Uv: Running a script with dependencies

Original Article ↗ Hacker News Discussion ↗

Enthusiasm for uv’s script mode

  • Many commenters call uv run --script a “killer feature” that revived their use of Python for one-off or small tools, especially in git hooks and ad‑hoc scripts.
  • Inline dependency blocks (PEP 723) are praised for letting a single .py file be self-contained: just ship the script, install uv, and run.
  • The shebang pattern #!/usr/bin/env -S uv run --script is widely used to make Python scripts feel like regular executables.

Shebang env -S portability discussion

  • Long subthread explaining why -S is needed: historically, different Unix variants handled argument splitting in shebangs inconsistently.
  • Modern BSDs, macOS, and recent coreutils support env -S for portable multi-arg shebangs; some systems (OpenBSD, NetBSD, Solaris, BusyBox) still don’t.
  • Consensus: using env -S is the safest cross-platform choice, even if it’s a no-op on some systems.

PEP 723, design trade-offs, and “magic comments”

  • Strong support for PEP 723 as a standard way to embed dependencies; several tools (uv, pipx, hatch, marimo, Jupyter kernels) now use it.
  • Debate over using “magic comments” vs real syntax:
    • Pro: tools don’t need a full Python parser; Python core stays packaging-agnostic; works across language changes.
    • Con: feels non-obvious as “syntax,” and people wish imports could carry version info directly.
  • Multiple replies explain why inferring deps from imports is fragile: import names don’t map cleanly to distribution names or versions.

Tooling comparisons and ecosystem impact

  • Some say uv halted plans to migrate scripts to Go, though Go binaries remain preferable for zero-runtime, airgapped use.
  • Conda is criticized as heavy; uv plus wheels is seen as enough for most, though some note conda is still useful for complex C/C++/GPU stacks.
  • uv is described as “pip + venv + pyenv + pipx” in one fast, coherent tool, leading some to wish it were the default Python toolchain.

Limitations, gotchas, and editor integration

  • Single-file only: multi-module projects still need pyproject.toml.
  • uv’s cache grows indefinitely; there’s uv cache clean, but no GC yet. Hard links reduce disk cost, but don’t help across filesystems.
  • Offline/infra scripts must pre-warm caches or avoid this pattern; relying on runtime downloads can fail when the network is down.
  • LSP/IDE integration is awkward: editors often don’t see uv’s transient venvs without manual configuration or helper scripts/extensions.
  • Specific pain points: PyTorch wheel variants, uv run project discovery from outside the project dir, and SCA tools missing inline deps.