ouroboros-ai 0.26.3.dev1__tar.gz → 0.26.3.dev2__tar.gz

{ouroboros_ai-0.26.3.dev1 → ouroboros_ai-0.26.3.dev2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: ouroboros-ai
-Version: 0.26.3.dev1
+Version: 0.26.3.dev2
 Summary: Specification-first workflow engine for AI coding agents. Works with Claude Code and Codex CLI.
 Author-email: Q00 <jqyu.lee@gmail.com>
 License-File: LICENSE

{ouroboros_ai-0.26.3.dev1 → ouroboros_ai-0.26.3.dev2}/src/ouroboros/_version.py

@@ -18,7 +18,7 @@ version_tuple: tuple[int | str, ...]
 commit_id: str | None
 __commit_id__: str | None
 
-__version__ = version = '0.26.3.dev1'
-__version_tuple__ = version_tuple = (0, 26, 3, 'dev1')
+__version__ = version = '0.26.3.dev2'
+__version_tuple__ = version_tuple = (0, 26, 3, 'dev2')
 
 __commit_id__ = commit_id = None

{ouroboros_ai-0.26.3.dev1 → ouroboros_ai-0.26.3.dev2}/src/ouroboros/evaluation/semantic.py

@@ -173,7 +173,6 @@ def parse_semantic_response(response_text: str) -> Result[SemanticResult, Valida
         "drift_score",
         "uncertainty",
         "reasoning",
-        "reward_hacking_risk",
     ]
     missing = [f for f in required_fields if f not in data]
     if missing:
@@ -185,6 +184,9 @@ def parse_semantic_response(response_text: str) -> Result[SemanticResult, Valida
             )
         )
 
+    if "reward_hacking_risk" not in data:
+        data["reward_hacking_risk"] = 0.0
+
     # Validate and clamp numeric ranges
     try:
         score = max(0.0, min(1.0, float(data["score"])))

{ouroboros_ai-0.26.3.dev1 → ouroboros_ai-0.26.3.dev2}/src/ouroboros/providers/claude_code_adapter.py

@@ -227,6 +227,25 @@ class ClaudeCodeAdapter:
         non_system_msgs = [m for m in messages if m.role != MessageRole.SYSTEM]
         system_prompt = system_msgs[0].content if system_msgs else None
 
+        # Claude Code's CLI path does not reliably honor json_schema structured
+        # output. When callers request a schema, reinforce the requirement in the
+        # prompt text and let downstream parsers extract JSON from the plain-text
+        # response rather than depending on SDK-level structured_output.
+        if config.response_format and config.response_format.get("type") == "json_schema":
+            schema = config.response_format.get("json_schema", {})
+            schema_instruction = (
+                "Respond with ONLY a valid JSON object that matches this schema. "
+                "Do not use markdown fences, headers, or explanatory text.\n\n"
+                f"JSON schema:\n{json.dumps(schema, indent=2, sort_keys=True)}"
+            )
+            if system_prompt:
+                system_prompt = f"{system_prompt}\n\n{schema_instruction}"
+            else:
+                non_system_msgs = [
+                    Message(role=MessageRole.USER, content=schema_instruction),
+                    *non_system_msgs,
+                ]
+
         # Build prompt from non-system messages only
         prompt = self._build_prompt(non_system_msgs)
 
@@ -417,13 +436,10 @@ class ClaudeCodeAdapter:
         if system_prompt:
             options_kwargs["system_prompt"] = system_prompt
 
-        # Pass structured output format if requested
-        # SDK expects: output_format={"type": "json_schema", "schema": {...}}
-        if config.response_format and config.response_format.get("type") == "json_schema":
-            options_kwargs["output_format"] = {
-                "type": "json_schema",
-                "schema": config.response_format.get("json_schema", {}),
-            }
+        # Do not pass output_format here. The Claude CLI path used by the Agent
+        # SDK currently ignores json_schema structured output constraints and may
+        # return plain text, so we enforce schema compliance via prompt text in
+        # complete() and parse the JSON from the response body.
 
         options = ClaudeAgentOptions(**options_kwargs)
 

{ouroboros_ai-0.26.3.dev1 → ouroboros_ai-0.26.3.dev2}/tests/unit/evaluation/test_semantic.py

@@ -119,8 +119,8 @@ class TestParseSemanticResponse:
         assert result.is_err
         assert "Missing required fields" in result.error.message
 
-    def test_missing_reward_hacking_risk_is_rejected(self) -> None:
-        """Omitting reward_hacking_risk must fail, not silently default to 0.0."""
+    def test_missing_reward_hacking_risk_defaults_to_zero(self) -> None:
+        """Omitting reward_hacking_risk should degrade gracefully to 0.0."""
         response = """{
             "score": 0.8,
             "ac_compliance": true,
@@ -131,8 +131,8 @@ class TestParseSemanticResponse:
         }"""
         result = parse_semantic_response(response)
 
-        assert result.is_err
-        assert "reward_hacking_risk" in str(result.error.details)
+        assert result.is_ok
+        assert result.value.reward_hacking_risk == 0.0
 
     def test_no_json_in_response(self) -> None:
         """Error when no JSON found."""

{ouroboros_ai-0.26.3.dev1 → ouroboros_ai-0.26.3.dev2}/tests/unit/providers/test_claude_code_adapter.py

@@ -229,3 +229,67 @@ class TestExecuteSingleRequestSystemPrompt:
 
         options_call_kwargs = mock_options_cls.call_args.kwargs
         assert "system_prompt" not in options_call_kwargs
+
+    @pytest.mark.asyncio
+    async def test_json_schema_is_enforced_via_prompt_not_output_format(self) -> None:
+        """json_schema requests should augment the prompt, not SDK output_format."""
+        adapter = ClaudeCodeAdapter()
+        messages = [Message(role=MessageRole.USER, content="Score this artifact")]
+        config = CompletionConfig(
+            model="claude-sonnet-4-6",
+            response_format={
+                "type": "json_schema",
+                "json_schema": {"type": "object", "properties": {"score": {"type": "number"}}},
+            },
+        )
+
+        mock_execute = AsyncMock()
+        mock_execute.return_value = MagicMock(is_ok=True)
+        adapter._execute_single_request = mock_execute
+
+        with patch.dict("sys.modules", {"claude_agent_sdk": MagicMock()}):
+            await adapter.complete(messages, config)
+
+        prompt_arg = mock_execute.call_args.args[0]
+        assert "Respond with ONLY a valid JSON object" in prompt_arg
+        assert '"score"' in prompt_arg
+
+    @pytest.mark.asyncio
+    async def test_execute_single_request_omits_output_format(self) -> None:
+        """SDK options should not include output_format for json_schema requests."""
+        adapter = ClaudeCodeAdapter()
+        config = CompletionConfig(
+            model="claude-sonnet-4-6",
+            response_format={
+                "type": "json_schema",
+                "json_schema": {"type": "object", "properties": {"score": {"type": "number"}}},
+            },
+        )
+
+        mock_options_cls = MagicMock()
+
+        async def fake_query(*args, **kwargs):
+            msg = MagicMock()
+            type(msg).__name__ = "ResultMessage"
+            msg.structured_output = None
+            msg.result = '{"score": 0.9}'
+            msg.is_error = False
+            yield msg
+
+        sdk_module = _make_sdk_mock(mock_options_cls, MagicMock(side_effect=fake_query))
+
+        with patch.dict(
+            "sys.modules",
+            {
+                "claude_agent_sdk": sdk_module,
+                "claude_agent_sdk._errors": sdk_module._errors,
+            },
+        ):
+            await adapter._execute_single_request(
+                "test prompt",
+                config,
+                system_prompt="Return JSON",
+            )
+
+        options_call_kwargs = mock_options_cls.call_args.kwargs
+        assert "output_format" not in options_call_kwargs