fix: address PR #97 review comments with clarifying comments and test dep

- Add comment in reset() explaining why _external_step_control is not reset
- Add comment on hasattr guard explaining MagicMock behavior is acceptable
- Add docstring note in TestFalseNegativeRegressions about VLM response limitation
- Add flask to test optional-dependencies for CI coverage

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: abrichr

openadapt_evals/agents/claude_computer_use_agent.py

@@ -362,6 +362,8 @@ def _clamp_coord(self, x_norm: float, y_norm: float) -> tuple[float, float]:
 
     def reset(self) -> None:
         """Reset agent state between episodes."""
+        # Note: _external_step_control is not reset here because the controller
+        # that set it persists across resets
         self._messages = []
         self._step_count = 0
         self._last_tool_use_id = None

openadapt_evals/demo_controller.py

@@ -159,6 +159,8 @@ def __init__(
         # step progression is driven exclusively by VLM verification here
         # in the controller.  This prevents drift between the agent's
         # keyword-based heuristic and the controller's verifier.
+        # hasattr works correctly for real agents; MagicMock auto-creates attrs
+        # but that's fine since we're setting the value anyway
         if hasattr(agent, "_external_step_control"):
             agent._external_step_control = True
 

pyproject.toml

@@ -94,6 +94,7 @@ all = [
 ]
 test = [
     "anthropic>=0.76.0",
+    "flask>=3.0.0",
 ]
 
 [project.scripts]

tests/test_plan_verify.py

@@ -622,6 +622,10 @@ class TestFalseNegativeRegressions:
 
     These verify that the updated prompts and status model correctly handle
     cases where the old verifier would produce false negatives.
+
+    Note: These tests validate the parsing pipeline, not the VLM's actual
+    response to the new prompts. Live eval is needed to validate prompt
+    effectiveness.
     """
 
     # -- Scenario 1: Header typed correctly, cursor moved after entry ------