• Run cache impact analysis using the REAL trimmer instead of estimation.

    Trims the session to a temp file, measures actual byte reduction, computes the token reduction ratio from real trimmer output, then cleans up. Much more accurate than estimatePostTrimTokens().

    Parameters

    • jsonlPath: string
    • model: "haiku" | "sonnet" | "opus" | "opus-4" = 'sonnet'
    • cacheHitRate: number = 0.90

    Returns Promise<RealTrimResult>