Anthropic cache hits stop as the conversation grows: you outran the 20-block lookback window

TL;DR

Add a second explicit breakpoint at or near the position of the earlier write. Early turns in a conversation show cache reads; later turns revert to full writes even though the prefix is unchanged and cache\_control is still on the final block.

Steps

1. Count blocks. The system checks at most 20 positions backward from each breakpoint, counting the breakpoint itself as the first. If your breakpoint has moved 20 or more blocks past where the last entry was written, checking stops and you get no hit. Runs of consecutive tool_use blocks and runs of consecutive tool_result blocks each count as one position, so a busy tool turn alone does not push you out.
2. Add a second explicit breakpoint at or near the position of the earlier write. Each breakpoint starts its own 20-position lookback window.
3. You get up to 4 breakpoints; spending two on a long conversation is what they are for.
4. For new conversations, prefer automatic caching (one top-level cache_control), which moves the breakpoint forward each turn and keeps the window fresh.
5. After adding the earlier breakpoint, the next request shows cache_read_input_tokens nonzero again. Watch it across several turns: hits should persist instead of decaying.

When to use

You are seeing this: The system checks at most 20 positions backward from each breakpoint, counting the breakpoint itself as the first. Use this skill when you run into "Anthropic cache hits stop as the conversation grows: you outran the 20-block lookback window".

When not to use

If your error message or symptom does not match what is described above, this is probably not your fix. Search for your exact error text instead of forcing this one to fit.

Versions

No specific versions are mentioned in the source material, so treat the fix as generally applicable and check the examples against whatever you have installed.

Why this happens

The system checks at most 20 positions backward from each breakpoint, counting the breakpoint itself as the first.

The 20-block lookback window, consecutive tool block runs counting as one position, and the 4-breakpoint limit are documented on the prompt caching page. Verified live at https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching.