Skip to content

CubeLoop — a Pythonic, async-native agent harness core designed for high performance, readability, and production-grade persistence.

fix: address 6 code review findings

xf gong5a8569686a0e
diff --git a/cubepi/agent/loop.py b/cubepi/agent/loop.py
index c67accd..1171505 100644
--- a/cubepi/agent/loop.py
+++ b/cubepi/agent/loop.py
@@ -30,6 +30,8 @@ from cubepi.providers.base import (
if TYPE_CHECKING:
from cubepi.providers.fallback import FallbackBoundModel
+_MAX_ON_RUN_END_CYCLES = 100
+
async def run_agent_loop(
*,
@@ -475,6 +477,8 @@ async def _run_loop_inner(
current_context.messages.append(msg)
new_messages.append(msg)
+ on_run_end_cycles = 0
+
while True:
has_more_tool_calls = True
@@ -663,6 +667,9 @@ async def _run_loop_inner(
if on_run_end:
inject = await on_run_end(current_context, signal=opts.signal)
if inject:
+ on_run_end_cycles += 1
+ if on_run_end_cycles > _MAX_ON_RUN_END_CYCLES:
+ break
for msg in inject:
await emit_event(emit, MessageStartEvent(message=msg))
await emit_event(emit, MessageEndEvent(message=msg))
diff --git a/cubepi/middleware/goal.py b/cubepi/middleware/goal.py
index 0f005fc..3ce9822 100644
--- a/cubepi/middleware/goal.py
+++ b/cubepi/middleware/goal.py
@@ -19,6 +19,8 @@ from cubepi.providers.base import (
BoundModel,
Message,
TextContent,
+ ToolCall,
+ ToolResultMessage,
UserMessage,
)
@@ -49,12 +51,21 @@ def _format_messages_for_eval(messages: list[Message], window: int) -> str:
parts: list[str] = []
for msg in tail:
role = getattr(msg, "role", "unknown")
- texts: list[str] = []
+ fragments: list[str] = []
+ if isinstance(msg, ToolResultMessage):
+ for block in msg.content:
+ if isinstance(block, TextContent):
+ fragments.append(block.text)
+ if fragments:
+ parts.append(f"[tool_result:{msg.tool_name}] {' '.join(fragments)}")
+ continue
for block in getattr(msg, "content", []):
if isinstance(block, TextContent):
- texts.append(block.text)
- if texts:
- parts.append(f"[{role}] {' '.join(texts)}")
+ fragments.append(block.text)
+ elif isinstance(block, ToolCall):
+ fragments.append(f"[called {block.name}]")
+ if fragments:
+ parts.append(f"[{role}] {' '.join(fragments)}")
return "\n".join(parts)
@@ -99,9 +110,11 @@ class GoalMiddleware(Middleware):
self._evaluations = 0
rewritten = last.model_copy(update={"content": [TextContent(text=condition)]})
- # Mutate ctx.messages so the stripped form is what gets stored in
- # agent state — the /goal prefix is a one-way control signal that
- # should not persist in the conversation history.
+ # Strip the /goal prefix from the in-memory message list so the
+ # worker model never sees it. Note: the checkpointer may have
+ # already persisted the raw message via MessageEndEvent — the
+ # prefix in stored history is harmless (it won't be the last
+ # message on restore, so transform_context won't re-trigger).
ctx.messages[-1] = rewritten
return [*messages[:-1], rewritten]
@@ -144,11 +157,13 @@ class GoalMiddleware(Middleware):
if result.achieved:
goal_state["status"] = "achieved"
ctx.extra["goal"] = goal_state
+ self._condition = None
return None
if self._evaluations >= self._max_evaluations:
goal_state["status"] = "exhausted"
ctx.extra["goal"] = goal_state
+ self._condition = None
return None
goal_state["status"] = "active"
diff --git a/cubepi/providers/anthropic.py b/cubepi/providers/anthropic.py
index 8dd1737..72fe3c1 100644
--- a/cubepi/providers/anthropic.py
+++ b/cubepi/providers/anthropic.py
@@ -192,7 +192,7 @@ class AnthropicProvider(BaseProvider):
]
else:
kwargs["system"] = system_prompt
- if tools:
+ if tools and tool_choice != "none":
api_tools = [self._convert_tool(t) for t in tools]
if cache_control and api_tools and self._cache_policy.mark_last_tool():
api_tools[-1]["cache_control"] = cache_control
diff --git a/cubepi/providers/base.py b/cubepi/providers/base.py
index 9103c6b..8af9a17 100644
--- a/cubepi/providers/base.py
+++ b/cubepi/providers/base.py
@@ -878,16 +878,19 @@ class BoundModel:
return output_type.model_validate(block.arguments)
except Exception as exc:
last_error = exc
+ error_text = (
+ f"Validation error: {exc}. Fix the data and try again."
+ )
attempt_messages = [
*attempt_messages,
response,
- UserMessage(
- content=[
- TextContent(
- text=f"Validation error: {exc}. Fix the data and try again."
- )
- ]
+ ToolResultMessage(
+ tool_call_id=block.id,
+ tool_name=block.name,
+ content=[TextContent(text=error_text)],
+ is_error=True,
),
+ UserMessage(content=[TextContent(text=error_text)]),
]
break
else:
diff --git a/tests/providers/test_tool_choice.py b/tests/providers/test_tool_choice.py
index 0a0b770..d5fb73f 100644
--- a/tests/providers/test_tool_choice.py
+++ b/tests/providers/test_tool_choice.py
@@ -247,6 +247,7 @@ class TestAnthropicStreamToolChoice:
call_kwargs = mock_client.messages.stream.call_args[1]
assert "tool_choice" not in call_kwargs
+ assert "tools" not in call_kwargs
# ---------------------------------------------------------------------------
diff --git a/tests/test_goal.py b/tests/test_goal.py
index 09bd4b6..76d06fd 100644
--- a/tests/test_goal.py
+++ b/tests/test_goal.py
@@ -5,12 +5,14 @@ from __future__ import annotations
from cubepi import Agent
from cubepi.agent.types import AgentContext
-from cubepi.middleware.goal import GoalMiddleware
+from cubepi.middleware.goal import GoalMiddleware, _format_messages_for_eval
from cubepi.providers.base import (
AssistantMessage,
ImageContent,
Message,
TextContent,
+ ToolCall,
+ ToolResultMessage,
UserMessage,
)
from cubepi.providers.faux import FauxProvider, faux_assistant_message, faux_tool_call
@@ -353,3 +355,67 @@ async def test_goal_state_restored_from_extra() -> None:
assert agent2.state.extra["goal"]["status"] == "achieved"
assert agent2.state.extra["goal"]["evaluations"] == 2
+
+
+# ---------------------------------------------------------------------------
+# Test 9: Stale _condition cleared after goal completes
+# ---------------------------------------------------------------------------
+
+
+async def test_condition_cleared_after_achieved() -> None:
+ """After goal achieved, subsequent non-goal prompt does NOT fire the evaluator."""
+ worker = FauxProvider(provider_id="worker")
+ worker.set_responses(
+ [
+ faux_assistant_message("goal work"),
+ faux_assistant_message("normal work"),
+ ]
+ )
+
+ evaluator_provider = FauxProvider(provider_id="evaluator")
+ evaluator_provider.set_responses(
+ [
+ faux_assistant_message(
+ faux_tool_call(
+ "structured_output",
+ {"achieved": True, "reason": "done"},
+ )
+ )
+ ]
+ )
+
+ goal = GoalMiddleware(evaluator=evaluator_provider.model("eval"))
+ agent = Agent(model=worker.model("test"), middleware=[goal])
+
+ await agent.prompt("/goal all tests pass")
+ assert agent.state.extra["goal"]["status"] == "achieved"
+ assert evaluator_provider.call_count == 1
+
+ await agent.prompt("refactor auth.py")
+ assert evaluator_provider.call_count == 1 # evaluator NOT called again
+
+
+# ---------------------------------------------------------------------------
+# Test 10: _format_messages_for_eval includes tool calls and results
+# ---------------------------------------------------------------------------
+
+
+def test_format_messages_includes_tool_calls() -> None:
+ """Evaluator transcript includes tool call names and tool result text."""
+ messages: list[Message] = [
+ UserMessage(content=[TextContent(text="run the tests")]),
+ AssistantMessage(
+ content=[ToolCall(id="tc-1", name="run_tests", arguments={"cmd": "pytest"})]
+ ),
+ ToolResultMessage(
+ tool_call_id="tc-1",
+ tool_name="run_tests",
+ content=[TextContent(text="5 passed, 0 failed")],
+ ),
+ AssistantMessage(content=[TextContent(text="All tests pass now.")]),
+ ]
+ transcript = _format_messages_for_eval(messages, window=10)
+
+ assert "[called run_tests]" in transcript
+ assert "[tool_result:run_tests] 5 passed, 0 failed" in transcript
+ assert "All tests pass now." in transcript
diff --git a/tests/test_structured_output.py b/tests/test_structured_output.py
index 45f7e7b..6a3a21b 100644
--- a/tests/test_structured_output.py
+++ b/tests/test_structured_output.py
@@ -137,3 +137,44 @@ async def test_generate_structured_retry_succeeds() -> None:
assert result.title == "Dune"
assert result.rating == 8
assert provider.call_count == 2
+
+
+async def test_generate_structured_retry_includes_tool_result_message() -> None:
+ """Retry path inserts a ToolResultMessage between the failed assistant message and the retry prompt."""
+ provider = FauxProvider()
+
+ captured_messages: list = []
+
+ def capture_and_respond(messages, model):
+ captured_messages.clear()
+ captured_messages.extend(messages)
+ return faux_assistant_message(
+ faux_tool_call(
+ "structured_output",
+ {"title": "Dune", "rating": 8, "summary": "ok"},
+ ),
+ stop_reason="tool_use",
+ )
+
+ invalid_response = faux_assistant_message(
+ faux_tool_call(
+ "structured_output",
+ {"title": "Dune", "rating": "bad", "summary": "ok"},
+ ),
+ stop_reason="tool_use",
+ )
+ provider.set_responses([invalid_response, capture_and_respond])
+
+ model = provider.model("faux-1")
+ messages = [UserMessage(content=[TextContent(text="Review Dune")])]
+ result = await model.generate_structured(MovieReview, messages, max_retries=1)
+
+ assert result.rating == 8
+ # The retry call should have: original UserMessage, failed AssistantMessage,
+ # ToolResultMessage (error), UserMessage (retry prompt)
+ from cubepi.providers.base import ToolResultMessage
+
+ tool_results = [m for m in captured_messages if isinstance(m, ToolResultMessage)]
+ assert len(tool_results) == 1
+ assert tool_results[0].is_error is True
+ assert "Validation error" in tool_results[0].content[0].text