[ryanduguid]

Summary

Fixes a Windows failure in the benchmark's retry: a retried cell raised Destination path ... already exists, skipped the retry and stranded its session in the scratch directory, which the run deletes at the end. Only benchmarks/agentic/run.py changes. No rule text, scoring or retry policy changes, and no paid session has run.

Cause: _one removed the kept workspace of the failed attempt with shutil.rmtree(..., ignore_errors=True). That workspace holds a git repository from _git_snapshot, whose objects are read-only, and Windows will not delete a read-only file. The delete failed silently, so shutil.move nested the new workspace inside the old directory.

_keep_cell now removes the old directory, clearing the read-only flag when needed, raises if anything remains, then moves. Seen in the first live pilot attempt on 31 October, where 2 cells failed this way.

Evidence

• Before: with the old delete-and-move behaviour, the new selftest case keep cell prints XX on Windows (Python 3.15).
  After: it prints ok.
• run.py --selftest passes at this head, with the new case.
• node --test tests/*.test.js gives 154 passed, 2 failed, 4 skipped, the same as on main (the 2 are correctness.test.js: email bare unfenced code, csv timeout; cause not established).

Merge danger

Door: two-way. Blast radius: benchmarks

Only the harness's cell storage changes.

Unverified

• The new case passes on Linux with or without the fix, because a read-only file does not block deletion there. It runs in CI, so it guards the contract but only fails on Windows without the fix.
• No live cell has run with the fix: the account is at its usage limit until 10pm.

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues
0 New dependency risks

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud