TokenPak v1.9.0 — TokenPak releases

Minor release: a guided onboarding & lifecycle pass on the CLI. Additive to existing behavior; no breaking changes to current workflows.

Added — onboarding & lifecycle

cli: tokenpak uninstall with --soft (remove TokenPak config/state, leave integrations) and --hard (full removal of TokenPak-owned state).
cli (doctor): tokenpak doctor --lifecycle summary plus a default route-state line so the proxy/route status is visible at a glance.
cli (permissions): permission-tier system for tokenpak integrate (strict / standard / auto, plus a multi-instance fleet tier) with a new tokenpak permissions verb, doctor rows, and an opt-in multi-instance fleet launcher mode — see tokenpak permissions --help.
config: read-only tokenpak config doctor diagnostics and a tokenpak config env provenance view (loaded env vars + precedence), with a documented config load-order spec.
cli (integrate): guided cross-shell tokenpak integrate flow with --apply / --revert, shell detection, and a --no-tui escape hatch for non-interactive use.
cli (menu): hardened interactive menu renderer (single alternate-screen session, per-command lifecycle, cached non-blocking status strip; honest unknown metrics render as a dash, never a fabricated $0.00).

Added

TokenPak Dispatch (v0.1-alpha, preview) — scoped, station-based, resumable work packages with a Decision Inbox and delivery receipts. Available on the main branch only; not yet included in a released pip install tokenpak package. See tokenpak dispatch --help.

Breaking — install footprint: heavy extras are now opt-in

Background: pip install tokenpak previously pulled ~5 GB of CUDA/ML wheels (torch, nvidia/*, transformers, sentence-transformers, scipy, tree-sitter-languages, pandas, litellm, llmlingua) as hard runtime dependencies. This made first-run installs impractical on machines without CUDA or a fast connection.

What changed: the six heavy packages listed below have been moved from [project.dependencies] to named [project.optional-dependencies] extras. The runtime behaviour is unchanged — every import site was already guarded with try/except ImportError before this release. Only the install metadata changed.

Migration: if your code uses any of the features below, add the corresponding extra to your install command:

Feature	Add to install command
Semantic search / vector embeddings (sentence-transformers)	`pip install tokenpak[retrieval]`
Tree-sitter code parsing	`pip install tokenpak[code-compression]`
A/B testing optimizer (scipy)	`pip install tokenpak[intelligence]`
Pandas data utilities	`pip install tokenpak[data]`
LLMLingua prompt compression	`pip install tokenpak[compression]`
LiteLLM Router integration	`pip install tokenpak[integrations-litellm]`
Everything (previous default)	`pip install tokenpak[full]`

If you previously ran pip install tokenpak and relied on retrieval / code-compression / intelligence / compression / integrations-litellm features, you must add the extra to your install. Features that use the guarded import will raise a clear ImportError with the correct pip install command if the extra is absent.

Slim install target: pip install tokenpak on a clean machine resolves in under 30 seconds and uses under 200 MB of disk. The [full] extra restores the previous behaviour for users who want everything.

Added — install footprint extras split

Named extras: tokenpak[retrieval], tokenpak[code-compression], tokenpak[intelligence], tokenpak[data], tokenpak[compression], tokenpak[integrations-litellm], tokenpak[full].
CI: slim-install smoke test — installs tokenpak with no extras, asserts venv site-packages < 200 MB, runs python -c "import tokenpak; from tokenpak.proxy import client".
CI: full-install matrix — pip install -e.[full,dev] + full test suite.
tests/test_dependencies_extras.py — slim-core invariant gate.
tests/test_extras_import_guard.py — lightweight post-demotion gate that asserts each heavy package is absent from [project.dependencies] and smoke-tests each guarded import path.

Changed — import error messages

tokenpak/integrations/litellm/proxy.py — error message updated to suggest pip install tokenpak[integrations-litellm] instead of bare pip install litellm.