October 8, 2026
Claude Code stops confusing prohibitions with permissions: instruction hooks fixed
Claude Code v2.1.294, released on Oct 8, fixes 2 bugs in action and task completion checks.
Claude Code could allow a command that a hook instruction required it to block. Anthropic fixed this bug in v2.1.294.
Hooks define automatic checks for an AI agent. In Claude Code, you can describe them in words: for example, prohibit certain commands or require the agent to fix the build before finishing a task.
Working through to completion. Anthropic also improved how prompt hooks evaluate instructions on Stop and SubagentStop events. With an instruction such as "Keep going if the build is broken," Claude should be less likely to stop working prematurely.
The check works the same way for both events. An `ok: false` response forces the agent to continue working, while the `reason` field supplies the next instruction. A prompt hook requires `type: "prompt"` and a `prompt` field, while `$ARGUMENTS` inserts the check's input data. Anthropic's documentation includes ready-to-use JSON for checking task completion on Stop.
Updating and configuring. An existing Claude Code installation can be updated with `claude update`. Hooks are stored in `.claude/settings.json` for a single project or `~/.claude/settings.json` for all projects.
Native installation on macOS, Linux and WSL starts with `curl -fsSL https://claude.ai/install.sh | bash`. The installer supports choosing a specific version. Claude Code belongs to the coding agents category.
Sources: [Claude Code v2.1.294 release](https://github.com/anthropics/claude-code/releases/tag/v2.1.294), Anthropic documentation dated Oct 8, 2026.
The `claude --version` command shows the installed version.