Claude Code 2.1.294 Fixes Instruction Hooks That Failed to Block

Claude Code 2.1.294 is a two-line release. The GitHub tag was published on 8 October 2026 at 05:03:54Z by ashwin-ant. The first line says the team fixed prompt and agent hooks written as instructions, with the example "Block commands that...", because those hooks were allowing what they should block.

That is a harness failure of a specific kind. A hook that reads like a policy sentence is still a control. If the matcher treats the sentence as advice instead of a deny rule, the command the operator tried to forbid still runs. The note does not publish the old matcher or a sample payload. It names the class of hook and says the allow-through behavior is fixed.

The stop hook is a separate judgment

The second line is about prompt hooks on Stop and SubagentStop, again written as instructions. The example in the notes is "Carry on if the build is broken." The release says those hooks are now judged so Claude is less likely to stop early. A stop hook and a block hook are different controls. One decides whether a command is permitted. The other decides whether the turn is finished. 2.1.294 says both were misfiring when the hook text was an instruction rather than a structured rule.

The release page does not list a pull-request number. The assets are the usual platform archives plus SHASUMS256.txt.

Source: Claude Code v2.1.294.