Reports

Everything is counted in tokens, ratios and time. Those mean the same thing whatever kind of access you have and whoever you have it from, so tokunseba reports nothing else and asks you to configure no rate table to read its own output.

statusCopy Anchor Link

tokunseba status

The smallest answer: is it running, and what happened today.

statsCopy Anchor Link

tokunseba stats
tokunseba stats --since 30d
tokunseba stats --project ~/work/api
tokunseba stats --json
tokunseba stats --ab

Tokens kept off the wire, by tool, with every signal the proxy logged.

--ab compares the tier 3 control and treatment arms: requests per session, input and output tokens per session, latency per request. This is the only figure that is about your traffic rather than someone else's, and it is the one worth trusting.

topCopy Anchor Link

tokunseba top
tokunseba top --since 24h --limit 30

Where the tokens went. Two tables:

• The biggest reductions. Transform kind, before, after, and the handle that holds the original.
• The largest passthroughs. What could not be reduced, and roughly why. This table is deliberately as prominent as the first one.

reportCopy Anchor Link

tokunseba report
tokunseba report --since 30d
tokunseba report --project ~/work/api
tokunseba report --json
tokunseba report --save report.txt

The whole window on one page:

• tokens kept off the wire per day and per hour
• where the compressible mass is, by tool-result size
• transforms by kind
• savings by tool and by model
• the biggest untouched blocks
• every signal with an explanation of what it means
• what you asked about, by domain and by difficulty
• context growth across the newest session

explain and expandCopy Anchor Link

tokunseba explain <request-id>
tokunseba expand <handle>

One request, original beside replacement, block by block. Then the full original text behind any handle. See Handles and recovery.

adviseCopy Anchor Link

tokunseba advise --since 30d --to claude-haiku-4-5

Runs the judge offline over conversations that already happened and reports what your prompts looked like and what routing them would have moved. It sends nothing, adds no latency and cannot change an answer.

It ends with a plain statement of where the judge is reliable on your machine and where it is not, which is more useful than the numbers above it.

HistoryCopy Anchor Link

tokunseba retention show      # the window, and how much is actually stored
tokunseba retention keep 1y   # 24h, 30d, 6w, 6m, 1y, or forever
tokunseba prune --days 30     # delete everything older than that, now

Default is 90 days for the ledger and 14 days for stored request bodies, pruned automatically once a day. See Retention.