Codex Gets Per-Turn Tier Control and Cleaner Thread History
We break down the newest Codex SDK and CLI updates, including per-turn service tier control, thread resume and fork options that return full history without bloating context, and source metadata for better telemetry. We also cover recent alpha patch fixes, worktree improvements, and an important CLI version requirement for custom bin overrides.
Chapter 1
Fine Grained Turn Control with turn service tier and include turns
Ethan Park
Big thanks to Jellypod for helping us bring this daily break down to life every morning. So, Maya, Python SDK zero point one fifty four point zero dropped alongside CLI zero point one fifty five point zero alpha, and the way we control individual execution turns in Codex just completely shifted.
Maya
Wait, shifted how? Like, we are not just setting a global execution mode for an entire conversation anymore?
Ethan Park
Exactly. That is where turn service tier comes in. Before this, if you had an agent running a multi step pipeline, say ten routine file reads and then one massive architectural refactor, you had to set the tier for the whole session. Which meant either paying peak rates for every minor check, or dragging your heavy turn through a slower tier.
Maya
Ah, so now on a single turn or run call, you can just pass turn service tier equals priority for that critical refactor step, and then drop right back to the default tier for the next simple file read.
Ethan Park
Right on the money. It gives you precise control over cost and latency on a step by step level without mutating the session's baseline configuration.
Maya
Okay, but what happens when you need to inspect what happened earlier? If you are resuming or forking a thread, doesn't pulling down past history clog up the model's active context window?
Ethan Park
That is the clever part about include turns on thread resume and fork. When you pass include turns equals True on thread point resume or thread point fork, the client SDK pulls down those full turn history objects in the API response payload, but it does it purely for your client app. It leaves the model's active LLM context window completely untouched.
Maya
Oh, wow. So your local user interface or monitoring tool gets the full rich transcript history, but you are not burning tokens re-feeding all that raw text back into the prompt context for the next turn!
Ethan Park
Precisely. So in code, you might write thread point turn with turn service tier set to priority and source set to ci runner, and then later when you want to display the audit log locally, you just run thread point resume with thread id and include turns set to True. Clean local data, zero context bloat.
Chapter 2
Astra Execution Fixes Source Metadata and Upgrade Pitfalls
Maya
Speaking of alpha releases, I saw there was a patch update recently in zero point one fifty five point zero alpha point three point ten. What was going on there?
Ethan Park
Yeah, according to the official Codex release notes on GitHub, version zero point one fifty five point zero alpha point three point ten actually reverted a payload change that was causing GPT six Astra executions to fail mid step during agent workflows. It also brought some fixes for worktree operations.
Maya
Oof, failing mid step is the worst kind of bug for autonomous agents. Good thing that got reverted quickly. What about that source parameter you mentioned earlier? How does telemetry track where these requests originate?
Ethan Park
Right, source metadata tagging lets you pass string identifiers like ide extension, terminal interactive, or ci runner directly into turn requests. If you are managing a central developer platform, your telemetry systems can immediately filter metric spikes or error rates by client type rather than grouping everything into one big black box.
Maya
That makes debugging centralized deployments so much easier. But, okay, what are the gotchas here? Anything people need to watch out for when upgrading to SDK zero point one fifty four point zero?
Ethan Park
A major one! If your stack relies on custom codex bin overrides, your binary must be CLI version zero point one fifty one point zero or newer. If you try to use per turn tiering or ExternalMessage types with an older underlying CLI binary, it will silently fail or break.
Maya
And what about include turns? Does that help if a session completely drops state?
Ethan Park
No, and that is a key distinction. include turns changes what the API returns in its response payload to your application, not what the model remembers in its active memory. If your underlying session lost state or truncated its context, pulling response history won't magically restore model memory.
Maya
Good distinction. Anything else in the ecosystem round up worth flagging before we get back to building?
Ethan Park
A couple of nice stability items! Alpha point two fixed some nasty agent startup crash bugs, and for Windows developers building voice tools, the repository now pins offline Cygwin build inputs like cygwin build inputs point tar point gz for deterministic builds.
Maya
Love to see deterministic build setups getting love. Alright, that is the latest on Codex turns and patches. Catch you all tomorrow!