uv（紫外线）

uv Package Manager

Prefer uv over pip/poetry for Python dependency and project management. Use the workflow below and choose the right mode (project, script, or tool).

Mode Decision

User needs to...
│
├─ Repo uses poetry / pipenv / conda (poetry.lock, Pipfile, environment.yml)
│  → Ask: "This repo uses [poetry/pipenv/conda]. Do you want to switch to uv?"
│  → If yes: prefer the migrate-to-uv tool (uvx migrate-to-uv in project root) to convert metadata and lockfile; then uv sync. Remove legacy files only after user confirms.
│  → If no: do not use uv for project management; use the existing tool or uv pip only if appropriate.
│
├─ Repo has requirements.txt (no pyproject.toml)
│  → Ask: "This repo uses requirements.txt. Do you want to convert to uv (pyproject.toml + uv.lock)?"
│  → If yes: guide conversion (uv init, uv add from requirements, then uv sync)
│  → If no: use Pip-Compatible workflow (uv pip)
│
├─ Manage a project (pyproject.toml, lockfile, team repo)
│  → Use PROJECT workflow (uv init, uv add, uv sync, uv run)
│
├─ Run a single script with dependencies
│  → Use SCRIPT workflow (uv add --script, uv run <script>)
│
├─ Run a one-off CLI tool (no project)
│  → Use uvx: uvx <package> [args] or uv tool install <package>
│
└─ User declined conversion; existing pip/requirements workflow
   → Use uv pip (uv venv, uv pip sync, uv pip compile)

Project Workflow

Use for apps, libraries, or any repo with pyproject.toml.

New project

uv init [project-name]
cd [project-name]
uv add <package> [package2...]
uv sync
uv run python main.py
# or: uv run <any command>

Existing project (already has pyproject.toml)

uv sync                    # install from lockfile (or resolve and lock)
uv add <package>           # add dependency, update lockfile and env
uv remove <package>       # remove dependency
uv run <command>          # run in project venv (creates .venv if needed)
uv lock                   # refresh lockfile only

Pin Python version (optional)

uv python pin 3.11        # writes .python-version
uv python install 3.11    # ensure that version is available

Rules:

• Use uv add for project dependencies; avoid editing pyproject.toml by hand for deps when uv can do it.
• After changing dependencies, run uv sync (or rely on uv add/uv remove which update the env).
• Run project commands via uv run so the correct venv and env are used; do not assume pip install or manual activate.
• When creating a new project, ensure .venv is in .gitignore (uv init usually adds it; add it if missing).
• Treat uv.lock as a mandatory source-controlled artifact: commit it so all environments and CI use the same dependency versions; the lockfile is universal (one file for Windows, macOS, Linux).

Script Workflow

For a single Python file that needs packages (no full project).

1. Add inline script metadata (or use uv add --script to add deps to the file):

# /// script
# requires-python = ">=3.10"
# dependencies = ["requests"]
# ///
import requests
print(requests.get("https://example.com"))

1. Add deps from CLI (updates the script file):

uv add --script example.py requests

1. Run with uv (creates an ephemeral env if needed):

uv run example.py

For executable scripts that run without typing uv run, use a shebang (PEP 723):

#!/usr/bin/env -S uv run --script
# /// script
# requires-python = ">=3.10"
# dependencies = ["requests"]
# ///

Rules:

• Use uv add --script <file> to declare dependencies for that script.
• Use uv run <script>.py to run it; do not tell the user to pip install first.
• For portability, scripts can use shebang #!/usr/bin/env -S uv run --script.

Tools (one-off CLI from PyPI)

Run without installing globally:

uvx <package> [args]
# e.g. uvx ruff check .
# e.g. uvx pycowsay "hello"

Install the tool for repeated use:

uv tool install <package>
# e.g. uv tool install ruff

Rules:

• Prefer uvx for one-off runs; use uv tool install when the user will call the tool often.

Repos Using requirements.txt

When the repo has requirements.txt (and no pyproject.toml):

1. Ask the user: "This repo uses requirements.txt. Do you want to convert to uv (pyproject.toml + uv.lock)?"
2. If yes: Guide conversion: uv init, then uv add -r requirements.txt (or, if you have requirements.in and locked requirements.txt, uv add -r requirements.in -c requirements.txt to preserve versions), then uv sync. Optionally remove or keep requirements.txt per user preference.
3. If no: Use the Pip-Compatible workflow below (no conversion).

Do not convert to uv without asking when the repo currently uses only requirements.txt.

Pip-Compatible Workflow

For repos that still use requirements.txt or pip (and user did not want conversion):

uv venv                           # create .venv
uv pip sync requirements.txt      # install from locked requirements
uv pip compile requirements.in -o requirements.txt  # compile/lock

Use uv pip instead of pip for the same commands when you want speed and better resolution.

CI and Docker

When the user asks about uv in CI (e.g. GitHub Actions) or uv in Docker, advise using the patterns below and point to references/docker-and-ci.md for full detail.

CI (e.g. GitHub Actions):

• Install uv with the official astral-sh/setup-uv action; use cache keyed by uv.lock and pyproject.toml.
• Run uv sync --locked so the job fails if the lockfile is out of sync (no silent deploy of untested deps).
• Optionally run uv cache prune --ci at end of job to keep cache lean.

Docker:

• Use a multi-stage build: builder stage with ghcr.io/astral-sh/uv, copy only uv.lock and pyproject.toml first, run uv sync (or equivalent), then copy .venv and app code into a slim runtime image (no uv/Rust in final image).
• Set UV_COMPILE_BYTECODE=1 and UV_LINK_MODE=copy in the build stage; use --no-editable when syncing so the final image doesn’t depend on source.
• Run the container as a non-root user when possible.

For step-by-step Dockerfiles and CI workflow examples, see references/docker-and-ci.md.

IDE Setup (Use uv.venv)

After uv creates or uses a project .venv, automatically try to configure the IDE to use that interpreter so run/debug and IntelliSense use the same environment.

1. Target: Workspace settings (project-scoped, shareable): .vscode/settings.json
2. Setting: python.defaultInterpreterPath pointing at the project’s .venv:

- Linux/macOS: "${workspaceFolder}/.venv/bin/python" - Windows: "${workspaceFolder}/.venv/Scripts/python.exe"

1. Behavior: If .vscode/settings.json exists, read it, add or update only python.defaultInterpreterPath, then write back. If it doesn’t exist, create .vscode/ and the file with this setting. Preserve all other keys.
2. When: After uv init, uv sync, or conversion from requirements.txt that creates/updates .venv.

Rules:

• Use workspace settings (.vscode/settings.json), not user settings, so the choice is per-project and can be committed.
• Do not overwrite or remove other settings in the file.
• If the workspace already has a Python interpreter selected, still add/update this key so the project’s .venv is the default.

Automation and Agent Behavior

1. New Python project: Run uv init (or uv init <name>), then uv add for initial deps, then uv sync. Suggest uv run for run/test commands.
2. Add dependency: Use uv add <package>. For dev/optional: uv add --dev <package>.
3. Run something: Use uv run <command> in a project; uv run script.py for scripts; uvx <tool> for one-off tools.
4. No manual venv activate: Prefer uv run so the agent and user don’t depend on source.venv/bin/activate.
5. Lockfile: Commit uv.lock; treat it as mandatory for parity. After pull or after editing deps, run uv sync. In CI, use uv sync --locked so the job fails if the lockfile is out of sync with pyproject.toml.
6. Detecting uv: If pyproject.toml exists and there is no poetry/pipenv/conda (no poetry.lock, Pipfile, environment.yml), assume uv is allowed and suggest uv commands. If the user said "use uv", always prefer uv over pip/poetry. If poetry/pipenv/conda is present, ask before switching to uv (see Mode Decision).
7. requirements.txt only: If the repo has requirements.txt but no pyproject.toml, ask whether the user wants to convert to uv before converting or using uv pip.
8. IDE interpreter: After uv creates or syncs .venv, try to set the workspace Python interpreter by adding or updating python.defaultInterpreterPath in .vscode/settings.json (see IDE Setup above).

Common Commands Reference

Additional Resources

• Full CLI and options: references/cli-and-commands.md
• Example flows: references/examples.md
• Common issues and fixes: references/troubleshooting.md
• Docker and CI: references/docker-and-ci.md
• Workspaces (monorepos) and resolution: references/workspaces-and-resolution.md