Pin an app to the fork
Pin a full GitHub commit SHA to install the same tested Turnspan code whenever you build your app. The installed package's source URL identifies the commit.
Lock the fork as an app dependency
Choose a version from the published releases
or a reviewed commit on the maintained branch. Open the commit on GitHub and
copy its full SHA. In your app's pyproject.toml, keep openai-chatkit in
the dependencies list. If you specify an exact version there, use the
version declared by the chosen commit. Add this uv source, replacing
COMMIT_SHA with the full SHA.
[tool.uv.sources]
openai-chatkit = { url = "https://github.com/michaelgiraldo/chatkit-python/archive/COMMIT_SHA.tar.gz" }
Choose extras for the pinned version. On the maintained branch, add agents
for chatkit.agents, fastapi for the FastAPI adapter, and templates for
loading exported .widget files. For example, use openai-chatkit[agents,fastapi]
for an Agents integration served by FastAPI.
Refresh the lock and install it with Python 3.14.
uv lock --python 3.14
uv sync --locked --python 3.14
The lockfile records the archive URL and its SHA-256 hash. Commit both
pyproject.toml and uv.lock in the app repository. Use uv sync --locked in
your build so it installs the reviewed dependency graph.
Verify the installed build
Replace COMMIT_SHA with the same SHA and run this from the app directory with
the interpreter that runs the app.
uv run --locked --no-sync python - <<'PY'
import json
from importlib.metadata import distribution
commit = "COMMIT_SHA"
expected = f"https://github.com/michaelgiraldo/chatkit-python/archive/{commit}.tar.gz"
package = distribution("openai-chatkit")
metadata = package.read_text("direct_url.json")
if metadata is None:
raise SystemExit("The installed package has no recorded direct source URL.")
installed = json.loads(metadata)["url"]
if installed != expected:
raise SystemExit(f"Expected {expected}, found {installed}.")
print(f"openai-chatkit {package.version} installed from {installed}")
PY
Verify what serves live traffic
- Follow the live route or alias to each serving revision. Record the resolved artifact, such as the image digest.
- Run the verification snippet with the app's Python interpreter inside that exact artifact. You can inspect a matching container image offline without starting the service. Compare the installed version and recorded source URL with the release you intended to deploy. If the package was installed from a wheel or an index, look for the source commit in the build provenance tied to the resolved artifact.
- Record the revision, artifact digest, package version, and source commit together. A lockfile records the intended build input. An image tag can move. The resolved artifact and its installed metadata establish what is serving traffic.