> But mostly I was assuming that the agents would mostly follow the skill and existing docs, so we don’t _need_ to do anything else for humans to fill it in, as by default they will do that anyway. Actually, I think this change is the first time we are adding that requirement. We never asked agents to stop and "expect" the PR description to be written by hand. We wrote that we expect humans to understand their changes and be responsible for it. But we never expected **agents** to stop and ask the operating human to write the description by hand.
Do you think it was written somewhere Ash? (I could not find it). And maybe its' the right time to **quickly** refactor our agents.md and others without doing the complete work that our intern is doing. One problem with those instructions for agents is that they don't read the full "long" context initially and might "miss" some instructions. I am recently doing optimizations in Magpie - and making sure that all the SKILLS and tools are written in the way that they follow some budgets, and limitations - I can definitely use the magpie skills to optimize our agentic instructions. So Aiurf Jl On Thu, Sep 24, 2026 at 4:53 PM Ash Berlin-Taylor <[email protected]> wrote: > Yes, the html comment would fork fine too. > > But mostly I was assuming that the agents would mostly follow the skill > and existing docs, so we don’t _need_ to do anything else for humans to > fill it in, as by default they will do that anyway. > > > On 24 Sep 2026, at 15:44, Shahar Epstein <[email protected]> wrote: > > > > Ash - I understand your approach, and not against the formulation that > you > > suggested (without headers + changing "Click here" to "AI Summary"). > > However, I think that it will be easier for contributors to comprehend > that > > they should write the description in their own words if there's an > explicit > > instruction for doing that. If you're against headers, it could also be > in > > the form of an HTML comment like we already do in other areas (i.e., <!-- > > YOU MUST WRITE THIS SECTION WITHOUT LLM -->). > > Personally, I think that headers will have a better effect as they are > more > > visible (for both author and reviewers altogether). > > > > > > > > Shahar > > > > On Thu, Sep 24, 2026 at 5:34 PM Jarek Potiuk <[email protected]> wrote: > > > >> Hello everyone, > >> > >> Following the objection in the [LAZY CONSENSUS] thread, I am turning > >> this into a discussion. The change (#73660) has already been reverted > >> in #73669 (thanks Ash) - and apologies for merging it before the lazy > >> consensus ended; I thought it was uncontroversial, which it clearly > >> is not. > >> > >> The original proposal, for PR descriptions drafted by AI agents: > >> > >> * The description starts with a short "Human Summary" - a few > >> sentences typed by the human author, in their own words: why the > >> change is needed, key decisions, caveats. > >> * The agent-written description stays below it, in an "AI Summary" > >> section folded by default. > >> * Agents are instructed to ask the author for the Human Summary and > >> paste it verbatim - never to write, suggest or paraphrase it. > >> > >> Ash's objection, as I understand it: the PR description should already > >> be written for humans, and it is how the author engages with the > >> community - so an explicitly labelled "Human Summary" reverses the > >> priority. His counter-proposal is to drop both the "Human Summary" > >> and "AI Summary" headers and put the agent text in a folded > >> <details><summary>AI Summary</summary> block. > >> > >> I see it the other way round. Following the ASF Generative Tooling > >> guidance, we already require every such PR to state that it was > >> generated with AI - the reader knows up front that an agent was > >> involved. Given that, an explicit "Human Summary" is exactly the > >> signal that matters: it shows that the human was driving the change, > >> and that they understand it well enough to summarize - in their own > >> words - what the PR does and why. Without the label, the reader of an > >> AI-disclosed PR has no way to tell which part the author actually > >> wrote. This is the "centaur" - human in control - not the "reverse > >> centaur" or "meat proxy" [1]. > >> > >> Looking at it again, the two proposals are closer than it seems: both > >> put the author's own words on top and fold the agent's text below. > >> The actual differences are: > >> > >> 1. Should the human-written part be explicitly labelled > >> ("Human Summary") so it is distinguishable in an AI-disclosed PR, > >> or just be the unlabelled top of the description? > >> 2. Should the folded agent text be marked "AI Summary" (Ash suggests > >> the <details> summary label - that part we seem to agree on)? > >> 3. Should agents be required to ask for the human part and refuse to > >> draft it themselves - or should that be left to the author? > >> > >> Just to remind - this is **only** for PRs that are agenically generated. > >> If someone creates their PRs manually, without the AI agent, none of > >> this applies. > >> > >> The template for PRs opened via GH UI is not supposed to have either > >> section. > >> > >> I'd love to hear what others think - especially those who review a lot > >> of agent-assisted PRs, and Shahar, who has been using this format > >> already. > >> > >> [1] > >> > https://doctorow.medium.com/https-pluralistic-net-2025-09-11-vulgar-thatcherism-there-is-an-alternative-f1428b42a8fd > >> > >> J. > >> > >> --------------------------------------------------------------------- > >> To unsubscribe, e-mail: [email protected] > >> For additional commands, e-mail: [email protected] > >> > >> > > > --------------------------------------------------------------------- > To unsubscribe, e-mail: [email protected] > For additional commands, e-mail: [email protected] > >
