Loading data…
Shallow clone of torvalds/linux with --shallow-since="2026-01-01", then git log --all --grep="Assisted-by:" -i. Each merged commit and each Assisted-by: line counted directly. Denominator is git log --since=2026-01-01 --oneline | wc -l on the same clone.
Submissions to lkml and subsystem lists pulled with lei from lore.kernel.org/all: lei q -d mid -f mboxrd 'b:"Assisted-by:" AND d:20260101..'. Replies, cover letters (0/N), and bot accounts (Patchwork, kernel test robot, syzbot, 0day) are dropped. Patch series respins (v1, v2, v3) collapse to one entry per (canonical subject, sender). The Assisted-by: line must appear in non-quoted body text.
Each tag string is parsed into {vendor, model, tool}. Claude variants (Opus 4.5, 4.6, 4.7; Sonnet 4.5, 4.6) collapse cleanly. Wrappers (Claude Code, Cursor, GitHub Copilot, OpenCode, Kiro, Cody) attribute the model to its actual lab; the wrapper appears under "By tool" instead. Tags that do not parse (free-text, joke names, FOO:BAR.baz) bucket as Unknown.
Merge rate (different humans use different tools for different patch types; the ratio is not a model quality signal). Submitted-versus-merged percentages (same reason). Authorial intent or motivation behind any tag string. Patches that landed without disclosure (this page measures policy compliance, not actual AI usage). Share of kernel-wide lines: GitHub's stats API refuses repos over 10k commits and computing diffstats from the partial clone would take hours; the AI-side line counts stand on their own.
Source on GitHub: snek-git/assisted-by. Three small Python scripts: parse_commits.py (merged side), parse_lei.py (submitted side), build_data.py (assembles the JSON payload). refresh.sh runs the full pipeline end to end.
Built with assistance from Anthropic's Claude (Opus 4.7, via Claude Code) by a human collaborator. Tag string normalisation choices are judgement calls; the parser source on GitHub is the authoritative answer for any "why was X bucketed as Y?" question. Where contributors have publicly explained their setup (e.g., Greg KH's local LLM fuzzer rig), the page links the source; where they have not, it does not guess at intent.