ouroboros-ai 0.28.4.dev2__tar.gz → 0.28.5.dev1__tar.gz

{ouroboros_ai-0.28.4.dev2 → ouroboros_ai-0.28.5.dev1}/.claude-plugin/marketplace.json

@@ -10,7 +10,7 @@
     {
       "name": "ouroboros",
       "description": "Transform vague ideas into validated seed specifications through Socratic questioning and ontological analysis",
-      "version": "0.28.3",
+      "version": "0.28.4",
       "author": {
         "name": "Q00",
         "email": "jqyu.lee@gmail.com"

{ouroboros_ai-0.28.4.dev2 → ouroboros_ai-0.28.5.dev1}/.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "ouroboros",
-  "version": "0.28.3",
+  "version": "0.28.4",
   "description": "Self-improving AI workflow system. Crystallize requirements before execution with Socratic interview, ambiguity scoring, and 3-stage evaluation.",
   "author": {
     "name": "Q00",

{ouroboros_ai-0.28.4.dev2 → ouroboros_ai-0.28.5.dev1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: ouroboros-ai
-Version: 0.28.4.dev2
+Version: 0.28.5.dev1
 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-0.28.4.dev2 → ouroboros_ai-0.28.5.dev1}/skills/setup/SKILL.md

@@ -246,7 +246,7 @@ A backup will be created: CLAUDE.md.bak
 **If "Preview first", show:**
 ````markdown
 <!-- ooo:START -->
-<!-- ooo:VERSION:0.28.3 -->
+<!-- ooo:VERSION:0.28.4 -->
 # Ouroboros — Specification-First AI Development
 
 > Before telling AI what to build, define what should be built.

{ouroboros_ai-0.28.4.dev2 → ouroboros_ai-0.28.5.dev1}/src/ouroboros/_version.py

@@ -18,7 +18,7 @@ version_tuple: tuple[int | str, ...]
 commit_id: str | None
 __commit_id__: str | None
 
-__version__ = version = '0.28.4.dev2'
-__version_tuple__ = version_tuple = (0, 28, 4, 'dev2')
+__version__ = version = '0.28.5.dev1'
+__version_tuple__ = version_tuple = (0, 28, 5, 'dev1')
 
 __commit_id__ = commit_id = None

{ouroboros_ai-0.28.4.dev2 → ouroboros_ai-0.28.5.dev1}/src/ouroboros/mcp/server/adapter.py

@@ -802,6 +802,15 @@ def create_ouroboros_server(
     )
 
     # Create shared LLM adapter for interview/seed/evaluation paths.
+    #
+    # NOTE: ``max_turns=1`` is appropriate for interview question generation
+    # and seed synthesis, where the adapter returns a single-shot response.
+    # Evaluation (Stage 2 semantic) DOES NOT use this adapter — see
+    # ``EvaluateHandler.handle`` in ``mcp/tools/evaluation_handlers.py`` which
+    # constructs a fresh adapter with ``max_turns=20`` so the semantic
+    # evaluator can issue tool calls when reading spec files. Do not bump
+    # this value without first auditing every caller of ``self.llm_adapter``
+    # — it affects interview latency and token usage.
     llm_adapter = create_llm_adapter(
         backend=llm_backend,
         max_turns=1,

{ouroboros_ai-0.28.4.dev2 → ouroboros_ai-0.28.5.dev1}/src/ouroboros/mcp/tools/evaluation_handlers.py

@@ -402,7 +402,23 @@ class EvaluateHandler:
             # Use injected or create services.
             # Evaluation reads multiple spec files (one Read call per AC), so
             # max_turns must be well above 1.
-            llm_adapter = self.llm_adapter or create_llm_adapter(
+            #
+            # IMPORTANT: The shared ``self.llm_adapter`` that may have been
+            # injected by ``build_mcp_server`` (see ``mcp/server/adapter.py``)
+            # is constructed with ``max_turns=1`` for interview / seed-generation
+            # use cases. That value is fatal for evaluation because the Stage 2
+            # semantic evaluator issues at least one ``Read`` tool call per AC to
+            # inspect spec files. With ``max_turns=1`` the Agent SDK returns
+            # ``error_max_turns`` on the very first tool call, surfacing as
+            # ``Command failed with exit code 1`` at the MCP tool boundary.
+            #
+            # To ensure the evaluator always has enough turns, construct a
+            # fresh adapter here rather than relying on the injected instance.
+            # The old ``self.llm_adapter or create_llm_adapter(...)`` pattern
+            # short-circuited on the injected adapter and silently preserved
+            # the ``max_turns=1`` ceiling — see issue #305 for the user-facing
+            # failure it produced.
+            llm_adapter = create_llm_adapter(
                 backend=self.llm_backend,
                 max_turns=20,
             )

{ouroboros_ai-0.28.4.dev2 → ouroboros_ai-0.28.5.dev1}/tests/unit/mcp/tools/test_evaluation_handler.py

@@ -135,14 +135,89 @@ class TestEvaluateHandlerAdapterCreation:
             "at least one turn per AC file read. Use max_turns >= 10."
         )
 
-    async def test_injected_adapter_skips_create_llm_adapter(self):
-        """When an adapter is injected (e.g. in tests), create_llm_adapter is not called."""
-        handler = EvaluateHandler(llm_adapter=_make_mock_adapter())
+    async def test_injected_adapter_is_ignored_for_evaluation(self):
+        """Evaluation ALWAYS creates a fresh adapter, even when one is injected.
+
+        Regression for issue #305 / related max_turns fix:
+
+        The shared adapter wired up in ``build_mcp_server``
+        (``mcp/server/adapter.py``) is constructed with ``max_turns=1`` because
+        interview and seed-generation paths only need a single-shot response.
+        If the evaluator reuses that adapter, the very first ``Read`` tool
+        call issued by the Stage 2 semantic evaluator hits ``error_max_turns``
+        and surfaces as ``Command failed with exit code 1`` at the MCP tool
+        boundary.
+
+        To prevent that, the handler now ignores ``self.llm_adapter`` and
+        always constructs a fresh adapter with ``max_turns=20``.
+        """
+        captured: dict = {}
+
+        def _capture(**kwargs):
+            captured.update(kwargs)
+            return _make_mock_adapter()
+
+        # Simulate the shared adapter that build_mcp_server would inject.
+        shared_adapter = _make_mock_adapter()
+        handler = EvaluateHandler(llm_adapter=shared_adapter)
 
         with (
-            patch("ouroboros.mcp.tools.evaluation_handlers.create_llm_adapter") as mock_create,
+            patch(
+                "ouroboros.mcp.tools.evaluation_handlers.create_llm_adapter",
+                side_effect=_capture,
+            ),
             _patch_pipeline(),
         ):
             await handler.handle(_BASE_ARGUMENTS)
 
-        mock_create.assert_not_called()
+        # Fresh adapter was created with a sufficient max_turns budget, even
+        # though one was already injected on the handler.
+        assert "max_turns" in captured, (
+            "create_llm_adapter was not called — the handler must build a "
+            "fresh adapter for evaluation instead of reusing the shared "
+            "interview-tuned adapter."
+        )
+        assert captured["max_turns"] >= 10, f"max_turns={captured['max_turns']} is too low."
+
+    async def test_injected_max_turns_1_adapter_does_not_leak(self):
+        """Regression for issue #305: injecting a max_turns=1 adapter must not
+        cause evaluation to inherit the pathological ceiling.
+
+        This guards against a future refactor re-introducing the
+        ``self.llm_adapter or create_llm_adapter(...)`` short-circuit pattern.
+        """
+        # The shared MCP adapter is built with max_turns=1; verify that the
+        # handler does not forward that adapter into the evaluation pipeline.
+        low_turn_adapter = _make_mock_adapter()
+        low_turn_adapter._max_turns = 1  # mirror ClaudeCodeAdapter attribute
+        handler = EvaluateHandler(llm_adapter=low_turn_adapter)
+
+        captured_pipeline_adapters: list = []
+
+        def _capture_pipeline(llm_adapter, config):  # noqa: ARG001
+            mock_pipeline = MagicMock()
+            mock_pipeline.evaluate = AsyncMock(
+                return_value=Result.ok(_make_approved_eval_result()),
+            )
+            captured_pipeline_adapters.append(llm_adapter)
+            return mock_pipeline
+
+        with (
+            patch(
+                "ouroboros.mcp.tools.evaluation_handlers.create_llm_adapter",
+                return_value=_make_mock_adapter(),
+            ),
+            patch(
+                "ouroboros.evaluation.EvaluationPipeline",
+                side_effect=_capture_pipeline,
+            ),
+        ):
+            await handler.handle(_BASE_ARGUMENTS)
+
+        assert captured_pipeline_adapters, "EvaluationPipeline was not constructed"
+        pipeline_adapter = captured_pipeline_adapters[0]
+        assert pipeline_adapter is not low_turn_adapter, (
+            "EvaluationPipeline was constructed with the injected max_turns=1 "
+            "adapter — this is the bug described in issue #305. The handler "
+            "must build a fresh adapter with a higher max_turns budget."
+        )