AI Coding Roundup — September 9, 2026: Claude Code Ships v2.1.265 as GitHub Copilot Adds JetBrains Sandbox Policies
Two releases dominate this week's AI coding tool news: Anthropic pushed out Claude Code's largest changelog in weeks and then had to ship a same-day hotfix, while GitHub rolled out enterprise-grade sandbox controls for Copilot in JetBrains IDEs. Here's what actually changed on September 8, 2026, and why it matters if you run either tool in a team setting.
Claude Code v2.1.265: A Big Release, Then a Quick Fix
On September 8, 2026, Anthropic shipped Claude Code v2.1.265 — documented in the official changelog — one of the tool's largest updates in recent weeks, spanning plugin loading, prompt-cache stability, Windows sandboxing, and VS Code session handling.
- Resuming a foreground-spawned subagent no longer breaks prompt-cache reuse by silently changing its tool list or system-prompt prefix.
- Tool results saved to disk are now capped at 1 GB, with truncation clearly flagged in the in-conversation preview.
- --plugin-dir can now point at a folder containing multiple plugins; folders added or removed while Claude Code is running are picked up automatically.
- VS Code and SDK sessions no longer force re-login when a session closes mid token-refresh.
- Windows sessions running inside an AppContainer or restricted-token sandbox no longer have Read, Write, and Edit blocked outright.
- Two-key keyboard shortcuts (common inside tmux) now wait three seconds before cancelling, instead of failing silently.
Hours later, a regression forced a follow-up release. In v2.1.265, the undocumented CLAUDE_CODE_USE_GATEWAY environment variable started forcing Cloud-gateway sign-in on its own — previously it was ignored unless both ANTHROPIC_BASE_URL and ANTHROPIC_AUTH_TOKEN were set. That broke any setup combining it with an API key, apiKeyHelper, or custom auth headers, failing every request with "Not signed in to the Cloud gateway." v2.1.266, shipped the same day, restores the old behavior. Teams running Claude Code through an LLM gateway or proxy should update immediately; no configuration change is needed.
GitHub Copilot for JetBrains: Enterprise Sandbox Policies Land
GitHub shipped a matching JetBrains update on September 8 (changelog): administrators can now enforce enterprise-managed sandbox policies for Copilot's local sandboxing feature, next-edit suggestions gained cross-file cursor jumps, and chat gained global project context. It builds directly on the plugin-, MCP-, and OpenTelemetry-level enterprise controls GitHub added to Copilot for JetBrains back in August 2026.
The direction is consistent across both vendors this week: rather than new models, September 8 was about governance — capping resource usage, locking down sandboxes, and giving admins levers to standardize how AI coding agents behave across a fleet of developer machines.
What About Codex CLI and OpenCode?
OpenAI's Codex CLI has held at v0.153.4 since September 4–5 (changelog), and OpenCode's newest build is still v1.18.29 from September 4 (changelog). Both were already covered in our September 6 AI Coding Roundup, and neither has shipped a new version since.
Bottom line: if you run Claude Code behind a gateway or proxy, get to v2.1.266 today. If you administer Copilot for a JetBrains fleet, the new sandbox policy controls are worth reviewing before rolling local sandboxing out further.
Resources & References
Comments
Share your thoughts and join the conversation