name	cmake-extract
description	Extract and analyze CMake build metadata from a repository, mapping it to Spack-relevant abstractions for use in recipe generation

Repository Extraction

Use this skill when a user initiates a request to create or update a Spack package recipe. The details extracted by this skill will be used in downstream tasks like package generation and repair.

Step 1 -- obtain repository

The user will provide a URL (tarball/Git) or local path to a codebase. Use standard tools (curl/wget/tar/zip/git) to download and open the repo.

Place downloaded files into a newly created temporary directory (mktemp -d "$TMPDIR/XXXXX").

Downloads should be efficient (i.e., use depth=1 with Git).

Once the repository is accessible, run the two extraction steps below. The scripts accept directory paths as their argument and expand shell variables (e.g., $TMPDIR).

Step 2 -- detect build system

Run: python scripts/detect_build_systems.py <repo_dir>

This outputs two values:

build_system: the detected build system (e.g. cmake, autotools, meson)
features: additional repository-level features (e.g. cuda, python)

If the detected build system is not cmake, stop immediately and inform the user: "Only CMake-based repositories are supported."

Step 3 -- extract CMake metadata

Run: python scripts/cmake_extract.py <repo_dir>

This script scans CMake files in the repository and outputs a structured summary of:

cache_sets: cache variables (type, default, docstring)
normal_sets: plain set() variables
commands: ordered list of CMake commands and arguments
options: option(NAME "help" DEFAULT) toggles
if blocks: conditional logic with true/false bodies

Because the script will read the CMake files, do not cat or grep the files.

Step 4 -- map CMake metadata to Spack abstractions

Your task is to translate the CMake metadata from step 3 into a concise manifest of Spack-relevant insights used for recipe generation.

Brevity and accuracy are critical; do not invent anything not explicitly present in the metadata.

Output format

Plain text, one mapping per line: → ← Example lines: WITH_ADIAK → variant('+adiak') ← cache_sets:WITH_ADIAK (BOOL, default ON) find_package(Python) → variant('python') + depends_on('python', when='+python') ← commands:find_package cmake_minimum_required(3.14) → depends_on('cmake@3.14:', type='build') ← commands:cmake_minimum_required project(LANGUAGES CXX) → depends_on('cxx', type='build') ← commands:project LANGUAGES

After all mapping lines, append a SUMMARY block:

SUMMARY variants: var1 var2 var3 dependencies: dep1 dep2 dep3 conflicts: '+tight_error when=round=never' compilers: cxx fortran

All identifiers must use Spack-style naming (lowercase, no cmake prefixes): cmake, cuda, mpi, python, fortran, c, cxx, openmpi, etc.

Mapping guidelines

Variants

WITH_*, ENABLE_*, USE_*, BUILD_* BOOL variables → variant(...)
STRING variables with enumerated STRINGS property → multi-valued variant Example: ZFP_ROUNDING_MODE (NEVER, FIRST, LAST) → variant('round', values=('never','first','last'))
Emit variant('shared') if and only if BUILD_SHARED_LIBS or a direct equivalent is present
Normalize enum values: strip CMake prefixes, lowercase Example: ZFP_ROUND_NEVER → 'never'
Use canonical names for language bindings: variant('python'), variant('fortran'), variant('c') Do not use: zfpy, zforp, cfp, or other non-standard names
Do not skip options that default to OFF, are marked ADVANCED, or appear internal, if they control behavior, optional outputs, or language bindings

Conflicts

If a CMake option is only valid when another has a specific value, emit a Spack conflict: Example: TIGHT_ERROR only valid when ROUNDING != NEVER → conflicts('+tight_error', when='round=never')

Compilers

Emit depends_on('', type='build') for any language that is actually used to compile sources. Use the following signals, in rough order of strength:

Strong signals (language is definitely used):

project(... LANGUAGES C/CXX/Fortran ...)
enable_language(C/CXX/Fortran)
source files with .c, .cpp, .f90, .f, .f95 extensions referenced in CMakeLists

Weak signals (language is likely used, emit with note):

check_cxx_compiler_flag, check_language, check_c_compiler_flag
target_link_libraries referencing C/C++/Fortran objects
CMAKE_CXX_FLAGS, CMAKE_C_FLAGS, CMAKE_Fortran_COMPILER variables set or tested

You must emit the dependency even if there is only a single weak signal.

Emit exactly one mapping line per language detected, citing the strongest evidence found.

Dependencies

find_package(X REQUIRED) or find_library(X) → depends_on('x')
find_package(X COMPONENTS foo bar) → depends_on('x') and note components
Wrapped in a condition tied to a variant → add when='+'
pkg_check_modules(X) → depends_on('x')
target_link_libraries referencing an external library → depends_on (if not already covered)
MPI is a virtual package: find_package(MPI) → depends_on('mpi')
OpenMP is never a dependency; if find_package(OpenMP) is present → variant('openmp') only
cmake_minimum_required(VERSION X.Y) → depends_on('cmake@X.Y:', type='build')
CUDA: find_package(CUDA) or enable_language(CUDA)
- GPU-specific files, target_link_libraries(... cuda) → implies depends_on('cuda'), gated on +cuda
Python : any find_package(Python* …) ⟶ variant('python'), associated deps

What not to emit

Do not guess version constraints for any dependency
Do not emit dependencies or variants that are not evidenced in the metadata
Do not impart knowledge that is not present in the CMake metadata. This will cause downstream tasks to fail.

Step 5 -- write output to disk

Write all results to: <repo_dir>/spack_metadata.txt

The file must contain the following sections in order. Include the Note: lines, they are for the user.

BUILD_SYSTEM: Note: Use this to select the correct Spack base class (e.g. cmake → CMakePackage). FEATURES: <comma-separated list from Step 2>

--- CMAKE MAPPINGS --- Note: Each line is a direct instruction: emit exactly the spack abstraction shown, and consult the evidence field to resolve ambiguity or determine constraints. <all mapping lines from Step 4>

--- SUMMARY --- variants: ... dependencies: ... conflicts: ... compilers: ...

This file is the bridge between this skill and other Spack-adjacent tasks.

name	cmake-extract
description	Extract and analyze CMake build metadata from a repository, mapping it to Spack-relevant abstractions for use in recipe generation

Repository Extraction

Use this skill when a user initiates a request to create or update a Spack package recipe. The details extracted by this skill will be used in downstream tasks like package generation and repair.

Step 1 -- obtain repository

The user will provide a URL (tarball/Git) or local path to a codebase. Use standard tools (curl/wget/tar/zip/git) to download and open the repo.

Place downloaded files into a newly created temporary directory (mktemp -d "$TMPDIR/XXXXX").

Downloads should be efficient (i.e., use depth=1 with Git).

Once the repository is accessible, run the two extraction steps below. The scripts accept directory paths as their argument and expand shell variables (e.g., $TMPDIR).

Step 2 -- detect build system

Run: python scripts/detect_build_systems.py <repo_dir>

This outputs two values:

build_system: the detected build system (e.g. cmake, autotools, meson)
features: additional repository-level features (e.g. cuda, python)

If the detected build system is not cmake, stop immediately and inform the user: "Only CMake-based repositories are supported."

Step 3 -- extract CMake metadata

Run: python scripts/cmake_extract.py <repo_dir>

This script scans CMake files in the repository and outputs a structured summary of:

cache_sets: cache variables (type, default, docstring)
normal_sets: plain set() variables
commands: ordered list of CMake commands and arguments
options: option(NAME "help" DEFAULT) toggles
if blocks: conditional logic with true/false bodies

Because the script will read the CMake files, do not cat or grep the files.

Step 4 -- map CMake metadata to Spack abstractions

Your task is to translate the CMake metadata from step 3 into a concise manifest of Spack-relevant insights used for recipe generation.

Brevity and accuracy are critical; do not invent anything not explicitly present in the metadata.

Output format

After all mapping lines, append a SUMMARY block:

SUMMARY variants: var1 var2 var3 dependencies: dep1 dep2 dep3 conflicts: '+tight_error when=round=never' compilers: cxx fortran

All identifiers must use Spack-style naming (lowercase, no cmake prefixes): cmake, cuda, mpi, python, fortran, c, cxx, openmpi, etc.

Mapping guidelines

Variants

WITH_*, ENABLE_*, USE_*, BUILD_* BOOL variables → variant(...)
STRING variables with enumerated STRINGS property → multi-valued variant Example: ZFP_ROUNDING_MODE (NEVER, FIRST, LAST) → variant('round', values=('never','first','last'))
Emit variant('shared') if and only if BUILD_SHARED_LIBS or a direct equivalent is present
Normalize enum values: strip CMake prefixes, lowercase Example: ZFP_ROUND_NEVER → 'never'
Use canonical names for language bindings: variant('python'), variant('fortran'), variant('c') Do not use: zfpy, zforp, cfp, or other non-standard names
Do not skip options that default to OFF, are marked ADVANCED, or appear internal, if they control behavior, optional outputs, or language bindings

Conflicts

If a CMake option is only valid when another has a specific value, emit a Spack conflict: Example: TIGHT_ERROR only valid when ROUNDING != NEVER → conflicts('+tight_error', when='round=never')

Compilers

Emit depends_on('', type='build') for any language that is actually used to compile sources. Use the following signals, in rough order of strength:

Strong signals (language is definitely used):

project(... LANGUAGES C/CXX/Fortran ...)
enable_language(C/CXX/Fortran)
source files with .c, .cpp, .f90, .f, .f95 extensions referenced in CMakeLists

Weak signals (language is likely used, emit with note):

check_cxx_compiler_flag, check_language, check_c_compiler_flag
target_link_libraries referencing C/C++/Fortran objects
CMAKE_CXX_FLAGS, CMAKE_C_FLAGS, CMAKE_Fortran_COMPILER variables set or tested

You must emit the dependency even if there is only a single weak signal.

Emit exactly one mapping line per language detected, citing the strongest evidence found.

Dependencies

find_package(X REQUIRED) or find_library(X) → depends_on('x')
find_package(X COMPONENTS foo bar) → depends_on('x') and note components
Wrapped in a condition tied to a variant → add when='+'
pkg_check_modules(X) → depends_on('x')
target_link_libraries referencing an external library → depends_on (if not already covered)
MPI is a virtual package: find_package(MPI) → depends_on('mpi')
OpenMP is never a dependency; if find_package(OpenMP) is present → variant('openmp') only
cmake_minimum_required(VERSION X.Y) → depends_on('cmake@X.Y:', type='build')
CUDA: find_package(CUDA) or enable_language(CUDA)
- GPU-specific files, target_link_libraries(... cuda) → implies depends_on('cuda'), gated on +cuda
Python : any find_package(Python* …) ⟶ variant('python'), associated deps

What not to emit

Do not guess version constraints for any dependency
Do not emit dependencies or variants that are not evidenced in the metadata
Do not impart knowledge that is not present in the CMake metadata. This will cause downstream tasks to fail.

Step 5 -- write output to disk

Write all results to: <repo_dir>/spack_metadata.txt

The file must contain the following sections in order. Include the Note: lines, they are for the user.

BUILD_SYSTEM: Note: Use this to select the correct Spack base class (e.g. cmake → CMakePackage). FEATURES: <comma-separated list from Step 2>

--- SUMMARY --- variants: ... dependencies: ... conflicts: ... compilers: ...

This file is the bridge between this skill and other Spack-adjacent tasks.

cmake-extract

Repository Extraction

Step 1 -- obtain repository

Step 2 -- detect build system

Step 3 -- extract CMake metadata

Step 4 -- map CMake metadata to Spack abstractions

Output format

Mapping guidelines

Variants

Conflicts

Compilers

Dependencies

What not to emit

Step 5 -- write output to disk

Plus depuis ce dépôt

Plus depuis ce dépôt

Repository Extraction

Step 1 -- obtain repository

Step 2 -- detect build system

Step 3 -- extract CMake metadata

Step 4 -- map CMake metadata to Spack abstractions

Output format

Mapping guidelines

Variants

Conflicts

Compilers

Dependencies

What not to emit

Step 5 -- write output to disk