104006: Session continuity not guaranteed

After an AI Transport session's channel first attached, it entered a failed, suspended, or detached state, or re-attached without resuming, so the SDK can no longer guarantee that active streams received every event.

What you should do

Once the channel re-attaches, new messages arrive normally. The only risk is that the session may not have received some events published to the channel during the disconnection.

What to do depends on which session you are running. An agent session has already aborted every in-flight run, so re-run them (a durable-execution framework does this for you). A client session surfaces the error on its error event, and any per-run stream you are consuming ends with it; reload the conversation from history, the durable record on the channel, to recover anything missed.

Why it happens

The SDK relies on an unbroken channel attachment to deliver every event in order. This error is raised when, after the initial attach, the channel enters FAILED, SUSPENDED, or DETACHED, or re-attaches with resumed: false, meaning it could not resume from where it left off, typically because the disconnection outlasted the roughly two-minute resume window. Either way some events may have been missed, so continuity can no longer be guaranteed; it does not mean events were definitely dropped. The state change that triggered it is preserved as the cause; for a failed resume that is the channel's own continuity error.

What you'll see

The error is reported with code 104006 and status 500. The message names the operation and the new state, for example unable to deliver events; channel continuity lost (suspended) or unable to continue; channel continuity lost (attached, resumed: false).