"""Prompt caching with the official Python SDK (anthropic 1.7): explicit breakpoint on a ~5.8k-token system
block, then a second identical call, then a third call using top-level *automatic* caching and a 1h TTL variant
on a fresh prefix.
STATUS: LIVE_VERIFIED 2026-09-18 (claude-haiku-4-5-20251001): cold probe write 5809 → read 5809 → auto: read + create 7;
example run (prefix already warm from the .sh example): read 5251 on the first two calls, auto read 5251 + create 7;
1h variant: cache_creation.ephemeral_1h_input_tokens 5809. ≈$0.02.
Run: .venv/bin/python examples/anthropic/prompt-caching/cache_system_prompt.py
"""
from __future__ import annotations

import os
import sys
from pathlib import Path

sys.path.insert(0, str(Path(__file__).resolve().parents[3]))
from scripts import live  # noqa: E402  loads .env, log_request

import anthropic  # noqa: E402

MODEL = os.environ.get("ANTHROPIC_MODEL", "claude-haiku-4-5-20251001")
client = anthropic.Anthropic(max_retries=1)


def filler(seed: str, n: int = 150) -> str:
    return " ".join(f"Section {seed}-{i}: The quarterly logistics review covers warehouse throughput, carrier performance, "
                    "customs documentation, cold-chain compliance, and regional demand forecasting for the coming period." for i in range(n))


def show(label: str, msg: anthropic.types.Message) -> None:
    u = msg.usage
    print(f"{label}: input={u.input_tokens} creation={u.cache_creation_input_tokens} read={u.cache_read_input_tokens} "
          f"detail={u.cache_creation.model_dump() if u.cache_creation else None}")
    live.log_request("anthropic", "POST", "/v1/messages", 200, 0.004, f"example prompt-caching/cache_system_prompt.py {label} {MODEL}")


system_a = [{"type": "text", "text": filler("A"), "cache_control": {"type": "ephemeral"}}]
user = [{"role": "user", "content": "Reply with OK."}]
show("explicit write", client.messages.create(model=MODEL, max_tokens=8, system=system_a, messages=user))
show("explicit read ", client.messages.create(model=MODEL, max_tokens=8, system=system_a, messages=user))
# Automatic caching: one top-level cache_control, no block markers. Lookback finds the system entry written above.
show("automatic     ", client.messages.create(model=MODEL, max_tokens=8, system=[{"type": "text", "text": filler("A")}],
                                              messages=user, cache_control={"type": "ephemeral"}))
# 1-hour TTL (2x write price) on a fresh prefix so the write is visible under cache_creation.ephemeral_1h_input_tokens.
system_b = [{"type": "text", "text": filler("B"), "cache_control": {"type": "ephemeral", "ttl": "1h"}}]
show("1h write      ", client.messages.create(model=MODEL, max_tokens=8, system=system_b, messages=user))