Kernel gains LLM agent skills for Fixes tags and stable backports

Sasha Levin has proposed a set of coding-assistant skills for the Linux kernel aimed at four chores that trip up both humans and models: naming the right Fixes commit, deciding stable tags, hunting the same bug elsewhere in the tree, and writing short commit messages.

The work answers a Maintainers Summit question from Ted Ts'o about bad Fixes tags. Hannes Reinecke has estimated that 15 to 20 percent are wrong, often because authors trust git blame and pick a commit that only moved the code. Levin's group already uses LLM workflows to find the commit that actually introduced a bug; other developers asked that those workflows be shared.

The skills live under an agents directory as short pointers into existing process docs, plus checks models commonly 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. A third looks for sibling copies of the same defect, optionally with Coccinelle, and only reports verified hits. A fourth drafts or reviews commit messages, cover letters, and pull-request text in the short imperative style Linus Torvalds has asked for. The skills inspect and report only; they do not edit trees, create tags, or send mail.

Alongside the skills, Levin clarifies submitting-patches guidance: a Fixes tag names the commit that introduced the bug, may add a second tag for a later commit that merely exposed it, and may list every independent introduction, because those tags tell the stable team which trees need the fix. Stable-kernel-rules text is brought in line with practice. The 100-line cap, the blanket bar on theoretical issues, and the "obviously correct" bar are dropped in favor of requiring fixes that are correct and tested. Greg Kroah-Hartman reviewed that change.

Miguel Ojeda asked that Rust soundness fixes be called out explicitly for backporting, noting an earlier agreement with the stable team. Bradley Morgan reviewed several of the patches and urged clearer expectations for what "tested" means when agents are involved. Levin plans to keep the agent notes trimmed against Documentation each release so the two do not drift.