upload-sarif: relative paths beginning with `./` are not handled correctly

upload-sarif エントリーポイントから始め、リンク先の SARIF パスに関するガイダンスと RFC 3986 の文脈を使って、foo/bar と ./foo/bar の報告された差異を再現します。関連する実装がこのリポジトリにあるのか、GitHub の SARIF 取り込み機構にあるのかを調査します。相対参照がリポジトリのファイルに解決され、プレビューが表示され、pull request に一貫して現れることが完了の条件です。

Hello!

I'm reporting what I believe to be a bug in GitHub's SARIF consumption. This bug is probably happening somewhere deeper in GitHub's SARIF consumption machinery but upload-sarif is the main user-facing entrypoint for that machinery, so I'm filing the report here. Please let me know if a better discussion venue exists and I'd be happy to continue this elsewhere 🙂

Description

GitHub's documentation says the following about paths in SARIF inputs:

Code scanning interprets results that are reported with relative paths as relative to the root of the repository analyzed. If a result contains an absolute URI, the URI is converted to a relative URI. The relative URI can then be matched against a file committed to the repository.

Ref: https://docs.github.com/en/code-security/code-scanning/integrating-with-code-scanning/sarif-support-for-code-scanning#specifying-the-root-for-source-files

This, combined with the fact that SARIF stipulates RFC 3986 for artifactLocation URIs suggests that relative paths should be able to include . and .., as these path components are well-defined in RFC 3986.

However, if a SARIF file contains a relative URI like ./foo/bar, GitHub appears to fail to resolve that URL to a path in the repository being scanned. This results in suboptimal finding presentation:

1. Findings are not rendered on PRs, since GitHub doesn't think the finding URLs match repository paths;
2. Findings are rendered under "Code scanning alerts" but show "Preview unavailable" instead of a proper synopsis/extraction of the source file.

Here is a screenshot of the above behavior, demonstrating the "Preview unavailable" behavior because the path (./tests/integration/test-data/issue-612-repro/action.yml) starts with a ./:

For contrast, here's a similar finding rendering correctly because it doesn't begin with ./:

Expected behavior

I expect GitHub's SARIF ingestion to handle URIs that begin with (or contain) relative references, since SARIF stipulates RFC 3986 for URIs and RFC 3986 permits relative references. More generally, I believe many code-scanning tools produce relative references (like ./foo) by default and would benefit from not having to specialize their handling for GitHub's SARIF consumer.

Actual behavior

Relative URIs like foo/bar work correctly (resolving relative to the repository root), while relative URIs like ./foo/bar do not work correctly.

Workarounds

The primary workarounds here are:

1. "Normalize" relative URIs from ./foo/bar to foo/bar. This is possible, but non-trivial in the general case (e.g. ./foo/bar/../baz), and requires more pre-processing on the generated SARIF than the specification stipulates.
2. Switch entirely to absolute URIs, and use SARIF's features (like invocations[0].workingDirectory.uri) to help GitHub ingest and transform paths into appropriate relative URIs. I've had limited success making this work: the SARIF ingestor appears to be very fickle about absolute paths.

Overall, I think both of these workarounds are non-ideal, and add additional burden to SARIF producers to pre-process their inputs/offer a GitHub specific "quirks mode" beyond what SARIF stipulates. I think it would be fantastic if GitHub could instead support these kinds of relative paths!

Additional context

I ran into this behavior within zizmor, which supports SARIF as an output format and encourages people to use it when integrating with GitHub. Some original issue context: https://github.com/woodruffw/zizmor/issues/604, https://github.com/woodruffw/zizmor/issues/571.

Additionally, I've observed that GitHub's behavior around equivalent relative paths (e.g. foo/bar and ./foo/bar) is somewhat mixed: ./foo/bar doesn't produce a preview, but results for foo/bar and ./foo/bar are deduplicated against each other. That makes me think that some degree of normalization/equivalence checking is happening in GitHub's SARIF processing, just not at the point needed to handle previews correctly.

Please let me know if there's any other information I can provide!

CC @kommendorkapten