Zed Delta opens to everyone, aiming to replace the pull request
Threads keep the full agent history, not just a diff; Zed already runs on zero PRs. Plus Copilot cost tiers, a 200 tok/s GLM API, and models that skip text.
Copy markdown
Threads replace the diff
Delta is a shared environment where you and agents work in threads that preserve the whole build — prompts, decisions, incremental edits — instead of a squashed final diff. Reviewers spin up isolated worktree subthreads to interrogate the agent and test fixes before merging. Free during public beta on macOS, Linux, Windows and the browser.
Zed already turned its own PRs off
Zed says it disabled pull requests internally and landed 570 changes to main across 33 developers without one. It runs on DeltaDB, a Git extension that records edits and chat as sub-commit ‘deltas’; paid plans are coming, but a free tier stays. CEO Nathan Sobo: ‘the thread will replace the commit or branch as the fundamental unit of software development.’
Elsewhere: Copilot lets you trade cost for smarts
GitHub Copilot’s auto model selection now exposes efficiency, balance and intelligence tiers across VS Code, the CLI and the app, so you can bias every request toward cheaper or stronger models. The same rollout auto-resolves addressed review comments and adds a Sentry canvas for triaging crashes before you open a fix.
Elsewhere: GLM-5.3-FlashX buys speed
Zhipu is now serving GLM-5.3-Flash through a faster tier, GLM-5.3-FlashX, at up to 200 tokens/sec, $0.37 / $1.25 per million tokens in/out and a 1M-token context. Same 320B / 18B-active model, just a quicker inference path — the base weights are open, this serving tier is not.
Build this weekend: let two models skip text
Cache-to-Cache (ICLR’26) wires LLMs to pass meaning through fused KV-caches instead of generating text between them — the authors report 3-5% higher accuracy than text hand-off and a 2x latency cut. Code and pre-trained fusers for Qwen and Llama pairs are on GitHub and Hugging Face under Apache-2.0.