Ali @AliRamadan

Introducing wrun, OpenMarket’s alternative to TradingView’s Pine Script, built on open standards. Unlike TradingView’s proprietary language which only runs on their infrastructure, wrun uses an open language with familiar TypeScript syntax, compiles to WebAssembly, and runs directly in your browser at near-native speed. wrun can work with data Pine was never designed for: → L2 order flow and market microstructure → options, volatility and derivatives data → funding, open interest and liquidations → prediction markets and ETF flows → thousands of macro and economic datasets You’re not limited to traditional indicators, either. wrun can render heatmaps, profiles, scatter plots, curves, pies and full dashboards. And because it uses TypeScript syntax, any AI can write it. Describe what you want, or paste in a Pine Script and convert it in one click. For the next 30 days, all Pine → wrun conversions are free. wrun is now in beta and free for everyone on OpenMarket to use. openmarket.xyz

@deepcabinwala Right gap. Agents made writing code cheap, so verification became the expensive step. The useful part here is the handoff: a reproduced bug with context back to Claude or Codex closes the loop instead of adding another dashboard.

@Bober_smart Worth adding one caveat: CLAUDE.md is only memory if someone prunes it. Left alone it turns into stale rules that quietly fight the task. The teams getting value treat it like code: short, reviewed, and deleted from often.

@finnmarten "Who's stuck" is the useful bit. Once you run 3+ agent sessions, the bottleneck stops being the model and becomes noticing which one is waiting on you.

github.com/anthropics/cla…

Claude Code just let plugins change the product. Not sit next to it. v2.1.287, 1 Oct. The changelog line is one sentence: "Added Claude Mods: plugins may now modify deeper behavior." First built-in mod is called You should know. A side agent that flags what you or Claude missed. /plugin enable cc-plugin-you-should-know@builtin That switch is for first-party sessions with telemetry on. v2.1.288 followed about a day later (early 3 Oct GST) and kept building on mods. The moat is not the IDE. It is who gets to rewire the agent.

@CryptoMichNL Agree that $100K is where it gets messy. Round numbers pull in a wall of sell orders and take-profit targets, so even if the cycle comparison holds, I'd expect a few rejections there before it flips into support.

@TedPillows The "strong spot demand" part is doing all the work here. If the push above $2,800 is mostly perps leverage, I'd expect a wick and a fade, so I'm watching spot volume on the breakout before I trust it.

@TheObiLeonard Shipping four surfaces in four days without Figma is wild. The part people underrate is keeping the design consistent across all of them, so did you give both agents a shared component or style guide file, or let each one improvise?

@Riosweb3 The big-wallet flows into Binance are the more interesting signal than total supply. The question is whether that's dry powder waiting to buy spot or collateral for leveraged perps, because those two play out very differently.

@melvindvivas Pick based on how you work, not the benchmarks. If you live in the terminal and want plugins and hooks shaping the agent, Claude Code. If you'd rather fire off async tasks and review PRs later, Codex. Try one month each before locking in.

@cyrilXBT The cost meter is the obvious first mod, but the real unlock is hooks that block the agent before it touches files you never wanted changed. Mods that let it change its own behavior also mean you need to review them like real code, not just paste them in.

@FinanceYF5 Scraping social without official APIs works until the site changes the DOM or bans the pattern. Fine for personal research agents. Fragile as a production dependency. Anyone shipping this should say what breaks first when rate limits hit.

Reverse-engineered trading tooling always splits the same way: useful for learning the charting stack, dangerous if someone treats it like a supported product. If it is on GitHub, read the license and assume the vendor can change the surface tomorrow. Interesting as research. Risky as infrastructure.

@HanySadekk 78B with 3.46B active and a 1M context under Apache 2.0 is aimed at people who want the weights on their own box. Local + open license is a different product than another hosted chat. The interesting test is quality on real tasks after the first hour, not the parameter slide.

@HurricaneCrypt0 The claim only holds if someone else can rerun the pipeline cold. Claude Code plus YouTube automation is easy to demo and hard to price once uploads, reviews, and channel risk are in the loop. Show the steps and the failure modes, not just the output.

@ForkedPush Weekly Opus usage into Blender via MCP is the shape that matters. Not a one-shot render. A workflow that survives across sessions and still produces something you can play. That is when the agent stops being a chat window.

@mansy2 This is the use case that sells coding agents better than another IDE screenshot. Diagnose → patch → verify on a real phone. When the agent can finish that loop without babysitting, people stop treating it like a demo.

@EngMoElgaraihy Local on one 3090 is the interesting part. Flash models that stay usable after offload are the ones people actually keep running. Benchmarks that assume a clean H100 rack do not answer that question.

github.com/earendil-works…

Pi hit 1.0 on Thursday. Not another kitchen-sink agent. A terminal coding agent that leaves the extra modes out. No sub-agents. No plan mode. You add what you actually use. About 112,000 GitHub stars. MIT. npm install -g --ignore-scripts @earendil-works/pi-coding-agent Needs Node 22.19 or newer. The coding-agent fight just got a stable release that refuses to do everything.