Claude Code's new mods can rewrite its interface and behavior; a built-in one flags what you or Claude might miss. Three more stories inside.
Anthropic shipped Claude Code v2.1.287 on Thursday, adding mods: plugins that can redraw the tool's own interface and step into a tool call before it runs, not just add skills on top of it. The same day, Earendil's Pi, an open-source coding agent, hit 1.0 and spun off an experimental framework for longer-running work, Pi Durable. OpenAI says GPT-6.1 Sol's slow first days are over and is resetting ChatGPT's usage limits for every paid account today, though an independent benchmark still clocks the model as slow through its API. And GitHub put computer use into public preview in Copilot CLI and the Copilot app, so Copilot can click and type inside other programs on your Mac or Windows PC.
What changed this week #
Claude Code adds mods, plugins that rewrite its own interface
Anthropic's Claude Code v2.1.287, released Thursday, adds mods: plugins built from JavaScript or TypeScript event handlers that can redraw the tool's interface or step into a tool call before it runs, not just add new skills. The feature is on by default. One of the built-in mods, You should know, ships switched off: it runs a second agent that watches your session and flags things you or Claude might miss, turned on with /plugin enable cc-plugin-you-should-know@builtin. The docs are blunt about the risk: a mod "runs with your permissions" and can "read and write your files, start processes, and make network requests." Mods "aren't sandboxed," so a process one starts runs outside Claude Code's sandbox even when you've turned sandboxing on. Anthropic's announcement on X passed 2.9 million views by Friday morning, and the first user-built mods are already on Reddit.
Pi 1.0 ships, paired with the new Pi Durable
Pi, Earendil's open-source terminal coding agent, hit 1.0 on Thursday. It adds native support for MCP, the standard agents use to reach outside tools (Earendil calls its version "Codemode"), plus extensions that can define virtual models and a full-screen mode by default. Earendil also shipped Pi Durable, an experimental framework of about 15,000 lines for agents that outlive one terminal session and take direction from more than one person. "It does not replace the Pi coding agent," Earendil says. Pi 1.0 topped Hacker News, passing 1,500 points.
OpenAI resets ChatGPT usage limits after Sol's rocky launch
OpenAI resets usage limits for every paid ChatGPT account at "10am PST" today, Tibo Sottiaux of its Codex team said on X, apologizing for GPT-6.1 Sol's "slow start" and calling it "back to running at expected speeds." Through the API, Artificial Analysis, an independent benchmarking firm, still measures Sol (Medium) at 53 tokens per second, "notably slow," down from 59.4 on Wednesday. On Reddit, a Codex thread posted since says Sol "seems much faster now."
GitHub's Copilot learns to click around your desktop
GitHub's Copilot can now see and control other apps on your desktop, not just write code. It reads the screen, clicks controls, types, and drags across whatever program is open — public preview, macOS and Windows. Run /computer on in Copilot CLI, or flip it on in the app's Settings. Copilot asks permission before controlling an app; you choose which ones it can always touch. Try it on something low-stakes first: this is Copilot driving your real desktop, not a sandbox.
Also worth your time #
- For two weeks, design, deck, and doc work you start in the Claude app uses 50% less of your usage limits, Anthropic says (X)
- Cloudflare launches Clef, open-weight decision models for fast classification hosted on Workers AI, plus a hands-on reinforcement-learning service to fine-tune them, with self-serve promised later (Cloudflare)
- Cursor adds GLM 5.3 and GLM 5.3 Flash; Cursor says GLM 5.3 Max is the best-scoring open-weight model on its own CursorBench 4.0 (X)
- Moving to Sonnet 5.5? Code that turns thinking off with "disabled" now gets a 400 error; send "between_tools" instead, at high effort or below (Claude docs)
- livenerf, the daily benchmark tracking whether Opus 5.5 quietly got worse, posted day 8 of 30 with no verdict yet; the first possible one is around October 24 (GitHub)
Know someone who'd want this in their inbox? Forward it — that's how this grows. And if we got something wrong, or you think we buried the real story today, hit reply. A person reads every one.
The New Way is written with AI. It gathers the day's stories, checks them against their sources and drafts every summary. A person decides what runs and reviews every issue before we hit send.