefill-policy</string><string>remote-required</string>
     <string>--network-http-host</string><string>127.0.0.1</string>
     <string>--network-http-port</string><string>8090</string>

docs/ops/distributed-prefill-kv-network.md

@@ -335,10 +335,11 @@ streams `generator>` and `critic>` tokens as they arrive, followed by KV hit
 rate, decode tok/s, generation latency and E2E tok/s. `/quit` exits; services
 remain running. Reports remain redacted and appear in the Benchmarks tab.
 Generation uses 64-token streaming chunks and automatically continues to the
-model's EOS. `--max-response-tokens` defaults to 512 as an explicit safety cap;
-reaching it marks the stage incomplete instead of presenting a truncated answer
-as successful. The Critic receives the Generator completion status and must not
-penalize an honest statement that an open problem has no accepted proof.
+model's EOS. `--max-response-tokens` defaults to `0` (no client-side cap);
+operators may explicitly set a positive cap, which marks the stage incomplete
+instead of presenting a truncated answer as successful. The Critic receives the
+Generator completion status and must not penalize an honest statement that an
+open problem has no accepted proof.
 The REPL ignores external `SIGTERM`; its shell supervisor restarts signal-based
 exits. Only `/quit`, `/exit`, or EOF is treated as approval to stop.
 

scripts/agent_gan_inference_demo.py

@@ -51,11 +51,19 @@ def _infer(
         append_done = time.perf_counter()
         first_at = None
         generated = []
-        response_limit = int(max_response_tokens or output_tokens)
+        response_limit = (
+            int(output_tokens)
+            if max_response_tokens is None
+            else int(max_response_tokens) or None
+        )
         stop_reason = "unknown"
-        while len(generated) < response_limit:
+        while response_limit is None or len(generated) < response_limit:
             before_count = len(generated)
-            chunk = min(output_tokens, response_limit - len(generated))
+            chunk = (
+                output_tokens
+                if response_limit is None
+                else min(output_tokens, response_limit - len(generated))
+            )
             for token in s.generate(max_tokens=chunk):
                 generated.append(int(token))
                 if on_token is not None:
@@ -73,7 +81,11 @@ def _infer(
             if len(generated) == before_count:
                 stop_reason = "no_progress"
                 break
-        if len(generated) >= response_limit and stop_reason == "max_tokens":
+        if (
+            response_limit is not None
+            and len(generated) >= response_limit
+            and stop_reason == "max_tokens"
+        ):
             stop_reason = "client_safety_limit"
         done = time.perf_counter()
     after = get_stats()
@@ -106,7 +118,12 @@ def main() -> int:
              "reliably; larger values require more worker memory or timeout.",
     )
     parser.add_argument("--output-tokens", type=int, default=64)
-    parser.add_argument("--max-response-tokens", type=int, default=512)
+    parser.add_argument(
+        "--max-response-tokens",
+        type=int,
+        default=0,
+        help="Optional client response cap; 0 means generate until model EOS.",
+    )
     parser.add_argument("--report", default="/tmp/kakeya-agent-gan-demo.json")
     parser.add_argument("--skip-ensure", action="store_true")
     args = parser.parse_args()

scripts/agent_gan_repl.py

@@ -76,7 +76,12 @@ def main() -> int:
     parser.add_argument("--api-key-file", default="~/.kakeya/network_api_key")
     parser.add_argument("--tokenizer-id", required=True)
     parser.add_argument("--output-tokens", type=int, default=64)
-    parser.add_argument("--max-response-tokens", type=int, default=512)
+    parser.add_argument(
+        "--max-response-tokens",
+        type=int,
+        default=0,
+        help="Optional client response cap; 0 means generate until model EOS.",
+    )
     parser.add_argument("--skip-ensure", action="store_true")
     args = parser.parse_args()
 

tests/inference_engine/bridge/test_agent_gan_demo.py

@@ -67,7 +67,7 @@ def test_infer_continues_chunks_until_eos():
         2,
         lambda: {},
         on_token=lambda values: streamed.append(list(values)),
-        max_response_tokens=10,
+        max_response_tokens=0,
     )
     assert tokens == [1, 2, 3]
     assert streamed == [[1], [1, 2], [1, 2, 3]]

tests/inference_engine/bridge/test_prefill_worker_launchd.py

@@ -48,6 +48,10 @@ def test_two_mac_deployment_uses_allens_as_prefill_only():
         "<string>--prefill-policy</string><string>remote-required</string>"
         in plist
     )
+    assert (
+        "<string>--prefill-worker-timeout-s</string><string>900</string>"
+        in plist
+    )
     assert (
         "<string>--cache-tenant-id</string><string>private-fleet</string>"
         in plist