jord0.hooks¶
Three Claude Code hooks. Each one exists because of a failure that already happened.
| Hook | Fires on | Stops this |
|---|---|---|
| DONE-GATE | Stop | "Done, all fixed" when no test has run since the last code edit, or the last one failed |
| RECOVERABLE | PreToolUse · Bash | a delete or a git discard that would destroy work git cannot give back |
| FLAG-PROBE | PreToolUse · Bash | --help on a script that does not parse it and has not been read |
They are seeds, not a harness.
Each one is small enough to read in a sitting. Each one says why it refuses, in words Claude can act on.
Two commands¶
Claude Code installs the two npm dependencies itself, with install scripts switched off. There is no second step.
Read Install first. Installing is silent, and the hooks fire in every session from then on.
Three rules they share¶
No guard fails open. A PreToolUse hook that crashes lets the tool run. So every way these can fail becomes a decision: ask on the main thread, deny inside a subagent, and a reason naming the fault.
No guard says allow. In Claude Code, allow skips your permission prompt. A guard with no objection says nothing at all, and your own settings decide.
Commands are parsed, never split. tree-sitter's bash grammar finds the commands inside ( … ), if … then, $( … ), a redirection and a heredoc fed to a shell. A commit message that mentions rm -rf is one quoted word. A command the grammar cannot parse counts as unread, never as clean.
Where they came from¶
A research subagent deleted an untracked project it took for clutter, then reverted two uncommitted fixes with git checkout.
A --help probe ran the script it was asking about. The script rewrote a config, and two models were evicted from memory.
A "done" stood on 92 green tests. The suite beside them never ran.
The hook pages tell each one properly. How hooks actually work covers the machinery the official docs leave thin, and How this is tested covers the tests, including the ones that were proven able to fail.