Proving it works
The difference between configured and actually in the path, and how to tell.
There are two separate questions, and confusing them is the single most common way to end up staring at a report full of zeros.
| Question | Command | What it reads |
|---|---|---|
| Is it configured? | tokunseba doctor | Settings files, environment variables, the service |
| Is it actually being used? | tokunseba verify | The ledger: requests that really arrived |
doctor
tokunseba doctorChecks the installation and reports anything that would quietly send more than it needs to.
Each row is ok, -- or a warning:
okthe thing is configured and working.--the thing exists but has not been asked to do anything. Not a failure.- warning something is actively working against you. These are the rows to read.
The most important warning it produces:
-- no conflicting base url ANTHROPIC_BASE_URL=https://api.anthropic.com
overrides the settings fileAn exported environment variable beats a settings file. If something in your shell profile,
or an application that spawned your terminal, exports a base URL pointing at the provider,
your tool bypasses the proxy no matter what init wrote. Unset it and start your tool from a
new shell.
verify
tokunseba verifyThis does not read configuration at all. It reads what happened:
- whether any request has arrived, and how recently
- which tools they came from
- how many tokens were kept off the wire
- which transforms actually fired
If nothing has arrived, it says so and tells you the likely reason rather than printing zeros. The usual causes, in order of how often they are the answer:
- An exported base URL wins over the settings file.
doctornames it. - The tool was opened from the Dock, so it never read your shell profile. Run
tokunseba apps on. - The shell block has not been picked up yet. Open a new terminal.
- The proxy is not running.
tokunseba status.
Seeing one request end to end
tokunseba top --since 24hgives you the transforms and the handles. For any request id:
tokunseba explain <request-id>prints original beside replacement, block by block. For any handle:
tokunseba expand <handle>prints the original text exactly as it arrived. If what comes back is not identical to what your tool produced, that is a bug worth reporting.
Measuring whether it helped
Every session is assigned at random to a control arm or a treatment arm. Both go into the same ledger:
tokunseba stats --abThat compares requests per session, input and output tokens per session, and latency per request, between the two arms. It is the only figure on this site that is about your own traffic rather than someone else's.
The arms only differ when tier 3 is on. With tier 3 off, both arms get identical treatment and the comparison will show no difference, which is correct.