linux-kernel - Re: [PATCH v2 2/2] agents: add unified agent coding assistant configuration

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite for Android: free password hash cracker in your pocket

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <aJirFxwuzxyau1j-@lappy>
Date: Sun, 10 Aug 2025 10:22:15 -0400
From: Sasha Levin <sashal@...nel.org>
To: Greg KH <gregkh@...uxfoundation.org>
Cc: corbet@....net, josh@...htriplett.org, kees@...nel.org,
	konstantin@...uxfoundation.org, linux-doc@...r.kernel.org,
	linux-kernel@...r.kernel.org, rostedt@...dmis.org,
	workflows@...r.kernel.org
Subject: Re: [PATCH v2 2/2] agents: add unified agent coding assistant
 configuration

On Sun, Aug 10, 2025 at 09:01:00AM +0200, Greg KH wrote:
>On Sat, Aug 09, 2025 at 07:40:08PM -0400, Sasha Levin wrote:
>> Create a single source of truth for agent instructions with
>> symlinks for all major coding agents:
>> - CLAUDE.md (Claude Code)
>> - .github/copilot-instructions.md (GitHub Copilot)
>> - .cursorrules (Cursor)
>> - .codeium/instructions.md (Codeium)
>> - .continue/context.md (Continue)
>> - .windsurfrules (Windsurf)
>> - GEMINI.md (Gemini)
>
>Having these non-dot-files in the root directory for just specific
>closed source tools feels a bit wrong to me.  I thought all of these
>tools were standardizing on "use README" and didn't need these extra
>helpers?  The fact that most of them just contain README seems like
>that's the case, they just need to work it out somehow?
>
>Anyway, I'd not like this patch to be added (patch 1/2 is great), as it
>feels like these agents need to get their act together in ways to not
>need these, we shouldn't be forced to deal with their inefficiencies.

No, agents will not read the README unless we explicitly tell them to
"study" the codebase (either via something like "/init" in the CLI tools
or a "Scan the codebase" button on the VSCode clones).

This is why the agent files only have a reference to README: their only
purpose is to get the agents to read the goddamn README file by default.

We don't strictly need this patch. Here are the options I came up with:

1. Multiple agent files: agents automatically read our rules, higher
chance of comlpiance, doesn't require user to do anything manually.

2. No agents files at all: this means that the user will need to
explicitly trigger an action, otherwise all the tools I've tested it
with will happily ignore the README file even when prompted to do other
actions on the codebase (analyze code/write a patch/etc).

3. Only "AGENTS.md" which seems like a few of the tools are trying to
standardize on. We could also use it in the future to add prompt cues
that might not be relevant within the README.

We could also start with most of these files, along with a statement
that they will be removed within 12 months to give the various tools
time to standardize around something.

-- 
Thanks,
Sasha