Known issues

For: cluster operators and application developers.

These issues affect the implementation described by the current documentation. They remain open; a workaround does not mean the underlying issue is fixed. Each entry names the configuration or operation affected so you can decide whether it matters to your deployment.

For platform and integration availability, use Compatibility. Deployment boundaries and security reporting are covered in Security assurance.

Jobs and accounting

Budget counters on separate edges

With management and edge running separately, repeated usage reports can temporarily inflate a budget counter and refuse calls early. Usage rows and quota charges are not duplicated by this issue.

Check recorded usage. The counter corrects after that budget has no open calls. See repeated reports on separate edges.

Quota charges after a crash

A crash can lose an owner or team quota charge when no recovery record survives. This affects an unreserved call interrupted between its usage and quota writes, or a failed charge kept only in memory.

Review usage and quota totals after an unclean stop. Durable reservations and saved retries repair other missing charges. See quota accounting after a crash.

Delayed Temporal tasks

On a slow database under load, a Temporal task can be recorded as started without reaching a node. The job waits for the task timeout before retrying.

Check database latency and allow the retry to complete. See capacity guidance.

Grant revocation during a segment

When a grant is revoked during a running segment, a node's stale grant view can make the edge fail a job, rather than refuse the call on the node. An OpenAI Agents SDK segment still lists the revoked delegation.

Adjust the agent's work or access before restarting. The revoked call remains refused. See revocation during a running segment.

Sandbox exemptions across nodes

If a plugin is exempted from sandboxing on only some eligible nodes, routing can select a node that refuses its tool at startup.

Apply the reviewed exemption consistently, or restrict routing with a label. See builds and mixed fleets.

Clients and deployment

Remote nodes with Compose

When you add a remote node to the Compose deployment, the console's command can name an unreachable container address.

Replace both occurrences with the host's published edge address. Follow the remote-node procedure.

Resuming AG-UI approvals

When you resume several AG-UI approvals, a job update can fail after earlier decisions were applied.

Retry with the same run ID or Idempotency-Key; already-applied decisions are retained. See resume rules.

Reconnecting to a completed job

If a long job ended without an active poller, reconstructing its final stream can fail and request a resync.

Fetch current job state and its trace. See feed limits.

Isolation and stored data

macOS processes after a session

When a macOS sandbox session ends, a process that left the original process group can remain running under the old profile.

Use Linux when all child processes must end with the sandbox. Check for survivors before reusing its working directory. See processes after a session.

Older invalid job IDs

For jobs created with older invalid IDs, some name-only lookups return no result. Erasure and export can report jobs_partly_found.

Inspect those reports before relying on completeness. Destroyed keys stay destroyed. See older job IDs.

Abandoned temporary folders

A runtime killed before temporary-folder tracking was available can leave untracked folders containing job data.

Remove abandoned folders only while no session runs on the node. See locations and cleanup.