Home / Context and cost

The auto-compact threshold does not do what you think

The auto-compact threshold triggers a summary of the conversation when the context reaches it, it is not a wall that closes the window: a value set above the real window of the active model is capped at that window, a documented behaviour whose lack of an on-screen warning still needs to be checked for yourself.

The auto-compact threshold is not a fence that would stop the context from exceeding a given size. It is a trigger point: when the conversation reaches this threshold, Claude Code launches an automatic summary of the history. Nothing in this mechanism closes the window itself, and a threshold value set above the real window of the active model produces no displayed error: it is simply capped at that window.

The cap, with no guaranteed warning

The official documentation states it plainly: the limit that applies is always the model's own, whatever the threshold value configured elsewhere. Setting a threshold to one million tokens while the active model only has a window of two hundred thousand tokens pushes nothing back, the calculation falls back to the real two hundred thousand tokens. The documentation does not explicitly say whether this capping comes with an on-screen warning or not, it only documents the capping itself: checking it on your own machine is worth more than assuming it is silent.

When the summary loops on itself

A second mechanism complicates the picture. After a compaction, if the context fills up again to the limit almost immediately, and this cycle repeats several times in a row, Claude Code stops relaunching compaction and displays an explicit error rather than continuing to loop.

Autocompact is thrashing: the context refilled to the limit...

This opening line, which the documentation itself quotes before cutting off with an ellipsis, signals an anti-thrashing guard, a mechanism that recognises that one summary after another solves nothing as long as the source of the filling, a large file reread on every tool call for example, remains in the conversation. The guard stops after a few attempts and surfaces this error rather than continuing to loop silently.

A compaction you did not start

Since October, a long conversation left idle in the terminal can be compacted shortly before the one hour cache expires, so that the next message does not reprocess the whole conversation. The documentation page on the cache says so and gives the idleCompaction setting to turn it off, and the notices shown after such a compaction say how. A discomfort felt as "the context window" can therefore be this compaction: note each time whether it was manual, at saturation or at rest before touching a threshold. The figure gives the delay and the size taken from the changelog.

Two other changes concern the threshold itself. The /autocompact setting is saved per model, so each model keeps its own when you switch. The CLAUDE_CODE_AUTO_COMPACT_WINDOW variable accepts the k and M suffixes: the value 200k used to silently set a window half the size, and the variable now warns about a value that is out of range or ignored.

What to do about it

First, a custom threshold set higher than the model's window is useless, you need to check the real window of the current model before configuring anything. Second, a session that displays a thrashing error signals a content problem, not a software fault: the right move is to remove what is filling the context rather than relaunching the same action hoping for a different result. The neighbouring lesson on compacting, clearing, reading only what is needed details how to choose between these two commands once this threshold is understood.