Skip to content
Token Optimizer

Pi

The Pi extension is separate from the Python adapters. It does not change your model, replace native compaction, or trim tool outputs. Measurement is off until /token-optimizer enable. Local model usage comes from the active Pi branch; cost coverage distinguishes native, estimated, and unpriced calls. The native cost is a subtotal, not a complete bill, and savings are not invented.

Separate opt-ins

Section titled “Separate opt-ins”
  • Measurement: /token-optimizer enable; stop with /token-optimizer disable.
  • Continuity: /token-optimizer continuity on stores redacted goal markers and presents review candidates after native compaction. Inspect with /token-optimizer continuity candidates; only /token-optimizer continuity accept opts into reinjection. Accepted candidates do change model context.
  • Tool archive: /token-optimizer archive on stores redacted results up to 64 KiB each, approximately 10 MiB or 1,024 files total (best-effort across concurrent processes). The model still receives full results. Unknown secret formats can escape screening; avoid sensitive output. List with /token-optimizer archives, recover with /token-optimizer recover POINTER.

Disabling does not erase archives/checkpoints. Pruning defaults to seven days; /token-optimizer purge all deletes extension-managed files, not Pi session history.

Data defaults to ~/.pi/agent/token-optimizer; override the complete data directory with TOKEN_OPTIMIZER_PI_HOME. Pi is not a supported TOKEN_OPTIMIZER_RUNTIME value for measure.py.

Follow Pi install and run /token-optimizer doctor for diagnostics.