If the agent bill went up and nothing errored, a prompt cache broke.

Replay Doctor reads the transcripts already on your disk and names the turn it broke on, the cause, and the tokens re-billed at write prices. Expiry is only one of the causes it can name.

On a subscription seat the dollars are list price for someone else. The tokens are yours: a re-billed token is context the work did not get. One machine proves the method and nothing about the market; the pool below is how that changes.

Set promptCacheTtl to 1h in Claude Code settings, keep the tool list stable mid-session, and you are done. This page is not for you, and that is fine.

Run it once. The diff tells you whether the wall is output, compaction, or silent re-reads.

This is for you. Every re-bill is an invoice line, and the tool names the turn.

Add your machine to the pool. Under 600 bytes, nothing about you.

Replay writes one file: sixteen counts and ratios, three strings naming the build that priced them, no paths and no account. You read it, then you post it yourself; the binary cannot. The pool is summed by replay pool in public, and every row on the roster is a file you can download and re-hash.

1 machine · 121 sessions · pooled 2026-09-12

mkdir -p ~/.config/replay && printf 'corpus_opt_in = true\n' > ~/.config/replay/corpus-consent.toml replay cost --contribute launch-2026-09 --contribute-dir . ~/.claude/projects curl -sS -X POST --data-binary @replay-corpus-*.json "https://replay.doctor/api/contribute?campaign=launch-2026-09"

--contribute needs 0.6.0 or later: go install github.com/RedRobotKK/Replay/cmd/replay@latest, or a signed release. Read the file before the third line; the reply names your digest.

Four commands. Each one prints what it measured, and refuses what it did not.

Every command, generated from the binaryNames the turn the cache broke on, what changed, and the tokens it cost.

What is filling your context, ranked, and by how much the ranking overstates after a compaction.

Prices the switch itself and names the turn where the cheaper model becomes cheaper.

Writes under 600 bytes: 16 numbers and 3 strings naming the build, no paths. This is how one machine stops being the sample.

Diagnosis is free and stays free. The week is the only thing for sale: you get the invoice annotated turn by turn, every break named with its cause, the layout changes worth making and the ones that are not, and the write-up, which is yours. As of 2026-09-13 the week is $22,000, quoted at $18,000 to $25,000 by scope, and none has been sold yet; that sentence changes the day one is.

Talk to DanielHear what happens next.

Replay is free and installs in one line, so there is nothing here to wait for. This list is for the two things that are not available yet: word when the tool gains something worth knowing about, and the forensic engagement, which is one person's calendar and is therefore genuinely finite.

No schedule, because there has never been one. One confirmation email, then nothing until there is something. Leaving is one click and needs no reply.

Or sign in, which skips the confirmation email because your provider has already verified the address.

Only your verified email address is read. No profile, no name, no repositories, and the token is used once and never stored.