Herki Parn

Composer 2.5 at the Try It Out event. It was the push I needed, and honestly I was surprised what I got done. Built a before/after upgrade to my engine's DevTools in a few hours. First time with OBS and editing too. The engine was the easier part. Quick context for anyone not watching the video. WebShell is the dev surface around my own Rust web engine. It does its own HTML and CSS rendering, but it is built for game and native hosts, not desktop browsing. Not Chromium, not Electron, not WebView. The rendering, layout and DOM all come from the engine itself. In a few hours with Composer 2.5 the DevTools went from crowded to something I would actually want to debug in: rebuilt onto Lit, a unified Elements inspector with a DOM tree and per-element Styles/Computed/Layout/Properties, an element picker that runs through the engine's own hit-test path, per-tab state for the inspector and console, and a cleaned-up console. Some of it is still partial and I am fine saying that. On Composer 2.5: comparable to Opus 4.7 for my work, fast, and it understood a fairly odd codebase that usually trips models up. The cost for what I got out of it is the part that stuck. I will be using it more. Good to see what everyone else was building too, some interesting demos in there. Thanks @cursor_ai and the hosts for running it: @KellehEyad @Khalidabd3laty @rasaljaya @tibor_tee @frankterpo @ftnabeelah

@benln What about London?

Cafe Cursor is coming up in these cities: • Florianópolis - 5/27 • Skopje - 5/28 • Tirana - 5/30 • Austin - 5/30 • Johannesburg - 6/5 • Tashkent - 6/6 • San Francisco - 6/10 • Manila - 6/16 • León - 6/17 • Dubai - 6/18 • Sofia - 6/26 • Cebu - 6/27

@evadne Same here today. Opus keeps editing files I never pointed it at. It drops the running discussion and starts making things up. Not a context issue at all. Happening in fresh sessions too, even under 50k.

Opus 4.7 is again, quite broken today, and has been for 3–5 days. Claude is acting like he got quantised to int8, forgetting instructions, using the wrong SSH user, and jumping to conclusions — under a maximum of 250k context (not 1m). This is a product stability issue

@orcdev Will wait to hear what you managed. Still testing more on my end, the DevTools rebuild went smoother than I thought and now I am moving onto the rendering layer.

ok I'm back to Cursor let's see what Composer 2.5 can really do

@KellehEyad Rebuilt the DevTools on my own Rust web engine in Cursor with Composer 2.5. Before and after, in a few hours. This week: pushing the model to its limits on the rendering layer. Display list as the output, Skia replays it on host GPU.

While we're reviewing your submissions. Head over to tryitout Find your country on the globe and show us where you're building from. What did you build? Drop it below.

@Andy_AJT Either way, I'll be there.

more notice for an event or to take advantage of the weather this week... that is the question...??? More notice, 28 degrees, more notice, sun....

Same here! Want more game devs on my timeline too. Curious what others are building. Right now I'm wiring scouting intel into Endgate, a smaller slice of my survival strategy game Valenar. Both run on a modified C# scripting language I'm building so they're moddable from day one. On the side I'm also building a Rust web UI engine for game hosts, Unity first. Here's what they look like:

I need more gamedevs on my timeline

@amix3k Composer for UI work yesterday and Codex on perf loops this week, so the stack keeps shifting. A clear winner would be nice right now, but then competition dies, prices move, and who knows what we can still access.

It’d be great to have one clear winner in AI coding. Constantly switching between models, environments, and harnesses is exhausting, and the grass is always greener on the other side! 😅

Quick note on Cursor if you're going to try it: Be careful with sub-agents. Even when you set the parent model to Composer 2.5, the harness can still pull in other models (like Opus) depending on what's enabled in your settings. I had to manually disable models to stop it from happening.

Try it and make your own call. Used it at a hackathon yesterday. I gave it mockups and images for DevTools UX changes on my Rust engine, and it implemented them. Even fixed a couple of bugs in the engine itself. Didn't have to fight it. The Cursor harness helps, but the model is doing the real work. Comparable to Opus 4.7 for my codebase. Still early to make sweeping claims. Same boat on the subscription pile.

is Composer 2.5 really that good? 😅 currently paying for: $100 Claude $200 Codex (open source program thank goodness ⚔️) $300 Grok Build (testing) and now I have to give Cursor another $20 too

No team. Codex is deep in a goal loop running perf optimization on my Rust web engine. Claude handles web design and frontend work. Composer has been strong for upgrading WebShell UX. Grok is the best at understanding images and video, but loses the thread when it needs to change or process them. So I use Grok to understand, Codex to execute the changes. I pick whatever ships the best result. This is just the last few days. Stack shifts every other week.

Are you team Claude or Codex?

@shafu0x Yesterday: shipped unified DevTools for my own Rust web engine. This week: display-list renderer for performance. Easier to debug now that the DevTools are unified. Not Chromium, not Electron, not WebView. End goal: drop into Unity, Unreal, or Godot for real HTML/CSS UI.

shill me what you are building

@RoundtableSpace This week: building Endgate. Tower defense on top of my own engine, SECS. Deliberately narrow, so the authority and transport layer get exercised in isolation. Currently wiring unit controls and selection.

It’s Monday, what are you building this week?

This is nice. Now do gym memberships. Those are built so you can't cancel without a phone call and a guilt trip, and half of them survive on people who gave up trying. Get computer use through that one and you can drop the gym you never go to and put the money on something you'd actually use, like Codex. 😁

Using computer use, you can ask codex to cancel subscriptions you don't need anymore. Very pleasant to watch. No particular one in mind, works on all of them. chatgpt.com/codex/

This is the bit I keep running into on my Rust web engine. The useful move isn't asking the model to port old code. It's turning old behavior into tests first: WPT cases, fixtures, perf gates, anything that makes the rewrite fail when it drifts. Then the model can help with the port. Without that I'm just reviewing cleaner-looking code by hand.

my main takeaway from the Bun rewrite is that legacy codebases will be incredibly valuable as a source for "distilling" code into new forms every game should be crossplatform, all legacy software should work on the web, we don't need COBOL anymore

My project is a desktop app built around my own web engine. The repo is private because it is a commercial project, and I do not have a live demo URL yet since it is not a hosted web app. Is a public GitHub repo + live demo URL mandatory to join, or is there flexibility for private desktop/native projects? I can share a short demo video if that works. Thanks.

@cursor_ai We're doing it @rasaljaya,@Khalidabd3laty,@Shriabhay1,@ojschwa, @ethan_leee9113 Online hackathon building only with Composer 2.5. Winners get Cursor credits. Join us live today at 4:00 PM CET. tryitout.io

Spent the whole night in Composer 2.Spent the whole night in Composer 2.5, just waking up now and honestly, hard to go back. The pace this thing moves at is something else. Same outputs as Codex and Opus, but way faster. Someone explain @cursor_ai ?

Yeah, I used "own" too loosely there. I do not mean fork the CLI. API is probably the right way to do it. What I mean is the harness owning goal state around the CLI: objective, pending work, changed files, checks, pass/fail, next step. The colony sim is just one workload I use to stress that loop. It is not feeding prompts back into it.

@herkiparn @boyuan__zheng so are you kind of an "enhancing loop" between prompts by auto feeding from the colony sim stress test? btw you can test the cli via api, there is no need to "own" ;)

One friction I keep hitting: the agent treats small edits as the end state. It makes one change, then I have to keep prompting "continue" to get back to the actual goal. A persistent goal system would help a lot here. Keep the high-level objective alive in the harness, even if the model cannot run forever yet. Then each edit is one step in the plan, not the finish line. Would make long-horizon agentic coding much more useful.

@GoogleAIStudio WebShell AppShell work for my Rust web engine. DevTools panels are turning into test surfaces now. Rust-rendered page, WebShell DevTools, evidence rows, export/report surfaces. Still debug-heavy. Right now I care more about evidence I can replay.

What are you vibe coding this weekend?

Tested on 0.1.218-alpha.1. This looks much better. I had an explore subagent read 16 files and return an aggregate: 18 tool calls, 1 turn. Parent only used spawn_subagent + get_command_or_subagent_output. Direct CLI resume of the child also worked. After /compact, the child reloaded the local skill and reran the chain. Only note: stable Linux updater still shows 0.1.217 for me.

> Question: does this apply across spawn_subagent / resume_from too, or only the main session after compaction? will double check but it does apply to subagents as well, in grok build world subagents are the same as main agent with the only gotcha that the main agent can interact with them and user can't directly send a prompt to them

Bug fixes shipping to Grok Build (release notes will be available in the TUI) 0.1.218 • Ctrl+X as default shortcuts help binding on Windows • Fix image pasting + shortcuts keybinding on Linux • User-specified duration for video generation tool • Ctrl+X as default shortcuts help binding on Windows • Support temporary screenshot images on mac • Validate image bytes to prevent retries • Improve compaction prompt (match training and rehydrate skills) • Increase ulimit on mac and linux, preventing ENOSPC errors from appearing and bricking the CLI • Multi-line image links are not clickable (and don't break anymore) We are focussing more on Linux and Windows so engineers on these platforms have a great experience.

@skcd42 That helps. The direct-to-subagent bit matches what I ran into. The parent had to keep repackaging tool results back through resume_from, which made larger orchestrator waves painful. I’ll rerun the same case on 0.1.218 and report back if it still breaks there.