[FEATURE]: Native infinite agent mode (auto-continue until done, with limits) #47019

Activity

1. github-actions

   assigned

   neriousy

   on Sep 3, 2026
2. github-actions commented on Sep 3, 2026

   

   github-actionsBot

   on Sep 3, 2026 – with GitHub Actions

   Contributor

   More actions

   This issue might be a duplicate of existing issues. Please check:

   • [FEATURE]: Auto-continue when model hits output token limit (finish_reason: "length") #17471: Auto-continue when model hits output token limit — overlaps on the "keep going" concept; that issue focuses specifically on finish_reason: length auto-continuation, while this request is broader (sentinel detection, todo tracking, per-session toggle, config limits). Worth coordinating to avoid divergent implementations.
3. AbhijitK20 commented on Sep 8, 2026

   

   AbhijitK20

   on Sep 8, 2026

   More actions

   I hit the same failure mode with GLM 5.3 in the normal build agent: the model brainstormed and hallucinated instead of executing, then returned a normal stop before the requested work was complete. I had to send continue three times; on the fourth prompt I explicitly told it to stop brainstorming, update the todos quickly, and execute the work.

   This suggests two useful acceptance criteria beyond an opt-in infinite mode:

   • The default build flow should make execution-first behavior explicit: do not restart planning or brainstorming when actionable todos remain; inspect the current todos and execute the next unfinished item.
   • A normal provider stop should not be treated as task completion when the session has unfinished todos, unless the user explicitly asked for a response-only turn.

   Automatic continuation still needs strict bounds and must stop for pending permissions/questions, provider errors, interruptions, or repeated no-progress turns. The current behavior makes users manually type continue even when the session state already shows unfinished work.
4. neriousy commented on Oct 21, 2026

   

   neriousy

   on Oct 21, 2026

   Member

   More actions

   Thanks, and for sharing your supervisor. This overlaps with #27167 (native session goals), which covers continuing automatically while a goal isn't finished, with limits and budgets. We'd rather track it in one place, so we're closing this as a duplicate. Please add your design points there (sentinel and todo-based completion, max iterations/hours, never continuing while a permission or question is pending), along with @AbhijitK20's point about unfinished todos. Until then, your supervisor can run as a plugin, since V2 plugins can subscribe to session events and send follow-up prompts: https://opencode.ai/v2/docs/build/plugins.
5. neriousy

   closed this as a duplicateof [FEATURE]: Add native session goals with /goal #27167 on Oct 21, 2026