diff --git a/agent/tts_provider.py b/agent/tts_provider.py
index c38469a3d2..873947e96d 100644
--- a/agent/tts_provider.py
+++ b/agent/tts_provider.py
@@ -34,7 +34,10 @@ class TTSProvider(CatalogProviderBase):
     and make ``stream(text, format="pcm", ...)`` yield raw int16 little-endian mono
     PCM at that rate. Both are read (with :meth:`is_available`) each time a reply
     resolves its streamer, so they may be properties reflecting live state; a
-    missing rate falls back to per-sentence :meth:`synthesize`.
+    missing rate falls back to per-sentence :meth:`synthesize`. Unlike
+    :meth:`synthesize`, ``stream()`` may run concurrently for consecutive
+    sentences of one reply (the speaker prefetches up to 3), so it must be
+    thread-safe.
     """
 
     streams_pcm: bool = False
diff --git a/tests/tools/test_tts_plugin_dispatch.py b/tests/tools/test_tts_plugin_dispatch.py
index 519825f875..c8936c5d01 100644
--- a/tests/tools/test_tts_plugin_dispatch.py
+++ b/tests/tools/test_tts_plugin_dispatch.py
@@ -268,6 +268,8 @@ def test_opted_in_plugin_streams_pcm_in_the_configured_voice(monkeypatch, stream
     (_PCMPlugin(streams_pcm=False), {}),
     (_PCMPlugin(rate=None), {}),
     (_PCMPlugin(rate=0), {}),
+    (_PCMPlugin(rate=0.5), {}),
+    (_PCMPlugin(rate=float("nan")), {}),
     (_PCMPlugin(available=False), {}),
     (_PCMPlugin(), {"providers": {"fake-pcm": {"type": "command", "command": "say {input_path}"}}}),
 ])
diff --git a/tools/tts_streaming.py b/tools/tts_streaming.py
index cd72c31ecc..9facab3039 100644
--- a/tools/tts_streaming.py
+++ b/tools/tts_streaming.py
@@ -175,10 +175,9 @@ class _PluginPCMStreamer(StreamingTTSProvider):
     """A plugin ``TTSProvider`` that opted into raw PCM (``streams_pcm``; see
     ``tools.tts_tool_plugins._plugin_pcm_streaming_provider``) behind this ABC."""
 
-    def __init__(self, provider: Any, tts_config: Dict):
+    def __init__(self, provider: Any, sample_rate: int, tts_config: Dict):
         super().__init__(tts_config, tts_config.get(provider.name) or {})
-        self._provider = provider
-        self.sample_rate = int(provider.stream_sample_rate)
+        self._provider, self.sample_rate = provider, sample_rate
 
     @staticmethod
     def available() -> bool:
@@ -191,8 +190,8 @@ class _PluginPCMStreamer(StreamingTTSProvider):
 
 
 def _plugin_streamer(name: str, tts_config: Dict) -> Optional[StreamingTTSProvider]:
-    provider = _plugin_pcm_streaming_provider(name, tts_config)
-    return _PluginPCMStreamer(provider, tts_config) if provider is not None else None
+    found = _plugin_pcm_streaming_provider(name, tts_config)
+    return _PluginPCMStreamer(*found, tts_config) if found is not None else None
 
 
 def resolve_streaming_provider(
diff --git a/tools/tts_tool_plugins.py b/tools/tts_tool_plugins.py
index 2d2624dbcf..10c7e14c45 100644
--- a/tools/tts_tool_plugins.py
+++ b/tools/tts_tool_plugins.py
@@ -7,11 +7,13 @@ tool module stays importable without the plugin machinery).
 from __future__ import annotations
 
 import logging
-from typing import Any, Dict, Optional
+import math
+from typing import Any, Dict, Optional, Tuple
 
 from tools.tts_command_provider import (
     BUILTIN_TTS_PROVIDERS, DEFAULT_COMMAND_TTS_OUTPUT_FORMAT, _get_named_provider_config,
     _is_command_provider_config)
+from tools.tool_backend_helpers import NOUS_MANAGED_PROVIDER
 
 logger = logging.getLogger("tools.tts_tool")
 
@@ -35,7 +37,7 @@ def _plugin_route_key(provider: str, tts_config: Dict[str, Any]) -> Optional[str
     """Normalized *provider* when a plugin may service it, else None: built-in names never reach the
     registry and a same-named ``type: command`` provider wins (config is more local than a plugin)."""
     key = (provider or "").lower().strip()
-    if not key or key in BUILTIN_TTS_PROVIDERS:
+    if not key or key in BUILTIN_TTS_PROVIDERS or key == NOUS_MANAGED_PROVIDER:  # nous = the openai path
         return None
     if _is_command_provider_config(_get_named_provider_config(tts_config, key)):
         return None
@@ -52,12 +54,12 @@ def _plugin_voice_kwargs(tts_config: Dict[str, Any]) -> Dict[str, Any]:
             "speed": float(speed) if isinstance(speed, (int, float)) else None}
 
 
-def _plugin_pcm_streaming_provider(provider: str, tts_config: Dict[str, Any]):
-    """The plugin that services *provider* when it opted into raw-PCM streaming, else None.
+def _plugin_pcm_streaming_provider(provider: str, tts_config: Dict[str, Any]) -> Optional[Tuple[Any, int]]:
+    """``(plugin, sample_rate)`` when the plugin servicing *provider* opted into raw-PCM streaming.
 
-    Opt-in = ``streams_pcm`` truthy AND a positive ``stream_sample_rate`` AND ``is_available()`` —
-    all read at resolve time, so per-profile state (key present, user opted out) belongs there.
-    A missing rate is refused loudly: PCM played at a guessed rate garbles speech.
+    Opt-in = ``streams_pcm`` truthy AND ``stream_sample_rate`` a finite number >= 1 Hz AND
+    ``is_available()`` — read once per resolve, so per-profile state (key present, user opted out)
+    belongs there. A bad rate is refused loudly: PCM played at a guessed rate garbles speech.
     """
     key = _plugin_route_key(provider, tts_config)
     if key is None:
@@ -67,14 +69,14 @@ def _plugin_pcm_streaming_provider(provider: str, tts_config: Dict[str, Any]):
         if plugin is None or not plugin.streams_pcm or not plugin.is_available():
             return None
         rate = plugin.stream_sample_rate
-    except Exception as exc:  # noqa: BLE001 — a broken plugin falls back to per-sentence synthesis
-        logger.debug("plugin TTS streaming lookup failed for '%s': %s", key, exc)
+    except Exception:  # third-party plugin code: a broken plugin falls back to per-sentence synthesis
+        logger.debug("plugin TTS streaming lookup failed for '%s'", key, exc_info=True)
         return None
-    if isinstance(rate, bool) or not isinstance(rate, (int, float)) or rate <= 0:
-        logger.warning("TTS provider '%s' sets streams_pcm without a positive stream_sample_rate; "
-                       "using per-sentence synthesis instead", key)
+    if isinstance(rate, bool) or not isinstance(rate, (int, float)) or not math.isfinite(rate) or rate < 1:
+        logger.warning("TTS provider '%s' sets streams_pcm without a valid stream_sample_rate (%r); "
+                       "using per-sentence synthesis instead", key, rate)
         return None
-    return plugin
+    return plugin, int(rate)
 
 
 def _dispatch_to_plugin_provider(text: str, output_path: str, provider: str, tts_config: Dict[str, Any]) -> Optional[str]:
diff --git a/website/docs/developer-guide/streaming-tts.md b/website/docs/developer-guide/streaming-tts.md
index 497011631c..dd8280b94a 100644
--- a/website/docs/developer-guide/streaming-tts.md
+++ b/website/docs/developer-guide/streaming-tts.md
@@ -98,6 +98,9 @@ plugin streams first. Under `tts.streaming.provider: auto` it is tried only afte
 the built-in priority list. `streams_pcm`, `stream_sample_rate` and
 `is_available()` are read every time a streamer is resolved. If the rate is
 missing or the provider reports unavailable, Hermes keeps per-sentence synthesis.
+As with the built-in streamers, the speaker pipeline prefetches up to three
+sentences, so a plugin's `stream()` must tolerate concurrent calls (the sync
+path serializes `synthesize()`; this path does not).
 
 ## Gateway streaming (platform adapters)
 
diff --git a/website/docs/user-guide/features/tts.md b/website/docs/user-guide/features/tts.md
index 167989fe8c..9536836ea9 100644
--- a/website/docs/user-guide/features/tts.md
+++ b/website/docs/user-guide/features/tts.md
@@ -469,7 +469,7 @@ Override these on your provider class for richer integration:
 - `list_models()` → list of `{id, display, languages, max_text_length}` dicts.
 - `get_setup_schema()` → return `{name, badge, tag, env_vars: [{key, prompt, url}]}` to power the picker row in `hermes tools` / `hermes setup`. Without this, the plugin still works but its row in the picker is minimal.
 - `stream(text, *, voice, model, format, **extra)` → iterator yielding audio bytes (default raises `NotImplementedError`).
-- `streams_pcm = True` + `stream_sample_rate` (Hz) → join the streaming voice path (CLI/TUI voice mode, desktop read-aloud, gateway streaming audio). Hermes then calls `stream(text, format="pcm", voice=..., model=..., speed=...)` with the same `tts.voice` / `tts.model` / `tts.speed` that `synthesize()` gets, and expects raw int16 little-endian mono PCM at that rate. Both attributes and `is_available()` are read each time a reply starts, so they can be properties that reflect live state. Without a positive `stream_sample_rate`, or when `is_available()` is `False`, Hermes keeps synthesizing one sentence at a time.
+- `streams_pcm = True` + `stream_sample_rate` (Hz) → join the streaming voice path (CLI/TUI voice mode, desktop read-aloud, gateway streaming audio). Hermes then calls `stream(text, format="pcm", voice=..., model=..., speed=...)` with the same `tts.voice` / `tts.model` / `tts.speed` that `synthesize()` gets, and expects raw int16 little-endian mono PCM at that rate. Both attributes and `is_available()` are read each time a reply starts, so they can be properties that reflect live state. Unlike `synthesize()`, `stream()` can be called for up to three consecutive sentences at once while earlier audio plays, so it must be thread-safe. Without a positive `stream_sample_rate`, or when `is_available()` is `False`, Hermes keeps synthesizing one sentence at a time.
 - `voice_compatible` property → set `True` if your output is Opus-compatible and the gateway should deliver it as a voice bubble (default `False` = regular audio attachment).
 - `warm()` / `release()` → called when a surface toggles speech output on / when the last lease across surfaces is released, while your provider is the configured `tts.provider` — preload or unload a local model server here. Both default to no-ops; exceptions are logged at debug and never fail the toggle.