[JacobStephens2]

Summary

A completed unit-test process left its Factory skills extraction in TMPDIR because static values are not dropped at process exit.

unit-test users acquire a shared cleanup guard
  → share the process's Factory skills extraction
  → last guard removes it
Command → existing finish cleanup

Closes #1913

Test seams: the existing Session subprocess entry point with a private TMPDIR; public Factory skills linking across overlapping test users; an ordinary Run through Scenario with its private TMPDIR.

Evidence

Execution evidence

• Before: cargo test --bin thirdshift session::execution_tests::session_subprocess_removes_its_factory_skills -- --exact --nocapture failed with Factory skills left behind: ["thirdshift-skills-acWAv6"] (0 passed; 1 failed).
  After: the same command passed (1 passed; 0 failed).
• The original reproduction ran the existing review-report Session fixture under a private TMPDIR: it passed but left 7 extractions before the fix (one per current Harness), and left 0 after it.
• cargo test --bin thirdshift skills::tests -- --nocapture: 6 passed, including overlapping users retaining one extraction until the last finishes. cargo fmt --check passed.
• Unit binary in one process, with a private TMPDIR, --test-threads=3: 1,684 passed, 3 ignored; 0 Factory skills entries left (666.79 seconds, execution only).
• All 15 tests named in Unit tests leave 24 Factory skills directories in the temp directory on every run #1913, each in its own process with a new private TMPDIR: passed; 0 Factory skills entries left by each.
• Validation uses a private HOME and the installed CARGO_HOME/RUSTUP_HOME. An initial run inherited the operator's admission limits; its toolchain fixture also needed RUSTUP_HOME set. That fixture passes with the installed root, and the full unit rerun above passes.
• Full Rust check: CI cargo nextest run passed on exact head ded15db6eb0d1992c8bd5262624963a23115a545: 3,960 passed, 92 skipped. Test execution took 1,297.727 seconds, plus 73 seconds of compilation (about 1,372 seconds for the command including compilation).
• scripts/affected-tests main selected that same full Rust suite locally. After CI passed the exact head, the duplicate local run was intentionally stopped, following docs/testing.md's CI-reuse guidance. Its /usr/bin/time -p wall time was 1,681.56 seconds including cached compilation; this interrupted run is not reported as a local pass.

Integration TMPDIR inventory

Each built test binary ran with --test-threads=3 and a fresh private TMPDIR (python3 /tmp/issue-1913-validation.py integrations). All 23 exited successfully.

Binary	Factory skills entries left
security_review	0
security_run	0
session_traces	0
setup	0
setup_lifetime	0
setup_offer	0
share_card	0
show_logs	0
site_deploy	0
skill	0
spec_1796	0
spec_push_race	0
spec_review_706_regressions	0
spec_run	0
standards_1795	0
status	0
stop	0
toolchain	0
triage_run	0
trial	0
trial_scope	0
update	0
user_config	0

Unaddressed findings

Standards

None. The Standards reviewer raised 0 findings.

Spec

None. The Spec reviewer raised 0 findings.

Security

None. The Security review found no unaddressed introduced findings.

Security review outcome

No unaddressed introduced findings.

Review coverage

Both independent reviewers read all six changed files: src/skills.rs, src/skills/testing.rs, src/skills/tests.rs, src/session/execution_tests.rs, src/session/private/tests.rs, and tests/fresh_run.rs. Unread changed files: none. Generated-output exclusions: none.

Merge Danger

Door: two-way

Blast Radius: tests

Built with codex codex-cli 0.164.0 · gpt-6.1-sol · xhigh