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
{{ message }}
Repository navigation
<system-reminder> wrapping fires on every tool-call iteration in follow-up messages, derailing model behavior #29659
In prompt.ts (~line 1415), user message text parts in follow-up messages get wrapped in tags with "Please address this message and continue with your tasks" whenever step > 1 && lastFinished. This fires on every tool-call round.
Models interpret "Please address this message and continue with your tasks" as a system-level instruction to keep going, and start outputting "The system is asking me to continue" and wastes tokens. The in-place mutation of p.text also causes the wrapping to nest on subsequent iterations, compounding the issue.
ghost changed the title [-]<system-reminder> tags nest on every loop iteration, causing model to loop "the system is asking me to continue[/-][+]<system-reminder> wrapping fires on every tool-call iteration in follow-up messages, derailing model behavior[/+]on May 28, 2026
Activity
github-actions commentedon May 28, 2026
This issue might be a duplicate of existing issues. Please check:
step > 1 && lastFinishedinprompt.tstriggering the<system-reminder>wrap and causing spurious extra LLM round-trips. That issue focuses on clock skew as the trigger, while this issue focuses on normal multi-tool-call sessions, but the wrapping/nesting mechanism and the affected code path are the same.<system-reminder>leaking into user messages and being treated as user-authored input — related symptom where the tag content influences model behavior unintentionally.[-]<system-reminder> tags nest on every loop iteration, causing model to loop "the system is asking me to continue[/-][+]<system-reminder> wrapping fires on every tool-call iteration in follow-up messages, derailing model behavior[/+]