Repository navigation
provider/streaming: reasoning degenerates into repetitive loop on long session #51503
Closed as not planned
Description
Summary
In a long desktop session (heavy MCP tool usage while testing wincontrol desktop-control), the assistant's reasoning stream degenerated into a repetitive near-empty loop — dozens of consecutive lines of "好。", occasionally "我输出。", "(写)", "(执行)", "嗯。" — until the turn was abandoned. Possibly model-side behavior, but filing for triage since it coincided with compaction / MCP catalog changes in the session.
Environment
- opencode version: 2.0.18
- OS: Windows_NT 10.0.26200 (win32 x64), OpenCode Desktop app
- Model: qwen3.8-flash (provider: alibaba-token-plan-cn, variant: xhigh)
- Active plugins: @prevalentware/opencode-goal-plugin, local handoff-on-compact plugin (fires on context compaction)
Reproduction
- Run a long session with many tool calls against an MCP server (wincontrol), including an MCP server restart / tool-catalog change mid-session.
- Session grows long enough to trigger context compaction (handoff-on-compact plugin active).
- On a subsequent turn, the reasoning part degenerates into a repetitive loop of "好。" and similar filler tokens, continuing indefinitely.
Expected Behavior
Either coherent reasoning output, or the provider/stream erroring cleanly. Not an endless degenerate loop.
Actual Behavior
The reasoning section rendered dozens of repeated short lines ("好。好。我输出。好。嗯。好。好。(写)好。好。...") in the desktop UI. Screenshot held by reporter; can attach on request.
Additional Context
- Uncertain whether this is an opencode bug (streaming/retry/compaction handling) or purely provider/model degeneration; happy to provide session JSON / logs if maintainers can tell which layer it belongs to.
- The session also exhibited a separate transient issue (404 "Message not found" toast after revert + resubmit), filed separately.
Reacted by rztz
Activity
github-actions commentedon Sep 26, 2026
This issue might be a duplicate of existing issues. Please check:
opencodejr commentedon Oct 28, 2026
Thanks for the detailed report, @eltPlay!
A reasoning stream that keeps repeating short filler like
好。is usually the model degenerating. OpenCode's built-in loop detection (doom_loop) only catches repeated identical tool calls, not repeated text. To tell whether OpenCode is involved (for example, a broken context sent to the model after compaction or the MCP tool-catalog change), could you share:opencode export <sessionID>, or the session JSON you mentioned). Please remove anything sensitive first. The parts we most need are the compaction summary and the messages just before the looping turn.handoff-on-compactplugin disabled? Does it only happen after compaction, or also in fresh sessions withqwen3.8-flashatxhigh?You can reply here with the details, or edit the issue. If you edit it, please also leave a comment saying you've updated it so we know you're done.
opencodejr commentedon Oct 31, 2026
Closing this for now, since we didn't get the details needed to look into it: the session export (especially the compaction summary and the messages before the looping turn), the screenshot, or whether the loop happens again with another model or with the
handoff-on-compactplugin disabled. Without these we can't tell whether OpenCode sent a broken context or whether the model (qwen3.8-flashatxhigh) degenerated on its own.Thanks for reporting it, @eltPlay! If it happens again, please reply here with those details or open a new issue, and we'll take another look.