Skip to content

FAQ ​

Why not just write tests? ​

You should. But a test only catches a regression if someone wrote a test for that exact bug, and many real fixes (an API date format, a rounding rule, a null check for one customer) never get one. fixlock covers the fix from the moment it's committed, tells the AI before it edits, and names the person to ask. When a fix does have a test, record it with mark --test, and you get both.

Why not CODEOWNERS and branch protection? ​

CODEOWNERS works on whole files or folders: every change to date.util.ts needs Abi, even a comment. fixlock works on the few lines that matter, explains why they matter, and tells the AI before the edit, not after the PR is open. fixlock also generates CODEOWNERS from your locks if you want GitHub to request the review automatically.

How is this different from code review? ​

Review relies on the reviewer remembering that a line fixed a bug months ago. With AI-sized diffs, that line is one of hundreds. fixlock remembers for them and puts the lock, the reason and the owner at the top of the PR.

TestsCODEOWNERSCode reviewfixlock
Catches a reverted fix with no testif rememberedYes
Warns the AI before it editsYes
Line-level, with the reasonYes
Names who must approveYesYes
Proves behaviourYes

Does it slow developers down? ​

Locally, fixlock only warns: commits are never blocked, and your own fresh work never triggers a prompt. The only hard stop is the CI check, and it fires only when a locked line is removed. Day to day, most developers never see it.

Does the AI have to cooperate? ​

There are three layers:

  1. Claude Code: the pre-edit hook checks every edit, whether or not the model remembers to, and Claude is denied permission to edit the hook settings.
  2. Other assistants: they're instructed (rules file) and equipped (MCP) to check. Good models follow this reliably, but it isn't enforced.
  3. Everyone: the commit warning and the CI check don't depend on the AI at all.

So the worst case is that a reverted fix is caught at the PR instead of before the edit.

What if I rename a variable inside a locked line? ​

fixlock compares code structure, not text. Reformatting, comments, quote style and line wrapping don't count as changes. A consistent rename of a variable (user → account everywhere in the lock) is reported as a rename warning: the PR isn't blocked, but reviewers see it. Renaming a field or object key (user.role → user.type) reads different data, so it counts as a real change.

Two cautious exceptions: in a very short lock (a few tokens), or when the old name is still used in the code around it (isAdmin → isGuest while isAdmin is still in use), a rename counts as a change.

You can make renames stricter or quieter with matching.renames. How often matching gets it right on real code: see the accuracy benchmark.

Does fixlock send my code anywhere? ​

No. There's no telemetry. The network is used only to read PR approvals from GitHub in CI and to update your tracker, and only when you configure them. The MCP server runs locally over stdio. See Security & privacy.

Which languages does it support? ​

All of them. Locks are lines of text, so fixlock works for TypeScript, Java, Python, SQL, YAML, anything in git.

Do teammates need to install anything? ​

In a JavaScript project, npm install (fixlock is a dev dependency) and, optionally, npx fixlock install-hook for the commit warning. In a repo set up with the pip-installed command, each developer runs pipx install fixlock once, then fixlock install-hook. Everything else (locks, rules files, assistant configs) is committed to the repo.

Do I need Node.js? ​

No. pipx install fixlock (or uv tool install fixlock, pip install fixlock) installs a standalone binary for Linux, macOS and Windows on x64 and ARM. It's the same fixlock as the npm package, with the Node.js runtime built in. See Using fixlock in any project.

Is it free? ​

Yes, MIT licensed.

MIT licensed · A product of Algotricz