Problem

collector/patterns/pattern_detector.py (441 lines) has zero test coverage.

PatternDetector implements cross-session pattern detection with pure logic (no I/O, no async):

• detect_all_patterns() — groups by agent, calls per-agent detection
• detect_error_rate_trends() — compares baseline vs recent error rates
• detect_tool_failure_frequency() — compares tool failure rates
• detect_new_failure_modes() — detects new error session patterns
• detect_confidence_drops() — replay value drop detection
• _calculate_average_error_rate() / _calculate_tool_failure_rate() — metric helpers

Pattern dataclass and to_dict() are also untested.

Scope

Add tests/test_pattern_detector.py covering:

Pattern dataclass

• to_dict() roundtrip has all expected keys

Helper methods

• Empty session list returns 0.0
• Error rate: sessions with errors / total sessions
• Tool failure rate: total_errors / total_tool_calls

detect_error_rate_trends()

• No pattern when recent rate <= baseline
• Warning when increase exceeds threshold
• Critical when increase exceeds 2x threshold
• Handles zero baseline rate

detect_tool_failure_frequency()

• No pattern below threshold
• Pattern detected when rate spikes
• Severity escalation to critical

detect_new_failure_modes()

• No pattern when recent errors <= 2x baseline
• Warning when recent exceeds 2x baseline
• Critical when recent exceeds 3x baseline

detect_confidence_drops()

• No pattern when drop is below threshold
• Pattern when replay_value drops significantly

detect_all_patterns()

• Empty list returns empty
• Groups sessions by agent correctly
• Sorts critical before warning
• Returns no patterns when baseline too small (< 3 sessions)

Validation

python3 -m pytest -q tests/test_pattern_detector.py