OpenAI Codex release notes: 0.143.0-alpha.38 makes code mode and proxy flows easier to test

OpenAI Codex has not received a new stable CLI entry after 0.142.5 in the public Codex changelog. But the alpha track has moved: GitHub release rust-v0.143.0-alpha.38 was published on July 7, 2026, and the npm registry shows alpha: 0.143.0-alpha.38, while latest remains 0.142.5.

So this is not a broad alpha rollout recommendation. It is a release notes signal for teams already testing OpenAI Codex close to real work. Codex CLI is OpenAI's local coding agent: an AI assistant that runs in the terminal and can read, change, and test code within the permissions you give it.

Source: OpenAI Codex changelog, GitHub release 0.143.0-alpha.38, npm @openai/codex

OpenAI Codex release notes: what changes in 0.143.0-alpha.37 and 0.143.0-alpha.38?

The two latest alpha versions are not a polished marketing release. There is no large headline in the official Codex changelog yet. But the comparison from 0.143.0-alpha.36 to 0.143.0-alpha.38 includes practical changes that are easy to miss if you only check the npm version.

Three items are worth picking up:

• The code mode host is now built, signed, and packaged with the primary Codex release binaries for Linux, macOS, and Windows. The PR says release artifacts previously bundled codex without codex-code-mode-host, which meant code mode could not start from installed packages.
• features.respect_system_proxy starts moving more of the real product path through the system proxy. PR #31335 covers the two HTTP endpoints for Responses API and compact, not WebSockets, model discovery, memories, realtime, or file uploads.
• Codex now warns at session startup when a configured service_tier is not supported by the selected model. Before this, Codex could silently omit the value, which made troubleshooting slower than it needed to be.

A system proxy is the operating system's or organization's route for network traffic, for example PAC/WPAD or a managed proxy. For Nordic teams behind company networks, school networks, or managed developer environments, that matters: login can work while the actual model traffic still needs the right route.

Source: PR #30202 bundle code mode host, PR #31335 route Responses API through system proxy, PR #31284 service tier warning

What does this mean for a Hammer team?

I would treat 0.143.0-alpha.38 as a test receipt, not as a new default version. The stable npm channel is still 0.142.5. The alpha channel shows where Codex is heading: better packaging, fewer silent network misses, and clearer warnings when configuration does not match the model.

For a team using Codex in a customer portal, internal app, or automation repo, the test is practical:

• Can you see exactly which channel and version is running?
• Does code mode start from the install path you actually use?
• Does Responses traffic take the right route in a system-proxy environment?
• Do you see a warning if someone has set a service_tier the model does not support?
• Can a cancelled /review be followed by a new review without MCP startup state getting stuck?

That last point comes from PR #31189, where a cancelled inline review could leave the TUI in a busy MCP startup state so the next /review was rejected. MCP, Model Context Protocol, is a way to give Codex access to external tools and data sources through scoped server connections.

Source: PR #31189 cancelled review fix, Codex CLI docs

Short example: use the new Codex feature

Human step first: choose a separate alpha test environment if you test 0.143.0-alpha.38. Do not ask the agent to install, switch channels, or change configuration in the prompt below.

Inspect this Codex workspace for 0.143.0-alpha.38 readiness. Do not edit files, install packages, or change configuration. Return one short release-readiness receipt with: visible Codex channel/version, whether code-mode host packaging is present in this install path, whether Responses API traffic should respect the system proxy, any unsupported service_tier warning risk, and one human review step before team rollout.

Good output should:

• separate stable 0.142.5 from alpha 0.143.0-alpha.38;
• state which install path was actually checked;
• mention proxy and service tier signals without guessing beyond the sources;
• end with a clear human decision before broader rollout.

What to watch next

The useful question now is whether these alpha changes land in the next stable Codex CLI release and whether OpenAI documents code mode packaging or system proxy behavior more broadly. Until then, one focused test is enough: version, install path, proxy, service tier warning, and review flow.

This is also a Tool Forge question. If Codex will connect to repos, issue trackers, internal tools, or customer-adjacent data, the team does not need more caution theater. It needs a working integration routine: scoped tokens, environment variables or a secret manager, bounded permissions, approval gates, and a short run log that someone actually reads.