Fail a build list that names software the baseline provides

A site layer holds only the software the European Environment for Scientific Software Installations (EESSI) does not install. stack solve --build-list-excludes-baseline writes that list from a solve. A later edit can put an EESSI module back on it, and the list becomes a full stack again. eb-stack stack provided reads the list and fails when a line’s package name is one the baseline already provides. That comparison ignores the version and the toolchain. The same command also fails a toolchain other than 2026.1 or GCCcore-15.2.0 below it, a per-line --hooks flag, and a name that only a --family-lock covers.

Check the fixture list

tests/provided_fixtures/site.list is a Jenkins build list. tests/provided_fixtures/installed.txt names FFTW and Python. tests/provided_fixtures/baseline.lock.json names GROMACS. VASP is on the list and in neither source, so it is the site layer.

eb-stack stack provided \
  --build-list tests/provided_fixtures/site.list \
  --installed tests/provided_fixtures/installed.txt \
  --baseline-lock tests/provided_fixtures/baseline.lock.json

That command printed:

provided GROMACS-2026.0-foss-2026.1.eb name=GROMACS list=tests/provided_fixtures/site.list line=3
provided FFTW-3.3.10-gompi-2026.1.eb name=FFTW list=tests/provided_fixtures/site.list line=4
provided fixtures/gromacs_2025_to_next/easyconfigs/foss-2025b/Python-3.12.3-foss-2025b.eb name=Python list=tests/provided_fixtures/site.list line=5
unparsed not-a-version.eb list=tests/provided_fixtures/site.list line=7
toolchain fixtures/gromacs_2025_to_next/easyconfigs/foss-2025b/Python-3.12.3-foss-2025b.eb name=Python list=tests/provided_fixtures/site.list line=5 uses=foss-2025b; a site-layer line on EESSI 2026.06 must use a 2026.1 toolchain or GCCcore-15.2.0 below it (GCCcore-15.2.0, GCC-15.2.0, gfbf-2026.1, gompi-2026.1, foss-2026.1)
hooks FFTW-3.3.10-gompi-2026.1.eb name=FFTW list=tests/provided_fixtures/site.list line=4 sets=--hooks=/sw/hooks/site.py; which replaces EESSI's hooks
hits=3 unparsed=1 allowed=0 provided=3 toolchain=1 hooks=1 family=0 licensed=0

The process exited 1. Standard error was:

error: 3 build-list lines name software the baseline already provides; 1 build-list line names no package; 1 build-list line uses a toolchain other than 2026.1 or GCCcore-15.2.0; 1 build-list line sets --hooks

Read a hit left to right. GROMACS-2026.0-foss-2026.1.eb is the line. The baseline lock has GROMACS 2025.2, and the names match, so a newer build of the same software still fails. The Python line is a path; only the filename is read. It is also toolchain: foss-2025b is not 2026.1 or GCCcore-15.2.0 below it. The allowed labels are GCCcore-15.2.0, GCC-15.2.0, gfbf-2026.1, gompi-2026.1, and foss-2026.1. The FFTW line is gompi-2026.1, which is allowed, and it sets --hooks=, which replaces EESSI’s hooks. # FFTW-3.3.10-gompi-2026.1.eb is a comment, so it is not a hit. not-a-version.eb names no package. VASP-6.5.1-foss-2026.1.eb is absent from the report: it is not in the baseline, its toolchain is foss-2026.1, and no licensed list was passed.

Keep a second version

Pass --allow for a package the site was asked to build beside the baseline’s version. The match ignores case.

eb-stack stack provided \
  --build-list tests/provided_fixtures/site.list \
  --installed tests/provided_fixtures/installed.txt \
  --baseline-lock tests/provided_fixtures/baseline.lock.json \
  --allow GROMACS

GROMACS drops out of the hits. The counts line is hits=2 unparsed=1 allowed=1 provided=3 toolchain=1 hooks=1 family=0 licensed=0. The process still exits 1, because FFTW, Python, the unparsed line, the foss-2025b toolchain, and the --hooks flag remain. --allow does not waive those.

A list that passes

printf 'VASP-6.5.1-foss-2026.1.eb\n# FFTW-3.3.10-gompi-2026.1.eb\n' > /tmp/site-only.list
eb-stack stack provided \
  --build-list /tmp/site-only.list \
  --installed tests/provided_fixtures/installed.txt

That command printed:

hits=0 unparsed=0 allowed=0 provided=2 toolchain=0 hooks=0 family=0 licensed=0

The process exited 0. --installed and --baseline-lock are combined when both are set. Passing neither is status 2. --json and --format json print the same fields as JSON. --format github prints one ::error or ::warning workflow command per finding, with file, line, and title. eb_stack_provided is the same check. A hit, an unparsed line, a wrong toolchain, a hooks flag, or a family-lock name is a tool error.

A family lock, and a group-restricted name

A foss baseline does not contain the LLVM family. Pass that family’s lock as --family-lock. A name only that lock has is not left as site layer: the line names the lock to use. --licensed-file data/site-licensed.txt warns for VASP, MATLAB, ANSYS, ORCA, and Gurobi. The warning does not fail the command unless --strict is set.

eb-stack stack provided \
  --build-list tests/provided_fixtures/site.list \
  --baseline-lock tests/provided_fixtures/baseline.lock.json \
  --family-lock tests/provided_fixtures/lfoss.lock.json \
  --licensed-file data/site-licensed.txt \
  --allow GROMACS --allow FFTW --allow Python

With those three names allowed, the provided hits are gone. VASP is a licensed warning. The foss-2025b line and the --hooks line still fail, and the status is 1. A build-list name that is only in tests/provided_fixtures/lfoss.lock.json (OpenBLAS and SciPy in that file) is a family line. It names lock=tests/provided_fixtures/lfoss.lock.json label=lfoss-2026.1 and says to use that lock instead of the site layer. --allow drops that line too. It does not drop a wrong toolchain or a hooks flag.