"""Phase W - The Abyss safety / moderation layer (self-hosted, from scratch). Layers: input moderation, output moderation, agent tool-call screening, and an audit trail - with tiered-by-severity enforcement (low -> log, medium -> soft refusal, high -> block). Engine = weighted lexical rules + co-occurrence, no external API. """ from server.safety import config from server.safety import rules from server.safety import classifier from server.safety import injection from server.safety import audit from server.safety.moderator import ( ModerationResult, moderate_input, moderate_output, ) from server.safety.toolguard import screen_tool_call SAFETY_ENABLED = config.SAFETY_ENABLED __all__ = [ "config", "rules", "classifier", "injection", "audit", "ModerationResult", "moderate_input", "moderate_output", "screen_tool_call", "SAFETY_ENABLED", ]