
Codex 0.147.0: Goodbye Full Auto, Hello Safer Automation
This episode breaks down the breaking changes in Codex 0.147.0, including the removal of full auto in favor of safer workspace write sandboxing and automated approvals. It also covers non-blocking MCP startup, secret redaction in logs, and how to standardize the new settings in config.
Chapter 1
Why Codex 0.147.0 Killed Full Auto
Ethan Park
So, uh, I was looking through the release notes for Codex zero point one forty seven point zero, and, well, they completely killed it. They removed the deprecated codex exec dash dash full auto flag entirely.
Maya
Wait, they actually deleted full auto? Oh man, that explain, that explains why our nightly test suite was, uh, completely exploding at three in the morning!
Ethan Park
Yeah, it is gone gone. The release explicitly says to remove the deprecated codex exec dash dash full auto flag, and use dash dash sandbox workspace write instead.
Maya
I, I spent two hours this morning trying to figure out why the CI runner suddenly started failing with an invalid flag error. I mean, we relied on full auto for all our automated scripts because it just, you know, bypassed all the prompts without stopping.
Ethan Park
Right, but that was always the danger, wasn't it? Full auto mashed together two very different things. It mixed execution permissions with approve, approval rights. So if an agent went rogue, it had full root access to your entire filesystem and network.
Maya
Okay, so by splitting them up, how does dash dash sandbox workspace write actually fix that?
Ethan Park
It creates a strict security boundary. Workspace write locks the agent down so it can only write files inside your local workspace directory. It can't touch system files, and it blocks unauthorized network calls out to random external hosts.
Maya
Ah, okay. So workspace write restricts where it can touch the disk. But what about the non interactive approval part? If we don't have full auto, does the script just hang waiting for a human to click yes?
Ethan Park
That is where the new flag comes in. You pair sandbox workspace write with dash dash approve for me. Approve for me handles the automated approvals so headless scripts keep running, but without giving subagents blanket host level access.
Maya
Wow, that is actually so much cleaner. So instead of one dangerous master key that lets a script delete your home directory, you decouple the filesystem sandbox from the auto approval mechanism.
Ethan Park
Exactly. You get headless automation for your pipelines, but if an agent tries to escape the workspace folder, the sandbox slams the door shut.
Chapter 2
Non Blocking MCP Startup and Bearer Token Redaction
Maya
You know, once I got past the breakages in zero point one forty seven point zero, there was actually a huge quality of life feature that blew me away. The, uh, the MCP protocol update!
Ethan Park
Oh, the, the twenty twenty six zero seven twenty eight update?
Maya
Yes! They adopted the non blocking server initialization from that spec. It used to be that if you loaded heavy Model Context Protocol tools, your terminal input would completely freeze until every single server booted up.
Ethan Park
Ugh, I remember that. You would launch a session and just sit there for ten seconds staring at a frozen cursor while optional background tools initialized.
Maya
Exactly! Now it is completely non blocking. The terminal is responsive instantly, and the MCP tools spin up in the background without locking your prompt.
Ethan Park
That is huge for agent velocity. And speaking of security and cleanliness, did you see what they added for logging? They added automatic secret redacting across command displays and replayed chat history.
Maya
Wait, really? It catches bearer tokens automatically?
Ethan Park
Yeah, bearer tokens, raw API credentials, private headers, it redacts them before they ever hit stdout logs or transcript history. So if a subagent runs a curl command with an authorization header, it won't leak into your terminal logs.
Maya
That is such a relief. I can't tell you how many times I've had to scrub raw bearer tokens out of shared terminal recordings or debug transcripts.
Ethan Park
Same here. Oh, and they also added persisted, manually ordered thread sections. So when you are managing complex, multi agent sessions over several days, you can organize your long running threads without losing your exact context or transcript positioning.
Maya
Oh, that is so nice for keeping context clean! So if a team wants to roll out zero point one forty seven point zero across all their developers today, what is the best way to standardize it?
Ethan Park
Simple. Update your local dot codex slash config dot toml file. Set your default flags in config dot toml to include sandbox workspace write and approve for me across the team, so nobody relies on old deprecated flags like full auto ever again.
Maya
Done and done. I am updating our team config right now. Good chat, Ethan!
Ethan Park
Talk soon, Maya.