Claude Code 2.1.259 Ships the Opposite of --dangerously-skip-permissions
Unattended agents have been running on auto-yes for two years because there was no other way to run them. The new flag makes the same silence mean no.
NeuroX AI · September 3, 2026
The Claude Code 2.1.259 changelog added one flag for headless hosts: --permission-prompts none, and the bullet defines it in seven words — anything that would prompt is denied automatically.
Look at what that replaces. Until today an unattended run had two options, both documented in the CLI reference. --dangerously-skip-permissions, which is exactly --permission-mode bypassPermissions — every prompt resolves to yes, nobody watching. Or --permission-prompt-tool, which hands each decision to an MCP server the session waits up to 30 seconds to reach before the first turn. One is a blank cheque. The other is infrastructure you have to build and keep online.
So most CI pipelines picked the blank cheque. The prompt was never a safety feature there; it was a thing to get rid of. A flag that turns "nobody is here to answer" into a denial closes the gap that made that a reasonable choice.
The cost lands where it should. Your allowlist stops being documentation and becomes the spec — anything you forgot to grant now fails the run instead of quietly happening. That is a better failure. Nine bypasses closed two releases ago were all cases of the harness saying yes when a rule said no. This is the same lesson from the other side: an agent with no one to ask should not be guessing.