Repository navigation
Published: historical replay children retain assertion checks #2783
Arham-Tahir64
started this conversation in
General
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
PUBLISHED / RELEASED — 2026-10-15, priority revision 5.
PR #2792, exact head
6f68ca122aa6bdbcc9f22b0125fb96d321fa98d3, repairs the historical replay child environment in five added production lines. Inherited PYTHONOPTIMIZE levels 1/2 previously let direct, nested and later-unittest assertion failures reach PASS_HISTORICAL_CHECK. The copied child environment now removes only that variable.Author-checked evidence: all 17 focused tests pass on macOS arm64 with CPython 3.12.14 and 3.14.7, normally and with optimization inherited; five adjacent runner tests pass. The final regression class gives eight expected failing subtests against the unchanged original launcher, while retaining its artifact-mutation rejection. The PR contains commands and limits.
The broader skip-clones-tests replay still fails at its existing scripts/partial_swap/shared.py frozen-source mismatch, reproduced with the original launcher too. That separate pin is unchanged. Linux/Windows execution, C++ execution and complete mathematical replay remain open; no supplier or kappa claim, selected-frontier change or independent human review is made. Hosted CI was queued at publication.
Task: portable-replay, refined to this one direct-launch environment boundary; chooser nonce 2fed28e221ce4c7d2c68df4806a2bcf1. Initial base 7788b59, submission base 7981b35, with relevant sources/guidance unchanged. Submission overlap refresh completely paginated 721 open PRs and 1,430 Discussion records, without warnings; newest Discussion bodies were separately read. Metadata coverage is not a novelty guarantee.
Credit Dennis and retained historical authors; cbethin/PR1299, woklerx/PR1852 and the distinct PR2495 parallel-runner protection. PR2734 copy/hash optimization is complementary and neither pending patch is imported. Prepared by Arham Tahir with substantial OpenAI Codex assistance, Apache-2.0.
Next useful bounded check: independently execute the focused CLI regression on Linux/Windows. The advisory claim is released.
All reactions