Analyzes unit economics of the closed-loop controller architecture:
Claude agent costs, VLM verifier costs, scaling projections, and a
three-phase strategy from loop-as-product to trained-model-as-product.

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

pyautogui.drag() uses relative coordinates that compound with starting
position errors, making it unreliable for small targets like LibreOffice
fill handles (~3x3 pixels). Replace with explicit mouseDown/moveTo/mouseUp
sequence with timing delays for reliable drag operations.

Also adds drag case to _build_pixel_command() for the pixel_action() path.

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

…al verification

Three issues addressed:

1. Heuristic/verifier step drift: The agent's keyword-based
   _advance_plan_steps() heuristic and the DemoController's VLM verifier
   operated on independent state, allowing them to disagree on which step
   was current. Fix: add _external_step_control flag to the agent that the
   DemoController sets at init, making _advance_plan_steps() a no-op when
   the controller manages step progression via VLM verification.

2. partially_verified invisible to goal verification: When steps were
   marked partially_verified, the final goal verification pass had no
   visibility into which steps had partial completions. Fix: _verify_goal()
   now builds a step verification summary and augments the goal text with
   it when noteworthy statuses (partially_verified, failed) exist.

3. Missing integration tests: Added TestHeuristicVerifierSync (4 tests)
   and TestGoalVerificationContext (5 tests) that verify the heuristic is
   properly disabled under controller management, step advancement is
   driven by VLM verification, and partial/failed step context reaches
   goal verification. Also added 2 agent-level tests for
   _external_step_control behavior.

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

When DemoController sets _external_step_control=True, the agent's
internal plan progress injection and done-override logic now become
no-ops. This prevents the agent from sending conflicting step-tracking
signals to the Claude model (agent says "step 1 in progress" while
controller says "step 3 is current").

Three specific suppressions:
1. _build_initial_messages skips plan progress text injection
2. Follow-up messages skip plan progress / demo re-injection
3. Premature "done" override is left to the controller

Adds integration tests exercising agent+controller interaction:
- Agent suppresses progress under external control
- Agent injects progress normally without external control
- Controller's augmented task instruction reaches the agent
- Done override handled by controller, not agent

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

After WAA setup (close_all → verify_apps → download → open), the
target application may be behind other windows, still loading, or
obscured by notifications. This wastes 6+ agent steps recovering.

Add _ensure_app_focused() with multi-strategy approach:
- Maps task related_apps to window title patterns
- Uses WAA /setup/activate_window endpoint (same as WAA postconfig)
- Falls back to Alt+Tab
- Retries 3x with increasing delays (2s, 3s, 5s)
- Verifies foreground window title via pygetwindow on VM
- Runs during reset(), does NOT count against agent step budget

Also adds _APP_WINDOW_PATTERNS mapping, _get_expected_window_patterns(),
_check_foreground_matches(), and _normalize_app_name() helpers.

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

Comprehensive analysis of GUI agent failure modes with taxonomy,
recording system design, training viability assessment, and
prioritized action plan. Key findings:

- 4-category taxonomy: Environment, Agent Planning, Grounding, Verifier
- Existing ExecutionTraceCollector needs only minor extensions
- SFT on 50-100 corrected trajectories expected 10-30pp improvement
- Deterministic infrastructure fixes should come first (Tier 1)

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

… 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>