VyroAI

Loading...

Plans, Questions and the Task Queue

In Chat mode the AI posts a plan card before it touches a file, and asks questions when your brief is thin. How Execute plan and the task queue work.

Plans belong to Chat mode

This is the first thing to get straight, because it decides whether you ever see a plan at all. Plan cards appear in Chat mode. In Build mode the AI gets on with it and writes the files.

So "plan it first" is not something the IDE decides for you by weighing up how big your request is — there is no such judgement. It is a mode you choose. Talk it through in Chat, get a plan, then run it. See Chat vs Build Mode.

The plan card

The AI posts a plan as a card in the chat, headed Proposed plan · N steps — nothing changed yet. Under that heading is the numbered list: each step names the file it will touch and what it intends to do there, and a small icon says whether that file is being created, updated or removed.

The words nothing changed yet are the point of the card. Your project is untouched while you read it.

The row along the bottom of the card. There is one button — Execute plan — and the way to change a step is to say so in the chat.

On this step card there is no Edit plan button and no Approve button: Execute plan is the one that starts the work, and if a step is wrong you simply reply in the chat and the AI proposes a better plan. (When a plan arrives as plain prose instead of a step card, a simpler fallback card appears with Approve plan and Edit plan buttons. Edit plan changes nothing today — it only folds the card away — so to change that plan, reply in the chat as well.) Once you press Execute plan, the card collapses to a short strip that keeps the agreed steps on the record.

Execute runs the plan you read

Pressing Execute plan switches the chat to Build and runs the plan the card was drawn from: the same steps and the same files. There is no second planning call, so you do not pay for the planning twice, and no file can quietly drop out. If an approved file still does not get written, the reply names it as a failure.

What the plan does not fix in advance is the code itself: the AI still writes each file. Read what it changed afterwards, the same as any other generation. Version History is the way back if a step went somewhere you did not want.

One exception: for an older plan card, or one from far back in a long conversation, the stored plan may not be found. Then the AI plans again from the card's lines, and any approved file the new plan leaves out is put back.

When it asks you questions

If your description leaves something genuinely open, the AI stops and asks instead of guessing. You get a short card with a few suggested answers you can tick, plus an Other… box for anything it did not think of. Apply sends your choices back as the next message and the work carries on.

The other button is Deny, and it does not mean "decide for me". Denying sends nothing at all: the card closes with the line "Dismissed — type your own answer below", and nothing else happens until you type an answer yourself. If you want the AI to use its own judgement, say so in the box — an empty chat is a stopped job.

A question gets an answer, not a plan

Asking something is not asking for work. "Is a database needed?", "what does this file do?", "does this bot need the members intent?", "how do I add a command?" — all of those come back as a written answer in the chat, with no plan card and no steps, and the answer uses your project's own files and values rather than a generic example.

The answer may end with one short offer, along the lines of "Want me to plan it?" — one, at the end, and only after you have already been given the answer you asked for. Nothing happens until you take it up.

This holds in both modes and at any size of project, including a brand-new project with no files in it at all. That case used to be the trap: an empty project normally makes the AI get on and build something rather than ask you what you meant, and a yes-or-no question was being treated as a vague build request — so a question came back as a whole generated project you never asked for. A question is answered instead, even on an empty project.

Where the line sits between a question and a request

A question that carries an instruction inside it is a request, and it is planned as usual. "Can you fix the bug?", "could you add a shop command?", "how about a ban command?" all ask for work, however politely they are phrased — and so does an order that follows the question mark, as in "is the config broken? fix it".

So if you genuinely want the work rather than the answer, say so plainly. And if you only wanted to know, leave the change word out — "is a database needed?" gets you the answer, "do I need a database, add one" gets you a plan.

A turn that finds nothing to fix will not rewrite your README instead

There is a related habit the app now blocks outright. A turn that looks at your project and concludes there is nothing wrong used to come back as a plan whose only change was a re-worded README.md — you paid for a fix and got documentation.

A plan whose only file change is the README is now dropped by the server unless you actually asked about the README or the documentation, and you get the written answer instead. Ask for the README by name and it is written as normal.

The task queue

Work that is waiting sits in a queue bar you can open. Tasks in it can be edited, reordered or removed before they run — so a five-step job is still yours to change after you have set it going.

The queue holds 10 waiting tasks. Line up an eleventh and it is refused with "Queue full" rather than quietly dropped, so nothing you typed disappears — send it once something ahead of it has run.

Read the plan properly. Correcting a wrong step while it is still a line of text costs nothing; correcting it after four files have been written costs a rebuild.

Also in Code Studio (IDE)

16 of 37 in Code Studio (IDE) · Last checked 11 September 2026 · Previous: Chrome Extension Studio