Added#
trazum init — the first five minutes. Everything since 1.41 raised the
ceiling; this lowers the floor. One command that does what a person would do on
their first afternoon: walks for prompt files, reads source files for which
provider the code calls (the where machinery from 1.7, which has always
refused to guess), notices a usage log or a connector credential, and writes a
trazum.config.json out of what it can actually justify. Then it prints the
single most valuable thing it found, arithmetic before the figure — the
calls, the label, the model, what they cost, and only then the saving. A tool
that opens with a dollar amount nobody can check gets closed.
It is a detection, not a wizard. Nothing is asked. --dry-run prints the config
and writes nothing; --yes replaces one already there; without it an existing
config is left alone; --json emits the whole proposal and writes nothing.
Four keys it refuses to write, which is the substance of the release.
- A budget, ever. A log says what your traffic was; a budget says what it may cost, which no log can answer. "The measured month plus twenty per cent" would be Trazum inventing a threshold and then grading somebody against it. The measured figure is handed over; the limit stays with the person who can set one.
- A monthly rate from a short window. Twenty-eight days minimum, so every weekday appears the same number of times. And a separate refusal for the quiet case: if any call in the log carries no timestamp, no rate is stated at all — those calls cannot be placed inside the span a rate would divide by, and dividing anyway makes the figure come out high with nothing to show why.
- A cache hit rate from a log with no cache columns. Not recorded is not not-happened. A zero there would tell every later caching advisory that caching is doing nothing — a finding invented out of a missing field. A log that does record cache writes and reports zero reads is a different thing entirely: that is a measurement, and it is written.
usage.batchEligible, in either direction. Whether the work tolerates a batch window is a product decision no log records.falsewould silently delete the batch lever from every report this config touches;truewould sell a saving on latency nobody agreed to give up.
It also declines a model when the source names a provider and no model.
trazum where prints a provider's default because a reader can see it is a
guess; a config file cannot — six weeks later it reads as somebody's decision,
and every price in every report rests on it.
Every refusal arrives with what would settle it, typed rather than as prose
— the rule spend_guard established for a call in 1.45, applied here to a file.
proposeInit in @trazum/core. The judgement is a pure function over
observations the CLI collected: no filesystem anywhere near it, so every rule
above is tested without one, and --dry-run is the same code path minus the
write rather than a second implementation that drifts.
docs/usage-logs.md. Worked examples for the four shapes people actually
have — an Anthropic response, an OpenAI response, a Vercel AI SDK onFinish
hook, an OTel collector — with real records rather than a schema dump, and a
table of which finding each optional field buys. Linked from init's output,
because the moment somebody needs it is the moment the tool says "no usage
found". Every example in it was run through trazum profile before it was
written down; one claim did not survive that (a JSON array in a .json file
is not read) and the page says so instead.
Fixed#
init no longer needs a working config to run. Every command loads
trazum.config.json before dispatch and throws when it will not parse —
correct for the rest of them, since "defaults" for a budget means "no budget"
and a silent revert is a green build that should have been red. But init is
the command somebody runs because their setup is broken, and it was the one
command a broken setup could stop from running. It now survives the failure
with nothing carried forward — no keys, no budgets, no locale — and refuses to
write over the file it could not read, naming it. The refusal had been written
two hours earlier in the same release and was unreachable code standing behind
a throw; it was found by a test that expected it to fire and watched a parse
error arrive instead.
Security#
Three guards on the first run, each proven with a planted probe. init has
the widest reach in this product and the least trust behind it: it runs in a
directory it has never seen, before anybody has read a page of documentation.
So the build fails if commandInit reaches the network or an LLM (the
deterministic core has been the entry point since 0.1.0, and a tool whose
introduction costs money is one nobody introduces); if it reads the value out
of findCredential rather than the variable name (checked as what is
destructured, not as what is printed — a version that pulls the key out and
happens not to log it today is one refactor away from logging it tomorrow); or
if it writes any file other than join(root, CONFIG_FILENAME).
A time-of-check/time-of-use race, found by CodeQL in code written the same
day. init bounds how large a source file it will read, and it took that
measurement with stat(path) and then read readFile(path) — two lookups of
the same name, where what arrives the second time need not be what was measured
the first. The bound would have been enforced against a file that was no longer
there. It now opens the file once and stats the handle, which is the same
inode by construction. A fourth guard holds the fix in place, because it is
invisible in the output: both versions print exactly the same thing, and only
one of them is checking the file it reads.
Documentation#
The first-run document is contracted. trazum init --json has a section in
docs/json-output.md with a two-direction parity test, bounded to its own
section — and the spend-guard harvest above it was bounded in the same commit,
because an unbounded harvest starts enforcing the next shape's fields the moment
a new contract is appended. That has now happened five times in this file, so
the bound is written before the section that would break it rather than after.
README leads with init. Getting started opened with optimize your-prompt.txt — the smallest thing in the tool — for twenty-two releases. It
opens with npx @trazum/cli init now, and the command table gains a row.