From 00b3a48900bbed866011bc6ef5738ddcfe31ca71 Mon Sep 17 00:00:00 2001 From: Philipp Schuster Date: Tue, 24 Mar 2026 18:40:42 +0100 Subject: [PATCH] docs: revisit AI/LLM policy in CONTRIBUTING.md This summarizes the latest state all relevant parties have agreed on in the last meeting. This commit was rephrased and improved with the help of Claude Code using the Sonnet 4.6 model. Signed-off-by: Philipp Schuster On-behalf-of: SAP philipp.schuster@sap.com --- CONTRIBUTING.md | 51 ++++++++++++++++++++++++++++++++++++++++++------- 1 file changed, 44 insertions(+), 7 deletions(-) diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md index 1518d0f1f..ead1cbe20 100644 --- a/CONTRIBUTING.md +++ b/CONTRIBUTING.md @@ -164,12 +164,49 @@ Signed-off-by: Sebastien Boeuf Then, after the corresponding PR is merged, GitHub will automatically close that issue when parsing the [commit message](https://help.github.com/articles/closing-issues-via-commit-messages/). -## AI Generated Code +## AI/LLM Assistance & Generated Code -Our policy is to decline any contributions known to contain contents -generated or derived from using Large Language Models (LLMs). This -includes ChatGPT, Gemini, Claude, Copilot and similar tools. +We recommend **a careful and conservative approach** to LLM usage, guided by +sound engineering judgment. Please use AI/LLM-assisted tooling thoughtfully and +responsibly to ensure efficient use of limited project resources, particularly +in code review and long-term maintenance. Our primary goals are to avoid +ambiguity in license compliance and to keep contributions clear and easy to +review. -The goal is to avoid ambiguity in license compliance and optimize the -use of limited project resources, especially for code review and -maintenance. This policy can be revisited as LLMs evolve and mature. +Or in other words: please apply common sense and don't blindly accept LLM +suggestions. + +This policy can be revisited as LLMs evolve and mature. + +### Code Review + +We generally recommend doing early coarse-grained reviews using state-of-the-art +LLMs. This can help identify rough edges, copy & paste errors, and typos early +on. This reduces review cycles for human reviewers. + +Please **do not** use GitHub Copilot directly in PRs to keep discussions clean. +Instead, ask an LLM of your choice for a review. A convenient way to do this is + +- appending `.patch` to the GitHub PR URL + (e.g., `https://github.com/cloud-hypervisor/cloud-hypervisor/pull/1234.patch`) + and pasting it into the LLM of your choice, or +- using a local agent in your terminal, such as `codex` or `claude`. + +### Contributions assisted by LLMs + +All contributions **must** be submitted by a human contributor. Automated or +bot-driven PRs are not accepted. + +You are responsible for every piece of code you submit, and you must understand +both the design and the implementation details. LLMs are useful for prototyping +and generating boilerplate code. However, large or complex logic must be +authored and fully understood by the contributor - LLM output should not be +submitted without careful review and comprehension. + +Please disclose LLM use in your commit message and PR description if it +meaningfully contributed to the submitted code. Again, we recommend careful and +conservative use of LLMs, guided by common sense. + +Maintainers reserve the right to request additional clarification or decline +contributions where LLM usage raises concerns. Ultimately, acceptance of any +contribution is at the maintainers' discretion.