Making It Wait for "Jobs That Run Over an Hour": Codex 0.152 Adds Ceiling Dials for MCP Output Volume and Execution Time, and Turns the Planning Tool Off by Default

OpenAI's coding agent Codex shipped v0.152.0 on August 31, followed by a fix release, v0.152.1, the next day on September 1. This isn't a flashy new-model story; the focus is on unglamorous tweaks that make it easier to run long jobs and workflows that lean heavily on external tools (MCP). Here's a rundown of the practical changes that matter most to anyone running agents unattended or semi-autonomously.

Capping Long-Running Jobs and Runaway Output

The core of this release is that it lets you, the user, decide the "time" and "volume" involved when an agent calls out to external tools to do its work. Areas that used to be left to implicit limits now come with explicit dials.

• You can extend the execution time limit: Clients going through the app server can set a timeout for thread/shellCommand, and you can even specify a deadline longer than an hour. Long-running jobs like builds or large test suites can be left to run without being cut off partway through.
• You can throttle output volume per tool: You can set an output_token_limit on a per-MCP-tool basis, and the same threshold trims output even after you resume a session. This lets you keep massive logs or search results from overflowing the agent's context, tool by tool.

Neither is about "intelligence"—both are about "handling." The more you rely on an agent for long jobs, the more room you now have to decide both the wait time and the input volume yourself.

You Can Give MCP Servers Package-Style Names

The MCP layer that bundles external tools has also gotten easier to work with. Server names can now use symbols like :, @, /, and ., so package-style naming such as @scope/name passes through as-is in CLI commands and authentication flows. In setups where you register many tools and switch between them, it's now easier to keep your naming consistent. Alongside this, a bug that made MCP tools inaccessible across cache refreshes or remote-side plugin updates has been fixed, improving stability for workflows that frequently swap out the tool shelf.

The Planning Tool Is Now Off by Default

A change that operators can easily overlook: the plan-update tool (update_plan), which drafts a work plan, is now disabled by default. If you want to use it, you'll need to explicitly enable it in your config with tools.update_plan.enabled = true. Workflows previously built around planning behavior may leave you puzzled by "missing plan steps" after the update. Since this is a behavioral change, it's wise to check your settings if you depend on it.

Small Quality-of-Life Improvements in the Terminal

The terminal experience has seen some touch-ups too. Since these come in a handful of items, let's break them out.

• Search within drafts: In Vim mode, / and ? search is available, matches are highlighted, and n / N move you forward and backward.
• Input mode fix: After sending a message or running a slash command, new drafts now correctly open in Insert mode.
• Guidance when you hit a limit: The rate-limit banner now surfaces concrete options such as checking usage, managing credits, the reset time, and changing your plan.
• Visibility into auth recovery: Both the terminal UI and codex exec now show the progress of re-acquiring authentication (including re-authenticating with Amazon Bedrock).

Behind the Convenience, Tightened Safeguards

Convenience improvements stand out this time, but the release also includes safety-side measures aimed at unattended operation. Cloud task requests now reject untrusted backend URLs and disable redirects to prevent credential leakage. In the next-day fix release, v0.152.1, Guardian's approval review was corrected to respect the Node REPL policy passed via model metadata. The more you delegate approvals to a machine, the more these "don't-just-wave-it-through" details start to matter. As you take the convenience, spelling out in your settings how much output and time to allow each tool—and whether or not to use the planning tool—becomes a prerequisite for running stably over the long haul. Note that this is a summary based on publicly available information; the surest way to know the actual behavior is to verify it in your own environment.

References: openai/codex Release v0.152.0 / openai/codex Release v0.152.1 / openai/codex Releases / Releasebot: Codex Updates