Repeated social interaction is rarely reducible to material payoff alone: the same outcome can differ in fairness, relationship meaning, vulnerability, affective activation, and regulatory cost. We present an interpretable dual-system architecture for emotionally bounded-rational agents in repeated games. A game-to-values adapter maps external outcomes into material, fairness, relationship, and safety/no-regret event values; delta-based appraisal, mood, fatigue, and computational wellbeing then update the agent’s internal state. A fast affective System 1 proposes action tendencies, while a slower System 2 can accept, override, or refocus them. We evaluate the implementation on 20 independent episodes of 200 rounds per condition in the Prisoner’s Dilemma, Battle of the Sexes, and Ultimatum Game. The experiments compare a diagnostic-control mode, where adapter/appraisal quantities are logged for audit, with an integrated mode, where they participate in state updates. Under the specified internal-state aggregation, the integrated mode yields higher state scores and lower negative-state share across the tested scenarios, while still exposing exploitation, coordination failure, and game-specific stress-test artifacts. The study is positioned as a reproducible layer of a broader programme: it defines the ablations, sensitivity checks, stronger baselines, and observer-validation steps needed to validate the proposed internal-state interpretation.