Back to rankings

Claude Code Idle Compaction Comes to VS Code, Claude Desktop and Bedrock, Vertex and Foundry

Claude CodeView original changelog

Claude Code 2.1.314 extends idle compaction, first shipped in the terminal in 2.1.299, to the VS Code extension and the Code tab of Claude Desktop. A conversation of 200K or more tokens left idle is summarized, and the summary is used on the next message if the prompt cache has expired. The same release also makes it run for subscribers with telemetry turned off and on Bedrock, Vertex, Foundry and LLM gateways once the one-hour prompt cache is on.

Key Takeaways

  • Idle compaction now covers the VS Code extension and Claude Desktop's Code tab, not just the terminal.
  • Conversations of 200K or more tokens left idle are summarized before the cache lapses.
  • The summary is used on the next message only if the prompt cache has expired.
  • The release also enables it for subscribers with telemetry turned off.
  • Bedrock, Vertex, Foundry and LLM gateway users get it once the one-hour prompt cache is on.
  • The cost is that the resumed session runs on a summary rather than the full transcript, so fine detail can be lost.

Idle compaction beyond the terminal

Idle compaction first appeared in the Claude Code terminal in 2.1.299. The idea is simple: a very long conversation that sits untouched is summarized before its prompt cache lapses, so coming back to it does not mean re-reading the entire history at full price. Claude Code 2.1.314 brings the same behavior to two more places: the VS Code extension and the Code tab in Claude Desktop.

Anthropic's description is specific. A conversation of 200K or more tokens that is left idle gets summarized. If the prompt cache has expired by the time the person sends their next message, the summary is used in place of the full history.

Wider eligibility

A second change in the same release widens who gets the feature. Idle compaction now also runs for subscribers who have telemetry turned off. It also runs on Amazon Bedrock, Google Vertex, Microsoft Foundry and LLM gateways, once the one-hour prompt cache is enabled there. That removes two reasons people on those setups could not benefit from it earlier.

The trade-off

Resuming becomes cheaper and faster because the cache is not rebuilt from a huge transcript, but the session then continues from a summary and not from the full original text. Detail that the summary leaves out is no longer available to the model. People who rely on exact earlier output in a long session should know that behavior before stepping away.

Who benefits

Developers who keep one VS Code or Desktop session open all day, leave for lunch or a meeting, and return to a conversation hundreds of thousands of tokens long are the intended audience. For them, the first message after a break should cost less and respond sooner.