--- title: "Model retirement week: what got switched off, and a radar to catch the next one" description: "Issue #002 of VibeKit Weekly: GitHub Copilot, Codex and Lovable retired or scheduled out models, Claude Code fixed a worktree bug that could lose commits, Cursor and Lovable gave agents more reach, plus the project of the week: a Model Retirement Radar." issue: 2 date: 2026-10-21 slug: issue-002 --- > **How to read this issue:** every news item links to its source, mostly official changelogs. Dates are as published by each vendor for the week of October 12 to 20, 2026. Version numbers and model names change fast, so click through before you change anything. Nothing here is sponsored. ## The short version - **Models got switched off.** GitHub Copilot deprecated six models on October 19, Codex retired GPT-5.5 for ChatGPT sign-ins on October 14, and Lovable set end dates for two Gemini Flash models used in app AI features. If you hard-coded a model name anywhere, this is your week to check. - **Claude Code 2.1.304** fixed a bug where `claude --worktree` could delete a worktree and its branch without asking, which could lose commits. Update. - **Agents got more reach.** Cursor's cloud agents can now build shareable canvases, and Lovable apps can expose tools to browser agents. More reach means more to review before you click "allow". ## This week's news: what changed, what breaks, what to do ### 1. The model retirement wave **What changed.** [GitHub deprecated six Copilot models on October 19](https://github.blog/changelog/2026-10-19-github-copilot-models-deprecated-on-october-19-2026/): Gemini 3.7 Flash, GPT-5.5, GPT-5.4, GPT-5.4 mini, GPT-5 mini and Grok 4.5, with suggested replacements Gemini 3.8 Flash, GPT-6.1 Sol, GPT-6 Luna and Grok 4.7. GitHub has already [announced another round for November 16](https://github.blog/changelog/month/10-2026/). Separately, the [ChatGPT and Codex changelog](https://learn.chatgpt.com/docs/changelog) says GPT-5.5 retired from Codex on October 14 for users signed in with ChatGPT; the OpenAI API and API-key Codex sessions are not affected. And the [Lovable changelog](https://docs.lovable.dev/changelog) marks Gemini 3.6 Flash as deprecated for app AI features, stopping on November 19, 2026, with Gemini 3.7 Flash following on January 28, 2027. **What breaks.** Anything that names a retired model directly: saved model settings, custom agents, scheduled tasks, scripts, and AI features inside apps you built. GitHub notes that Copilot Enterprise admins may need to enable the replacement models in their model policies before people can pick them. **What to do.** Search your projects for model IDs (this week's project automates it). Swap to the suggested replacement, re-run a few saved test prompts, and put the November dates on your calendar now. ### 2. Claude Code: a commit-losing bug fixed, and help with usage limits **What changed.** [Claude Code v2.1.304](https://github.com/anthropics/claude-code/releases/tag/v2.1.304) (October 19) fixed `claude --worktree` removing the worktree and its branch without asking at exit when HEAD was detached or on another branch, "which could lose commits". It also added `/auto-mode-setup`, a guided setup that reads your project to teach auto mode about your environment, a safety prompt for `rm -r` on mounted drives, and a warning on Pro and Max plans when resuming an old conversation would use a noticeable share of your 5-hour usage limit. Earlier, [v2.1.301](https://github.com/anthropics/claude-code/releases/tag/v2.1.301) (October 15) added a setting so unattended retry mode stops at a usage limit that resets too far away instead of waiting. **What breaks.** If you use worktrees and run an older version, unpushed work in a worktree is at risk when you exit. **What to do.** Update Claude Code, and push or commit worktree branches before closing a session. If you keep hitting limits, take the resume warning seriously: a fresh conversation is often cheaper than reviving a huge old one. ### 3. Cursor: canvases, bigger environments, stricter delete checks **What changed.** The [Cursor changelog](https://cursor.com/changelog) says cloud agents can now build interactive canvases you share by link (October 16), cloud environments support up to 300 repositories and large ones load 5x faster, and an Environments API is in beta. On October 13 Cursor opened an Origin App Marketplace with over 20 integrations, including Cloudflare, Linear, Sentry and Vercel. The [Cursor CLI v2026.10.08 notes](https://cursor.com/docs/cli/changelog) make auto-review ask before risky recursive, force or wildcard deletes, and make `/skills` flag skills that break the Agent Skills limits, such as a name over 64 characters or a description over 1,024 characters. **What breaks.** Skills you wrote by hand may now show warnings in `/skills`. Nothing stops working, but it is a sign the skill may not load the way you expect. **What to do.** Open `/skills`, fix any yellow warnings, and keep skill names short, lowercase and hyphenated. ### 4. Lovable: let browser agents use your app (carefully) **What changed.** [Lovable](https://docs.lovable.dev/changelog) can now add WebMCP tools to your app, so an AI agent built into a browser can act on the page a person has open, for example marking a task done. Lovable says it starts with tools that read information or are easy to undo, and adds tools that delete, pay, send, publish or share only when you ask. There is also a new "Always allow all tools from this connection" option for chat connectors, which covers tools the server adds later, and new Growth tools that turn a goal like "my first 100 signups" into a plan. **What breaks.** Your safety assumptions. "Always allow all" means future tools you have never seen also run without asking. **What to do.** Keep "ask each time" for any connector that can write, send or pay. Only ask Lovable for destructive browser tools when you really need them. ### 5. GitHub: autofix goes GA, and public clones may need a login **What changed.** [Agentic autofix for code scanning alerts](https://github.blog/changelog/month/10-2026/) and AI Scan for pull requests became generally available on October 15 (for GitHub Code Security and Advanced Security customers). On October 19 GitHub also said that during periods of unusual traffic, some HTTPS clones and fetches of public repositories may temporarily require authentication. **What breaks.** Scripts, CI jobs or tutorials that clone public repos anonymously could fail intermittently. **What to do.** In CI, authenticate clones (the built-in token works), and add a retry to setup scripts that fetch public repos. ### 6. Also worth a click - **Replit** lets you [work across projects from one chat](https://docs.replit.com/updates/2026/10/09/changelog) (October 9): find projects, read their files and send them tasks, with nothing published automatically. A Windows desktop preview is on a waitlist. - **Microsoft Defender for Endpoint** documents [discovery and runtime protection for local AI agents](https://learn.microsoft.com/en-us/defender-endpoint/ai-agent-support-matrix), including Claude Code and OpenAI Codex. Runtime protection, which can audit or block agent actions, is in preview; see the [overview](https://learn.microsoft.com/en-us/defender-endpoint/ai-agent-runtime-protection-overview). Useful if your company is nervous about agents on laptops. - **Meta** [announced Muse Wonder](https://dev.meta.ai/resources/blog/build-with-muse-wonder) on its developer site, pitched for long-running coding agents and available through the Meta Model API and Muse Code. Read the [model card](https://research.meta.ai/static/muse-wonder-model-card) before trusting launch claims. ## Project of the week: build a Model Retirement Radar **Level:** beginner to intermediate · **Time:** an afternoon · **Category:** dev tools Last week you built a Changelog Translator. This week's project is its sidekick: a small tool that scans a folder for AI model names and tells you which ones are retired or about to be. It turns this week's headline into a two-minute check you can run every Monday. **What it's for, in one sentence:** find every hard-coded AI model name in my projects and warn me before the vendor switches it off. ### Starter prompt ``` Build a small command-line tool (Node.js, no paid services) called model-radar. 1. It reads a file models.json that I maintain by hand, a list of entries like: { "id": "gpt-5.5", "status": "retired", "date": "2026-10-19", "replacement": "gpt-6.1-sol", "source": "" } 2. It scans a folder I pass in (default: current folder) for those ids in code, .env.example, JSON, YAML and Markdown files. Skip node_modules, .git and anything listed in .gitignore. Never read or print the contents of .env files. 3. It prints a table: file, line, model id, status, date, suggested replacement, and the source link. Retired = red, scheduled = yellow. 4. It exits with code 1 if it finds a retired model, so I can use it in CI. 5. Add a --json flag and a short README. Rules: do not guess retirement dates. Only use what is in models.json. Put the file name above every code block you write, and make one change at a time. ``` ### Checklist - [ ] `models.json` has at least this week's entries, each with a source link - [ ] The scan skips `node_modules`, `.git` and `.env` files - [ ] Output shows file, line, status and replacement - [ ] A retired model makes the tool exit with code 1 - [ ] You ran it on a real project and fixed at least one hit - [ ] You committed before each big AI edit, so you can undo **Stretch goals:** run it in a GitHub Action every Monday, fuzzy-match near misses like `gpt5.5`, or pipe a vendor changelog through last week's Changelog Translator to suggest new `models.json` entries for you to approve. ## Builder tips from the week *(Collected from public posts by other builders. Treat them as ideas to try, not facts.)* 1. **Ask for the file name above every code block.** A builder [suggested it on X](https://x.com/GetPlugzz/status/2112444266911760780), and it saves a lot of "where does this go?" confusion. 2. **Stretch your plan.** One [widely shared post](https://x.com/marouane53/status/2112501048711508232) described hitting usage limits on two Max subscriptions within days. Ask for one change per prompt, use a smaller model for simple steps, and start a fresh chat when a conversation gets long. 3. **Check "will people come back?" before you scale.** Another builder [posted](https://x.com/ThisIsRajat24/status/2112564751238877590) that a launch got 500 visitors in two days and two paying users, and both cancelled a week later. Before you buy ads, ask five real users to come back tomorrow and watch what they do. 4. **Skip money and health apps as a first project.** A [post this week](https://x.com/realFatCat1/status/2112536715265028198) warned about vibe-coding your own trading platform, where bad data can flow in unnoticed. Pick something where a bug costs you a laugh, not savings. ## Next week We'll keep tracking the November 16 Copilot round and the November 19 Lovable cutoff. Tell us which tools you use, and we'll prioritise their changes. --- **Sources & transparency:** every news item above links to its source, mostly vendor changelogs and release notes, checked on October 20, 2026. Builder tips come from public posts and are opinions, not verified facts. This issue has no sponsor. VibeKit Weekly is not affiliated with any tool vendor mentioned.