Codex CLI 0.157.0 adds GPT-6 Sol/Luna on Bedrock, locks down network calls

OpenAI's Codex CLI moved to rust-v0.157.0 on September 25, adding GPT-6 Sol and Luna as model choices, including a path through Amazon Bedrock. The release notes also list fullscreen transcripts turned on by default, a keyboard shortcut to fork a conversation, and enforced network restrictions that now cover redirects and live HTTP and WebSocket traffic. For a tool that developers run against real codebases and, increasingly, real company infrastructure, the network change is the one worth reading twice.

Codex is a command-line agent that can execute code and make outbound calls on a developer's behalf. Before this release, restricting what it could reach over the network was reportedly less reliable at the edges, specifically redirects and persistent connections like WebSockets, where a policy could be set but not fully enforced. The release notes state the restrictions now apply across those paths. OpenAI's announcement does not say what the prior failure mode looked like in practice, or whether any incidents prompted the fix. That gap matters for anyone trying to assess risk retroactively.

For a manager who approved Codex for a team, the practical shift is that network policy is now a control worth auditing rather than a setting to configure and forget. If redirects and streaming connections were previously a soft spot, any existing allowlist or denylist should be retested against this version before trusting it in a regulated environment. For a policy watcher, the question is evidentiary: does enforcement across redirects and live traffic hold up under adversarial testing, or does it close one gap while leaving others. OpenAI's release notes assert the behavior; they do not provide test results.

Bedrock support for Sol and Luna changes the calculus for teams already standardized on AWS. Routing model calls through Bedrock can mean existing AWS billing, IAM roles, and data residency commitments apply, instead of a separate API relationship with OpenAI directly. That is a procurement and compliance detail as much as a technical one, and it is the kind of change that shows up in a vendor risk review before it shows up in a changelog. Whether Sol and Luna perform differently from prior models on coding tasks is not addressed in the release notes.

The interface changes are smaller but not trivial for daily use. Fullscreen transcripts by default change how much screen a Codex session occupies and how output gets reviewed during a working session. The conversation fork shortcut lets a developer branch a session without losing the original thread, useful when testing two approaches against the same prompt. Neither change touches security posture, but both affect how a team's engineers actually work with the tool day to day, which shapes whether adoption sticks.

Anyone running Codex in an environment with outbound restrictions should pull the 0.157.0 release notes directly and compare them against the specific network policy rules already in place, since enforcement behavior is the part OpenAI changed without fully detailing the before and after.