Auto-compact fails at context limit and forces a new session, while manual /compact succeeds
Summary
In a very large, long-lived session, auto-compact triggered near the context limit, failed during summarization, and the CLI told me to start a new session. Running /compact manually in the same session worked fine and immediately dropped context usage.
Steps to reproduce
- Run a long-lived session with many large file reads/writes (session transcript grew to ~171 MB / ~20,900 events)
- Let context usage climb to ~97-100%
- Auto-compact triggers → summarization fails → prompted to start a new session
- In the same session, run
/compactmanually → succeeds, usage drops immediately
Expected behavior
Auto-compact should trigger with enough headroom to actually produce the summary, and should fall back (retry / partial summarization) instead of dead-ending the session.
Actual behavior
Auto-compact fails at the limit. Manual /compact at the same point works, which suggests summarization was still possible — the automatic trigger just fired too late, after the headroom was already consumed.
Additional observation
Context usage stayed pinned at ~97% across turns and would not come down. A scratch directory containing ~90 files was being re-injected as file-change notifications every turn, steadily consuming context. Removing those files stopped the re-injection.
Deduplicating or capping repeated file-content injection would help sessions like this stay usable.
Environment
- Claude Code, desktop app
- Model: Opus 5
- OS: Windows 11