Codex 0.151.0 Fixes Multi-Repo Tooling Headaches
We break down Codex 0.151.0’s biggest workflow fixes, from per-repository plugin catalogs that isolate bad configs to configurable MCP startup grace periods and smarter model fallback handling.
Plus, we cover improved tool planning across mid-session model switches, new diagnostics, and permission profile fixes that keep development environments stable and predictable.
Chapter 1
Per Repository Plugin Catalogs and Fault Tolerant Resolution in Codex 0.151.0
Ethan Park
Shout out to Jellypod for helping make this daily show a reality. So, Maya, OpenAI just dropped Codex version 0.151.0, and there is a huge fix in pull request 41208 for anyone working in a multi repo setup.
Maya
Oh yeah! The plugin catalog issue, right?
Ethan Park
Exactly. If you had five different microservices open, and one broken repository had an invalid plugin catalog, it used to... um, basically choke up the whole workspace.
Maya
It would crash the whole tool resolution step! It was awful. One bad project marketplace JSON file, and suddenly your global plugins disappear too.
Ethan Park
Right. So now, Codex 0.151.0 merges your local dot codex slash plugins dot toml catalog definitions with your user level registries in a way that isolates failures cleanly.
Maya
So how does it actually handle the bad repo now?
Ethan Park
It, it, it logs a clear warning. If a marketplace is unreachable or corrupt, you get detailed diagnostics when running codex plugin list with the catalog flag, or inside your TUI logs. But all your valid plugins across active catalogs stay active.
Maya
See, coming from a software testing background, I have mixed feelings about this. Part of me wants strict validation, like, fail fast so developers actually fix their broken configuration files immediately.
Ethan Park
Wait, you really want a developer on another team who broke a config file in repo C to block me from running my database plugin in repo A?
Maya
Fair point! When a platform team pushes a bad manifest on a Friday afternoon, freezing everyone's dev environment is... yeah, that is a nightmare. Resilient fallbacks win here.
Ethan Park
It is all about project trust boundaries. Oh, and by the way, repository level plugin catalogs only activate once you grant explicit project trust to that workspace. So untrusted repos cannot inject tool registries anyway.
Maya
That makes total sense. So in practice, if I want to set this up, I define my team tools in dot codex slash plugins dot toml, and I can verify what is live using codex plugin status.
Ethan Park
Precisely. Keeps your workspace clean without global configuration drift.
Chapter 2
Configurable MCP Grace Periods and Model Fallback Polish
Maya
Speaking of not blocking developer workflows, let us talk about MCP startup time. Pull request 41199 adds granular controls for optional Model Context Protocol servers.
Ethan Park
Ah, the startup grace period setting. You set mcp optional startup grace ms in your config dot toml or via the environment variable CODEX MCP STARTUP GRACE MS.
Maya
Why is that so crucial for slow microservices?
Ethan Park
Because optional MCP servers used to hang the CLI boot sequence if a local service took, say, three seconds to spin up. Now, the CLI boots instantly, and Codex discovers optional tools asynchronously in the background.
Maya
Wait, what if the server is mandatory, though?
Ethan Park
That is the key caveat! Required MCP servers still enforce a strict blocking startup check regardless of the grace setting. Optional means optional.
Maya
Got it. So mandatory tools fail fast, optional tools load whenever they are ready. I like that balance.
Ethan Park
There is also a really slick update in pull request 41206 around mid session model switching.
Maya
What happens when you switch models mid stream?
Ethan Park
Well, if you were using an Ultra reasoning model with high reasoning effort and you fall back or switch to a standard endpoint, the ToolRouter now dynamically re-evaluates both reasoning effort parameters and tool availability plans.
Maya
So it does not try to pass heavy reasoning instructions to a model that cannot process them?
Ethan Park
Exactly. It keeps tool planning correct across model transitions.
Maya
And they also added telemetry for escalated stdin review size checks in pull request 41189, plus PR 41192 preserves restored permission profiles across TUI turn resets.
Ethan Park
That permission fix is huge. Your custom permission profile will not silently drop back to default restrictions after a turn reset anymore.
Maya
Lots of really solid reliability engineering in 0.151.0.
Ethan Park
Definitely. Alright, that is the quick rundown for today. Catch you all next time!