fix(thunderagent): support topology-only wrapped backends

components/src/dynamo/common/configuration/groups/runtime_args.py

@@ -266,7 +266,10 @@ def add_arguments(self, parser: argparse.ArgumentParser) -> None:
             env_var="DYN_ENDPOINT_TYPES",
             default="chat,completions",
             obsolete_flag="--dyn-endpoint-types",
-            help="Comma-separated list of endpoint types to enable. Options: 'chat', 'completions'. Use 'completions' for models without chat templates.",
+            help="Comma-separated list of endpoint types to enable. Options: "
+            "'chat', 'completions', or 'none'. Use 'completions' for models "
+            "without chat templates. Use 'none' for topology-only workers "
+            "fronted by another Dynamo service.",
         )
 
         add_argument(

components/src/dynamo/common/utils/endpoint_types.py

@@ -11,14 +11,15 @@ def parse_endpoint_types(endpoint_types_str: str) -> ModelType:
 
     Args:
         endpoint_types_str: Comma-separated list of endpoint types.
-                          Valid values: 'chat', 'completions'
-                          Examples: 'chat', 'completions', 'chat,completions'
+                          Valid values: 'chat', 'completions', or 'none'
+                          Examples: 'chat', 'completions', 'chat,completions', 'none'
 
     Returns:
         ModelType flags combined with bitwise OR
 
     Raises:
-        ValueError: If any invalid endpoint type is provided or string is empty
+        ValueError: If any invalid endpoint type is provided, 'none' is mixed
+            with other endpoint types, or the string is empty
 
     Examples:
         >>> parse_endpoint_types("chat")
@@ -27,6 +28,8 @@ def parse_endpoint_types(endpoint_types_str: str) -> ModelType:
         ModelType.Completions
         >>> parse_endpoint_types("chat,completions")
         ModelType.Chat | ModelType.Completions
+        >>> parse_endpoint_types("none")
+        ModelType.Empty
     """
     if not endpoint_types_str or not endpoint_types_str.strip():
         raise ValueError("Endpoint types string cannot be empty")
@@ -36,6 +39,13 @@ def parse_endpoint_types(endpoint_types_str: str) -> ModelType:
     if not types:
         raise ValueError("No valid endpoint types provided")
 
+    if "none" in types:
+        if len(types) > 1:
+            raise ValueError(
+                "Endpoint type 'none' cannot be combined with other endpoint types"
+            )
+        return ModelType.Empty
+
     result: ModelType | None = None
     for t in types:
         if t == "chat":
@@ -44,7 +54,8 @@ def parse_endpoint_types(endpoint_types_str: str) -> ModelType:
             flag = ModelType.Completions
         else:
             raise ValueError(
-                f"Invalid endpoint type: '{t}'. Valid options: 'chat', 'completions'"
+                f"Invalid endpoint type: '{t}'. Valid options: "
+                "'chat', 'completions', 'none'"
             )
 
         result = flag if result is None else result | flag

components/src/dynamo/common/utils/tests/test_endpoint_types.py

@@ -0,0 +1,36 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+"""Unit tests for endpoint type parsing."""
+
+import pytest
+
+from dynamo.common.utils.endpoint_types import parse_endpoint_types
+from dynamo.llm import ModelType
+
+pytestmark = [
+    pytest.mark.unit,
+    pytest.mark.gpu_0,
+    pytest.mark.pre_merge,
+]
+
+
+def test_parse_endpoint_types_allows_topology_only_workers():
+    assert parse_endpoint_types("none") == ModelType.Empty
+
+
+@pytest.mark.parametrize("value", ["chat,none", "none,completions"])
+def test_parse_endpoint_types_rejects_mixing_none_with_public_surfaces(value):
+    with pytest.raises(ValueError, match="'none' cannot be combined"):
+        parse_endpoint_types(value)
+
+
+@pytest.mark.parametrize("value", ["", "   ", ",,,"])
+def test_parse_endpoint_types_rejects_empty_input(value):
+    with pytest.raises(ValueError):
+        parse_endpoint_types(value)
+
+
+def test_parse_endpoint_types_invalid_option_lists_none():
+    with pytest.raises(ValueError, match="'chat', 'completions', 'none'"):
+        parse_endpoint_types("responses")

components/src/dynamo/thunderagent_router/README.md

@@ -36,6 +36,7 @@ uv pip install -e .
 # 1. Start your Dynamo workers (vLLM example, with KV events on)
 python -m dynamo.vllm \
     --model <model> --tensor-parallel-size <N> \
+    --endpoint-types none \
     --kv-events-config '{"publisher":"zmq","topic":"kv-events",
                          "endpoint":"tcp://*:20080",
                          "enable_kv_cache_events":true}'
@@ -51,6 +52,12 @@ python -m dynamo.thunderagent_router \
 python -m dynamo.frontend --router-mode round-robin
 ```
 
+Use `--endpoint-types none` on workers wrapped by ThunderAgent so they register
+only for topology and readiness. ThunderAgent registers the public
+chat/completions surface; if the wrapped backend also advertises that surface
+for the same model, the frontend may route requests directly to the backend and
+bypass ThunderAgent lifecycle handling such as `x-dynamo-session-final`.
+
 The control-loop knobs (`--pause-threshold`, `--pause-target`,
 `--resume-hysteresis`, `--scheduler-interval-seconds`, …) and their defaults are
 documented in [docs/agents/thunderagent-router.md](../../../../docs/fern/agents/thunderagent-router.md#utilization-driven-control-loop).