Does this replace auto-compact?
No. Claude Code’s autocompact still runs when the context nears the window limit; keep it on as a safety net. Compactly uses Claude Code’s compaction engine. It adds a decision about when to compact, tailored handoff instructions, and automatic resend of the held prompt.
See the full comparison →
How is this different from idle compaction?
Claude Code’s idle compaction runs after about 55 minutes idle once the context is at least 200K tokens, and writes a general summary. Compactly compacts at a natural break in the work, before your next prompt, with a handoff for the next task. Idle compaction depends on time, not on work.
What happens to unfinished work?
The handoff includes open work, decisions with their reasons, and relevant file paths. Recent wording that suggests paused or blocked work suppresses signal-triggered compaction; this is a heuristic.
Does it keep the cache warm?
Compaction reduces the context a later turn needs to rebuild. It does not extend the cache lifetime.
What does the judge see?
At or above your context threshold, it receives a short digest: the opening task context, recent prompts, edited file names, the start of the latest answer, and your next prompt. Claude Code sends this digest to the selected model. Ratings and savings are stored in Claude Code’s local plugin store; digest examples are reused only in the project where you rated them.
Can I review the decision myself?
Compactly checks at or above your context threshold (150K by default). If context is below it or the judge is not confident a break has happened, your prompt continues without compaction. Switch to suggest mode for boundary notifications instead; the threshold and judge settings are configurable.