AI Coding News — August 28, 2026: Claude Code Ships a Restricted Mode

Two days ago we left Claude Code fresh off its SendFeedback tool, OpenCode's Cloudflare Gateway fix settling in, and GitHub Copilot's changelog gone quiet since August 10. None of that held. Copilot broke its silence with three releases in three days, Claude Code shipped a --restricted flag aimed at safer unattended automation, Codex CLI added @ mentions for referencing other tasks, and OpenCode picked up Azure Entra ID support while fixing Bedrock and Azure CLI sign-in. Here's what actually shipped since our last check-in.

Claude Code ships a --restricted flag

The headline release is v2.1.248 (Aug 27) in the official Claude Code changelog. Its --restricted flag strips out the built-in tools for running commands or code and WebFetch, keeps file tools scoped to the working directory, and refuses to combine with bypassPermissions — a locked-down mode built for CI pipelines and untrusted automation rather than interactive sessions.

• v2.1.248 (Aug 27): --restricted flag; experimental.cacheTtl in agent frontmatter for per-agent prompt cache TTL; claude self-hosted-runner --client-label; settings diagnostics in /doctor and /status; a /web-setup warning when the GitHub CLI token lacks workflow scope; /usage-credits for Enterprise AWS Marketplace orgs; cross-session messaging (SendMessage/ListAgents) now works on Bedrock, Vertex, and Foundry.
• Also in v2.1.248: two prompt-cache fixes worth knowing about if you run long or scheduled sessions — one for an hourly cache miss caused by OAuth token refresh re-rendering tool definitions, and one for ScheduleWakeup definition changes forcing a full cache miss on resumed sessions.
• v2.1.250 (Aug 28): logged simply as "bug fixes and reliability improvements."

That follows v2.1.247 (Aug 26), which added the SendFeedback tool and /claude-api cost-optimize command covered in our last roundup.

Codex CLI adds @ mentions, tightens image-token accounting

OpenAI's Codex CLI moved from v0.149.1 (Aug 24) to v0.150.0 (Aug 26) to v0.150.1 (Aug 27), per the ChatGPT & Codex changelog.

• v0.150.0 (Aug 26): reference other Codex tasks inline with @ mentions; auto-generated descriptive titles for unnamed terminal tasks instead of generic labels; /copy now offers a picker for full responses, code blocks, or blockquotes; markdown links render as clickable labels in supported terminals; new Interrupt hooks fire commands when a turn gets interrupted.
• v0.150.1 (Aug 27): a narrower but useful fix — remote compaction now counts retained images toward the token budget by default and trims older ones as needed, so long sessions full of screenshots stop quietly blowing past context limits.

OpenCode adds Azure Entra ID, ships twice in one day

Two releases landed the same day this time: v1.18.24 and v1.18.25, both August 28, per the project's GitHub releases. v1.18.24 brought Bedrock fixes, Azure Entra ID integration, V1/V2 config compatibility, and a desktop fix for archived-session visibility. v1.18.25 followed almost immediately with a fix for Azure authentication so Azure CLI sign-in works without requiring Bun to be installed. Together they extend the Cloudflare AI Gateway fix from v1.18.23 (Aug 25) covered two days ago — this is a good week to be an Azure or Bedrock user of OpenCode.

GitHub Copilot breaks its quiet streak

Our last roundup noted that Copilot's changelog had gone quiet since August 10. That changed within 48 hours. Three releases landed back to back: the GitHub Copilot app's Customize tab — bringing MCP servers, plugins, skills, and canvases into one place — went generally available on August 25. Global model policy went GA on August 26, giving Business and Enterprise admins a single default policy that newly GA models inherit automatically, with open-weight and data-retention models off by default. And on August 27, Copilot code review added resolution reasons — tagging a resolved comment as Addressed, Won't fix, or Incorrect — plus expanded coverage for bot-authored pull requests, cloud agent pull requests, and large PRs that previously hit a 300-file/20,000-line ceiling.

What it means for developers

Three changes are worth acting on rather than filing away. If you run Claude Code in CI or any untrusted context, try v2.1.248's --restricted flag before reaching for a custom sandbox. If your Codex CLI sessions lean on screenshots, v0.150.1's image-token accounting keeps long sessions from silently overflowing context. And if you administer Copilot for a team, the new global model policy is rolling out through September 1 — check which models it's enabling or disabling before that date. Everything else this cycle is incremental: OpenCode's Azure and Bedrock fixes matter if you're on those clouds, and Copilot's Customize tab and code-review resolution reasons are quality-of-life rather than new capability.

Sources & further reading

• Claude Code changelog (Anthropic)
• ChatGPT & Codex changelog (OpenAI)
• OpenCode releases (GitHub)
• Global model policy generally available (GitHub Changelog)
• Copilot code review: Resolution reasons and expanded capabilities (GitHub Changelog)