Document dual-step best-effort run-store cleanup behavior

Co-authored-by: Eric Allam <eric@trigger.dev>

Author: cursoragent

docs/tasks/streams.mdx

@@ -623,7 +623,8 @@ optional `runId`, and the underlying `error` (non-Error throws are normalized to
 instances).
 
 Run-store cleanup is handled as best effort, and cleanup failures won't mask the original
-transport failure that triggered `onError`.
+transport failure that triggered `onError`. Cleanup still attempts both persistence steps
+(`set` inactive state and `delete`) even when one step fails.
 
 ```ts
 import type { TriggerChatRunState, TriggerChatRunStore } from "@trigger.dev/ai";

packages/ai/CHANGELOG.md

@@ -16,3 +16,4 @@
 - Added phase-aware `onError` reporting across send, stream-subscribe, reconnect, and stream-consumption paths.
 - Added normalization of non-Error throw values into Error instances before `onError` reporting.
 - Added best-effort run-store cleanup so cleanup failures do not mask root transport errors.
+- Improved best-effort run-store cleanup to attempt both inactive-state writes and deletes even if one step fails.

packages/ai/README.md

@@ -141,7 +141,8 @@ The callback receives:
 - `error`
 
 Cleanup operations against custom `runStore` implementations are best-effort. If store cleanup
-fails, the original transport error is still preserved and surfaced.
+fails, the original transport error is still preserved and surfaced. The transport also attempts
+both cleanup steps (`set` inactive state and `delete`) even if one of them fails.
 
 ## Reconnect semantics