        sized = self.tokenize(messages)
        total = sum(tokens for tokens, _msg in sized)
        if total <= self.max_tokens and depth == 0:
            return messages

        min_split = 4
        if len(messages) <= min_split or depth > 3:
            return self.summarize_all(messages)

        tail_tokens = 0
        split_index = len(messages)
        half_max_tokens = self.max_tokens // 2

        # Iterate over the messages in reverse order
        for i in range(len(sized) - 1, -1, -1):
            tokens, _msg = sized[i]
            if tail_tokens + tokens < half_max_tokens:
                tail_tokens += tokens
                split_index = i
            else:
                break

        # Ensure the head ends with an assistant message
        while messages[split_index - 1]["role"] != "assistant" and split_index > 1:
            split_index -= 1

        if split_index <= min_split:
            return self.summarize_all(messages)

        # Split head and tail
        tail = messages[split_index:]
…
        summary = self.summarize_all(keep)

        # If the combined summary and tail still fits, return directly
        summary_tokens = self.token_count(summary)
        tail_tokens = sum(tokens for tokens, _ in sized[split_index:])
        if summary_tokens + tail_tokens < self.max_tokens:
            return summary + tail

        # Otherwise recurse with increased depth
        return self.summarize_real(summary + tail, depth + 1)

Warm-up