Codex 0.142.0: Delegation Controls, Budget Caps, and TUI Fixes

Ethan Park

Welcome to the show, and thanks to Jellypod for helping make this daily show a reality. I'm Ethan Park, here with Maya, and today we are looking directly at some highly practical updates in Codex 0.142.0. If you have ever run a multi-agent workflow and watched in horror as your primary agent unilaterally decided to spawn five sub-agents, which then spawned their own sub-agents, all burning through your API credits, [chuckles] you are going to appreciate this release.

Maya

Oh, the classic agent sprawl. It is a genuine nightmare for budget predictability. You start a simple refactoring task, and suddenly you have a recursive tree of agents discussing variable names on your dime. It is basically the software engineering version of Mickey Mouse and the magic brooms in Fantasia. [chuckles]

Ethan Park

Exactly. [matter-of-fact] And to fix this, Codex has introduced three distinct delegation modes in 0.142.0: disabled, explicit-request-only, and proactive. They give you direct, granular control over how and when an agent can hand off tasks to a sub-agent.

Maya

Let us break those down. "Disabled" is pretty self-explanatory—it completely cuts off the agent's ability to spawn or delegate to any sub-agents, right? It keeps everything strictly single-agent.

Ethan Park

Right. The agent has to solve the problem with the tools it has locally, or fail. Then you have "explicit-request-only". In this mode, the agent can only delegate if it explicitly asks you, the operator, for permission first, or if your prompt directly instructs it to spin up a helper. It cannot make that call on its own mid-run.

Maya

And then "proactive" is basically the old behavior? Where the agent can autonomously decide, "Hey, this sub-task is better suited for a specialized code-writer agent, I am going to spin one up right now"?

Ethan Park

Yes, but now it is a deliberate choice you make, rather than the default wildcard behavior. And setting this up is incredibly straightforward. It is all managed in your `.codex/config.toml` file. You just add a `[delegation]` block and set the `mode` key. So, for example, `mode = "explicit-request-only"`.

Maya

[thoughtfully] That is great for global settings, but what if I have a specific script where I want to allow proactive delegation just for that run, without changing my global config? Can we override it dynamically?

Ethan Park

Absolutely. You can pass the mode dynamically on a turn-by-turn basis through the CLI or the API wrapper. This means you can keep a tight global default like "disabled" to prevent accidental spend, and only open the gate to "proactive" when you are running complex, trusted pipelines.

Maya

Speaking of preventing runaway spend, let us talk about the new rollout token budgets. Because even if you allow delegation, you still need a hard ceiling. Under the hood, how does Codex actually track this now? [curious]

Ethan Park

This is a big architectural improvement. Previously, tracking was often localized to a specific agent's run. Now, Codex tracks cumulative token usage across entire agent threads, including all sub-agents. You set a hard limit—a rollout token budget—and as the run progresses, Codex calculates the cumulative burn. If you start running low, it actually inserts remaining-budget warnings directly into the system prompts to warn the agent.

Maya

Wait, so the agent itself is made aware that it is running out of money? [chuckles] "Hey, you have fifty cents left, wrap it up"?

Ethan Park

[laughs] Literally, yes! It tells the agent, "Your remaining budget is X tokens," which encourages it to wrap up its work or stop spawning deep search loops. And if the agent ignores that and hits the hard ceiling, Codex doesn't just let it hang—it automatically aborts the turn cleanly, saving you from those infinite loop scenarios where an agent tries the same failing tool call fifty times.

Maya

That is a massive relief for anyone who has left an autonomous run active while grabbing lunch, only to return to a massive bill. Now, let us pivot to some of the terminal and network resilience updates in this release, because there are some really nice quality-of-life fixes here. Specifically, the Linux TUI and Ctrl+Z.

Ethan Park

Oh, this one was highly requested. If you were running Codex's Text User Interface on Linux and wanted to temporarily pause it—say, to run a quick git command in your shell—you would press Ctrl+Z to suspend it. But when you typed `fg` to bring it back to the foreground, the terminal rendering would completely break. Text would overlap, and the layout would turn into a scrambled mess.

Maya

I remember that. You had to basically close the terminal window or run a manual `reset` command to get your shell looking normal again. [sighs]

Ethan Park

Exactly. In 0.142.0, they fixed the signal handling for the Linux TUI. Now, when you bring it back with `fg`, it triggers a clean redraw of the interface automatically. It sounds minor, but for command-line purists, it makes the environment feel so much more native and stable.

Maya

And speaking of stability, what about those network disconnects? I saw something about stdio session survival.

Ethan Park

Yes, Stdio MCP—Model Context Protocol—sessions and the underlying exec-servers are now resilient to transient disconnects. If you are on a flaky Wi-Fi connection or switching networks, and the stdio connection drops momentarily, Codex doesn't just crash the session and throw away your active state. It keeps the session alive in a holding pattern and cleanly reconnects once the pipeline is restored.

Maya

That is huge for anyone working on laptops while traveling. You do not lose your entire context just because the train went through a tunnel. Now, what about the `/usage` command update? [curious]

Ethan Park

They have completely refactored `/usage`. It is no longer just a static output of your current token consumption. You can now inspect your usage-limit reset credits directly from the terminal. And more importantly, you can redeem those credits in-line.

Maya

In-line redemption? So you do not have to open a browser, log into a dashboard, click three buttons, copy a code, and come back?

Ethan Park

None of that. You just type `/usage` in your terminal. If you have earned reset credits, it will show them to you. You can trigger a redemption right there with built-in confirmation prompts. And they even built in retry states, so if the billing or credit server has a momentary hiccup during the redemption request, it won't fail silently or break your session.

Maya

[thoughtfully] That is incredibly clean. It really feels like Codex is maturing from an experimental tool into something designed for highly resilient, professional local development.

Ethan Park

It really is. Well, that wraps up our quick look at Codex 0.142.0. Dynamic delegation modes, hard token guardrails, and much-needed terminal resilience. Thanks for listening, and we will see you tomorrow.

Maya

See you then! [warmly]