This case study records a maintainer-led local pilot completed on 2026-09-07 for gradle/gradle#38843. The issue already contained a valid public reproduction, but recreating it required two complete repositories and a manual settings change. The goal was to test whether ReproMin could turn that workflow into smaller, replayable evidence. It was not a root-cause investigation, an upstream submission, or evidence of non-maintainer adoption.
The prepared source combined these Apache-2.0 snapshots:
autonomousapps/dedebugcommitd26e30980489a316fb824c52c217ef7aaf5d942bfrom branchtrobalik.gradle-issue-38843; andautonomousapps/dependency-analysis-gradle-plugincommit93b52302ef0f48ad7aa0776ae2a1e479a9055f32.
The only upstream-source edit enabled the composite build in
dedebug/settings.gradle.kts:
includeBuild("../dependency-analysis-gradle-plugin")The pilot root added an oracle script, a provenance record, and generated-file ignore rules. Both upstream licenses were protected and retained. Gradle's contribution policy requires a human to understand an AI-assisted submission and sign its DCO. No comment or pull request was sent upstream; the project owner must review the artifact and disclosure before any delivery.
The reproduce.sh oracle runs dedebug/gradlew :help, unsets CI, and emits
REPOMIN_GRADLE_38843_TARGET with exit code 23 only when Gradle reports all of
these signals:
- failure to resolve
com.gradle.develocityversion4.2.2; - the exact
:dependency-analysis-gradle-plugin -> :dependency-analysis-gradle-pluginplugin-build cycle; and - the statement that this cycle is unsupported with Isolated Projects.
The script also requires active, non-commented settings for the outer
includeBuild, org.gradle.isolated-projects=true, the included build name,
and its includeBuild(".") self-include. Negative controls confirmed that a
disabled outer include, a disabled self-include, and an unrelated plugin
resolution failure cannot emit the target marker.
The default command permits dependency downloads so a new user can start from
an empty Gradle cache. REPOMIN_GRADLE_OFFLINE=1 is available for repeated
runs after warming the same checkout, but the pilot does not claim that a
single warm-up makes newly copied paths work offline.
The source contained 854 reduction-visible files and 3,817,947 bytes. The run
used the v0.1.0.dev10 GitHub Release wheel, macOS ARM64, OpenJDK 17.0.20, the
Gradle 9.7.0 wrapper, and the host backend. The oracle, provenance record, both
licenses, and the outer Gradle wrapper were protected. Source reduction was
disabled.
A global 80-attempt budget was consumed by the Gradle structured reducer before file reduction could start. Because the budget is part of session identity and cannot be raised on resume, the accepted tree became the input to a separate file-only run. A third file-only run reached a fixed point, and a short final run re-certified the portable oracle change.
| Stage | Files before | Files after | Bytes before | Bytes after | Attempts | Accepted |
|---|---|---|---|---|---|---|
| Gradle model | 854 | 854 | 3,817,947 | 3,806,016 | 80 | 9 |
| File reduction | 854 | 20 | 3,806,016 | 121,130 | 80 | 13 |
| File fixed point | 20 | 11 | 121,130 | 87,521 | 66 | 6 |
| Portable re-certification | 11 | 11 | 87,862 | 87,862 | 19 | 0 |
The final presentation is 341 bytes larger than the preceding fixed-point payload because the protected oracle and provenance text were updated to support and explain first-run online resolution. Across the prepared source and final presentation, 843 files and 3,730,085 bytes were removed, leaving about 2.3% of the original bytes.
The final 11 files are the oracle and provenance record, two license files, the outer wrapper script/JAR/properties, and the two builds' settings and properties files. ReproMin could not remove another eligible file while preserving the configured failure and protected-file boundary.
report validate confirmed the exact final payload fingerprint. With a cache
populated by the cold-start run, report replay reproduced the recorded
match-and-exit-code oracle in 5 of 5 online-enabled fresh copies; each took
between 3.86 and 4.17 seconds. A separate first run with an empty Gradle user
home downloaded its requirements and reproduced the target in 96 seconds.
The trial also exposed a cache boundary worth preserving in the record. Twice,
a Doctor run started immediately after a reduction or export reported zero
passing fresh copies; subsequent direct or replay runs succeeded, followed by
Doctor results of 3/3 and 5/5. A cache warmed in one checkout also did not make
freshly copied paths reliable with --offline. Final evidence therefore uses
online-enabled fresh-copy replay, and the artifact does not claim a
network-free cold start.
- A strict marker plus exact exit code rejected ordinary Gradle failures; the active-line preconditions prevented commented settings from passing setup.
- Build-tool caches are part of the reproducibility boundary. A successful warm run is not evidence that a copied checkout will work offline.
- A global budget can be consumed by an early structured reducer. Since a completed bounded session cannot accept a larger resume budget, practical pilots may need staged runs and separate reports.
- Aggregate Doctor output reports pass counts but not per-sample mismatch reasons, which made transient baseline failures harder to diagnose.
--gitignore-recursiverequired a root.gitignorefor this directory that combined two repositories, even though both nested repositories had their own files.- The release wheel installed correctly from GitHub Releases, while neither the configured package mirror nor official PyPI exposed this project. That extra installation URL is adoption friction.
These are observations from one maintainer-run workflow, not sufficient reason on their own to add new budget, cache, or diagnostic features. The roadmap's two-independent-trial threshold still applies.
OpenAI Codex performed the local reproduction, oracle construction, reduction, and documentation under delegated repository maintenance. Any future upstream message must disclose that material involvement and be reviewed and sent by a human who can make the required authorship and DCO attestations.