Reference

Budgets

Budgets put limits on how much work Nawa does at once and how large each model request may be. The shipped values suit most setups; change them when a provider or a machine needs it. For what the composer's token numbers mean, see Tokens and budgets.

Orchestration budgets

Max parallel agents

How many agents, such as researchers and the coder, may run at the same time within a task. From 1 to 8; the shipped value is 4. It applies to cloud providers; agents that share a model server on your own machine or network run at most two at a time, whatever this value says. Lower it if a provider slows down or rejects requests under parallel load.

Progress timeout (minutes)

How long work may go without progress before Nawa stops it as stalled. From 1 to 120; the shipped value is 40. The clock restarts whenever a real operation finishes or a verified workspace result changes, so long but productive work is never cut off. Model text alone does not restart it.

Stage repair attempts

Nawa saves this value (1 to 6, shipped as 3), but the current version does not use it.

Model and context budgets

Composer context tokens

The context size the composer's gauge assumes when a provider cannot report the model's real window. Nawa normally uses the window it measured for the selected model, so this is only a fallback. The Windows app ships 262,144.

Output tokens per model call

The most a single model reply may contain. Blank, shown as Auto (model limit), uses the model's own allowance; any value must be at least 256. A low limit can cut useful replies short.

Document budgets

Max attachment text (characters) and Document analysis threshold (characters)

Nawa saves these values (shipped as 100,000 and 15,000), but the current version does not use them. The note under them describes an earlier behavior.

The point at which a long conversation is compacted is set on the LLM / Image gen tab, under Advanced.