Hi Guillaume,
I share Matthias' and Maarten's reservation about Maven core "repairing"
a misconfiguration from the shadows -- but the underlying need (building
an older project on a modern JDK) is real and worth solving. I think the
disagreement dissolves if we split it across three responsibilities
instead of putting everything in core:
1. Core = diagnose, don't repair.
When the running JDK cannot honour the requested source/release level,
emit one clear, actionable error: the max JDK that supports the level,
and a single command to fix it. That satisfies the "read the error and
act" position without a cryptic javac failure.
2. Discovery + selection stays in maven-toolchains-plugin.
The plugin already owns this: `select-jdk-toolchain` discovers and
selects a matching JDK at build time, `display-discovered-jdk-toolchains`
and `generate-jdk-toolchains-xml` round it out, and 3.3.0 just extended
discovery (asdf, more Windows locations, dedup across env vars). Moving
the discoverer into core would duplicate -- and eventually orphan --
code the plugin is actively investing in. So to your question in the PR
("does this make the plugin's discovery redundant?"): I'd argue the
opposite -- keep discovery in one place, the plugin, and have core only
diagnose.
3. The real gap none of the current pieces fill: what to do when *no*
compatible JDK is installed at all.
This is where a "propose" (and, opt-in, "provision") capability would
help -- and it belongs in tooling, not core. Concretely, a new goal in
maven-toolchains-plugin that, given the required level, queries the
foojay Disco API for a matching distribution. Default output is
actionable only: an `sdk install java <id>` line and/or a download URL,
plus optionally writing the toolchains.xml entry. Actually downloading a
JDK stays opt-in and checksum-verified. This mirrors what Gradle already
does via its foojay-resolver convention, and foojay is vendor-neutral
(Temurin/Zulu/Liberica/Corretto/...).
Your mvnup idea fits cleanly on top: mvnup can write the <toolchain>
declaration into the project using that same discover/propose engine, so
the project becomes self-describing for the next contributor.
For the project-less case (no reactor at hand), the same proposal action
would sit naturally in the Maveniverse Toolbox / mvnx family, which is
already in this space (MWM came up on the #12646 thread).
Cheers,
Gerd
> On 3. Aug 2026, at 23:10, Guillaume Nodet <[email protected]> wrote:
>
> The starting point was to be able to build old projects with Maven 4.
> Would enhancing mvnup to add a toolchain declaration on the project,
> thereby leveraging the m-toolchain-p be a better solution ?
>
> Le lun. 3 août 2026 à 21:10, Maarten Mulders <[email protected]> a
> écrit :
>
>> Hi,
>>
>> Agree with Matthias. Earlier ideas on "troubleshooting common errors
>> through Maven Core" have been rejected for reasons like "people should
>> read the error messages and take corresponding action, rather than Maven
>> trying to diagnose or even repair it". It would be strange to now
>> include such features. I would prefer having clear(er) error messages if
>> that is the source of problem.
>>
>> Also, I see many people blindly ignore warnings, so that would
>> contribute to the scenario that Matthias outlined: "it works on my
>> machine", while in fact I ignored the warning that says Maven switched
>> to an auto-discovered JDK 11 that someone else on my project does not have.
>>
>> Thanks,
>>
>> Maarten
>>
>> On August 3, 2026 at 17:14, Matthias Bünger wrote:
>>> Hi,
>>>
>>> I'm not convinced in trying to "fix" configuration errors in users
>>> system/project from the "shadows". I think a clear error (and if
>>> possible giving information what's wrong/how to fix) supports the user
>>> more in really fixing the problem. Otherwise we support "it worked on
>>> my machine". I can also think that it's massively increasing
>>> maintenance effort on our side due question why it works here, but not
>>> there or paths to look at on different OS.
>>>
>>> Matthias
>>>
>>> Am 03.08.2026 um 14:18 schrieb Guillaume Nodet:
>>>> Hi all,
>>>>
>>>> I've opened a draft PR for a quality-of-life feature in Maven 4.1.0:
>>>> automatic JDK toolchain selection when the running JDK cannot compile
>>>> the
>>>> project's declared source/release level.
>>>>
>>>> PR: https://github.com/apache/maven/pull/12633
>>>>
>>>> The problem
>>>>
>>>> When a project declares <maven.compiler.source>6</maven.compiler.source>
>>>> (or uses <release>, <targetVersion>, or compiler plugin
>>>> <configuration><source>), and you run Maven with JDK 21 — which dropped
>>>> --source 6 support — the build fails with a cryptic javac error. The
>>>> user
>>>> has to figure out they need to install a compatible JDK and either
>>>> configure toolchains.xml or add the maven-toolchains-plugin to their
>>>> build.
>>>> This is a frequent stumbling block, especially when maintaining older
>>>> projects.
>>>>
>>>> The solution
>>>>
>>>> Maven now automatically detects the incompatibility and searches for a
>>>> compatible JDK — first in configured toolchains (toolchains.xml),
>>>> then by
>>>> lazily discovering JDK installations on the filesystem. If a
>>>> compatible JDK
>>>> is found, it's selected as the compilation toolchain and a warning is
>>>> emitted:
>>>>
>>>> [WARNING] Project requires --source 6 which is not supported by JDK 21.
>>>> [WARNING] Automatically selected JDK 11 (discovered at
>>>> /usr/lib/jvm/java-11) for compilation.
>>>>
>>>>
>>>> This is a zero-cost feature: the auto-selection logic only runs when the
>>>> running JDK genuinely cannot handle the project's source level. Normal
>>>> builds are completely unaffected.
>>>>
>>>> Design decisions worth discussing
>>>>
>>>> 1. Filesystem discovery — The discoverer scans well-known locations
>>>> (SDKMAN, IntelliJ .jdks/, Gradle, jEnv, JBang, asdf, mise, OS-specific
>>>> paths like /usr/lib/jvm). Version is read from the JDK release file — no
>>>> java processes are spawned. This is essentially what
>>>> maven-toolchains-plugin's ToolchainDiscoverer does, but moved into
>>>> core and
>>>> made lazy. Does this make the plugin's auto-discovery redundant?
>>>> Should we
>>>> deprecate it?
>>>>
>>>> 2. Source level detection — We read the source level from multiple
>>>> places
>>>> in priority order: Model 4.1.0 <source><targetVersion>, then
>>>> maven.compiler.release/maven.compiler.source properties, then compiler
>>>> plugin <configuration><release>/<source>. Is this the right
>>>> precedence? Are
>>>> there other places we should check?
>>>>
>>>> 3. "Newest compatible" strategy — When multiple compatible JDKs are
>>>> found,
>>>> we pick the newest one (highest major version that still supports the
>>>> required source level). The rationale is that a newer JDK gives better
>>>> performance and diagnostics. Should this be configurable?
>>>>
>>>> 4. Compat layer — The v3 ToolchainManagerFactory bridge now passes the
>>>> discoverer through, so plugins using the v3 API also benefit. This felt
>>>> important for the transition period.
>>>>
>>>> CI is green on all platforms (Linux/macOS/Windows × JDK 17/21/25).
>>>>
>>>> Feedback and reviews welcome.
>>>>
>>>> Guillaume
>>>>
>>>
>>> ---------------------------------------------------------------------
>>> To unsubscribe, e-mail: [email protected]
>>> For additional commands, e-mail: [email protected]
>>>
>>
>> ---------------------------------------------------------------------
>> To unsubscribe, e-mail: [email protected]
>> For additional commands, e-mail: [email protected]
>>
>>
>
> --
> ------------------------
> Guillaume Nodet
--
Gerd Aschemann (er/he) --- Veröffentlichen heißt Verändern (Carmen Thomas)
+49/173/3264070 -- [email protected] -- https://aschemann.net
