Codex 0.150.0: Cross-Task Mentions and Safer Agents
This episode breaks down Codex 0.150.0’s new cross-task mentions, showing how agents can message, query, and spawn subtasks without the usual copy-paste chaos. It also covers key security hardening for untrusted repos, Interrupt lifecycle hooks for cleanup, and quality-of-life upgrades like smarter task naming and copy controls.
Chapter 1
Inter Task Mentions and Terminal Agent Coordination
Ethan Park
So if you are running multiple terminal tasks in Codex right now, things just got a whole lot cleaner in version zero point one fifty point zero.
Maya
Wait, you mean we can finally stop doing the whole copy paste dance between different terminal windows?
Ethan Park
Exactly that. Thanks to Jellypod sponsoring today's daily update, we get to dive straight into this. OpenAI just dropped cross task mentions with @ syntax in Codex zero point one fifty point zero.
Maya
Oh, interesting. So like, how you @ mention someone in Slack, but for running agent tasks?
Ethan Park
Right. In the prompt input, you can literally type at task one zero two, or whatever task ID was generated. And Codex can read from that task, send a message to it, or even spawn sub tasks off it directly from your terminal session.
Maya
Oh wow. Okay, walk me through what that actually looks like when you are sitting there writing code. Give me a concrete scenario.
Ethan Park
Okay, say you are refactoring a primary module in one main terminal task, right? But in the background, you have a secondary task, let us call it at task eighty eight, running your integration test suite.
Maya
Right, right, and tests start failing.
Ethan Park
Exactly. Instead of stopping what you are doing, scrolling up through a thousand lines of terminal output, copying the error stack, and pasting it back into your primary prompt... you just tell Codex, hey, check at task eighty eight and figure out why the user auth test failed.
Maya
That is so nice! Because manually dumping terminal logs into context always feels so messy. And honestly, it burns through tokens like crazy.
Ethan Park
Well, and that is the clever part here. Codex does not just dump the entire transcript of at task eighty eight into your current window. It pulls structured task summaries, keeping context tight while giving the agent enough info to fix the issue.
Maya
So it turns individual terminal prompt loops into actual multi task orchestration. That feels like a pretty big shift in how developers will coordinate background agents.
Chapter 2
Security Hardening Interrupt Hooks and Patch Polish
Ethan Park
It really is. But speaking of running agents across different tasks, version zero point one fifty point zero also brings some big security hardening, especially around untrusted repositories.
Maya
Ah, untrusted repos. Let me guess, prompt injection risks through project config files?
Ethan Park
Spot on. In previous builds, if a project root had an AGENTS dot md file, Codex would just execute those instructions automatically. Now, if the project is untrusted, Codex ignores those project level AGENTS dot md instructions completely.
Maya
That is a huge relief for anyone reviewing open source pull requests. You do not want a cloned repo secretly overriding your agent's behavior behind your back.
Ethan Park
Exactly. Plus, managed deny read path rules now stay strictly enforced even if you cycle your TUI permission modes mid session. Security controls stay locked down no matter what.
Maya
Okay, what about when an agent goes off on a tangent or gets stuck in a loop? Anything new for handling interruptions?
Ethan Park
Actually, yes! They introduced Interrupt lifecycle hooks. So when you hit Control C to interrupt a top level turn, Codex can automatically run custom shell commands or Model Context Protocol handlers that you set up.
Maya
Wait, really? So if my agent spun up a local dev server or an active cloud execution, Control C can clean up those background processes automatically?
Ethan Park
Yep, no more orphaned node processes lingering in the background eating up your RAM!
Maya
Okay, as a former tester, that speaks directly to my soul. That is fantastic.
Ethan Park
There are a few really nice developer quality of life polish items in this release too. Like slash copy now opens an interactive picker where you can choose whether to copy the full response, just isolated code blocks, or blockquotes.
Maya
Oh, sweet! And didn't I see something about automatic task naming?
Ethan Park
Yes! Unnamed terminal tasks now get descriptive titles automatically based on what you are doing, and slash rename will actually suggest an editable title for you.
Maya
That makes tracking multiple at mentions so much easier. Oh, and markdown links now show up as clickable labels in supported terminals!
Ethan Park
Right. And we should definitely highlight the follow up patch that dropped right after, version zero point one fifty point one.
Maya
Oh, what did zero point one fifty point one fix?
Ethan Park
It fixed remote compaction for multimodal sessions. Retained images now get counted toward the token budget by default, trimming older images as needed so your context window does not quietly balloon during long image heavy debugging sessions.
Maya
That is super solid. Between cross task at mentions, safer untrusted repo handling, and interrupt hooks, zero point one fifty point zero is a monster release.
Ethan Park
It really is. Alright, that is the latest update for today. Catch you all on the next one!