Internal installation¶
The supported product is a source checkout opened in Claude Code or Codex. Its Python package is studio/, with command shims under bin/. Keep the checkout intact: doctrine, exported context and a few existing engine helpers resolve relative to it. This internal release does not publish a standalone wheel or ask users to pip install the repository. Public distribution/extraction remains later work.
The internal Python baseline is CPython 3.13 or newer. The verified environment for this branch is CPython 3.13.14, Windows 11 build 26200, AMD64. Other Python versions and macOS/Linux are not certified by that Windows run; the portable code paths still need their own acceptance runs. Git and one supported session client are required. Provider accounts, Node, Pillow, app servers and GPU software are optional for setup/planning and supplied-media decisions.
From the repository root:
python --version
python -m venv .venv
.venv/Scripts/python.exe bin/studio_admin.py --help
On a POSIX shell the virtual-environment executable is .venv/bin/python. Use that environment's Python consistently, including in a client terminal. Core commands need no pip installation: requirements/core.txt is deliberately empty. Run from the checkout so shared instructions and doctrine are available. Store production workspaces outside Git; the internal onboarding/setup commands document their selection.
Optional capabilities¶
Install only the profile needed for the chosen workflow:
| Capability | Installation / prerequisite | Core impact when absent |
|---|---|---|
| Identity measurement | python -m pip install -r requirements/identity.txt |
Instrument unavailable; no fabricated identity score or automatic approval |
| Video/audio probe and export | ffmpeg and ffprobe on PATH | Planning, metadata/decisions and byte landing still work; encoding/probing need the executables |
| Tests | python -m pip install -r requirements/test.txt |
User commands are unaffected |
These optional profiles declare package names, not a frozen environment. The clean-core acceptance does not certify a newly resolved optional ML or SDK stack. Keep the tested environment's package lock separately when deploying a particular optional instrument; do not add its packages to the core profile. Model weights and provider credentials are configured explicitly after choosing that capability, never downloaded or tested through a generation during setup.
bin/identity_drift.py --help works without NumPy. A measurement without the required stack reports the missing capability and profile install command. Its weights remain under the explicitly configured STUDIO_FACE_MODEL_DIR (the old machine-specific fallback is retained for compatibility); diagnostics must identify missing weights rather than silently fetching them.
The minimal-environment and import-boundary checks are in tests/studio/test_minimal_installation.py and test_boundary.py. Run python bin/check_session_first.py for the full supplied-media lifecycle acceptance. The actual teammate/client pilot remains a separate check; passing imports or fixture operations alone is not adoption evidence.