saw fix

Clean detected findings on a branch. By default fix prepares security/auto-clean locally and stops — no push, no PR, no network — for you to review. On a confirmed infection it also clears the checkout you are standing in: the files it confirms, the installed tree, generated build outputs, and the lockfile (the lockfile is kept on CI). A merge finding that is still live in the working tree is restored there; the merge commit is left in history. A merge commit that carries a known payload reads infected until saw fix amend takes it out of history; for such a merge alone, saw fix changes nothing in your checkout. --pr pushes and opens or updates one rolling PR per repository. Bare saw fix publishes nothing.

Those removals happen in your working tree, as the run happens. They are not staged on the branch and they do not wait for a pull request: the files are gone from the checkout you are standing in whether the pull request is merged, closed, or never opened.

On a confirmed infection everything a package manager put on disk is removed, whole — every installed tree in the repository, the dependency caches and resolver files a project uses instead of one, the lockfiles, and the generated output directories. Whichever package manager the project uses, and however many, the same holds. Nothing inside any of them is kept. Reinstall and rebuild to restore them; whatever installed them can install them again.

A package manager's own directory keeps its patches, plugins and pinned releases, and only what the resolver writes there is removed. A generated output directory goes whole, including what your repository commits inside it: nothing reads those directories, so nothing in one can be called clean. Name it under keep_dirs to keep it.

A known-malicious package your project declares stays declared unless you say otherwise. On a terminal, saw fix names each one and the package.json files that declare it, and asks: yes removes it from your checkout and from the branch it prepares, Enter keeps it, since your code may rely on it. Without a terminal it is never removed: it is named with the command to remove it, and your checkout is not called clean while it is still declared.

The directories a scan leaves out are read too, at any depth, and what they confirm is removed. An installed tree is read once the others have confirmed an infection, after it is removed. When part of your checkout cannot be read, the confirmed files are still removed one by one, but the installed trees are left and the checkout is not called clean. A link saw confirms is removed; the file it points at goes only when that file carries the payload itself.

What is not this repository's, it does not take. Anything linked to a location outside the repository loses the link, and what it points at is left alone. Directories you listed under keep_dirs are never removed, and neither is anything inside them. Write each one relative to the repository root — data, or dist/assets for one directory of a generated tree. A name alone keeps that directory at the root, not every directory sharing the name. A generated tree is cleared whatever it holds.

Your uncommitted work is kept. Once the cleanup is done, what your working tree holds and has not committed is recorded on a local branch named saw/uncommitted-…, as it then stands, under the user.name and user.email you set; with none set, it says how to set them and makes no branch. It is never pushed, and your branch is not touched. What the run removed is not on it, and neither is a file saw confirmed, nor one holding the same bytes as a file it confirmed; a file it cleaned in place is saved cleaned. A file your .gitignore covers, or one saw could not read, stays on disk only.

What it confirmed leaves your staged changes too. Your staged changes are read as well as the files on disk, so a staged file you have since changed or deleted is still found. A staged file saw confirmed goes back to what your current commit holds, or leaves your staged changes when that commit holds nothing clean there. A file in the middle of a merge keeps its other versions and stays unresolved, so git will not let you commit it until you choose one. Everything else you staged is left as it was.

Your history is reported on its own. When the commit you are on, another local branch or a stash entry still stores what the run removed, or a payload your commit holds that you have since deleted or changed, on disk or in your staged changes, the run says so and names the remedy: the fix pull request for a commit the base branch already has, and saw fix amend for your own commits. A stash entry is named only, because it also holds your own work.

Files added in the same commit as the malware are yours to decide. When that commit added other files with no finding of their own, saw never removes them by itself.

• On a terminal, for one repository, saw fix asks about them the same way saw fix amend does. A file you take out is removed in the branch it prepares, as the payload is; your earlier commits keep it until saw fix amend rewrites them, and the run says where it is still stored. A file you keep is not asked about again. Your answers are kept for both commands.
• Anywhere else — no terminal, CI, several repositories, --remote — the run removes none of them, names them, and is not called done or clean.

fix cleans your working tree and records that as a new commit. What the repository already stored stays stored: the payload is still there in the earlier commit, and one git show puts it back on disk. Anyone who cloned or forked the repository still has it too, and nothing you do to your own copy reaches theirs. Clearing it needs a history rewrite and the hosting provider's collection — deliberate work, not something fix decides for you. Run saw scan --history to see what is still stored.

saw fix amend is a different act: it amends the infected commits and force-updates the branches that carried them. Read that page before you run it.

See the safety envelope for what fix will and will not touch.

saw fix [TARGETS...] [--pr] [--json] [-r] [--user U] [--org O] [-p PATH] [-c FILE] [-j N] [--no-stream]

Option	Description
TARGETS... / -p / -c / -r / --user / --org / -j / --no-stream	As for saw scan. A missing explicit --config path is a clear error, never a crash.
--pr, --open-pr	Also push the branch and open/update one rolling, de-duplicated PR per repository. Needs a credential with repo + PR write; the API is pre-flighted before any push. Not accepted with amend.
--json	Print what each repository's fix came to as JSON on stdout, one record per repository: repository; remediation, one of pr, fork-pr, patch, issue, prepared, skipped or failed; cause, why it came to less than a pull request, such as actions-may-not-open-pull-requests; and the pull_request number and url. Progress stays on stderr. Not accepted with amend.
amend	Replace past commits that still carry the payload and force-update each branch they sat on — see saw fix amend. Not accepted with --pr, --branch, --user or --org.