hermes: harden routed vision normalization

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: "20260816-routed-vision-v3"
+        ai.bstein.dev/config-rev: "20260816-routed-vision-v4"
         prometheus.io/scrape: "true"
         prometheus.io/path: /metrics
         prometheus.io/port: "9010"

services/hermes/scripts/codex_broker.py

@@ -226,13 +226,14 @@ def _normalize_input_images(response_input: list[Any]) -> None:
 
     def image_url_value(part: dict[str, Any]) -> tuple[str | None, str | None]:
         """Return a Responses URL from URL-object or base64-source forms."""
-        raw = part.get("image_url")
-        if isinstance(raw, str):
-            return raw, None
-        candidates = [raw, part.get("source")]
-        for candidate in candidates:
-            if not isinstance(candidate, dict):
-                continue
+
+        def search(candidate: Any, depth: int = 0) -> tuple[str | None, str | None]:
+            if isinstance(candidate, str):
+                if candidate.startswith(("data:image/", "https://", "http://")):
+                    return candidate, None
+                return None, None
+            if not isinstance(candidate, dict) or depth > 4:
+                return None, None
             detail = candidate.get("detail")
             url = candidate.get("url")
             if isinstance(url, str) and url.strip():
@@ -249,8 +250,39 @@ def _normalize_input_images(response_input: list[Any]) -> None:
                 return f"data:{media_type};base64,{data}", (
                     detail if isinstance(detail, str) else None
                 )
+            for nested in candidate.values():
+                nested_url, nested_detail = search(nested, depth + 1)
+                if nested_url:
+                    return nested_url, (
+                        detail if isinstance(detail, str) else nested_detail
+                    )
+            return None, None
+
+        for candidate in (part.get("image_url"), part.get("source"), part):
+            image_url, detail = search(candidate)
+            if image_url:
+                return image_url, detail
         return None, None
 
+    def image_shape(part: dict[str, Any]) -> str:
+        """Describe only structural types when an upstream image is malformed."""
+
+        def describe(value: Any, depth: int = 0) -> Any:
+            if depth > 3:
+                return type(value).__name__
+            if isinstance(value, dict):
+                return {
+                    str(key)[:40]: describe(nested, depth + 1)
+                    for key, nested in list(value.items())[:12]
+                }
+            if isinstance(value, list):
+                return [describe(nested, depth + 1) for nested in value[:4]]
+            if isinstance(value, str):
+                return f"str[{len(value)}]"
+            return type(value).__name__
+
+        return json.dumps(describe(part), sort_keys=True, separators=(",", ":"))
+
     for item in response_input:
         if not isinstance(item, dict):
             continue
@@ -268,7 +300,10 @@ def _normalize_input_images(response_input: list[Any]) -> None:
             if isinstance(detail, str) and detail and "detail" not in part:
                 part["detail"] = detail
             if not isinstance(image_url, str) or not image_url.strip():
-                raise ValueError("non-empty Responses image URL required")
+                raise ValueError(
+                    "non-empty Responses image URL required; shape="
+                    + image_shape(part)
+                )
             part["type"] = "input_image"
             part["image_url"] = image_url
             part.pop("source", None)

testing/tests/test_hermes_chat_quality.py

@@ -824,7 +824,26 @@ def test_codex_broker_auth_and_request_contract(tmp_path: Path, monkeypatch):
         "type": "input_image",
         "image_url": "data:image/png;base64,cHJpdmF0ZQ==",
     }
-    with pytest.raises(ValueError, match="non-empty Responses image URL"):
+    nested_switchyard_items = [
+        {
+            "role": "user",
+            "content": [
+                {
+                    "type": "input_image",
+                    "image_url": {
+                        "image": {
+                            "original_url": "data:image/png;base64,cHJpdmF0ZQ=="
+                        }
+                    },
+                }
+            ],
+        }
+    ]
+    nested_image = module._validate_payload(
+        {"model": "gpt-5.6-terra", "input": nested_switchyard_items}
+    )["input"][0]["content"][0]
+    assert nested_image["image_url"] == "data:image/png;base64,cHJpdmF0ZQ=="
+    with pytest.raises(ValueError, match=r"non-empty Responses image URL.*str\[4\]"):
         module._validate_payload(
             {
                 "model": "gpt-5.6-terra",
@@ -833,7 +852,10 @@ def test_codex_broker_auth_and_request_contract(tmp_path: Path, monkeypatch):
                         "type": "message",
                         "role": "user",
                         "content": [
-                            {"type": "input_image", "image_url": {"detail": "high"}}
+                            {
+                                "type": "input_image",
+                                "image_url": {"detail": "high"},
+                            }
                         ],
                     }
                 ],