stm32-installer

Installs embedded C libraries into an STM32 project.

Copying two files into a project is easy. Doing it so the build actually picks them up, without clobbering the settings you changed last month, is the part this handles.

It works with any library that has an installer.yml at the root of its repository. I wrote it for my own NimaLTD libraries, and they all have one, but nothing in it is tied to them: add an installer.yml to yours, as shown below, and it installs the same way. The same file serves zephyr-installer, esp32-installer, arduino-installer and pio-installer, so one library can be installed into an STM32 project, a Zephyr application, an ESP-IDF project, an Arduino sketch and a PlatformIO project alike.

Every command below uses owner/example, the library's GitHub owner and name, such as nimaltd/ee24. Put the real ones in its place.

Installing a library

Install the installer once. This needs internet:

pip install stm32-installer

Then, from the root of your STM32 project, install a library by its GitHub owner and name. To see it work first, take stm32-installer-example, a library made for trying the installer on:

stm32-installer nimaltd/stm32-installer-example

It builds in any STM32 project CubeMX made, with nothing to set up, and blinks an LED without HAL_Delay(). Its README walks through an install, an update that keeps a setting you changed, an option, and an uninstall. Any other library goes in the same way, as stm32-installer owner/example.

To update it later, pip install --upgrade stm32-installer. To remove it, pip uninstall stm32-installer.

On Windows, if stm32-installer is not found after installing, pip put it in a folder that is not on your PATH. python -m stm32_installer owner/example runs the same thing.

On Linux or macOS, if pip refuses with externally-managed-environment, use pipx install stm32-installer instead, and pipx upgrade stm32-installer to update.

When a newer stm32-installer is out, the last line of a run says so, with the command that updates it. It asks PyPI at most once a day, waits no more than two seconds for an answer, and says nothing when there is none. --no-update-check turns it off, and so does INSTALLER_NO_UPDATE_CHECK set to anything. On a build server, where CI is set, it is off already.

A library can need a newer installer than the one you have. It then says so and changes nothing:

Error: This library needs stm32-installer 1.2.0 or newer, and this one is 1.1.1. Nothing was changed.
Update it with:
    pip install --upgrade stm32-installer

Every command

Each command with owner/example standing for the library. An update is the install again. Every option is explained under Options.

On a bad connection

A file that does not arrive is tried again, five times in all, 1, 2, 4 and 8 seconds apart. The tries take turns between raw.githubusercontent.com and api.github.com, two hosts on different addresses, so one that is filtered or failing does not stop the install. Every file of every library is fetched before anything is installed, so a run that fails part way leaves your project as it was: run it again.

Behind a proxy, or a VPN client that offers one, point the installer at it. In Command Prompt:

set HTTPS_PROXY=http://127.0.0.1:10809

In PowerShell:

$env:HTTPS_PROXY = "http://127.0.0.1:10809"

On Linux and macOS, export HTTPS_PROXY=http://127.0.0.1:10809. Use your proxy's own address and port. On Windows, a proxy set in the system settings is used without this.

pip has its own retries, and on a slow link it helps to give it more time: pip install --retries 10 --timeout 60 stm32-installer. When nothing gets through, install from zips, as below.

Without internet

On a machine that has internet, save the installer as a file:

pip download stm32-installer

That saves one file, such as stm32_installer-1.1.2-py3-none-any.whl, where the number is the installer's version. Download the library too, from the green Code button, Download ZIP. Copy both to the machine that has no internet, then from your project:

pip install stm32_installer-1.1.2-py3-none-any.whl
stm32-installer D:/Downloads/example-main.zip

Neither line needs the internet. The installer needs nothing but itself, and the library comes from the zip.

If the library needs another one, such as osal, download that zip as well and give both on one line:

stm32-installer D:/Downloads/example-main.zip D:/Downloads/osal-main.zip

Forget it, and the installer stops before writing anything, naming the zip it is missing.

A private repository

A library in a private repository installs the same way, given a GitHub token that can read it. Make one once:

1. Open github.com/settings/personal-access-tokens and click Generate new token.

2. Give it a name and an expiry. Under Resource owner, pick the account or the organisation that owns the repository.

3. Under Repository access, choose Only select repositories and pick the library's repository.

4. Under Permissions, set Contents to Read-only:

   • Older layout: open Repository permissions and change Contents from No access to Read-only.
   • Newer layout: click + Add permissions, tick Contents, then set its access to Read-only.

   Metadata comes with it by itself. Contents is the one that matters: without it the token can see the repository but not read a file of it.

5. Click Generate token, and copy it. GitHub shows it only once.

Then give it with --key:

stm32-installer owner/example --key github_pat_...

Or put it in GITHUB_TOKEN once, and install by name as usual. That keeps it out of your shell's history, where --key leaves it:

Windows, PowerShell:

$env:GITHUB_TOKEN = "github_pat_..."
stm32-installer owner/example

Windows, Command Prompt:

set GITHUB_TOKEN=github_pat_...
stm32-installer owner/example

Linux and macOS:

export GITHUB_TOKEN=github_pat_...
stm32-installer owner/example

--key comes first when both are there. GH_TOKEN works as well, and if you use the gh command, gh auth token prints the token you are signed in with. Whichever way it comes, the token is sent to GitHub and nowhere else.

When the token cannot read the library, the installer says which of these it is:

• It can see the repository, but cannot read its files. The token has no Contents permission, only Metadata. Edit the token, and its value stays the same, so the command you ran keeps working: open the tokens page, click the token's name, click Edit, set Contents to Read-only as in step 4, and click Update at the bottom of the page. Then run the same command again.
• It cannot see the repository. The repository is not in the token's Repository access. Edit the token as above and add it.
• GitHub did not accept the token. It has expired, been revoked, or was copied with a character missing. Make a new one.

A repository of an organisation can need one more step: an organisation that asks for fine-grained tokens to be approved shows the token nothing of its repositories until an owner of the organisation approves it, under its settings, Personal access tokens, Pending requests.

What to install

The argument can be any of these, and you can give several at once:

A bare name with no owner, like example, is a NimaLTD library, nimaltd/example on GitHub, unless a folder of that name exists where you run it, in which case the folder is used. A path that does not exist is reported as missing rather than looked up on GitHub, so a typo in D:/Downloads/... gets a straight answer.

The zip does not need unpacking first. If you did unpack it with Windows' Extract All, which puts example-main inside another example-main, either folder works. A repository still on master gives example-master, which works the same.

From GitHub, a library comes from its repository's default branch, the one GitHub shows first, whatever the repository calls it. After an @, a tag, a branch or a commit takes exactly that instead, and so does a GitHub address ending in /tree/v2.0.0. Every file comes from one commit, even when the branch moves on while they download, and the record keeps that commit, so you can always tell what you built with.

Options

--ref holds a project on one release, which is useful when you need exactly what you built with last time:

stm32-installer owner/example --ref v2.0.0         # a tag
stm32-installer owner/example --ref develop        # a branch
stm32-installer owner/example --ref 00949e695e16   # an exact commit
stm32-installer owner/example@v2.0.0               # the same, for this library alone

Where the files go

From GitHub, from a zip, or from a folder outside your project, the files the library's installer.yml lists are copied into a folder of your project. Nothing else comes with them, so the library's tests never reach your build. The zip or folder you gave is left exactly as it was.

A folder already inside your project becomes the library where it stands. The header and source move to its top, your config file is created, and everything that belongs to the repository rather than your firmware is removed from it, the test folder above all. --dir does not apply here, since the folder is already where it is going. Rename it first if you want it called something else.

The folder question

Without --dir, you are asked where the library should go:

Folder to install example into [example]:

Press Enter for the default: the library's name, or the folder it is already in when this is an update. In a script or on a build server, where nobody is there to answer, the default is taken without asking, rather than waiting for ever.

Parts you choose

Some libraries have a part not every project wants, such as LittleFS support in a flash driver. The first install asks about each one:

Options
  littlefs   LittleFS file system on the flash. Add it? [y/N]:

Anything but y or yes leaves it out. A yes brings the part's files, and any library it needs, which goes in first like any other. In a script, where nobody can answer, it is left out, and --with littlefs takes it without a question.

The answer is kept in .stm32-installer.json in your project, so an update does not ask again. It says what it kept, and how to change it:

Options
  littlefs   yes, as before. Change it with --without littlefs

Change your mind with --with or --without on any later run. --without littlefs takes the part's files out of the library's folder and out of your IDE project. A library the part brought stays, since something else may use it by then. A part that a new version adds is asked about on that update. Keep .stm32-installer.json with your project, in version control: without it the installer remembers nothing, and asks again.

Files that become yours

A library can bring files that belong to your project rather than to the library: a template of a port layer for you to fill in, a linker script section, a note for your own documentation. Each is a start, made for you to change, so it is created in your project's folder only when it is not there:

Files
  written example/example.h
  created Core/Src/example_port.c  yours from now on

From then on it is yours. An update never writes over it, and says so with kept Core/Src/example_port.c yours. An uninstall leaves it where it is, with kept Core/Src/example_port.c yours, not removed. It is not added to any IDE project either: it is your project's own file. STM32CubeIDE builds every file in its source folders, Core/Src among them, so there it is built with the rest. In a CMake, Keil or IAR project, which list their files one by one, add it the way you add a file of your own. A library can never put such a file over one the installer manages or CubeMX writes, such as CMakeLists.txt, .cproject, the .ioc, the Makefile or stm32xxxx_hal_conf.h, nor outside your project's folder.

Steps a library runs

A library can need something made that no file can carry, such as a key made fresh for each product. It names a step for that, and once its files are in, the installer shows you exactly what will run, and asks:

Run
  make-key     Make the product's key, product.key
               python scripts/make_key.py product.key
               Run it? [Y/n]:

Enter runs it, in your project's folder. A Python script runs with the same Python as the installer, so it works the same on every computer. n leaves it, and the output says how to run it later. In a script, where nobody can answer, it is not run.

Nothing in a library's installer.yml can skip the question. Only you can: --run make-key runs that step without asking, --no-run make-key leaves it, and --yes runs every step and takes the offered answer to every other question, for a script or a build server.

A step that makes a file can name it, and when that file is already in your project the step is skipped without a question, so a second install never makes a second key. A step that names no file and ran before is not asked about again either, until the library changes what it runs. When a step fails, what was installed stays installed, and the installer ends with exit code 1, so a script notices.

Updating a library

Run the same command again. Every file of the library is replaced, so an update updates, and a library you broke by accident is put right. What you wrote between USER CODE BEGIN and USER CODE END is kept, the way STM32CubeMX keeps your code when it generates again:

/* USER CODE BEGIN EXAMPLE_CONFIGURATION */
#define EXAMPLE_SIZE  32
/* USER CODE END EXAMPLE_CONFIGURATION */

Files
  written example/example.h
  written example/example_config.h
  written example/example.c
  kept    example/example_config.h  your USER CODE EXAMPLE_CONFIGURATION

This was an update. The library replaced, what was yours kept.

Put your settings, and anything else you add to a library file, inside those sections. Outside them, the next install replaces it. When that would lose something of yours, the file as it was is saved beside it first, and the report says why:

  saved   example/example.c.20261006-165940.bak  it was changed outside its USER CODE sections

That happens when you edited the library outside a section, when its markers no longer pair, or when the new version no longer has a section you had written in. Take what you need from the .bak file, then delete it. A file nobody touched is left as it is, so your build does not compile it again for nothing.

When a new version keeps its files somewhere else, say in src/, the update moves your project along with it. Your example_config.h goes where the new version looks for it, with your settings between its USER CODE lines still in it, rather than a fresh default turning up next to the header. Files the old version installed and the new one no longer has are removed, so an old example.c cannot end up compiled beside the new one. Keil, IAR and the Makefile are pointed at the new places, and a file the library gained joins its group there. Anything you put in the library's folder yourself is left alone.

Files
  written example/src/example.h
  written example/src/example_config.h
  written example/src/example.c
  moved   example/example_config.h -> example/src/example_config.h  your USER CODE EXAMPLE_CONFIGURATION went with it
  removed example/example.h  no longer part of the library
  removed example/example.c  no longer part of the library

Uninstalling a library

stm32-installer --uninstall example

From your project's root, or with --project. It takes the library out of every IDE project the installer put it in: its block in CMakeLists.txt and the CMakeLists.txt in its folder, its include path, source folder and defines in STM32CubeIDE's .cproject, its group, include path and defines in Keil and IAR, and its lines in the Makefile. Then its files go:

example 2.0.0
out of example

Project
  updated CMake  Removed example from CMakeLists.txt.
  updated STM32CubeIDE  Removed example from the include path, source folders and defines in .cproject.

Files
  removed example/example.h
  removed example/example.c
  saved   example/example_config.h.20261014-163211.bak  it was yours to edit

Done. example is out of the project.

What was yours is saved before it goes, as a .bak beside it: a configuration file the library created, and any file of the library you wrote in between its USER CODE lines or changed by hand. A file you put in the library's folder yourself is left where it is, and the folder with it, and so are the files that became yours. A define another installed library also asks for stays.

A library another one needs stays, and the installer says which needs it:

Error: spif needs osal, so it stays. Uninstall spif first, or name them all on one line.

Name them all, in any order, and each goes after what needs it: stm32-installer --uninstall osal spif. When the library you took out brought another with it that nothing left in the project needs, the last line names it, with the command that takes it out too.

Uninstall works from .stm32-installer.json, the record of what was installed and where. For each library it also keeps where it came from, down to the commit, the blocks written into your files, and the steps that ran. A library put in by hand, or a record that was deleted, leaves it nothing to go on. A library installed by a version before 1.11.0 comes out the same way, but its record does not say what it needed, so nothing is named at the end.

A library that needs another

Some libraries build on another one, such as osal for the RTOS. You install only the one you want, and what it needs goes in first:

example 2.0.0 needs osal >= 1.0.0, which is not in this project. Installing osal first.
...
Done. #include "osal.h" and you are away.
...
Done. #include "example.h" and you are away.

You are asked for its folder like any library. Install a second library that needs osal, and the one already in your project is used:

spif 3.0.0 needs osal >= 1.0.0. osal 1.0.0 is already in this project, kept.

Nothing is asked about it then. A library that goes in this way comes from the release the library needing it names, or else from its default branch. If a library needs a newer osal than the one you have, it is updated where it is, and your setting in osal_config.h, between its USER CODE lines, is kept as always. Everything is fetched before anything is installed, so when a library cannot be had, your project is left as it was.

What you see

A first install from a downloaded zip, into a CubeMX project that builds with both CMake and STM32CubeIDE:

Reading example-main.zip ...

example 2.0.0
What this library does, in one line
middleware / system

Files
  written example/example.h
  written example/example.c
  written example/LICENSE.md
  written example/NOTICE
  written example/README.md
  created example/example_config.h  yours to edit

Project
  updated CMake  Added example to CMakeLists.txt, linked to ${CMAKE_PROJECT_NAME}.
          backup: CMakeLists.txt.20260923-193649.bak
  updated STM32CubeIDE  Added example to the include path and source folders in .cproject.
          Refresh the project in CubeIDE (F5) so it picks up the new files.
          backup: .cproject.20260923-193649.bak

Done. #include "example.h" and you are away.

The first line says where the library came from: Fetching owner/example ... from GitHub, Reading ... from a zip or a folder.

What it does for you

Finds your IDE and wires the library in. CMake, STM32CubeIDE, Keil MDK, IAR and CubeMX's Makefile are all handled, and a project can be more than one of them at once.

CubeMX can write its Makefile again when it generates code, and the library's lines go with it. Run the installer again after that and they are back.

Never loses your settings. What you write inside USER CODE BEGIN and USER CODE END survives every install. Reinstall as often as you like: the library is replaced, your sections are not, and anything else of yours that would be lost is saved to a .bak file first.

Leaves the file looking like you wrote it. An added line copies the indentation, the line ending and the path separator of the lines around it. These files go into version control, so a correct edit that shows up as a whole file diff is still a bad edit.

Keeps your CMakeLists.txt short. Your project gets two lines that never change, and the library's file list lives with the library:

# >>> stm32-installer: example >>>
add_subdirectory(example)
target_link_libraries(${CMAKE_PROJECT_NAME} PRIVATE example_lib)
# <<< stm32-installer: example <<<

The target is called example_lib rather than example, because a project is often named after the library being tried out in it and CMake allows only one target per name. The PRIVATE keyword is dropped when your project links its libraries without one, since CMake refuses to mix the two forms.

The generated example/CMakeLists.txt declares an INTERFACE target, not a STATIC one. That matters: an INTERFACE target hands its sources to whoever links it, so they are compiled as part of your application and inherit its defines and include paths. That is what lets a driver's .c file find main.h and the HAL headers. A STATIC library would not see them and would fail to compile.

Backs up before editing. Every project file is copied to a timestamped .bak first. If the file is not one it recognises, it changes nothing and prints what to click instead.

Edits the project, not a copy of it. IAR leaves a Backup of <name>.ewp beside the real project when it upgrades one, and that copy is a valid project file whose name sorts first. The workspace file is asked which project is the real one, and backups are skipped.

Keeps the repository out of your build. CubeIDE compiles every .c under your project, and a library repository ships a test suite with its own main(). Left in place, that breaks your build with an error that points nowhere useful. So only the listed files are copied, and a repository that sits inside your project has everything else removed from it.

Tells you what the library needs. It reads your stm32xxxx_hal_conf.h and .ioc and warns before you hit a confusing compile error:

Requirements
  check   HAL SPI module is not enabled
          Uncomment HAL_SPI_MODULE_ENABLED in stm32f4xx_hal_conf.h
  ok      GPIO is configured
  check   This library needs an RTOS and none was found
          It expects freertos. Enable it in CubeMX under Middleware.

It cannot switch a peripheral on for you. The .ioc belongs to CubeMX, and editing it behind CubeMX's back goes wrong quietly.

For someone writing a library

Put an installer.yml at the root of the repository. Any library can have one, whoever wrote it, and the one file serves stm32-installer and the other installers alike: each reads its own keys and leaves the others' alone. Releases before 1.9.0 called it library.yml, and the installer still reads that name when there is no installer.yml, so a version pinned with --ref from before the rename installs as it always did. A library that renames its file says requires.installer: 1.9.0.

Every key the installer reads is below, each with what it does. Only name and files are required; leave out what your library does not need. For a working one to copy, see stm32-installer-example.

name: example                      # the library's name, and its folder in the project
description: Driver for the X chip # one line, shown while installing
repository: https://github.com/owner/example  # where it comes from
license: Apache-2.0                # the licence's SPDX name
version: 1.0.0                     # only when the first header has no @version tag

kind: driver                       # driver, middleware, rtos, protocol, filesystem,
                                   # utility or bsp
category: sensor                   # sensor, storage, display, communication, wireless,
                                   # rtos, timing, power, input, math or system
provides: []                       # what it gives other libraries: [rtos] for an RTOS

requires:
  stm32_installer: 2.0.0           # the oldest stm32-installer that reads this file
  libraries:                       # other libraries, installed first when missing
    - osal >= 1.0.0                # no owner means nimaltd; >= the oldest that will do
    - owner/lib                    # anyone else's, as owner/name
    - owner/lib@v1.2.0             # a tag, branch or commit, not its default branch
    - source: owner/other@v2.9.3   # a repository with no installer.yml of its own,
      manifest: manifests/other.yml  # with one this library carries for it
      minimum: 2.9.0               # the oldest version that will do
  hal: true                        # uses the STM32 HAL
  cmsis: false                     # uses CMSIS directly
  rtos: optional                   # none, optional, any, freertos, cmsis-os2 or threadx
  hal_modules: [spi]               # must be enabled in stm32xxxx_hal_conf.h, checked
  peripherals:                     # must be set up in CubeMX, checked in the .ioc
    - type: spi                    # the peripheral, as CubeMX names it
      count: 1                     # how many the library needs
      note: Master mode, 8 bit     # how to set it up, shown when it is missing
  c_standard: c11                  # the C standard the code is written for

install:
  layout: mirror                   # mirror keeps the repository's folders,
                                   # flat (the default) puts every file at the top
  include_dirs: [src]              # include folders, only when the guess is wrong
  defines:                         # given to the whole project, beside USE_HAL_DRIVER
    - EXAMPLE_FAST                 # NAME, or NAME=value

files:
  headers:                         # .h files; the first is the one to #include
    - src/example.h
    - src/example_config.h         # settings in USER CODE sections, kept on update
    - src/port/*.h                 # a wildcard, matched the way a shell would
  sources:                         # .c files, added to the build
    - src/example.c
    - from: src/port/spi.c         # a file that goes somewhere of its own,
      to: port/spi.c               # whatever the layout says

options:                           # parts the user chooses at install
  littlefs:                        # the name typed after --with
    description: LittleFS file system on the flash  # the question asked
    libraries:                     # libraries this part needs, installed first
      - littlefs >= 2.11.0
    files:                         # its files, written like files above
      sources:
        - src/example_lfs.c

once:                              # the project's own files: made when missing,
                                   # then the user's, never replaced or removed
  - from: templates/example_port.c # the file in the repository, to the project's folder
    to: Core/Src/example_port.c    # its place in the project, when not its name

run:                               # steps run once the files are in, each asked about first
  - label: make-key                # the name typed after --run
    description: Make the product's key, product.key  # shown with the question
    script: scripts/make_key.py    # Python, run with the installer's own Python,
    args: [product.key]            # or command: [program, arguments], with no shell
    unless: product.key            # skipped when this file is in the project already

extras:                            # copied as they are, never compiled
  - LICENSE.md                     # LICENSE.md and NOTICE are the default,
  - NOTICE                         # so leave this out if that is all
  - README.md
  - docs                           # a folder is copied whole

The sections below say more about the keys that need it.

Leave version out. The installer reads the version from the @version tag in the file comment of the first header listed, so it is written in the code and nowhere else:

/**
 * @file        example.h
 * @version     2.0.0
 */

requires.stm32_installer is the oldest stm32-installer that reads the file correctly. It is checked before anything else in the file is read, so when a manifest starts using something only a newer installer understands, raising this makes an older one stop with a message saying to update, rather than read the file wrongly without a word. Write it as three numbers, 1.1.0: unquoted, YAML reads 1.10 as the decimal number 1.1, and that is refused.

Before 2.0.0 the key was requires.installer, and it still works. It is named after the installer now, as every other installer's key is, so a reader of the file knows which installer it is for. One thing to know: an installer older than 2.0.0 reads only requires.installer, so a library that has to stop those older ones writes installer: 2.0.0, instead of the new key or as well.

A library that uses anything 2.0.0 brings asks for 2.0.0: once files in the project's folder, run, or a library it needs given with an @ or a manifest. The version also decides where once files go, as below.

requires.libraries lists the libraries yours needs, each as its name, or its name with the oldest version that will do:

requires:
  stm32_installer: 2.0.0
  libraries:
    - osal >= 1.0.0                    # no owner means nimaltd: github.com/nimaltd/osal
    - someone/lib                      # anyone else's, by owner and name
    - someone/lib@v1.2.0               # exactly this tag, branch or commit
    - source: littlefs-project/littlefs@v2.11.0
      manifest: manifests/littlefs.yml # an installer.yml for a repository without one
      minimum: 2.11.0

What is missing or older than that goes in first, from its repository's default branch, or from the tag, branch or commit after its @. What is already in the project at a version that will do is left alone. An installer older than 1.5.0 only names them at the end and installs nothing.

The last form installs a library whose repository has no installer.yml of its own: yours carries one for it, at the path given, written as if it sat at the root of that repository.

defines

Some code reads its settings from a define the compiler is given, not from a file. littlefs is one: it reads a settings header only when built with -DLFS_DEFINES=lfs_defines.h. List such a define under install.defines, and the installer puts it where the project keeps USE_HAL_DRIVER: the generated CMakeLists.txt, every configuration in .cproject, every target in Keil, every configuration in IAR, and C_DEFS in a CubeMX Makefile.

It goes to the whole project, not only to the library's files, because a define can change what the library's header declares. littlefs built with LFS_THREADSAFE has two more members in struct lfs_config, and a file of yours that did not see the define would put every member after them in the wrong place, with no error from the compiler.

Write each as NAME or NAME=value, with letters, digits and _ . + - / only. A space or a quote is refused, since each define is written as it stands into XML, a Makefile and a Keil list. An update takes out a define the library no longer lists and changes one whose value it changed. A define the project already sets to another value is left alone and reported, because it may be the user's own choice.

A library that lists defines must say requires.installer: 1.8.0. An older installer would ignore them without a word, and the library would build with its defaults.

options

A part of your library that not every project wants, such as spif's port for LittleFS, goes under options. Each is asked about at the first install, and its files and the libraries it needs come only with a yes:

options:
  littlefs:                         # the name typed after --with
    description: LittleFS file system on the flash
    libraries:
      - littlefs >= 2.11.0          # installed first, like requires.libraries
    files:
      headers:
        - src/spif_lfs.h
      sources:
        - src/spif_lfs.c

The description is the question, so write it as what the part is. The name is lower case letters, digits, _ and -, since people type it.

An option's files are your library's files like any other. They land in the same folder, are replaced on every install with their USER CODE sections kept, and are removed by the update that turns the option off. So a file belongs to the library or to one option, never both, and a manifest that lists one twice is refused. Every option's files are downloaded whichever are taken, so an answer never needs a second download. Installed in place, the files of an option not taken are deleted from the folder, since STM32CubeIDE would otherwise compile them.

--with and --without apply to the libraries named on the command line. A library that comes in because another needs it keeps the answer it had, and an option of it that was never chosen stays out.

A library with options must say requires.installer: 1.8.0. An older installer would install it without them, and never ask.

USER CODE sections

Mark what the user is meant to change, the way STM32CubeMX marks its generated files. A configuration header, for instance, lists like any other header and carries its settings in a section:

/* USER CODE BEGIN EXAMPLE_CONFIGURATION */
#define EXAMPLE_SIZE  8
/* USER CODE END EXAMPLE_CONFIGURATION */

Every install copies all the library's files again and carries each section's text from the user's copy into the section of the same name. So keep a section's name from one version to the next. A section a new version adds starts with what you ship in it, and what is inside a section is never changed by an update, defaults included. Each name is one word, used once per file. Markers that do not pair, a BEGIN with no END or one inside another, stop the install before anything is written.

A library that moves a file from once into files this way must say requires.installer: 1.7.1. An older installer replaces the file without keeping its sections, and the user's settings would be lost. 1.7.0 keeps them, but takes a marker mentioned in a sentence for a real one, so a README or a comment that names the markers stops it.

once

A file that belongs to the project rather than to the library goes under once: a port layer someone fills in, a template of their own code, anything that is the library's to start and theirs from then on. It is created in the project's folder when it is missing, and is the user's from then on: never replaced by an update, never removed by an uninstall.

once:
  - from: templates/example_port.c     # the file in the repository
    to: Core/Src/example_port.c        # its place from the project's folder

from is a file of the library's repository, and to its place from the project's folder, its own name when left out. The file is the project's own, so the installer adds it to no IDE project. STM32CubeIDE builds every file in its source folders, such as CubeMX's Core/Src, and has Core/Inc on its include path; a CMake, Keil or IAR project lists its files one by one, and the user adds it there. Say so in your README.

Each is checked before anything is written. to stays inside the project, names none of the files the installer manages or CubeMX writes, CMakeLists.txt, .cproject, .project, the Keil and IAR projects, the Makefile, the .ioc, stm32xxxx_hal_conf.h and the record, and does not point into the library's own folder, which every install replaces. Two entries with one to are refused. And from, like every file a manifest names, stays inside the library: no .., no absolute path, no link leading out of it.

Settings the user changes belong in USER CODE sections of the library's own files, not in once.

Libraries written for an installer before 2.0.0 keep the meaning once had then, so a release tagged before still installs and builds as it did: the file lands in the library's own folder, at its top whatever the layout, and is added to the build. The version requires asks for decides which meaning a manifest gets. The rest of this section is about that older meaning.

With mirror, give a configuration header a to that puts it beside the header including it. At the top, the include path the layout gives would not reach it.

One catch worth knowing: #include "x.h" searches the including file's own folder before any include path, so a config sitting beside the header that includes it will always win. That is fine when the library's tests use the same config, and it is why the NimaLTD libraries test against a setting such as EXAMPLE_MAX_TASKS rather than a fixed number.

The include path follows the layout on its own: flat gives the library folder, mirror gives wherever the headers landed. Add install.include_dirs only when that guess is wrong.

run

Something no file can carry, such as a key made fresh for each product, is made by a step under run. It runs once the library's files are in, in the project's folder:

run:
  - label: make-key
    description: Make the product's key, product.key
    script: scripts/make_key.py
    args: [product.key]
    unless: product.key

• label names the step, typed after --run, in lower case letters, digits, _ and -.
• description is shown with the question.
• script is a Python file of the library, run with the installer's own Python, so it works on every computer the installer does. Or command, a program and its arguments as a list, [openssl, rand, -out, product.key, "32"], run as it is, with no shell. One of the two, never both.
• args are added after either.
• unless names a file the step makes, from the project's folder. When it is there, the step is skipped without a question.

The step sees where things are in its environment: INSTALLER_PROJECT, the project's folder, INSTALLER_LIBRARY, the library's folder in it, INSTALLER_NAME and INSTALLER_VERSION, the library's, and INSTALLER, stm32-installer.

The user is always asked, and is shown exactly what will run. Nothing in the manifest can turn that off, only the user, with --run or --yes. A step with no unless that ran before is not asked about again until what it runs changes, so make it safe to run twice anyway. A step that fails leaves the install as it is and ends the run with exit code 1.

Easy to get wrong

A .c file is built only when it is listed. A port layer under files.sources is built like any other source. One under once, for a manifest asking for 2.0.0, is the project's and is added to no IDE project, so a port layer every user needs belongs in files, with its settings in USER CODE sections. Getting this wrong fails at link time with undefined references and no clue why.

kind: rtos implies provides: [rtos]. A library that is an operating system satisfies another library's RTOS requirement, whether or not you remembered to write it down.

LICENSE.md and NOTICE travel with the code. The Apache licence requires it, and the NOTICE file is what carries your attribution into someone else's product. They are never removed during cleanup.

That is everything the repository needs. There is no installer script to copy in: put the install lines from the top of this page in your README, with your library's name in them.

Running the tests

The shared core, core-installer, is a git submodule in src/core, so clone with it:

git clone --recursive https://github.com/nimaltd/stm32-installer.git

In a clone made without --recursive, git submodule update --init fetches it. Then:

python test/run_tests.py

Support

I write these tools and libraries in my own time and give them away, because good tools should be easy to get. If this one saved you an afternoon, there are two things that genuinely help:

Star the repo. It costs you one click, it helps other engineers find the tool, and it is the main reason I keep going.

Buy me a coffee on Ko-fi. Any amount is a real motivation to keep writing, documenting and maintaining this work.

License

Apache License 2.0. See LICENSE.