LLM Description written by Claude Code:claude-opus-5-5
Refs #823 (slice 18: the e2e_maven suite). This uses Refs, so #823 stays open for the remaining PENDING_RAW_SPAWNS files.
Summary
e2e_maven.rs now starts socket-patch through tests/common/hermetic.rs (hermetic::binary_command()), scrubs any activated venv (Extra::Venv) and isolates the JVM env with jvm_env::isolate_cli, the same way e2e_cargo.rs does. After that call it pins its own fixture-local homes: MAVEN_REPO_LOCAL, HOME/USERPROFILE, XDG_CACHE_HOME, LOCALAPPDATA, and now GRADLE_USER_HOME. It also removes M2_HOME. The file comes off PENDING_RAW_SPAWNS (9 → 8 on main). Once #1515 and #1518 also land, 3 files remain: e2e_hosted_production, e2e_redirect_cargo_shapes and e2e_vendored_production. The change touches tests only: no product code and no workflows.
This PR removes the e2e_maven.rs entry, which is the only remaining entry whose neighbours #1515 and #1518 don't remove. The three PRs merge cleanly in any order.
Root cause
run() spawned the CLI with a bare Command::new and removed only 6 SOCKET_* names by hand. Every other ambient SOCKET_* reached the child:
- On main, a shell that exports
SOCKET_DRY_RUN=true SOCKET_ECOSYSTEMS=cargo SOCKET_ORG_SLUG=hostile SOCKET_NO_CONFIG=0 SOCKET_JSON=false SOCKET_SILENT=true SOCKET_NO_UPDATE_CHECK=0 fails 2 of the 22 tests: scan_discovers_maven_artifacts and scan_discovers_gradle_project_artifacts.
SOCKET_NO_CONFIG=0 let a developer's real socket login reach the child. The update check and telemetry were never forced off.GRADLE_USER_HOME, GRADLE_OPTS, GRADLE_RO_DEP_CACHE and VIRTUAL_ENV were inherited by the unscoped scans.
Test evidence
- Red with only the
PENDING_RAW_SPAWNS removal: spawn_env_hygiene fails no_new_bare_binary_spawns (15 passed, 1 failed).
- Red on main under the hostile shell above (plus
VIRTUAL_ENV=/nonexistent): e2e_maven passes 20 and fails 2.
- Green at be57c3d under the same hostile shell (plus
GRADLE_USER_HOME=/nonexistent M2_HOME=/nonexistent): e2e_maven 22/22 and spawn_env_hygiene 16/16. With a clean env, e2e_maven is 22/22. The suite has no #[ignore]d tests.
cargo fmt --all -- --check and cargo clippy --locked --workspace --all-features -- -D warnings are clean, and clippy --test e2e_maven -- -D warnings is clean too.- CI at be57c3d: all 86 check runs completed. 35 passed (including
ci-ok and clippy), 50 were skipped (the full tier), 1 was neutral (Bugbot) and none failed. Bugbot found no issues.
/code-review high
No correctness findings in this diff. Fixed in be57c3d:
- The shared stand-in
GRADLE_USER_HOME (from isolate_cli) outlives test runs. run() now pins it next to the fixture's HOME, so a stray cache can't change the Gradle scan's count.
- Added a comment saying the home pins must stay after
isolate_cli, because the last write for a name wins.
run() now calls hermetic::binary_command() instead of hermetic::command(&binary()).
Not fixed here. These change the shared common harness, so they belong in a follow-up slice under #823:
jvm_env::AMBIENT doesn't scrub M2_HOME / MAVEN_REPO_LOCAL, and isolate_cli doesn't pin XDG_CACHE_HOME / LOCALAPPDATA. Other suites that only use isolate_cli (for example e2e_cargo and common::run) can still see a developer's real ~/.m2 or Coursier cache. e2e_maven pins all of these itself.- The proxied async
run() helper is now repeated almost verbatim across e2e_maven, e2e_cargo, e2e_composer and others. A shared common::run_proxied would remove the copies.
CI cost
No workflow changes. A lean PR run costs the same before and after: about 28 Linux jobs and about 8 minutes.
🤖 Generated with Claude Code
https://claude.ai/code/session_015ra9QiqqVWRsp6FEqw8BXb