src/detection/manager.rs

This file owns the DetectionManager facade: one LoopDetector and one ConvergenceDetector plus stats, and it is your engine’s single entry point into detection.

Key items

DetectionConfig ships loop threshold 3, stop threshold 10, max history 100, convergence threshold 0.95, convergence count 3, both enables true, and on_converge: Warn.

Your constructors are new, new_with_config, new_with_loop_detector, and new_with_signature, where the last copies the signature’s per-tool thresholds.

Recording runs through record_operation, which returns a DetectedPattern, plus record_tool_call and record_tool_call_with_result, plus record_response for terminal replies only.

Your pure queries are check_loop_pattern, check_convergence_pattern, and check_current_pattern, which checks loop first.

consume_pending_loop_stop() clears an unfired stop at run end so the next run isn’t killed by repetitions it never produced. Convergence state is deliberately untouched.

Finish with acknowledge_loop_warning, stats(), and reset(), which leaves config alone.

Behavior notes

A should_stop pattern recording auto-marks warned, and your engine’s fired stop clears the window.

Poison handling splits here. The manager’s own mutexes propagate LockPoisoned for convergence_detector and detection_stats, while the inner LoopDetector degrades gracefully.

Your engine’s response to a detection poison is to disable detection for the session, since detection is advisory, and never fail the run.

Stop counts are window-relative, so eviction by max_history can reset them. Raise that bound for slow loops.

Deep dives: loop · convergence.