Codex Shipped Its Biggest Feature Drop Since the AWS Bedrock Launch — and Nobody Noticed
OpenAI published one of the most substantial Codex releases in months on April 30, and the internet barely registered it. That's the story — not the features, but the reception context that makes the features more interesting, not less.
The release covers six major areas: persisted /goal workflows with app-server APIs and TUI controls, an overhauled MultiAgentV2 configuration system with thread caps and wait-time controls, a plugin marketplace with remote bundle caching, external agent session import, expanded permission profiles with sandbox CLI profile selection, and Bedrock model support fixes. None of that is small. All of it dropped late afternoon US time on a Thursday with no accompanying blog post, no press release, and no explanatory video from the product team. The result was almost no external coverage despite the release containing more significant new functionality than the AWS Bedrock announcement 48 hours earlier.
The most immediately useful feature is /goal persistence. If you've been using Codex for anything beyond single-session tasks, you know the failure mode: a long refactor gets abandoned because you had to restart your terminal, or a multi-step implementation gets halfway done before context degrades. The new system lets you create, pause, resume, and clear goal threads across sessions with runtime continuation. That's not a gimmick — it's the feature that makes Codex behave like a real project management layer rather than a session-bound chatbot. This was one of the most-requested workflow improvements in the Codex user community, and the implementation is more complete than a typical v1 feature: app-server APIs, model tools, runtime continuation, and TUI controls all ship together.
The MultiAgentV2 changes are less immediately visceral but potentially more significant for enterprise buyers. Thread caps and wait-time controls are the configuration surface that production deployment requires. Without them, multi-agent orchestration in regulated environments is governance theater — you can't explain to a compliance team why your agent system has unbounded concurrent threads. The explicit v2 depth handling is what makes multi-agent Codex production-ready in environments that require audit trails. If you've been watching the multi-agent space and wondering when it becomes actually governable rather than just technically possible, this release is your answer. The configuration surface has significantly expanded since the initial v2 rollout, and the granularity being exposed tells you the team is thinking seriously about enterprise deployment, not just individual developer experience.
The plugin marketplace launch is the longer-term strategic move. Remote bundle caching means plugins load faster and can be served from a centralized distribution point. External-agent config import means Codex can now consume sessions from other agentic systems, which is an interoperability signal that matters as the agent ecosystem fragments across vendors. The practical implication for teams building agent stacks with Codex at the center: you're getting the tooling to treat it as a platform rather than a point solution. That framing — platform versus product — is how OpenAI is going to need to position Codex going forward if the enterprise push through systems integrators is going to stick.
The Bedrock-specific fixes are the quiet confirmation that the AWS integration is moving faster than the announcement implied. apply_patch support being fixed, GPT-5.4 reasoning levels corrected, updated Bedrock GPT-5.4 endpoint and model metadata — these tell you the Bedrock API surface is being actively developed, not just announced. If your organization is evaluating Codex via Bedrock, this release is more relevant than the headline partnership announcement. The fact that the team is shipping endpoint-level patches alongside feature work is a signal of product velocity that the initial announcement didn't convey.
There's a pattern worth naming here: OpenAI's best Codex releases tend to arrive with minimal fanfare while the marketing-heavy announcements get more coverage. The April 28 Bedrock announcement generated dozens of articles. The April 30 feature release that actually ships theBedrock integration fixes, plus goal persistence, plus the plugin marketplace, plus the multi-agent governance surface — that got picked up by nobody. That's either a timing problem or a signal about where OpenAI's press strategy is versus its engineering velocity. Given that the enterprise rollout through Accenture, Capgemini, CGI, Cognizant, Infosys, PwC, and TCS is now live, the gap between what OpenAI's PR machine talks about and what its engineering team ships is increasingly a story in itself.
For practitioners, the immediate actionable item is this: if you've been waiting for /goal persistence before treating Codex as a reliable long-horizon coding partner, the wait is over. If you're deploying multi-agent Codex in a regulated environment, the thread caps and wait-time controls are the governance controls you've been asking for. And if you're evaluating Codex via Bedrock, the endpoint-level fixes in this release tell you the integration is further along than the announcement suggested. The release notes are on GitHub. Read them. The features are real and they shipped without anyone telling you to pay attention.
Sources: OpenAI Codex Releases, Reuters on Bedrock