fix(llm): stop DeepSeek enrichment failing on json_schema rejection

Enriching a model with `llm_provider=deepseek` failed outright with
HTTP 400 "This response_format type is unavailable now".  Probing the
endpoint shows why:

    response_format absent      -> 200
    {"type": "json_object"}     -> 200
    {"type": "json_schema",...} -> 400

`chat_completion_json` preferred `json_schema` for a real reason -- LM
Studio and other local OpenAI-compatible servers reject `json_object`
but accept `json_schema` -- and guarded the fallback with a substring
test for `'response_format.type'` (the wording of those servers'
rejection).  DeepSeek's message is "This response_format type is
unavailable now", which does not contain that substring, so the guard
re-raised and the retry never ran.

Make the format a per-provider chain instead of a single guess:

- `_JSON_OBJECT_ONLY_PROVIDERS` lists providers known to reject
  json_schema (currently just deepseek).  They ask for `json_object`
  first, so the common case costs one request and no wasted retry.
- Everyone else keeps `json_schema` first, then downgrades through
  `json_object` and finally prompt-only mode.
- A downgrade now happens on any error mentioning `response_format`,
  which covers wording variants without swallowing unrelated failures:
  auth errors, unknown models, and rate limits still surface unchanged
  because their messages never name the parameter.

`json_object` is sufficient here: the skill prompt already specifies the
exact JSON shape, and `_try_salvage_json` repairs imperfect output.

Verified against the real configured endpoint with the real
`enrich_hf_metadata` prompt, prompt renderer, and ModelScope model card
for jj3550945163/Krea-2-LORA: a 9,815-character prompt returns
parseable JSON (base_model "Flux.1 Krea", description, tags, notes).

Three regression tests cover the DeepSeek ordering, the
json_schema -> json_object downgrade, and the no-retry-on-unrelated-400
path.  Full backend suite: 2856 passed.

py/services/llm_service.py

@@ -267,6 +267,16 @@ _PROVIDER_DEFAULTS: Dict[str, str] = {
 # Request timeout for LLM calls (seconds)
 _LLM_TIMEOUT = aiohttp.ClientTimeout(total=120)
 
+# Providers that do NOT implement ``response_format: {"type": "json_schema"}``
+# and reject it with HTTP 400.  For these the weaker, widely supported
+# ``json_object`` mode is used instead (the prompt already specifies the
+# expected JSON shape, and ``_try_salvage_json`` repairs imperfect output).
+# DeepSeek answers a json_schema request with
+# ``{"error":{"message":"This response_format type is unavailable now"}}``.
+# LM Studio and some other local OpenAI-compatible servers reject
+# ``json_object`` but accept ``json_schema``, so they are not listed here.
+_JSON_OBJECT_ONLY_PROVIDERS = frozenset({"deepseek"})
+
 
 class LLMService:
     """Centralized LLM API client.
@@ -614,47 +624,61 @@ class LLMService:
         if effective_max is None:
             effective_max = 4096
 
-        # Use json_schema (not json_object) for broader provider compatibility:
-        # LM Studio and some other OpenAI-compatible servers reject
-        # json_object but accept json_schema.  {"type": "object"} is
-        # functionally equivalent — it accepts any JSON object without
-        # constraining specific fields.
-        response_format = {
+        # Structured-output format.  ``json_schema`` is preferred because LM
+        # Studio and other local OpenAI-compatible servers reject
+        # ``json_object`` but accept ``json_schema``; ``{"type": "object"}``
+        # accepts any JSON object without constraining specific fields, so the
+        # two modes are functionally equivalent here.  Providers known to
+        # reject json_schema (see _JSON_OBJECT_ONLY_PROVIDERS) get
+        # ``json_object`` instead.
+        schema_format: Dict[str, Any] = {
             "type": "json_schema",
             "json_schema": {
                 "name": "metadata",
                 "schema": {"type": "object"},
             },
         }
+        json_object_format: Dict[str, Any] = {"type": "json_object"}
 
-        try:
-            result = await self.chat_completion(
-                messages=messages,
-                model=model,
-                temperature=temperature,
-                response_format=response_format,
-                max_tokens=effective_max,
-            )
-        except LLMResponseError as e:
-            # Only fall back when the provider rejects the response_format
-            # type value (e.g. "'response_format.type' must be...").  Avoid
-            # catching unrelated 400 errors whose body happens to mention
-            # "response_format" (e.g. "model does not support
-            # response_format restrictions on this endpoint").
-            if "'response_format.type'" not in str(e).lower():
-                raise
-            logger.info(
-                "Provider rejected response_format, retrying without it. "
-                "Falling back to prompt-only JSON mode. Error: %s",
-                e,
-            )
-            result = await self.chat_completion(
-                messages=messages,
-                model=model,
-                temperature=temperature,
-                response_format=None,
-                max_tokens=effective_max,
-            )
+        if self._get_config()["provider"] in _JSON_OBJECT_ONLY_PROVIDERS:
+            format_chain: List[Optional[Dict[str, Any]]] = [
+                json_object_format,
+                None,
+            ]
+        else:
+            format_chain = [schema_format, json_object_format, None]
+
+        result: Optional[Dict[str, Any]] = None
+        for index, fmt in enumerate(format_chain):
+            try:
+                result = await self.chat_completion(
+                    messages=messages,
+                    model=model,
+                    temperature=temperature,
+                    response_format=fmt,
+                    max_tokens=effective_max,
+                )
+                break
+            except LLMResponseError as e:
+                message = str(e).lower()
+                if index + 1 >= len(format_chain):
+                    raise
+                # Only downgrade when the failure is about ``response_format``.
+                # Everything else (auth, unknown model, rate limits) must
+                # surface unchanged.  Matching on the bare parameter name also
+                # covers variants such as DeepSeek's "This response_format
+                # type is unavailable now" without swallowing unrelated 400s.
+                if "response_format" not in message:
+                    raise
+                logger.info(
+                    "Provider rejected response_format=%s, retrying with %s. "
+                    "Error: %s",
+                    (fmt or {}).get("type", "none"),
+                    (format_chain[index + 1] or {}).get("type", "none"),
+                    e,
+                )
+
+        assert result is not None  # non-empty chain always sets or raises
 
         content = result.get("content", "") or ""
         if not content:

tests/services/test_llm_service.py

@@ -76,6 +76,25 @@ class MockSession:
         pass
 
 
+class RecordingSession:
+    """Mock session that records each request payload and replays responses."""
+
+    def __init__(self, responses):
+        self._responses = list(responses)
+        self.payloads = []
+
+    def post(self, url, json=None, headers=None):
+        self.payloads.append(json)
+        index = min(len(self.payloads) - 1, len(self._responses) - 1)
+        return self._responses[index]
+
+    async def __aenter__(self):
+        return self
+
+    async def __aexit__(self, *args):
+        pass
+
+
 @pytest.fixture
 def llm_service():
     """Create an LLMService with mock settings."""
@@ -298,6 +317,107 @@ class TestLLMServiceChatCompletionJson:
         assert result == {"key": "value"}
         assert call_index == 2
 
+    @pytest.mark.asyncio
+    async def test_chat_completion_json_prefers_json_object_for_deepseek(self):
+        """DeepSeek rejects json_schema, so json_object is used first.
+
+        Regression: DeepSeek answers json_schema with
+        "This response_format type is unavailable now", which the old
+        substring check did not recognise, so enrichment failed outright.
+        """
+        settings = MockSettings(
+            llm_enabled=True,
+            llm_provider="deepseek",
+            llm_api_key="sk-test-key",
+            llm_api_base="https://api.deepseek.com/v1",
+            llm_model="deepseek-v4-flash",
+        )
+        service = LLMService(settings)
+
+        session = RecordingSession(
+            [
+                MockResponse(
+                    200,
+                    json_data={
+                        "choices": [{"message": {"content": '{"key": "value"}'}}],
+                        "usage": {},
+                    },
+                )
+            ]
+        )
+
+        with mock.patch("aiohttp.ClientSession", return_value=session):
+            result = await service.chat_completion_json(
+                system_prompt="You are helpful.",
+                user_prompt="Return JSON.",
+            )
+
+        assert result == {"key": "value"}
+        assert len(session.payloads) == 1
+        assert session.payloads[0]["response_format"] == {"type": "json_object"}
+
+    @pytest.mark.asyncio
+    async def test_chat_completion_json_downgrades_from_json_schema(
+        self, llm_service,
+    ):
+        """json_schema → json_object when the provider rejects json_schema."""
+        session = RecordingSession(
+            [
+                MockResponse(
+                    400,
+                    text_data=(
+                        '{"error":{"message":"This response_format type is '
+                        'unavailable now","type":"invalid_request_error"}}'
+                    ),
+                ),
+                MockResponse(
+                    200,
+                    json_data={
+                        "choices": [{"message": {"content": '{"key": "value"}'}}],
+                        "usage": {},
+                    },
+                ),
+            ]
+        )
+
+        with mock.patch("aiohttp.ClientSession", return_value=session):
+            result = await llm_service.chat_completion_json(
+                system_prompt="You are helpful.",
+                user_prompt="Return JSON.",
+            )
+
+        assert result == {"key": "value"}
+        assert [p.get("response_format") for p in session.payloads] == [
+            {
+                "type": "json_schema",
+                "json_schema": {"name": "metadata", "schema": {"type": "object"}},
+            },
+            {"type": "json_object"},
+        ]
+
+    @pytest.mark.asyncio
+    async def test_chat_completion_json_does_not_retry_unrelated_errors(
+        self, llm_service,
+    ):
+        """Unrelated 400s are surfaced unchanged, without format downgrades."""
+        session = RecordingSession(
+            [
+                MockResponse(
+                    400,
+                    text_data='{"error":{"message":"Model not found"}}',
+                )
+            ]
+        )
+
+        with mock.patch("aiohttp.ClientSession", return_value=session):
+            with pytest.raises(LLMResponseError, match="HTTP 400"):
+                await llm_service.chat_completion_json(
+                    system_prompt="You are helpful.",
+                    user_prompt="Return JSON.",
+                )
+
+        assert len(session.payloads) == 1
+
     @pytest.mark.asyncio
     async def test_chat_completion_json_raises_on_non_json(self, llm_service):
         # Non-JSON content raises LLMResponseError (salvage also fails)