Repository navigation
Compaction keeps an unbounded recent-context in agent sessions with few user messages #60391
Description
Description
Summary compaction keeps the newest compaction.keep.tokens (default 15,000) of conversation verbatim beside the summary. But recentStart then moves the start back to the latest type: "user" message at or before the cut, so tool calls stay with their results (packages/core/src/session/compaction.ts:1004-1027, v2.2.2).
In a long agent session, real user messages can be far apart. Synthetic, skill and shell rows don't count as user messages. So the kept range can run from a weeks-old user message to now: thousands of tool steps, whatever keep says. Nothing bounds it against the model's context window.
The kept text isn't stored. withRecentText (packages/core/src/session/history.ts:226-263) rebuilds it from every row between metadata.keepID and the compaction on each load. So the oversized checkpoint is resent on every request, and the session can't recover by itself.
We saw this on 2.2.2 in a session of about 5,900 messages. The compaction was triggered by /goal when its reviewer request overflowed. The rendered checkpoint was about 9 MB, almost all of it in <recent-context>, and every later request failed at about 2.2M tokens.
Suggested fix: cap the kept range in tokens. If the nearest user message would push it past the cap, start at an assistant step boundary instead, keeping each tool call with its result. Or summarize everything and keep no recent context.
Plugins
Magic Context supplied the summary through the compaction hook. The kept range comes from the host either way: summarize adds keepID to a hook-supplied result too (compaction.ts:523-525).
OpenCode version
2.2.2
Steps to reproduce
- Run an agent session with one user message followed by many tool steps, well past 15,000 tokens.
- Compact it, with
/compactor/goaloverflow. - Inspect the next request:
<recent-context>contains everything since that user message, not about 15,000 tokens.
Operating System
macOS
Activity
opencodejr commentedon Oct 29, 2026
Thanks for the detailed write-up and the numbers from your 5,900-message session.
I think this is a duplicate of #43250. It reports the same problem: after the
keep.tokenscut,recentStartmoves the start back to the nearesttype: "user"message with no limit, and synthetic and shell rows are skipped. So a long agent run after one prompt is kept verbatim in the recent context. That issue already has measured cases from autonomous sessions on 2.0.15 through 2.0.24, and the code is still the same in 2.2.2 and ondev. Your/goaltrigger and the rebuild inwithRecentTexton every load make the impact worse, but the root cause is the same.Possible duplicates and related issues, ordered by likelihood:
keep.tokensin agent-driven sessions (open)/compactkeeps a long final turn verbatim in the next request, from the same causekeep.tokensand summarizes an oversized latest exchange insteadYour measurements, especially the ~9 MB checkpoint and the session that couldn't recover, would be useful on #43250.
The bot will close this issue as a duplicate in 1 day unless you reply explaining how your problem is different.
opencodejr commentedon Oct 30, 2026
We haven't heard back, so I'm closing this as a duplicate of #43250. That issue covers the same problem: the kept range moves back to the latest user message with no limit, so agent sessions with few user messages keep far more than
compaction.keep.tokensverbatim beside the summary. Please follow #43250 for updates; open PR #53233 is the fix aimed at it. Your measurements, especially the ~9 MB checkpoint from/goaland the session that couldn't recover, would be useful there. If your case turns out to be different, reply here and we can reopen.