BTW.

The dashboard (link:
https://htmlpreview.github.io/?https://gist.githubusercontent.com/potiuk/ef446baff0fa2972951b5eb5a3826dcf/raw/pr-dashboard-v3.html)
is a first draft (even if it is already pretty good, usable and actionable)
created literally this morning, from the scratch - through less than hour
of agent iteration.

I have submitted PRs to Airflow (
https://github.com/apache/airflow/pull/66464) and airflow-steward (
https://github.com/apache/airflow-steward/pull/68). I am currently syncing
these but will soon switch Airflow to use steward skills exclusively.

I also built the capacity for projects to keep their local modifications by
default, with the option to use a separate agent-driven skill to
automatically upstream generalized features back to steward.

We are truly living in a new era of software development.

Thanks,
Jarek Potiuk

On Wed, May 6, 2026 at 12:02 PM Jarek Potiuk <[email protected]> wrote:

> Hi all,
>
> Two related things this morning, both as a starting point for discussion:
>
>
> 1. NEXT ROUND OF TRIAGE
> =======================
>
> I ran another full pass of the auto-triage skill against the open-PR
> queue. The skill closed the obvious cases on its own (stale-triaged
> drafts past the 7-day no-response window, a few rebase-able PRs whose
> "failing CI" was actually stale-main pollution, etc.) and surfaced
> contributor PRs into the "ready for maintainer review" label so we
> don't have to keep re-discovering them.
>
> The big picture after this round:
>
>   - 459 open PRs (non-bot), 342 from contributors, 187 currently
>     labelled "ready for maintainer review" (54% of the contributor
>     queue).
>   - The triage funnel is doing its job: only ~98 contributor non-draft
>     PRs are still untriaged, and only 1 is older than 4 weeks.
>   - The rest of the open queue is mostly drafts (where the author is
>     in control) and the growing review backlog.
>
> What changed since last round: the queue of *bad* PRs (mergeable
> conflicts not addressed, abandoned drafts, missing CI signals) is
> mostly gone — those are now either auto-drafted, auto-closed via
> stale-sweep, or rebased. What is *growing* is the backlog of PRs
> that are ready and waiting for a human maintainer to actually
> review the code.
>
>
> 2. PROTOTYPE TRIAGE DASHBOARD
> =============================
>
> To make the above legible at a glance, I asked my agent this morning
> to build a maintainer dashboard on top of the same data the triage
> skill already fetches. It took about an hour, end to end. This is
> explicitly a first attempt — not a finished product.
>
>   Live (snapshot, dated):
>
> https://htmlpreview.github.io/?https://gist.githubusercontent.com/potiuk/ef446baff0fa2972951b5eb5a3826dcf/raw/pr-dashboard-v3.html
>
>   Source HTML (single self-contained page, no JS):
>   https://gist.github.com/potiuk/ef446baff0fa2972951b5eb5a3826dcf
>
> What's currently on it:
>
>   - A health rating + four hero cards (queue size, ready count,
>     untriaged-non-drafts).
>   - A "What needs attention" panel — deterministic recommendations
>     derived from rules in the skill spec, each with the exact
>     slash-command to run for the action.
>   - Closure velocity for the last 6 weeks (per-week stacked bars).
>   - Opened-vs-closed momentum (line chart) — answers "is the backlog
>     growing or shrinking" (last 6 weeks: -56 PRs, shrinking).
>   - Ready-for-review trend by top-pressure areas (line chart, one line
>     per area) — surfaces where the review backlog is climbing fastest.
>   - Closed-by-triage-reason per week (stacked bars) — green = merged,
>     amber = author-engaged-then-closed, red = stale-sweep close,
>     grey = closed without ever being triaged.
>   - Pressure ranking by area (top 8) so the maintainer can pick a
>     focused 30-minute session.
>   - Triage-funnel coverage / response-rate metrics.
>   - Original per-area tables collapsed into "Detailed tables" at the
>     bottom for anyone who wants the raw numbers.
>
> Opening discussion: what *else* should be on this dashboard so we
> can use it day to day? What's missing, what's noise, what would you
> look at first thing in the morning to know what to do?
>
>
> 3. WHAT MAINTAINERS NEED TO DO NOW
> ==================================
>
> The dashboard already encodes my proposed answer for "what to do",
> but in plain text:
>
>   - Look at the "What needs attention" panel. Each card is a concrete
>     slash-command (or pointer to one) that you can paste.
>   - For most maintainers right now, that's going to be
>     "/maintainer-review" filtered by an area:* label you care about
>     — which is also what the "Pressure by area" panel is sorted to
>     surface.
>   - The triage skill has handled the bottom of the queue. The top of
>     the queue (the 187 ready-for-review PRs) needs us, the humans.
>
> If each of us picks an area we care about and runs a maintainer-review
> session against just that area's "ready for maintainer review" PRs,
> we move the queue down measurably and the dashboard reflects it on
> the next snapshot.
>
>
> 4. FEEDBACK + PRS WELCOME
> =========================
>
> The dashboard skill is in the same place as the triage skill — both
> are in the apache/airflow repo under .github/skills/, with mirrored
> copies in apache/airflow-steward. PRs against either are welcome.
>
> There's also a Slack channel for auto-triage feedback — the dashboard
> recommendations and triage decisions are deterministic rules, so if
> something feels wrong, it's a rule worth changing.
>
> And — riffing on the "Intention is all you need" framing
> ( https://potiuk.com/intention-is-all-you-need-1379ec41a19c ) —
> nothing about either the dashboard or the triage process is fixed
> in stone. Both are agent-readable skills, which means the modification
> loop is now: notice what you'd like different → describe the
> intention to your agent → it edits the skill → propose the change
> back as a PR. You don't need to file a feature request, wait for me
> to find time, or write the rendering code yourself. If you'd like
> the dashboard to show "PRs that *I* requested review on", or
> "per-author response rate", or "Helm-chart age distribution" — tell
> your agent and it will do it.
>
> That applies to the triage process itself too. If a triage rule is
> producing comments you don't like, or the wrong PRs are being
> auto-drafted, the rules live in
> .github/skills/pr-triage/classify-and-act.md and are about as
> agent-modifiable as anything in this repo.
>
> Looking forward to feedback — replies on this thread, Slack, or PR.
>
> Cheers,
> J.
>

Reply via email to