Kernel adds agent skills for Fixes tags and stable backports

The Linux kernel is gaining a dedicated home for coding-assistant “skills” aimed at the paperwork that bug fixes often get wrong: Fixes tags, stable backport tags, related copies of the same bug, and commit message text.

Sasha Levin posted the fourth version of the series after the Maintainers Summit raised a concrete pain point. Incorrect Fixes tags are common; Hannes estimated that 15 to 20 percent are wrong, often because an author ran blame and grabbed the first commit that touched the line rather than the change that introduced the bug. Levin’s group already uses LLM workflows to chase the real introducing commit, and others asked for those processes to be shared.

The series adds an agents area with short skill definitions that point assistants at existing process docs and spell out checks models routinely skip. One skill walks history past moves and refactors to choose a Fixes target, or returns nothing if the origin cannot be established. Another decides whether a fix belongs in stable, finds the oldest affected release, and checks that the fix applies there, guided by how stable trees actually accept or reject changes. A third searches for the same defect in sibling drivers, copied code, and shared interfaces, optionally with Coccinelle, and reports candidates without applying fixes. A fourth drafts or reviews commit messages, cover letters, and pull request text in the short imperative style Linus Torvalds has asked for, without inventing tags the human did not supply.

Skills are required to inspect and report only. They must not edit files, create commits or tags, or send mail, and they must not guess when evidence is missing. Documentation updates wire the skills into the assistant bug-fix path and add a MAINTAINERS entry with Levin as the contact.

Greg Kroah-Hartman reviewed the stable-tag skill. On the list, Thorsten Leemhuis urged clearer treatment of user-reported regressions at the top of the practical acceptance list, and noted that models often misread “regression” as any worsening rather than something that used to work and broke.