Omar

Pick a date, tell the list, and let the date be the constraint that decides what gets built. The logo has never once been the reason.

The bit that trips everyone up: the picker stopped being a model picker and became a compute picker. Instant is literally GPT-5.5 Instant, the low end of the same control, not some old model you accidentally selected. Medium and above are where Sol lives.

So if your effort setting landed on Instant, or got reset to it when 5.6 rolled out, every new thread genuinely starts on 5.5 Instant and the product is doing exactly what it says, it just is not labelled the way you expect. Move the thinking level to Medium or higher and see whether the label on the next reply changes.

The other documented half: on eligible paid plans it can step up from Instant on its own when a request looks like it needs more reasoning. That is why a long, gnarly prompt sometimes feels like Sol and a one liner never does.

Do not pin it globally. Pin it where the work is.

The global default has to be right for both "explain this stack trace across four services" and "convert this to kebab case", and there is no setting that is right for both. Per project settings let the serious project run at high effort and leave the default low, which also means the fast tier stays fast for the throwaway questions, which is the actual thing you like about it.

Direct answer to (1) and (2), from moving a similar codebase: your problem is retrieval, not the model, and swapping models will not fix it.

On a large multi service repo, every assistant is guessing which files are relevant from a fairly small window into the tree. The single biggest improvement I have made was not a model change, it was writing an architecture map: one file that lists each service, what it owns, the entry points, and where the shared code lives. Point the assistant at it in the rules or the prompt. After that Composer stopped inventing a new module next to an existing one, and so did everything else I tried.

On (2): assume a long design doc will not survive a multi file edit intact regardless of model. Break the work into steps small enough that the relevant part of the doc fits in the step, and restate the constraint in each step. That is annoying, and it is also what makes the cheap model viable.

Yes to the second one, and it is the answer to your whole question. Every turn re-sends the conversation, so cost per turn grows with the length of the thread. Turn 40 of a debugging session is not the same price as turn 4, it is substantially more, and the growth is roughly linear in how much history and how many files are in the window.

Practical consequence: start new chats aggressively. When you have finished a subproblem, summarise the outcome in three lines, open a new thread, paste the three lines. I halved my spend doing only that, without changing model.

On agent runs: it is not a flat multiplier per file, but a run that reads forty files pulls all of that into context and then keeps re-sending it for the rest of the run. The expensive thing is the reading, not the editing.

Since there is no notification from Apple and no gradual degradation, you find out from users. So make it not be users:

1. A scheduled job that decodes the configured client secret and alerts when exp is less than 30 days away. Fifteen lines, and it turns a silent cliff into a ticket.
2. A synthetic login that runs the real token exchange every few minutes from outside your network and alerts on a non-200. That catches this and roughly ten other things.

Both are cheap. The reason nobody has them is that this failure happens once every six months, which is exactly long enough to forget.

Shape test, in the order I would run it.

Vendor outage looks like: 5xx responses, timeouts, or wildly variable latency. Affects every flow including native. Starts for everyone within the same couple of minutes. Error bodies are inconsistent between attempts.

Your config looks like: instant failure, the identical error string every single time, one flow affected and not others, and it started at a suspiciously round moment.

First thirty seconds: open Apple's developer system status page and look for the Sign in with Apple entry specifically, not the general one. Next thirty: retry and check whether the error body is byte identical. Deterministic identical errors are almost never an outage.

It does not stop there. This is the single most expensive misunderstanding in this product and it is worth being blunt about.

Budgets and budget alerts notify, they do not cap usage or charges. The docs say so directly. On top of that the cost data itself lags, by up to a few days for some services, so the moment your $5 alert arrives you may already be well past $5 with more still unreported.

A budget is a smoke alarm. It is not a sprinkler. If you want something to actually stop, you have to build the stopping part, which is the Pub/Sub plus disable-billing pattern above, plus limits inside your own app.

That's a fair point I hadn't thought about, who gets picked, not whether something gets picked.

And a dev tab left open overnight with a live listener attached will quietly consume a day's allowance while you sleep, which is the single most demoralising way to discover this. Check that before you conclude your users are doing something weird.

Demo workspace regardless. It's the cheapest of the three, it works while you sleep, and it converts the 'show me before I do work' segment, which is most people. Sample CSV second. Calls are research, not a channel.

just drop mem_limit, the box has 2gb and the kernel will sort it out

you'll know it's time for CI the first time you try to deploy from a train on a phone hotspot and it dies halfway. that's the actual trigger for most people

The numbers on that kit are genuinely impressive for what it is: the transmitters are around ten grams, the quoted line-of-sight range is four hundred metres, and the whole system does something like forty-eight hours of runtime with the charging case. None of which matters at one metre in your own bedroom. Buy it for interviews and walk-and-talks, not for a static desk.

Fair challenge and the alternating test is the right call. I would only note that the direction of the effect is very consistent across everyone I know who has tested it, so the prior is not neutral. Test it, but do not be surprised.

Second this. Reply rate went from about 1% to 9% for me purely by swapping the pitch line for a question about their current workflow. Same list, same domain, same week.

One thing to decide up front is whether lint blocks the merge or just annotates it. We moved the slow type-aware pass to a scheduled run against main and kept only the fast checks on PRs. Fewer people waiting on a queue, and the rules still catch things within a few hours.

Fair, retracting. I'd used it for a legitimate case (reading config inside an effect without subscribing to it) and over-generalised.

you can wrap the write in untrack() and it'll stop complaining

Pin. A floating alias is an unannounced deploy of somebody else's code into the middle of your critical path, scheduled at their convenience. You would not accept that from a library dependency and this one is less deterministic than a library.

If you do buy a camera, the specific feature that helps your situation is manual exposure control, which most of the mid-range ones expose through a companion app. Lock the exposure to your face and let the window blow out to white, which looks completely normal to everyone else and is what a human eye does anyway. Auto exposure is what is producing the silhouette; it is doing exactly what it was told and averaging the frame.

Raising the price while conversion is 0.4% feels completely insane, and yet I keep hearing it.

The 19% are almost all freelance translators and they hit 200 docs in about nine days. Everyone else does six documents and leaves. That is uncomfortably clarifying, thank you.

Windows first. Sealing draughts is cheap, permanent and does not have a running cost. Then if you still want one, buy on clean air delivery rate for your room area rather than on the claimed square footage, which is always quoted at the top fan speed nobody can stand to sleep next to.

I got specific about it because vague advice to 'move more' never survived a busy week. Timer every 30 minutes, and the rule is only that I change something: stand up, recline, or just get water. That's roughly sixteen changes in a working day and it took my three o'clock ache from routine to occasional over about a month. Standing I limit to two half-hour blocks, because when I tried four hours a day standing I swapped a back ache for foot and hip aches within a fortnight.