fix: redact Realtime audio format diagnostics (#3992)

Author: seratch

src/agents/realtime/audio_formats.py

@@ -10,6 +10,7 @@
     RealtimeAudioFormats,
 )
 
+from .. import _debug
 from ..logger import logger
 
 
@@ -25,6 +26,8 @@ def to_realtime_audio_format(
                 format = AudioPCMU(type="audio/pcmu")
             elif input_audio_format in ["g711_alaw", "audio/pcma", "pcma"]:
                 format = AudioPCMA(type="audio/pcma")
+            elif _debug.DONT_LOG_MODEL_DATA:
+                logger.debug("Unknown input audio format")
             else:
                 logger.debug("Unknown input_audio_format: %s", input_audio_format)
         elif isinstance(input_audio_format, Mapping):
@@ -37,15 +40,21 @@ def to_realtime_audio_format(
                 elif rate is None:
                     pcm_rate = 24000
                 else:
-                    logger.debug(
-                        "Unknown pcm rate in input_audio_format mapping: %s", input_audio_format
-                    )
+                    if _debug.DONT_LOG_MODEL_DATA:
+                        logger.debug("Unknown PCM rate in input audio format mapping")
+                    else:
+                        logger.debug(
+                            "Unknown pcm rate in input_audio_format mapping: %s",
+                            input_audio_format,
+                        )
                     pcm_rate = 24000
                 format = AudioPCM(type="audio/pcm", rate=pcm_rate)
             elif fmt_type == "audio/pcmu":
                 format = AudioPCMU(type="audio/pcmu")
             elif fmt_type == "audio/pcma":
                 format = AudioPCMA(type="audio/pcma")
+            elif _debug.DONT_LOG_MODEL_DATA:
+                logger.debug("Unknown input audio format mapping")
             else:
                 logger.debug("Unknown input_audio_format mapping: %s", input_audio_format)
         else:

tests/realtime/test_audio_formats_unit.py

@@ -1,5 +1,10 @@
+import logging
+from typing import Any
+
+import pytest
 from openai.types.realtime.realtime_audio_formats import AudioPCM, AudioPCMA, AudioPCMU
 
+from agents import _debug
 from agents.realtime.audio_formats import to_realtime_audio_format
 
 
@@ -51,3 +56,96 @@ def test_to_realtime_audio_format_from_mapping():
     assert alaw.type == "audio/pcma"
 
     assert to_realtime_audio_format({"type": "audio/unknown", "rate": 8000}) is None
+
+
+@pytest.mark.parametrize("tool_data_redacted", [False, True])
+@pytest.mark.parametrize(
+    ("input_audio_format", "expected_message", "expected_type"),
+    [
+        ("format-secret", "Unknown input audio format", None),
+        (
+            {"type": "audio/pcm", "rate": "rate-secret"},
+            "Unknown PCM rate in input audio format mapping",
+            AudioPCM,
+        ),
+        (
+            {"type": "format-secret", "nested": "mapping-secret"},
+            "Unknown input audio format mapping",
+            None,
+        ),
+    ],
+)
+def test_to_realtime_audio_format_redacts_unknown_values(
+    monkeypatch: pytest.MonkeyPatch,
+    caplog: pytest.LogCaptureFixture,
+    tool_data_redacted: bool,
+    input_audio_format: Any,
+    expected_message: str,
+    expected_type: type[AudioPCM] | None,
+) -> None:
+    monkeypatch.setattr(_debug, "DONT_LOG_MODEL_DATA", True)
+    monkeypatch.setattr(_debug, "DONT_LOG_TOOL_DATA", tool_data_redacted)
+    caplog.set_level(logging.DEBUG, logger="openai.agents")
+
+    result = to_realtime_audio_format(input_audio_format)
+
+    if expected_type is None:
+        assert result is None
+    else:
+        assert isinstance(result, expected_type)
+    record = caplog.records[-1]
+    assert record.msg == expected_message
+    assert record.args == ()
+    assert record.exc_info is None
+    assert record.exc_text is None
+    assert all(value is not input_audio_format for value in record.__dict__.values())
+    assert logging.Formatter().format(record) == expected_message
+
+
+@pytest.mark.parametrize("tool_data_redacted", [False, True])
+def test_to_realtime_audio_format_preserves_diagnostic_mapping(
+    monkeypatch: pytest.MonkeyPatch,
+    caplog: pytest.LogCaptureFixture,
+    tool_data_redacted: bool,
+) -> None:
+    input_audio_format = {"type": "format-secret", "nested": "mapping-secret"}
+    monkeypatch.setattr(_debug, "DONT_LOG_MODEL_DATA", False)
+    monkeypatch.setattr(_debug, "DONT_LOG_TOOL_DATA", tool_data_redacted)
+    caplog.set_level(logging.DEBUG, logger="openai.agents")
+
+    assert to_realtime_audio_format(input_audio_format) is None
+
+    record = caplog.records[-1]
+    assert record.msg == "Unknown input_audio_format mapping: %s"
+    assert record.args is input_audio_format
+    assert record.exc_info is None
+    assert record.exc_text is None
+    assert "format-secret" in logging.Formatter().format(record)
+    assert "mapping-secret" in logging.Formatter().format(record)
+
+
+def test_to_realtime_audio_format_redaction_does_not_render_hostile_mapping(
+    monkeypatch: pytest.MonkeyPatch,
+    caplog: pytest.LogCaptureFixture,
+) -> None:
+    class HostileMapping(dict[str, object]):
+        def __str__(self) -> str:
+            raise AssertionError("redacted logging must not call __str__")
+
+        def __repr__(self) -> str:
+            raise AssertionError("redacted logging must not call __repr__")
+
+    input_audio_format = HostileMapping(type="unknown")
+    monkeypatch.setattr(_debug, "DONT_LOG_MODEL_DATA", True)
+    monkeypatch.setattr(_debug, "DONT_LOG_TOOL_DATA", False)
+    caplog.set_level(logging.DEBUG, logger="openai.agents")
+
+    assert to_realtime_audio_format(input_audio_format) is None
+
+    record = caplog.records[-1]
+    assert record.msg == "Unknown input audio format mapping"
+    assert record.args == ()
+    assert record.exc_info is None
+    assert record.exc_text is None
+    assert all(value is not input_audio_format for value in record.__dict__.values())
+    assert logging.Formatter().format(record) == "Unknown input audio format mapping"