You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Classify prompt-too-long errors and run bounded emergency recovery that removes old tool bodies or selects source-bearing compact input without mutating the transcript.
Acceptance criteria
Recovery retries are capped.
Recent unresolved tool results remain protected.
Emergency changes do not modify RuntimeEvent history.
Invalid recovery summaries fail explicitly.
Normal strategy and emergency metrics are separate.
Required tests
Provider error variants
retry cap
recent tool protection
invalid summary
impossible recovery
Benchmark
Fault-injected long tool loops; recovery >=95%, normal benchmark overflow zero.
Superseded by #1615 (automatic long-term memory lifecycle, SQLite-based; accepted by maintainers 2026-07-30, PR ① #1770 merged). This MM-26 work item never landed outside draft PR #820 and is replaced by the #1615 design and its LongMemEval/MSC evaluation plan.
Activity
Astro-Han commentedon Aug 2, 2026
Superseded by #1615 (automatic long-term memory lifecycle, SQLite-based; accepted by maintainers 2026-07-30, PR ① #1770 merged). This MM-26 work item never landed outside draft PR #820 and is replaced by the #1615 design and its LongMemEval/MSC evaluation plan.