OpenShell #4062 reproduction (Go SDK)

Minimal reproduction for NVIDIA/OpenShell#4062: a committed per-sandbox policy revision does not always advance the sandbox resource version, so a stale ExpectedResourceVersion is still accepted and silently overwrites a newer policy.

Uses only the public Go SDK (github.com/NVIDIA/OpenShell/sdk/go, pinned in go.mod) against a running Gateway. No database or internal API access.

What it does

1. Creates one disposable sandbox with an inline policy (no providers, no annotations), waits for Ready, stops it, and waits for Stopped plus a stable resource version. Stopping avoids supervisor status writes that could advance the version and mask the bug.
2. Reads resource version R and policy P.
3. Builds policy A (P + rule writer-a) and policy B (P + rule writer-b).
4. Config().Update with A, ExpectedResourceVersion = R; rereads the sandbox version.
5. Config().Update with B, the same stale R.
6. Reads the final policy and revision list. On the fixed path it also rereads, merges B into A, and retries.
7. Deletes the sandbox.

Sequential calls are sufficient: they model two clients that read the same snapshot.

Run

go run . --address localhost:8080 --tls-dir /path/to/certs --image <sandbox-image>

--tls-dir holds ca.crt, tls.crt, tls.key (omit for plaintext). --image is optional. Exit code: 0 expected behavior, 2 bug reproduced, 1 anything else.

Observed

	Bug (unpatched)	Expected (fixed)
Sandbox resource version after A	unchanged (R)	greater than R
B with stale R	accepted	rejected, gRPC Aborted
Latest policy	writer-b only; A lost	writer-a only until retry
Revisions inserted by B	yes	no

Unpatched Gateway

Gateway image ghcr.io/nvidia/openshell/gateway:374c035962dada63852ec1dea073fea618ed76c5 (upstream main at that commit), SQLite backend (--db-url sqlite:...), SDK v0.0.0-20260930202246-912a077bd641:

snapshot:  resource_version=27 policy_version=2
writer A:  accepted, policy_version=3, sandbox resource_version=27 (was 27)
writer B:  ACCEPTED with stale resource_version=27, policy_version=4
final:     policy_version=4 revisions=4 writer-a=false writer-b=true
RESULT: BUG REPRODUCED - stale write accepted; writer A's policy change was lost

Gateway with the proposed fix (local build)

snapshot:  resource_version=28 policy_version=2
writer A:  accepted, policy_version=3, sandbox resource_version=29 (was 28)
writer B:  rejected: Conflict: persist policy revision failed due to concurrent modification (current resource_version: 29)
           grpc_code=Aborted
final:     policy_version=3 revisions=3 writer-a=true writer-b=false
retry:     accepted after reread; writer-a=true writer-b=true
RESULT: EXPECTED BEHAVIOR - stale write rejected with Aborted; fresh merge accepted

Notes

• Annotations are omitted on purpose: changing them advances the version independently.
• A and B differ in content on purpose: identical content can take an idempotent path with no new revision.
• The check uses the exact gRPC code (Aborted) rather than v1.IsConflict, which also matches FailedPrecondition.