Skip to main content
Hiroshi utilizes a predictive character-to-token estimator to monitor active session history files and trigger context rotations before hit limits.

​
📉 1. The Compaction Pipeline

If a conversation’s total token count exceeds the configured safety threshold (e.g. 70% of model context):
  1. Runs an asynchronous evaluation turn using the NO_REPLY protocol.
  2. Prompts the model to synthesize a dense, semantic summary of all previous thread turns.
  3. Overwrites the history database with the generated summary, rolling forward the context state cleanly.
// Character-to-token safety boundary checking
if char_count * 3 / 4 > threshold_tokens {
    run_compaction().await;
}