Plan your project

Planning mode is where you shape work before anything runs. Nothing in it can change your code.

It holds two kinds of conversation, and the control above the message box says which one you are in.

Project chat

A Project chat manages the project itself: your Inbox, your roadmap, your open questions, and the list of plans. It does not write a plan of its own.

It changes those records directly rather than describing changes back to you. Ask it to add something to your Inbox and the row appears, with an Undo beside it for the next minute. Ask "what should we do next?" and it answers from what is actually recorded, not from reading your files.

It can still read your code when that helps it answer — "why does the port poller flake?" is a fair question. It cannot edit anything it reads.

Roadmap changes need your approval

Your roadmap is a set of commitments, so an agent does not rearrange it on its own. Ask for a reordering and you get one card showing the resulting order — not a list of individual nudges — with Apply, Edit and Cancel. Nothing is written until you apply it.

The card arrives once, when the agent finishes its turn, so you are never asked to approve half a thought. It appears at the top of Planning, and on your phone, because proposals travel with everything else.

Your Inbox and your open questions are different: those are a scratchpad, not a commitment, so a capture lands straight away.

Plan chat

A This plan chat works on one saved plan. The plan opens in Preview beside the conversation and updates as you refine it. This is the mode to use when you know what you are building and want the approach written down properly.

Creating a plan does not start it. "Create a plan for the Windows tray" gives you a plan record and opens it — nothing runs.

The download icon beside Execute… saves the current plan as a Markdown file.

Bring more than one agent to a plan

A plan gets better when more than one mind works on it. Switch the model or persona in the composer and send: a Reviewer to poke holes, a Designer for the screens, Codex for a second opinion, then back to Claude. Queue several follow-ups, each on a different model, and they take their turns one after another. Each one builds on the plan rather than starting over.

What each one is given:

• The full plan, always. Whenever the model answering did not write the latest version itself — a different provider, a new chat opened on the plan, or after you edited it — the whole current plan goes with your message. It is never shortened.
• The conversation, as much as fits. Switching to a different model family or persona within the same provider keeps the whole conversation. Switching provider (Claude ↔ Codex) passes along what the other one missed, up to roughly the last 48,000 characters; anything older is left out.
• Nothing is dropped silently. Before you send, the composer says who takes the next turn and what they will get. After you send, a line under your message says exactly what that model was given — for example "GPT-5.5 was given the full plan (revision 4) and the latest 18 of 30 messages it hadn't seen."

The plan's header lists everyone who has shaped it, model and persona.

Starting a build is always yours

An agent cannot start a build. Ask one to build something and it opens the execution review — the phases, which agent handles each one, and the Execute button — and then waits for you to press it. That review is also where you choose per-phase models and personas, so it is a better place to end up than a run that started itself.

Providers

Project chat needs to call Nocturnal's own tools, which requires Claude today. Pick a Claude model for a Project chat; Codex works normally for plan chats and for building. If you pick a Codex model for a Project chat, Nocturnal tells you why it cannot start rather than opening a chat that can only talk.

Nothing needs setting up. The tools are there whether or not you have ever configured an MCP server.