Stay under a token budget

Goal: keep a fusion within a fixed token budget so it fits the context window of the model you are feeding.

Cap with a hard limit

fuse dotnet --directory ./src --skeleton --max-tokens 100000

--max-tokens stops emission when the running total reaches the limit. In a scoped fusion the most relevant files are emitted first, so the files closest to the seed survive the cut and the least relevant drop. Use it when the fusion must fit the window and partial content is acceptable.

Split instead of dropping

fuse dotnet --directory ./src --all --split-tokens 200000

--split-tokens (default 800,000) divides the output into multiple parts when a part would exceed the threshold, rather than dropping anything. Use it when you want the whole fusion, delivered in pieces a model reads in sequence.

Find what to trim

fuse dotnet --directory ./src --all --track-top-token-files

--track-top-token-files reports the largest files after the run, so you know which to scope out or reduce further when you are over budget.

Match the tokenizer to the model

fuse dotnet --directory ./src --all --tokenizer o200k_base --max-tokens 150000

A token count is only accurate against the tokenizer that produced it, so set --tokenizer to match the model the fusion is destined for. Supported names are in Tokenizers.

Suggested budgets

When to use it

Set a budget whenever the fusion is destined for a model with a fixed window or when you are capping the cost of a run. --max-tokens enforces a ceiling by dropping content; --split-tokens keeps everything by dividing it.

Related concept

How parts and budgets are written is in the Output specification; choosing the smallest reduction that fits is in Reduction levels.