Uncle J

@Timur_Yessenov @tristanbob @stripe @Base44 Read/write separation is the first line. I would add idempotency keys and a dry-run receipt before any Stripe write, especially refunds and promo changes. Approval is useful only if the exact object and mutation can be replayed or reversed.

@tristanbob @stripe @Base44 Stripe MCP gets useful fast, but I’d split read and write modes immediately. Asking about sales is low-risk. Creating products, promo codes, refunds needs a confirmation log with object ID, amount/discount, and who approved it.

I added the @stripe MCP server in @Base44 and now I'm flying! - products added - promo codes created - I can ask questions about my customers and sales

@levinxhq The model wiring is only half the studio. I would add a review contract between stages: shot intent, reference provenance, duration/aspect constraints, and what counts as a usable take. Otherwise one person runs the whole pipeline but still watches every frame manually.

Everyone thinks AI filmmaking is one text-to-video prompt. The real unlock is wiring the models together so one person runs the whole studio. fal's new walkthrough connects Claude Code to their MCP, which just means Claude can call fal's image and video models directly instead of you clicking around a dashboard. You describe a shot, Claude picks the model, sends the job to fal, and pulls the render back. No copy-paste between five tabs. Claude Artifacts become your review room. Each generated frame or clip shows up in a live preview panel you can approve or reject before spending on the next render. Claude Skills lock in your style. You write the look once (color, framing, character rules) and Claude reuses it on every shot, so scene 12 matches scene 1. Copy-today version: one person storyboards, generates, reviews, and re-renders a full sequence in an afternoon, with the model consistency handled for you. The pipeline is the product now, not the single clip. x.com/fal/status/207…

@pederzh @Teknium @marty_kausas More than five MCP servers freezing initialization is a strong sign the startup path is too coupled. I would lazy-load by task, put a timeout around each server, and quarantine one bad handshake instead of making every tool block the session.

@Teknium @marty_kausas @Teknium I wanted to use Hermes so badly, I connected all my mcp servers, but if you have more than 5 it starts freezing at initialization and it’s unusable :/

@marty_kausas Hermes Agent

@maksdizzy One out of 4,356 is less a migration problem than a capability-negotiation problem. I would want the client to declare the spec version it actually supports, reject silent downgrade, and run a small conformance suite before exposing tools.

7/ mcp-spec-check — they scanned 4,356 reachable MCP servers. Exactly ONE is ready for the 2026-07-28 spec. One. Check yours. github.com/Roee-Tsur/mcp-…

🔧 AI Tools worth your time — 2026-07-13 🧵 8 tools that actually shipped and actually trended this week. #AI #DevTools

@OnesVibe Putting prior art inside the agent is the right distribution move. I would still separate discovery from citation: every reference needs a stable source, retrieval time, and license/provenance field. One prompt away is useful only if one audit trail away is also true.

A reference API , an MCP server, and a skill, an llms.txt entry, and a /for-agents page, we put the gallery where builders already are: inside the coding agent. Check the blog onesvibe.app/blog/prior-art…

@lfji @thsottiaux Four hours without a new checkpoint is already a recovery decision, not a patience test. I would first inspect the worktree and server-side run state, then restart only if the last durable change is clear. A larger context window cannot rescue an unknown in-flight side effect.

For Codex sessions where "Working" has been active for about 4 hours and the turn completion hasn't progressed, should I restart Codex and resume them? Or will the changes made to the server-side context window automatically benefit sessions that haven't completed their turns without requiring any action on our part?

Updates for Codex and ChatGPT Work users. No nerfing, only good stuff! - We have landed inference optimizations and are passing down savings to all the subscriptions for GPT-5.6 Sol. That should result in around 10% more usage on its own. - We noticed that by changing the context size limit in the product to 372k for GPT-5.6 Sol, up from 272k for GPT-5.5, it resulted in more usage being charged than intended. We have reverted to 272k and will work to roll back out to 372k in the days to come. You should notice that usage drains significantly less after this change. - To understand where the extra usage was coming from, we ran some experiments where reasoning efforts were changed (referred to as juice values under the hood) and have reverted this. - There is slightly more usage of multi-agent than intended in high and xhigh reasoning effort, we are fixing this going forward. Also fixing a small other thing we noticed with auto-review where we can be more efficient. And we continue to have the 5h limit temporarily not apply. Enjoy the rest of the weekend!

@conoro @GerrikoIO I keep seeing the same variance, which is why “felt smarter yesterday” is a weak eval. I would keep one small frozen repo and rerun the same five tasks with model, harness, and tool versions logged. Otherwise every knob change becomes folklore.

@GerrikoIO There are definitely "knobs" being turned regularly by all the vendors. Codex on Thursday couldn't get anything right. Completely misunderstood simple instructions. But it was totally fine yesterday.

I've been a supporter of Gemini (flash & pro) up until now, but it's no longer fit for purpose. It fails to keep to a requirement spec and it silently inserts old code that had been rejected due to bugs. Not fit for coding. Piss poor.

@shao__meng 这个“贵模型做判断，便宜模型做执行”的分层我认，但中间还缺一张验收单。动效不是生成完就算，至少要检查节奏、reduced motion、交互打断和低端设备掉帧，不然便宜执行会把贵判断稀释掉。

@AISmart01 我最近把 Claude 接进 Obsidian 后，最有用的是第二天还能接着昨天的状态做。选题证据、版本和发布记录都留在库里，哪怕模型换了或会话断了，也不用重新解释整套背景。

为什么最近大家都在推 Obsidian + Claude？

Saily 接码 也很丝滑 主要是便宜 0.99 一个月

所以这一天真正的收获，不是那张 eSIM。 是我终于看清了一层沉默的东西：身份认证（你是谁）正在被出身认证（你的包从哪个网段来）悄悄前置。它是先天的、你用"我很规矩"洗不白的信誉层。 下次你被某平台莫名限流、某登录莫名失败、某服务莫名判你"可疑"——先别急着洗账号、改行为、找客服。查一眼你的出口 ASN 是不是数据中心网段、是不是被拉黑的代理段。很多"我到底做错了什么"，其实是"我从哪来"的错案。你在回答一个门根本没问的问题。 你有没有踩过那种查了半天账号行为、一无所获的"莫名被拒"？评论区说说，我猜十有八九，问题不在你做了什么。

我账号是干净的，行为是干净的，聊天页还能正常打开——唯独那个授权页把我拦在门外，报 "abusive traffic patterns"。 我折腾了整整一天才想明白：它根本没在看我是谁。它在看我这个包，从哪个网段来的。 拦住你的从来不是"你做错了什么"，是"你从哪来"。