fix(deriver): document explicit observation JSON shape

src/deriver/prompts.py

@@ -76,6 +76,10 @@ def minimal_deriver_prompt(
 - EXPLICIT: "I took my dog for a walk in NYC" → "alice has a dog", "alice lives in NYC"
 - EXPLICIT: "alice attended college" + general knowledge → "alice completed high school or equivalent"
 
+Respond with a json object in the following format:
+{{"explicit": [{{"content": "fact 1"}}, {{"content": "fact 2"}}]}}
+Each item in "explicit" must be an object with a "content" field, not a bare string.
+
 {custom_instructions_section}
 
 Target peer:

src/utils/representation.py

@@ -114,7 +114,12 @@ class PromptRepresentation(BaseModel):
     """
 
     explicit: list[ExplicitObservationBase] = Field(
-        description="Facts LITERALLY stated by the user - direct quotes or clear paraphrases only, no interpretation or inference. Example: ['The user is 25 years old', 'The user has a dog named Rover']",
+        description=(
+            "Facts LITERALLY stated by the user - direct quotes or clear "
+            "paraphrases only, no interpretation or inference. "
+            'Example: [{"content": "The user is 25 years old"}, '
+            '{"content": "The user has a dog named Rover"}]'
+        ),
         default_factory=list,
     )
 

tests/deriver/test_prompts.py

@@ -1,12 +1,27 @@
+from collections.abc import AsyncGenerator, Generator
 from unittest.mock import patch
 
 import pytest
+from pydantic import ValidationError
 
 from src.deriver.prompts import (
     estimate_deriver_prompt_tokens,
     estimate_minimal_deriver_prompt_tokens,
     minimal_deriver_prompt,
 )
+from src.utils.representation import PromptRepresentation
+
+
+@pytest.fixture(autouse=True)
+async def clean_queue_tables() -> AsyncGenerator[None, None]:
+    """Prompt-only tests do not need queue table cleanup."""
+    yield
+
+
+@pytest.fixture(autouse=True)
+def mock_tracked_db() -> Generator[None, None, None]:
+    """Prompt-only tests do not need tracked_db patching."""
+    yield
 
 
 def test_minimal_deriver_prompt_includes_custom_instructions_when_present() -> None:
@@ -30,6 +45,30 @@ def test_minimal_deriver_prompt_omits_custom_instructions_when_absent() -> None:
     assert "CUSTOM INSTRUCTIONS:" not in prompt
 
 
+def test_minimal_deriver_prompt_describes_explicit_json_object_shape() -> None:
+    prompt = minimal_deriver_prompt(
+        peer_id="alice",
+        messages="alice: I just had my 25th birthday",
+        custom_instructions=None,
+    )
+
+    assert '"explicit": [{"content": "fact 1"}' in prompt
+    assert 'Each item in "explicit" must be an object with a "content" field' in prompt
+    assert "not a bare string" in prompt
+    with pytest.raises(ValidationError):
+        PromptRepresentation.model_validate_json(
+            '{"explicit": ["alice is 25 years old"]}'
+        )
+    assert PromptRepresentation.model_validate_json(
+        '{"explicit": [{"content": "alice is 25 years old"}]}'
+    ).explicit[0].content == "alice is 25 years old"
+    explicit_schema_description = PromptRepresentation.model_json_schema()["properties"][
+        "explicit"
+    ]["description"]
+    assert '{"content": "The user is 25 years old"}' in explicit_schema_description
+    assert "['The user is 25 years old'" not in explicit_schema_description
+
+
 def test_estimate_deriver_prompt_tokens_increases_with_custom_instructions() -> None:
     base_tokens = estimate_minimal_deriver_prompt_tokens()
     custom_tokens = estimate_deriver_prompt_tokens(

tests/llm/test_backends/test_openai.py

@@ -813,6 +813,38 @@ async def test_structured_output_json_object_mode_request_shape() -> None:
     assert result.content.answer == "ok"
 
 
+@pytest.mark.asyncio
+async def test_structured_output_json_object_mode_prompt_representation_schema_shape() -> None:
+    """PromptRepresentation schema hints explicit observations as content objects."""
+    client = Mock()
+    client.chat.completions.parse = AsyncMock()
+    client.chat.completions.create = AsyncMock(
+        return_value=_structured_create_return(
+            '{"explicit": [{"content": "alice likes coffee"}]}'
+        )
+    )
+
+    backend = OpenAIBackend(client)
+    result = await backend.complete(
+        model="glm-4.6",
+        messages=[{"role": "user", "content": "Hello"}],
+        max_tokens=100,
+        response_format=PromptRepresentation,
+        extra_params={"structured_output_mode": "json_object"},
+    )
+
+    call = _await_kwargs(client.chat.completions.create)
+    system_messages = [m for m in call["messages"] if m["role"] == "system"]
+    assert system_messages, "expected a system message carrying the schema"
+    system_content = system_messages[0]["content"]
+    schema = json.loads(system_content.split("JSON schema:\n", 1)[1])
+    explicit_description = schema["properties"]["explicit"]["description"]
+    assert '{"content": "The user is 25 years old"}' in explicit_description
+    assert "['The user is 25 years old'" not in explicit_description
+    assert isinstance(result.content, PromptRepresentation)
+    assert result.content.explicit[0].content == "alice likes coffee"
+
+
 @pytest.mark.asyncio
 async def test_structured_output_json_object_mode_repairs_markdown() -> None:
     """A provider that ignores json_object and returns prose must not crash —