Auto-mode permission classifier blocks kill on orphaned process even after explicit user confirmation, with no override

Summary

An orphaned sibling Claude Code process (no reachable window to close it from) needed to be killed. The user explicitly confirmed "yes, kill it" in response to Claude's own prompt. Claude then attempted kill <pid> and was blocked by the auto-mode permission classifier ("Blocked by classifier"), with no path to override even given direct, explicit user confirmation.

Impact

The user had to manually open a separate Terminal window and run kill/kill -9 themselves. Several follow-up checks showed the process still hadn't died, extending what should have been a one-step fix into a multi-turn manual workaround.

Request

• Allow explicit, in-context user confirmation to override this specific classifier gate for process-kill actions, at least when the target is a Claude-owned/orphaned process rather than an arbitrary system process.
• If that's not desired, surface a clearer reason than the generic "Blocked by classifier" denial, so the safety intent (e.g. "kill is never auto-approved regardless of confirmation") is legible to both the model and the user in the moment, rather than looking like a transient/retryable failure.

Environment

• Claude Code CLI version: 2.1.231
• Platform: macOS (Darwin 24.6.0)