Guides

Plan and Build modes

When to plan, when to build, and how the agent's questions work.

Every run has a mode. You switch it with the Plan / Build toggle next to the send button.

Build mode#

Build is the default. The agent:

1. reads the project and your request,
2. writes and edits files,
3. runs a Maven build,
4. reads the log and fixes compile errors,
5. finishes by committing the result.

When it succeeds you get updated files, a Files changed summary and a downloadable jar. Build runs spend credits for the AI work and build minutes for the compile time.

Plan mode#

Plan is strictly read-only. The agent has no tools that write files, run commands or build. It can read your code, think, and ask you questions, and it ends with a written plan. Nothing in your project changes and no commit is made.

Plan is the right choice when:

• the idea is still fuzzy and you want to be asked the right questions,
• you want to review an approach before any code is written,
• you want the agent to look over existing code and tell you what it would change.

Answering questions#

In Plan mode the agent may ask between one and five questions. Each is single choice, multiple choice or free text, with a few options to pick from.

• The composer is replaced by a tabbed form, one tab per question, plus a final overview.
• Every question has a Tell the agent something different field if none of the options fit.
• Send your answers and the run continues from where it paused.
• Cancelling while it waits cancels the run.

From plan to build#

When the plan looks right, switch to Build and send a short message such as "Go ahead with that plan." The agent has the plan in the thread and builds from it.

Things that apply to both modes#

• One active run per project. Start a new message after the current run finishes, or cancel it first.
• Runs have budgets. A run can stop early if it uses its allowed number of turns, its time limit or its cost budget. The thread tells you which. Break big requests into smaller ones if this happens.
• Long conversations are compacted. When the thread gets very long the agent summarizes older parts to stay within the model's context. A notice appears when this happens.
• You can cancel. Cancelling stops the run. Work already written stays in the project.