Repository navigation
sandbox start fails after initial command exits: 'supervisor authentication bundle does not match runtime session' #4505
Description
Summary
Starting a sandbox whose initial command already finished (phase Completed) fails deterministically and moves the sandbox to Error. Creating a sandbox and running a command via sandbox create works fine; only restarting a stopped/completed sandbox is affected.
Environment
- OpenShell 0.1.3 (CLI, gateway, sandbox, driver-vm)
- Install method: Arch-based distro (CachyOS). The official
install.shsupports only snap/deb/rpm, so I used the AUR packageopenshell-bin0.1.3-1, which repackages the official release tarballs (sha256 verified against the release checksums file). - Gateway: local, native systemd user service (
systemctl --user), mTLS - Registration:
openshell gateway add https://127.0.0.1:17670 --local --name openshell→ Connected / Authenticated (mTLS) - Compute driver: Docker 29.9.0 (user is in the
dockergroup);openshell doctor checkpasses - Host: Linux x86_64, kernel 7.2.9, CachyOS
- Default sandbox image:
nvcr.io/nvidia/base/ubuntu:24.04
Steps to reproduce
openshell sandbox create --name repro1 -- echo done→ runs fine; sandbox reaches phaseCompletedopenshell sandbox start repro1
Actual
Error: code: 'The service is currently unavailable', message: "start sandbox failed: Docker supervisor exited before becoming ready; log tail: 2026-10-13T21:45:18.451Z INFO openshell_supervisor: Starting sandbox supervision
Error: supervisor authentication bundle does not match runtime session"
The sandbox transitions to phase Error. A subsequent openshell sandbox exec repro1 -- ... reports not ready (phase: Error). Reproduced with two independent sandboxes. The gateway was NOT restarted between creation and start.
Relevant sandbox-side log lines from the failed start (openshell logs repro1):
[sandbox] CONFIG:ENABLED TLS termination enabled: ephemeral CA generated
[sandbox] WARN host.openshell.internal maps to a non-link-local IP; trusted-gateway SSRF exemption disabled ip=127.0.0.1
[sandbox] NET:OPEN 127.0.0.1:17670
[gateway] supervisor session: superseded by reconnect, shutting down
[gateway] supervisor session: ended (already superseded)
[gateway] Stopped Docker sandbox after control supervisor failure
[gateway] Container exited
openshell sandbox get repro1 after the failure:
Phase: Error
Configuration: ConfigurationPending: Waiting for effective configuration validation before workload activation
Main process instance: e9924d02-442b-4c85-943b-efd169c12405
Annotations:
internal.openshell.ai/auth-epoch: 2
internal.openshell.ai/runtime-generation: 4ae09559-7423-4685-9bf4-0ae105d8ee5b
Conditions:
- Ready: False (ContainerExited) - Container exited
- ConfigurationReady: False (ConfigurationPending)
Expected
sandbox create → Completed → sandbox start → Ready.
Workaround
Delete the sandbox and create a new one (works immediately, image is cached).
Activity
mdusserebetclic commentedon Oct 20, 2026
Reproduced on 0.1.4, on a different platform:
install.sh(Homebrew, gateway asbrew services)linux/arm64, kernel7.0.12-linuxkit)nvcr.io/nvidia/base/ubuntu:24.04sandbox create -- sh -c 'exit 0'→Completed→sandbox startgives the samesupervisor authentication bundle does not match runtime sessionerror, and the sandbox can't be started again.stop/starton aReadysandbox works.We'd welcome a fix too.
vibexp-agent commentedon Oct 25, 2026
Same thing on 0.1.5 with the Docker driver, and it also happens after a failed main process, not only
Completed.Repro: connect to a sandbox, hit Ctrl-C so the main process exits with 130. The sandbox goes to
Error(MainProcessFailed) but the container keeps running ("retaining control-mode access plane").sandbox startthen fails with the same "authentication bundle does not match runtime session" error.I think the cause is in
start_sandbox_with_lifecycle_fencein the Docker driver. The gateway mints a new runtime identity (new session_id) for the start. The container is still running, so the driver takes the "gateway restart" branch and callsrefresh_docker_supervisor_authentication. That writes the new auth bundle but leavesruntime-descriptor.jsonwith the old session_id, andSelectedBackend::selectrejects the mismatch.Workaround without deleting: once the failed start has stopped the container,
sandbox stopthensandbox startworks and keeps the workspace. With the container exited, start goes throughrefresh_docker_boundary_authentication, which rotates the descriptor too.