Fixed#
An Action pin could point at a commit that is not on main, and every check
passed. The guard on the README's uses: lines is thorough — a 40-character
SHA, never a tag, with a version comment verified against that commit's own
manifest rather than the working tree. All of it is satisfied by the pre-squash
head of a feature branch, which says exactly the right version and is deleted
the moment the pull request merges. GitHub can garbage-collect it, and a
workflow pinned there stops resolving with no warning to anybody.
This was caught while preparing this very release: the sha to hand was the
branch commit, not the squash-merge on main. It is now also gone from the
clone, which is the hazard demonstrating itself. The pin must be an ancestor of
origin/main, and a clone without that ref reports the pin as unverified rather
than passing over it.
Two commands had no options section, and one had two. trazum ladder takes
--since, --until and --label; trazum owners takes --since and
--until. Neither had an OPTIONS FOR section, so a reader had no way to learn
from the help that they take a window at all — and both are commands whose whole
point is judging a period.
eval had two sections under the same heading, with different content. One
listed --export promptfoo and -o; the other carried the paragraph explaining
that it costs three provider calls per case — the original twice to measure
the model's own run-to-run variance, and the optimised once — and exits 1 on
divergence. Whichever a reader scrolled to, they got half, and the duplicate
heading meant neither half announced itself as one. Merged into a single
section, in both locales.
A guard now derives the commands that take flags of their own from
COMMAND_FLAGS, minus the globals, and requires each to have exactly one
section. A command that gains its first flag fails until somebody writes what it
does; a heading that appears twice fails naming itself. The duplicate check is
handed the shape it exists for rather than only today's corrected help.
trazum profile was missing from --help. Not a small one: profile is
the command almost every refusal in this product points a reader at — "trazum
profile prices a mistral log you export", the --max-usd gate that fails a
build on the bill, the --json documents history reads. It had a full flag
allowlist and its own OPTIONS FOR profile section, and it was absent from the
list of commands the help presents.
Nothing noticed because the "thirty-two commands" figure the README states is
guarded against COMMAND_FLAGS, which had all thirty-two. The USAGE block had
thirty-one, and no check compared the product's own two lists with each
other. A guard now does, in both directions: a command that is dispatched and
absent from USAGE fails, and so does a USAGE line promising a command the CLI
does not have — because a fix for the first that invented the second would
satisfy one check and mislead every reader.
--help said the gateway fronts two providers while the gateway itself said
four. The USAGE block read trazum gateway <anthropic|openai>; running
trazum gateway with no argument answers "Known: anthropic, openai, deepseek,
google." A reader meets both within a minute of each other, and the wrong one
is the first.
It went stale during the release whose entire subject was that list. Nothing
caught it, because the guard that exists for exactly this — the one comparing
docs/gateway.md against the compiled upstream table — was pointed at the
documentation, and the same sentence lived inside the product.
The fix is not a longer list kept in sync. It is <provider>, in both
commands and both locales, with the enumeration left to the refusal that
derives it. A list of providers typed beside the thing has now gone stale in
docs/gateway.md, in ROADMAP.md and in the help text itself; the answer each
time was to stop writing the list, not to correct it.
A guard now refuses any USAGE line whose <a|b> group names two or more
providers — the provider set derived from the catalogue, the upstream table and
the connector list together, so it covers a name defined in any of them. It is
handed the exact line it was written for, and separately the innocent shapes
(<file|dir|->, [--dry-run | --yes]) it must not fire on. And the other half
is asserted too: both refusals still name every provider they support, because
removing the list from USAGE would be a regression if nothing else told a reader
what to pass.