feat: inject Pydantic JSON schema into Deepseek prompt (Phase 6)

Follows Deepseek JSON Output guide: the prompt now includes the word 'json' and a format example derived from the Pydantic model schema. Added _pydantic_to_json_instruction() helper that builds a human-readable schema description with EXAMPLE JSON OUTPUT.

Ultraworked with [Sisyphus](https://github.com/code-yeongyu/oh-my-openagent)

Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>

backend/app/services/llm_client_dp.py

@@ -39,6 +39,53 @@ def _truncate_prompt_for_log(prompt: str, first_chars: int = 100, last_chars: in
     )
 
 
+def _pydantic_to_json_instruction(model: Any) -> str:
+    """Build a JSON-format instruction from a Pydantic model's schema.
+
+    Follows the Deepseek JSON Output guide: the prompt must contain the word
+    "json" and an example of the expected shape.  The model schema is
+    converted into a human-readable text description with a filled-in example.
+    """
+    schema = model.model_json_schema()
+    props = schema.get("properties", {})
+    title = schema.get("title", model.__name__)
+
+    parts: list[str] = []
+    parts.append(f"Output the result in JSON format as a {title} object.")
+
+    # Build an example by filling each field with a representative value.
+    example: dict[str, Any] = {}
+    for name, info in props.items():
+        t = info.get("type", "any")
+        desc = info.get("description", "")
+        if t == "array":
+            items = info.get("items", {})
+            item_type = items.get("type", "string")
+            min_items = info.get("minItems", 1)
+            parts.append(
+                f'- "{name}": array of {item_type} '
+                f"(min {min_items}"
+                + (f", max {info['maxItems']}" if info.get("maxItems") else "")
+                + f") — {desc}"
+            )
+            example[name] = [f"<{item_type}_1>", f"<{item_type}_2>"]
+        elif t == "string":
+            parts.append(f'- "{name}": {t} — {desc}')
+            example[name] = f"<{desc[:40]}>"
+        elif t == "integer" or t == "number":
+            parts.append(f'- "{name}": {t} — {desc}')
+            example[name] = 0
+        else:
+            parts.append(f'- "{name}": {t} — {desc}')
+            example[name] = f"<{t}>"
+
+    parts.append("")
+    parts.append("EXAMPLE JSON OUTPUT:")
+    parts.append(json.dumps(example, indent=2, ensure_ascii=False))
+
+    return "\n".join(parts)
+
+
 class LLMClientDP:
     """Async Deepseek API client for query decomposition.
 
@@ -126,17 +173,21 @@ class LLMClientDP:
 
         Deepseek supports ``response_format={"type": "json_object"}`` (which
         guarantees valid JSON) but not OpenAI's ``json_schema`` mode (which
-        would validate against a specific schema).  We use the JSON mode to
+        would validate against a specific schema).  We inject a JSON format
-        get a guaranteed-valid JSON response, then validate it client-side
+        instruction derived from *pydantic_model* into the prompt (per the
-        against *pydantic_model*.
+        Deepseek JSON Output guide), then validate client-side.
         """
         prompt_preview = _truncate_prompt_for_log(prompt, first_chars=300, last_chars=100)
         logger.info("[%s] Deepseek structured request started. Prompt: %s", step_name, prompt_preview)
         start_time = time.perf_counter()
 
+        # Inject JSON format instruction from the Pydantic model.
+        json_instruction = _pydantic_to_json_instruction(pydantic_model)
+        full_prompt = f"{prompt}\n\n{json_instruction}"
+
         try:
             response = await self.complete(
-                prompt=prompt,
+                prompt=full_prompt,
                 temperature=0.0,
                 step_name=step_name,
                 response_format={"type": "json_object"},

backend/app/test/test_phase6_llm_client_dp.py

@@ -191,11 +191,16 @@ class TestLLMClientDPCompleteStructured:
         assert result == expected
         assert result.questions == ["Q1", "Q2", "Q3"]
         mock_complete.assert_called_once()
-        # Verify Deepseek JSON mode is used
         call_kwargs = mock_complete.call_args.kwargs
         assert call_kwargs["response_format"] == {"type": "json_object"}
         assert call_kwargs["temperature"] == 0.0
 
+        # The prompt must contain the JSON format instruction per Deepseek docs.
+        full_prompt = call_kwargs["prompt"]
+        assert "json" in full_prompt.lower()
+        assert "EXAMPLE JSON OUTPUT" in full_prompt
+        assert '"questions"' in full_prompt
+
     @pytest.mark.asyncio
     async def test_complete_structured_with_markdown_fence(self, client):
         """Should strip markdown code fences before JSON parsing."""