Replay

Every setting comes with the same question, and no documentation can answer it: what would this have done to my work? A number measured on somebody else's repository is not an answer.

tokunseba replay --since 7d --tier 2

312 requests replayed under tier 2.
                    tokens removed
as it ran                  1.41M
as configured here         2.08M
difference                  +674k

transform        blocks
outline:python       46
outline:ts           18
summary:pytest       31

84 of 312 requests would come out different.

What it doesCopy Anchor Link

Re-runs traffic that already happened under a configuration you have not committed to, from the request bodies already on disk.

• Nothing is sent anywhere. No provider is contacted. There is no network involved at all.
• Nothing of yours is written. The replay gets its own ledger and its own blob store in a temporary directory, which is deleted when it finishes. It cannot leave behind a handle, an event, a frozen replacement, or a learned token ratio.

OptionsCopy Anchor Link

tokunseba replay --tier 1 --tier 2
tokunseba replay --set thresholds.truncate_lines=120
tokunseba replay --set thresholds.truncate_tokens=3000 --verbose

--verbose gives you request ids, so you can read any one of them in full:

tokunseba explain 1f4c9a20e8bb

Reading the baseline honestlyCopy Anchor Link

With nothing overridden, the two rows will usually still differ, and the output says why.

A turn whose earlier identical copy falls outside --since or --limit has nothing to refer back to inside the window, so it gets summarised instead of deduplicated, and removes less. That is the window talking, not the configuration.

Compare two runs of --set against each other rather than against a baseline of zero.

What is not replayedCopy Anchor Link

Tier 3 is deliberately out of scope. It decides which model answers, and no offline pass can know what a different model would have said.

For that question:

tokunseba advise --since 30d

which judges the opening prompts of conversations you have already had and reports how many would have moved.

What it needsCopy Anchor Link

Replay reads stored request bodies. Those have their own retention window (retention.keep_bodies_days, 14 days by default) because bodies are the bulkiest thing tokunseba stores, and they are not stored at all if proxy.store_bodies is off.

tokunseba retention show

tells you how far back you can actually replay.