Stops "fake green": blocks an agent from making a suite pass by deleting, renaming-away, or skip-disabling tests instead of fixing the code.
A PreToolUse hook fires before every Bash, Edit, MultiEdit, and Write call and denies the ones that remove or disable tests: rm / git rm of a test file or test directory, mv of a test file to a disabled name (.bak, .disabled, ...), an edit that introduces a skip/xfail/ignore marker into an existing test (through Edit, a whole-file Write, or a shell in-place edit or redirect such as sed -i, perl -pi, >, >> or tee), and (at the strict level) writing a brand-new test file that is already skipped. Test paths and skip markers are matched across pytest, unittest, jest/mocha/vitest, Go, JUnit/TestNG, NUnit/MSTest, RSpec, and Rust. It does NOT block writing new, real tests, refactor-renaming a test to another test name, or editing test bodies: only removal and disabling.
/plugin marketplace add karanb192/claude-code-hooks # once per machine
/plugin install protect-tests@claude-code-hooks
Restart Claude Code, done. (Or from a shell: claude plugin install protect-tests@claude-code-hooks.)
codex plugin marketplace add karanb192/claude-code-hooks
codex plugin add protect-tests@claude-code-hooksOpen /hooks in Codex and approve this plugin. Re-approve after every update; Codex skips unapproved or changed hooks. Do not install it alongside guard-pack, which includes this guard.
This entry point checks Bash and native apply_patch calls. Patch files are checked individually; unreadable patches are blocked. Shell patches must use a plain command with a quoted heredoc. Literal searches and commit messages mentioning apply_patch pass. See the shared adapter's limits, including shell sessions and dynamically constructed commands.
Export configuration variables before launching Codex. An ask-mode match becomes a block because Codex hooks cannot ask. Logs stay in ~/.claude/hooks-logs/ with host: "codex", including on Codex-only machines.
| Tool | Blocked when | Level | Deny id |
|---|---|---|---|
Bash |
rm / unlink / shred / trash / git rm targets a test file or test directory (tests/ or a bare tests), find deletes test files (-delete, -exec rm), or mv moves tests into /tmp, $TMPDIR, ~/.Trash or /dev/null |
critical |
delete-test |
Bash |
mv renames a test file to a disabled name (.bak, .old, .orig, .disabled, .skip, .ignore, .tmp, ~) |
high |
rename-test |
Edit / MultiEdit |
the edit adds a skip/xfail/ignore marker to an existing test file | high |
skip-test |
Write |
it rewrites an existing test file and the new content has more skip/xfail/ignore markers than the file on disk | high |
skip-test |
Bash |
sed -i, perl -pi or ruby -pi substitutes a marker into a test file named in the same command (a glob such as find -name '*.test.js' | xargs sed -i counts), or >, >> or tee writes one into a test file |
high |
skip-test |
Write |
a brand-new test file's content is already skipped/ignored | strict |
write-skipped-test |
Markers are counted, not just detected, so adding a second it.skip( to a file that already has one still counts. Direct edits that remove skips are allowed. For Write and shell redirects the hook reads the target file (relative paths resolve against the session's working directory; files over 2 MB are not read). Each complete command joined by ;, &&, || or & is judged on its own. Groups and continued pipelines, including a comment after |, keep their output with the destination. Literal braces in command arguments do not join unrelated commands.
For a small, literal sed substitution, perl -pe substitution, or grep filter with one source file and one redirect or tee destination, the hook reads the source and compares the resulting content. This allows a command that removes all skips while still denying a positive grep that copies skipped tests. Output inference supports a restricted pattern syntax and ASCII text, with limits on command length, matching work and output size. Unsupported commands use the command-text heuristic, so they may be conservatively denied. Search-looking text inside printf arguments or heredocs is always treated as literal content, not stripped away.
An in-place replacement counts when it completes a marker with the text around it (s/test/test.skip/ on test(). This remains a guardrail, not a shell evaluator or sandbox: arbitrary language scripts, cp of a skipped file over a test, and a heredoc that creates a brand-new skipped file still pass at high.
Anything at or below the active safety level is denied with a reason Claude sees ("Fix the code, don't disable the test"); everything else passes through. On malformed input the hook fails open (emits {}), so it can never wedge a session.
| Level | Blocks |
|---|---|
critical |
deleting test files or whole test directories (rm / git rm) |
high (default) |
+ renaming a test file to a disabled name, + adding a skip/xfail/ignore marker to an existing test, by any of Edit, Write or a shell edit |
strict |
+ writing a whole new test file that is already skipped |
One optional environment variable:
| Variable | Default | Purpose |
|---|---|---|
HOOK_SAFETY_LEVEL |
high |
Safety level: critical, high, or strict. Any other value falls back to high. Set it via env (shell profile, or inline in a hook command), not by editing the installed script: plugin updates overwrite plugin files. |
HOOK_FAIL_OPEN |
unset (deny) | The literal string true lets the call through when the guard's own check throws, instead of denying it. See When the guard fails. |
hooks/hooks.json registers one PreToolUse hook with matcher Bash|Edit|MultiEdit|Write, running synchronously (a deny decision has to land before the tool executes).
From 1.3.0, a guard that cannot finish its check refuses the call instead of allowing it. A check that throws is denied with a [guard-error] reason. A timeout (10 s) or a missing node is blocked by "onFailure": "block" on Claude Code v2.1.295 or later; older versions ignore that key and let the call run. On Codex a missing node is blocked by the command's || exit 2, but a timeout still lets the call run, because Codex has no setting for it. Set HOOK_FAIL_OPEN=true to let the call through when the check throws; When a guard fails covers the rest of the way back.
Blocked calls and parse errors are logged as JSONL to ~/.claude/hooks-logs/YYYY-MM-DD.jsonl (deny id, level, tool name, session id, cwd, permission mode; no commands or file contents, though a parse error's message can quote the first few characters of a malformed payload). The script reads the target test file and, for supported literal shell transformations, the source file. It never runs the proposed command, and nothing leaves your machine.
/plugin uninstall protect-tests@claude-code-hooks
Part of claude-code-hooks, a marketplace of production-ready Claude Code hooks.