Topics covered

• what problem are we trying to fix
• b4 review: triage, reply, apply, test, in one place
• the maintainer container
  • local-first mirror of all you need
• korgalore: the alternative to subscribing to lists
• b4 bugs: local-first bug tracking using git-bug

About me

• Keeper of kernel.org for 10+ years
• I try to make infrastructure useful for maintainers
• Caveat: I am not a kernel maintainer myself
• That said:
  • I hear maintainers complain
  • I can drive infra changes
• Based in Montréal, Québec

Let's pretend to be Jon Corbet

Start the container

localhost:11043

• 11043 = LLORE
  • "Local Lore"
  • Not that 👉
• local copy of a subset of:
  • git.kernel.org
  • lore.kernel.org
• offers services:
  • git pull/clone
  • b4 interactions
  • nntp/imap
  • mcp
• We'll see it in action shortly

Reviewing series today

The (manual) labour of love

• Look at the series in your MUA/patchwork, one at a time
• Did anyone comment on this?
• What does CI say?
• What did the bots say?
• Let me shazam it
• Wait, wrong branch
• Did they make the changes I asked for?
• Oh really, what does range-diff say?
• This patch still looks wrong
• Let me find it in the mail client
• Dammit, it wrapped the lines wrong
• Delete the series, but don't forget about it

b4 review streamlines it

a TUI with a shallow learning curve

Textual TUI

• Advanced and responsive lib
• Well-tested and robust
• Shipped by default on many distros
• Good support for keybindings
• Familiar-ish interface

All under one roof

• CI results
• Email client
• Improved code review
• Quick actions
  • Send a review trailer
  • Send a quick NACK
  • Apply-and-thank
  • Check for newer revisions
  • Park/snooze a series
• Still in beta and changes rapidly

Live demo

• The maintainer-container should be set up by now
  • Continuously mirroring with live data
• It runs an MCP server that can let us query mail traffic
• We'll be using Claude Sonnet (for speed)
  • Local models work well, too (tried Qwen 3.5 4B)
• My local repo is preconfigured to pretend to be Jon
• MCP and b4 talk to localhost
• There are some old series already tracked

"P" stands for "Prompt Injection"

"S" in "MCP" stands for "Security"

We try to hedge against that

• We return structured json data with no bodies
  • subjects
  • trailers
  • follow-up authors
  • message-ids
• The agent is instructed to grab full mailboxes externally
  • helps keep context smaller anyway
• However, the attack surface is very large
  • LLMs emulate reasoning
  • with enough effort, the reasoning can be poisoned

Hopefully, it worked!

• b4 keeps track of prior revisions
  • the agent can help figure out if your points were addressed
• Helps better manage the firehose
  • park a series till later
  • mark as waiting on a new revision
  • quickly send off trailers
  • apply-and-thank as one action
• New features coming soon (see plan.otl)
  • workbook support to bundle series
  • arrange the order how you want it
  • review within the context of the whole bundle

lei/korgalore behind the scenes

• The container continuously queries lore.kernel.org
• using korgalore as the driving tool
  • delivers to a local "dummy" target
  • understands F:/N: entries in MAINTAINERS
  • monitors patches for changes you care about
• You can run korgalore on your workstation
  • delivers mail to your inbox
  • sidesteps Gmail's poor performance
  • can deliver to IMAP/JMAP/maildir
  • no need to subscribe any more

Honesty break

What you'll like

• Doesn't force your hand
• No "flag day" switch
• No central infra dependency
• Use agents if you like
• Don't use agents if you don't want to
• Completely ignore the whole thing

What it doesn't fix

• Burnout is not a workflow problem
• Burnout is linked to decision density
  • ​Do I take the series?
  • Do I nitpick the thing that I kinda don't like?
  • Will this person even send a v2?
  • Or was it a drive-by contribution?
  • Will it break something important?
• This workflow should help you be more efficient
• But it will actually increase the decision density

LLM-contributed series

• LLMs make incoming series look reasonably good
  • (That's literally what LLMs do)
  • It's now hard to identify and dismiss junk
  • LLM-assisted patches receive good LLM-reviewed scores
    • code is well-formatted
    • commit messages are verbose and well-reasoned
    • tests are abundant
    • and yet it can still be junk
• Maintainers' cognitive overload is not a tooling problem

Still just git and mail

• You can still party like it's 1999!
• But please don't, we need new features in submissions
  • change-id
  • base-commit
  • dependencies
  • other utility footers
• Stays decentralized and resilient

Where to find:

• b4.docs.kernel.org
• korgalore.docs.kernel.org
• grokmirror.docs.kernel.org
• llore.docs.kernel.org