feat(mvn): add Maven (Java) filter module — test, compile, checkstyle:check, dependency:tree by mariuszs · Pull Request #1089 · rtk-ai/rtk

mariuszs · 2026-04-08T20:07:59Z

Summary

Maven/Java ecosystem support for RTK.

Subcommands

Command	Savings
`rtk mvn test`	99%
`rtk mvn verify`	95%
`rtk mvn compile`	85%
`rtk mvn checkstyle:check`	90%
`rtk mvn dependency:tree`	70%
`rtk mvn clean`	95%

Plus rtk mvnd — same filters for Maven Daemon. Unrecognized goals pass through with metrics.

For test/verify, RTK reads target/surefire-reports/ and target/failsafe-reports/
XML to surface structured failure details (class, method, exception, message) with
compact stack traces — Caused-by chains preserved, framework frames
(org.junit.*, java.base/*, org.apache.maven.*) collapsed to keep app frames
visible. Auto-detects mvnw wrapper.

Test plan

cargo fmt --all && cargo clippy --all-targets && cargo test --all passes
Snapshot tests (insta) for every filter
Token savings verified on real Maven fixtures
Manual test on real Maven project

CLAassistant · 2026-04-08T20:08:06Z

All committers have signed the CLA.

Ckram · 2026-04-17T01:29:08Z

@mariuszs could you add support for mvnd (maven daemon) ?

Adds a parallel 'rtk mvnd' subcommand that reuses the mvn filters (test, compile, checkstyle:check, dependency:tree, passthrough) but invokes the mvnd binary directly (bypassing mvnw, which the daemon does not use). - New MvnBinary enum (Mvn auto-detects mvnw; Mvnd always uses mvnd) - Threaded through run_test/run_compile/run_checkstyle/run_dep_tree/run_other - Tracking labels keep 'mvn' and 'mvnd' separate in rtk gain - Discover rule added for mvnd → rtk mvnd rewrites Addresses review comment on PR rtk-ai#1089.

Addresses review items from sibling PR rtk-ai#368 that apply to this PR: - P1 #4 (code duplication): run_checkstyle, run_dep_tree, and the compile-like runner shared ~15 lines of near-identical cmd construction + runner::run_filtered plumbing. Extract run_simple_goal(binary, goal, tee_slug, filter, args, verbose) as the shared shell. Drops the compile_like_labels indirection (slug lookup now inline in run_compile_like). run_tests_like stays separate because XML enrichment needs cwd + app_pkgs + closure capture. - Minor #subcmd_savings: mvn and mvnd discover rules had empty subcmd_savings. Populate with measured per-goal ratios (test 99%, verify 95%, checkstyle 90%, dependency:tree 70%, compile 85%) so `rtk discover` reports accurate opportunity per goal instead of the rule-level 90%.

`rtk mvn clean` previously fell through to Passthrough, producing ~33 lines of ANSI-colored startup noise (prefix loading, model warnings, plugin separators) for a goal whose useful output is a single `Deleting <path>` line plus BUILD SUCCESS. Filter now collapses to one line: mvn clean: deleted /path/to/target (1.4 s) Multi-module reactor reports `deleted N targets`; zero-target case says `nothing to clean`. When clean is combined with a failing goal (e.g. `mvn clean compile`), `[ERROR]` lines are preserved so the compile failure reason stays visible. Savings measured at ≥90% on the real fixture. - New MvnCommands::Clean variant; dispatch_mvn routes to run_clean via the existing run_simple_goal helper. - Per-goal savings (95%) added to MVN_SUBCMD_SAVINGS for rtk discover.

Multi-module compile output was dominated by the `Reactor Build Order:` block and per-module `Reactor Summary` status lines (~40 redundant tokens on a green 5-module build), and javac `[ERROR] <path>:[line,col]` locations were emitted twice — inline during compilation and again in the trailing help block — nearly doubling the failure-path output. Changes in filter_mvn_compile: - Skip `Reactor Build Order:` block entirely (redundant with per-module Building headers that are already stripped). - Collapse Reactor Summary: emit a compact `Reactor: N modules — M SUCCESS, K FAILURE (name, ...)` line only when something failed; all-green reactors rely on the surviving `BUILD SUCCESS` line. - Dedup `[ERROR] <path>:[line,col]` occurrences by (path, line, col). When a duplicate fires, swallow the indented `[ERROR] symbol:/location:/ required:/found:/reason:` context lines that mirror an earlier javac explanation emitted without the `[ERROR]` prefix. Also: broaden `deprecated` check to `deprecat` (catches both `deprecated` and `deprecation` variants); strip `/pom.xml` path references (previously only literal `from pom.xml` matched); drop the Help-footer line `For more information about the errors...` which wasn't in the boilerplate list. Savings on fixtures adapted from rtk-ai#782 (attribution in filenames): mvn_pr782_compile_success_raw 74.6% → 97.8% (+23pp) mvn_pr782_compile_fail_raw 49.1% → 79.7% (+31pp) mvn_pr782_test_pass_raw 98.7% → 98.7% (unchanged, correct) mvn_pr782_test_fail_raw 93.4% → 93.4% (unchanged, correct) Four new regression tests validate multi-module accumulation (20 tests across 6 modules must not report only the first module's count), failure enumeration uniqueness, reactor collapse, and javac error dedup.

Addresses the simplify review on 4e8db62 — behaviour and test output unchanged (savings on PR782 fixtures identical: 98.7 / 93.4 / 97.8 / 79.7). - `result: Vec<String>` → `String` with capacity hint. Previous code allocated one heap `String` per kept line and joined at the end; the original iterator-chain body (pre-state-machine) was zero-alloc via `Vec<&str>`. Incremental append on a preallocated `String` restores zero-allocation-per-line in the hot path. - Dedup key `HashSet<(String, String, String)>` → `HashSet<&str>` keyed on the javac error prefix slice (borrows `clean`). Drops 3 String allocations per `[ERROR] path:[L,C]` line to 0. - `reactor_modules: Vec<(String, String)>` → `Vec<(&str, &str)>` tied to `clean`'s lifetime. Drops 2 allocations per reactor summary line. - Gate `COMPILE_ERROR_LOCATION_RE` with `line.starts_with(ERROR_TAG)` so the regex doesn't fire on every surviving [INFO]/bare line in a green build. - Swap `COMPILE_ERROR_LOCATION_RE.captures()` for `.find()` — we only need the matched substring for the dedup key, not the capture groups. - `REACTOR_SUMMARY_LINE_RE`: drop the dead `time` capture group (`format_reactor_summary` only reads name+status). - `format_reactor_summary`: use `write!` macro against the output `String` instead of `push_str(&format!(...))` — drops 2 intermediate allocations on the BUILD FAILURE path. - Tighten new `is_maven_boilerplate` entry from the generic "For more information about" to the Maven-canonical "For more information about the errors" — reduces false-positive surface on plugin log lines. - Move `use std::collections::HashSet` to top-of-file imports (matches the rest of the module).

Running our filter against the competing JVM PR's fixtures surfaced real gaps on a multi-module Spring Boot success build with pgpverify-maven-plugin: compile savings were only 50.8% (below our 60% target). This pushes them to 98.1% and closes a cosmetic stack-bleed on simple test failures. filter_mvn_compile: - Route is_mvn_startup_noise for the banner emitted by `mvn -V` (Apache Maven x.y.z, Maven home, Java version, Default locale, OS name) and for SLF4J static-binder complaints — previously all kept as bare text. - Add INFO_NOISE_PATTERNS for pgpverify-maven-plugin (`Verifying`, `Key server(s)`, `Create cache directory`, `Artifacts were already validated`, `artifact(s) in repository`), maven-resources-plugin (`encoding to copy`, `skip non existing resourceDirectory`) and maven-checkstyle-plugin clean-audit output (`Starting audit`, `Audit done`, `Checkstyle violations`). REACTOR_BUILD_ORDER_RE: - Also match the mvn 3.9.x default `<name> <version>` format (version token starting with a digit), not only the classic `[pom|jar|war|ear]` suffix variant. Previously the per-module listing escaped the Build Order collapse. filter_mvn_tests_with_goal: - Terminate the current FAILURE stack collection when the next class emits its `Running <class>` marker, so the marker does not bleed into the failure's details block. Tests: - 2 fixtures adapted from rtk-ai#1241 (pgp+multimodule success, single-module test failure). - 3 regression tests pinning ≥85% savings on the compile fixture, absence of banner/JVM/plugin noise, and no stack bleed. Full suite: 1714 passed.

…dvisory Observed on a real build: a ~100-word maven-resources-plugin `[INFO]` advisory ("The encoding used to copy filtered properties files have not been set…") was slipping through the filter because our pattern was `encoding to copy`, which does not match the `encoding used to copy` variant. Broaden the substring to `copy filtered` so both forms collapse: - "Using 'UTF-8' encoding to copy filtered resources." - "The encoding used to copy filtered properties files have not been set…" Adds a 3-line regression fixture and a test that pins the advisory disappears while `BUILD SUCCESS` and `Total time` are preserved. Full suite: 1715 passed.

When `mvn test` is run against a project that fails to compile, the build never reaches the `T E S T S` marker and our test filter was returning "mvn test: no tests run" — hiding the real compile errors and leaving the user with no signal about why the build failed. filter_mvn_tests_with_goal now checks for `BUILD FAILURE` before returning the cheerful short-circuit. When present, it falls back to filter_mvn_compile, which correctly renders the javac error block. Other early-exit cases (validate-only phases, plugin-only runs that genuinely produce no tests) still return the concise "no tests run" line, since those do not carry `BUILD FAILURE`. Observed on the user's real `api` repo `mvn test` log (180 lines, 1029 tokens): before: `mvn test: no tests run` (hides 18 compile errors) after: 67 lines, 338 tokens, 67.2% savings, full error block shown Adds the 180-line raw log as a regression fixture and a test pinning the new behaviour. Full suite: 1716 passed.

Observed on a real Spring Boot project using Google Cloud's private Maven repo via `artifactregistry-maven-wagon`. Compile-failure logs carry large amounts of non-actionable chatter that drove savings down to 41.6% on a 259-line fixture. Adds four detectors: - INFO_NOISE_PATTERNS: * `is present in the local repository, but cached` — dozens of ~300-char `[INFO]` lines explaining which remote repos each cached artifact would otherwise resolve from. Never needs action. * `Initializing Credentials`, `Application Default Credentials`, `Refreshing Credentials` — GCP auth lifecycle lines emitted by the wagon during download. - is_mvn_startup_noise: * `JUL_LOG_HEADER_RE` — detects java.util.logging header format (`Apr 18, 2026 12:19:27 AM com.google.auth.oauth2.X warnY`) so that the next-line WARNING body has the right context. * `BARE_PLUGIN_WARNING_PREFIXES = ["WARNING: Your application has authenticated"]` — body of the GCP end-user-credentials warning. Fixture: 259-line anonymized capture of the user's real failing build. Savings: 41.6% -> 86.0% (2419 -> 339 tokens). All compile errors preserved; BUILD FAILURE + Total time preserved. Full suite: 1718 passed.

On every successful build, three maven plugins emit one-line notifications that the user never acts on: - maven-enforcer-plugin: one `[INFO] Rule N: <fqcn> passed` line per enforced rule (typically RequireMavenVersion, RequireJavaVersion, RequirePluginVersions). - githook-maven-plugin: `[INFO] Installing commit-msg hook into <.git/hooks>` on every build. - maven-compiler-plugin: `[INFO] Changes detected - recompiling the module!` immediately before the real compile step. Adds ENFORCER_RULE_PASSED_RE (regex, anchored and requires the trailing ` passed`) and substring patterns for the other two. Saves 3–8 lines on a typical build. On the 259-line artifactregistry fixture, savings go from 86.0% → 86.8% (2419 → 319 tokens, 64 lines). Compile errors and `BUILD FAILURE` still preserved. Full suite: 1719 passed.

Three large fixtures were padded with redundant noise that did not strengthen any assertion — the filter strips a *pattern*, not a count. Cut the repetitions while preserving every signal the assertions check: - mvn_compile_artifactregistry.txt: 259 -> 107 (-58%) - mvn_test_compile_failure.txt: 180 -> 52 (-71%) - mvn_dep_tree_beacon.txt: 652 -> 281 (-57%) Beacon snapshot regenerated (28 direct deps with intact transitive counts). All three preserve their savings thresholds (>=80%, >=60%). Renamed PR-tagged fixtures to behavior-based names so their purpose stays clear once the PR number is forgotten: - mvn_pr782_*_raw.txt -> mvn_{compile,test}_reactor_*.txt - mvn_pr1241_compile_pgp_* -> mvn_compile_pgp_multimodule.txt - mvn_pr1241_test_failure_* -> mvn_test_failure_stack_isolation.txt Test fn names updated to match. PR provenance moved into module-level comments. cargo test --all: 1719 passed, 6 ignored.

…l review Two HIGH findings from Codex adversarial review on PR rtk-ai#1089: 1) Reactor / `-pl <module>` runs lost failure details `enrich_with_reports` only probed `<cwd>/target/{surefire,failsafe}-reports`, so reactor builds and `mvn -pl <module>` from the repo root silently fell back to "no XML reports found" despite fresh per-module reports existing under `<module>/target/...`. Added `discover_report_dirs()` walking depth-1 module dirs (skipping `.*`, `target`, `src`, `node_modules`, `build`, `out`) and a `collect_reports()` merger that aggregates SurefireResults across every module that produced output. 2) Forked-VM crashes / surefire aborts reported as `0 passed` When the parser entered Testing but BUILD FAILURE arrived before a `Results:` block (forked-VM crash, surefire timeout, plugin abort), `cumulative` stayed at zero and the success branch emitted a synthetic `0 passed` summary — silently hiding a hard failure. Added a guard: when `BUILD FAILURE` is present and no failure was parsed, fall back to `filter_mvn_compile` so the actual error block surfaces to the user. Tests: - enrich_reactor_finds_per_module_reports — depth-1 module walk - enrich_reactor_real_world_anliksim_multi_module — real fixture from anliksim/maven-multi-project-example (4-module reactor, module1 fails, module2 SKIPPED) using full pipeline (filter -> enrich) - enrich_reactor_skips_dot_dirs_and_node_modules — walker boundary - test_forked_vm_crash_never_emits_synthetic_pass — surefire booter exit-137 scenario cargo test --all: 1723 passed, 6 ignored.

The reactor multi-module fixtures landed with personal paths (`/home/mariusz/...`) and a project-specific package name (`anliksim`). Anonymize and compact: - anliksim -> com.example.app (package + groupId) - /home/mariusz/projects/sample/... -> /workspace/example-project/ Compaction (filter still asserts the same signal — class.method, ComparisonFailure message, BUILD FAILURE, no-reports-hint absent): - mvn_test_reactor_module_failure.txt: 130 -> 56 (-57%) Drops repeated WARNING boilerplate, "Loaded N auto-discovered prefixes" repeats, BoM "Copying" pom lines, "skip non existing resourceDirectory" trivia, and the 30-line JUnit reflection stack trace (5-frame essence kept). - TEST-com.example.app.SpeakerTest.xml: ~70 -> 13 (-81%) - TEST-com.example.app.AppTest.xml: ~67 -> 4 (-94%) Drops the surefire <properties> block (60+ properties full of absolute paths and personal info — the parser never reads them). cargo test --all: 1723 passed, 6 ignored.

XML reports are authoritative (stack trace, captured output), so strip the text-filter's `Failures:` block whenever surefire/failsafe XML has failures. Fallback path (XML missing) still surfaces the text block so users keep context from stdout parsing.

When XML is missing and the text parser is the only failure source, its output now matches the XML-enriched format: 1. Shorten exception FQN on the first detail line (`org.junit.ComparisonFailure:` -> `ComparisonFailure:`) via `shorten_exception_header`, mirroring `failure_kind_label`'s `rsplit('.').next()` logic. 2. Drop non-app stack frames when pom groupId is known (`app_packages` threaded through `filter_mvn_tests_with_goal`). `org.junit.Assert.*`, `org.hamcrest.*`, and similar assertion-lib frames were leaking through the legacy whitelist — now gated by `is_framework_frame_ext` which also rejects any `at <pkg>...` frame not in `app_packages`. Empty `app_packages` falls back to the legacy whitelist — no regression for fixtures and callers that do not detect a groupId.

Upstream introduced src/cmds/jvm/ (gradlew). Consolidate the Maven filter (mvn_cmd, pom_groupid, stack_trace, surefire_reports) under the same jvm/ umbrella so the module layout matches upstream. - move src/cmds/java/* -> src/cmds/jvm/* - crate::cmds::java:: -> crate::cmds::jvm:: - rename insta snapshots __java__ -> __jvm__ and fix source: headers - drop `pub mod java;` (jvm/mod.rs automod already covers the dir) - docs: src/cmds README, jvm README title + gradlew note, module map

The old prose described the pre-multi-goal Clap sub-enum with external_subcommand routing, which no longer exists. Describe the actual dispatch -> route_goal flow (0 goals passthrough, 1 goal dedicated filter, >=2 goals multi-goal aggregator).

mariuszs changed the base branch from master to develop April 8, 2026 20:11

mariuszs force-pushed the feat/mvn-rust-module branch 9 times, most recently from 5d7c9a2 to 8baf6bd Compare April 14, 2026 12:46

mariuszs changed the title ~~feat(mvn): add Maven filter module — test, build, dependency:tree~~ feat(mvn): add Maven (Java) filter module — test, build, dependency:tree Apr 14, 2026

mariuszs force-pushed the feat/mvn-rust-module branch from 6edc416 to b44c8f7 Compare April 14, 2026 13:10

mariuszs marked this pull request as draft April 14, 2026 15:23

mariuszs changed the title ~~feat(mvn): add Maven (Java) filter module — test, build, dependency:tree~~ feat(mvn): add Maven (Java) filter module — test, compile, checkstyle:check, dependency:tree Apr 14, 2026

mariuszs force-pushed the feat/mvn-rust-module branch from b604ddb to 80da844 Compare April 14, 2026 19:16

mariuszs marked this pull request as ready for review April 14, 2026 19:26

mariuszs marked this pull request as draft April 15, 2026 20:43

mariuszs mentioned this pull request Apr 15, 2026

feat(mvn): enrich test output with Surefire/Failsafe XML reports mariuszs/rtk-java#3

Open

6 tasks

mariuszs force-pushed the feat/mvn-rust-module branch from 80da844 to 91a6534 Compare April 16, 2026 06:55

mariuszs changed the base branch from develop to master April 16, 2026 06:55

mariuszs marked this pull request as ready for review April 16, 2026 07:32

mariuszs force-pushed the feat/mvn-rust-module branch 3 times, most recently from 10fd699 to 98907f1 Compare April 16, 2026 12:37

mariuszs force-pushed the feat/mvn-rust-module branch from 98907f1 to 781ec7b Compare April 17, 2026 14:15

mariuszs force-pushed the feat/mvn-rust-module branch from 864932d to 3d993bf Compare April 17, 2026 18:55

mariuszs added 28 commits May 29, 2026 14:36

feat(mvn): add parse_goals for multi-goal detection

3d71ed1

feat(mvn): add chain_runs_tests enrichment gate

9b99c48

feat(mvn): add marker-based segment splitter

ef1128b

feat(mvn): add multi-goal segment filter (success path)

9400913

test(mvn): multi-goal compile-failure snapshot

952c169

feat(mvn): add run_multi_goal with -q strip and XML enrichment

4b4f533

feat(mvn): route all goals via dispatch, enable multi-goal filtering

49e7a48

refactor(mvn): drop orphaned run_compile wrapper

a062116

fix(mvn): classify_marker handles short plugin goal-prefix form

d197906

test(mvn): multi-goal verify-failure (multi-module) snapshot

f1e125b

fix(mvn): dedupe BUILD line in multi-goal output

f027f28

docs(mvn): document multi-goal filtering

1c3fd86

docs(mvn): fix stale run_other references after cutover

a0f5bde

mariuszs force-pushed the feat/mvn-rust-module branch from f79f814 to 6b203cb Compare May 29, 2026 12:41

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

feat(mvn): add Maven (Java) filter module — test, compile, checkstyle:check, dependency:tree#1089

feat(mvn): add Maven (Java) filter module — test, compile, checkstyle:check, dependency:tree#1089
mariuszs wants to merge 32 commits into
rtk-ai:developfrom
mariuszs:feat/mvn-rust-module

mariuszs commented Apr 8, 2026 •

edited

Loading

Uh oh!

CLAassistant commented Apr 8, 2026 •

edited

Loading

Uh oh!

Ckram commented Apr 17, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

mariuszs commented Apr 8, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Subcommands

Test plan

Uh oh!

CLAassistant commented Apr 8, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Ckram commented Apr 17, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

mariuszs commented Apr 8, 2026 •

edited

Loading

CLAassistant commented Apr 8, 2026 •

edited

Loading