I've slashed Codex's token startup cost 4 times while retaining full functionality. It works with ChatGPT desktop and is ready to tackle the quota abuse that the Dots will probably cause me. Works for me, will probably work for you. AMA. Compiled downloads and source code here : What cut the context overhead: - Stripped bloated system prompts and repeated tool instructions. - Replaced upfront tool inventories with compact usage hints and on-demand help. - Made skill discovery and loading on demand instead of injecting the whole catalog. - Load nested AGENTS.md by default instructions when reaching the relevant files. - Trimmed tool-output metadata and slimmed bundled skill instructions. Other highlights - /settings - you can switch some optimisations on and off from there. - all current experimental flags exposed in settings so you don't have to dig them up and then edit toml manually. - Persistent JavaScript/TypeScript Notebook with reusable variables and helpers across calls and context resets. Basically a persistent shell. - Notes-based OAI context management enabled by default for all models and several optimisations around that. - Thanks to this, all models can use their full 272k ctx. I made some precautions so they should hardly ever exceed it. - When you send a prompt after 25 mins, ctx window automatically rolls over before the prompt so your chat is never cold. Tried and tested, no risk to continuity at all. - Overall CTX management implementation vastly improved - I've been working on it for my Pi version since Astra released. - Agent board enabled by default. - Voice progress updates, context continuity (summarisation before engagement), and dropped-call recovery. - One personality file for text and voice (with a script adding voice personality to the desktop app - please note this changes every release so you'll have to vibe-update it), separate from AGENTS.md. OAI does not provide any of this. - Works as the backend for ChatGPT Desktop, locally or over SSH. Server works normally. You do not have to totally replace the Codex installation for this to work, the change to the ChatGPT desktop routing are 100% reversible. I suggest using .codex-lean as the global dir. Skills will need copying over as the skill tool supports categories. Instructions attached in the repo. Why do this? 1. Because I think Codex CLI could be more efficient 2. This is pretty much the best way to use Dots without messing around with herdr etc. 3. Yall like to whine about quotas. Yours truly has no issues with them, running basically this version of Codex, but in Pi.