Agent Sandbox Registry

python-sdk

Verification: real-cloud E2E suite foundation (python-sdk)

无中文版

.agents/evidence/verify/2026-10-12-e2e-foundation.md

Verification: real-cloud E2E suite foundation (python-sdk)

What was verified

The new e2e/ pytest suite itself: that it authenticates against the real platform, creates, uses and destroys real sandboxes (envd data plane and the in-sandbox agent-sandbox-server from the python-hello template), tags and cleans up everything it creates, skips with an exact reason without credentials, and pins known product defects as strict xfails.

Commands run

make e2e                                   # fast path
make e2e-all                               # slow tests + Docker/ACR flow
python scripts/e2e-cleanup.py --all-tagged # dry run: lists leftovers (none)

Results

Run Selection Result Duration
1 not e2e_slow (fast path) 67 passed, 13 xfailed (known issues KI-001 to KI-010), 0 failed 8 min 8 s
2 e2e_slow and e2e_docker 1 passed, 4 skipped (Docker/ACR flow), 1 xfailed (KI-008) 4 min 41 s

Skip reason of the Docker module, verbatim: e2e_docker skipped: ACR push not usable from this host: a 2000 KB layer did not finish uploading within 120s (upload throughput <17 KB/s to the ACR public endpoint); real template images are far larger (KI-011).

The 13 xfails are strict: when a defect is fixed the test XPASSes and fails the run, forcing the marker to be removed. They are catalogued in e2e/KNOWN-ISSUES.md.

Cleanup

The session finalizer deleted every tracked sandbox. python scripts/e2e-cleanup.py --all-tagged after both runs reported no sandbox carrying an e2e tag; no e2e- template remained. Nothing leaked.

Problems found

Not verified