Skip to content

Validate coordinates and repository precedence during artifact resolution - #12952

Draft
slachiewicz wants to merge 6 commits into
apache:maven-3.10.xfrom
slachiewicz:pr/artifact-resolution-3.10.x
Draft

Validate coordinates and repository precedence during artifact resolution#12952
slachiewicz wants to merge 6 commits into
apache:maven-3.10.xfrom
slachiewicz:pr/artifact-resolution-3.10.x

Conversation

@slachiewicz

@slachiewicz slachiewicz commented Aug 30, 2026

Copy link
Copy Markdown
Member

Coordinates and repository lists arriving from a resolved artifact descriptor are validated before use.

  • Repository precedence. Repositories declared by a resolved model no longer replace a same-id repository supplied by the request or session. A repository the model itself declared is still refreshed in place, for example once its URL is interpolated.
  • Relocation coordinates. Relocation groupId/artifactId/version are checked against the artifact-coordinate character set before a RelocatedArtifact is built; malformed components fail the descriptor read rather than entering resolution. A follow-up commit covers the remaining rejected character classes.
  • Metadata version tokens. Version and snapshot-timestamp tokens read from maven-metadata.xml are validated before being spliced into a resolved version, so malformed metadata fails the read rather than producing a malformed coordinate. The rejection message names the field.

Each change is a separate commit.


Draft while related code paths are reviewed — the same validation may be needed in sibling implementations of these interfaces, and I would rather establish that before asking for review time.

A systemPath change was removed from this PR: it broke MavenITmng4379TransitiveSystemPathInterpolatedWithEnvVarTest, which covers deliberate MNG-4379 behaviour — a dependency resolved from a repository declaring its own system-scope dependency with the path interpolated from an environment variable.

@slachiewicz slachiewicz added this to the 3.10.0 milestone Aug 30, 2026
@slachiewicz
slachiewicz marked this pull request as draft August 30, 2026 18:54
slachiewicz and others added 5 commits August 30, 2026 21:00
Repositories declared by a resolved model no longer replace a same-id repository
supplied by the request or session; the ids supplied at resolver construction keep
their precedence on the replace pass. A repository the model itself declared is
still refreshed in place, e.g. once its URL has been interpolated.
Relocation groupId/artifactId/version taken from a resolved model are checked
against the artifact-coordinate character set before a RelocatedArtifact is
built; invalid components fail the descriptor read instead of entering
resolution.
Version tokens adopted from repository metadata are checked before use, in both
the version resolver and the version range resolver. Metadata carrying a token
that is not a valid coordinate component is treated as invalid, so resolution
falls back to the requested base version and other repositories still resolve
normally.
Applies the same version-token check used by the aether version and
version-range resolvers to the maven-compat metadata read path: a token
that is not a valid coordinate component is treated as invalid, and the
read is skipped rather than merged in.

Also carries the version-range resolver's own check onto this branch and
adds a resolution test for it, since its sibling change on the aether
side shipped without one.

Claude-Session: https://claude.ai/code/session_011PdF7DSMbwHn1U9xMSrKN9
@slachiewicz
slachiewicz force-pushed the pr/artifact-resolution-3.10.x branch from 95b5890 to e5a65b3 Compare August 30, 2026 19:01
…data

DefaultMetadataReader now checks parsed metadata the same way
DefaultVersionResolver already does: version tokens (release, latest,
versions, snapshot versions, snapshot timestamp) and plugin entries'
artifactId and prefix must be valid coordinate components, rejecting
"..", '/', '\\', ':' and ISO control characters. Callers that read
metadata through this class already treat a read failure as invalid
metadata, so a rejected document degrades gracefully instead of being
adopted as a resolved version or plugin artifactId.
@slachiewicz
slachiewicz marked this pull request as ready for review August 30, 2026 21:27

@gnodet gnodet left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Solid security-hardening PR that adds correct validation of metadata tokens and relocation coordinates against path-traversal and coordinate-injection attacks. The 3.10.x adaptation is well done — validation is correctly placed in the legacy code paths (DefaultRepositoryMetadataManager, DefaultMetadataReader) appropriate for this branch.

Minor suggestions (non-blocking):

  • DefaultVersionResolverTest (line 105): A comment appears truncated mid-sentence ("so the") — the master variant has the full text: "so the metadata is treated as invalid and resolution falls back to the requested base version." Likely a copy-paste artifact during the backport.
  • The validation logic duplication across files is acceptable for a stabilized branch where minimizing cross-module changes is desirable.

The validation character set (rejecting /, \, :, ISO control chars, exact ..) is well-chosen for the threat model. Each resolver validates exactly the fields it consumes, which avoids unnecessary strictness. All CI checks pass.

This review was generated by an AI agent and may contain inaccuracies. Please verify all suggestions before applying.

Claude Code on behalf of Guillaume Nodet

@slachiewicz

Copy link
Copy Markdown
Member Author

Related work in maven-resolver, for reviewer awareness: apache/maven-resolver#2087 adds the same kind of component checking one layer down — PathUtils.validatePathComponent gains : rejection, and a new validateDotSeparatedPathComponent covers group ids that get dot-expanded into path segments.

The overlap is deliberate for now. This PR checks in Maven because the pinned resolver does not, and a Maven line only picks up the resolver-side change when its resolverVersion moves — this line currently pins 2.0.22, and 4.0.0-rc-7 is expected to pin 2.0.23, which would include #2087.

Worth a reviewer decision on whether the check should stay in both places once that pin lands, or move down to the resolver.

This comment was created with AI assistance.

@gnodet
gnodet marked this pull request as draft August 31, 2026 07:27
@slachiewicz
slachiewicz marked this pull request as ready for review August 31, 2026 07:46
@slachiewicz
slachiewicz marked this pull request as draft August 31, 2026 13:03
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants