Cache rebuilds

Claude Code keeps the conversation in a prompt cache. A rebuild is a request that found the cache gone and had to write the whole conversation into it again. The pane lists each rebuild and the tokens it wrote.

The dimension

In the pane	Cache rebuilds
Key	none: it is not a detector
Tier	a statistic, computed by the free checks from token figures Claude Code records
On by default	yes, computed for every session; listed when a rebuild happened
Confidence	none. It is a fact about the session, never a finding or a score.

Description

Each model request reads the conversation so far from the prompt cache and writes anything new to it. A rebuild is a request that found the cache gone and wrote the whole conversation so far again.

Only rebuilds are shown; ordinary cache use is not.

What it is not:

• Not a compaction. A compaction replaces the conversation with a summary on purpose. A drop right after one is not counted.
• Not a finding. It has no confidence and no Hide button. It is listed beside the findings.

Example conversation

From the public test data of claude-code-log (MIT).

Developer: “New feature: we support .zip, but I’ve seen an archive that contains embedded tar.gz. We should support those (.tar, .tar.gz) as well…”

Agent: explores the code, asks the developer a question, writes a plan and asks for its approval, and edits the code, all in one long turn.

Turn 134, inside that turn: a model request made 69 minutes after the previous one. The previous request had read 65,720 tokens from cache. This one found them gone and wrote 54,357.

In the pane, under Cache rebuilds, the rebuild is listed as:

turn 134 · a7da6a22 · 69 min after the previous request · 54,357 written

a7da6a22 is the start of the session id. The summary line gives the number of rebuilds and the tokens written again, and opening it lists the three largest.

How it’s detected

It reads only the token figures Claude Code records for each response, which thread each response belongs to, compactions and timestamps. No message text.

Each response is counted once, on its own thread (the main conversation, or a subagent). A request is a rebuild when all three hold:

1. The previous request on that thread read more than 20,000 tokens from cache. Below that, early turns swing too much to tell.
2. This request read less than half of that, and wrote to cache.
3. There was no compaction on that thread since the previous request.

Each rebuild is recorded with its turn, the tokens written, and the time since the previous request.

Applicable fixes

There is no fix: nothing Claude Code loads or enforces prevents a rebuild. Claude Code’s one-hour cache setting (ENABLE_PROMPT_CACHING_1H) is not offered either: Claude Code lists it for API key, Bedrock, Vertex and Foundry users, not subscriptions.