fix(voice): redact TTS instructions from speech spans

src/agents/voice/result.py

@@ -124,7 +124,11 @@ async def _stream_audio(
             input=text if self._voice_pipeline_config.trace_include_sensitive_data else "",
             model_config={
                 "voice": self.tts_settings.voice,
-                "instructions": self.instructions,
+                "instructions": (
+                    self.instructions
+                    if self._voice_pipeline_config.trace_include_sensitive_data
+                    else None
+                ),
                 "speed": self.tts_settings.speed,
             },
             output_format="pcm",

tests/voice/test_pipeline.py

@@ -15,7 +15,7 @@
 
 import agents._debug as _debug
 from agents import trace
-from tests.testing_processor import fetch_events, fetch_span_errors
+from tests.testing_processor import fetch_events, fetch_ordered_spans, fetch_span_errors
 
 try:
     from agents.voice import (
@@ -587,6 +587,46 @@ async def run(self, text: str, settings: TTSModelSettings):
     ]
 
 
+@pytest.mark.asyncio
+@pytest.mark.parametrize("trace_include_sensitive_data", [True, False])
+async def test_speech_span_redacts_tts_instructions(
+    trace_include_sensitive_data: bool,
+) -> None:
+    """TTS instructions are author-written prompt text, so they follow the same gate as input."""
+    instructions = "sensitive-style-instructions"
+    received: list[str] = []
+
+    class RecordingTTS(ZeroPcmTTSModel):
+        async def run(self, text: str, settings: TTSModelSettings) -> AsyncIterator[bytes]:
+            received.append(settings.instructions)
+            yield np.zeros(2, dtype=np.int16).tobytes()
+
+    result = StreamedAudioResult(
+        RecordingTTS(),
+        TTSModelSettings(instructions=instructions),
+        VoicePipelineConfig(trace_include_sensitive_data=trace_include_sensitive_data),
+    )
+    local_queue: asyncio.Queue[VoiceStreamEvent | None] = asyncio.Queue()
+
+    with trace("tts-instructions"):
+        await result._stream_audio("spoken text", local_queue)
+
+    speech_spans = [span for span in fetch_ordered_spans() if span.span_data.type == "speech"]
+    assert len(speech_spans) == 1
+    model_config = cast(dict[str, Any], speech_spans[0].span_data.model_config)
+
+    if trace_include_sensitive_data:
+        assert model_config["instructions"] == instructions
+        assert speech_spans[0].span_data.input == "spoken text"
+    else:
+        assert model_config["instructions"] is None
+        assert speech_spans[0].span_data.input == ""
+
+    # Non-sensitive settings stay visible either way, and the model still gets the real value.
+    assert model_config["speed"] == TTSModelSettings().speed
+    assert received == [instructions]
+
+
 @pytest.mark.asyncio
 async def test_streamed_audio_dispatcher_handles_stream_failure() -> None:
     """A failed _stream_audio task must not leave _dispatch_audio blocked forever."""