AI Coding News: July 23, 2026 — Claude Code Backgrounds Code Review, Copilot CLI Adds Gemini 3.6 Flash
Two releases stand out today. Claude Code's v2.1.218 moves /code-review out of the main conversation and into a background subagent, and quietly closes a security gap where agent frontmatter hooks could run from folders that had never been trusted. GitHub Copilot CLI, meanwhile, is iterating fast on a v1.0.74 pre-release train that adds support for Gemini 3.6 Flash and introduces a first-run splash screen explaining the CLI's default sandbox.
Claude Code v2.1.218: Code Review Moves to the Background
Claude Code v2.1.218 landed July 22 and continues the run of near-daily point releases this month. The headline change is that /code-review now runs as a background subagent instead of filling up the current conversation, so review output no longer crowds out the work you were doing — and stacked slash commands can now target the review as it runs. The same background-by-default pattern extends to skills: any skill declared with context: fork now runs in the background unless it opts out with background: false.
The more consequential fix is a security one: agent frontmatter hooks were able to run from folders that had never gone through workspace trust. As of this release, a hook defined in an agent file only runs if that agent file's own folder has been explicitly trusted — closing a path where a malicious or careless agent definition dropped into an untrusted directory could execute code on load. Anyone who loads custom agents from shared repos or downloaded skill packs should treat this as a reason to update promptly.
Accessibility also gets attention: --ax-screen-reader mode now announces deleted text for word and line deletions (Option+Delete, Ctrl+W, Cmd+Backspace, Ctrl+U, Ctrl+K), and VoiceOver no longer misreads a trailing space as "new line." On the reliability side, the release fixes Windows paths with \u-prefixed segments (like C:\Users\unicorn) getting corrupted into CJK characters and made inaccessible; stops the left arrow key from silently discarding a conversation without a confirmation prompt; adds HTTP status and error detail to claude mcp list and /mcp when a server fails to connect; and resolves a retry loop that kept re-sending doomed requests after a context-overflow error. Bedrock users get a metering fix so application-inference-profile ARNs and other config-mapped model IDs are priced at the correct configured rate, and the Bedrock setup wizard no longer fails profile verification for assume-role profiles in partitioned AWS regions.
Smaller but still notable: /deep-research now only starts when invoked manually rather than being launched automatically by Claude; plan mode with auto no longer prompts for Bash commands its static analyzer can't already prove read-only, deferring to the auto-mode classifier instead; and a race in remote session heartbeats that left long-lived desktop and IDE processes endlessly retrying a rejected request has been fixed.
GitHub Copilot CLI v1.0.74: Gemini 3.6 Flash Arrives, Sandbox Gets a Splash Screen
While v1.0.73 remains the latest tagged stable build, GitHub has been shipping a fast train of v1.0.74 pre-releases (-0 through -3) since July 21, with -1 on July 22 carrying the most substantive changes. It adds a first-run splash screen that explains the CLI's default sandbox and lets new users opt in deliberately rather than discovering it mid-session, plus support for the gemini-3.6-flash model as a selectable option alongside Copilot's existing model lineup.
The same pre-release fixes a session-multiplexing bug where an open dialog in one session could leak into another instead of reopening cleanly when you switched back, and makes the $ interactive-shell shortcut usable even while the agent is actively working. It also preserves = characters — the kind found in base64-padded secrets and tokens — when the /mcp add and /mcp edit wizard writes environment variables, stops remote session uploads from retrying permanent 400/404 responses, and downscales oversized tool-result images so CAPI Responses requests keep going through. The earlier v1.0.74-0 build added a /model plan command for picking a model specifically for plan mode.
Why It Matters
Both releases point at the same underlying trend: as these tools take on more autonomous, background-running work, vendors are having to retrofit the trust boundaries that autonomy needs. Claude Code backgrounding /code-review is a workflow convenience, but pairing it with a fix that stops untrusted-folder hooks from executing is the more important half of the release — it's a reminder that anything with hook or agent-definition support deserves the same trust-boundary scrutiny as running arbitrary code. Copilot CLI's sandbox opt-in splash screen is a smaller version of the same idea: making a security-relevant default visible to the user instead of silent. If you run Claude Code with custom agents pulled from external sources, update to v2.1.218 without delay. If you're tracking Copilot CLI's pre-release channel for early model support, v1.0.74-1 is the build to grab for Gemini 3.6 Flash.
Sources & Further Reading
Comments
Share your thoughts and join the conversation
Leave a Comment
Keep reading.
Daily Dev Stack Release Audit — July 28, 2026: Redis, MongoDB, and Nuxt Rush Out Critical Security Patches
AI Coding News: July 28, 2026 — Model Context Protocol Ships Its Final 2026-07-28 Spec
