[myl7]

Problem

getVersion() parses the cuda input with new SemVer(versionString), so any major.minor-only pin throws before installation starts:

##[error]TypeError: Invalid Version: 12.6

Strict semver requires all three components, but a major.minor pin ("install whatever patch of 12.6 is current") is a natural way to use this action, and it matches how the Linux network path actually installs (cuda-nvcc-<major>-<minor> via apt, with apt resolving the patch).

Change

• A full semver string keeps its exact behavior: install that version or fail with Version not available.
• A major.minor pin (12.6) resolves to the newest listed patch of that series (currently 12.6.3).
• A major-only pin (12) resolves to the newest listed release of that major series.
• Anything that is neither full semver nor a bare major(.minor) number still throws the same TypeError: Invalid Version: ....
• The cuda output now reports the resolved installed version instead of the raw input.

Jest covers pin resolution (major.minor and major-only), unavailable pins, and a full-but-unlisted version; dist is rebuilt.

Complements #448 (CUDA 13.4 links) — with both, cuda: 13.4 works end to end.

[Copilot]

🟡 Changes recommended

Major-only pins select the wrong minor series, and the action output reports the unresolved input.

2 open findings

• Major-only CUDA pins incorrectly select only version 11.0.x · New
• CUDA output reports requested version instead of resolved version · New

What changed in this PR

Adds partial CUDA version resolution to the action while preserving exact-version lookup for full semver inputs.

Changes:

• Resolves partial pins against available download versions.
• Adds regression tests and documents major.minor pins.
• Updates the shipped bundle.

File	Description
src/​version.ts	Parses partial pins and selects matching releases.
README.md	Documents version-pin behavior.
dist/​index.js	Includes the updated resolver.
tests/​version.test.ts	Tests pin resolution and unavailable versions.

🧠 Review effort: Balanced

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.