stale synthetic continuation rows can overwrite the session model #33044

Problem

When an interrupted or failed turn leaves behind an internal/synthetic continuation user row, later session resume paths can treat that synthetic row's model as the durable session model. In practice this can make a normal user follow-up continue with a fallback/retry model even though the user did not choose that model and the fallback should have been scoped to the failed retry path.

Related edge cases:

• provider streams that finish without content can leave ambiguous state instead of a retryable failure
• retry-advisory errors such as Retry Error: An error occurred while processing your request. You can retry your request... may not be classified as retryable if the SDK marks them non-retryable or omits status metadata
• Web session model sync can restore a stale/synthetic-only model into local state
• a stale stored output model can be reused when the UI sends no explicit model on the next turn

Expected

Internal/synthetic continuation rows should not become the durable user-selected model for a normal follow-up turn. If a turn fails before content is produced, recovery should preserve the previous real user-selected model and retry only on verified retryable failures.

Actual

A synthetic/internal continuation row can become the most recent model source, causing later normal user input to reuse a fallback/retry model without a fresh reason.

Notes

I have a small patch that:

• hides synthetic-only internal continuation users from Web timeline model sync
• adds authoritative session-model sync for the active session page
• treats empty provider streams and retry-advisory messages as retryable failures
• closes blank/interrupted assistant rows without emitting a new error event that can re-trigger fallback
• recovers stale explicit continuation models back to the previous real user model

Prepared with AI assistance.