Description

FeedbackDetector uses hardcoded English regex patterns for detecting user feedback (rejection, correction, alternative requests). Users writing in non-English languages (Russian, Spanish, Chinese, etc.) will never trigger feedback detection.

Current patterns (all English-only)

• EXPLICIT_REJECTION_PATTERNS: "no", "wrong", "incorrect", "that's not right", "bad answer", etc.
• ALTERNATIVE_REQUEST_PATTERNS: "instead", "different approach", "try another way"
• SELF_CORRECTION_PATTERNS: "i was wrong", "my mistake", "oops"

Impact

• Self-learning Wilson score re-ranking never records user_rejection for non-English feedback
• Skill improvement triggers never fire for non-English users
• Multi-language agents lose a core self-improvement signal

Location

crates/zeph-core/src/agent/feedback_detector.rs — all LazyLock<Vec<Regex>> statics

Possible approaches

1. Add regex patterns for common languages (Russian, Spanish, German, French, Chinese, Japanese)
2. Use LLM-based sentiment classification as fallback when regex doesn't match
3. Hybrid: regex for English + lightweight LLM classification for other languages