hermes: preserve Switchyard response content

services/hermes/agent-deployment.yaml

@@ -25,7 +25,7 @@ spec:
         ai.bstein.dev/execution: Hermes Kanban with durable direct Codex and Claude Code CLI workers
         ai.bstein.dev/model-policy: Jetson-assisted AUTO routing, low through xhigh, cross-provider fallback
         ai.bstein.dev/placement: rpi5 preferred; Jetson deferred until state storage is available
-        ai.bstein.dev/config-rev: "20260811-switchyard-responses-adapter"
+        ai.bstein.dev/config-rev: "20260811-switchyard-responses-output"
         vault.hashicorp.com/agent-inject: "true"
         vault.hashicorp.com/role: hermes-agent
         vault.hashicorp.com/agent-inject-secret-anthropic-token: kv/data/atlas/hermes/agent-tokens

services/hermes/model-gate-configmap.yaml

@@ -73,15 +73,12 @@ data:
         # unambiguous and safe default.
         payload["model"] = "qwen2.5:14b-instruct-q4_0"
         changed = True
-      for key in ("reasoning_effort", "reasoning"):
+      # This non-thinking Qwen build rejects hosted-provider reasoning fields.
+      # Effort remains a Switchyard routing concern, not an Ollama request.
+      for key in ("reasoning_effort", "reasoning", "thinking", "output_config"):
         value = payload.get(key)
-        if isinstance(value, str) and value.lower() in {"xhigh", "max"}:
-          payload[key] = "high"
-          changed = True
-        elif isinstance(value, dict):
-          effort = value.get("effort")
-          if isinstance(effort, str) and effort.lower() in {"xhigh", "max"}:
-            value["effort"] = "high"
+        if value is not None:
+          payload.pop(key, None)
           changed = True
       if not changed:
         return body

services/hermes/model-gate-deployment.yaml

@@ -15,7 +15,7 @@ spec:
   template:
     metadata:
       annotations:
-        ai.bstein.dev/config-rev: "20260811-switchyard-model-default"
+        ai.bstein.dev/config-rev: "20260811-switchyard-model-fields"
       labels:
         app: hermes-model-gate
     spec:

services/hermes/scripts/codex_broker.py

@@ -133,6 +133,7 @@ def _validate_payload(payload: Any) -> dict[str, Any]:
 def _completed_response(lines: Iterable[str]) -> dict[str, Any]:
     """Collapse a Codex SSE stream for a non-streaming Responses caller."""
     terminal_response: dict[str, Any] | None = None
+    output_items: dict[int, dict[str, Any]] = {}
     upstream_error = ""
     for line in lines:
         if not line.startswith("data:"):
@@ -148,6 +149,11 @@ def _completed_response(lines: Iterable[str]) -> dict[str, Any]:
             continue
         event_type = str(event.get("type") or "")
         response = event.get("response")
+        if event_type == "response.output_item.done":
+            item = event.get("item")
+            output_index = event.get("output_index")
+            if isinstance(item, dict) and isinstance(output_index, int):
+                output_items[output_index] = item
         if event_type in {
             "response.completed",
             "response.failed",
@@ -161,6 +167,13 @@ def _completed_response(lines: Iterable[str]) -> dict[str, Any]:
             else:
                 upstream_error = str(error or "")
     if terminal_response is not None:
+        # The subscription Codex endpoint streams complete output items but
+        # currently leaves the terminal response's output array empty.  Public
+        # Responses clients, including Switchyard, expect those items there.
+        if output_items and not terminal_response.get("output"):
+            terminal_response["output"] = [
+                output_items[index] for index in sorted(output_items)
+            ]
         return terminal_response
     raise RuntimeError(upstream_error or "Codex stream ended without a terminal response")
 

services/hermes/switchyard-configmap.yaml

@@ -85,12 +85,10 @@ data:
     [targets.local_qwen_low]
     id = "route/local/qwen2.5-14b/low"
     llm_client = "local_low"
-    extra_body = { reasoning_effort = "low" }
 
     [targets.local_qwen_medium]
     id = "route/local/qwen2.5-14b/medium"
     llm_client = "local_medium"
-    extra_body = { reasoning_effort = "medium" }
 
     [targets.codex_luna_low]
     id = "route/codex/luna/low"

services/hermes/switchyard-deployment.yaml

@@ -19,7 +19,7 @@ spec:
       labels:
         app: hermes-switchyard
       annotations:
-        ai.bstein.dev/config-rev: "20260811-switchyard-authority-v6"
+        ai.bstein.dev/config-rev: "20260811-switchyard-authority-v7"
         prometheus.io/scrape: "true"
         prometheus.io/port: "9005"
         prometheus.io/path: /metrics

testing/tests/test_hermes_chat_quality.py

@@ -548,16 +548,31 @@ def test_codex_broker_auth_and_request_contract(tmp_path: Path, monkeypatch):
         "status": "completed",
         "output": [],
     }
+    completed_item = {
+        "type": "message",
+        "role": "assistant",
+        "status": "completed",
+        "content": [{"type": "output_text", "text": "done"}],
+    }
     assert module._completed_response(
         [
             "event: response.created",
             'data: {"type":"response.created","response":{}}',
+            "event: response.output_item.done",
+            "data: "
+            + json.dumps(
+                {
+                    "type": "response.output_item.done",
+                    "output_index": 0,
+                    "item": completed_item,
+                }
+            ),
             "event: response.completed",
             "data: "
             + json.dumps({"type": "response.completed", "response": completed}),
             "data: [DONE]",
         ]
-    ) == completed
+    )["output"] == [completed_item]
     with pytest.raises(RuntimeError, match="provider unavailable"):
         module._completed_response(
             [

testing/tests/test_hermes_model_gate.py

@@ -18,24 +18,21 @@ def _model_gate_namespace() -> dict:
     return namespace
 
 
-def test_model_gate_clamps_hosted_only_reasoning_efforts():
+def test_model_gate_removes_hosted_only_reasoning_fields():
     normalize = _model_gate_namespace()["_normalize_reasoning"]
 
     top_level = json.loads(normalize(b'{"reasoning_effort":"xhigh"}'))
     string_reasoning = json.loads(normalize(b'{"reasoning":"max"}'))
     nested = json.loads(normalize(b'{"reasoning":{"effort":"xhigh"}}'))
 
-    assert top_level["reasoning_effort"] == "high"
-    assert string_reasoning["reasoning"] == "high"
-    assert nested["reasoning"]["effort"] == "high"
+    assert "reasoning_effort" not in top_level
+    assert "reasoning" not in string_reasoning
+    assert "reasoning" not in nested
 
 
 def test_model_gate_preserves_supported_and_non_json_requests():
     normalize = _model_gate_namespace()["_normalize_reasoning"]
-    supported = (
-        b'{"model":"qwen2.5:14b-instruct-q4_0",'
-        b'"reasoning_effort":"medium","messages":[]}'
-    )
+    supported = b'{"model":"qwen2.5:14b-instruct-q4_0","messages":[]}'
     non_json = b"streamed-body"
 
     assert normalize(supported) == supported
@@ -53,7 +50,7 @@ def test_model_gate_translates_switchyard_local_target_alias():
     )
 
     assert routed["model"] == "qwen2.5:14b-instruct-q4_0"
-    assert routed["reasoning_effort"] == "medium"
+    assert "reasoning_effort" not in routed
 
 
 def test_model_gate_supplies_its_single_model_when_switchyard_omits_it():