@ferueda

Cool pattern I've been using with GPT-5.6 today Advisor thread orchestrates (GPT-5.6 Sol High+) with instructions for the executor threads (Terra or Luna) to checkpoint back or ask for feedback to the original orchestrator. This way, the orchestrator, which is usually a more expensive model, doesn't have to continually poll executor threads and can steer them on demand.

@janstevens One big advantage I see to sticking with one sub is that you can customize the codebase specifically for that agent, which makes it run way more smoothly than trying to accommodate many of them Especially for Sol, that needs especial treatment imo

I agree with Matthew. If you can only choose one subscription, ChatGPT is the one to go for. Pros/cons anyone?

It should come with a ChatGPT sub for 1 month or smth am I right? or credits

@mattpocockuk The codebase is the harness

The code is the environment your agent runs in Ignore it, and the world around your agent crumbles If you still think ignoring the code - i.e. vibe coding - is the future, I'd love to hear a counter to this

@coleschfr I've had it spin off for hours and turn a task into a huge 10k+ diff 5.6 needs taming and explicit instructions around overengineering and scope

Codex limits and 5.6 make absolutely no sense > start Sol Ultra with 80% 5 hour window remaining > works for 12+ hours, I have to "Continue" twice and limits are at 0% > still going for 11+ hours, limit is still at 0% > spun up 20+ child threads to review it's work > it's still going > should've been 2K LOC for 2 small issues > it decides to refactor everything for maximal code health > finds infinite bugs > edited 480 files > +80,000 LOC and -30,000 LOC > it's still going what the fuck did they feed this model

@tylermayberry agreed, /goal is obsolete with 5.6

/goal is a liability now Just makes Sol overengineer unless you're ultra specific

yep same experience. I've had multiple implement - review loops spiral down to 10k+ diffs bc of this. I had to add rules to some of my skills to tame these - Every abstraction, phase, compatibility layer, test, or command needs an acceptance criterion, invariant, or verified risk. - Reject future-proofing, optional hardening, nearby cleanup, generic “best practices,” and unrelated refactoring. - “Minimum sufficient plan”; smallest coherent change; every item traces to acceptance/invariant/risk; no preserved-behavior prose, speculative hardening, future-proofing, or repetition. - Prefer explicit, boring, repo-native code. Simplification must be materially smaller and behavior-preserving; no broad rewrites or optional refactors.

It seems like 5.6 Sol has a tendency to over engineer, and even go off the rails even more so than 5.5. 5.5 would usually just go overboard on tests. 5.6 Sol flagged an issue during a review that a Semantic Version string parser was problematic because it parsed each component as a plan Number, so it would fail and throw if the version number got larger than 9007199254740991. I was speechless.

I've experienced 1 and 2 heavily as well. I've controlled them by adding a couple rules: 1.- Use clear, simple language. Avoid jargon and unnecessary complexity. 2.- I've had multiple implement - review loops spiral down to 10k+ diffs bc of this. I have specific rules for ti as well now General: - Repository invariants/project intent → accepted user requirements and decisions → verified repository facts → proposals, patterns, and reviewer preferences. - Every abstraction, phase, compatibility layer, test, or command needs an acceptance criterion, invariant, or verified risk. - Reject future-proofing, optional hardening, nearby cleanup, generic “best practices,” and unrelated refactoring. Planning: - Prefer the smallest repo-native design. - “Minimum sufficient plan”; smallest coherent change; every item traces to acceptance/invariant/risk; no preserved-behavior prose, speculative hardening, future-proofing, or repetition. Implementation review: - Follow reviewed plan or direct request; stop on plan/repository conflict or scope expansion. - Pre-existing debt, nearby cleanup, alternative architecture, optional hardening, and unrelated refactors are out of scope. Recommend the smallest correction. - Prefer explicit, boring, repo-native code. Simplification must be materially smaller and behavior-preserving; no broad rewrites or optional refactors. These have helped a lot over the past days. Note that these are not all rule sin my agents.md, most of these are in skills. My agents.md is pretty small since 5.6

after a few more days of using gpt 5.6 sol, i started noticing some issues - if you have good solutions, please share! 1. it uses technical jargons a lot. it's almost speaking its own language that looks like English but you can't understand it until you ask "what do you really mean by this" 2. it can over-engineer and spiral out of control. something that can be done with a few lines of changes can often result in a massive diff fixing everything in the codebase 3. it's overly conservative in terms of touching live environment, so much so that it overly relied mocks for development and validation and build things that don't really work in production these things can likely get tweaked in system prompts, but so far gpt 5.6 is the only model family that don't get these things right out of the box, and it's a bit annoying i love overall how intelligent and fast the model is though! just curious if anyone's run into similar challenges with the model and has good solutions

@sama is the growth coming from codex or chatgpt work?

2.5x increase in usage of our agentic products (codex and chatgpt work) in the last week! welcome.

@sama solmaxxing

GPT-5.6 sol is half the price and ~twice as token efficient as fable in many cases for accomplishing the same task. happy to deliver at one-quarter of the price.

@reach_vb @OpenAIDevs @work_louder asking the right questions

@OpenAIDevs @work_louder which button for resets??

Meet kbd-1.0-codex-micro, built with @work_louder. Map the buttons and joystick to your workflow, and keep your pinned chats in view. Get yours before stock returns 410.

@DanielSmidstrup codex

If you had $200, what would you choose? - codex - cursor - claude code

👇👇👇👇

If you haven't tried Mole, do yourself and your laptop a favor and check it out github.com/tw93/Mole

@promotions24214 harness will make you spend more tokens tbh, as it optimizes for code correctness over cost (burns more tokens) sessions can help you analyze how can you make your sessions more efficient if you put an agent to analyze your threads

@ferueda Hello! I am trying to get better at coding using AI (Codex app). I've pretty much used 95% of my weekly limit in 7 hours. Should I try to download both of your relevant repos (harness & sessions) to be more efficient? Any advice would be greatly appreciated :)

just realized I need a serious project for the week of the 26th

@zachlloydtweets Great read, I have something similar going on in my local setup. A scheduled task extracts insights from my sessions through a CLI I built for this, analyzes my workflows, grades results and make suggestions.

@prd_008 Wife is pregnant, I taught Codex how to fill claim forms and submit my receipts

@btaylor Sometimes I catch myself rewording phrases just to sound less AI. Same with typos, I let them by for the same reason

I deeply resent that AI has forced me to eliminate em dashes from my writing for fear of signaling slop

@DavidOndrej1 😂😂

gemini prediction on point

@trikcode Loop a couple architecture reviews and you get the most beautiful and overengineered piece of software ever created

AI coding assistants are getting scary good at fixing isolated bugs, but they are absolutely terrible at software architecture. you accept 10 consecutive "perfect" PRs, only to realize the model quietly introduced three different state management patterns. I think we are trading short-term speed for massive, invisible architectural drift

@reach_vb @jxnlco Lmao 😂

@jxnlco ask codex

Anyone got any last-minute acting tips?