Skip to content
loopctl

`src/detection/convergence.rs` — repeated-answer detection

Flags the agent’s terminal replies becoming near-identical, turn after turn.

Key items

  • ConvergenceConfig { enabled: true, window_size: 3, similarity_threshold: 0.95, on_converge: Warn } — validated (window ≥ 2; threshold in [0,1]).
  • ConvergenceDetector::add_response(&str) -> ConvergenceStatus — the feed; check_convergence() — pure read; clear().
  • ConvergenceStatus { detected, consecutive_count, similarity_score, similar_responses, action }.
  • ConvergenceAction — Warn (default), Stop, AskUser, Compact, SwitchPhase.
  • Similarity: word-set overlap (case-folded, punctuation stripped) computed via unit_ratio.

Behavior notes

  • The streak counts similarity to the immediately previous reply only — A/B/A/B never converges; one dissimilar reply resets to 1; an empty reply resets to 0.
  • Only terminal replies feed this detector — acting-turn preambles never count (engine-side filter).
  • Word overlap is semantically blind: paraphrases evade it, boilerplate false-positives it — hence Warn-by-default and strictly opt-in Stop/AskUser.
  • The engine maps Stop → LoopDetected, AskUser → UserInputRequired; Compact/SwitchPhase are host-executed.

Deep dive: Convergence.