Skip to main content

needs_compression

atomcode_core::ctx::render

Function needs_compression 

Source 
pub fn needs_compression(
    conv: &Conversation,
    system_prompt_tokens: usize,
    token_budget: usize,
) -> bool
Expand description

Check if context needs compression.

Threshold derived from auto_compact_threshold — fires when fewer than buffer tokens remain (5K for ≤100K windows, 13K for >100K). Buffer scales with the deployment: self-hosted GLM at 65K trips at 60K (4K runway is plenty for one round); Anthropic at 200K trips at 187K, matching CC’s behaviour.

The messages.len() < 12 guard stays — needs a non-trivial backlog before compression is worthwhile, and 1 user msg can produce 15+ messages so message count is the right unit.