v0.2.0 Linux desktop — release notes draft and build recipe

This is the maintainer packet for the Linux half of v0.2.0. It does not publish anything. Windows (WSL) and macOS images are built on those hosts and attached later; one annotated tag and one draft release cover all three.

Release notes draft (v0.2.0)

OpenAI4S 0.2.0 is the first release that ships a Linux desktop bundle next to the existing macOS image and the Windows/WSL launcher.

Downloads (once published)

Asset	What it is
`OpenAI4S-0.2.0-linux-x86_64.tar.gz`	Relocatable Linux desktop bundle: embedded CPython 3.13, pre-baked science stack, loose source tree, `./OpenAI4S` launcher, `.desktop` template, per-user `install.sh`. Not an AppImage and not a `.deb` — a FUSE squashfs would nest user namespaces inside a FUSE mount and fail at cell-execution time.
`OpenAI4S-0.2.0-macos-arm64.dmg`	Apple Silicon app (built on macOS). Ad-hoc signed unless a Developer ID identity is configured; notarization is still unreachable, so `--mode release` still refuses the DMG.
`OpenAI4S-0.2.0-windows-x86_64.zip`	Windows launcher around the exact Linux tarball above. Requires WSL2. Built from the Linux artifact, not a second compile.
`openai4s-0.2.0-py3-none-any.whl` / `openai4s-0.2.0.tar.gz`	Zero-dependency wheel and sdist for any supported platform that already has Python ≥ 3.10.

What landed since v0.1.0

Linux and Windows desktop packages — same embedded interpreter and science stack as the macOS image; Linux unpacks anywhere and runs ./OpenAI4S.
Read-only session sharing — openai4s share / openai4s relay over an outbound tunnel you operate.
Seven source-attributed public-database connectors — UniProt, RCSB PDB, Ensembl, ChEMBL, PubChem, arXiv, OpenAlex.
Versioned /api/v1 — keyset pagination, one error envelope, resumable WebSocket cursor.
Environments as a transaction — openai4s env plan|apply|list|rollback|recover.
Support surfaces — redacted openai4s doctor / openai4s diagnostics, consent-gated revocable telemetry.
Benchmark — workflows against the real Store, kernels, and dispatcher.
CLI --version — prints openai4s 0.2.0 without requiring a subcommand.

Linux install (end user)

tar -xzf OpenAI4S-0.2.0-linux-x86_64.tar.gz
cd OpenAI4S-0.2.0-linux-x86_64
./OpenAI4S                 # daemon + browser at http://127.0.0.1:8760/
./install.sh               # optional: CLI on PATH + application-menu entry
# recommended:
#   Debian/Ubuntu: sudo apt install bubblewrap
#   Fedora:        sudo dnf install bubblewrap
#   Arch:          sudo pacman -S bubblewrap

Headless: ./bin/openai4s serve --no-open. Data lives in ~/.openai4s. The R kernel is not bundled; add it with ./bin/openai4s setup after installing micromamba/mamba/conda. Only x86_64 is published; arm64 Linux installs from PyPI (pip install openai4s).

Known limitations to state in the public notes

The Linux sandbox tier is beta. The full filesystem-and-egress bubblewrap smoke is still a manual run (OPENAI4S_KERNEL_SANDBOX=enforce uv run python -m harness.smoke.linux_sandbox). CI proves the private-PID interrupt path.
OPENAI4S_REQUIRE_TOKEN=0 remains a loopback-only escape hatch until 0.3.0 (LEGACY_TOKEN_OPT_OUT_REMOVED_IN). Default is still token-required.
macOS --mode release still cannot pass: Developer ID + notarization is unreachable in this tree. Do not flip the GitHub release public until the macOS owner decides how to handle that gate.
Do not tag, gh release create, or publish to PyPI from a packaging branch. One annotated v0.2.0 on main after all three platforms are ready.

Reproducible Linux build

Run this on Linux x86_64 (native). The script can cross-build from macOS, but verify_linux_bundle.py's import probe only executes on a matching Linux host.

System packages

Debian/Ubuntu (this is what the Linux packaging VM used):

sudo apt-get update
sudo apt-get install -y --no-install-recommends \
  ca-certificates curl tar gzip rsync coreutils
# optional, for doctor/sandbox smoke on the build host:
# sudo apt-get install -y bubblewrap

Install uv if it is not already on PATH:

curl -LsSf https://astral.sh/uv/install.sh | sh
# then ensure ~/.local/bin is on PATH

Fedora: sudo dnf install tar gzip rsync curl. Arch: sudo pacman -S tar gzip rsync curl.

No extra compilers are required. The bundle embeds a relocatable CPython from uv python install and only manylinux wheels (scripts/bundled_packages.txt). Pillow is pulled ephemerally by uv run --no-project --with pillow when the icon ladder is sliced.

Commands

From a checkout whose [project] version and openai4s.__version__ are both 0.2.0:

# 0) identity
git rev-parse HEAD
uv run --locked python scripts/verify_release_tag.py v0.2.0

# 1) lightweight control env (pytest / mypy / uv build)
uv sync --locked --extra science

# 2) wheel + sdist
uv run --locked python scripts/source_secret_scan.py
uv build --no-sources --out-dir dist --clear
uv run --locked python scripts/verify_release_artifacts.py dist

# 3) Linux desktop tarball (slow: downloads CPython 3.13 + science wheels)
bash scripts/build_linux_bundle.sh
python3 scripts/verify_linux_bundle.py dist/OpenAI4S-0.2.0-linux-x86_64.tar.gz

# 4) optional receipt, same shape the release job writes
uv run --locked python scripts/release_receipts.py \
  --kind linux --source-sha "$(git rev-parse HEAD)" \
  dist/OpenAI4S-0.2.0-linux-x86_64.tar.gz

On macOS, step 3 is bash scripts/build_macos_dmg.sh then python3 scripts/verify_macos_bundle.py dist/OpenAI4S-*.dmg. On a machine that already has the Linux tarball, the Windows zip is bash scripts/build_windows_zip.sh (it wraps that exact .tar.gz).

Smoke the unpacked Linux bundle

tar -xzf dist/OpenAI4S-0.2.0-linux-x86_64.tar.gz -C /tmp
APP=/tmp/OpenAI4S-0.2.0-linux-x86_64
test "$(cat "$APP/VERSION")" = "0.2.0"
"$APP/bin/openai4s" --version          # openai4s 0.2.0
"$APP/bin/openai4s" doctor
OPENAI4S_DATA_DIR=/tmp/openai4s-smoke \
  "$APP/bin/openai4s" serve --host 127.0.0.1 --port 8760 --no-open &
# wait for listen, then:
curl -fsS http://127.0.0.1:8760/ >/tmp/index.html
"$APP/bin/openai4s" stop

The daemon mints a loopback access token by default. curl of / without it is expected to be 401 or the HTML recovery page; the workbench document is still served to a browser that presents the token. For an unauthenticated liveness check use /health if the gate allows it, otherwise open the URL printed on stderr (http://localhost:8760/?token=…).

Offline suite on the build host

OPENAI4S_SECRET_STORE=plaintext \
  uv run pytest -n auto --maxprocesses=4 --dist loadfile
uv run mypy

One artifact-capture test (test_capture_detects_same_length_rewrite_that_restores_mtime) can fail on kernels whose ctime resolution is coarse. That is a host filesystem artifact, not a product bug — do not "fix" it in the tree.