Fork a subagent instead of spawning fresh

Claude Code 2.1.117 introduced forked subagents. Enable with CLAUDE_CODE_FORK_SUBAGENT=1, then /fork spawns a subagent that:

1. Inherits the full parent conversation context.
2. Shares the parent's prompt cache (no cache-miss cost on the first turn).
3. Runs independently — its edits come back but its context doesn't pollute the parent.

Use it when: the subagent needs to know why you're doing what you're doing. E.g., you've spent 30 turns debugging a race condition, and now you want a subagent to write the fix without re-explaining the codebase.

Don't use it when: you want a clean context (use a fresh general-purpose agent). Or when the task is small enough that inheriting the whole parent is wasteful.

The prompt-cache part is not cosmetic. A fresh subagent on a long parent context pays the full input-token price on its first turn. A fork pays nothing new — the cache hit is ~90% cheaper and several times faster. For agent-heavy workflows, this compounds.

Check it's enabled with /fork at the prompt. If the command isn't available, the env var isn't set or the CLI is older than 2.1.117.