Modes and Reduction Pipeline

Modes and Reduction Pipeline

Token Terminator uses one runtime mode to control which reduction paths are active.

Mode	RTK terminal rewrite	Temporal delta	Native search/process	Native read_file	Request compiler
balanced	✓	✓	✓	—	✓
aggressive	✓	✓	✓	✓	✓
native	—	—	✓	—	—
terminal	✓	—	—	—	—
suggest	measure only	—	—	—	—
off	—	—	—	—	—

Balanced

The default. It combines the broadest useful reduction set without applying aggressive structured reads.

Aggressive

Adds large read_file reduction to the balanced pipeline. Use it when tool-output volume justifies the extra transformation path and exact recovery remains available to the model.

Native

Disables RTK and request compilation. Useful when you only want deterministic large tool-result compression.

Terminal

Only terminal rewriting is active.

Suggest

Measures rewrite candidates without replacing the command.

Off

No provider-visible optimization is attempted.

Acceptance is per transformation

A mode only makes a path eligible. It does not force a reduction. Every candidate can still be rejected because it is not smaller, expands under the exact tokenizer, cannot be recovered safely, is malformed, violates backend constraints, or hits another fail-open condition.

This is why a healthy Token Terminator session can contain many untouched results.