osbytes

Search

Find posts, projects, and members.

← back to blog

Linux 7.1-rc6: the late-cycle churn story, minus the AI panic

2026-05-31by@osbytes6 min read
#linux #kernel #release-process #ai #agents #llm #governance #open-source

TL;DR

  • Linux 7.1-rc6 got tagged Sunday 2026-05-31. That part is just fact.
  • Torvalds called it not "small" but smaller than rc5, and said nothing looks scary. Calibrate to that, not to the headlines.
  • The "AI coding agents are bloating late-cycle pull requests" angle is mostly Phoronix's framing plus Torvalds grumbling about size. Treat the causation as reported, not measured.
  • The durable thing isn't this -rc. It's the policy that already merged in early April: Documentation/process/coding-assistants.rst, the Assisted-by: tag, and the rule that only a human signs the DCO. That's the part worth copying.

What actually happened

Linus Torvalds tagged 7.1-rc6 on Sunday, 2026-05-31. His read, per the announcement quoted across coverage: he "wouldn't call this 'small', but it is certainly smaller than rc5 was," and there's "nothing particularly scary here," so maybe the cycle is still on track for a normal release. Driver changes (GPU, networking, USB, sound, SCSI) plus core networking and selftests — a normal late-cycle shape.

So at the fact level: an -rc tagged on schedule, trending down in size, no fire. If you don't ship kernels, there is nothing here that touches your week.

The reason this -rc got written up at all is the size story running through the whole 7.1 cycle. Phoronix titled the rc6 piece around a "Larger-Than-I'd-Wish-For Size" week, and ran earlier pieces on networking pull requests being "significantly bigger" than normal. That follows last week's rc5 thread, where Torvalds said the fifth candidate had grown larger than usual, a lot of it tiny driver fixes, and that he'd get stricter about pulls that aren't real regressions, naming AI code review as one source of the extra series.

Separate the three claims

This story is easy to over-read, so split it:

  1. rc6 tagged 2026-05-31, smaller than rc5. Fact. From the tag and the announcement.
  2. The cycle's pull requests, networking especially, ran bigger than normal. Reported as Torvalds's own characterization ("larger than usual," "not small"). I haven't seen a hard diffstat baseline in the coverage; it reads as asserted by the people doing the merging, not a measured delta against prior cycles. Believe it, but hold it as a maintainer's impression, not a metric.
  3. AI/LLM coding agents (Claude Code, ChatGPT Codex) caused the bloat. This is the part that gets stretched into "AI is breaking the kernel." What's actually on record is narrower: Torvalds tied some of the extra rc5 series to AI code review, and Phoronix has framed the bigger-PR trend around AI-assisted contribution. That's an attribution from the merge boss plus an aggregator's framing, not a measured "AI added N% churn." Label it that way.

The honest summary: late-cycle volume is up, the person doing integration is annoyed about it, and AI tooling is one named contributor to the volume. Everything past that is interpretation.

The durable payload merged in April, not today

The governance answer isn't new and didn't ship with rc6. It landed in early April 2026, with the 7.0 release: Documentation/process/coding-assistants.rst. Worth reading even if you never touch the kernel, because it's a small, concrete take on the exact problem every repo now has.

The mechanics:

  • AI-assisted contributions are allowed, and disclosed with an Assisted-by: trailer. Format: Assisted-by: AGENT_NAME:MODEL_VERSION [TOOL1] [TOOL2] — e.g. Assisted-by: Claude:claude-3-opus coccinelle sparse.
  • AI agents must not add Signed-off-by. Only a human can certify the Developer Certificate of Origin.
  • The human submitter owns the patch: reviews it, checks licensing, signs off, carries responsibility. A tool can write the diff; it can't be the accountable party.

Why a new tag and not Co-developed-by: Co-developed-by requires a paired Signed-off-by, and a model can't legally sign the DCO. So they coined a trailer that records the assist without faking accountability. That's the whole design: keep the chain of "a named human stands behind this" intact while letting the tooling in the door.

What a kernel-adjacent engineer takes from this

Two things, neither of them "AI bad."

The freeze-risk lesson is old and tool-agnostic. Volume late in a cycle is its own hazard, every "tiny" patch is nonzero risk, and the bottleneck near a freeze is who gets to say no. AI tooling raises suggestion rate faster than humans raise merge-judgment capacity, so the pressure shows up at the maintainer, exactly where 7.1 is feeling it. If one person merges many parallel "small" late branches in your repo, that's your bell too.

The Assisted-by: / DCO split is the copyable part. If you're writing an AI-contribution policy, the kernel already did the hard thinking: disclose the assist, never let the tool be the signer, pin accountability on a named human. You don't need the panic to take the pattern.

rc7 is expected next week. The interesting artifact already shipped six weeks ago.

Sources