Simon Carter
Weekly Digest

AI product notes: week of 15 June 2026

A weekly roundup of smaller AI product updates, release-note items, and admin changes worth tracking.

Weekly AI product update digest

Update, 17 June 2026: Codex scheduled automations now correctly honour selected approval mode

The latest Codex app release fixes a bug where scheduled automations ignored the approval mode configured by the user or admin and fell back to a default instead. The fix is now live; update the app to receive it.

This is governance-relevant for Enterprise teams. When your organisation policy permits it, automations default to approval_policy = "never". If admin-enforced requirements disallow that, automations are supposed to fall back to the behaviour of your explicitly selected mode. Before this fix, those stricter admin policies were not being applied to scheduled runs, meaning configured approval boundaries were silently bypassed. The automations reference covers how approval policy interacts with admin requirements in detail.

The risk is sharpest for teams running automations with full-access sandbox mode, where Codex can modify files, run commands, and access the network without prompting.

Three things worth doing now:

  1. Update the Codex app.
  2. Review approval-mode settings on any existing scheduled automations, particularly in managed workspaces where requirements.toml restrictions were configured.
  3. Confirm sandbox mode for automations touching production repositories.

The fix shipped alongside a batch of other corrections tracked in the Codex changelog, including manual project ordering, sidebar pull request status updates, and remote SSH connection issues.

Update, 17 June 2026: Claude Code v2.1.178+ bug-fix wave: mid-stream connection drops preserved, WSL2 scroll fixed, Bedrock GovCloud 400 errors resolved, and sub-agent denylist silently-ignored bug patched

A cluster of patch releases around v2.1.178 landed on June 17, addressing four reliability regressions introduced in earlier builds. Details are confirmed across the GitHub releases page and the Claude Code docs changelog.

Mid-stream connection drops now preserve partial responses rather than surfacing raw errors or leaving the spinner stuck on “running tool.” API drops during thinking now trigger an automatic retry instead of showing “Connection closed while thinking.” WSL2 mouse-wheel scrolling under Windows Terminal and VS Code is restored after a regression introduced in v2.1.172. Bedrock GovCloud regions (us-gov-*) were deriving the wrong inference profile prefix (global instead of us-gov), causing 400 errors on derived model IDs; that is now fixed.

The most security-relevant fix patches a silent-ignore bug in sub-agent tool denylists. MCP server-level specs (mcp__server, mcp__server__*, mcp__*) passed to disallowedTools were not being enforced at all before this patch. Teams that assumed they were blocking an entire MCP server’s tools from sub-agents were getting no enforcement. Enterprise teams using server-level deny rules should treat this update as urgent and audit existing configs after upgrading.

Run claude update to install, then claude --version to confirm you are on v2.1.178 or later.

Smaller product updates do not always need a full article, but they can still matter if you own the platform, run the admin console, or build against the API.

Updates