A large compaction.buffer leaves small models no room before compaction #58146

Activity

1. opencode-agent commented on Oct 22, 2026

   

   opencode-agentBot

   on Oct 22, 2026 – with OpenCode Agent

   Contributor

   More actions

   Thanks for the clear report and for porting this from #54994. I think this is a duplicate of #56783, which reports the same root cause on v2 and has already been reproduced on v2.0.32 and on dev. When one global compaction.buffer is at or above a model's window, window - buffer comes out at zero or below, so automatic compaction is due on every request. Your buffer: 100000 on a 100k window (and on a 64k window) is that case.

   Possible duplicates and closely related issues, ordered from most to least likely:

   • V2: compaction fires on first request / loops when compaction.buffer >= model context limit #56783: same problem, compaction.buffer >= the model's context limit makes auto-compaction fire on every turn (open, reproduced)
   • V2: compaction fires on the first request and repeats every turn when compaction.buffer >= model context limit #56813: the same report again, already closed as a duplicate of V2: compaction fires on first request / loops when compaction.buffer >= model context limit #56783
   • fix(core): bound auto-compaction when the buffer exceeds the window #56786: open PR for V2: compaction fires on first request / loops when compaction.buffer >= model context limit #56783. It treats a buffer at or above the window as unset, so the default reserve applies
   • fix(core): cap a configured compaction buffer at a quarter of the window #58147: your PR, which caps the buffer at a quarter of the window instead. It would be worth linking it to V2: compaction fires on first request / loops when compaction.buffer >= model context limit #56783 so maintainers can compare the two approaches
   • [FEATURE]: per-model compaction threshold (reserved/buffer) #43703 / feat(core): allow per-model compaction buffer #56443: request and PR for a per-model compaction buffer, which would cover the case of a large buffer that is still below a smaller window

   The bot will close this issue as a duplicate in 1 day unless you reply explaining how your problem is different from #56783.
2. opencode-agent

   added

   pending close

   triaging

   on Oct 22, 2026
3. rekram1-node

   closed this as completedin #58313on Oct 23, 2026
4. opencode-agent

   closed this as duplicateon Oct 23, 2026
5. opencode-agent commented on Oct 23, 2026

   

   opencode-agentBot

   on Oct 23, 2026 – with OpenCode Agent

   Contributor

   More actions

   Closing this as a duplicate of #56783, since there's been no reply saying how this one is different. Both report the same problem: when compaction.buffer is at or above a model's context window, window - buffer comes out at zero or below, so automatic compaction is due on every request. Please follow #56783 for updates.

   Two PRs would fix it in different ways: #56786 treats an oversized buffer as unset, and your #58147 caps the buffer at a quarter of the window. Linking #58147 to #56783 would help maintainers compare them. Per-model buffers are tracked in #43703 and #56443.

   Until a fix lands, keeping compaction.buffer below the smallest context window among your models avoids the problem. If this turns out to be a different problem, reply here and we'll take another look.