Token savers

Avo includes a deterministic, opt-in token saver for long tool-heavy chats. It is an internal implementation inspired by terse-output and context compression patterns; it does not require an external rtk or caveman binary.

Choose a preset

avo saver list
avo saver show compact
avo saver use full

The setting is stored in ~/.config/avo/saver.json (or the directory selected by AVO_CONFIG_DIR). Existing chats keep their provider wrapper, so restart a chat after changing the preset.

Preset	Behavior
terse	Adds the caveman-terse concise-answer style guide.
yagni	Adds a minimal-build style guide for implementation tasks.
compact	Minifies JSON tool output and removes exact duplicate results.
full	Combines concise output, compression, verbose-output elision, and pre-trimming.

Disable the persisted setting with:

avo saver off

For one process or CI invocation, use the environment variable. It takes precedence over saver.json:

AVO_SAVER=compact avo

Unknown preset names fail closed and list the valid choices. The original messages remain in the event history; compression happens only on the request sent to the provider. A SAVER_APPLIED trace event reports the preset, stages, estimated tokens before/after, and percentage saved.

What the numbers mean

The pipeline is deterministic and provider-independent. Its token estimates use len(text)//4, so the measurements are useful for comparing presets but are not billing data or exact tokenizer counts. See the committed benchmark results.