Repository navigation
Roadmap: v2.14.0 — and the one question that would change it #912
divshekhar
announced in
Announcements
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Nearly every issue in the v2.14.0 milestone was filed by an agent, not a human.
They come in through
reticle_feedback— an agent drives a real app, hits something Reticle gets wrong, and files it with the trace attached. The titles are theirs. That is why they read the way they do: not "improve error handling" but "A crashed destination and a found-but-forbidden element both return unknown, where the evidence supported a no."So this roadmap is less a plan than a readout. Here is what the agents are telling us, grouped, and then the one question that would change it most.
1.
inithas to prove it worked, not report that it ranThe largest cluster, and the most embarrassing one. An install that says green and isn't costs someone their entire first session.
initnever checked that the plugin actually landed invite.config. It didn't, and nothing said so.initforever, silently. (The stated cause didn't reproduce; the silence is being fixed regardless.)initstarts a second dev server beside the running one, leaving three sessions and no way to pick.connect()never runs, the install reports green, the daemon waits forever.The shape of the fix: every step of
initends in an observation, not an assertion about what it wrote. The install gate exists precisely because nothing else in the repo can see a first install — every app inapps/is already instrumented, so re-runninginitover one proves nothing.2. "unknown" is usually a "no" we already had the evidence for
The verdict is the product. A verdict that says unknown when the evidence supported no is a verdict that trains agents to stop reading verdicts.
unknown.already_true, proving nothing.nodes: 0, which reads exactly like a page that hasn't rendered yet.sessionId.3. The network channel drops the bytes that explain the failure
bodyContainsassertions permanently undecidable, and nothing says where the cap lives.DOCUMENTinitiator, with no synthesised status, because a guessed 200 is worse than an honest silence.)4. What Reticle simply cannot see
The category that matters most for where this goes next, because each one is a place an agent had to give up and reach for a screenshot.
WheelEventbursts against aquerySelector'd canvas just to move a camera. Once that escape hatch is open, agents start using it for things we do cover.observation_lost, so an OAuth handoff can't be verified even as a handoff.reticle_viewportneeds CDP, so mobile-only UI is unreachable on an SDK-only install.5. And the meta-fix
#891 — a corpus of deliberately awkward apps in CI. Every blocking defect in the last field export would have been caught by one. Five of the six critical fixes in #902 were reproduced from a reporter's trace rather than from a running app with the property that broke — there was no >20s endpoint, no file-upload flow, no WebGL subtree here to test against. The unit tests pin the logic; the end-to-end proof that the symptom is gone does not exist. That gap is the argument for the corpus, seen from the fixing side.
Want to help?
There are 19 issues labelled
good first issuein this milestone right now, most of them a contained fix plus a regression test. Two things that will save you time:#NNNfinds nothing while a green PR for the same bug is already sitting there.Thanks to everyone already shipping against this milestone — there is a real queue of external PRs open right now, and several of them are fixes I would not have prioritised correctly on my own.
The question
Every item in section 4 arrived the same way: an agent gave up on Reticle and reached for something else. That moment is the highest-signal thing anyone can tell us, and it is invisible from inside this repo — every fixture in
apps/is one we wrote, so it is already shaped like something we know how to see.So: what did your agent have to reach outside Reticle to do?
A screenshot because the snapshot was empty. Raw JavaScript because the element wasn't addressable. A
sleepbecause nothing would settle. A verdict you didn't believe, so you checked by hand.Reply with the app shape and what broke — a framework, a widget, a rendering strategy, a login flow. You don't need a repro; the symptom alone is enough to be useful, and the diagnosis is genuinely optional. Four reports this week named a cause that pointed at correct code, and every one of them was still worth acting on, because the symptom was precise.
If your agent is already connected,
reticle_feedbackfiles it straight into the queue with the trace attached, and it will show up here labelledagent-reported, like most of the list above.All reactions