Handing the folder to someone else

AGENTS.md §2 is how you install it; this is how someone else does, from a
copy you made. share.sh is one command, but the command is not the procedure.

The thing to understand before anything else: share.sh has its own exclusion
list, and it is not .gitignore. They are maintained separately and have drifted
before. share.sh now excludes ./data and .git wholesale, but it did not
always, and the report patterns it relies on for reports written outside data/
have stopped matching the tool's filenames twice. Check the bundle; do not assume
the list is current.

1. Scrub, then build, then copy — in that order

ls data/imported data/reports 2>/dev/null   # your exports and your reports

data/imported/*.csv is the raw export — every campaign name, every day of spend
— and data/reports/*_models.html names the same campaigns. Both are excluded by
./data now, and TestShareScriptExcludesGeneratedFiles asserts that exclusion
with all three real import filenames. Note what that pattern does not cover: a
report forecast wrote next to a CSV somewhere else is caught only by
*_forecast*.html, which is a filename pattern and has failed twice before.

Verify rather than trust — list the copy before you hand it over (step 3).

rm -rf dist                                  # or the old binaries get re-shipped
./share.sh ~/somewhere/predict-marketing
ls ~/somewhere/predict-marketing/data 2>/dev/null   # must print nothing

rm -rf dist matters: the guard in share.sh is [ ! -d dist ], so it builds
the prebuilt binaries only when there are none. With dist/ already present it
re-ships whatever is there, however old. dist/BUILT.txt records the date.

.git is excluded too, so their build reports no commit. If you want them to
have the history, copy it in deliberately afterwards — but
it is the whole history.

2. Check the copy before you send it

cd ~/somewhere/predict-marketing
ls data 2>/dev/null && echo "STOP: your data is in the bundle"
find . -name '*.db' -o -name '*_forecast*.html' -o -name '*_models.html' \
     -o -path './models/finetuned*'

That find must print nothing. A fine-tuned adapter in particular is fitted to
your numbers and registered by a path on your machine: it would arrive broken,
and if it worked it would be the wrong model for their data (AGENTS.md §4c).

Then prove the bundle builds and runs its own checks:

go build -o predictmarketing . && go vet ./... && go test -count=1 .

Expect fifteen skips, and expect the suite to still say ok. A copy has no
models/.venv, so everything that needs a worker disappears: all eight
protocol tests, both weights tests, the multivariate perturbation test, and the
binary/pycache tests. A green run in a fresh bundle proves the CSV reader, the
database and the HTML — it proves nothing about the model protocol, which
AGENTS.md §4 calls the one thing not to break. That check only happens after
./install.sh on their machine, or on yours.

3. Prebuilt binaries: who they are for

build-dist.sh builds four: darwin/arm64, darwin/amd64, linux/amd64,
linux/arm64. They exist only so somebody with no Go toolchain can install.
Anyone with Go gets a fresh build from source instead, and that is the path
install.sh prefers.

There is no Windows binary on purpose. The program cross-compiles to Windows
cleanly, but install.sh and share.sh are POSIX sh, so there would be
nothing on Windows to run them. Send a Windows user to the by-hand steps in
README.md, or to WSL.

4. What to tell them

One command, in a terminal opened in the folder:

./install.sh

Three things on their first run look like faults and are not, and they will ask
about all three — AGENTS.md §2b has the detail: the Hugging Face
"unauthenticated requests" warning (both models are public and ungated), the
Loading weights progress bar on stderr, and the database arriving at mode
0600.

Tell them the third model is not in the box. chronos2ft is trained on their
own numbers and has nothing to learn from until they have a CSV; install.sh
says so at the end, and .claude/skills/finetune/ is the procedure.

5. Done means

You looked inside the copy. Say what find printed and what go test reported,
including the skip count. "share.sh ran" is not done — the script has shipped
real campaign names before, and it is the one failure mode that cannot be taken
back once the folder is sent.