Hooks

One at a time. Tell it whether you knew, and the ones you did not stay in your pile.

1 / 7

Exit code 2 blocks the action; any other non-zero code does not.

On exit 2 the stderr is fed back to Claude. Choose exit codes or JSON output, not both.

The whole topic7

Exit code 2 blocks the action; any other non-zero code does not.
On exit 2 the stderr is fed back to Claude. Choose exit codes or JSON output, not both.
Only command and http hooks are deterministic.
prompt and agent hooks ask another model, so they are judgment, not a guarantee.
A Stop hook can keep Claude fixing until the typecheck passes.
Exit 2 sends the errors back and the turn continues. Redirect the output to stderr so Claude sees it.
A PostToolUse hook on Edit|Write removes the "remember to format" rule.
Deterministic formatting after every edit beats a line in CLAUDE.md that Claude might skip.
A Stop hook that blocks eight times in a row is overridden.
The turn then ends. CLAUDE_CODE_STOP_HOOK_BLOCK_CAP raises the limit.
A PreToolUse hook can rewrite a tool's input before it runs.
For example, turn a test command into one that greps only failures, so far less output enters your context.
Hooks have no CLAUDE_MODEL variable.
A PreModelSwitch hook receives the from and to models. The input JSON always carries session_id, cwd, permission_mode and effort.