New environment variables for subagent handback and host mid-turn prompts

What

Environment variables are settings you set in your shell before starting Claude Code. Several were added to the lists of variables Claude Code passes along or accepts (its allowlists):

• CLAUDE_CODE_DISABLE_SUBAGENT_HANDBACK is new. It turns off the SendMessage handback, where a subagent (a helper Claude runs for part of a task) hands back to the main conversation. Otherwise handback follows CLAUDE_CODE_SENDMESSAGE_HANDBACK, then the tengu_lively_waffle switch, which falls back to on.
• CLAUDE_CODE_DEEP_TWILIGHT is new and was added to two allowlists. When set, its value decides whether a prompt counts as input_host_mid_turn_prompt, meaning a prompt the host app asked for in the middle of a turn. When unset, that needs the tengu_deep_twilight switch (falls back to off; not read) and a further unresolved check. The prompt must also come from local input, declare no sender, have hostAsked and ownWord set, and not be a verified Slack human turn.
• CLAUDE_CODE_PARSED_WILLOW (switch tengu_parsed_willow, falls back to on; not read) gained the same kind of allowlist entry and also gates the mid-turn path.
• CLAUDE_CODE_DAPPER_CUPCAKE, the override for showing artifacts beside tool results (switch tengu_dapper_cupcake, falls back to on; not read), gained the same kind of allowlist entry.
• CLAUDE_CODE_DISABLE_HTTPS_AGENT_KEEPING was added to a list next to CLAUDE_CODE_PROXY_AUTH_HELPER_TTL_MS and ANTHROPIC_UNIX_SOCKET.

The captured SendMessage tool description for this release is unchanged.

Why

You can now opt out of subagent handback with CLAUDE_CODE_DISABLE_SUBAGENT_HANDBACK. The handling of mid-turn host prompts stays off unless its switch or CLAUDE_CODE_DEEP_TWILIGHT turns it on.