TokenPak

Release

TokenPak v1.19.0

v1.19.0 · Aug 12, 2026

This backward-compatible minor release adds a versioned session-economics contract and a deterministic runway view while preserving unknown usage and pricing facts instead of inventing certainty.

Added

  • The immutable session-economics/1 contract exposes session facts, measured values, estimates, provenance, forecasts, guard state, and runway through 29 additive Python symbols. Value states distinguish measured, estimated, unavailable, and error results; existing public symbols are unchanged.
  • The proxy exposes /v1/messages/session-economics, a deterministic local view of session runway that does not call a model provider. Context soft and hard limits, request or dollar budgets, and rolling caps are evaluated conservatively; the binding constraint is reported explicitly.
  • Provider usage ledgers retain the source and quality of token-usage facts supplied by supported providers. Missing usage remains unknown rather than being coerced to zero.

Fixed

  • Re-ingesting an already wrapped Pak envelope preserves a single stable envelope instead of nesting another wrapper around it.
  • Proxy monitor writers now keep their database association and lifecycle stable across concurrent requests and shutdown flushing.
  • Proxy test fixtures use isolated ephemeral ports, preventing unrelated test processes from colliding on fixed listeners.

Upgrade

python -m pip install --upgrade "tokenpak==1.19.0"

No configuration migration is required.

Rollback

python -m pip install --upgrade "tokenpak==1.18.5"

The release introduces no destructive state migration. Artifact-level upgrade and rollback verification is part of the release gate.

Compatibility

  • The public API snapshot contains 4,663 symbols: 29 additive session-economics exports and zero removals relative to v1.18.5.
  • Existing proxy request and response routes remain compatible. The new endpoint and Python contract are additive, and unknown or insufficient facts produce explicit learning, unavailable, or error states instead of fabricated numeric values.