Repository navigation
@agent-native/core@0.153.6
Patch Changes
-
9f0ae16: Stop the chat continuation guards from ending turns that were still working: an explicit
recoverable: falseon an error event now outranks the transient message sniff (a repeat-guard stop naming a*connection*tool auto-continued the loop it was meant to break),loop_limitwork boundaries are bounded by their own ceiling instead of the transient-failure one, and the stall signal is read from each round's delta so an unresolved "Preparing" card cannot make later rounds look stuck. History trimming also prices object tool results by what the request actually sends instead of[object Object]. -
9f0ae16: Fix stale agent runs never being reaped in production. The periodic stale reaper
lived on an in-process timer thatshouldDisableInProcessSweeps()turns off for
every production serverless function, and nothing durable replaced it, so a run
whose producer died stayed "running" until an unrelated request path happened to
notice. Stale reaping now rides the signed, platform-scheduled recurring-job
sweep — one site-wide reap per tick instead of one per warm container.Also corrects what the reapers record:
completed_atis now the run's last
liveness basis rather than the time a reaper noticed, so a reaped run reports its
real duration instead of detection latency; andterminal_reasonis the
normalizederror:stale_runthat the event reconciler already writes for the
same outcome, so one failure no longer splits across two permanent
agent_run_outcome_dailybuckets. Heartbeat writes are bounded to one attempt
inside a third of the stale window, so a live run can no longer be reaped while
its own heartbeat is still in flight holding a pooler connection. -
9f0ae16: Use a clear loading message for the route transition indicator instead of exposing the destination URL.
-
9f0ae16: Keep agent runs alive when the SQL abort state is briefly unreadable. A few consecutive failed abort-state reads (roughly 9s of database unreadability) used to self-abort the run with
aborted_abort_check_unavailable, killing in-flight work the user was waiting on. The check now fails open and reports the outage to Sentry instead; the run stays bounded by the soft timeout, no-progress backstop, and iteration limits. -
9f0ae16: Stop the new repeat-loop guards from killing turns that would have succeeded: network "connect … before/first" failures and retention-window "only available in …" errors no longer read as permanent preconditions, per-item errors that differ only by id (
Record 41 not found) are counted separately again, the permanent-precondition stop keeps the raw tool error indetailslike every other terminal stop, an explicitrecoverable: falseoutranks the message sniff, a single ledger-read blip is retried before the turn stops, and the persisted-thread rebuild scopes retry clears exactly like the live client so narration no longer vanishes on reload.