gateway: find the model where the endpoint actually names it

/v1/ocr names its engine in a multipart field called `engine`, not `model`.
The gateway only ever looked for `model`, so an OCR request carried no model as
far as it was concerned: per-model placement resolved correctly, the test
reported 'target: its own pod', and then the gateway refused with 'nothing
configures ocr to run remotely'. Two layers disagreeing about the same request.

_MODEL_FIELD maps the endpoints whose field is named something else, and the
extractor takes the field name for JSON, multipart and urlencoded alike.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0182DhR1eFNPQyGmDrpHbedY

codai/api/remote_gateway.py

@@ -258,8 +258,14 @@ def _any_remote_configured_uncached() -> bool:
     return False
 
 
-def _model_from_body(body: bytes, content_type: str) -> str:
-    """Pull the requested model out of a JSON or multipart body."""
+def _model_from_body(body: bytes, content_type: str, field: str = "model") -> str:
+    """Pull the requested model out of a JSON or multipart body.
+
+    ``field`` because not every endpoint calls it "model": /v1/ocr names its
+    engine in a field called `engine`, so a per-model placement for OCR resolved
+    correctly and was then refused by the gateway, which could not see which
+    engine the request was for.
+    """
     if not body:
         return ""
     ctype = (content_type or "").lower()
@@ -268,18 +274,26 @@ def _model_from_body(body: bytes, content_type: str) -> str:
             obj = json.loads(body)
         except Exception:
             return ""
-        return str(obj.get("model") or "") if isinstance(obj, dict) else ""
+        return str(obj.get(field) or "") if isinstance(obj, dict) else ""
     if "multipart/form-data" in ctype:
-        m = _MULTIPART_MODEL.search(body[:262144])
+        pat = (_MULTIPART_MODEL if field == "model"
+               else re.compile(rb'name="' + field.encode() +
+                               rb'"\r?\n(?:[^\r\n]*\r?\n)*?\r?\n([^\r\n]*)',
+                               re.IGNORECASE))
+        m = pat.search(body[:262144])
         if m:
             return m.group(1).decode("utf-8", "replace").strip()
     if "application/x-www-form-urlencoded" in ctype:
         from urllib.parse import parse_qs
-        vals = parse_qs(body.decode("utf-8", "replace")).get("model")
+        vals = parse_qs(body.decode("utf-8", "replace")).get(field)
         return (vals or [""])[0]
     return ""
 
 
+#: Endpoints whose "which model" field is not called `model`.
+_MODEL_FIELD = {"/v1/ocr": "engine", "/v1/ocr/batch": "engine"}
+
+
 class Target:
     """Where a request is going: a fixed URL, or a pod pool to borrow one from."""
 
@@ -333,10 +347,11 @@ def resolve_target(path: str, method: str, query: str, body: bytes,
         base = file_origin(path[len("/v1/files/"):])
         return Target(url=base, reason="generated file") if base else None
     cap = capability_for(path)
-    model = _model_from_body(body, content_type)
+    field = _MODEL_FIELD.get(path, "model")
+    model = _model_from_body(body, content_type, field)
     if not model and query:
         from urllib.parse import parse_qs
-        model = (parse_qs(query).get("model") or [""])[0]
+        model = (parse_qs(query).get(field) or [""])[0]
 
     # Per model first, so two models of the same kind can be placed differently.
     where, detail = model_placement(model)

tests/test_remote_execution.py

@@ -1615,3 +1615,23 @@ def test_a_200_carrying_no_media_is_not_a_pass():
     assert _empty_result(json.dumps({"url": "/v1/files/x.mp4"}).encode(), "video") == ""
     for cap in ("voice", "audio_gen", "video"):
         assert "no media" in _empty_result(json.dumps({"ok": True}).encode(), cap)
+
+
+def test_the_gateway_finds_the_ocr_engine_field():
+    """/v1/ocr names its engine in a field called `engine`, not `model`. The
+    gateway could not see it, so a per-model OCR placement resolved correctly
+    and was then refused: 'nothing configures ocr to run remotely'."""
+    from codai.api.remote_gateway import _model_from_body, _MODEL_FIELD
+
+    assert _MODEL_FIELD["/v1/ocr"] == "engine"
+
+    body = (b'--B\r\nContent-Disposition: form-data; name="engine"\r\n\r\nsurya\r\n'
+            b'--B\r\nContent-Disposition: form-data; name="file"; '
+            b'filename="a.png"\r\n\r\n\x89PNG\r\n--B--\r\n')
+    ct = "multipart/form-data; boundary=B"
+    assert _model_from_body(body, ct, "engine") == "surya"
+    # and the default field is still `model`, absent here
+    assert _model_from_body(body, ct) == ""
+
+    # JSON bodies honour the field too.
+    assert _model_from_body(b'{"engine":"doctr"}', "application/json", "engine") == "doctr"