Colton QiandClaude
9ed616c135
Publish administrator termination to session readers and reject new input before rebuilding compute. Preserve normal recovery after sandbox reclamation, explicit manual recovery, and completed history. Cover both lifecycle paths with live tests and give deployments a stable hook for styling the recovery control. Co-Authored-By: Claude <noreply@anthropic.com>
Python live E2E execution contract
These tests drive a real deployment: a live backend, a real sandbox, a real
model. They carry the e2e mark and are deselected by default, so a plain
pytest run never starts one.
scripts/e2e_live.sh is the entry point. Point it at a deployment that is
already running and let it select the lane:
ASTRABOX_E2E_BASE_URL=http://127.0.0.1:8088 ./scripts/e2e_live.sh
It stops at the first failed report and cleans up nothing, so the failing scene — the sandbox, its logs, the deployment state — is there when you look. A selected pass is not lane evidence; the lane is what the two phases below finish.
The runner finishes two phases in order:
The canonical PostgreSQL + Agent lane excludes the deployment-specific marker
assistant_live, then runs:
e2e and not backend_restartacross xdist workers withloadgroup.e2e and backend_restartonly after every parallel worker has exited, without xdist.
An Assistant/Hermes contract carries assistant_live and runs against its
actual deployment. It must not skip inside the Agent lane or be counted as one
of that lane's passes.
The frozen node inventory, engine matrix, and storage partitions live in
suite-contract.json.
main runs ordinary and restart tests for each configured Agent program;
assistant is a separate serial lane. E2E_LANE selects one of them for a
direct diagnostic:
E2E_LANE=main ./scripts/e2e_live.sh
E2E_LANE=assistant \
ASTRABOX_E2E_ASSISTANT_ENVIRONMENT=astrabox-e2e-assistant \
ASTRABOX_E2E_ASSISTANT_IMAGE=<exact-hermes-image> \
ASTRABOX_E2E_ASSISTANT_MODEL=<configured-openai-chat-model> \
./scripts/e2e_live.sh
For a direct diagnostic after a fix, select exact nodes instead of rerunning the complete lane:
./scripts/e2e_live.sh \
--parallel-node tests/e2e/test_delete_and_archive.py::test_delete_makes_session_unreadable
Use --restart-node for a backend-restart test. Repeated node arguments retain
the ordinary/restart separation. Preserve failed reports and sandbox state
before another invocation; a targeted pass is not complete-lane evidence.
Any test that stops or restarts the shared backend, bundled sandbox server, or
database must carry both backend_restart and
xdist_group("backend-restart"). The first marker moves it into the exclusive
runner phase. The second keeps restart tests together in targeted xdist runs;
the top-level conftest rejects any other scheduler whenever a live E2E
invocation uses workers, so callers cannot silently select one that ignores the
group. The canonical runner supplies --dist loadgroup itself.
loadgroup alone is not an exclusive lock: other workers may still run
unmarked lifecycle operations. Therefore a bare full-suite
pytest -m e2e -n ... invocation does not satisfy this contract even though
the group itself is serialized. Use the runner for the full lane, or select
only the backend_restart phase and run it without -n.
Markers are strict repository-wide. A misspelled or unregistered isolation marker is a collection error rather than a silently parallel test.