From f2ec56103128daf05bc9f148d7cbf53a72b3be40 Mon Sep 17 00:00:00 2001
From: Calin Teodor <45018530+TeoSlayer@users.noreply.github.com>
Date: Wed, 28 Oct 2026 09:01:44 -0700
Subject: [PATCH 1/2] fix: reject calls without declared tools

---
 src/inference_endpoint/evaluation/bfcl_v4_execution.py |  2 +-
 tests/unit/evaluation/test_bfcl_v4_execution.py        | 10 +++++++---
 2 files changed, 8 insertions(+), 4 deletions(-)

diff --git a/src/inference_endpoint/evaluation/bfcl_v4_execution.py b/src/inference_endpoint/evaluation/bfcl_v4_execution.py
index e6b51293c..d7d57e228 100644
--- a/src/inference_endpoint/evaluation/bfcl_v4_execution.py
+++ b/src/inference_endpoint/evaluation/bfcl_v4_execution.py
@@ -231,7 +231,7 @@ def process_response(
             model_responses_bfcl = []
             for tc in tool_calls:
                 name = tc["name"]
-                if allowed_names and name not in allowed_names:
+                if name not in allowed_names:
                     logger.warning(
                         "Dropping tool call to undeclared function %r for %s turn "
                         "%d (not in the entry's tool list); refusing to execute an "
diff --git a/tests/unit/evaluation/test_bfcl_v4_execution.py b/tests/unit/evaluation/test_bfcl_v4_execution.py
index 86856adc3..b0bbf0941 100644
--- a/tests/unit/evaluation/test_bfcl_v4_execution.py
+++ b/tests/unit/evaluation/test_bfcl_v4_execution.py
@@ -40,7 +40,7 @@ def _state(entry_id="multi_turn_base_0", max_steps=25):
         current_step=0,
         current_turn_results=[],
         messages=[],
-        tools=[],
+        tools=[{"type": "function", "function": {"name": "f"}}],
         force_terminated=False,
         completed=False,
         model_results_per_turn=[],
@@ -112,14 +112,18 @@ def test_process_response_no_tool_calls_advances_turn():
 
 
 @pytest.mark.unit
-def test_process_response_drops_undeclared_tool_name():
+@pytest.mark.parametrize(
+    "tools",
+    [[{"type": "function", "function": {"name": "f"}}], [], [{"type": "function"}]],
+)
+def test_process_response_drops_undeclared_tool_name(tools):
     """A tool call to a name the entry never declared must not reach bfcl-eval's
     executor (guards against injection like __import__('os').system)."""
     executed = []
     bridge = _make_bridge()
     state = _state()
     state.entry.num_turns = 1
-    state.tools = [{"type": "function", "function": {"name": "f"}}]
+    state.tools = tools
     with (
         patch.object(ex, "convert_to_function_call", lambda calls: list(calls)),
         patch.object(ex, "is_empty_execute_response", lambda d: len(d) == 0),

From 2fd9015710b8d3971b896b0f46bd445dff817151 Mon Sep 17 00:00:00 2001
From: Calin Teodor <45018530+TeoSlayer@users.noreply.github.com>
Date: Thu, 5 Nov 2026 09:38:59 -0800
Subject: [PATCH 2/2] test: wait for benchmark startup before wedging metrics

---
 tests/integration/commands/test_sigint.py | 13 +++++++++++--
 1 file changed, 11 insertions(+), 2 deletions(-)

diff --git a/tests/integration/commands/test_sigint.py b/tests/integration/commands/test_sigint.py
index 1197744dc..d6986981e 100644
--- a/tests/integration/commands/test_sigint.py
+++ b/tests/integration/commands/test_sigint.py
@@ -276,6 +276,7 @@ def test_sigint_grace_expiry_abandons_wedged_drain(mock_http_echo_server, tmp_pa
     """
     report_dir = tmp_path / "report"
     config_path = tmp_path / "bench.yaml"
+    log_file = tmp_path / "run.log"
     _write_config(
         report_dir,
         mock_http_echo_server.url,
@@ -286,10 +287,18 @@ def test_sigint_grace_expiry_abandons_wedged_drain(mock_http_echo_server, tmp_pa
     agg_pid: int | None = None
     try:
         with _benchmark_proc(
-            [_cli(), "benchmark", "from-config", "-c", str(config_path)]
+            [_cli(), "benchmark", "from-config", "-c", str(config_path)],
+            log_file,
         ) as proc:
             _wait_services_ready(proc, report_dir)
-            time.sleep(3.0)  # comfortably inside the ~120 s performance phase
+            # Service readiness precedes client startup. Wait for the session
+            # before allowing live metrics ticks to reach the parent.
+            deadline = time.monotonic() + 60.0
+            while "Starting phase:" not in log_file.read_text(errors="replace"):
+                assert proc.poll() is None, log_file.read_text(errors="replace")
+                assert time.monotonic() < deadline, "performance phase never started"
+                time.sleep(0.05)
+            time.sleep(3.0)
 
             agg_pid = _pid_of_child("metrics_aggregator", str(report_dir))
             assert agg_pid is not None, "aggregator child not found"