Cache warm-up on resume for remote sessions, with CLAUDE_CODE_RESUME_CACHE_WARM

What

When a remote session resumes, Claude Code can now send a short one-turn background request labelled resume_cache_warm. A session is one conversation with Claude, and resuming means picking it up again later. The request fills the prompt cache before your first message. The prompt cache is a store on the server that saves reprocessing the conversation each turn.

• It runs only when CLAUDE_CODE_REMOTE is set.
• The environment variable CLAUDE_CODE_RESUME_CACHE_WARM forces it on (true) or off (false). An environment variable is a setting passed to Claude Code from your shell.
• Otherwise it follows the remote setting tengu_resume_cache_warm_config. The fallback values in the code are enabled: false, a delay of 2000 milliseconds, at least 10,000 tokens of context, and only the remote_cowork entry point. Nothing has been read about what this setting holds.
• It is skipped for many reasons, for example if your own message gets there first, a command is queued, a budget cap applies, or the thinking-budget setup is not safe for it.
• The request is allowed through a check that normally stops bare forked requests early.
• Resume also times a read-ahead of the transcript, counted in the transcript scan time.

Why

The first reply after resuming a remote or cowork session can come back sooner. The warm-up spends tokens on a background request, so CLAUDE_CODE_RESUME_CACHE_WARM=false is how to turn it off.