Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-08-17 · high
Version 0.21.13 adds worktree locks and session IDs so concurrent AI reviewers stop destroying each other's work.
Why now: August 17's release notes are the first detailed public account of what breaks in review agents run at scale.
Qwen Code shipped version 0.21.13 on August 17, capping review suggestions to Critical findings after five rounds, per GitHub's release notes. Teams running review agents at scale pay for that gap: infinite loops, sessions overwriting each other's review state, corrupted checkouts that kill a run. Few vendors document these failures publicly. This changelog does.
The /review skill got rebuilt around dedicated platform subcommands, meta and fetch-diff, instead of raw gh commands issued through prompt prose. Stringing shell commands together inside natural-language instructions turned out to be fragile in ways a typed subcommand isn't.
Worktree leases now act as locks, so two concurrent review sessions can't destroy the same active review state. A corrupted checkout triggers an automatic wipe and retry instead of leaving a broken session behind. Session IDs plus diff hashes track which agent did what, so work survives an interrupted run.
That worktree lock fixes a symptom, not the cause: raw gh commands issued from prompt text, which any agent-review tool will hit eventually. Worth watching: whether other agent-review tools ship the same fix, or find the same failure first.
These release notes are the first detailed public account of what breaks in review agents run at scale, published August 17.
Each link below shares sources, entities, or timing with this story.
A spec is a press release until someone who didn't write it implements it. GitHub made Agent Plugins 1.0 generally available on August 12 across VS Code, Copilot CLI, the Copilot SDK, and the Copilot app on all plans. The spec, published August 6, was co-authored by AWS, Anysp...
Qwen 3.5 is the first major model pretrained specifically for agentic multimodal workflows from the first training stage, not fine-tuned after the fact. 397B total / 17B active parameters (MoE architecture), 256K context window, 201 languages. Ships with Qwen Code (terminal ag...
qwen-code 0.22.2 shipped three review changes on August 26 that deliberately pull opposite ways: a do-not-refute list and constructible rejection bar in the /review verifier to stop it invalidly rejecting speculative-but-real findings, test acceptance criteria attached to ever...
Qwen released Qwen-AgentWorld-35B-A3B on June 24: 35B total parameters, 3B active in an MoE, 256K context, Apache 2.0. It ships with AgentWorldBench. (GitHub) The idea is the interesting part. It's a "language world model," trained to simulate the environment an agent acts in....
cua-driver-rs v0.20.0, published August 24, ships a codesigned and notarized macOS universal binary plus a QwenCuaDriver.app, unsigned Linux x86_64/arm64 builds on a glibc 2.31 floor, and Windows builds, with a single @qwen-code/cua-sdk npm package built against those assets....
The repo appeared August 24, opening the weights of a multimodal MoE that Qwen frames explicitly as an architecture preview, the same role Qwen3-Next played for Qwen3.5 (GitHub). The hybrid Gated DeltaNet plus Gated Attention design it previews already carried through the Qwen...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.