src/detection/convergence.rs

This file flags your agent’s terminal replies becoming near-identical, turn after turn. Read it when your runs keep answering themselves.

Key items

Your ConvergenceConfig { enabled: true, window_size: 3, similarity_threshold: 0.95, on_converge: Warn } is validated: window must be at least 2, and threshold must land in [0,1].

ConvergenceDetector::add_response(&str) -> ConvergenceStatus is your feed. check_convergence() is a pure read, and clear() restarts the streak.

ConvergenceStatus { detected, consecutive_count, similarity_score, similar_responses, action } is what you read back.

ConvergenceAction offers Warn, the default, plus Stop, AskUser, Compact, and SwitchPhase.

Your similarity score measures word-set overlap, case-folded with punctuation stripped, computed through unit_ratio.

Behavior notes

The streak counts similarity to the immediately previous reply only, so A/B/A/B never converges. One dissimilar reply resets the count to 1, and an empty reply resets it to 0.

Only terminal replies feed this detector. Your engine’s filter keeps acting-turn preambles out.

Word overlap is semantically blind: paraphrases evade it and boilerplate false-positives it, which is why Warn is the default and Stop or AskUser stay strictly opt-in.

Your engine maps Stop to LoopDetected and AskUser to UserInputRequired. It leaves Compact and SwitchPhase host-executed, which means you run them.

Deep dive: Convergence.